Add files using upload-large-folder tool

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-12ac3267-b673-44fd-8ea3-37e3e74cb0101755540018956-2025_08_18-20.00.27.475/source.csv

@@ -0,0 +1,566 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,788,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"8:00:27 PM [info] Activating crowd-code\n8:00:27 PM [info] Recording started\n8:00:27 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,1099,"TERMINAL",0,0,"python3",,terminal_focus
+4,1105,"TERMINAL",0,0,"bash",,terminal_focus
+5,2785,"test/test_nan.ipynb",0,0,"# Restore a dynamics checkpoint and enable sowing\nimport os\nfrom typing import Dict\n\nimport jax\nimport jax.numpy as jnp\nimport flax.nnx as nnx\nimport optax\nimport orbax.checkpoint as ocp\nimport grain\n\nfrom utils.dataloader import get_dataloader\nfrom models.lam import LatentActionModel\n\n# Adjust to your checkpoint directory, dataset directory, and dynamics type\nckpt_dir = ""/fast/project/HFMI_SynergyUnit/jafar_ws/checkpoints/coinrun/lam/train_lam_coinrun_reproduction_20067/100000_ckpt""\ndata_dir = ""/fast/project/HFMI_SynergyUnit/jafar_ws/data/coinrun/array_records_10m""\nnum_steps: int = 200_000\nseed: int = 0\nseq_len: int = 16\nimage_channels: int = 3\nimage_height: int = 64\nimage_width: int = 64\nsave_ckpt: bool = False\nrestore_ckpt: bool = False\n# Optimization\nbatch_size: int = 36\nvq_beta: float = 0.25\ninit_lr: float = 0.0\nmax_lr: float = 3e-5\ndecay_end: float = 0.0\nwsd_decay_steps: int = (\n    10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n)\nwarmup_steps: int = 5000\nlr_schedule: str = ""wsd""  # supported options: wsd, cos\nvq_reset_thresh: int = 50\n# LAM\nmodel_dim: int = 512\nffn_dim: int = 2048\nlatent_dim: int = 32\nnum_latents: int = 6\npatch_size: int = 16\nnum_blocks: int = 4\nnum_heads: int = 8\ndropout: float = 0.0\ncodebook_dropout: float = 0.0\nparam_dtype = jnp.float32\ndtype = jnp.bfloat16\n# Logging\nlog_interval: int = 5\nlog_image_interval: int = 250\nuse_flash_attention: bool = True\n\n# Build model graph matching the checkpoint\nrng = jax.random.key(seed)\nrng, _rng = jax.random.split(rng)\nrngs = nnx.Rngs(_rng)\nlam = LatentActionModel(\n    in_dim=image_channels,\n    model_dim=model_dim,\n    ffn_dim=ffn_dim,\n    latent_dim=latent_dim,\n    num_latents=num_latents,\n    patch_size=patch_size,\n    num_blocks=num_blocks,\n    num_heads=num_heads,\n    dropout=dropout,\n    codebook_dropout=codebook_dropout,\n    param_dtype=param_dtype,\n    dtype=dtype,\n    use_flash_attention=use_flash_attention,\n    rngs=rngs,\n)\n\n# Optimizer (matches training opt hyperparams; lr value is irrelevant for restore)\ntx = optax.adamw(\n    learning_rate=max_lr,\n    b1=0.9,\n    b2=0.9,\n    weight_decay=1e-4,\n    mu_dtype=param_dtype,\n)\noptimizer = nnx.Optimizer(lam, tx)\n",python,tab
+6,325278,"test/test_nan.ipynb",186,0,"",python,selection_mouse
+7,325279,"test/test_nan.ipynb",185,0,"",python,selection_command
+8,331804,"test/test_nan.ipynb",1007,8,"schedule",python,selection_command
+9,334786,"test/test_nan.ipynb",968,8,"schedule",python,selection_command
+10,338868,"test/test_nan.ipynb",1007,8,"schedule",python,selection_command
+11,341147,"test/test_nan.ipynb",1014,0,"",python,selection_command
+12,341595,"test/test_nan.ipynb",1004,0,"",python,selection_command
+13,343378,"test/test_nan.ipynb",0,0,"# Restore latest checkpoint: optimizer and dataloader state, like in training\nfrom typing import cast\n\nhandler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\nhandler_registry.add(\n    ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n)\nhandler_registry.add(\n    ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n)\nhandler_registry.add(\n    ""dataloader_state"",\n    grain.checkpoint.CheckpointSave,\n    cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n)\nhandler_registry.add(\n    ""dataloader_state"",\n    grain.checkpoint.CheckpointRestore,\n    cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n)\n\nckpt_mgr = ocp.CheckpointManager(\n    directory=ckpt_dir,\n    options=ocp.CheckpointManagerOptions(step_format_fixed_length=6),\n    handler_registry=handler_registry,\n)\n\n# Recreate dataloader and iterator exactly like training\narray_record_files = [\n    os.path.join(data_dir, x)\n    for x in os.listdir(data_dir)\n    if x.endswith("".array_record"")\n]\ngrain_dataloader = get_dataloader(\n    array_record_files,\n    seq_len,\n    batch_size,\n    image_height,\n    image_width,\n    image_channels,\n    num_workers=8,\n    prefetch_buffer_size=1,\n    seed=seed,\n)\ninitial_state = grain_dataloader._create_initial_state()\nloader_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n# Restore optimizer and dataloader iterator states\nabstract_optimizer = nnx.eval_shape(lambda: optimizer)\nabstract_optimizer_state = nnx.state(abstract_optimizer)\nrestored = ckpt_mgr.restore(\n    ckpt_mgr.latest_step(),\n    args=ocp.args.Composite(\n        model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n        dataloader_state=grain.checkpoint.CheckpointRestore(loader_iterator),  # type: ignore\n    ),\n)\n\nnnx.update(optimizer, restored[""model_state""])  # type: ignore\nloader_iterator = restored[""dataloader_state""]\nstep = ckpt_mgr.latest_step() or 0\nckpt_mgr.close()\n\n# Convenience handle\nlam = optimizer.model\nprint(f""Restored optimizer and dataloader at step {step}."")\n",python,tab
+14,347677,"test/test_nan.ipynb",0,0,"",python,tab
+15,348788,"test/test_nan.ipynb",2041,0,"",python,selection_command
+16,350687,"test/test_nan.ipynb",2059,0,"",python,selection_command
+17,350907,"test/test_nan.ipynb",2085,0,"",python,selection_command
+18,350927,"test/test_nan.ipynb",2097,0,"",python,selection_command
+19,350996,"test/test_nan.ipynb",2109,0,"",python,selection_command
+20,351167,"test/test_nan.ipynb",2132,0,"",python,selection_command
+21,351315,"test/test_nan.ipynb",2158,0,"",python,selection_command
+22,351707,"test/test_nan.ipynb",2160,0,"",python,selection_command
+23,352067,"test/test_nan.ipynb",2158,0,"",python,selection_command
+24,360027,"test/test_nan.ipynb",2081,0,"",python,selection_mouse
+25,361167,"test/test_nan.ipynb",2077,0,"",python,selection_command
+26,361336,"test/test_nan.ipynb",2077,6,"",python,content
+27,364567,"test/test_nan.ipynb",2076,0,"",python,selection_command
+28,366495,"test/test_nan.ipynb",2077,0,"max_lr",python,content
+29,366497,"test/test_nan.ipynb",2077,0,"",python,selection_command
+30,376216,"test/test_nan.ipynb",0,0,"%pwd",python,tab
+31,377207,"test/test_nan.ipynb",0,0,"",python,tab
+32,379207,"test/test_nan.ipynb",2057,0,"",python,selection_command
+33,380927,"test/test_nan.ipynb",2040,0,"\n",python,content
+34,381867,"test/test_nan.ipynb",2040,0,"\n",python,content
+35,384135,"test/test_nan.ipynb",2041,0,"l",python,content
+36,384136,"test/test_nan.ipynb",2042,0,"",python,selection_keyboard
+37,384176,"test/test_nan.ipynb",2042,0,"r",python,content
+38,384177,"test/test_nan.ipynb",2043,0,"",python,selection_keyboard
+39,384387,"test/test_nan.ipynb",2043,0,"_",python,content
+40,384388,"test/test_nan.ipynb",2044,0,"",python,selection_keyboard
+41,384707,"test/test_nan.ipynb",2044,0,"f",python,content
+42,384709,"test/test_nan.ipynb",2045,0,"",python,selection_keyboard
+43,384792,"test/test_nan.ipynb",2045,0,"n",python,content
+44,384793,"test/test_nan.ipynb",2046,0,"",python,selection_keyboard
+45,384911,"test/test_nan.ipynb",2046,0," ",python,content
+46,384912,"test/test_nan.ipynb",2047,0,"",python,selection_keyboard
+47,385055,"test/test_nan.ipynb",2047,0,"=",python,content
+48,385056,"test/test_nan.ipynb",2048,0,"",python,selection_keyboard
+49,385127,"test/test_nan.ipynb",2048,0," ",python,content
+50,385128,"test/test_nan.ipynb",2049,0,"",python,selection_keyboard
+51,386627,"test/test_nan.ipynb",2049,0,"g",python,content
+52,386628,"test/test_nan.ipynb",2050,0,"",python,selection_keyboard
+53,386707,"test/test_nan.ipynb",2050,0,"e",python,content
+54,386708,"test/test_nan.ipynb",2051,0,"",python,selection_keyboard
+55,386787,"test/test_nan.ipynb",2051,0,"t",python,content
+56,386788,"test/test_nan.ipynb",2052,0,"",python,selection_keyboard
+57,387024,"test/test_nan.ipynb",2052,0,"_",python,content
+58,387025,"test/test_nan.ipynb",2053,0,"",python,selection_keyboard
+59,387227,"test/test_nan.ipynb",2053,0,"l",python,content
+60,387228,"test/test_nan.ipynb",2054,0,"",python,selection_keyboard
+61,387302,"test/test_nan.ipynb",2054,0,"r",python,content
+62,387303,"test/test_nan.ipynb",2055,0,"",python,selection_keyboard
+63,387576,"test/test_nan.ipynb",2055,0,"_",python,content
+64,387577,"test/test_nan.ipynb",2056,0,"",python,selection_keyboard
+65,387696,"test/test_nan.ipynb",2056,0,"s",python,content
+66,387697,"test/test_nan.ipynb",2057,0,"",python,selection_keyboard
+67,387791,"test/test_nan.ipynb",2057,0,"c",python,content
+68,387792,"test/test_nan.ipynb",2058,0,"",python,selection_keyboard
+69,387847,"test/test_nan.ipynb",2058,0,"h",python,content
+70,387848,"test/test_nan.ipynb",2059,0,"",python,selection_keyboard
+71,388047,"test/test_nan.ipynb",2059,0,"e",python,content
+72,388048,"test/test_nan.ipynb",2060,0,"",python,selection_keyboard
+73,388049,"test/test_nan.ipynb",2060,0,"d",python,content
+74,388049,"test/test_nan.ipynb",2061,0,"",python,selection_keyboard
+75,388258,"test/test_nan.ipynb",2061,0,"u",python,content
+76,388258,"test/test_nan.ipynb",2062,0,"",python,selection_keyboard
+77,388259,"test/test_nan.ipynb",2062,0,"l",python,content
+78,388260,"test/test_nan.ipynb",2063,0,"",python,selection_keyboard
+79,388260,"test/test_nan.ipynb",2063,0,"e",python,content
+80,388261,"test/test_nan.ipynb",2064,0,"",python,selection_keyboard
+81,388447,"test/test_nan.ipynb",2064,0,"{}",python,content
+82,388448,"test/test_nan.ipynb",2065,0,"",python,selection_keyboard
+83,388795,"test/test_nan.ipynb",2064,2,"",python,content
+84,389027,"test/test_nan.ipynb",2064,0,"()",python,content
+85,389028,"test/test_nan.ipynb",2065,0,"",python,selection_keyboard
+86,389975,"test/test_nan.ipynb",2065,0,"\n    \n",python,content
+87,390187,"test/test_nan.ipynb",2070,0,"l",python,content
+88,390188,"test/test_nan.ipynb",2071,0,"",python,selection_keyboard
+89,390547,"test/test_nan.ipynb",2071,0,"r",python,content
+90,390548,"test/test_nan.ipynb",2072,0,"",python,selection_keyboard
+91,391007,"test/test_nan.ipynb",2072,0,"_",python,content
+92,391008,"test/test_nan.ipynb",2073,0,"",python,selection_keyboard
+93,391190,"test/test_nan.ipynb",2073,0,"s",python,content
+94,391191,"test/test_nan.ipynb",2074,0,"",python,selection_keyboard
+95,391192,"test/test_nan.ipynb",2074,0,"c",python,content
+96,391193,"test/test_nan.ipynb",2075,0,"",python,selection_keyboard
+97,391283,"test/test_nan.ipynb",2075,0,"h",python,content
+98,391284,"test/test_nan.ipynb",2076,0,"",python,selection_keyboard
+99,391387,"test/test_nan.ipynb",2076,0,"e",python,content
+100,391388,"test/test_nan.ipynb",2077,0,"",python,selection_keyboard
+101,391407,"test/test_nan.ipynb",2077,0,"d",python,content
+102,391408,"test/test_nan.ipynb",2078,0,"",python,selection_keyboard
+103,391535,"test/test_nan.ipynb",2078,0,"u",python,content
+104,391536,"test/test_nan.ipynb",2079,0,"",python,selection_keyboard
+105,391608,"test/test_nan.ipynb",2079,0,"l",python,content
+106,391609,"test/test_nan.ipynb",2080,0,"",python,selection_keyboard
+107,391695,"test/test_nan.ipynb",2080,0,"e",python,content
+108,391696,"test/test_nan.ipynb",2081,0,"",python,selection_keyboard
+109,391827,"test/test_nan.ipynb",2081,0,",",python,content
+110,391828,"test/test_nan.ipynb",2082,0,"",python,selection_keyboard
+111,392035,"test/test_nan.ipynb",2082,0,"\n    ",python,content
+112,394907,"test/test_nan.ipynb",2087,0,"i",python,content
+113,394908,"test/test_nan.ipynb",2088,0,"",python,selection_keyboard
+114,394947,"test/test_nan.ipynb",2088,0,"n",python,content
+115,394948,"test/test_nan.ipynb",2089,0,"",python,selection_keyboard
+116,395075,"test/test_nan.ipynb",2089,0,"i",python,content
+117,395076,"test/test_nan.ipynb",2090,0,"",python,selection_keyboard
+118,395187,"test/test_nan.ipynb",2090,0,"t",python,content
+119,395188,"test/test_nan.ipynb",2091,0,"",python,selection_keyboard
+120,395415,"test/test_nan.ipynb",2091,0,"_",python,content
+121,395416,"test/test_nan.ipynb",2092,0,"",python,selection_keyboard
+122,395607,"test/test_nan.ipynb",2092,0,"l",python,content
+123,395608,"test/test_nan.ipynb",2093,0,"",python,selection_keyboard
+124,395755,"test/test_nan.ipynb",2093,0,"r",python,content
+125,395756,"test/test_nan.ipynb",2094,0,"",python,selection_keyboard
+126,395835,"test/test_nan.ipynb",2094,0,",",python,content
+127,395836,"test/test_nan.ipynb",2095,0,"",python,selection_keyboard
+128,396307,"test/test_nan.ipynb",2095,0,"\n    ",python,content
+129,396576,"test/test_nan.ipynb",2100,0,"m",python,content
+130,396577,"test/test_nan.ipynb",2101,0,"",python,selection_keyboard
+131,396656,"test/test_nan.ipynb",2101,0,"a",python,content
+132,396657,"test/test_nan.ipynb",2102,0,"",python,selection_keyboard
+133,396829,"test/test_nan.ipynb",2102,0,"x",python,content
+134,396830,"test/test_nan.ipynb",2103,0,"",python,selection_keyboard
+135,396987,"test/test_nan.ipynb",2103,0,"_",python,content
+136,396988,"test/test_nan.ipynb",2104,0,"",python,selection_keyboard
+137,397223,"test/test_nan.ipynb",2104,0,"l",python,content
+138,397224,"test/test_nan.ipynb",2105,0,"",python,selection_keyboard
+139,397315,"test/test_nan.ipynb",2105,0,"r",python,content
+140,397316,"test/test_nan.ipynb",2106,0,"",python,selection_keyboard
+141,397415,"test/test_nan.ipynb",2106,0,",",python,content
+142,397416,"test/test_nan.ipynb",2107,0,"",python,selection_keyboard
+143,398196,"test/test_nan.ipynb",2107,0,"\n    ",python,content
+144,398675,"test/test_nan.ipynb",2112,0,"d",python,content
+145,398676,"test/test_nan.ipynb",2113,0,"",python,selection_keyboard
+146,398835,"test/test_nan.ipynb",2113,0,"e",python,content
+147,398836,"test/test_nan.ipynb",2114,0,"",python,selection_keyboard
+148,398907,"test/test_nan.ipynb",2114,0,"c",python,content
+149,398908,"test/test_nan.ipynb",2115,0,"",python,selection_keyboard
+150,399015,"test/test_nan.ipynb",2115,0,"a",python,content
+151,399016,"test/test_nan.ipynb",2116,0,"",python,selection_keyboard
+152,399115,"test/test_nan.ipynb",2116,0,"y",python,content
+153,399116,"test/test_nan.ipynb",2117,0,"",python,selection_keyboard
+154,399356,"test/test_nan.ipynb",2117,0,"_",python,content
+155,399357,"test/test_nan.ipynb",2118,0,"",python,selection_keyboard
+156,399542,"test/test_nan.ipynb",2118,0,"e",python,content
+157,399543,"test/test_nan.ipynb",2119,0,"",python,selection_keyboard
+158,399667,"test/test_nan.ipynb",2119,0,"n",python,content
+159,399668,"test/test_nan.ipynb",2120,0,"",python,selection_keyboard
+160,399739,"test/test_nan.ipynb",2120,0,"d",python,content
+161,399740,"test/test_nan.ipynb",2121,0,"",python,selection_keyboard
+162,399867,"test/test_nan.ipynb",2121,0,",",python,content
+163,399868,"test/test_nan.ipynb",2122,0,"",python,selection_keyboard
+164,400175,"test/test_nan.ipynb",2122,0,"\n    ",python,content
+165,400447,"test/test_nan.ipynb",2127,0,"n",python,content
+166,400448,"test/test_nan.ipynb",2128,0,"",python,selection_keyboard
+167,400676,"test/test_nan.ipynb",2128,0,"u",python,content
+168,400677,"test/test_nan.ipynb",2129,0,"",python,selection_keyboard
+169,400943,"test/test_nan.ipynb",2129,0,"m",python,content
+170,400944,"test/test_nan.ipynb",2130,0,"",python,selection_keyboard
+171,401175,"test/test_nan.ipynb",2130,0,"_",python,content
+172,401175,"test/test_nan.ipynb",2131,0,"",python,selection_keyboard
+173,401267,"test/test_nan.ipynb",2131,0,"s",python,content
+174,401268,"test/test_nan.ipynb",2132,0,"",python,selection_keyboard
+175,401435,"test/test_nan.ipynb",2132,0,"t",python,content
+176,401436,"test/test_nan.ipynb",2133,0,"",python,selection_keyboard
+177,401507,"test/test_nan.ipynb",2133,0,"p",python,content
+178,401508,"test/test_nan.ipynb",2134,0,"",python,selection_keyboard
+179,401515,"test/test_nan.ipynb",2134,0,"e",python,content
+180,401515,"test/test_nan.ipynb",2135,0,"",python,selection_keyboard
+181,401767,"test/test_nan.ipynb",2135,0,"s",python,content
+182,401768,"test/test_nan.ipynb",2136,0,"",python,selection_keyboard
+183,402067,"test/test_nan.ipynb",2135,0,"",python,selection_command
+184,404783,"train_lam.py",0,0,"from dataclasses import dataclass, field\nimport os\nfrom typing import cast\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.lam import LatentActionModel\nfrom utils.dataloader import get_dataloader\nfrom utils.lr_utils import get_lr_schedule\nfrom utils.parameter_utils import count_parameters_by_component\n\njax.config.update(""jax_transfer_guard"", ""allow"")\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    vq_beta: float = 0.25\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    vq_reset_thresh: int = 50\n    # LAM\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 6\n    patch_size: int = 16\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.0\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_lam""\n    tags: list[str] = field(default_factory=lambda: [""lam""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_checkpoint_keep_period: int = 20000\n    wandb_id: str = """"\n    use_flash_attention: bool = True\n\n\nargs = tyro.cli(Args)\n\n\ndef lam_loss_fn(\n    model: LatentActionModel, inputs: dict\n) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n    # --- Compute loss ---\n    gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n    inputs[""videos""] = gt.astype(args.dtype)\n    model.train()\n    outputs = model(inputs, training=True)\n    outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n    gt_future_frames = gt[:, 1:]\n    mse = jnp.square(gt_future_frames - outputs[""recon""]).mean()\n    q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n    commitment_loss = jnp.square(\n        outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n    ).mean()\n    loss = mse + q_loss + args.vq_beta * commitment_loss\n\n    # --- Compute validation metrics ---\n    gt = gt_future_frames.clip(0, 1).reshape(-1, *gt_future_frames.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt, recon)).mean()\n    count_fn = jax.vmap(lambda i: (outputs[""indices""] == i).sum())\n    index_counts = count_fn(jnp.arange(args.num_latents))\n    metrics = dict(\n        loss=loss,\n        mse=mse,\n        q_loss=q_loss,\n        commitment_loss=commitment_loss,\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage=(index_counts != 0).mean(),\n    )\n    return loss, (outputs[""recon""], index_counts, metrics)\n\n\n@nnx.jit\ndef train_step(\n    lam: LatentActionModel,\n    optimizer: nnx.Optimizer,\n    inputs: dict,\n    action_last_active: jax.Array,\n    rng: jax.Array,\n) -> tuple[jax.Array, jax.Array, jax.Array, dict]:\n    def loss_fn(\n        model: LatentActionModel,\n    ) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n        return lam_loss_fn(model, inputs)\n\n    # --- Update model ---\n    (loss, (recon, idx_counts, metrics)), grads = nnx.value_and_grad(\n        loss_fn, has_aux=True\n    )(lam)\n    optimizer.update(grads)\n\n    # --- Reset inactive latent actions ---\n    codebook = lam.vq.codebook\n    num_codes = len(codebook)\n    active_codes = idx_counts != 0.0\n    action_last_active = jnp.where(active_codes, 0, action_last_active + 1)\n    p_code = active_codes / active_codes.sum()\n    reset_idxs = jax.random.choice(rng, num_codes, shape=(num_codes,), p=p_code)\n    do_reset = action_last_active >= args.vq_reset_thresh\n    new_codebook = jnp.where(\n        jnp.expand_dims(do_reset, -1), codebook[reset_idxs], codebook.value\n    )\n    lam.vq.codebook.value = new_codebook\n    action_last_active = jnp.where(do_reset, 0, action_last_active)\n    return loss, recon, action_last_active, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    lam = LatentActionModel(\n        in_dim=args.image_channels,\n        model_dim=args.model_dim,\n        ffn_dim=args.ffn_dim,\n        latent_dim=args.latent_dim,\n        num_latents=args.num_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.num_blocks,\n        num_heads=args.num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.codebook_dropout,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        rngs=rngs,\n    )\n\n    # Count parameters\n    _, params, _ = nnx.split(lam, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype, # moments in full precision\n    )\n    optimizer = nnx.Optimizer(lam, tx)\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n    # --- Initialize checkpoint manager ---\n    step = 0\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    # --- Create DataLoaderIterator from dataloader ---\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n    # --- Restore checkpoint ---\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    print(f""Starting training from step {step}..."")\n    action_last_active = jnp.zeros(args.num_latents, dtype=jnp.int32)\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng = jax.random.split(rng)\n\n            inputs = dict(videos=videos, rng=_rng)\n            rng, _rng = jax.random.split(rng)\n            loss, recon, action_last_active, metrics = train_step(\n                lam, optimizer, inputs, action_last_active, _rng\n            )\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0, 1:].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n",python,tab
+185,409447,"train_lam.py",502,0,"",python,selection_command
+186,409727,"train_lam.py",6619,0,"",python,selection_command
+187,409907,"train_lam.py",502,0,"",python,selection_command
+188,410427,"train_lam.py",6619,0,"",python,selection_command
+189,411011,"train_lam.py",6601,34,"    lr_schedule = get_lr_schedule(",python,selection_command
+190,411091,"train_lam.py",6601,60,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,",python,selection_command
+191,411349,"train_lam.py",6601,82,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,",python,selection_command
+192,411371,"train_lam.py",6601,103,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,",python,selection_command
+193,411412,"train_lam.py",6601,127,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,",python,selection_command
+194,411451,"train_lam.py",6601,151,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,",python,selection_command
+195,411467,"train_lam.py",6601,178,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,",python,selection_command
+196,411491,"train_lam.py",6601,208,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,",python,selection_command
+197,411651,"train_lam.py",6601,214,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )",python,selection_command
+198,412171,"train_lam.py",6601,208,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,",python,selection_command
+199,412387,"train_lam.py",6601,214,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )",python,selection_command
+200,412987,"train_lam.py",6601,0,"",python,selection_command
+201,413529,"test/test_nan.ipynb",0,0,"",python,tab
+202,416723,"test/test_nan.ipynb",2075,0,"",python,selection_mouse
+203,416847,"test/test_nan.ipynb",2070,11,"lr_schedule",python,selection_mouse
+204,416967,"test/test_nan.ipynb",2066,17,"    lr_schedule,\n",python,selection_mouse
+205,417544,"test/test_nan.ipynb",2048,0,"",python,selection_mouse
+206,417727,"test/test_nan.ipynb",2048,1," ",python,selection_mouse
+207,417827,"test/test_nan.ipynb",2041,25,"lr_fn = get_lr_schedule(\n",python,selection_mouse
+208,417967,"test/test_nan.ipynb",2041,42,"lr_fn = get_lr_schedule(\n    lr_schedule,\n",python,selection_mouse
+209,418007,"test/test_nan.ipynb",2041,55,"lr_fn = get_lr_schedule(\n    lr_schedule,\n    init_lr,\n",python,selection_mouse
+210,418047,"test/test_nan.ipynb",2041,67,"lr_fn = get_lr_schedule(\n    lr_schedule,\n    init_lr,\n    max_lr,\n",python,selection_mouse
+211,418055,"test/test_nan.ipynb",2041,82,"lr_fn = get_lr_schedule(\n    lr_schedule,\n    init_lr,\n    max_lr,\n    decay_end,\n",python,selection_mouse
+212,418155,"test/test_nan.ipynb",2041,96,"lr_fn = get_lr_schedule(\n    lr_schedule,\n    init_lr,\n    max_lr,\n    decay_end,\n    num_stpes\n",python,selection_mouse
+213,418407,"test/test_nan.ipynb",2041,98,"lr_fn = get_lr_schedule(\n    lr_schedule,\n    init_lr,\n    max_lr,\n    decay_end,\n    num_stpes\n)\n",python,selection_mouse
+214,419235,"test/test_nan.ipynb",2041,97,"\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n",python,content
+215,419247,"test/test_nan.ipynb",2046,0,"",python,selection_command
+216,420707,"test/test_nan.ipynb",2042,34,"    lr_schedule = get_lr_schedule(",python,selection_command
+217,420827,"test/test_nan.ipynb",2042,60,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,",python,selection_command
+218,421087,"test/test_nan.ipynb",2042,82,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,",python,selection_command
+219,421107,"test/test_nan.ipynb",2042,103,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,",python,selection_command
+220,421147,"test/test_nan.ipynb",2042,127,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,",python,selection_command
+221,421167,"test/test_nan.ipynb",2042,151,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,",python,selection_command
+222,421207,"test/test_nan.ipynb",2042,178,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,",python,selection_command
+223,421227,"test/test_nan.ipynb",2042,208,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,",python,selection_command
+224,421375,"test/test_nan.ipynb",2042,214,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )",python,selection_command
+225,421543,"test/test_nan.ipynb",2042,215,"    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n",python,selection_command
+226,421775,"test/test_nan.ipynb",2251,4,"",python,content
+227,421775,"test/test_nan.ipynb",2221,8,"    ",python,content
+228,421775,"test/test_nan.ipynb",2194,8,"    ",python,content
+229,421775,"test/test_nan.ipynb",2170,8,"    ",python,content
+230,421775,"test/test_nan.ipynb",2146,8,"    ",python,content
+231,421775,"test/test_nan.ipynb",2125,8,"    ",python,content
+232,421775,"test/test_nan.ipynb",2103,8,"    ",python,content
+233,421775,"test/test_nan.ipynb",2077,8,"    ",python,content
+234,421775,"test/test_nan.ipynb",2042,4,"",python,content
+235,421776,"test/test_nan.ipynb",2042,0,"",python,selection_command
+236,422167,"test/test_nan.ipynb",2041,0,"",python,selection_command
+237,422475,"test/test_nan.ipynb",2041,1,"",python,content
+238,422707,"test/test_nan.ipynb",2072,0,"",python,selection_command
+239,422947,"test/test_nan.ipynb",2094,0,"",python,selection_command
+240,422987,"test/test_nan.ipynb",2112,0,"",python,selection_command
+241,423027,"test/test_nan.ipynb",2129,0,"",python,selection_command
+242,423075,"test/test_nan.ipynb",2149,0,"",python,selection_command
+243,423087,"test/test_nan.ipynb",2169,0,"",python,selection_command
+244,423127,"test/test_nan.ipynb",2192,0,"",python,selection_command
+245,423147,"test/test_nan.ipynb",2218,0,"",python,selection_command
+246,423267,"test/test_nan.ipynb",2220,0,"",python,selection_command
+247,423867,"test/test_nan.ipynb",2220,1,"",python,content
+248,424007,"test/test_nan.ipynb",2218,0,"",python,selection_command
+249,424248,"test/test_nan.ipynb",2192,0,"",python,selection_command
+250,424295,"test/test_nan.ipynb",2169,0,"",python,selection_command
+251,424327,"test/test_nan.ipynb",2149,0,"",python,selection_command
+252,424355,"test/test_nan.ipynb",2129,0,"",python,selection_command
+253,424387,"test/test_nan.ipynb",2112,0,"",python,selection_command
+254,424415,"test/test_nan.ipynb",2094,0,"",python,selection_command
+255,424447,"test/test_nan.ipynb",2072,0,"",python,selection_command
+256,424487,"test/test_nan.ipynb",2041,0,"",python,selection_command
+257,424767,"test/test_nan.ipynb",1958,0,"",python,selection_command
+258,424927,"test/test_nan.ipynb",1960,0,"",python,selection_command
+259,425187,"test/test_nan.ipynb",1970,0,"",python,selection_command
+260,425207,"test/test_nan.ipynb",1971,0,"",python,selection_command
+261,425247,"test/test_nan.ipynb",1979,0,"",python,selection_command
+262,425275,"test/test_nan.ipynb",1988,0,"",python,selection_command
+263,425307,"test/test_nan.ipynb",1992,0,"",python,selection_command
+264,425335,"test/test_nan.ipynb",2003,0,"",python,selection_command
+265,425567,"test/test_nan.ipynb",2005,0,"",python,selection_command
+266,425807,"test/test_nan.ipynb",2004,0,"",python,selection_command
+267,426127,"test/test_nan.ipynb",2003,0,"",python,selection_command
+268,426208,"test/test_nan.ipynb",2003,1,";",python,selection_command
+269,426227,"test/test_nan.ipynb",2003,4,"; lr",python,selection_command
+270,426387,"test/test_nan.ipynb",2003,10,"; lr value",python,selection_command
+271,426575,"test/test_nan.ipynb",2003,13,"; lr value is",python,selection_command
+272,426727,"test/test_nan.ipynb",2003,24,"; lr value is irrelevant",python,selection_command
+273,427107,"test/test_nan.ipynb",2003,28,"; lr value is irrelevant for",python,selection_command
+274,427527,"test/test_nan.ipynb",2003,36,"; lr value is irrelevant for restore",python,selection_command
+275,427747,"test/test_nan.ipynb",2003,36,"",python,content
+276,428547,"test/test_nan.ipynb",1958,0,"",python,selection_command
+277,428907,"test/test_nan.ipynb",2005,0,"",python,selection_command
+278,429187,"test/test_nan.ipynb",2017,0,"",python,selection_command
+279,430215,"test/test_nan.ipynb",0,0,"",python,selection_command
+280,431287,"test/test_nan.ipynb",50,0,"",python,selection_command
+281,431527,"test/test_nan.ipynb",60,0,"",python,selection_command
+282,431567,"test/test_nan.ipynb",84,0,"",python,selection_command
+283,431590,"test/test_nan.ipynb",85,0,"",python,selection_command
+284,431607,"test/test_nan.ipynb",96,0,"",python,selection_command
+285,431647,"test/test_nan.ipynb",120,0,"",python,selection_command
+286,431687,"test/test_nan.ipynb",143,0,"",python,selection_command
+287,431727,"test/test_nan.ipynb",156,0,"",python,selection_command
+288,431887,"test/test_nan.ipynb",187,0,"",python,selection_command
+289,432027,"test/test_nan.ipynb",200,0,"",python,selection_command
+290,432167,"test/test_nan.ipynb",201,0,"",python,selection_command
+291,433475,"test/test_nan.ipynb",199,0,"\nfrom utils.lr_schedule import get_lr_schedule",python,content
+292,434547,"test/test_nan.ipynb",246,0,"",python,selection_command
+293,434727,"test/test_nan.ipynb",200,0,"",python,selection_command
+294,435147,"test/test_nan.ipynb",246,0,"",python,selection_command
+295,435467,"test/test_nan.ipynb",246,1,"",python,content
+296,435527,"test/test_nan.ipynb",200,0,"",python,selection_command
+297,435667,"test/test_nan.ipynb",187,0,"",python,selection_command
+298,435967,"test/test_nan.ipynb",200,0,"",python,selection_command
+299,436315,"test/test_nan.ipynb",199,0,"\n",python,content
+300,436895,"test/test_nan.ipynb",201,0,"",python,selection_command
+301,437067,"test/test_nan.ipynb",206,0,"",python,selection_command
+302,437247,"test/test_nan.ipynb",211,0,"",python,selection_command
+303,437429,"test/test_nan.ipynb",212,0,"",python,selection_command
+304,438267,"test/test_nan.ipynb",211,0,"",python,selection_command
+305,438415,"test/test_nan.ipynb",206,0,"",python,selection_command
+306,442867,"test/test_nan.ipynb",206,0,"L",python,content
+307,442868,"test/test_nan.ipynb",207,0,"",python,selection_keyboard
+308,443555,"test/test_nan.ipynb",206,1,"",python,content
+309,443739,"test/test_nan.ipynb",206,0,"l",python,content
+310,443740,"test/test_nan.ipynb",207,0,"",python,selection_keyboard
+311,443863,"test/test_nan.ipynb",207,0,"r",python,content
+312,443863,"test/test_nan.ipynb",208,0,"",python,selection_keyboard
+313,444355,"test/test_nan.ipynb",208,0,"_",python,content
+314,444355,"test/test_nan.ipynb",209,0,"",python,selection_keyboard
+315,444884,"test/test_nan.ipynb",208,0,"",python,selection_command
+316,445667,"test/test_nan.ipynb",214,0,"",python,selection_command
+317,445967,"test/test_nan.ipynb",206,0,"",python,selection_command
+318,448816,"utils/lr_utils.py",0,0,"import optax\n\n\ndef get_lr_schedule(\n    lr_schedule: str,\n    init_lr: float,\n    max_lr: float,\n    decay_end: float,\n    total_steps: int,\n    warmup_steps: int,\n    wsd_decay_steps: int,\n) -> optax.Schedule:\n    supported_schedules = [""wsd"", ""cos""]\n    if lr_schedule == ""cos"":\n        assert (\n            warmup_steps <= total_steps\n        ), ""Warmup steps can't be greater than total steps.""\n        return optax.warmup_cosine_decay_schedule(\n            init_value=init_lr,\n            peak_value=max_lr,\n            warmup_steps=warmup_steps,\n            decay_steps=total_steps,  # Note: decay_steps includes the warmup steps, so we need to pass total value\n            end_value=decay_end,\n        )\n    elif lr_schedule == ""wsd"":\n        assert (\n            warmup_steps + wsd_decay_steps <= total_steps\n        ), ""Warmup and decay period is longer than total steps.""\n        schedules = [\n            optax.linear_schedule(\n                init_value=init_lr, end_value=max_lr, transition_steps=warmup_steps\n            ),\n            optax.constant_schedule(value=max_lr),\n            optax.linear_schedule(\n                init_value=max_lr, end_value=decay_end, transition_steps=wsd_decay_steps\n            ),\n        ]\n        boundaries = [warmup_steps, total_steps - wsd_decay_steps]\n        return optax.join_schedules(schedules, boundaries)\n    else:\n        raise ValueError(\n            f""Learning rate schedule not supported. Please use one of {supported_schedules}""\n        )\n",python,tab
+319,449457,"test/test_nan.ipynb",0,0,"",python,tab
+320,449847,"test/test_nan.ipynb",214,0,"",python,selection_command
+321,450007,"test/test_nan.ipynb",215,0,"",python,selection_command
+322,450807,"test/test_nan.ipynb",214,0,"",python,selection_command
+323,450947,"test/test_nan.ipynb",206,0,"",python,selection_command
+324,451295,"test/test_nan.ipynb",206,1,"l",python,selection_command
+325,451367,"test/test_nan.ipynb",206,8,"lr_utils",python,selection_command
+326,451527,"test/test_nan.ipynb",206,9,"lr_utils.",python,selection_command
+327,451947,"test/test_nan.ipynb",206,20,"lr_utils.lr_schedule",python,selection_command
+328,452140,"test/test_nan.ipynb",206,20,"",python,content
+329,452675,"test/test_nan.ipynb",206,0,"u",python,content
+330,452675,"test/test_nan.ipynb",207,0,"",python,selection_keyboard
+331,452787,"test/test_nan.ipynb",207,0,"t",python,content
+332,452787,"test/test_nan.ipynb",208,0,"",python,selection_keyboard
+333,452915,"test/test_nan.ipynb",208,0,"i",python,content
+334,452916,"test/test_nan.ipynb",209,0,"",python,selection_keyboard
+335,453095,"test/test_nan.ipynb",209,0,"s",python,content
+336,453095,"test/test_nan.ipynb",210,0,"",python,selection_keyboard
+337,453535,"test/test_nan.ipynb",209,1,"",python,content
+338,453707,"test/test_nan.ipynb",209,0,"l",python,content
+339,453708,"test/test_nan.ipynb",210,0,"",python,selection_keyboard
+340,453807,"test/test_nan.ipynb",210,0,"s",python,content
+341,453807,"test/test_nan.ipynb",211,0,"",python,selection_keyboard
+342,453907,"test/test_nan.ipynb",211,0,".",python,content
+343,453908,"test/test_nan.ipynb",212,0,"",python,selection_keyboard
+344,454103,"test/test_nan.ipynb",212,0,"l",python,content
+345,454104,"test/test_nan.ipynb",213,0,"",python,selection_keyboard
+346,454147,"test/test_nan.ipynb",213,0,"r",python,content
+347,454147,"test/test_nan.ipynb",214,0,"",python,selection_keyboard
+348,454355,"test/test_nan.ipynb",214,0,"_",python,content
+349,454355,"test/test_nan.ipynb",215,0,"",python,selection_keyboard
+350,454555,"test/test_nan.ipynb",215,0,"s",python,content
+351,454555,"test/test_nan.ipynb",216,0,"",python,selection_keyboard
+352,454816,"test/test_nan.ipynb",215,1,"",python,content
+353,454995,"test/test_nan.ipynb",215,0,"u",python,content
+354,454995,"test/test_nan.ipynb",216,0,"",python,selection_keyboard
+355,455043,"test/test_nan.ipynb",216,0,"t",python,content
+356,455044,"test/test_nan.ipynb",217,0,"",python,selection_keyboard
+357,455147,"test/test_nan.ipynb",217,0,"i",python,content
+358,455147,"test/test_nan.ipynb",218,0,"",python,selection_keyboard
+359,455251,"test/test_nan.ipynb",218,0,"s",python,content
+360,455252,"test/test_nan.ipynb",219,0,"",python,selection_keyboard
+361,455252,"test/test_nan.ipynb",219,0,"l",python,content
+362,455252,"test/test_nan.ipynb",220,0,"",python,selection_keyboard
+363,455811,"test/test_nan.ipynb",219,1,"",python,content
+364,455944,"test/test_nan.ipynb",218,1,"",python,content
+365,456131,"test/test_nan.ipynb",218,0,"l",python,content
+366,456132,"test/test_nan.ipynb",219,0,"",python,selection_keyboard
+367,456147,"test/test_nan.ipynb",219,0,"s",python,content
+368,456147,"test/test_nan.ipynb",220,0,"",python,selection_keyboard
+369,456356,"test/test_nan.ipynb",219,0,"",python,selection_command
+370,458754,"test/test_nan.ipynb",221,0,"",python,selection_command
+371,458761,"test/test_nan.ipynb",228,0,"",python,selection_command
+372,459227,"test/test_nan.ipynb",2062,0,"",python,selection_command
+373,460370,"test/test_nan.ipynb",2093,0,"",python,selection_command
+374,460617,"test/test_nan.ipynb",2115,0,"",python,selection_command
+375,460650,"test/test_nan.ipynb",2133,0,"",python,selection_command
+376,460695,"test/test_nan.ipynb",2150,0,"",python,selection_command
+377,460803,"test/test_nan.ipynb",2170,0,"",python,selection_command
+378,461115,"test/test_nan.ipynb",2150,0,"",python,selection_command
+379,461416,"test/test_nan.ipynb",2170,0,"",python,selection_command
+380,461533,"test/test_nan.ipynb",2190,0,"",python,selection_command
+381,461863,"test/test_nan.ipynb",2213,0,"",python,selection_command
+382,462229,"test/test_nan.ipynb",2225,0,"",python,selection_command
+383,463971,"test/test_nan.ipynb",2227,0,"",python,selection_command
+384,464681,"test/test_nan.ipynb",2225,0,"",python,selection_command
+385,464934,"test/test_nan.ipynb",2213,0,"",python,selection_command
+386,464959,"test/test_nan.ipynb",2190,0,"",python,selection_command
+387,465000,"test/test_nan.ipynb",2170,0,"",python,selection_command
+388,465019,"test/test_nan.ipynb",2150,0,"",python,selection_command
+389,465189,"test/test_nan.ipynb",2133,0,"",python,selection_command
+390,465345,"test/test_nan.ipynb",2115,0,"",python,selection_command
+391,465482,"test/test_nan.ipynb",2093,0,"",python,selection_command
+392,465661,"test/test_nan.ipynb",2062,0,"",python,selection_command
+393,469342,"test/test_nan.ipynb",0,0,"",python,tab
+394,470114,"test/test_nan.ipynb",0,0,"",python,tab
+395,470720,"test/test_nan.ipynb",0,0,"",python,tab
+396,470912,"test/test_nan.ipynb",0,0,"",python,tab
+397,477040,"test/test_nan.ipynb",2093,0,"",python,selection_command
+398,477041,"test/test_nan.ipynb",2079,0,"",python,selection_command
+399,477042,"test/test_nan.ipynb",2083,0,"",python,selection_command
+400,477042,"test/test_nan.ipynb",2083,1,"a",python,selection_command
+401,477043,"test/test_nan.ipynb",2083,2,"ar",python,selection_command
+402,477043,"test/test_nan.ipynb",2083,3,"arg",python,selection_command
+403,477044,"test/test_nan.ipynb",2083,4,"args",python,selection_command
+404,477044,"test/test_nan.ipynb",2083,4,"args",python,selection_command
+405,477045,"test/test_nan.ipynb",2083,5,"args.",python,selection_command
+406,477046,"test/test_nan.ipynb",2083,5,"args.",python,selection_command
+407,477046,"test/test_nan.ipynb",2083,5,"args.",python,selection_command
+408,477046,"test/test_nan.ipynb",2083,5,"args.",python,selection_command
+409,477047,"test/test_nan.ipynb",2083,5,"args.",python,selection_command
+410,477047,"test/test_nan.ipynb",2083,5,"args.",python,selection_command
+411,477047,"test/test_nan.ipynb",2080,4,"   a",python,selection_command
+412,477048,"test/test_nan.ipynb",2083,5,"args.",python,selection_command
+413,477122,"test/test_nan.ipynb",2203,5,"",python,content
+414,477122,"test/test_nan.ipynb",2180,5,"",python,content
+415,477122,"test/test_nan.ipynb",2160,5,"",python,content
+416,477122,"test/test_nan.ipynb",2140,5,"",python,content
+417,477122,"test/test_nan.ipynb",2123,5,"",python,content
+418,477122,"test/test_nan.ipynb",2105,5,"",python,content
+419,477122,"test/test_nan.ipynb",2083,5,"",python,content
+420,477143,"test/test_nan.ipynb",2083,0,"",python,selection_command
+421,485066,"test/test_nan.ipynb",2052,0,"",python,selection_command
+422,485112,"test/test_nan.ipynb",2060,0,"",python,selection_command
+423,485287,"test/test_nan.ipynb",2062,0,"",python,selection_command
+424,485804,"utils/lr_utils.py",0,0,"",python,tab
+425,487249,"test/test_nan.ipynb",0,0,"",python,tab
+426,488274,"test/test_nan.ipynb",2093,0,"",python,selection_command
+427,488285,"test/test_nan.ipynb",2083,0,"",python,selection_command
+428,491359,"test/test_nan.ipynb",2048,11,"lr_schedule",python,selection_command
+429,491916,"test/test_nan.ipynb",2058,0,"",python,selection_command
+430,492057,"test/test_nan.ipynb",2083,0,"",python,selection_command
+431,492512,"test/test_nan.ipynb",2052,0,"",python,selection_command
+432,494140,"test/test_nan.ipynb",2058,0,"",python,selection_command
+433,494543,"test/test_nan.ipynb",2059,0,"",python,selection_command
+434,494855,"test/test_nan.ipynb",2059,0,"_",python,content
+435,494856,"test/test_nan.ipynb",2060,0,"",python,selection_keyboard
+436,495058,"test/test_nan.ipynb",2060,0,"f",python,content
+437,495058,"test/test_nan.ipynb",2061,0,"",python,selection_keyboard
+438,495198,"test/test_nan.ipynb",2061,0,"n",python,content
+439,495198,"test/test_nan.ipynb",2062,0,"",python,selection_keyboard
+440,495273,"test/test_nan.ipynb",2061,0,"",python,selection_command
+441,495408,"test/test_nan.ipynb",2095,0,"",python,selection_command
+442,495657,"test/test_nan.ipynb",2110,0,"",python,selection_command
+443,495693,"test/test_nan.ipynb",2122,0,"",python,selection_command
+444,495725,"test/test_nan.ipynb",2137,0,"",python,selection_command
+445,495756,"test/test_nan.ipynb",2152,0,"",python,selection_command
+446,495814,"test/test_nan.ipynb",2167,0,"",python,selection_command
+447,495821,"test/test_nan.ipynb",2185,0,"",python,selection_command
+448,495859,"test/test_nan.ipynb",2193,0,"",python,selection_command
+449,495907,"test/test_nan.ipynb",2195,0,"",python,selection_command
+450,495929,"test/test_nan.ipynb",2209,0,"",python,selection_command
+451,496180,"test/test_nan.ipynb",2227,0,"",python,selection_command
+452,496339,"test/test_nan.ipynb",2250,0,"",python,selection_command
+453,496517,"test/test_nan.ipynb",2227,0,"",python,selection_command
+454,496722,"test/test_nan.ipynb",2209,0,"",python,selection_command
+455,496987,"test/test_nan.ipynb",2195,0,"",python,selection_command
+456,497031,"test/test_nan.ipynb",2193,0,"",python,selection_command
+457,497040,"test/test_nan.ipynb",2185,0,"",python,selection_command
+458,497071,"test/test_nan.ipynb",2167,0,"",python,selection_command
+459,497097,"test/test_nan.ipynb",2152,0,"",python,selection_command
+460,497129,"test/test_nan.ipynb",2137,0,"",python,selection_command
+461,497166,"test/test_nan.ipynb",2122,0,"",python,selection_command
+462,497203,"test/test_nan.ipynb",2110,0,"",python,selection_command
+463,497229,"test/test_nan.ipynb",2095,0,"",python,selection_command
+464,497264,"test/test_nan.ipynb",2061,0,"",python,selection_command
+465,497296,"test/test_nan.ipynb",2014,0,"",python,selection_command
+466,497385,"test/test_nan.ipynb",2061,0,"",python,selection_command
+467,497669,"test/test_nan.ipynb",2095,0,"",python,selection_command
+468,497926,"test/test_nan.ipynb",2110,0,"",python,selection_command
+469,497952,"test/test_nan.ipynb",2122,0,"",python,selection_command
+470,497979,"test/test_nan.ipynb",2137,0,"",python,selection_command
+471,498017,"test/test_nan.ipynb",2152,0,"",python,selection_command
+472,498075,"test/test_nan.ipynb",2167,0,"",python,selection_command
+473,498098,"test/test_nan.ipynb",2185,0,"",python,selection_command
+474,498117,"test/test_nan.ipynb",2193,0,"",python,selection_command
+475,498144,"test/test_nan.ipynb",2195,0,"",python,selection_command
+476,498182,"test/test_nan.ipynb",2209,0,"",python,selection_command
+477,498226,"test/test_nan.ipynb",2227,0,"",python,selection_command
+478,498362,"test/test_nan.ipynb",2250,0,"",python,selection_command
+479,498820,"test/test_nan.ipynb",2227,0,"",python,selection_command
+480,498940,"test/test_nan.ipynb",2231,0,"",python,selection_command
+481,499088,"test/test_nan.ipynb",2232,0,"",python,selection_command
+482,499440,"test/test_nan.ipynb",2232,6,"",python,content
+483,499929,"test/test_nan.ipynb",2232,0,"l",python,content
+484,499929,"test/test_nan.ipynb",2233,0,"",python,selection_keyboard
+485,500026,"test/test_nan.ipynb",2233,0,"r",python,content
+486,500026,"test/test_nan.ipynb",2234,0,"",python,selection_keyboard
+487,500321,"test/test_nan.ipynb",2234,0,"_",python,content
+488,500322,"test/test_nan.ipynb",2235,0,"",python,selection_keyboard
+489,500456,"test/test_nan.ipynb",2235,0,"s",python,content
+490,500457,"test/test_nan.ipynb",2236,0,"",python,selection_keyboard
+491,500543,"test/test_nan.ipynb",2236,0,"c",python,content
+492,500544,"test/test_nan.ipynb",2237,0,"",python,selection_keyboard
+493,500557,"test/test_nan.ipynb",2237,0,"h",python,content
+494,500557,"test/test_nan.ipynb",2238,0,"",python,selection_keyboard
+495,500684,"test/test_nan.ipynb",2238,0,"e",python,content
+496,500685,"test/test_nan.ipynb",2239,0,"",python,selection_keyboard
+497,500722,"test/test_nan.ipynb",2239,0,"d",python,content
+498,500723,"test/test_nan.ipynb",2240,0,"",python,selection_keyboard
+499,500829,"test/test_nan.ipynb",2240,0,"u",python,content
+500,500830,"test/test_nan.ipynb",2241,0,"",python,selection_keyboard
+501,500958,"test/test_nan.ipynb",2241,0,"l",python,content
+502,500959,"test/test_nan.ipynb",2242,0,"",python,selection_keyboard
+503,501100,"test/test_nan.ipynb",2242,0,"e",python,content
+504,501101,"test/test_nan.ipynb",2243,0,"",python,selection_keyboard
+505,501400,"test/test_nan.ipynb",2243,0,"_",python,content
+506,501401,"test/test_nan.ipynb",2244,0,"",python,selection_keyboard
+507,501576,"test/test_nan.ipynb",2244,0,"f",python,content
+508,501577,"test/test_nan.ipynb",2245,0,"",python,selection_keyboard
+509,501635,"test/test_nan.ipynb",2245,0,"n",python,content
+510,501636,"test/test_nan.ipynb",2246,0,"",python,selection_keyboard
+511,501797,"test/test_nan.ipynb",2245,0,"",python,selection_command
+512,502157,"test/test_nan.ipynb",2214,0,"",python,selection_command
+513,504365,"utils/lr_utils.py",0,0,"",python,tab
+514,507250,"test/test_nan.ipynb",0,0,"",python,tab
+515,509332,"train_lam.py",0,0,"",python,tab
+516,512271,"test/test_nan.ipynb",0,0,"",python,tab
+517,512870,"train_lam.py",0,0,"",python,tab
+518,513510,"test/test_nan.ipynb",0,0,"",python,tab
+519,2086061,"train_lam.py",0,0,"",python,tab
+520,2086423,"train_lam.py",0,13598,"from dataclasses import dataclass, field\nimport os\nfrom typing import cast\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.lam import LatentActionModel\nfrom utils.dataloader import get_dataloader\nfrom utils.lr_utils import get_lr_schedule\nfrom utils.parameter_utils import count_parameters_by_component\n\njax.config.update(""jax_transfer_guard"", ""allow"")\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    vq_beta: float = 0.25\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    vq_reset_thresh: int = 50\n    # LAM\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 6\n    patch_size: int = 16\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.0\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_lam""\n    tags: list[str] = field(default_factory=lambda: [""lam""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_checkpoint_keep_period: int = 20000\n    wandb_id: str = """"\n    use_flash_attention: bool = True\n\n\nargs = tyro.cli(Args)\n\n\ndef lam_loss_fn(\n    model: LatentActionModel, inputs: dict\n) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n    # --- Compute loss ---\n    gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n    inputs[""videos""] = gt.astype(args.dtype)\n    model.train()\n    outputs = model(inputs, training=True)\n    outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n    gt_future_frames = gt[:, 1:]\n    mse = jnp.square(gt_future_frames - outputs[""recon""]).mean()\n    q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n    commitment_loss = jnp.square(\n        outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n    ).mean()\n    loss = mse + q_loss + args.vq_beta * commitment_loss\n\n    # --- Compute validation metrics ---\n    gt = gt_future_frames.clip(0, 1).reshape(-1, *gt_future_frames.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt, recon)).mean()\n    count_fn = jax.vmap(lambda i: (outputs[""indices""] == i).sum())\n    index_counts = count_fn(jnp.arange(args.num_latents))\n    metrics = dict(\n        loss=loss,\n        mse=mse,\n        q_loss=q_loss,\n        commitment_loss=commitment_loss,\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage=(index_counts != 0).mean(),\n    )\n    return loss, (outputs[""recon""], index_counts, metrics)\n\n\n@nnx.jit\ndef train_step(\n    lam: LatentActionModel,\n    optimizer: nnx.Optimizer,\n    inputs: dict,\n    action_last_active: jax.Array,\n    rng: jax.Array,\n) -> tuple[jax.Array, jax.Array, jax.Array, dict]:\n    def loss_fn(\n        model: LatentActionModel,\n    ) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n        return lam_loss_fn(model, inputs)\n\n    # --- Update model ---\n    (loss, (recon, idx_counts, metrics)), grads = nnx.value_and_grad(\n        loss_fn, has_aux=True\n    )(lam)\n    optimizer.update(grads)\n\n    # --- Reset inactive latent actions ---\n    codebook = lam.vq.codebook\n    num_codes = len(codebook)\n    active_codes = idx_counts != 0.0\n    action_last_active = jnp.where(active_codes, 0, action_last_active + 1)\n    p_code = active_codes / active_codes.sum()\n    reset_idxs = jax.random.choice(rng, num_codes, shape=(num_codes,), p=p_code)\n    do_reset = action_last_active >= args.vq_reset_thresh\n    new_codebook = jnp.where(\n        jnp.expand_dims(do_reset, -1), codebook[reset_idxs], codebook.value\n    )\n    lam.vq.codebook.value = new_codebook\n    action_last_active = jnp.where(do_reset, 0, action_last_active)\n    return loss, recon, action_last_active, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    lam = LatentActionModel(\n        in_dim=args.image_channels,\n        model_dim=args.model_dim,\n        ffn_dim=args.ffn_dim,\n        latent_dim=args.latent_dim,\n        num_latents=args.num_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.num_blocks,\n        num_heads=args.num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.codebook_dropout,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        rngs=rngs,\n    )\n\n    # Count parameters\n    _, params, _ = nnx.split(lam, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype, # moments in full precision\n    )\n    optimizer = nnx.Optimizer(lam, tx)\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n    # --- Initialize checkpoint manager ---\n    step = 0\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    # --- Create DataLoaderIterator from dataloader ---\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n    # --- Restore checkpoint ---\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    print(f""Starting training from step {step}..."")\n    action_last_active = jnp.zeros(args.num_latents, dtype=jnp.int32)\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng = jax.random.split(rng)\n\n            inputs = dict(videos=videos, rng=_rng)\n            rng, _rng = jax.random.split(rng)\n            loss, recon, action_last_active, metrics = train_step(\n                lam, optimizer, inputs, action_last_active, _rng\n            )\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0, 1:].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n",python,selection_command
+521,2086571,"train_lam.py",13598,0,"",python,selection_command
+522,2248550,"utils/nn.py",0,0,"import math\nfrom typing import Tuple, Callable, List\n\nfrom flax import nnx\nimport jax\nimport jax.numpy as jnp\nimport einops\n\n\nclass SpatioTemporalPositionalEncoding(nnx.Module):\n    """"""\n    Applies separate sinusoidal positional encodings to the temporal and spatial dimensions.\n    """"""\n    def __init__(self, d_model: int, max_len: int = 5000):\n        self.d_model = d_model\n        self.max_len = max_len\n\n        pe = jnp.zeros((self.max_len, self.d_model))\n        position = jnp.arange(0, self.max_len, dtype=jnp.float32)[:, None]\n        div_term = jnp.exp(\n            jnp.arange(0, self.d_model, 2) * (-math.log(10000.0) / self.d_model)\n        )\n        pe = pe.at[:, 0::2].set(jnp.sin(position * div_term))\n        pe = pe.at[:, 1::2].set(jnp.cos(position * div_term))\n        self.pe = nnx.Variable(pe)\n\n    def __call__(self, x: jax.Array) -> jax.Array:\n        """"""\n        Args:\n            x: The input tensor of shape (Batch, Time, Space, Dimension).\n\n        Returns:\n            The input tensor with positional encodings added.\n        """"""\n        assert x.ndim == 4, f""Input must be 4-dimensional, but got shape {x.shape}""\n\n        num_timesteps = x.shape[1]\n        num_spatial_patches = x.shape[2]\n\n        # Temporal positional encoding: (1, T, 1, D)\n        temporal_pe = self.pe.value[None, :num_timesteps, None, :]\n        x = x + temporal_pe\n\n        # Spatial positional encoding: (1, 1, S, D)\n        spatial_pe = self.pe.value[None, None, :num_spatial_patches, :]\n        x = x + spatial_pe\n\n        return x\n\n\nclass STBlock(nnx.Module):\n    def __init__(\n        self,\n        dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool,\n        sow_activations: bool,\n    ):\n        self.dim = dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=False\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(self, x_BTNM: jax.Array) -> jax.Array:\n        # --- Spatial attention ---\n        z_BTNM = self.spatial_norm(x_BTNM)\n        z_BTNM = self.spatial_attention(z_BTNM, sow_weights=self.sow_weights)\n        x_BTNM = x_BTNM + z_BTNM\n\n        # --- Temporal attention ---\n        x_BNTM = x_BTNM.swapaxes(1, 2)\n        z_BNTM = self.temporal_norm(x_BNTM)\n        z_BNTM = self.temporal_attention(z_BNTM, sow_weights=self.sow_weights)\n        x_BNTM = x_BNTM + z_BNTM\n        x_BTNM = x_BNTM.swapaxes(1, 2)\n\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n        if self.sow_activations:\n            self.sow(nnx.Intermediate, 'activations', x_BTNM)\n        return x_BTNM\n\n\nclass STTransformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        V: vocabulary size\n    """"""\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool = False,\n        sow_activations: bool = False,\n        sow_logits: bool = False,\n        max_len: int = 5000,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n        self.sow_logits = sow_logits\n\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n\n        self.pos_enc = SpatioTemporalPositionalEncoding(self.model_dim, max_len=max_len)\n\n        self.blocks = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                STBlock(\n                    dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    rngs=rngs,\n                    sow_weights=self.sow_weights,\n                    sow_activations=self.sow_activations,\n                )\n            )\n\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n        x_BTNM = self.pos_enc(x_BTNM)\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM)\n\n        x_BTNV = self.output_dense(x_BTNM)\n        if self.sow_logits:\n            self.sow(nnx.Intermediate, 'logits', x_BTNV)\n        return x_BTNV\n\nclass TransformerBlock(nnx.Module):\n    def __init__(\n        self,\n        model_dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool,\n        sow_activations: bool,\n    ):\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.decode = decode\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    # @nnx.remat\n    def __call__(self, x_BTNM: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None) -> jax.Array:\n        # --- Spatial attention ---\n        B, T, N, M = x_BTNM.shape\n        z_FNM = einops.rearrange(x_BTNM, ""b t n m -> (b t) n m"")\n        z_FNM = self.spatial_norm(z_FNM)\n        z_FNM = self.spatial_attention(z_FNM, sow_weights=self.sow_weights)\n        z_BTNM = einops.rearrange(z_FNM, ""(b t) n m -> b t n m"", t=T)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Temporal attention ---\n        z_PTM = einops.rearrange(x_BTNM, ""b t n m -> (b n) t m"")\n        z_PTM = self.temporal_norm(z_PTM)\n        z_PTM = self.temporal_attention(z_PTM, sow_weights=self.sow_weights)\n        z_BTNM = einops.rearrange(z_PTM, ""(b n) t m -> b t n m"", n=N)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n        if self.sow_activations:\n            self.sow(nnx.Intermediate, 'activations', x_BTNM)\n\n        return x_BTNM\n\nclass Transformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        V: vocabulary size\n        F: number of frames in batch\n        P: number of patch positions in batch\n    """"""\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n        sow_logits: bool = False,\n        sow_weights: bool = False,\n        sow_activations: bool = False,\n        max_len: int = 5000,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_logits = sow_logits\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype, # layer norm in full precision\n            rngs=rngs,\n        )\n\n        self.pos_enc = SpatioTemporalPositionalEncoding(self.model_dim, max_len=max_len)\n\n        self.blocks: List[TransformerBlock] = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                TransformerBlock(\n                    model_dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    decode=decode,\n                    sow_weights=self.sow_weights,\n                    sow_activations=self.sow_activations,\n                    rngs=rngs,\n                )\n            )\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n        x_BTNM = self.pos_enc(x_BTNM)\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM, pos_index)\n\n        x_BTNV = self.output_dense(x_BTNM)\n        if self.sow_logits:\n            self.sow(nnx.Intermediate, 'logits', x_BTNV)\n        return x_BTNV\n\ndef normalize(x: jax.Array) -> jax.Array:\n    return x / (jnp.linalg.norm(x, ord=2, axis=-1, keepdims=True) + 1e-8)\n\n\nclass VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n    def __init__(\n        self, latent_dim: int, num_latents: int, dropout: float, rngs: nnx.Rngs\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n\n        self.codebook = nnx.Param(\n            normalize(\n                nnx.initializers.lecun_uniform()(\n                    rngs.params(), (self.num_latents, self.latent_dim)\n                )\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = normalize(x_DL)\n        normalized_codebook_KL = normalize(self.codebook.value)\n        distance_DK = -jnp.matmul(x_DL, normalized_codebook_KL.T)\n        if training:\n            distance_DK = self.drop(distance_DK)\n\n        # --- Get indices and embeddings ---\n        indices_D = jnp.argmin(distance_DK, axis=-1)\n        z_DL = self.codebook[indices_D]\n\n        # --- Straight through estimator ---\n        z_q_DL = x_DL + jax.lax.stop_gradient(z_DL - x_DL)\n        return z_q_DL, z_DL, x_DL, indices_D\n\n    def get_codes(self, indices_E: jax.Array) -> jax.Array:\n        return self.codebook[indices_E]\n\n\ndef _create_flash_attention_fn(use_flash_attention: bool, is_causal: bool) -> Callable:\n    """"""\n    Create an attention function that uses flash attention if enabled.\n\n    flax.nnx.MultiHeadAttention provides tensors with shape (batch..., length, num_heads, head_dim),\n    but jax.nn.dot_product_attention expects (batch, length, num_heads, head_dim). We reshape to\n    ensure compatibility. cuDNN's flash attention additionally requires a sequence length that\n    is a multiple of 4. We pad the sequence length to the nearest multiple of 4 and mask\n    accordingly. Note that cuDNN requires the mask to be broadcast before calling the attention\n    function due to strict shape checking.\n    """"""\n\n    def attention_fn(query_BTHD, key_BSHD, value_BSHD, bias=None, mask_B111=None, **kwargs):\n        implementation = ""cudnn"" if use_flash_attention else None\n\n        def _merge_batch_dims(x):\n            return einops.rearrange(x, ""... l h k -> (...) l h k"")\n\n        def _pad(x, pad_size):\n            return jnp.pad(x, ((0, 0), (0, pad_size), (0, 0), (0, 0)))\n\n        original_shape = query_BTHD.shape\n        T = query_BTHD.shape[-3]\n        S = key_BSHD.shape[-3]\n\n        # Pad to nearest multiple of 4\n        Q = ((T + 3) // 4) * 4\n        pad_size_Q = Q - T\n        K = ((S + 3) // 4) * 4\n        pad_size_K = K - S\n\n        query_BQHD = _pad(_merge_batch_dims(query_BTHD), pad_size_Q)\n        key_BKHD = _pad(_merge_batch_dims(key_BSHD), pad_size_K)\n        value_BKHD = _pad(_merge_batch_dims(value_BSHD), pad_size_K)\n\n        attention_mask = jnp.ones((Q, K), dtype=jnp.bool_)\n        attention_mask = attention_mask.at[T:, :].set(False)\n        attention_mask = attention_mask.at[:, S:].set(False)\n\n        mask_11TS = attention_mask[jnp.newaxis, jnp.newaxis, :, :]\n\n        bias_4d = jnp.pad(_merge_batch_dims(bias), ((0, 0), (0, 0), (0, pad_size_Q), (0, pad_size_K))) if bias is not None else None\n\n        # NOTE: jax.nn.dot_product_attention does not support dropout\n        output_4d = jax.nn.dot_product_attention(\n            query=query_BQHD,\n            key=key_BKHD,\n            value=value_BKHD,\n            bias=bias_4d,\n            mask=mask_11TS,\n            implementation=implementation,\n            is_causal=is_causal,\n        )\n        return output_4d[..., :T, :, :].reshape(original_shape)\n\n    return attention_fn\n",python,tab
+523,2252301,"utils/nn.py",7559,0,"",python,selection_command
+524,2252303,"utils/nn.py",7528,0,"",python,selection_command
+525,2252303,"utils/nn.py",7462,0,"",python,selection_command
+526,2252304,"utils/nn.py",7424,0,"",python,selection_command
+527,2252305,"utils/nn.py",7374,0,"",python,selection_command
+528,2252306,"utils/nn.py",7332,0,"",python,selection_command
+529,2252307,"utils/nn.py",7286,0,"",python,selection_command
+530,2252307,"utils/nn.py",7244,0,"",python,selection_command
+531,2252308,"utils/nn.py",7204,0,"",python,selection_command
+532,2252309,"utils/nn.py",7179,0,"",python,selection_command
+533,2252503,"utils/nn.py",1547,0,"",python,selection_command
+534,2253323,"utils/nn.py",1574,0,"",python,selection_command
+535,2253579,"utils/nn.py",1592,0,"",python,selection_command
+536,2253605,"utils/nn.py",1606,0,"",python,selection_command
+537,2253631,"utils/nn.py",1624,0,"",python,selection_command
+538,2253664,"utils/nn.py",1646,0,"",python,selection_command
+539,2253700,"utils/nn.py",1670,0,"",python,selection_command
+540,2253735,"utils/nn.py",1694,0,"",python,selection_command
+541,2253761,"utils/nn.py",1726,0,"",python,selection_command
+542,2253903,"utils/nn.py",1752,0,"",python,selection_command
+543,2254063,"utils/nn.py",1787,0,"",python,selection_command
+544,2254650,"utils/nn.py",1811,0,"",python,selection_command
+545,2255170,"utils/nn.py",1813,0,"",python,selection_command
+546,2255577,"utils/nn.py",2124,0,"",python,selection_command
+547,2270643,"utils/nn.py",2138,0,"",python,selection_command
+548,2271361,"utils/nn.py",4408,0,"",python,selection_command
+549,2277137,"utils/nn.py",4406,0,"",python,selection_command
+550,2277319,"utils/nn.py",4400,0,"",python,selection_command
+551,2277484,"utils/nn.py",4399,0,"",python,selection_command
+552,2277755,"utils/nn.py",4382,0,"",python,selection_command
+553,2280545,"train_lam.py",0,0,"",python,tab
+554,2282114,"utils/nn.py",0,0,"",python,tab
+555,2553194,"utils/nn.py",4985,0,"",python,selection_command
+556,2555180,"utils/nn.py",5018,0,"",python,selection_command
+557,2556838,"utils/nn.py",4985,0,"",python,selection_command
+558,2556969,"utils/nn.py",4952,0,"",python,selection_command
+559,2557302,"utils/nn.py",4949,0,"",python,selection_command
+560,2557470,"utils/nn.py",4947,0,"",python,selection_command
+561,2557644,"utils/nn.py",4940,0,"",python,selection_command
+562,2571436,"utils/nn.py",5757,0,"",python,selection_command
+563,2572184,"utils/nn.py",6274,0,"",python,selection_command
+564,2573118,"utils/nn.py",6287,0,"",python,selection_command
+565,2573318,"utils/nn.py",8258,0,"",python,selection_command
+566,2613940,"utils/nn.py",8257,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-3f2b1a99-0d75-466c-970c-4deff62cba851753462933379-2025_07_25-19.02.23.245/source.csv

@@ -0,0 +1,94 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"genie.py",0,0,"from typing import Dict, Any\n\nimport optax\nimport jax\nimport jax.numpy as jnp\nimport flax.nnx as nnx\nfrom flax.training.train_state import TrainState\nimport orbax.checkpoint as ocp\n\nfrom models.dynamics import DynamicsMaskGIT\nfrom models.lam import LatentActionModel\nfrom models.tokenizer import TokenizerVQVAE\n\nimport grain\n\n\nclass Genie(nnx.Module):\n    """"""Genie model""""""\n\n    def __init__(\n        self,\n        in_dim: int,\n        tokenizer_dim: int,\n        tokenizer_ffn_dim: int,\n        latent_patch_dim: int,\n        num_patch_latents: int,\n        patch_size: int,\n        tokenizer_num_blocks: int,\n        tokenizer_num_heads: int,\n        lam_dim: int,\n        lam_ffn_dim: int,\n        latent_action_dim: int,\n        num_latent_actions: int,\n        lam_patch_size: int,\n        lam_num_blocks: int,\n        lam_num_heads: int,\n        lam_co_train: bool,\n        dyna_dim: int,\n        dyna_ffn_dim: int,\n        dyna_num_blocks: int,\n        dyna_num_heads: int,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n        dropout: float = 0.0,\n        mask_limit: float = 0.0,\n    ):\n        # --- Tokenizer ---\n        self.in_dim = in_dim\n        self.tokenizer_dim = tokenizer_dim\n        self.tokenizer_ffn_dim = tokenizer_ffn_dim\n        self.latent_patch_dim = latent_patch_dim\n        self.num_patch_latents = num_patch_latents\n        self.patch_size = patch_size\n        self.tokenizer_num_blocks = tokenizer_num_blocks\n        self.tokenizer_num_heads = tokenizer_num_heads\n        # --- LAM ---\n        self.lam_dim = lam_dim\n        self.lam_ffn_dim = lam_ffn_dim\n        self.latent_action_dim = latent_action_dim\n        self.num_latent_actions = num_latent_actions\n        self.lam_patch_size = lam_patch_size\n        self.lam_num_blocks = lam_num_blocks\n        self.lam_num_heads = lam_num_heads\n        self.lam_co_train = lam_co_train\n        # --- Dynamics ---\n        self.dyna_dim = dyna_dim\n        self.dyna_ffn_dim = dyna_ffn_dim\n        self.dyna_num_blocks = dyna_num_blocks\n        self.dyna_num_heads = dyna_num_heads\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.dropout = dropout\n        self.mask_limit = mask_limit\n\n        self.tokenizer = TokenizerVQVAE(\n            in_dim=self.in_dim,\n            model_dim=self.tokenizer_dim,\n            ffn_dim=self.tokenizer_ffn_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_patch_latents,\n            patch_size=self.patch_size,\n            num_blocks=self.tokenizer_num_blocks,\n            num_heads=self.tokenizer_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            use_flash_attention=self.use_flash_attention,\n            rngs=rngs,\n        )\n        self.lam = LatentActionModel(\n            in_dim=self.in_dim,\n            model_dim=self.lam_dim,\n            ffn_dim=self.lam_ffn_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_latent_actions,\n            patch_size=self.lam_patch_size,\n            num_blocks=self.lam_num_blocks,\n            num_heads=self.lam_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            use_flash_attention=self.use_flash_attention,\n            rngs=rngs,\n        )\n        self.dynamics = DynamicsMaskGIT(\n            model_dim=self.dyna_dim,\n            ffn_dim=self.dyna_ffn_dim,\n            num_latents=self.num_patch_latents,\n            latent_action_dim=self.latent_action_dim,\n            num_blocks=self.dyna_num_blocks,\n            num_heads=self.dyna_num_heads,\n            dropout=self.dropout,\n            mask_limit=self.mask_limit,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            use_flash_attention=self.use_flash_attention,\n            rngs=rngs,\n        )\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        tokenizer_outputs = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        lam_outputs = self.lam.vq_encode(batch[""videos""], training=False)\n        latent_actions = jax.lax.cond(\n            self.lam_co_train,\n            lambda: lam_outputs[""z_q""],\n            lambda: jax.lax.stop_gradient(lam_outputs[""z_q""]),\n        )\n        outputs = dict(\n            video_tokens=jax.lax.stop_gradient(tokenizer_outputs[""indices""]),\n            latent_actions=latent_actions,\n        )\n        outputs[""mask_rng""] = batch[""mask_rng""]\n        dyna_outputs = self.dynamics(outputs, training)\n        outputs.update(dyna_outputs)\n        mle_indices = jnp.argmax(outputs[""token_logits""], axis=-1)\n        outputs[""recon""] = self.tokenizer.decode(\n            mle_indices, batch[""videos""].shape[2:4]\n        )\n        outputs[""lam_indices""] = lam_outputs[""indices""]\n        return outputs\n\n    def sample(\n        self,\n        batch: Dict[str, Any],\n        seq_len: int,\n        steps: int = 25,\n        temperature: float = 1,\n        sample_argmax: bool = False,\n    ) -> Any:\n        """"""\n        Autoregressively samples up to `seq_len` future frames, following Figure 8 of the paper.\n\n        - Input frames are tokenized once.\n        - Future frames are generated autoregressively in token space.\n        - All frames are detokenized in a single pass.\n\n        Note:\n        - For interactive or step-wise sampling, detokenization should occur after each action.\n        - To maintain consistent tensor shapes across timesteps, all current and future frames are decoded at every step.\n        - Temporal causal structure is preserved by\n            a) reapplying the mask before each decoding step.\n            b) a temporal causal mask is applied within each ST-transformer block.\n\n        Dimension keys:\n            B: batch size\n            T: number of input (conditioning) frames\n            N: patches per frame\n            S: sequence length\n            A: action space\n            D: model latent dimension\n        """"""\n        # --- Encode videos and actions ---\n        tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        token_idxs = tokenizer_out[""indices""]  # (B, T, N)\n        B, T, N = token_idxs.shape\n        pad_shape = (B, seq_len - T, N)\n        pad = jnp.zeros(pad_shape, dtype=token_idxs.dtype)\n        token_idxs = jnp.concatenate([token_idxs, pad], axis=1)  # (B, S, N)\n        action_tokens = self.lam.vq.get_codes(batch[""latent_actions""])\n\n        # Define the inner MaskGIT loop using nnx.scan\n        maskgit_step = MaskGITStep(\n            dynamics=self.dynamics,\n            tokenizer=self.tokenizer,\n            temperature=temperature,\n            sample_argmax=sample_argmax,\n            steps=steps,\n        )\n\n        def maskgit_scan_fn(module, carry, x):\n            new_carry, _ = module(carry, x)\n            return new_carry, None\n\n        MaskGITLoop = nnx.scan(\n            maskgit_scan_fn,\n            in_axes=(None, nnx.Carry, 0),  # (module, carry, x)\n            out_axes=(nnx.Carry, None),    # (new_carry, None)\n        )\n\n        # Define the outer autoregressive loop's body function\n        def generation_step_fn(carry, step_t):\n            rng, current_token_idxs = carry\n            rng, step_rng = jax.random.split(rng)\n\n            # Mask current and future frames (i.e., t >= step_t)\n            mask = jnp.arange(seq_len) >= step_t  # (S,)\n            mask = jnp.broadcast_to(mask[None, :, None], (B, seq_len, N)).astype(bool)  # (B, S, N)\n            masked_token_idxs = current_token_idxs * ~mask\n\n            # --- Initialize and run MaskGIT loop ---\n            init_carry_maskgit = (\n                step_rng,\n                masked_token_idxs,\n                mask,\n                action_tokens,\n            )\n            final_carry_maskgit, _ = MaskGITLoop(\n                maskgit_step, init_carry_maskgit, jnp.arange(steps)\n            )\n            updated_token_idxs = final_carry_maskgit[1]\n            new_carry = (rng, updated_token_idxs)\n            return new_carry, None\n\n        # --- Run the autoregressive generation using jax.lax.scan ---\n        initial_carry = (batch[""rng""], token_idxs)\n        timesteps_to_scan = jnp.arange(T, seq_len)\n        final_carry, _ = jax.lax.scan(\n            generation_step_fn, initial_carry, timesteps_to_scan\n        )\n        final_token_idxs = final_carry[1]\n\n        # --- Decode all tokens at once at the end ---\n        final_frames = self.tokenizer.decode(\n            final_token_idxs,\n            video_hw=batch[""videos""].shape[2:4],\n        )\n        return final_frames\n\n    def vq_encode(self, batch, training) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\n\nclass MaskGITStep(nnx.Module):\n    def __init__(\n        self,\n        dynamics: DynamicsMaskGIT,\n        tokenizer: TokenizerVQVAE,\n        temperature: float,\n        sample_argmax: bool,\n        steps: int,\n    ):\n        self.dynamics = dynamics\n        self.tokenizer = tokenizer\n        self.temperature = temperature\n        self.sample_argmax = sample_argmax\n        self.steps = steps\n\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs)  # (B, S, N, D)\n        mask_token = self.dynamics.mask_token.value  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None]  # (B, S, N, 1)\n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jax.random.categorical(_rng, final_logits)\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n\n\n# FIXME (f.srambical): add conversion script for old checkpoints\ndef restore_genie_components(\n    optimizer: nnx.Optimizer,\n    sharding: jax.sharding.NamedSharding,\n    rng: jax.Array,\n    args,\n):\n    """"""Restore pre-trained Genie components""""""\n    rngs = nnx.Rngs(rng)\n\n    # dummy values since we only use tx to initialize the dummy train states\n    dummy_tx = optax.adamw(\n        learning_rate=optax.constant_schedule(args.max_lr),\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n    )\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        step_format_fixed_length=6,\n    )\n    tokenizer_checkpoint_manager = ocp.CheckpointManager(\n        directory=args.tokenizer_checkpoint,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    dummy_tokenizer = TokenizerVQVAE(\n        in_dim=args.image_channels,\n        model_dim=args.tokenizer_dim,\n        ffn_dim=args.tokenizer_ffn_dim,\n        latent_dim=args.latent_patch_dim,\n        num_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.tokenizer_num_blocks,\n        num_heads=args.tokenizer_num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.dropout,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        rngs=rngs,\n    )\n    dummy_tokenizer_optimizer = nnx.Optimizer(dummy_tokenizer, dummy_tx)\n    dummy_tokenizer_optimizer_state = nnx.state(dummy_tokenizer_optimizer)\n    abstract_sharded_tokenizer_optimizer_state = _create_abstract_sharded_pytree(\n        dummy_tokenizer_optimizer_state, sharding\n    )\n    restored_tokenizer = tokenizer_checkpoint_manager.restore(\n        step=tokenizer_checkpoint_manager.latest_step(),\n        args=ocp.args.Composite(\n            model_state=ocp.args.PyTreeRestore(\n                abstract_sharded_tokenizer_optimizer_state\n            ),\n        ),\n    )[""model_state""]\n    nnx.update(dummy_tokenizer_optimizer.model, restored_tokenizer.model)\n    optimizer.model.tokenizer = dummy_tokenizer_optimizer.model\n    tokenizer_checkpoint_manager.close()\n\n    if args.lam_checkpoint:\n        lam_checkpoint_manager = ocp.CheckpointManager(\n            directory=args.lam_checkpoint,\n            options=checkpoint_options,\n            handler_registry=handler_registry,\n        )\n        dummy_lam = LatentActionModel(\n            in_dim=args.image_channels,\n            model_dim=args.lam_dim,\n            ffn_dim=args.lam_ffn_dim,\n            latent_dim=args.latent_patch_dim,\n            num_latents=args.num_latent_actions,\n            patch_size=args.lam_patch_size,\n            num_blocks=args.lam_num_blocks,\n            num_heads=args.lam_num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n            use_flash_attention=args.use_flash_attention,\n            rngs=rngs,\n        )\n        dummy_lam_optimizer = nnx.Optimizer(dummy_lam, dummy_tx)\n        dummy_lam_optimizer_state = nnx.state(dummy_lam_optimizer)\n        abstract_sharded_lam_optimizer_state = _create_abstract_sharded_pytree(\n            dummy_lam_optimizer_state, sharding\n        )\n        restored_lam_optimizer = lam_checkpoint_manager.restore(\n            step=lam_checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(\n                    abstract_sharded_lam_optimizer_state\n                ),\n            ),\n        )[""model_state""]\n        nnx.update(dummy_lam_optimizer.model, restored_lam_optimizer.model)\n        optimizer.model.lam = dummy_lam_optimizer.model\n        # Remove the LAM decoder to save memory and avoid unnecessary computation.\n        del optimizer.model.lam.decoder\n        lam_checkpoint_manager.close()\n\n    return optimizer\n\n\ndef _create_abstract_sharded_pytree(pytree_template, sharding_spec):\n    """"""Replaces arrays in a pytree with ShapeDtypeStructs having the given sharding.""""""\n\n    def map_fn(leaf_template):\n        if hasattr(leaf_template, ""shape"") and hasattr(leaf_template, ""dtype""):\n            return jax.ShapeDtypeStruct(\n                leaf_template.shape, leaf_template.dtype, sharding=sharding_spec\n            )\n        return leaf_template\n\n    return jax.tree_util.tree_map(map_fn, pytree_template)\n",python,tab
+2,357,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"7:02:23 PM [info] Activating crowd-code\n7:02:23 PM [info] Recording started\n7:02:23 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,587,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"7:02:23 PM [info] Git repository found\n7:02:23 PM [info] Git provider initialized successfully\n7:02:23 PM [info] Initial git state: [object Object]\n",Log,content
+4,2898,"genie.py",0,0,"",python,tab
+5,3378,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+6,5988,"genie.py",0,0,"",python,tab
+7,19218,"genie.py",0,0,"",python,tab
+8,19298,"genie.py",6670,0,"",python,selection_command
+9,56549,"/fast/home/franz.srambical/jafar/sample.py",0,0,"from dataclasses import dataclass\nimport time\nimport os\nimport optax\n\nimport dm_pix as pix\nimport einops\nimport jax\nimport jax.numpy as jnp\nimport flax.linen as nn\nimport numpy as np\nimport orbax.checkpoint as ocp\nfrom PIL import Image, ImageDraw\nimport tyro\nfrom flax import nnx\n\nfrom genie import Genie\nfrom utils.dataloader import get_dataloader\n\n\n@dataclass\nclass Args:\n    # Experiment\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = ""data/coinrun_episodes""\n    checkpoint: str = """"\n    # Sampling\n    batch_size: int = 1\n    maskgit_steps: int = 25\n    temperature: float = 1.0\n    sample_argmax: bool = True\n    start_frame: int = 0\n    # Tokenizer checkpoint\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    # LAM checkpoint\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    # Dynamics checkpoint\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n\n\nargs = tyro.cli(Args)\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n\n    rng = jax.random.PRNGKey(args.seed)\n\n    # --- Load Genie checkpoint ---\n    rngs = nnx.Rngs(rng)\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_latent_actions=args.num_latent_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=False,\n        # Dynamics\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        rngs=rngs,\n    )\n\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        step_format_fixed_length=6,\n    )\n    checkpoint_manager = ocp.CheckpointManager(\n        args.checkpoint,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    dummy_tx = optax.adamw(\n            learning_rate=optax.linear_schedule(0.0001, 0.0001, 10000),\n            b1=0.9,\n            b2=0.9,\n            weight_decay=1e-4,\n            mu_dtype=args.dtype,\n        )\n    dummy_optimizer = nnx.Optimizer(genie, dummy_tx)\n\n    abstract_optimizer = nnx.eval_shape(lambda: dummy_optimizer)\n    abstract_optimizer_state = nnx.state(abstract_optimizer)\n    restored = checkpoint_manager.restore(\n                checkpoint_manager.latest_step(),\n                args=ocp.args.Composite(\n                    model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),\n                ),\n            )\n    restored_optimizer_state = restored[""model_state""]\n    nnx.update(dummy_optimizer, restored_optimizer_state)\n\n    # --- Define sampling function ---\n    # @nnx.jit\n    # @jax.jit\n    def _sampling_fn(model, batch):\n        """"""Runs Genie.sample with pre-defined generation hyper-parameters.""""""\n        return model.sample(\n            batch,\n            args.seq_len,\n            args.maskgit_steps,\n            args.temperature,\n            args.sample_argmax,\n        )\n\n\n    # --- Define autoregressive sampling loop ---\n    def _autoreg_sample(rng, video_batch, action_batch):\n        vid = video_batch[:, : args.start_frame + 1]\n        rng, _rng = jax.random.split(rng)\n        batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n        generated_vid = genie.sample(batch, args.seq_len, args.maskgit_steps, args.temperature, args.sample_argmax)\n        return generated_vid\n\n\n    # --- Get video + latent actions ---\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        args.batch_size,\n        args.image_height,\n        args.image_width,\n        args.image_channels,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    video_batch = next(iter(dataloader))\n    # Get latent actions for all videos in the batch\n    batch = dict(videos=video_batch)\n    action_batch = genie.vq_encode(batch, training=False)  # type: ignore[arg-type]\n    action_batch = jnp.asarray(action_batch).reshape(video_batch.shape[0], args.seq_len - 1, 1)\n\n    # --- Sample + evaluate video ---\n    vid = _autoreg_sample(rng, video_batch, action_batch)\n    gt = video_batch[:, : vid.shape[1]].clip(0, 1).reshape(-1, *video_batch.shape[2:])\n    recon = vid.clip(0, 1).reshape(-1, *vid.shape[2:])\n    # FIXME (f.srambical): investigate why this is needed\n    gt = gt.astype(jnp.float32)\n    ssim = pix.ssim(gt[:, args.start_frame + 1 :], recon[:, args.start_frame + 1 :]).mean()\n    print(f""SSIM: {ssim}"")\n\n    # --- Construct video ---\n    # true_videos = (video_batch * 255).astype(np.uint8)\n    # pred_videos = (vid * 255).astype(np.uint8)\n    # video_comparison = np.zeros((2, *vid.shape), dtype=np.uint8)\n    # video_comparison[0] = true_videos[:, : args.seq_len]\n    # video_comparison[1] = pred_videos\n    # frames = einops.rearrange(video_comparison, ""n b t h w c -> t (b h) (n w) c"")\n\n    # # --- Save video ---\n    # imgs = [Image.fromarray(img) for img in frames]\n    # # Write actions on each frame, on each row (i.e., for each video in the batch, on the GT row)\n    # for t, img in enumerate(imgs[1:]):\n    #     d = ImageDraw.Draw(img)\n    #     for row in range(action_batch.shape[0]):\n    #         action = action_batch[row, t, 0]\n    #         y_offset = row * video_batch.shape[2] + 2\n    #         d.text((2, y_offset), f""{action}"", fill=255)\n    # imgs[0].save(\n    #     f""generation_{time.time()}.gif"",\n    #     save_all=True,\n    #     append_images=imgs[1:],\n    #     duration=250,\n    #     loop=0,\n    # )\n",python,tab
+10,56549,"/fast/home/franz.srambical/jafar/sample.py",5611,0,"",python,selection_command
+11,58027,"/fast/home/franz.srambical/jafar/sample.py",4644,0,"",python,selection_command
+12,59949,"/fast/home/franz.srambical/jafar/genie.py",0,0,"from typing import Dict, Any\n\nimport optax\nimport jax\nimport jax.numpy as jnp\nimport flax.nnx as nnx\nfrom flax.training.train_state import TrainState\nimport orbax.checkpoint as ocp\n\nfrom models.dynamics import DynamicsMaskGIT\nfrom models.lam import LatentActionModel\nfrom models.tokenizer import TokenizerVQVAE\n\nimport grain\n\n\nclass Genie(nnx.Module):\n    """"""Genie model""""""\n\n    def __init__(\n        self,\n        in_dim: int,\n        tokenizer_dim: int,\n        tokenizer_ffn_dim: int,\n        latent_patch_dim: int,\n        num_patch_latents: int,\n        patch_size: int,\n        tokenizer_num_blocks: int,\n        tokenizer_num_heads: int,\n        lam_dim: int,\n        lam_ffn_dim: int,\n        latent_action_dim: int,\n        num_latent_actions: int,\n        lam_patch_size: int,\n        lam_num_blocks: int,\n        lam_num_heads: int,\n        lam_co_train: bool,\n        dyna_dim: int,\n        dyna_ffn_dim: int,\n        dyna_num_blocks: int,\n        dyna_num_heads: int,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n        dropout: float = 0.0,\n        mask_limit: float = 0.0,\n    ):\n        # --- Tokenizer ---\n        self.in_dim = in_dim\n        self.tokenizer_dim = tokenizer_dim\n        self.tokenizer_ffn_dim = tokenizer_ffn_dim\n        self.latent_patch_dim = latent_patch_dim\n        self.num_patch_latents = num_patch_latents\n        self.patch_size = patch_size\n        self.tokenizer_num_blocks = tokenizer_num_blocks\n        self.tokenizer_num_heads = tokenizer_num_heads\n        # --- LAM ---\n        self.lam_dim = lam_dim\n        self.lam_ffn_dim = lam_ffn_dim\n        self.latent_action_dim = latent_action_dim\n        self.num_latent_actions = num_latent_actions\n        self.lam_patch_size = lam_patch_size\n        self.lam_num_blocks = lam_num_blocks\n        self.lam_num_heads = lam_num_heads\n        self.lam_co_train = lam_co_train\n        # --- Dynamics ---\n        self.dyna_dim = dyna_dim\n        self.dyna_ffn_dim = dyna_ffn_dim\n        self.dyna_num_blocks = dyna_num_blocks\n        self.dyna_num_heads = dyna_num_heads\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.dropout = dropout\n        self.mask_limit = mask_limit\n\n        self.tokenizer = TokenizerVQVAE(\n            in_dim=self.in_dim,\n            model_dim=self.tokenizer_dim,\n            ffn_dim=self.tokenizer_ffn_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_patch_latents,\n            patch_size=self.patch_size,\n            num_blocks=self.tokenizer_num_blocks,\n            num_heads=self.tokenizer_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            use_flash_attention=self.use_flash_attention,\n            rngs=rngs,\n        )\n        self.lam = LatentActionModel(\n            in_dim=self.in_dim,\n            model_dim=self.lam_dim,\n            ffn_dim=self.lam_ffn_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_latent_actions,\n            patch_size=self.lam_patch_size,\n            num_blocks=self.lam_num_blocks,\n            num_heads=self.lam_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            use_flash_attention=self.use_flash_attention,\n            rngs=rngs,\n        )\n        self.dynamics = DynamicsMaskGIT(\n            model_dim=self.dyna_dim,\n            ffn_dim=self.dyna_ffn_dim,\n            num_latents=self.num_patch_latents,\n            latent_action_dim=self.latent_action_dim,\n            num_blocks=self.dyna_num_blocks,\n            num_heads=self.dyna_num_heads,\n            dropout=self.dropout,\n            mask_limit=self.mask_limit,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            use_flash_attention=self.use_flash_attention,\n            rngs=rngs,\n        )\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        tokenizer_outputs = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        lam_outputs = self.lam.vq_encode(batch[""videos""], training=False)\n        latent_actions = jax.lax.cond(\n            self.lam_co_train,\n            lambda: lam_outputs[""z_q""],\n            lambda: jax.lax.stop_gradient(lam_outputs[""z_q""]),\n        )\n        outputs = dict(\n            video_tokens=jax.lax.stop_gradient(tokenizer_outputs[""indices""]),\n            latent_actions=latent_actions,\n        )\n        outputs[""mask_rng""] = batch[""mask_rng""]\n        dyna_outputs = self.dynamics(outputs, training)\n        outputs.update(dyna_outputs)\n        mle_indices = jnp.argmax(outputs[""token_logits""], axis=-1)\n        outputs[""recon""] = self.tokenizer.decode(\n            mle_indices, batch[""videos""].shape[2:4]\n        )\n        outputs[""lam_indices""] = lam_outputs[""indices""]\n        return outputs\n\n    def sample(\n        self,\n        batch: Dict[str, Any],\n        seq_len: int,\n        steps: int = 25,\n        temperature: float = 1,\n        sample_argmax: bool = False,\n    ) -> Any:\n        """"""\n        Autoregressively samples up to `seq_len` future frames, following Figure 8 of the paper.\n\n        - Input frames are tokenized once.\n        - Future frames are generated autoregressively in token space.\n        - All frames are detokenized in a single pass.\n\n        Note:\n        - For interactive or step-wise sampling, detokenization should occur after each action.\n        - To maintain consistent tensor shapes across timesteps, all current and future frames are decoded at every step.\n        - Temporal causal structure is preserved by\n            a) reapplying the mask before each decoding step.\n            b) a temporal causal mask is applied within each ST-transformer block.\n\n        Dimension keys:\n            B: batch size\n            T: number of input (conditioning) frames\n            N: patches per frame\n            S: sequence length\n            A: action space\n            D: model latent dimension\n        """"""\n        # --- Encode videos and actions ---\n        tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        token_idxs = tokenizer_out[""indices""]  # (B, T, N)\n        B, T, N = token_idxs.shape\n        pad_shape = (B, seq_len - T, N)\n        pad = jnp.zeros(pad_shape, dtype=token_idxs.dtype)\n        token_idxs = jnp.concatenate([token_idxs, pad], axis=1)  # (B, S, N)\n        action_tokens = self.lam.vq.get_codes(batch[""latent_actions""])\n\n        # Define the inner MaskGIT loop using nnx.scan\n        maskgit_step = MaskGITStep(\n            dynamics=self.dynamics,\n            tokenizer=self.tokenizer,\n            temperature=temperature,\n            sample_argmax=sample_argmax,\n            steps=steps,\n        )\n\n        def maskgit_scan_fn(module, carry, x):\n            new_carry, _ = module(carry, x)\n            return new_carry, None\n\n        MaskGITLoop = nnx.scan(\n            maskgit_scan_fn,\n            in_axes=(None, nnx.Carry, 0),  # (module, carry, x)\n            out_axes=(nnx.Carry, None),    # (new_carry, None)\n        )\n\n        # Define the outer autoregressive loop's body function\n        def generation_step_fn(carry, step_t):\n            rng, current_token_idxs = carry\n            rng, step_rng = jax.random.split(rng)\n\n            # Mask current and future frames (i.e., t >= step_t)\n            mask = jnp.arange(seq_len) >= step_t  # (S,)\n            mask = jnp.broadcast_to(mask[None, :, None], (B, seq_len, N)).astype(bool)  # (B, S, N)\n            masked_token_idxs = current_token_idxs * ~mask\n\n            # --- Initialize and run MaskGIT loop ---\n            init_carry_maskgit = (\n                step_rng,\n                masked_token_idxs,\n                mask,\n                action_tokens,\n            )\n            final_carry_maskgit, _ = MaskGITLoop(\n                maskgit_step, init_carry_maskgit, jnp.arange(steps)\n            )\n            updated_token_idxs = final_carry_maskgit[1]\n            new_carry = (rng, updated_token_idxs)\n            return new_carry, None\n\n        # --- Run the autoregressive generation using jax.lax.scan ---\n        initial_carry = (batch[""rng""], token_idxs)\n        timesteps_to_scan = jnp.arange(T, seq_len)\n        final_carry, _ = jax.lax.scan(\n            generation_step_fn, initial_carry, timesteps_to_scan\n        )\n        final_token_idxs = final_carry[1]\n\n        # --- Decode all tokens at once at the end ---\n        final_frames = self.tokenizer.decode(\n            final_token_idxs,\n            video_hw=batch[""videos""].shape[2:4],\n        )\n        return final_frames\n\n    def vq_encode(self, batch, training) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\n\nclass MaskGITStep(nnx.Module):\n    def __init__(\n        self,\n        dynamics: DynamicsMaskGIT,\n        tokenizer: TokenizerVQVAE,\n        temperature: float,\n        sample_argmax: bool,\n        steps: int,\n    ):\n        self.dynamics = dynamics\n        self.tokenizer = tokenizer\n        self.temperature = temperature\n        self.sample_argmax = sample_argmax\n        self.steps = steps\n\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs)  # (B, S, N, D)\n        mask_token = self.dynamics.mask_token.value  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None]  # (B, S, N, 1)\n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jax.random.categorical(_rng, final_logits)\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n\n\n# FIXME (f.srambical): add conversion script for old checkpoints\ndef restore_genie_components(\n    optimizer: nnx.Optimizer,\n    sharding: jax.sharding.NamedSharding,\n    rng: jax.Array,\n    args,\n):\n    """"""Restore pre-trained Genie components""""""\n    rngs = nnx.Rngs(rng)\n\n    # dummy values since we only use tx to initialize the dummy train states\n    dummy_tx = optax.adamw(\n        learning_rate=optax.constant_schedule(args.max_lr),\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n    )\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        step_format_fixed_length=6,\n    )\n    tokenizer_checkpoint_manager = ocp.CheckpointManager(\n        directory=args.tokenizer_checkpoint,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    dummy_tokenizer = TokenizerVQVAE(\n        in_dim=args.image_channels,\n        model_dim=args.tokenizer_dim,\n        ffn_dim=args.tokenizer_ffn_dim,\n        latent_dim=args.latent_patch_dim,\n        num_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.tokenizer_num_blocks,\n        num_heads=args.tokenizer_num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.dropout,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        rngs=rngs,\n    )\n    dummy_tokenizer_optimizer = nnx.Optimizer(dummy_tokenizer, dummy_tx)\n    dummy_tokenizer_optimizer_state = nnx.state(dummy_tokenizer_optimizer)\n    abstract_sharded_tokenizer_optimizer_state = _create_abstract_sharded_pytree(\n        dummy_tokenizer_optimizer_state, sharding\n    )\n    restored_tokenizer = tokenizer_checkpoint_manager.restore(\n        step=tokenizer_checkpoint_manager.latest_step(),\n        args=ocp.args.Composite(\n            model_state=ocp.args.PyTreeRestore(\n                abstract_sharded_tokenizer_optimizer_state\n            ),\n        ),\n    )[""model_state""]\n    nnx.update(dummy_tokenizer_optimizer.model, restored_tokenizer.model)\n    optimizer.model.tokenizer = dummy_tokenizer_optimizer.model\n    tokenizer_checkpoint_manager.close()\n\n    if args.lam_checkpoint:\n        lam_checkpoint_manager = ocp.CheckpointManager(\n            directory=args.lam_checkpoint,\n            options=checkpoint_options,\n            handler_registry=handler_registry,\n        )\n        dummy_lam = LatentActionModel(\n            in_dim=args.image_channels,\n            model_dim=args.lam_dim,\n            ffn_dim=args.lam_ffn_dim,\n            latent_dim=args.latent_patch_dim,\n            num_latents=args.num_latent_actions,\n            patch_size=args.lam_patch_size,\n            num_blocks=args.lam_num_blocks,\n            num_heads=args.lam_num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n            use_flash_attention=args.use_flash_attention,\n            rngs=rngs,\n        )\n        dummy_lam_optimizer = nnx.Optimizer(dummy_lam, dummy_tx)\n        dummy_lam_optimizer_state = nnx.state(dummy_lam_optimizer)\n        abstract_sharded_lam_optimizer_state = _create_abstract_sharded_pytree(\n            dummy_lam_optimizer_state, sharding\n        )\n        restored_lam_optimizer = lam_checkpoint_manager.restore(\n            step=lam_checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(\n                    abstract_sharded_lam_optimizer_state\n                ),\n            ),\n        )[""model_state""]\n        nnx.update(dummy_lam_optimizer.model, restored_lam_optimizer.model)\n        optimizer.model.lam = dummy_lam_optimizer.model\n        # Remove the LAM decoder to save memory and avoid unnecessary computation.\n        del optimizer.model.lam.decoder\n        lam_checkpoint_manager.close()\n\n    return optimizer\n\n\ndef _create_abstract_sharded_pytree(pytree_template, sharding_spec):\n    """"""Replaces arrays in a pytree with ShapeDtypeStructs having the given sharding.""""""\n\n    def map_fn(leaf_template):\n        if hasattr(leaf_template, ""shape"") and hasattr(leaf_template, ""dtype""):\n            return jax.ShapeDtypeStruct(\n                leaf_template.shape, leaf_template.dtype, sharding=sharding_spec\n            )\n        return leaf_template\n\n    return jax.tree_util.tree_map(map_fn, pytree_template)\n",python,tab
+13,59950,"/fast/home/franz.srambical/jafar/genie.py",7076,0,"",python,selection_command
+14,77428,"/fast/home/franz.srambical/jafar/sample.py",0,0,"",python,tab
+15,79472,"genie.py",0,0,"",python,tab
+16,96262,"genie.py",7213,0,"",python,selection_command
+17,97679,"genie.py",7201,63,"            out_axes=(nnx.Carry, 0),       # (new_carry, None)\n",python,content
+18,113555,"/fast/home/franz.srambical/jafar/genie.py",0,0,"",python,tab
+19,113557,"/fast/home/franz.srambical/jafar/genie.py",7201,0,"",python,selection_command
+20,116189,"genie.py",0,0,"",python,tab
+21,117566,"genie.py",7258,0,"",python,selection_mouse
+22,155903,"/fast/home/franz.srambical/jafar/genie.py",0,0,"",python,tab
+23,157626,"/fast/home/franz.srambical/jafar/genie.py",7201,62,"            out_axes=(nnx.Carry, None),    # (new_carry, None)",python,content
+24,159605,"genie.py",0,0,"",python,tab
+25,179385,"/fast/home/franz.srambical/jafar/genie.py",0,0,"",python,tab
+26,179385,"/fast/home/franz.srambical/jafar/genie.py",7979,0,"",python,selection_command
+27,213168,"/fast/home/franz.srambical/jafar/genie.py",7979,117,"            final_carry_maskgit, _ = jax.lax.scan(\n                maskgit_step_fn, init_carry_maskgit, jnp.arange(steps)",python,content
+28,213169,"/fast/home/franz.srambical/jafar/genie.py",7076,199,"",python,content
+29,213169,"/fast/home/franz.srambical/jafar/genie.py",6949,90,"            new_carry = (rng, token_idxs, new_mask, action_tokens)",python,content
+30,213169,"/fast/home/franz.srambical/jafar/genie.py",6670,277,"        # Define the inner MaskGIT loop function\n        def maskgit_step_fn(carry, step):\n            rng, token_idxs, mask, action_tokens = carry\n            N = token_idxs.shape[2]\n\n            # --- Construct + encode video ---\n            vid_embed = self.dynamics.patch_embed(token_idxs)  # (B, S, N, D)\n            mask_token = self.dynamics.mask_token.value  # (1, 1, 1, D,)\n            mask_expanded = mask[..., None]  # (B, S, N, 1)\n            vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n            # --- Predict transition ---\n            act_embed = self.dynamics.action_up(action_tokens)\n            vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n            unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (steps * 2))\n            step_temp = temperature * (1.0 - unmasked_ratio)\n            final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n            # --- Sample new tokens for final frame ---\n            if sample_argmax:\n                sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n            else:\n                rng, _rng = jax.random.split(rng)\n                sampled_token_idxs = jax.random.categorical(_rng, final_logits)\n            gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n            final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n            final_token_probs += ~mask\n            # Update masked tokens only\n            token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n            # --- Update mask ---\n            num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n            idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n            sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n            mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n            new_mask = mask_update_fn(mask, sorted_idxs)",python,content
+31,223860,"/fast/home/franz.srambical/jafar/genie.py",9430,121,"            final_carry_maskgit, _ = MaskGITLoop(\n                maskgit_step, init_carry_maskgit, jnp.arange(steps)",python,content
+32,223860,"/fast/home/franz.srambical/jafar/genie.py",8726,0,"        MaskGITLoop = nnx.scan(\n            maskgit_scan_fn,\n            in_axes=(None, nnx.Carry, 0),  # (module, carry, x)\n            out_axes=(nnx.Carry, None),    # (new_carry, None)\n        )\n\n",python,content
+33,223860,"/fast/home/franz.srambical/jafar/genie.py",8623,66,"        def maskgit_scan_fn(module, carry, x):\n            new_carry, _ = module(carry, x)",python,content
+34,223860,"/fast/home/franz.srambical/jafar/genie.py",6670,1951,"        # Define the inner MaskGIT loop using nnx.scan\n        maskgit_step = MaskGITStep(\n            dynamics=self.dynamics,\n            tokenizer=self.tokenizer,\n            temperature=temperature,\n            sample_argmax=sample_argmax,\n            steps=steps,\n        )",python,content
+35,224039,"/fast/home/franz.srambical/jafar/genie.py",11216,94,"",python,content
+36,224039,"/fast/home/franz.srambical/jafar/genie.py",9013,2202,"",python,content
+37,224039,"/fast/home/franz.srambical/jafar/genie.py",7979,117,"            final_carry_maskgit, _ = jax.lax.scan(\n                maskgit_step_fn, init_carry_maskgit, jnp.arange(steps)",python,content
+38,224039,"/fast/home/franz.srambical/jafar/genie.py",7076,197,"            # --- Predict transition ---\n            act_embed = self.dynamics.action_up(action_tokens)\n            vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n            unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (steps * 2))\n            step_temp = temperature * (1.0 - unmasked_ratio)\n            final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n            # --- Sample new tokens for final frame ---\n            if sample_argmax:\n                sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n            else:\n                rng, _rng = jax.random.split(rng)\n                sampled_token_idxs = jax.random.categorical(_rng, final_logits)\n            gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n            final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n            final_token_probs += ~mask\n            # Update masked tokens only\n            token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n            # --- Update mask ---\n            num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n            idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n            sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n            mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n            new_mask = mask_update_fn(mask, sorted_idxs)\n\n            new_carry = (rng, token_idxs, new_mask, action_tokens)\n            return new_carry, None",python,content
+39,224039,"/fast/home/franz.srambical/jafar/genie.py",6949,125,"            # --- Construct + encode video ---\n            vid_embed = self.dynamics.patch_embed(token_idxs)  # (B, S, N, D)\n            mask_token = self.dynamics.mask_token.value  # (1, 1, 1, D,)\n            mask_expanded = mask[..., None]  # (B, S, N, 1)\n            vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)",python,content
+40,224039,"/fast/home/franz.srambical/jafar/genie.py",6670,277,"        # Define the inner MaskGIT loop function\n        def maskgit_step_fn(carry, step):\n            rng, token_idxs, mask, action_tokens = carry\n            N = token_idxs.shape[2]",python,content
+41,272905,"/fast/home/franz.srambical/jafar/genie.py",6670,0,"",python,selection_command
+42,380998,"sample.py",0,0,"from dataclasses import dataclass\nimport time\nimport os\nimport optax\n\nimport dm_pix as pix\nimport einops\nimport jax\nimport jax.numpy as jnp\nimport flax.linen as nn\nimport numpy as np\nimport orbax.checkpoint as ocp\nfrom PIL import Image, ImageDraw\nimport tyro\nfrom flax import nnx\n\nfrom genie import Genie\nfrom utils.dataloader import get_dataloader\n\n\n@dataclass\nclass Args:\n    # Experiment\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = ""data/coinrun_episodes""\n    checkpoint: str = """"\n    # Sampling\n    batch_size: int = 1\n    maskgit_steps: int = 25\n    temperature: float = 1.0\n    sample_argmax: bool = True\n    start_frame: int = 0\n    # Tokenizer checkpoint\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    # LAM checkpoint\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    # Dynamics checkpoint\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n\n\nargs = tyro.cli(Args)\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n\n    rng = jax.random.PRNGKey(args.seed)\n\n    # --- Load Genie checkpoint ---\n    rngs = nnx.Rngs(rng)\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_latent_actions=args.num_latent_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=False,\n        # Dynamics\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        rngs=rngs,\n    )\n\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        step_format_fixed_length=6,\n    )\n    checkpoint_manager = ocp.CheckpointManager(\n        args.checkpoint,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    dummy_tx = optax.adamw(\n            learning_rate=optax.linear_schedule(0.0001, 0.0001, 10000),\n            b1=0.9,\n            b2=0.9,\n            weight_decay=1e-4,\n            mu_dtype=args.dtype,\n        )\n    dummy_optimizer = nnx.Optimizer(genie, dummy_tx)\n\n    abstract_optimizer = nnx.eval_shape(lambda: dummy_optimizer)\n    abstract_optimizer_state = nnx.state(abstract_optimizer)\n    restored = checkpoint_manager.restore(\n                checkpoint_manager.latest_step(),\n                args=ocp.args.Composite(\n                    model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),\n                ),\n            )\n    restored_optimizer_state = restored[""model_state""]\n    nnx.update(dummy_optimizer, restored_optimizer_state)\n\n    # --- Define sampling function ---\n    # @nnx.jit\n    # @jax.jit\n    def _sampling_fn(model, batch):\n        """"""Runs Genie.sample with pre-defined generation hyper-parameters.""""""\n        return model.sample(\n            batch,\n            args.seq_len,\n            args.maskgit_steps,\n            args.temperature,\n            args.sample_argmax,\n        )\n\n\n    # --- Define autoregressive sampling loop ---\n    def _autoreg_sample(rng, video_batch, action_batch):\n        vid = video_batch[:, : args.start_frame + 1]\n        rng, _rng = jax.random.split(rng)\n        batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n        generated_vid = genie.sample(batch, args.seq_len, args.maskgit_steps, args.temperature, args.sample_argmax)\n        return generated_vid\n\n\n    # --- Get video + latent actions ---\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        args.batch_size,\n        args.image_height,\n        args.image_width,\n        args.image_channels,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    video_batch = next(iter(dataloader))\n    # Get latent actions for all videos in the batch\n    batch = dict(videos=video_batch)\n    action_batch = genie.vq_encode(batch, training=False)  # type: ignore[arg-type]\n    action_batch = jnp.asarray(action_batch).reshape(video_batch.shape[0], args.seq_len - 1, 1)\n\n    # --- Sample + evaluate video ---\n    vid = _autoreg_sample(rng, video_batch, action_batch)\n    gt = video_batch[:, : vid.shape[1]].clip(0, 1).reshape(-1, *video_batch.shape[2:])\n    recon = vid.clip(0, 1).reshape(-1, *vid.shape[2:])\n    # FIXME (f.srambical): investigate why this is needed\n    gt = gt.astype(jnp.float32)\n    ssim = pix.ssim(gt[:, args.start_frame + 1 :], recon[:, args.start_frame + 1 :]).mean()\n    print(f""SSIM: {ssim}"")\n\n    # --- Construct video ---\n    # true_videos = (video_batch * 255).astype(np.uint8)\n    # pred_videos = (vid * 255).astype(np.uint8)\n    # video_comparison = np.zeros((2, *vid.shape), dtype=np.uint8)\n    # video_comparison[0] = true_videos[:, : args.seq_len]\n    # video_comparison[1] = pred_videos\n    # frames = einops.rearrange(video_comparison, ""n b t h w c -> t (b h) (n w) c"")\n\n    # # --- Save video ---\n    # imgs = [Image.fromarray(img) for img in frames]\n    # # Write actions on each frame, on each row (i.e., for each video in the batch, on the GT row)\n    # for t, img in enumerate(imgs[1:]):\n    #     d = ImageDraw.Draw(img)\n    #     for row in range(action_batch.shape[0]):\n    #         action = action_batch[row, t, 0]\n    #         y_offset = row * video_batch.shape[2] + 2\n    #         d.text((2, y_offset), f""{action}"", fill=255)\n    # imgs[0].save(\n    #     f""generation_{time.time()}.gif"",\n    #     save_all=True,\n    #     append_images=imgs[1:],\n    #     duration=250,\n    #     loop=0,\n    # )\n",python,tab
+43,382134,"sample.py",7041,0,"",python,selection_command
+44,382137,"sample.py",7023,0,"",python,selection_command
+45,382152,"sample.py",6999,0,"",python,selection_command
+46,382225,"sample.py",6965,0,"",python,selection_command
+47,382231,"sample.py",6940,0,"",python,selection_command
+48,382323,"sample.py",6897,0,"",python,selection_command
+49,382340,"sample.py",6877,0,"",python,selection_command
+50,382345,"sample.py",6818,0,"",python,selection_command
+51,382381,"sample.py",6762,0,"",python,selection_command
+52,382416,"sample.py",6715,0,"",python,selection_command
+53,382427,"sample.py",6664,0,"",python,selection_command
+54,382469,"sample.py",6630,0,"",python,selection_command
+55,382583,"sample.py",6589,0,"",python,selection_command
+56,382584,"sample.py",6489,0,"",python,selection_command
+57,382658,"sample.py",6435,0,"",python,selection_command
+58,382658,"sample.py",6408,0,"",python,selection_command
+59,382659,"sample.py",6407,0,"",python,selection_command
+60,382692,"sample.py",6323,0,"",python,selection_command
+61,382707,"sample.py",6283,0,"",python,selection_command
+62,382729,"sample.py",6224,0,"",python,selection_command
+63,382778,"sample.py",6157,0,"",python,selection_command
+64,382822,"sample.py",6108,0,"",python,selection_command
+65,382836,"sample.py",6051,0,"",python,selection_command
+66,383110,"sample.py",6021,0,"",python,selection_command
+67,409372,"sample.py",0,0,"",python,tab
+68,423163,"sample.py",6020,0,"",python,selection_command
+69,423280,"sample.py",5993,0,"",python,selection_command
+70,456144,"genie.py",0,0,"",python,tab
+71,456145,"genie.py",6670,0,"",python,selection_command
+72,461257,"sample.py",0,0,"",python,tab
+73,464425,"sample.py",6969,0,"",python,selection_command
+74,464665,"sample.py",7049,0,"",python,selection_command
+75,465095,"sample.py",6228,0,"",python,selection_command
+76,465860,"sample.py",5207,0,"",python,selection_command
+77,466191,"sample.py",4424,0,"",python,selection_command
+78,466192,"sample.py",3753,0,"",python,selection_command
+79,466193,"sample.py",2998,0,"",python,selection_command
+80,466194,"sample.py",2219,0,"",python,selection_command
+81,466211,"sample.py",1453,0,"",python,selection_command
+82,466211,"sample.py",917,0,"",python,selection_command
+83,466304,"sample.py",351,0,"",python,selection_command
+84,466305,"sample.py",917,0,"",python,selection_command
+85,466306,"sample.py",1453,0,"",python,selection_command
+86,466306,"sample.py",2219,0,"",python,selection_command
+87,466542,"sample.py",2998,0,"",python,selection_command
+88,468353,"/fast/home/franz.srambical/jafar/genie.py",0,0,"",python,tab
+89,470581,"/fast/home/franz.srambical/jafar/genie.py",7766,0,"",python,selection_command
+90,470693,"/fast/home/franz.srambical/jafar/genie.py",8930,0,"",python,selection_command
+91,470894,"/fast/home/franz.srambical/jafar/genie.py",9838,0,"",python,selection_command
+92,471054,"/fast/home/franz.srambical/jafar/genie.py",10437,0,"",python,selection_command
+93,475752,"sample.py",0,0,"",python,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-accd586c-9376-4507-a888-197a6c40bdf51757184416102-2025_09_06-20.47.03.130/source.csv

@@ -0,0 +1,4 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"utils/nn.py",0,0,"import math\nfrom typing import Tuple, Callable, List\n\nfrom flax import nnx\nimport jax\nimport jax.numpy as jnp\nimport einops\n\n\nclass SpatioTemporalPositionalEncoding(nnx.Module):\n    """"""\n    Applies separate sinusoidal positional encodings to the temporal and spatial dimensions.\n    """"""\n\n    def __init__(self, d_model: int, max_len: int = 5000):\n        self.d_model = d_model\n        self.max_len = max_len\n\n        pe = jnp.zeros((self.max_len, self.d_model))\n        position = jnp.arange(0, self.max_len, dtype=jnp.float32)[:, None]\n        div_term = jnp.exp(\n            jnp.arange(0, self.d_model, 2) * (-math.log(10000.0) / self.d_model)\n        )\n        pe = pe.at[:, 0::2].set(jnp.sin(position * div_term))\n        pe = pe.at[:, 1::2].set(jnp.cos(position * div_term))\n        self.pe = nnx.Variable(pe)\n\n    def __call__(self, x: jax.Array) -> jax.Array:\n        """"""\n        Args:\n            x: The input tensor of shape (Batch, Time, Space, Dimension).\n\n        Returns:\n            The input tensor with positional encodings added.\n        """"""\n        assert x.ndim == 4, f""Input must be 4-dimensional, but got shape {x.shape}""\n\n        num_timesteps = x.shape[1]\n        num_spatial_patches = x.shape[2]\n\n        # Temporal positional encoding: (1, T, 1, D)\n        temporal_pe = self.pe.value[None, :num_timesteps, None, :]\n        x = x + temporal_pe\n\n        # Spatial positional encoding: (1, 1, S, D)\n        spatial_pe = self.pe.value[None, None, :num_spatial_patches, :]\n        x = x + spatial_pe\n\n        return x\n\n\nclass STBlock(nnx.Module):\n    def __init__(\n        self,\n        dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool,\n        sow_activations: bool,\n    ):\n        self.dim = dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=False\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.dim,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=False,\n        )\n\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(self, x_BTNM: jax.Array) -> jax.Array:\n        # --- Spatial attention ---\n        z_BTNM = self.spatial_norm(x_BTNM)\n        z_BTNM = self.spatial_attention(z_BTNM, sow_weights=self.sow_weights)\n        x_BTNM = x_BTNM + z_BTNM\n\n        # --- Temporal attention ---\n        x_BNTM = x_BTNM.swapaxes(1, 2)\n        z_BNTM = self.temporal_norm(x_BNTM)\n        z_BNTM = self.temporal_attention(z_BNTM, sow_weights=self.sow_weights)\n        x_BNTM = x_BNTM + z_BNTM\n        x_BTNM = x_BNTM.swapaxes(1, 2)\n\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n        if self.sow_activations:\n            self.sow(nnx.Intermediate, ""activations"", x_BTNM)\n        return x_BTNM\n\n\nclass STTransformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        V: vocabulary size\n    """"""\n\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool = False,\n        sow_activations: bool = False,\n        sow_logits: bool = False,\n        max_len: int = 5000,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_logits = sow_logits\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n\n        self.pos_enc = SpatioTemporalPositionalEncoding(self.model_dim, max_len=max_len)\n\n        self.blocks = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                STBlock(\n                    dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    rngs=rngs,\n                    sow_weights=self.sow_weights,\n                    sow_activations=self.sow_activations,\n                )\n            )\n\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(self, x_BTNI: jax.Array) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n        x_BTNM = self.pos_enc(x_BTNM)\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM)\n\n        x_BTNV = self.output_dense(x_BTNM)\n        if self.sow_logits:\n            self.sow(nnx.Intermediate, ""logits"", x_BTNV)\n        return x_BTNV\n\n\nclass TransformerBlock(nnx.Module):\n    def __init__(\n        self,\n        model_dim: int,\n        ffn_dim: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n        sow_weights: bool,\n        sow_activations: bool,\n    ):\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.decode = decode\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.temporal_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.spatial_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.ffn_norm = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.temporal_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.spatial_attention = nnx.MultiHeadAttention(\n            num_heads=self.num_heads,\n            in_features=self.model_dim,\n            qkv_features=self.model_dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            rngs=rngs,\n            decode=self.decode,\n        )\n        self.ffn_dense1 = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.ffn_dense2 = nnx.Linear(\n            in_features=self.ffn_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    @nnx.remat\n    def __call__(\n        self, x_BTNM: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None\n    ) -> jax.Array:\n        # --- Spatial attention ---\n        B, T, N, M = x_BTNM.shape\n        z_FNM = einops.rearrange(x_BTNM, ""b t n m -> (b t) n m"")\n        z_FNM = self.spatial_norm(z_FNM)\n        z_FNM = self.spatial_attention(z_FNM, sow_weights=self.sow_weights)\n        z_BTNM = einops.rearrange(z_FNM, ""(b t) n m -> b t n m"", t=T)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Temporal attention ---\n        z_PTM = einops.rearrange(x_BTNM, ""b t n m -> (b n) t m"")\n        z_PTM = self.temporal_norm(z_PTM)\n        z_PTM = self.temporal_attention(z_PTM, sow_weights=self.sow_weights)\n        z_BTNM = einops.rearrange(z_PTM, ""(b n) t m -> b t n m"", n=N)\n        x_BTNM = x_BTNM + z_BTNM\n        # --- Feedforward ---\n        z_BTNM = self.ffn_norm(x_BTNM)\n        z_BTND = self.ffn_dense1(z_BTNM)\n        z_BTND = jax.nn.gelu(z_BTND)\n        z_BTNM = self.ffn_dense2(z_BTND)\n        x_BTNM = x_BTNM + z_BTNM\n        if self.sow_activations:\n            self.sow(nnx.Intermediate, ""activations"", x_BTNM)\n\n        return x_BTNM\n\n\nclass Transformer(nnx.Module):\n    """"""\n    Dimension keys:\n        B: batch size\n        T: number of frames\n        N: number of patches per frame\n        I: number of input features\n        M: model dimension\n        D: FFN dimension\n        V: vocabulary size\n        F: number of frames in batch\n        P: number of patch positions in batch\n    """"""\n\n    def __init__(\n        self,\n        input_dim: int,\n        model_dim: int,\n        ffn_dim: int,\n        out_dim: int,\n        num_blocks: int,\n        num_heads: int,\n        dropout: float,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n        sow_logits: bool = False,\n        sow_weights: bool = False,\n        sow_activations: bool = False,\n        max_len: int = 5000,\n    ):\n        self.input_dim = input_dim\n        self.model_dim = model_dim\n        self.ffn_dim = ffn_dim\n        self.out_dim = out_dim\n        self.num_blocks = num_blocks\n        self.num_heads = num_heads\n        self.dropout = dropout\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.sow_logits = sow_logits\n        self.sow_weights = sow_weights\n        self.sow_activations = sow_activations\n\n        self.input_norm1 = nnx.LayerNorm(\n            num_features=self.input_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n        self.input_dense = nnx.Linear(\n            in_features=self.input_dim,\n            out_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n        self.input_norm2 = nnx.LayerNorm(\n            num_features=self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.param_dtype,  # layer norm in full precision\n            rngs=rngs,\n        )\n\n        self.pos_enc = SpatioTemporalPositionalEncoding(self.model_dim, max_len=max_len)\n\n        self.blocks: List[TransformerBlock] = []\n        for _ in range(self.num_blocks):\n            self.blocks.append(\n                TransformerBlock(\n                    model_dim=self.model_dim,\n                    ffn_dim=self.ffn_dim,\n                    num_heads=self.num_heads,\n                    dropout=self.dropout,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                    use_flash_attention=self.use_flash_attention,\n                    decode=decode,\n                    sow_weights=self.sow_weights,\n                    sow_activations=self.sow_activations,\n                    rngs=rngs,\n                )\n            )\n        self.output_dense = nnx.Linear(\n            in_features=self.model_dim,\n            out_features=self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            rngs=rngs,\n        )\n\n    def __call__(\n        self, x_BTNI: jax.Array, pos_index: Tuple[jax.Array, jax.Array] | None = None\n    ) -> jax.Array:\n        x_BTNI = self.input_norm1(x_BTNI)\n        x_BTNM = self.input_dense(x_BTNI)\n        x_BTNM = self.input_norm2(x_BTNM)\n        x_BTNM = self.pos_enc(x_BTNM)\n        for block in self.blocks:\n            x_BTNM = block(x_BTNM, pos_index)\n\n        x_BTNV = self.output_dense(x_BTNM)\n        if self.sow_logits:\n            self.sow(nnx.Intermediate, ""logits"", x_BTNV)\n        return x_BTNV\n\n\ndef normalize(x: jax.Array) -> jax.Array:\n    return x / (jnp.linalg.norm(x, ord=2, axis=-1, keepdims=True) + 1e-8)\n\n\nclass VectorQuantizer(nnx.Module):\n    """"""\n    Dimension keys:\n        D: B * T * N\n        K: number of latents\n        L: latent dimension\n    """"""\n\n    def __init__(\n        self,\n        latent_dim: int,\n        num_latents: int,\n        dropout: float,\n        dtype: jnp.dtype,\n        rngs: nnx.Rngs,\n    ):\n        self.latent_dim = latent_dim\n        self.num_latents = num_latents\n        self.dropout = dropout\n        self.dtype = dtype\n\n        self.codebook = nnx.Param(\n            nnx.initializers.lecun_uniform()(\n                rngs.params(), (self.num_latents, self.latent_dim)\n            )\n        )\n        self.drop = nnx.Dropout(self.dropout, rngs=rngs)\n\n    def __call__(\n        self, x_DL: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x_DL = x_DL.astype(self.dtype)\n        codebook = self.codebook.value.astype(self.dtype)\n\n        x_normalized_DL = normalize(x_DL)\n        normalized_codebook_KL = normalize(codebook)\n        distance_DK = -jnp.matmul(x_normalized_DL, normalized_codebook_KL.T)\n        if training:\n            distance_DK = self.drop(distance_DK)\n\n        # --- Get indices and embeddings ---\n        indices_D = jnp.argmin(distance_DK, axis=-1)\n        z_DL = codebook[indices_D]\n\n        # --- Straight through estimator ---\n        z_q_DL = x_normalized_DL + jax.lax.stop_gradient(z_DL - x_normalized_DL)\n        return z_q_DL, z_DL, x_DL, indices_D\n\n    def get_codes(self, indices_E: jax.Array) -> jax.Array:\n        return self.codebook[indices_E]\n\n\ndef _create_flash_attention_fn(use_flash_attention: bool, is_causal: bool) -> Callable:\n    """"""\n    Create an attention function that uses flash attention if enabled.\n\n    flax.nnx.MultiHeadAttention provides tensors with shape (batch..., length, num_heads, head_dim),\n    but jax.nn.dot_product_attention expects (batch, length, num_heads, head_dim). We reshape to\n    ensure compatibility. cuDNN's flash attention additionally requires a sequence length that\n    is a multiple of 4. We pad the sequence length to the nearest multiple of 4 and mask\n    accordingly. Note that cuDNN requires the mask to be broadcast before calling the attention\n    function due to strict shape checking.\n    """"""\n\n    def attention_fn(\n        query_BTHD, key_BSHD, value_BSHD, bias=None, mask_B111=None, **kwargs\n    ):\n        implementation = ""cudnn"" if use_flash_attention else None\n\n        def _merge_batch_dims(x):\n            return einops.rearrange(x, ""... l h k -> (...) l h k"")\n\n        def _pad(x, pad_size):\n            return jnp.pad(x, ((0, 0), (0, pad_size), (0, 0), (0, 0)))\n\n        original_shape = query_BTHD.shape\n        T = query_BTHD.shape[-3]\n        S = key_BSHD.shape[-3]\n\n        # Pad to nearest multiple of 4\n        Q = ((T + 3) // 4) * 4\n        pad_size_Q = Q - T\n        K = ((S + 3) // 4) * 4\n        pad_size_K = K - S\n\n        query_BQHD = _pad(_merge_batch_dims(query_BTHD), pad_size_Q)\n        key_BKHD = _pad(_merge_batch_dims(key_BSHD), pad_size_K)\n        value_BKHD = _pad(_merge_batch_dims(value_BSHD), pad_size_K)\n\n        attention_mask = jnp.ones((Q, K), dtype=jnp.bool_)\n        attention_mask = attention_mask.at[T:, :].set(False)\n        attention_mask = attention_mask.at[:, S:].set(False)\n\n        mask_11TS = attention_mask[jnp.newaxis, jnp.newaxis, :, :]\n\n        bias_4d = (\n            jnp.pad(\n                _merge_batch_dims(bias),\n                ((0, 0), (0, 0), (0, pad_size_Q), (0, pad_size_K)),\n            )\n            if bias is not None\n            else None\n        )\n\n        # NOTE: jax.nn.dot_product_attention does not support dropout\n        output_4d = jax.nn.dot_product_attention(\n            query=query_BQHD,\n            key=key_BKHD,\n            value=value_BKHD,\n            bias=bias_4d,\n            mask=mask_11TS,\n            implementation=implementation,\n            is_causal=is_causal,\n        )\n        return output_4d[..., :T, :, :].reshape(original_shape)\n\n    return attention_fn\n",python,tab
+2,164,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"8:47:03 PM [info] Activating crowd-code\n8:47:03 PM [info] Recording started\n8:47:03 PM [info] Initializing git provider using file system watchers...\n8:47:03 PM [info] Git repository found\n8:47:03 PM [info] Git provider initialized successfully\n8:47:03 PM [info] Initial git state: [object Object]\n",Log,tab
+3,3295,"utils/nn.py",0,0,"",python,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-dedac322-1282-4d89-8a49-f3a5624493ea1762171752270-2025_11_03-13.09.19.936/source.csv

@@ -0,0 +1,255 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,366,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"1:09:19 PM [info] Activating crowd-code\n1:09:19 PM [info] Recording started\n1:09:19 PM [info] Initializing git provider using file system watchers...\n1:09:20 PM [error] Not a git repository: EntryNotFound (FileSystemError): Error: ENOENT: no such file or directory, stat '/home/franz.srambical/jafar/slurm/dev/franz/berlin/crowd-pilot/.git'\n",Log,tab
+3,4284,"TERMINAL",0,0,"",,terminal_command
+4,7859,"start_sglang",0,0,"",plaintext,tab
+5,11243,"TERMINAL",0,0,"",,terminal_command
+6,12735,"start_sglang_server.py",0,0,"",python,tab
+7,27073,"start_sglang_server.sh",0,0,"",shellscript,tab
+8,27890,"start_sglang_server.sh",0,0,"python3 -m sglang.launch_server --model-path qwen/qwen2.5-0.5b-instruct --host 0.0.0.0\n",shellscript,content
+9,28380,"start_sglang_server.sh",86,0,"\n",shellscript,content
+10,29134,"start_sglang_server.sh",87,1,"",shellscript,content
+11,30095,"start_sglang_server.sh",0,0,"",shellscript,selection_command
+12,30473,"start_sglang_server.sh",0,0,"\n",shellscript,content
+13,31314,"start_sglang_server.sh",0,0,"s",shellscript,content
+14,31314,"start_sglang_server.sh",1,0,"",shellscript,selection_keyboard
+15,31358,"start_sglang_server.sh",1,0,"o",shellscript,content
+16,31358,"start_sglang_server.sh",2,0,"",shellscript,selection_keyboard
+17,31448,"start_sglang_server.sh",2,0,"u",shellscript,content
+18,31449,"start_sglang_server.sh",3,0,"",shellscript,selection_keyboard
+19,31476,"start_sglang_server.sh",3,0,"r",shellscript,content
+20,31476,"start_sglang_server.sh",4,0,"",shellscript,selection_keyboard
+21,31919,"start_sglang_server.sh",4,0,"c",shellscript,content
+22,31919,"start_sglang_server.sh",5,0,"",shellscript,selection_keyboard
+23,32140,"start_sglang_server.sh",5,0,"e",shellscript,content
+24,32140,"start_sglang_server.sh",6,0,"",shellscript,selection_keyboard
+25,32304,"start_sglang_server.sh",6,0," ",shellscript,content
+26,32304,"start_sglang_server.sh",7,0,"",shellscript,selection_keyboard
+27,32546,"start_sglang_server.sh",7,0,".venv/bin/activate",shellscript,content
+28,32775,"start_sglang_server.sh",24,0,"",shellscript,selection_command
+29,33216,"start_sglang_server.sh",25,0,"\n",shellscript,content
+30,35383,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+31,37385,"TERMINAL",0,0,"",,terminal_focus
+32,37386,"start_sglang_server.sh",0,0,"",shellscript,tab
+33,43060,"TERMINAL",0,0,"source .venv/bin/activate",,terminal_command
+34,49399,"TERMINAL",0,0,"ls",,terminal_command
+35,52175,"TERMINAL",0,0,"cd",,terminal_command
+36,56006,"TERMINAL",0,0,"cd crowd-pilot/",,terminal_command
+37,57470,"TERMINAL",0,0,"source .venv/bin/activate",,terminal_command
+38,61752,"TERMINAL",0,0,"uv pip show sglang",,terminal_command
+39,61798,"TERMINAL",0,0,"]633;C",,terminal_output
+40,62576,"TERMINAL",0,0,"Name: sglang\r\nVersion: 0.5.4.post1\r\nLocation: /fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages\r\nRequires: aiohttp, anthropic, blobfile, build, compressed-tensors, cuda-python, datasets, decord2, einops, fastapi, flashinfer-python, gguf, grpcio, grpcio-health-checking, grpcio-reflection, grpcio-tools, hf-transfer, huggingface-hub, interegular, ipython, llguidance, modelscope, msgspec, ninja, numpy, nvidia-cutlass-dsl, nvidia-ml-py, openai, openai-harmony, orjson, outlines, packaging, partial-json-parser, pillow, prometheus-client, psutil, py-spy, pybase64, pydantic, python-multipart, pyzmq, requests, scipy, sentencepiece, setproctitle, sgl-kernel, soundfile, tiktoken, timm, torch, torch-memory-saver, torchao, torchaudio, torchvision, tqdm, transformers, uvicorn, uvloop, xgrammar\r\nRequired-by:\r\n]0;franz.srambical@hai-login1:~/crowd-pilot",,terminal_output
+41,71102,"TERMINAL",0,0,"deactivate",,terminal_command
+42,77402,"TERMINAL",0,0,"bash /home/franz.srambical/jafar/slurm/dev/franz/berlin/crowd-pilot/start_sglang_server.sh",,terminal_command
+43,77468,"TERMINAL",0,0,"]633;C",,terminal_output
+44,85588,"TERMINAL",0,0,"^CTraceback (most recent call last):\r\n  File ""<frozen runpy>"", line 198, in _run_module_as_main\r\n  File ""<frozen runpy>"", line 88, in _run_code\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/sglang/launch_server.py"", line 7, in <module>\r\n    from sglang.srt.server_args import prepare_server_args\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/sglang/srt/server_args.py"", line 29, in <module>\r\n    from sglang.srt.connector import ConnectorType\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/sglang/srt/connector/__init__.py"", line 6, in <module>\r\n    from sglang.srt.connector.base_connector import (\r\n    ...<3 lines>...\r\n    )\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/sglang/srt/connector/base_connector.py"", line 10, in <module>\r\n    import torch\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/torch/__init__.py"", line 2150, in <module>\r\n    from torch import _VF as _VF, functional as functional  # usort: skip\r\n    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/torch/functional.py"", line 8, in <module>\r\n    import torch.nn.functional as F\r\n  File ""<frozen importlib._bootstrap>"", line 1360, in _find_and_load\r\n  File ""<frozen importlib._bootstrap>"", line 1331, in _find_and_load_unlocked\r\n  File ""<frozen importlib._bootstrap>"", line 935, in _load_unlocked\r\n  File ""<frozen importlib._bootstrap_external>"", line 1022, in exec_module\r\n  File ""<frozen importlib._bootstrap_external>"", line 1118, in get_code\r\n  File ""<frozen importlib._bootstrap_external>"", line 1217, in get_data\r\nKeyboardInterrupt\r\n",,terminal_output
+45,85639,"TERMINAL",0,0,"^C\r\n]0;franz.srambical@hai-login1:~/crowd-pilot",,terminal_output
+46,85825,"TERMINAL",0,0,"^C",,terminal_command
+47,86834,"start_sglang_server.sh",27,0,"",shellscript,selection_command
+48,87198,"start_sglang_server.sh",27,86,"python3 -m sglang.launch_server --model-path qwen/qwen2.5-0.5b-instruct --host 0.0.0.0",shellscript,selection_command
+49,88964,"start_sglang_server.sh",27,0,"",shellscript,selection_command
+50,89979,"TERMINAL",0,0,"/home/franz.srambical/jafar/slurm/dev/franz/berlin/crowd-pilot/start_sglang_server.sh",,terminal_command
+51,93031,"start_sglang_server.sh",27,0,"/home/franz.srambical/jafar/slurm/dev/franz/berlin/crowd-pilot/start_sglang_server.sh",shellscript,content
+52,93032,"start_sglang_server.sh",112,0,"",shellscript,selection_keyboard
+53,93769,"start_sglang_server.sh",27,85,"",shellscript,content
+54,93771,"start_sglang_server.sh",112,0,"",shellscript,selection_command
+55,94569,"start_sglang_server.sh",114,0,"",shellscript,selection_command
+56,95993,"start_sglang_server.sh",27,0,"",shellscript,selection_command
+57,96402,"start_sglang_server.sh",27,86,"python3 -m sglang.launch_server --model-path qwen/qwen2.5-0.5b-instruct --host 0.0.0.0",shellscript,selection_command
+58,96799,"start_sglang_server.sh",27,0,"",shellscript,selection_command
+59,99144,"TERMINAL",0,0,"source .venv/bin/activate",,terminal_command
+60,102053,"TERMINAL",0,0,"python3 -m sglang.launch_server --model-path qwen/qwen2.5-0.5b-instruct --host 0.0.0.0",,terminal_command
+61,102098,"TERMINAL",0,0,"]633;C",,terminal_output
+62,118857,"TERMINAL",0,0,"",,terminal_command
+63,123331,"TERMINAL",0,0,"2025-11-03 13:11:23.179652: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.\r\n",,terminal_output
+64,126934,"TERMINAL",0,0,"2025-11-03 13:11:26.785110: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.\r\nTo enable the following instructions: AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX512_FP16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.\r\n",,terminal_output
+65,140724,"TERMINAL",0,0,"2025-11-03 13:11:40.573353: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.\r\n",,terminal_output
+66,151901,"TERMINAL",0,0,"Traceback (most recent call last):\r\n  File ""<frozen runpy>"", line 198, in _run_module_as_main\r\n  File ""<frozen runpy>"", line 88, in _run_code\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/sglang/launch_server.py"", line 11, in <module>\r\n    server_args = prepare_server_args(sys.argv[1:])\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/sglang/srt/server_args.py"", line 3850, in prepare_server_args\r\n    return ServerArgs.from_cli_args(raw_args)\r\n           ~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/sglang/srt/server_args.py"", line 3472, in from_cli_args\r\n    return cls(**{attr: getattr(args, attr) for attr in attrs})\r\n  File ""<string>"", line 268, in __init__\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/sglang/srt/server_args.py"", line 538, in __post_init__\r\n    self._handle_missing_default_values()\r\n    ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/sglang/srt/server_args.py"", line 623, in _handle_missing_default_values\r\n    self.device = get_device()\r\n                  ~~~~~~~~~~^^\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/sglang/srt/utils/common.py"", line 1781, in get_device\r\n    raise RuntimeError(""No accelerator (CUDA, XPU, HPU) is available."")\r\nRuntimeError: No accelerator (CUDA, XPU, HPU) is available.\r\n",,terminal_output
+67,155163,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/crowd-pilot",,terminal_output
+68,246640,"TERMINAL",0,0,"id",,terminal_command
+69,246640,"TERMINAL",0,0,"]633;Cuid=961800067(franz.srambical) gid=961800067(franz.srambical) groups=961800067(franz.srambical),961800017(helmholtz-member),961800019(helmholtz-all),961800033(hmgu),961900525(hfmi_synergyunit)\r\n]0;franz.srambical@hai-login1:~/crowd-pilot",,terminal_output
+70,278403,"TERMINAL",0,0,"squeue",,terminal_command
+71,278419,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  33317   xiao.liu interacti     1   128  R 2025-11-02T17:43:38 2025-11-02T17:43:38 19:30:20   23:59:00 hai006\r\n  33328 kalyan.nad  standard     1    64  R 2025-11-03T11:56:23 2025-11-03T11:56:38  1:17:20 1-00:00:00 hai002\r\n  33320 kalyan.nad  standard     1    64  R 2025-11-03T11:36:55 2025-11-03T11:36:55  1:37:03 1-00:00:00 hai001\r\n  33318   xiao.liu  standard     1   128  R 2025-11-02T19:29:40 2025-11-02T19:30:38 17:43:20   23:59:00 hai004\r\n]0;franz.srambical@hai-login1:~/crowd-pilot",,terminal_output
+72,281131,"TERMINAL",0,0,"salloc --gpus=1 --ntasks-per-node=1 --cpus-per-task=10 --mem=100G",,terminal_command
+73,281187,"TERMINAL",0,0,"]633;Csalloc: Granted job allocation 33333\r\n",,terminal_output
+74,281282,"TERMINAL",0,0,"salloc: Nodes hai003 are ready for job\r\n",,terminal_output
+75,281642,"TERMINAL",0,0,"Running inside SLURM, Job ID 33333.\r\n",,terminal_output
+76,281742,"TERMINAL",0,0,"]0;franz.srambical@hai-login1:~/crowd-pilot[?2004h[franz.srambical@hai003.haicore.berlin:~/crowd-pilot] $ ",,terminal_output
+77,283584,"TERMINAL",0,0,"l",,terminal_output
+78,283694,"TERMINAL",0,0,"s",,terminal_output
+79,283773,"TERMINAL",0,0,"\r\n[?2004l\rLICENSE  README.md  crowd-pilot  maxtext  pyproject.toml  slurm  uv.lock\r\n]0;franz.srambical@hai-login1:~/crowd-pilot[?2004h[franz.srambical@hai003.haicore.berlin:~/crowd-pilot] $ ",,terminal_output
+80,284758,"TERMINAL",0,0,"\r(reverse-i-search)`': ",,terminal_output
+81,284889,"TERMINAL",0,0,"s': lso': . ""/fast/home/franz.srambical/.cursor-server/bin/3ccce8f55d8cca49f6d28b491a844c699b8719a0/out/vs/workbench/contrib/terminal/common/scripts/shellIntegration-bash.sh""",,terminal_output
+82,284952,"TERMINAL",0,0,"\ru': source .venv/bin/activate\r\n\r",,terminal_output
+83,285009,"TERMINAL",0,0,"[1@r': sour",,terminal_output
+84,285651,"TERMINAL",0,0,"\r[30@[franz.srambical@hai003.haicore.berlin:~/crowd-pilot] $ sour\r\n[?2004l\r]0;franz.srambical@hai-login1:~/crowd-pilot[?2004h(crowd-pilot) [franz.srambical@hai003.haicore.berlin:~/crowd-pilot] $ ",,terminal_output
+85,287855,"start_sglang_server.sh",27,86,"python3 -m sglang.launch_server --model-path qwen/qwen2.5-0.5b-instruct --host 0.0.0.0",shellscript,selection_command
+86,288194,"start_sglang_server.sh",27,0,"",shellscript,selection_command
+87,289101,"TERMINAL",0,0,"python3 -m sglang.launch_server --model-path qwen/qwen2.5-0.5b-instruct --host 0.0.0.0",,terminal_output
+88,289327,"TERMINAL",0,0,"\rpython3 -m sglang.launch_server --model-path qwen/qwen2.5-0.5b-instruct --host 0.0.0.0\r\n[?2004l\r",,terminal_output
+89,306946,"TERMINAL",0,0,"2025-11-03 13:14:26.791489: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.\r\n",,terminal_output
+90,308085,"TERMINAL",0,0,"2025-11-03 13:14:27.923401: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.\r\nTo enable the following instructions: AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX512_FP16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.\r\n",,terminal_output
+91,312755,"TERMINAL",0,0,"2025-11-03 13:14:32.604499: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.\r\n",,terminal_output
+92,326515,"TERMINAL",0,0,"[2025-11-03 13:14:46] WARNING server_args.py:1104: Attention backend not explicitly specified. Use fa3 backend by default.\r\n[2025-11-03 13:14:46] INFO trace.py:48: opentelemetry package is not installed, tracing disabled\r\n",,terminal_output
+93,328112,"TERMINAL",0,0,"[2025-11-03 13:14:47] server_args=ServerArgs(model_path='qwen/qwen2.5-0.5b-instruct', tokenizer_path='qwen/qwen2.5-0.5b-instruct', tokenizer_mode='auto', tokenizer_worker_num=1, skip_tokenizer_init=False, load_format='auto', model_loader_extra_config='{}', trust_remote_code=False, context_length=None, is_embedding=False, enable_multimodal=None, revision=None, model_impl='auto', host='0.0.0.0', port=30000, grpc_mode=False, skip_server_warmup=False, warmups=None, nccl_port=None, checkpoint_engine_wait_weights_before_ready=False, dtype='auto', quantization=None, quantization_param_path=None, kv_cache_dtype='auto', enable_fp32_lm_head=False, modelopt_quant=None, modelopt_checkpoint_restore_path=None, modelopt_checkpoint_save_path=None, modelopt_export_path=None, quantize_and_serve=False, mem_fraction_static=0.835, max_running_requests=None, max_queued_requests=None, max_total_tokens=None, chunked_prefill_size=8192, max_prefill_tokens=16384, schedule_policy='fcfs', enable_priority_scheduling=False, abort_on_priority_when_disabled=False, schedule_low_priority_values_first=False, priority_scheduling_preemption_threshold=10, schedule_conservativeness=1.0, page_size=1, hybrid_kvcache_ratio=None, swa_full_tokens_ratio=0.8, disable_hybrid_swa_memory=False, radix_eviction_policy='lru', device='cuda', tp_size=1, pp_size=1, pp_max_micro_batch_size=None, stream_interval=1, stream_output=False, random_seed=541394942, constrained_json_whitespace_pattern=None, constrained_json_disable_any_whitespace=False, watchdog_timeout=300, dist_timeout=None, download_dir=None, base_gpu_id=0, gpu_id_step=1, sleep_on_idle=False, log_level='info', log_level_http=None, log_requests=False, log_requests_level=2, crash_dump_folder=None, show_time_cost=False, enable_metrics=False, enable_metrics_for_all_schedulers=False, tokenizer_metrics_custom_labels_header='x-custom-labels', tokenizer_metrics_allowed_custom_labels=None, bucket_time_to_first_token=None, bucket_inter_token_latency=None, bucket_e2e_request_latency=None, collect_tokens_histogram=False, prompt_tokens_buckets=None, generation_tokens_buckets=None, gc_warning_threshold_secs=0.0, decode_log_interval=40, enable_request_time_stats_logging=False, kv_events_config=None, enable_trace=False, oltp_traces_endpoint='localhost:4317', api_key=None, served_model_name='qwen/qwen2.5-0.5b-instruct', weight_version='default', chat_template=None, completion_template=None, file_storage_path='sglang_storage', enable_cache_report=False, reasoning_parser=None, tool_call_parser=None, tool_server=None, sampling_defaults='model', dp_size=1, load_balance_method='round_robin', load_watch_interval=0.1, prefill_round_robin_balance=False, dist_init_addr=None, nnodes=1, node_rank=0, json_model_override_args='{}', preferred_sampling_params=None, enable_lora=None, max_lora_rank=None, lora_target_modules=None, lora_paths=None, max_loaded_loras=None, max_loras_per_batch=8, lora_eviction_policy='lru', lora_backend='triton', max_lora_chunk_size=16, attention_backend='fa3', decode_attention_backend=None, prefill_attention_backend=None, sampling_backend='flashinfer', grammar_backend='xgrammar', mm_attention_backend=None, nsa_prefill_backend='flashmla_sparse', nsa_decode_backend='fa3', speculative_algorithm=None, speculative_draft_model_path=None, speculative_draft_model_revision=None, speculative_draft_load_format=None, speculative_num_steps=None, speculative_eagle_topk=None, speculative_num_draft_tokens=None, speculative_accept_threshold_single=1.0, speculative_accept_threshold_acc=1.0, speculative_token_map=None, speculative_attention_mode='prefill', speculative_ngram_min_match_window_size=1, speculative_ngram_max_match_window_size=12, speculative_ngram_min_bfs_breadth=1, speculative_ngram_max_bfs_breadth=10, speculative_ngram_match_type='BFS', speculative_ngram_branch_length=18, speculative_ngram_capacity=10000000, ep_size=1, moe_a2a_backend='none', moe_runner_backend='auto', flashinfer_mxfp4_moe_precision='default', enable_flashinfer_allreduce_fusion=False, deepep_mode='auto', ep_num_redundant_experts=0, ep_dispatch_algorithm='static', init_expert_location='trivial', enable_eplb=False, eplb_algorithm='auto', eplb_rebalance_num_iterations=1000, eplb_rebalance_layers_per_chunk=None, eplb_min_rebalancing_utilization_threshold=1.0, expert_distribution_recorder_mode=None, expert_distribution_recorder_buffer_size=1000, enable_expert_distribution_metrics=False, deepep_config=None, moe_dense_tp_size=None, elastic_ep_backend=None, mooncake_ib_device=None, max_mamba_cache_size=None, mamba_ssm_dtype='float32', mamba_full_memory_ratio=0.9, enable_hierarchical_cache=False, hicache_ratio=2.0, hicache_size=0, hicache_write_policy='write_through', hicache_io_backend='kernel', hicache_mem_layout='layer_first', hicache_storage_backend=None, hicache_storage_prefetch_policy='best_effort', hicache_storage_backend_extra_config=None, enable_lmcache=False, kt_amx_weight_path=None, kt_amx_method='AMXINT4', kt_cpuinfer=None, kt_threadpool_count=2, kt_num_gpu_experts=None, enable_double_sparsity=False, ds_channel_config_path=None, ds_heavy_channel_num=32, ds_heavy_token_num=256, ds_heavy_channel_type='qk', ds_sparse_decode_threshold=4096, cpu_offload_gb=0, offload_group_size=-1, offload_num_in_group=1, offload_prefetch_step=1, offload_mode='cpu', multi_item_scoring_delimiter=None, disable_radix_cache=False, cuda_graph_max_bs=256, cuda_graph_bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256], disable_cuda_graph=False, disable_cuda_graph_padding=False, enable_profile_cuda_graph=False, enable_cudagraph_gc=False, enable_nccl_nvls=False, enable_symm_mem=False, disable_flashinfer_cutlass_moe_fp4_allgather=False, enable_tokenizer_batch_encode=False, disable_tokenizer_batch_decode=False, disable_outlines_disk_cache=False, disable_custom_all_reduce=False, enable_mscclpp=False, enable_torch_symm_mem=False, disable_overlap_schedule=False, enable_mixed_chunk=False, enable_dp_attention=False, enable_dp_lm_head=False, enable_two_batch_overlap=False, enable_single_batch_overlap=False, tbo_token_distribution_threshold=0.48, enable_torch_compile=False, enable_piecewise_cuda_graph=False, torch_compile_max_bs=32, piecewise_cuda_graph_max_tokens=4096, piecewise_cuda_graph_tokens=[4, 8, 12, 16, 20, 24, 28, 32, 48, 64, 80, 96, 112, 128, 144, 160, 176, 192, 208, 224, 240, 256, 288, 320, 352, 384, 416, 448, 480, 512, 640, 768, 896, 1024, 1152, 1280, 1408, 1536, 1664, 1792, 1920, 2048, 2176, 2304, 2432, 2560, 2688, 2816, 2944, 3072, 3200, 3328, 3456, 3584, 3712, 3840, 3968, 4096], piecewise_cuda_graph_compiler='eager', torchao_config='', enable_nan_detection=False, enable_p2p_check=False, triton_attention_reduce_in_fp32=False, triton_attention_num_kv_splits=8, triton_attention_split_tile_size=None, num_continuous_decode_steps=1, delete_ckpt_after_loading=False, enable_memory_saver=False, enable_weights_cpu_backup=False, allow_auto_truncate=False, enable_custom_logit_processor=False, flashinfer_mla_disable_ragged=False, disable_shared_experts_fusion=False, disable_chunked_prefix_cache=False, disable_fast_image_processor=False, keep_mm_feature_on_device=False, enable_return_hidden_states=False, scheduler_recv_interval=1, numa_node=None, enable_deterministic_inference=False, rl_on_policy_target=None, enable_dynamic_batch_tokenizer=False, dynamic_batch_tokenizer_batch_size=32, dynamic_batch_tokenizer_batch_timeout=0.002, debug_tensor_dump_output_folder=None, debug_tensor_dump_input_file=None, debug_tensor_dump_inject=False, disaggregation_mode='null', disaggregation_transfer_backend='mooncake', disaggregation_bootstrap_port=8998, disaggregation_decode_tp=None, disaggregation_decode_dp=None, disaggregation_prefill_pp=1, disaggregation_ib_device=None, disaggregation_decode_enable_offload_kvcache=False, num_reserved_decode_tokens=512, disaggregation_decode_polling_interval=1, custom_weight_loader=[], weight_loader_disable_mmap=False, remote_instance_weight_loader_seed_instance_ip=None, remote_instance_weight_loader_seed_instance_service_port=None, remote_instance_weight_loader_send_weights_group_ports=None, enable_pdmux=False, pdmux_config_path=None, sm_group_num=8)\r\n",,terminal_output
+94,329705,"TERMINAL",0,0,"[2025-11-03 13:14:49] Using default HuggingFace chat template with detected content format: string\r\n",,terminal_output
+95,346598,"TERMINAL",0,0,"[2025-11-03 13:15:06] INFO trace.py:48: opentelemetry package is not installed, tracing disabled\r\n",,terminal_output
+96,348586,"TERMINAL",0,0,"[2025-11-03 13:15:08] Init torch distributed begin.\r\n",,terminal_output
+97,348951,"TERMINAL",0,0,"[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0\r\n",,terminal_output
+98,349007,"TERMINAL",0,0,"[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0\r\n[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0\r\n[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0\r\n[2025-11-03 13:15:08] Init torch distributed ends. mem usage=0.00 GB\r\n",,terminal_output
+99,349159,"TERMINAL",0,0,"[2025-11-03 13:15:09] MOE_RUNNER_BACKEND is not initialized, the backend will be automatically selected\r\n",,terminal_output
+100,349970,"TERMINAL",0,0,"[2025-11-03 13:15:09] INFO trace.py:48: opentelemetry package is not installed, tracing disabled\r\n",,terminal_output
+101,354706,"TERMINAL",0,0,"[2025-11-03 13:15:14] Load weight begin. avail mem=78.68 GB\r\n",,terminal_output
+102,355343,"TERMINAL",0,0,"[2025-11-03 13:15:15] TensorFlow version 2.20.0 available.\r\n",,terminal_output
+103,360584,"TERMINAL",0,0,"[2025-11-03 13:15:20] Using model weights format ['*.safetensors']\r\n",,terminal_output
+104,361111,"TERMINAL",0,0,"[2025-11-03 13:15:20] No model.safetensors.index.json found in remote.\r\n\rLoading safetensors checkpoint shards:   0% Completed | 0/1 [00:00<?, ?it/s]\r\n",,terminal_output
+105,362166,"TERMINAL",0,0,"\rLoading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  1.03it/s]\r\n\rLoading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  1.03it/s]\r\n\r\n[2025-11-03 13:15:21] Load weight end. type=Qwen2ForCausalLM, dtype=torch.bfloat16, avail mem=77.61 GB, mem usage=1.07 GB.\r\n[2025-11-03 13:15:21] Using KV cache dtype: torch.bfloat16\r\n[2025-11-03 13:15:22] KV Cache is allocated. #tokens: 5647121, K size: 32.31 GB, V size: 32.31 GB\r\n[2025-11-03 13:15:22] Memory pool end. avail mem=12.31 GB\r\n",,terminal_output
+106,362703,"TERMINAL",0,0,"[2025-11-03 13:15:22] Capture cuda graph begin. This can take up to several minutes. avail mem=12.21 GB\r\n[2025-11-03 13:15:22] Capture cuda graph bs [1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256]\r\n",,terminal_output
+107,363169,"TERMINAL",0,0,"\r  0%|                                                                                                                                            | 0/36 [00:00<?, ?it/s]\rCapturing batches (bs=256 avail_mem=12.00 GB):   0%|                                                                                             | 0/36 [00:00<?, ?it/s]",,terminal_output
+108,364442,"TERMINAL",0,0,"\rCapturing batches (bs=256 avail_mem=12.00 GB):   3%|██▎                                                                                  | 1/36 [00:00<00:32,  1.09it/s]\rCapturing batches (bs=248 avail_mem=11.84 GB):   3%|██▎                                                                                  | 1/36 [00:00<00:32,  1.09it/s]\rCapturing batches (bs=240 avail_mem=11.83 GB):   3%|██▎                                                                                  | 1/36 [00:00<00:32,  1.09it/s]\rCapturing batches (bs=232 avail_mem=11.83 GB):   3%|██▎                                                                                  | 1/36 [00:00<00:32,  1.09it/s]\rCapturing batches (bs=232 avail_mem=11.83 GB):  11%|█████████▍                                                                           | 4/36 [00:01<00:06,  4.88it/s]\rCapturing batches (bs=224 avail_mem=11.82 GB):  11%|█████████▍                                                                           | 4/36 [00:01<00:06,  4.88it/s]\rCapturing batches (bs=216 avail_mem=11.81 GB):  11%|█████████▍                                                                           | 4/36 [00:01<00:06,  4.88it/s]\rCapturing batches (bs=208 avail_mem=11.81 GB):  11%|█████████▍                                                                           | 4/36 [00:01<00:06,  4.88it/s]\rCapturing batches (bs=208 avail_mem=11.81 GB):  19%|████████████████▌                                                                    | 7/36 [00:01<00:03,  8.61it/s]\rCapturing batches (bs=200 avail_mem=11.81 GB):  19%|████████████████▌                                                                    | 7/36 [00:01<00:03,  8.61it/s]\rCapturing batches (bs=192 avail_mem=11.80 GB):  19%|████████████████▌                                                                    | 7/36 [00:01<00:03,  8.61it/s]\rCapturing batches (bs=184 avail_mem=11.80 GB):  19%|████████████████▌                                                                    | 7/36 [00:01<00:03,  8.61it/s]\rCapturing batches (bs=184 avail_mem=11.80 GB):  28%|███████████████████████▎                                                            | 10/36 [00:01<00:02, 12.28it/s]\rCapturing batches (bs=176 avail_mem=11.79 GB):  28%|███████████████████████▎                                                            | 10/36 [00:01<00:02, 12.28it/s]\rCapturing batches (bs=168 avail_mem=11.79 GB):  28%|███████████████████████▎                                                            | 10/36 [00:01<00:02, 12.28it/s]",,terminal_output
+109,365219,"TERMINAL",0,0,"\rCapturing batches (bs=160 avail_mem=11.78 GB):  28%|███████████████████████▎                                                            | 10/36 [00:01<00:02, 12.28it/s]\rCapturing batches (bs=160 avail_mem=11.78 GB):  36%|██████████████████████████████▎                                                     | 13/36 [00:01<00:01, 14.47it/s]\rCapturing batches (bs=152 avail_mem=11.78 GB):  36%|██████████████████████████████▎                                                     | 13/36 [00:01<00:01, 14.47it/s]\rCapturing batches (bs=144 avail_mem=11.77 GB):  36%|██████████████████████████████▎                                                     | 13/36 [00:01<00:01, 14.47it/s]\rCapturing batches (bs=136 avail_mem=11.77 GB):  36%|██████████████████████████████▎                                                     | 13/36 [00:01<00:01, 14.47it/s]\rCapturing batches (bs=136 avail_mem=11.77 GB):  44%|█████████████████████████████████████▎                                              | 16/36 [00:01<00:01, 17.34it/s]\rCapturing batches (bs=128 avail_mem=11.76 GB):  44%|█████████████████████████████████████▎                                              | 16/36 [00:01<00:01, 17.34it/s]\rCapturing batches (bs=120 avail_mem=11.76 GB):  44%|█████████████████████████████████████▎                                              | 16/36 [00:01<00:01, 17.34it/s]\rCapturing batches (bs=112 avail_mem=11.75 GB):  44%|█████████████████████████████████████▎                                              | 16/36 [00:01<00:01, 17.34it/s]\rCapturing batches (bs=112 avail_mem=11.75 GB):  53%|████████████████████████████████████████████▎                                       | 19/36 [00:01<00:00, 19.06it/s]\rCapturing batches (bs=104 avail_mem=11.75 GB):  53%|████████████████████████████████████████████▎                                       | 19/36 [00:01<00:00, 19.06it/s]\rCapturing batches (bs=96 avail_mem=11.75 GB):  53%|████████████████████████████████████████████▊                                        | 19/36 [00:01<00:00, 19.06it/s]\rCapturing batches (bs=88 avail_mem=11.74 GB):  53%|████████████████████████████████████████████▊                                        | 19/36 [00:01<00:00, 19.06it/s]\rCapturing batches (bs=88 avail_mem=11.74 GB):  61%|███████████████████████████████████████████████████▉                                 | 22/36 [00:01<00:00, 20.36it/s]\rCapturing batches (bs=80 avail_mem=11.73 GB):  61%|███████████████████████████████████████████████████▉                                 | 22/36 [00:01<00:00, 20.36it/s]\rCapturing batches (bs=72 avail_mem=11.73 GB):  61%|███████████████████████████████████████████████████▉                                 | 22/36 [00:01<00:00, 20.36it/s]\rCapturing batches (bs=64 avail_mem=11.72 GB):  61%|███████████████████████████████████████████████████▉                                 | 22/36 [00:01<00:00, 20.36it/s]\rCapturing batches (bs=64 avail_mem=11.72 GB):  69%|███████████████████████████████████████████████████████████                          | 25/36 [00:01<00:00, 21.65it/s]\rCapturing batches (bs=56 avail_mem=11.72 GB):  69%|███████████████████████████████████████████████████████████                          | 25/36 [00:01<00:00, 21.65it/s]\rCapturing batches (bs=48 avail_mem=11.72 GB):  69%|████████████████████████████████████���██████████████████████                          | 25/36 [00:01<00:00, 21.65it/s]\rCapturing batches (bs=40 avail_mem=11.71 GB):  69%|███████████████████████████████████████████████████████████                          | 25/36 [00:01<00:00, 21.65it/s]\rCapturing batches (bs=40 avail_mem=11.71 GB):  78%|██████████████████████████████████████████████████████████████████                   | 28/36 [00:02<00:00, 22.85it/s]\rCapturing batches (bs=32 avail_mem=11.71 GB):  78%|██████████████████████████████████████████████████████████████████                   | 28/36 [00:02<00:00, 22.85it/s]\rCapturing batches (bs=24 avail_mem=11.70 GB):  78%|██████████████████████████████████████████████████████████████████                   | 28/36 [00:02<00:00, 22.85it/s]\rCapturing batches (bs=16 avail_mem=11.70 GB):  78%|██████████████████████████████████████████████████████████████████                   | 28/36 [00:02<00:00, 22.85it/s]",,terminal_output
+110,365457,"TERMINAL",0,0,"\rCapturing batches (bs=16 avail_mem=11.70 GB):  86%|█████████████████████████████████████████████████████████████████████████▏           | 31/36 [00:02<00:00, 21.01it/s]\rCapturing batches (bs=12 avail_mem=11.69 GB):  86%|█████████████████████████████████████████████████████████████████████████▏           | 31/36 [00:02<00:00, 21.01it/s]\rCapturing batches (bs=8 avail_mem=11.69 GB):  86%|██████████████████████████████████████████████████████████████████████████            | 31/36 [00:02<00:00, 21.01it/s]\rCapturing batches (bs=4 avail_mem=11.68 GB):  86%|██████████████████████████████████████████████████████████████████████████            | 31/36 [00:02<00:00, 21.01it/s]\rCapturing batches (bs=2 avail_mem=11.68 GB):  86%|██████████████████████████████████████████████████████████████████████████            | 31/36 [00:02<00:00, 21.01it/s]\rCapturing batches (bs=2 avail_mem=11.68 GB):  97%|███████████████████████████████████████████████████████████████████████████████████▌  | 35/36 [00:02<00:00, 24.38it/s]\rCapturing batches (bs=1 avail_mem=11.67 GB):  97%|███████████████████████████████████████████████████████████████████████████████████▌  | 35/36 [00:02<00:00, 24.38it/s]\rCapturing batches (bs=1 avail_mem=11.67 GB): 100%|██████████████████████████████████████████████████████████████████████████████████████| 36/36 [00:02<00:00, 15.53it/s]\r\n",,terminal_output
+111,365803,"TERMINAL",0,0,"[2025-11-03 13:15:25] Capture cuda graph end. Time elapsed: 3.10 s. mem usage=0.54 GB. avail mem=11.67 GB.\r\n",,terminal_output
+112,366523,"TERMINAL",0,0,"[2025-11-03 13:15:26] max_total_num_tokens=5647121, chunked_prefill_size=8192, max_prefill_tokens=16384, max_running_requests=4096, context_len=32768, available_gpu_mem=11.67 GB\r\n",,terminal_output
+113,367067,"TERMINAL",0,0,"[2025-11-03 13:15:26] INFO:     Started server process [1848778]\r\n[2025-11-03 13:15:26] INFO:     Waiting for application startup.\r\n[2025-11-03 13:15:26] Using default chat sampling params from model generation config: {'repetition_penalty': 1.1, 'temperature': 0.7, 'top_k': 20, 'top_p': 0.8}\r\n",,terminal_output
+114,367150,"TERMINAL",0,0,"[2025-11-03 13:15:27] Using default chat sampling params from model generation config: {'repetition_penalty': 1.1, 'temperature': 0.7, 'top_k': 20, 'top_p': 0.8}\r\n[2025-11-03 13:15:27] INFO:     Application startup complete.\r\n[2025-11-03 13:15:27] INFO:     Uvicorn running on http://0.0.0.0:30000 (Press CTRL+C to quit)\r\n",,terminal_output
+115,368184,"TERMINAL",0,0,"[2025-11-03 13:15:28] INFO:     127.0.0.1:57018 - ""GET /get_model_info HTTP/1.1"" 200 OK\r\n[2025-11-03 13:15:28] Prefill batch, #new-seq: 1, #new-token: 6, #cached-token: 0, token usage: 0.00, #running-req: 0, #queue-req: 0, \r\n",,terminal_output
+116,370619,"TERMINAL",0,0,"[2025-11-03 13:15:30] INFO:     127.0.0.1:57028 - ""POST /generate HTTP/1.1"" 200 OK\r\n[2025-11-03 13:15:30] The server is fired up and ready to roll!\r\n",,terminal_output
+117,913264,"TERMINAL",0,0,"",,terminal_focus
+118,934101,"TERMINAL",0,0,"cd",,terminal_command
+119,936634,"TERMINAL",0,0,"cd crowd-pilot/",,terminal_command
+120,938611,"TERMINAL",0,0,"source .venv/bin/activate",,terminal_command
+121,940201,"TERMINAL",0,0,"python",,terminal_command
+122,940251,"TERMINAL",0,0,"]633;C",,terminal_output
+123,940308,"TERMINAL",0,0,"Python 3.13.5 (main, Jul  1 2025, 18:37:36) [Clang 20.1.4 ] on linux\r\nType ""help"", ""copyright"", ""credits"" or ""license"" for more information.\r\n",,terminal_output
+124,940480,"TERMINAL",0,0,"[?2004h[?1h=[?25l\n>>> [?12l[?25h",,terminal_output
+125,953679,"TERMINAL",0,0,"[?25l\n>>> import requests... [?12l[?25h[?25l\n... [?12l[?25h[?25l... url = f""http://localhost:{port}/v1/chat/completions""[?12l[?25h",,terminal_output
+126,954821,"TERMINAL",0,0,"",,terminal_output
+127,954978,"TERMINAL",0,0,"",,terminal_output
+128,955111,"TERMINAL",0,0,"",,terminal_output
+129,955280,"TERMINAL",0,0,"",,terminal_output
+130,956069,"TERMINAL",0,0,"",,terminal_output
+131,956484,"TERMINAL",0,0,"[?25l... url = f""http://localhost:{port/v1/chat/completions""[?12l[?25h",,terminal_output
+132,957029,"TERMINAL",0,0,"[?25l... url = f""http://localhost:{por/v1/chat/completions""[?12l[?25h[?25l... url = f""http://localhost:{po/v1/chat/completions""[?12l[?25h[?25l... url = f""http://localhost:{p/v1/chat/completions""[?12l[?25h[?25l... url = f""http://localhost:{/v1/chat/completions""[?12l[?25h[?25l... url = f""http://localhost:/v1/chat/completions""[?12l[?25h[?25l... url = f""http://localhost/v1/chat/completions""[?12l[?25h[?25l... url = f""http://localhos/v1/chat/completions""[?12l[?25h[?25l... url = f""http://localho/v1/chat/completions""[?12l[?25h[?25l... url = f""http://localh/v1/chat/completions""[?12l[?25h[?25l... url = f""http://local/v1/chat/completions""[?12l[?25h",,terminal_output
+133,957146,"TERMINAL",0,0,"[?25l... url = f""http://loca/v1/chat/completions""[?12l[?25h",,terminal_output
+134,957298,"TERMINAL",0,0,"[?25l... url = f""http://loc/v1/chat/completions""[?12l[?25h",,terminal_output
+135,957432,"TERMINAL",0,0,"[?25l... url = f""http://lo/v1/chat/completions""[?12l[?25h",,terminal_output
+136,957569,"TERMINAL",0,0,"[?25l... url = f""http://l/v1/chat/completions""[?12l[?25h",,terminal_output
+137,958025,"TERMINAL",0,0,"[?25l... url = f""http:///v1/chat/completions""[?12l[?25h",,terminal_output
+138,959244,"TERMINAL",0,0,"",,terminal_focus
+139,960619,"TERMINAL",0,0,"squeue",,terminal_command
+140,960623,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  33333 franz.sram interacti     1    20  R 2025-11-03T13:14:01 2025-11-03T13:14:01    11:19 1-00:00:00 hai003\r\n  33317   xiao.liu interacti     1   128  R 2025-11-02T17:43:38 2025-11-02T17:43:38 19:41:42   23:59:00 hai006\r\n  33328 kalyan.nad  standard     1    64  R 2025-11-03T11:56:23 2025-11-03T11:56:38  1:28:42 1-00:00:00 hai002\r\n  33318   xiao.liu  standard     1   128  R 2025-11-02T19:29:40 2025-11-02T19:30:38 17:54:42   23:59:00 hai004\r\n]0;franz.srambical@hai-login1:~/jafar/slurm/dev/franz/berlin/crowd-pilot",,terminal_output
+141,965159,"TERMINAL",0,0,"python",,terminal_focus
+142,965738,"TERMINAL",0,0,"srun",,terminal_focus
+143,966961,"TERMINAL",0,0,"python",,terminal_focus
+144,968671,"TERMINAL",0,0,"[?25l... url = f""http://h/v1/chat/completions""[?12l[?25h",,terminal_output
+145,968869,"TERMINAL",0,0,"[?25l... url = f""http://ha/v1/chat/completions""[?12l[?25h[?25l... url = f""http://hai/v1/chat/completions""[?12l[?25h",,terminal_output
+146,969460,"TERMINAL",0,0,"[?25l... url = f""http://hai0/v1/chat/completions""[?12l[?25h",,terminal_output
+147,969519,"TERMINAL",0,0,"[?25l... url = f""http://hai00/v1/chat/completions""[?12l[?25h",,terminal_output
+148,969617,"TERMINAL",0,0,"[?25l... url = f""http://hai003/v1/chat/completions""[?12l[?25h",,terminal_output
+149,970391,"TERMINAL",0,0,"[?25l... url = f""http://hai003:/v1/chat/completions""[?12l[?25h",,terminal_output
+150,973152,"TERMINAL",0,0,"srun",,terminal_focus
+151,975372,"TERMINAL",0,0,"python",,terminal_focus
+152,977464,"TERMINAL",0,0,"[?25l... url = f""http://hai003:3/v1/chat/completions""[?12l[?25h",,terminal_output
+153,977548,"TERMINAL",0,0,"[?25l... url = f""http://hai003:30/v1/chat/completions""[?12l[?25h",,terminal_output
+154,978092,"TERMINAL",0,0,"[?25l... url = f""http://hai003:300/v1/chat/completions""[?12l[?25h",,terminal_output
+155,978256,"TERMINAL",0,0,"[?25l... url = f""http://hai003:3000/v1/chat/completions""[?12l[?25h",,terminal_output
+156,978414,"TERMINAL",0,0,"[?25l... url = f""http://hai003:30000/v1/chat/completions""[?12l[?25h",,terminal_output
+157,978848,"TERMINAL",0,0,"",,terminal_output
+158,979078,"TERMINAL",0,0,"[?25l\n... [?12l[?25h",,terminal_output
+159,981853,"TERMINAL",0,0,"\n\r[?2004l[?1l>",,terminal_output
+160,982027,"TERMINAL",0,0,"[?2004h[?1h=[?25l\n>>> [?12l[?25h",,terminal_output
+161,986222,"TERMINAL",0,0,"[?25l\n>>> data = {... [?12l[?25h[?25l\n...     ""model"": ""qwen/qwen2.5-0.5b-instruct"",... [?12l[?25h[?25l\n...     ""messages"": [{""role"": ""user"", ""content"": ""What is the capital of France?""}],... [?12l[?25h[?25l... }[?12l[?25h",,terminal_output
+162,986446,"TERMINAL",0,0,"\n\r[?2004l[?1l>[?2004h[?1h=[?25l\n>>> [?12l[?25h",,terminal_output
+163,986621,"TERMINAL",0,0,"\n\r[?2004l[?1l>[?2004h[?1h=[?25l\n>>> [?12l[?25h",,terminal_output
+164,990025,"TERMINAL",0,0,"[?25l>>> response = requests.post(url, json=data)[?12l[?25h",,terminal_output
+165,990225,"TERMINAL",0,0,"\n\r[?2004l[?1l>",,terminal_output
+166,990235,"TERMINAL",0,0,"[2025-11-03 13:25:50] Prefill batch, #new-seq: 1, #new-token: 36, #cached-token: 0, token usage: 0.00, #running-req: 0, #queue-req: 0, \r\n",,terminal_output
+167,990372,"TERMINAL",0,0,"\r\n[?2004h[?1h=[?25l\n>>> [?12l[?25h\n\r[?2004l[?1l>[?2004h[?1h=[?25l\n>>> [?12l[?25h",,terminal_output
+168,990374,"TERMINAL",0,0,"[2025-11-03 13:25:50] INFO:     10.86.2.251:49732 - ""POST /v1/chat/completions HTTP/1.1"" 200 OK\r\n",,terminal_output
+169,992718,"TERMINAL",0,0,"[?25l>>> p[?12l[?25h",,terminal_output
+170,992795,"TERMINAL",0,0,"[?25l>>> pr[?12l[?25h",,terminal_output
+171,992936,"TERMINAL",0,0,"[?25l>>> pri[?12l[?25h[?25l>>> prin[?12l[?25h",,terminal_output
+172,993015,"TERMINAL",0,0,"[?25l>>> print[?12l[?25h",,terminal_output
+173,993292,"TERMINAL",0,0,"[?25l>>> print([?12l[?25h",,terminal_output
+174,993402,"TERMINAL",0,0,"[?25l>>> print()[?12l[?25h",,terminal_output
+175,993786,"TERMINAL",0,0,"",,terminal_output
+176,994396,"TERMINAL",0,0,"[?25l>>> print(r)[?12l[?25h",,terminal_output
+177,994490,"TERMINAL",0,0,"[?25l>>> print(re)[?12l[?25h",,terminal_output
+178,994656,"TERMINAL",0,0,"[?25l>>> print(res)[?12l[?25h",,terminal_output
+179,994799,"TERMINAL",0,0,"[?25l>>> print(response)[?12l[?25h",,terminal_output
+180,996827,"TERMINAL",0,0,"[?25l>>> print(response.)[?12l[?25h",,terminal_output
+181,996908,"TERMINAL",0,0,"[?25l>>> print(response.s)[?12l[?25h",,terminal_output
+182,997008,"TERMINAL",0,0,"[?25l>>> print(response.sj)[?12l[?25h",,terminal_output
+183,997447,"TERMINAL",0,0,"[?25l>>> print(response.s)[?12l[?25h",,terminal_output
+184,997591,"TERMINAL",0,0,"[?25l>>> print(response.)[?12l[?25h",,terminal_output
+185,997715,"TERMINAL",0,0,"[?25l>>> print(response.j)[?12l[?25h",,terminal_output
+186,997801,"TERMINAL",0,0,"[?25l>>> print(response.js)[?12l[?25h",,terminal_output
+187,997933,"TERMINAL",0,0,"[?25l>>> print(response.json()[?12l[?25h",,terminal_output
+188,998729,"TERMINAL",0,0,"",,terminal_output
+189,998898,"TERMINAL",0,0,"[?25l\n... [?12l[?25h",,terminal_output
+190,1000184,"TERMINAL",0,0,"[?25l... _[?12l[?25h",,terminal_output
+191,1000759,"TERMINAL",0,0,"[?25l... [?12l[?25h",,terminal_output
+192,1001046,"TERMINAL",0,0,"[?25l... )[?12l[?25h",,terminal_output
+193,1001598,"TERMINAL",0,0,"\n\r[?2004l[?1l>{'id': 'b76c04917df44d9aaf99082ae9769ded', 'object': 'chat.completion', 'created': 1762172750, 'model': 'qwen/qwen2.5-0.5b-instruct', 'choices': [{'index': 0, 'message': {'role': 'assistant', 'content': 'The capital of France is Paris.', 'reasoning_content': None, 'tool_calls': None}, 'logprobs': None, 'finish_reason': 'stop', 'matched_stop': 151645}], 'usage': {'prompt_tokens': 36, 'total_tokens': 44, 'completion_tokens': 8, 'prompt_tokens_details': None, 'reasoning_tokens': 0}, 'metadata': {'weight_version': 'default'}}\r\n[?2004h[?1h=[?25l\n>>> [?12l[?25h",,terminal_output
+194,1024494,"TERMINAL",0,0,"srun",,terminal_focus
+195,1055248,"TERMINAL",0,0,"python",,terminal_focus
+196,1241705,"TERMINAL",0,0,"[2025-11-03 13:30:01] SIGTERM received. signum=None frame=None. Draining requests and shutting down...\r\n",,terminal_output
+197,1243235,"TERMINAL",0,0,"[2025-11-03 13:30:03] Gracefully exiting... Remaining number of requests 0. Remaining requests remaining_rids=[].\r\n",,terminal_output
+198,1243545,"TERMINAL",0,0,"Killed\r\n]0;franz.srambical@hai-login1:~/crowd-pilot[?2004h(crowd-pilot) [franz.srambical@hai003.haicore.berlin:~/crowd-pilot] $ ",,terminal_output
+199,1381345,"TERMINAL",0,0,"[?25l\n>>> print(response.json()... )[?12l[?25h",,terminal_output
+200,1381698,"TERMINAL",0,0,"[?25l>>> response = requests.post(url, json=data)[?12l[?25h",,terminal_output
+201,1382817,"TERMINAL",0,0,"[?25l\n\n\n>>> data = {...     ""model"": ""qwen/qwen2.5-0.5b-instruct"",...     ""messages"": [{""role"": ""user"", ""content"": ""What is the capital of France?""}],... }[?12l[?25h",,terminal_output
+202,1383740,"TERMINAL",0,0,"[?25l>>> import requests... ... url = f""http://hai003:30000/v1/chat/completions""[?12l[?25h",,terminal_output
+203,1385966,"TERMINAL",0,0,"[?25l\n... [?12l[?25h",,terminal_output
+204,1386716,"TERMINAL",0,0,"\n\r[?2004l[?1l>[?2004h[?1h=[?25l\n>>> [?12l[?25h",,terminal_output
+205,1387796,"TERMINAL",0,0,"[?25l\n\n>>> import requests... ... url = f""http://hai003:30000/v1/chat/completions""[?12l[?25h",,terminal_output
+206,1387869,"TERMINAL",0,0,"[?25l>>> print(response.json()... )[?12l[?25h",,terminal_output
+207,1388047,"TERMINAL",0,0,"[?25l>>> response = requests.post(url, json=data)[?12l[?25h",,terminal_output
+208,1388536,"TERMINAL",0,0,"[?25l\n\n\n>>> data = {...     ""model"": ""qwen/qwen2.5-0.5b-instruct"",...     ""messages"": [{""role"": ""user"", ""content"": ""What is the capital of France?""}],... }[?12l[?25h",,terminal_output
+209,1390082,"TERMINAL",0,0,"[?25l>>> import requests... ... url = f""http://hai003:30000/v1/chat/completions""[?12l[?25h",,terminal_output
+210,1390618,"TERMINAL",0,0,"[?25l\n>>> data = {...     ""model"": ""qwen/qwen2.5-0.5b-instruct"",...     ""messages"": [{""role"": ""user"", ""content"": ""What is the capital of France?""}],... }[?12l[?25h",,terminal_output
+211,1391047,"TERMINAL",0,0,"\n\r[?2004l[?1l>[?2004h[?1h=[?25l\n>>> [?12l[?25h",,terminal_output
+212,1391308,"TERMINAL",0,0,"\n\r[?2004l[?1l>[?2004h[?1h=[?25l\n>>> [?12l[?25h",,terminal_output
+213,1391930,"TERMINAL",0,0,"[?25l\n\n\n>>> data = {...     ""model"": ""qwen/qwen2.5-0.5b-instruct"",...     ""messages"": [{""role"": ""user"", ""content"": ""What is the capital of France?""}],... }[?12l[?25h",,terminal_output
+214,1392014,"TERMINAL",0,0,"[?25l>>> import requests... ... url = f""http://hai003:30000/v1/chat/completions""[?12l[?25h",,terminal_output
+215,1392471,"TERMINAL",0,0,"[?25l>>> print(response.json()... )[?12l[?25h",,terminal_output
+216,1392947,"TERMINAL",0,0,"[?25l>>> response = requests.post(url, json=data)[?12l[?25h",,terminal_output
+217,1393731,"TERMINAL",0,0,"\n\r[?2004l[?1l>Traceback (most recent call last):\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/urllib3/connection.py"", line 198, in _new_conn\r\n    sock = connection.create_connection(\r\n        (self._dns_host, self.port),\r\n    ...<2 lines>...\r\n        socket_options=self.socket_options,\r\n    )\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/urllib3/util/connection.py"", line 85, in create_connection\r\n    raise err\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/urllib3/util/connection.py"", line 73, in create_connection\r\n    sock.connect(sa)\r\n    ~~~~~~~~~~~~^^^^\r\nConnectionRefusedError: [Errno 111] Connection refused\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nTraceback (most recent call last):\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/urllib3/connectionpool.py"", line 787, in urlopen\r\n    response = self._make_request(\r\n        conn,\r\n    ...<10 lines>...\r\n        **response_kw,\r\n    )\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/urllib3/connectionpool.py"", line 493, in _make_request\r\n    conn.request(\r\n    ~~~~~~~~~~~~^\r\n        method,\r\n        ^^^^^^^\r\n    ...<6 lines>...\r\n        enforce_content_length=enforce_content_length,\r\n        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n    )\r\n    ^\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/urllib3/connection.py"", line 494, in request\r\n    self.endheaders()\r\n    ~~~~~~~~~~~~~~~^^\r\n  File ""/home/franz.srambical/.local/share/uv/python/cpython-3.13.5-linux-x86_64-gnu/lib/python3.13/http/client.py"", line 1333, in endheaders\r\n    self._send_output(message_body, encode_chunked=encode_chunked)\r\n    ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n  File ""/home/franz.srambical/.local/share/uv/python/cpython-3.13.5-linux-x86_64-gnu/lib/python3.13/http/client.py"", line 1093, in _send_output\r\n    self.send(msg)\r\n    ~~~~~~~~~^^^^^\r\n  File ""/home/franz.srambical/.local/share/uv/python/cpython-3.13.5-linux-x86_64-gnu/lib/python3.13/http/client.py"", line 1037, in send\r\n    self.connect()\r\n    ~~~~~~~~~~~~^^\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/urllib3/connection.py"", line 325, in connect\r\n    self.sock = self._new_conn()\r\n                ~~~~~~~~~~~~~~^^\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/urllib3/connection.py"", line 213, in _new_conn\r\n    raise NewConnectionError(\r\n        self, f""Failed to establish a new connection: {e}""\r\n    ) from e\r\nurllib3.exceptions.NewConnectionError: <urllib3.connection.HTTPConnection object at 0x7f4217d02fd0>: Failed to establish a new connection: [Errno 111] Connection refused\r\n\r\nThe above exception was the direct cause of the following exception:\r\n\r\nTraceback (most recent call last):\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/requests/adapters.py"", line 644, in send\r\n    resp = conn.urlopen(\r\n        method=request.method,\r\n    ...<9 lines>...\r\n        chunked=chunked,\r\n    )\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/urllib3/connectionpool.py"", line 841, in urlopen\r\n    retries = retries.increment(\r\n        method, url, error=new_e, _pool=self, _stacktrace=sys.exc_info()[2]\r\n    )\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/urllib3/util/retry.py"", line 519, in increment\r\n    raise MaxRetryError(_pool, url, reason) from reason  # type: ignore[arg-type]\r\n    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\nurllib3.exceptions.MaxRetryError: HTTPConnectionPool(host='hai003', port=30000): Max retries exceeded with url: /v1/chat/completions (Caused by NewConnectionError('<urllib3.connection.HTTPConnection object at 0x7f4217d02fd0>: Failed to establish a new connection: [Errno 111] Connection refused'))\r\n\r\nDuring handling of the above exception, another exception occurred:\r\n\r\nTraceback (most recent call last):\r\n  File ""<python-input-9>"", line 1, in <module>\r\n    response = requests.post(url, json=data)\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/requests/api.py"", line 115, in post\r\n    return request(""post"", url, data=data, json=json, **kwargs)\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/requests/api.py"", line 59, in request\r\n    return session.request(method=method, url=url, **kwargs)\r\n           ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/requests/sessions.py"", line 589, in request\r\n    resp = self.send(prep, **send_kwargs)\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/requests/sessions.py"", line 703, in send\r\n    r = adapter.send(request, **kwargs)\r\n  File ""/fast/home/franz.srambical/crowd-pilot/.venv/lib/python3.13/site-packages/requests/adapters.py"", line 677, in send\r\n    raise ConnectionError(e, request=request)\r\nrequests.exceptions.ConnectionError: HTTPConnectionPool(host='hai003', port=30000): Max retries exceeded with url: /v1/chat/completions (Caused by NewConnectionError('<urllib3.connection.HTTPConnection object at 0x7f4217d02fd0>: Failed to establish a new connection: [Errno 111] Connection refused'))\r\n[?2004h[?1h=[?25l\n>>> [?12l[?25h",,terminal_output
+218,1398331,"TERMINAL",0,0,"srun",,terminal_focus
+219,1403264,"TERMINAL",0,0,"python3 -m sglang.launch_server --model-path qwen/qwen2.5-0.5b-instruct --host 0.0.0.0",,terminal_output
+220,1403498,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+221,1413006,"TERMINAL",0,0,"2025-11-03 13:32:52.861020: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.\r\n",,terminal_output
+222,1413064,"TERMINAL",0,0,"2025-11-03 13:32:52.916612: I tensorflow/core/platform/cpu_feature_guard.cc:210] This TensorFlow binary is optimized to use available CPU instructions in performance-critical operations.\r\nTo enable the following instructions: AVX2 AVX512F AVX512_VNNI AVX512_BF16 AVX512_FP16 AVX_VNNI AMX_TILE AMX_INT8 AMX_BF16 FMA, in other operations, rebuild TensorFlow with the appropriate compiler flags.\r\n",,terminal_output
+223,1415760,"TERMINAL",0,0,"2025-11-03 13:32:55.587656: I tensorflow/core/util/port.cc:153] oneDNN custom operations are on. You may see slightly different numerical results due to floating-point round-off errors from different computation orders. To turn them off, set the environment variable `TF_ENABLE_ONEDNN_OPTS=0`.\r\n",,terminal_output
+224,1422741,"TERMINAL",0,0,"[2025-11-03 13:33:02] WARNING server_args.py:1104: Attention backend not explicitly specified. Use fa3 backend by default.\r\n[2025-11-03 13:33:02] INFO trace.py:48: opentelemetry package is not installed, tracing disabled\r\n",,terminal_output
+225,1423202,"TERMINAL",0,0,"[2025-11-03 13:33:03] server_args=ServerArgs(model_path='qwen/qwen2.5-0.5b-instruct', tokenizer_path='qwen/qwen2.5-0.5b-instruct', tokenizer_mode='auto', tokenizer_worker_num=1, skip_tokenizer_init=False, load_format='auto', model_loader_extra_config='{}', trust_remote_code=False, context_length=None, is_embedding=False, enable_multimodal=None, revision=None, model_impl='auto', host='0.0.0.0', port=30000, grpc_mode=False, skip_server_warmup=False, warmups=None, nccl_port=None, checkpoint_engine_wait_weights_before_ready=False, dtype='auto', quantization=None, quantization_param_path=None, kv_cache_dtype='auto', enable_fp32_lm_head=False, modelopt_quant=None, modelopt_checkpoint_restore_path=None, modelopt_checkpoint_save_path=None, modelopt_export_path=None, quantize_and_serve=False, mem_fraction_static=0.835, max_running_requests=None, max_queued_requests=None, max_total_tokens=None, chunked_prefill_size=8192, max_prefill_tokens=16384, schedule_policy='fcfs', enable_priority_scheduling=False, abort_on_priority_when_disabled=False, schedule_low_priority_values_first=False, priority_scheduling_preemption_threshold=10, schedule_conservativeness=1.0, page_size=1, hybrid_kvcache_ratio=None, swa_full_tokens_ratio=0.8, disable_hybrid_swa_memory=False, radix_eviction_policy='lru', device='cuda', tp_size=1, pp_size=1, pp_max_micro_batch_size=None, stream_interval=1, stream_output=False, random_seed=417737316, constrained_json_whitespace_pattern=None, constrained_json_disable_any_whitespace=False, watchdog_timeout=300, dist_timeout=None, download_dir=None, base_gpu_id=0, gpu_id_step=1, sleep_on_idle=False, log_level='info', log_level_http=None, log_requests=False, log_requests_level=2, crash_dump_folder=None, show_time_cost=False, enable_metrics=False, enable_metrics_for_all_schedulers=False, tokenizer_metrics_custom_labels_header='x-custom-labels', tokenizer_metrics_allowed_custom_labels=None, bucket_time_to_first_token=None, bucket_inter_token_latency=None, bucket_e2e_request_latency=None, collect_tokens_histogram=False, prompt_tokens_buckets=None, generation_tokens_buckets=None, gc_warning_threshold_secs=0.0, decode_log_interval=40, enable_request_time_stats_logging=False, kv_events_config=None, enable_trace=False, oltp_traces_endpoint='localhost:4317', api_key=None, served_model_name='qwen/qwen2.5-0.5b-instruct', weight_version='default', chat_template=None, completion_template=None, file_storage_path='sglang_storage', enable_cache_report=False, reasoning_parser=None, tool_call_parser=None, tool_server=None, sampling_defaults='model', dp_size=1, load_balance_method='round_robin', load_watch_interval=0.1, prefill_round_robin_balance=False, dist_init_addr=None, nnodes=1, node_rank=0, json_model_override_args='{}', preferred_sampling_params=None, enable_lora=None, max_lora_rank=None, lora_target_modules=None, lora_paths=None, max_loaded_loras=None, max_loras_per_batch=8, lora_eviction_policy='lru', lora_backend='triton', max_lora_chunk_size=16, attention_backend='fa3', decode_attention_backend=None, prefill_attention_backend=None, sampling_backend='flashinfer', grammar_backend='xgrammar', mm_attention_backend=None, nsa_prefill_backend='flashmla_sparse', nsa_decode_backend='fa3', speculative_algorithm=None, speculative_draft_model_path=None, speculative_draft_model_revision=None, speculative_draft_load_format=None, speculative_num_steps=None, speculative_eagle_topk=None, speculative_num_draft_tokens=None, speculative_accept_threshold_single=1.0, speculative_accept_threshold_acc=1.0, speculative_token_map=None, speculative_attention_mode='prefill', speculative_ngram_min_match_window_size=1, speculative_ngram_max_match_window_size=12, speculative_ngram_min_bfs_breadth=1, speculative_ngram_max_bfs_breadth=10, speculative_ngram_match_type='BFS', speculative_ngram_branch_length=18, speculative_ngram_capacity=10000000, ep_size=1, moe_a2a_backend='none', moe_runner_backend='auto', flashinfer_mxfp4_moe_precision='default', enable_flashinfer_allreduce_fusion=False, deepep_mode='auto', ep_num_redundant_experts=0, ep_dispatch_algorithm='static', init_expert_location='trivial', enable_eplb=False, eplb_algorithm='auto', eplb_rebalance_num_iterations=1000, eplb_rebalance_layers_per_chunk=None, eplb_min_rebalancing_utilization_threshold=1.0, expert_distribution_recorder_mode=None, expert_distribution_recorder_buffer_size=1000, enable_expert_distribution_metrics=False, deepep_config=None, moe_dense_tp_size=None, elastic_ep_backend=None, mooncake_ib_device=None, max_mamba_cache_size=None, mamba_ssm_dtype='float32', mamba_full_memory_ratio=0.9, enable_hierarchical_cache=False, hicache_ratio=2.0, hicache_size=0, hicache_write_policy='write_through', hicache_io_backend='kernel', hicache_mem_layout='layer_first', hicache_storage_backend=None, hicache_storage_prefetch_policy='best_effort', hicache_storage_backend_extra_config=None, enable_lmcache=False, kt_amx_weight_path=None, kt_amx_method='AMXINT4', kt_cpuinfer=None, kt_threadpool_count=2, kt_num_gpu_experts=None, enable_double_sparsity=False, ds_channel_config_path=None, ds_heavy_channel_num=32, ds_heavy_token_num=256, ds_heavy_channel_type='qk', ds_sparse_decode_threshold=4096, cpu_offload_gb=0, offload_group_size=-1, offload_num_in_group=1, offload_prefetch_step=1, offload_mode='cpu', multi_item_scoring_delimiter=None, disable_radix_cache=False, cuda_graph_max_bs=256, cuda_graph_bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256], disable_cuda_graph=False, disable_cuda_graph_padding=False, enable_profile_cuda_graph=False, enable_cudagraph_gc=False, enable_nccl_nvls=False, enable_symm_mem=False, disable_flashinfer_cutlass_moe_fp4_allgather=False, enable_tokenizer_batch_encode=False, disable_tokenizer_batch_decode=False, disable_outlines_disk_cache=False, disable_custom_all_reduce=False, enable_mscclpp=False, enable_torch_symm_mem=False, disable_overlap_schedule=False, enable_mixed_chunk=False, enable_dp_attention=False, enable_dp_lm_head=False, enable_two_batch_overlap=False, enable_single_batch_overlap=False, tbo_token_distribution_threshold=0.48, enable_torch_compile=False, enable_piecewise_cuda_graph=False, torch_compile_max_bs=32, piecewise_cuda_graph_max_tokens=4096, piecewise_cuda_graph_tokens=[4, 8, 12, 16, 20, 24, 28, 32, 48, 64, 80, 96, 112, 128, 144, 160, 176, 192, 208, 224, 240, 256, 288, 320, 352, 384, 416, 448, 480, 512, 640, 768, 896, 1024, 1152, 1280, 1408, 1536, 1664, 1792, 1920, 2048, 2176, 2304, 2432, 2560, 2688, 2816, 2944, 3072, 3200, 3328, 3456, 3584, 3712, 3840, 3968, 4096], piecewise_cuda_graph_compiler='eager', torchao_config='', enable_nan_detection=False, enable_p2p_check=False, triton_attention_reduce_in_fp32=False, triton_attention_num_kv_splits=8, triton_attention_split_tile_size=None, num_continuous_decode_steps=1, delete_ckpt_after_loading=False, enable_memory_saver=False, enable_weights_cpu_backup=False, allow_auto_truncate=False, enable_custom_logit_processor=False, flashinfer_mla_disable_ragged=False, disable_shared_experts_fusion=False, disable_chunked_prefix_cache=False, disable_fast_image_processor=False, keep_mm_feature_on_device=False, enable_return_hidden_states=False, scheduler_recv_interval=1, numa_node=None, enable_deterministic_inference=False, rl_on_policy_target=None, enable_dynamic_batch_tokenizer=False, dynamic_batch_tokenizer_batch_size=32, dynamic_batch_tokenizer_batch_timeout=0.002, debug_tensor_dump_output_folder=None, debug_tensor_dump_input_file=None, debug_tensor_dump_inject=False, disaggregation_mode='null', disaggregation_transfer_backend='mooncake', disaggregation_bootstrap_port=8998, disaggregation_decode_tp=None, disaggregation_decode_dp=None, disaggregation_prefill_pp=1, disaggregation_ib_device=None, disaggregation_decode_enable_offload_kvcache=False, num_reserved_decode_tokens=512, disaggregation_decode_polling_interval=1, custom_weight_loader=[], weight_loader_disable_mmap=False, remote_instance_weight_loader_seed_instance_ip=None, remote_instance_weight_loader_seed_instance_service_port=None, remote_instance_weight_loader_send_weights_group_ports=None, enable_pdmux=False, pdmux_config_path=None, sm_group_num=8)\r\n",,terminal_output
+226,1424166,"TERMINAL",0,0,"[2025-11-03 13:33:04] Using default HuggingFace chat template with detected content format: string\r\n",,terminal_output
+227,1440312,"TERMINAL",0,0,"[2025-11-03 13:33:20] INFO trace.py:48: opentelemetry package is not installed, tracing disabled\r\n",,terminal_output
+228,1445360,"TERMINAL",0,0,"[2025-11-03 13:33:25] INFO trace.py:48: opentelemetry package is not installed, tracing disabled\r\n",,terminal_output
+229,1447450,"TERMINAL",0,0,"[2025-11-03 13:33:27] Init torch distributed begin.\r\n",,terminal_output
+230,1447715,"TERMINAL",0,0,"[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0\r\n[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0\r\n[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0\r\n[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0\r\n[2025-11-03 13:33:27] Init torch distributed ends. mem usage=0.00 GB\r\n",,terminal_output
+231,1447771,"TERMINAL",0,0,"[2025-11-03 13:33:27] MOE_RUNNER_BACKEND is not initialized, the backend will be automatically selected\r\n",,terminal_output
+232,1448948,"TERMINAL",0,0,"[2025-11-03 13:33:28] Load weight begin. avail mem=78.68 GB\r\n",,terminal_output
+233,1449042,"TERMINAL",0,0,"[2025-11-03 13:33:28] TensorFlow version 2.20.0 available.\r\n",,terminal_output
+234,1450092,"TERMINAL",0,0,"[2025-11-03 13:33:29] Using model weights format ['*.safetensors']\r\n",,terminal_output
+235,1450641,"TERMINAL",0,0,"[2025-11-03 13:33:30] No model.safetensors.index.json found in remote.\r\n\rLoading safetensors checkpoint shards:   0% Completed | 0/1 [00:00<?, ?it/s]\r\n",,terminal_output
+236,1450848,"TERMINAL",0,0,"\rLoading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  6.08it/s]\r\n\rLoading safetensors checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  6.07it/s]\r\n\r\n[2025-11-03 13:33:30] Load weight end. type=Qwen2ForCausalLM, dtype=torch.bfloat16, avail mem=77.61 GB, mem usage=1.07 GB.\r\n[2025-11-03 13:33:30] Using KV cache dtype: torch.bfloat16\r\n[2025-11-03 13:33:30] KV Cache is allocated. #tokens: 5647121, K size: 32.31 GB, V size: 32.31 GB\r\n[2025-11-03 13:33:30] Memory pool end. avail mem=12.31 GB\r\n",,terminal_output
+237,1450942,"TERMINAL",0,0,"[2025-11-03 13:33:30] Capture cuda graph begin. This can take up to several minutes. avail mem=12.21 GB\r\n[2025-11-03 13:33:30] Capture cuda graph bs [1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64, 72, 80, 88, 96, 104, 112, 120, 128, 136, 144, 152, 160, 168, 176, 184, 192, 200, 208, 216, 224, 232, 240, 248, 256]\r\n",,terminal_output
+238,1451331,"TERMINAL",0,0,"\r  0%|                                                                                                                        | 0/36 [00:00<?, ?it/s]\rCapturing batches (bs=256 avail_mem=12.00 GB):   0%|                                                                         | 0/36 [00:00<?, ?it/s]",,terminal_output
+239,1451596,"TERMINAL",0,0,"\rCapturing batches (bs=256 avail_mem=12.00 GB):   3%|█▊                                                               | 1/36 [00:00<00:08,  4.22it/s]\rCapturing batches (bs=248 avail_mem=11.84 GB):   3%|█▊                                                               | 1/36 [00:00<00:08,  4.22it/s]\rCapturing batches (bs=240 avail_mem=11.83 GB):   3%|█▊                                                               | 1/36 [00:00<00:08,  4.22it/s]",,terminal_output
+240,1452423,"TERMINAL",0,0,"\rCapturing batches (bs=232 avail_mem=11.83 GB):   3%|█▊                                                               | 1/36 [00:00<00:08,  4.22it/s]\rCapturing batches (bs=232 avail_mem=11.83 GB):  11%|███████▏                                                         | 4/36 [00:00<00:02, 13.09it/s]\rCapturing batches (bs=224 avail_mem=11.82 GB):  11%|███████▏                                                         | 4/36 [00:00<00:02, 13.09it/s]\rCapturing batches (bs=216 avail_mem=11.81 GB):  11%|███████▏                                                         | 4/36 [00:00<00:02, 13.09it/s]\rCapturing batches (bs=208 avail_mem=11.81 GB):  11%|███████▏                                                         | 4/36 [00:00<00:02, 13.09it/s]\rCapturing batches (bs=208 avail_mem=11.81 GB):  19%|████████████▋                                                    | 7/36 [00:00<00:01, 17.78it/s]\rCapturing batches (bs=200 avail_mem=11.81 GB):  19%|████████████▋                                                    | 7/36 [00:00<00:01, 17.78it/s]\rCapturing batches (bs=192 avail_mem=11.80 GB):  19%|████████████▋                                                    | 7/36 [00:00<00:01, 17.78it/s]\rCapturing batches (bs=184 avail_mem=11.80 GB):  19%|████████████▋                                                    | 7/36 [00:00<00:01, 17.78it/s]\rCapturing batches (bs=184 avail_mem=11.80 GB):  28%|█████████████████▊                                              | 10/36 [00:00<00:01, 21.11it/s]\rCapturing batches (bs=176 avail_mem=11.79 GB):  28%|█████████████████▊                                              | 10/36 [00:00<00:01, 21.11it/s]\rCapturing batches (bs=168 avail_mem=11.79 GB):  28%|█████████████████▊                                              | 10/36 [00:00<00:01, 21.11it/s]\rCapturing batches (bs=160 avail_mem=11.78 GB):  28%|█████████████████▊                                              | 10/36 [00:00<00:01, 21.11it/s]\rCapturing batches (bs=160 avail_mem=11.78 GB):  36%|███████████████████████                                         | 13/36 [00:00<00:01, 22.38it/s]\rCapturing batches (bs=152 avail_mem=11.78 GB):  36%|███████████████████████                                         | 13/36 [00:00<00:01, 22.38it/s]\rCapturing batches (bs=144 avail_mem=11.77 GB):  36%|███████████████████████                                         | 13/36 [00:00<00:01, 22.38it/s]\rCapturing batches (bs=136 avail_mem=11.77 GB):  36%|███████████████████████                                         | 13/36 [00:00<00:01, 22.38it/s]\rCapturing batches (bs=136 avail_mem=11.77 GB):  44%|████████████████████████████▍                                   | 16/36 [00:00<00:00, 23.98it/s]\rCapturing batches (bs=128 avail_mem=11.76 GB):  44%|████████████████████████████▍                                   | 16/36 [00:00<00:00, 23.98it/s]\rCapturing batches (bs=120 avail_mem=11.76 GB):  44%|████████████████████████████▍                                   | 16/36 [00:00<00:00, 23.98it/s]\rCapturing batches (bs=112 avail_mem=11.75 GB):  44%|████████████████████████████▍                                   | 16/36 [00:00<00:00, 23.98it/s]\rCapturing batches (bs=112 avail_mem=11.75 GB):  53%|█████████████████████████████████▊                              | 19/36 [00:00<00:00, 24.01it/s]\rCapturing batches (bs=104 avail_mem=11.75 GB):  53%|█████████████████████████████████▊                              | 19/36 [00:00<00:00, 24.01it/s]\rCapturing batches (bs=96 avail_mem=11.75 GB):  53%|██████████████████████████████████▎                              | 19/36 [00:00<00:00, 24.01it/s]\rCapturing batches (bs=88 avail_mem=11.74 GB):  53%|██████████████████████████████████▎                              | 19/36 [00:01<00:00, 24.01it/s]\rCapturing batches (bs=88 avail_mem=11.74 GB):  61%|█████████��█████████████████████████████▋                         | 22/36 [00:01<00:00, 24.03it/s]\rCapturing batches (bs=80 avail_mem=11.73 GB):  61%|███████████████████████████████████████▋                         | 22/36 [00:01<00:00, 24.03it/s]\rCapturing batches (bs=72 avail_mem=11.73 GB):  61%|███████████████████████████████████████▋                         | 22/36 [00:01<00:00, 24.03it/s]",,terminal_output
+241,1452687,"TERMINAL",0,0,"\rCapturing batches (bs=64 avail_mem=11.72 GB):  61%|███████████████████████████████████████▋                         | 22/36 [00:01<00:00, 24.03it/s]\rCapturing batches (bs=64 avail_mem=11.72 GB):  69%|█████████████████████████████████████████████▏                   | 25/36 [00:01<00:00, 24.45it/s]\rCapturing batches (bs=56 avail_mem=11.72 GB):  69%|█████████████████████████████████████████████▏                   | 25/36 [00:01<00:00, 24.45it/s]\rCapturing batches (bs=48 avail_mem=11.72 GB):  69%|█████████████████████████████████████████████▏                   | 25/36 [00:01<00:00, 24.45it/s]\rCapturing batches (bs=40 avail_mem=11.71 GB):  69%|█████████████████████████████████████████████▏                   | 25/36 [00:01<00:00, 24.45it/s]\rCapturing batches (bs=40 avail_mem=11.71 GB):  78%|██████████████████████████████████████████████████▌              | 28/36 [00:01<00:00, 25.01it/s]\rCapturing batches (bs=32 avail_mem=11.71 GB):  78%|██████████████████████████████████████████████████▌              | 28/36 [00:01<00:00, 25.01it/s]\rCapturing batches (bs=24 avail_mem=11.70 GB):  78%|██████████████████████████████████████████████████▌              | 28/36 [00:01<00:00, 25.01it/s]\rCapturing batches (bs=16 avail_mem=11.70 GB):  78%|██████████████████████████████████████████████████▌              | 28/36 [00:01<00:00, 25.01it/s]",,terminal_output
+242,1452903,"TERMINAL",0,0,"\rCapturing batches (bs=16 avail_mem=11.70 GB):  86%|███████████████████████████████████████████████████████▉         | 31/36 [00:01<00:00, 23.88it/s]\rCapturing batches (bs=12 avail_mem=11.69 GB):  86%|███████████████████████████████████████████████████████▉         | 31/36 [00:01<00:00, 23.88it/s]\rCapturing batches (bs=8 avail_mem=11.69 GB):  86%|████████████████████████████████████████████████████████▊         | 31/36 [00:01<00:00, 23.88it/s]\rCapturing batches (bs=4 avail_mem=11.68 GB):  86%|████████████████████████████████████████████████████████▊         | 31/36 [00:01<00:00, 23.88it/s]\rCapturing batches (bs=2 avail_mem=11.68 GB):  86%|████████████████████████████████████████████████████████▊         | 31/36 [00:01<00:00, 23.88it/s]\rCapturing batches (bs=2 avail_mem=11.68 GB):  97%|████████████████████████████████████████████████████████████████▏ | 35/36 [00:01<00:00, 26.88it/s]\rCapturing batches (bs=1 avail_mem=11.67 GB):  97%|████████████████████████████████████████████████████████████████▏ | 35/36 [00:01<00:00, 26.88it/s]\rCapturing batches (bs=1 avail_mem=11.67 GB): 100%|██████████████████████████████████████████████████████████████████| 36/36 [00:01<00:00, 22.95it/s]\r\n",,terminal_output
+243,1453253,"TERMINAL",0,0,"[2025-11-03 13:33:33] Capture cuda graph end. Time elapsed: 2.32 s. mem usage=0.54 GB. avail mem=11.67 GB.\r\n",,terminal_output
+244,1454157,"TERMINAL",0,0,"[2025-11-03 13:33:34] max_total_num_tokens=5647121, chunked_prefill_size=8192, max_prefill_tokens=16384, max_running_requests=4096, context_len=32768, available_gpu_mem=11.67 GB\r\n",,terminal_output
+245,1454646,"TERMINAL",0,0,"[2025-11-03 13:33:34] INFO:     Started server process [1856857]\r\n[2025-11-03 13:33:34] INFO:     Waiting for application startup.\r\n[2025-11-03 13:33:34] Using default chat sampling params from model generation config: {'repetition_penalty': 1.1, 'temperature': 0.7, 'top_k': 20, 'top_p': 0.8}\r\n[2025-11-03 13:33:34] Using default chat sampling params from model generation config: {'repetition_penalty': 1.1, 'temperature': 0.7, 'top_k': 20, 'top_p': 0.8}\r\n[2025-11-03 13:33:34] INFO:     Application startup complete.\r\n[2025-11-03 13:33:34] INFO:     Uvicorn running on http://0.0.0.0:30000 (Press CTRL+C to quit)\r\n",,terminal_output
+246,1455651,"TERMINAL",0,0,"[2025-11-03 13:33:35] INFO:     127.0.0.1:51886 - ""GET /get_model_info HTTP/1.1"" 200 OK\r\n[2025-11-03 13:33:35] Prefill batch, #new-seq: 1, #new-token: 6, #cached-token: 0, token usage: 0.00, #running-req: 0, #queue-req: 0, \r\n",,terminal_output
+247,1456318,"TERMINAL",0,0,"[2025-11-03 13:33:36] INFO:     127.0.0.1:51898 - ""POST /generate HTTP/1.1"" 200 OK\r\n[2025-11-03 13:33:36] The server is fired up and ready to roll!\r\n",,terminal_output
+248,1459037,"TERMINAL",0,0,"[2025-11-03 13:33:38] Prefill batch, #new-seq: 1, #new-token: 36, #cached-token: 0, token usage: 0.00, #running-req: 0, #queue-req: 0, \r\n[2025-11-03 13:33:38] INFO:     10.86.2.252:38622 - ""POST /v1/chat/completions HTTP/1.1"" 200 OK\r\n",,terminal_output
+249,4563086,"TERMINAL",0,0,"[2025-11-03 14:25:22] INFO:     10.86.2.251:48316 - ""POST /v1/chat/completions HTTP/1.1"" 200 OK\r\n[2025-11-03 14:25:22] Prefill batch, #new-seq: 1, #new-token: 8192, #cached-token: 5, token usage: 0.00, #running-req: 0, #queue-req: 0, \r\n[2025-11-03 14:25:22] Prefill batch, #new-seq: 1, #new-token: 1418, #cached-token: 0, token usage: 0.00, #running-req: 0, #queue-req: 0, \r\n",,terminal_output
+250,4563164,"TERMINAL",0,0,"[2025-11-03 14:25:22] Decode batch, #running-req: 1, #token: 9640, token usage: 0.00, cuda graph: True, gen throughput (token/s): 0.01, #queue-req: 0, \r\n",,terminal_output
+251,4576851,"TERMINAL",0,0,"[2025-11-03 14:25:36] INFO:     10.86.2.251:53276 - ""POST /v1/chat/completions HTTP/1.1"" 200 OK\r\n[2025-11-03 14:25:36] Prefill batch, #new-seq: 1, #new-token: 164, #cached-token: 9654, token usage: 0.00, #running-req: 0, #queue-req: 0, \r\n",,terminal_output
+252,4576930,"TERMINAL",0,0,"[2025-11-03 14:25:36] Decode batch, #running-req: 1, #token: 9843, token usage: 0.00, cuda graph: True, gen throughput (token/s): 2.91, #queue-req: 0, \r\n",,terminal_output
+253,4603575,"TERMINAL",0,0,"[2025-11-03 14:26:03] INFO:     10.86.2.251:56392 - ""POST /v1/chat/completions HTTP/1.1"" 200 OK\r\n[2025-11-03 14:26:03] Prefill batch, #new-seq: 1, #new-token: 164, #cached-token: 9850, token usage: 0.00, #running-req: 0, #queue-req: 0, \r\n",,terminal_output
+254,4603624,"TERMINAL",0,0,"[2025-11-03 14:26:03] Decode batch, #running-req: 1, #token: 10046, token usage: 0.00, cuda graph: True, gen throughput (token/s): 1.50, #queue-req: 0, \r\n",,terminal_output
+255,4619746,"TERMINAL",0,0,"[2025-11-03 14:26:19] INFO:     10.86.2.251:33794 - ""POST /v1/chat/completions HTTP/1.1"" 200 OK\r\n[2025-11-03 14:26:19] Prefill batch, #new-seq: 1, #new-token: 167, #cached-token: 10046, token usage: 0.00, #running-req: 0, #queue-req: 0, \r\n",,terminal_output

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-0f5513f7-8bc9-4c5d-856d-79d92f75113d1751284706913-2025_06_30-13.59.01.459/source.csv

@@ -0,0 +1,52 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,263,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"1:59:01 PM [info] Activating crowd-code\n1:59:01 PM [info] Recording started\n1:59:01 PM [info] Initializing git provider using file system watchers...\n1:59:01 PM [info] Git repository found\n1:59:01 PM [info] Git provider initialized successfully\n1:59:01 PM [info] Initial git state: [object Object]\n",Log,tab
+3,3941,"TERMINAL",0,0,"/bin/python3 /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt",,terminal_command
+4,3980,"TERMINAL",0,0,"]633;E;2025-06-30 13:59:05 /bin/python3 /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt;ae710212-bbd7-466b-8215-e56dfc4f7a88]633;C",,terminal_output
+5,4004,"TERMINAL",0,0,"]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash]633;D;0",,terminal_output
+6,31538,"TERMINAL",0,0,"source .venv/bin/activate",,terminal_command
+7,33235,"TERMINAL",0,0,"bash",,terminal_focus
+8,35303,"scripts_horeka/train_tokenizer.sh",0,0,"#!/usr/bin/env bash\n\nmodule unload mpi/openmpi/5.0\nmodule unload devel/cuda/12.4\nsource .venv/bin/activate\n\ntf_records_dir=$ws_dir/knoms_tfrecords_500_shards\nws_dir='/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/'\n\njob_name=""debug""\nslurm_job_id=""0000""\n\nCHECKPOINT_DIR=$ws_dir/checkpoints/$job_name_$slurm_job_id\nmkdir -p $CHECKPOINT_DIR\n\nenv | grep SLURM\n\nsrun python train_tokenizer.py \\n    --ckpt_dir $CHECKPOINT_DIR \\n    --batch_size=16 \\n    --min_lr=4.24e-4 \\n    --max_lr=4.24e-4 \\n    --log_image_interval=100 \\n    --log \\n    --name=test-wandb-tags-$slurm_job_id \\n    --tags test tokenizer debug \\n    --entity instant-uv \\n    --project jafar \\n    --data_dir $tf_records_dir",shellscript,tab
+9,35920,"TERMINAL",0,0,"bash",,terminal_focus
+10,39024,"TERMINAL",0,0,"queue",,terminal_command
+11,39105,"TERMINAL",0,0,"]633;E;2025-06-30 13:59:40 queue;79ec3af7-6a10-4dac-bb07-e3b50f56ded4]633;C[?1049h(B[?7hEvery 1.0s: squeue --mehkn1991.localdomain: Mon Jun 30 13:59:40 2025JOBID PARTITION     NAME     USER ST\tTIME  NODES NODELIST(REASON)",,terminal_output
+12,39625,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/jafar]633;D;0",,terminal_output
+13,1168748,"TERMINAL",0,0,"salloc --time=01:00:00 --partition=accelerated --nodes=4 --ntasks-per-node=4 --gres=gpu:4 --cpus-per-task=5 --mem=200G --mail-user=mihir@pdoom.org --mail-type=ALL",,terminal_command
+14,1168797,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:30 salloc --time=01:00:00 --partition=accelerated --nodes=4 --ntasks-per-node=4 --gres=gpu:4 --cpus-per-task=5 --mem=200G --mail-user=mihir@pdoom.org --mail-type=ALL;79ec3af7-6a10-4dac-bb07-e3b50f56ded4]633;C",,terminal_output
+15,1168893,"TERMINAL",0,0,"salloc: Pending job allocation 3306136\r\nsalloc: job 3306136 queued and waiting for resources\r\n",,terminal_output
+16,1169921,"TERMINAL",0,0,"bash",,terminal_focus
+17,1170955,"TERMINAL",0,0,"queue",,terminal_command
+18,1171010,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:32 queue;ead59344-49db-4336-9336-47fae706e637]633;C[?1049h(B[?7hEvery 1.0s: squeue --mehkn1991.localdomain: Mon Jun 30 14:18:32 2025JOBID PARTITION     NAME     USER ST\tTIME  NODES NODELIST(REASON)3306136 accelerat interact tum_cte0 PD\t0:00\t  4 (Resources)",,terminal_output
+19,1172154,"TERMINAL",0,0,"3\t ",,terminal_output
+20,1173084,"TERMINAL",0,0,"4\t ",,terminal_output
+21,1174201,"TERMINAL",0,0,"5\t ",,terminal_output
+22,1174691,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D;0",,terminal_output
+23,1175459,"TERMINAL",0,0,"idle",,terminal_command
+24,1175485,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:36 idle;ead59344-49db-4336-9336-47fae706e637]633;CPartition dev_cpuonly             :     10 nodes idle\r\nPartition cpuonly                 :    154 nodes idle\r\nPartition dev_accelerated         :      0 nodes idle\r\nPartition accelerated             :      1 nodes idle\r\nPartition dev_accelerated-h100    :      0 nodes idle\r\nPartition accelerated-h100        :      0 nodes idle\r\nPartition large                   :      8 nodes idle\r\n]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D;0",,terminal_output
+25,1181058,"TERMINAL",0,0,"^C",,terminal_command
+26,1181077,"TERMINAL",0,0,"^C[?2004l\r[?2004h[?2004l\r\r\n]633;E;;ead59344-49db-4336-9336-47fae706e637]633;C]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D",,terminal_output
+27,1183520,"TERMINAL",0,0,"idle",,terminal_command
+28,1183538,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:44 idle;ead59344-49db-4336-9336-47fae706e637]633;CPartition dev_cpuonly             :     10 nodes idle\r\nPartition cpuonly                 :    154 nodes idle\r\nPartition dev_accelerated         :      2 nodes idle\r\nPartition accelerated             :      1 nodes idle\r\nPartition dev_accelerated-h100    :      0 nodes idle\r\nPartition accelerated-h100        :      0 nodes idle\r\nPartition large                   :      8 nodes idle\r\n]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D;0",,terminal_output
+29,1185175,"TERMINAL",0,0,"^C",,terminal_command
+30,1185191,"TERMINAL",0,0,"^C[?2004l\r[?2004h[?2004l\r\r\n]633;E;;ead59344-49db-4336-9336-47fae706e637]633;C]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D",,terminal_output
+31,1186579,"TERMINAL",0,0,"idle",,terminal_command
+32,1186594,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:47 idle;ead59344-49db-4336-9336-47fae706e637]633;CPartition dev_cpuonly             :     10 nodes idle\r\nPartition cpuonly                 :    154 nodes idle\r\nPartition dev_accelerated         :      2 nodes idle\r\nPartition accelerated             :      1 nodes idle\r\nPartition dev_accelerated-h100    :      0 nodes idle\r\nPartition accelerated-h100        :      0 nodes idle\r\nPartition large                   :      8 nodes idle\r\n]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D;0",,terminal_output
+33,1187256,"TERMINAL",0,0,"idle",,terminal_command
+34,1187270,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:48 idle;ead59344-49db-4336-9336-47fae706e637]633;CPartition dev_cpuonly             :     10 nodes idle\r\nPartition cpuonly                 :    154 nodes idle\r\nPartition dev_accelerated         :      2 nodes idle\r\nPartition accelerated             :      1 nodes idle\r\nPartition dev_accelerated-h100    :      0 nodes idle\r\nPartition accelerated-h100        :      0 nodes idle\r\nPartition large                   :      8 nodes idle\r\n]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D;0",,terminal_output
+35,1187998,"TERMINAL",0,0,"idle",,terminal_command
+36,1188014,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:49 idle;ead59344-49db-4336-9336-47fae706e637]633;CPartition dev_cpuonly             :     10 nodes idle\r\nPartition cpuonly                 :    154 nodes idle\r\nPartition dev_accelerated         :      2 nodes idle\r\nPartition accelerated             :      1 nodes idle\r\nPartition dev_accelerated-h100    :      0 nodes idle\r\nPartition accelerated-h100        :      0 nodes idle\r\nPartition large                   :      8 nodes idle\r\n]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D;0",,terminal_output
+37,1188374,"TERMINAL",0,0,"idle",,terminal_command
+38,1188379,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:49 idle;ead59344-49db-4336-9336-47fae706e637]633;CPartition dev_cpuonly             :     10 nodes idle\r\nPartition cpuonly                 :    154 nodes idle\r\nPartition dev_accelerated         :      2 nodes idle\r\nPartition accelerated             :      1 nodes idle\r\nPartition dev_accelerated-h100    :      0 nodes idle\r\nPartition accelerated-h100        :      0 nodes idle\r\nPartition large                   :      8 nodes idle\r\n]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D;0",,terminal_output
+39,1188713,"TERMINAL",0,0,"idle",,terminal_command
+40,1188724,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:50 idle;ead59344-49db-4336-9336-47fae706e637]633;CPartition dev_cpuonly             :     10 nodes idle\r\nPartition cpuonly                 :    154 nodes idle\r\nPartition dev_accelerated         :      2 nodes idle\r\nPartition accelerated             :      1 nodes idle\r\nPartition dev_accelerated-h100    :      0 nodes idle\r\nPartition accelerated-h100        :      0 nodes idle\r\nPartition large                   :      8 nodes idle\r\n]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D;0",,terminal_output
+41,1189024,"TERMINAL",0,0,"idle",,terminal_command
+42,1189039,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:50 idle;ead59344-49db-4336-9336-47fae706e637]633;CPartition dev_cpuonly             :     10 nodes idle\r\nPartition cpuonly                 :    154 nodes idle\r\nPartition dev_accelerated         :      2 nodes idle\r\nPartition accelerated             :      1 nodes idle\r\nPartition dev_accelerated-h100    :      0 nodes idle\r\nPartition accelerated-h100        :      0 nodes idle\r\nPartition large                   :      8 nodes idle\r\n]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D;0",,terminal_output
+43,1189680,"TERMINAL",0,0,"",,terminal_command
+44,1189694,"TERMINAL",0,0,"\r\n[?2004l\r]633;E;;ead59344-49db-4336-9336-47fae706e637]633;C]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D",,terminal_output
+45,1190195,"TERMINAL",0,0,"idle",,terminal_command
+46,1190210,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:51 idle;ead59344-49db-4336-9336-47fae706e637]633;CPartition dev_cpuonly             :     10 nodes idle\r\nPartition cpuonly                 :    154 nodes idle\r\nPartition dev_accelerated         :      1 nodes idle\r\nPartition accelerated             :      1 nodes idle\r\nPartition dev_accelerated-h100    :      0 nodes idle\r\nPartition accelerated-h100        :      0 nodes idle\r\nPartition large                   :      8 nodes idle\r\n]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D;0",,terminal_output
+47,1192901,"TERMINAL",0,0,"idle",,terminal_command
+48,1192917,"TERMINAL",0,0,"]633;E;2025-06-30 14:18:54 idle;ead59344-49db-4336-9336-47fae706e637]633;CPartition dev_cpuonly             :     10 nodes idle\r\nPartition cpuonly                 :    152 nodes idle\r\nPartition dev_accelerated         :      1 nodes idle\r\nPartition accelerated             :      1 nodes idle\r\nPartition dev_accelerated-h100    :      0 nodes idle\r\nPartition accelerated-h100        :      0 nodes idle\r\nPartition large                   :      8 nodes idle\r\n]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/Projects/jafar]633;D;0",,terminal_output
+49,1194207,"TERMINAL",0,0,"salloc",,terminal_focus
+50,1198907,"TERMINAL",0,0,"^Csalloc: Job allocation 3306136 has been revoked.\r\nsalloc: Job aborted due to signal\r\n]0;tum_cte0515@hkn1991:~/Projects/jafar]633;D;1",,terminal_output
+51,1199160,"TERMINAL",0,0,"^C",,terminal_command
+52,1199175,"TERMINAL",0,0,"^C[?2004l\r[?2004h[?2004l\r\r\n]633;E;;79ec3af7-6a10-4dac-bb07-e3b50f56ded4]633;C]0;tum_cte0515@hkn1991:~/Projects/jafar]633;D",,terminal_output

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-28f4aa5c-0534-40eb-ae05-51501d68e4871752860706222-2025_07_18-19.45.48.539/source.csv

@@ -0,0 +1,90 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,667,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"7:45:48 PM [info] Activating crowd-code\n7:45:48 PM [info] Recording started\n7:45:48 PM [info] Initializing git provider using file system watchers...\n7:45:48 PM [info] Git repository found\n7:45:48 PM [info] Git provider initialized successfully\n7:45:48 PM [info] Initial git state: [object Object]\n",Log,tab
+3,228069,"TERMINAL",0,0,"queue",,terminal_command
+4,228163,"TERMINAL",0,0,"]633;E;2025-07-18 19:49:36 queue;9dd50732-d6d7-4d7e-a22c-51ae92e646cb]633;C[?1049h(B[?7hEvery 1.0s: squeue --mehkn1991.localdomain: Fri Jul 18 19:49:36 2025JOBID PARTITION     NAME     USER ST\tTIME  NODES NODELIST(REASON)3357893 accelerat interact tum_cte0  R    3:08:29\t  2 hkn[0436,0708]3357894 accelerat interact tum_cte0  R    3:45:12\t  1 hkn0715\t\t",,terminal_output
+5,229211,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/jafar]633;D;0",,terminal_output
+6,231917,"TERMINAL",0,0,"idling",,terminal_command
+7,232013,"TERMINAL",0,0,"]633;E;2025-07-18 19:49:40 idling;9dd50732-d6d7-4d7e-a22c-51ae92e646cb]633;C[?1049h(B[?7hEvery 1.0s: sinfo_t_idlehkn1991.localdomain: Fri Jul 18 19:49:40 2025Partition dev_cpuonly:     12 nodes idle\rPartition cpuonly:\t 7 nodes idle\rPartition dev_accelerated:\t 3 nodes idle\rPartition accelerated:\t 1 nodes idle\rPartition dev_accelerated-h100    :\t 1 nodes idle\rPartition accelerated-h100:\t 0 nodes idle\rPartition large:\t 8 nodes idle",,terminal_output
+8,233034,"TERMINAL",0,0,"1\t",,terminal_output
+9,233586,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/jafar]633;D;0",,terminal_output
+10,237086,"TERMINAL",0,0,"queue",,terminal_command
+11,237134,"TERMINAL",0,0,"]633;E;2025-07-18 19:49:45 queue;9dd50732-d6d7-4d7e-a22c-51ae92e646cb]633;C[?1049h(B[?7hEvery 1.0s: squeue --mehkn1991.localdomain: Fri Jul 18 19:49:45 2025JOBID PARTITION     NAME     USER ST\tTIME  NODES NODELIST(REASON)3357893 accelerat interact tum_cte0  R    3:08:38\t  2 hkn[0436,0708]3357894 accelerat interact tum_cte0  R    3:45:21\t  1 hkn0715\t\t",,terminal_output
+12,237968,"TERMINAL",0,0,"692",,terminal_output
+13,238857,"TERMINAL",0,0,"7403",,terminal_output
+14,239544,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/jafar]633;D;0",,terminal_output
+15,241735,"TERMINAL",0,0,"scancel 3357893",,terminal_command
+16,241795,"TERMINAL",0,0,"]633;E;2025-07-18 19:49:50 scancel 3357893;9dd50732-d6d7-4d7e-a22c-51ae92e646cb]633;C",,terminal_output
+17,241927,"TERMINAL",0,0,"]0;tum_cte0515@hkn1991:~/Projects/jafar]633;D;0]633;P;Cwd=/home/hk-project-p0023960/tum_cte0515/Projects/jafar",,terminal_output
+18,243399,"TERMINAL",0,0,"scancel 3357894",,terminal_command
+19,243459,"TERMINAL",0,0,"]633;E;2025-07-18 19:49:51 scancel 3357894;9dd50732-d6d7-4d7e-a22c-51ae92e646cb]633;C]0;tum_cte0515@hkn1991:~/Projects/jafar]633;D;0]633;P;Cwd=/home/hk-project-p0023960/tum_cte0515/Projects/jafar",,terminal_output
+20,245208,"TERMINAL",0,0,"idling",,terminal_command
+21,245252,"TERMINAL",0,0,"]633;E;2025-07-18 19:49:53 idling;9dd50732-d6d7-4d7e-a22c-51ae92e646cb]633;C",,terminal_output
+22,245311,"TERMINAL",0,0,"[?1049h(B[?7hEvery 1.0s: sinfo_t_idlehkn1991.localdomain: Fri Jul 18 19:49:53 2025Partition dev_cpuonly:     12 nodes idle\rPartition cpuonly:\t 7 nodes idle\rPartition dev_accelerated:\t 3 nodes idle\rPartition accelerated:\t 3 nodes idle\rPartition dev_accelerated-h100    :\t 1 nodes idle\rPartition accelerated-h100:\t 0 nodes idle\rPartition large:\t 8 nodes idle",,terminal_output
+23,246324,"TERMINAL",0,0,"4\t",,terminal_output
+24,246706,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/jafar]633;D;0",,terminal_output
+25,247830,"models/dynamics.py",0,0,"from typing import Dict, Any\n\nimport jax\nimport jax.numpy as jnp\nimport flax.linen as nn\nimport einops\n\nfrom utils.nn import STTransformer\n\n\nclass DynamicsMaskGIT(nn.Module):\n    """"""MaskGIT dynamics model""""""\n\n    model_dim: int\n    num_latents: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    mask_limit: float\n    param_dtype: jnp.dtype\n    dtype: jnp.dtype\n\n    def setup(self):\n        self.dynamics = STTransformer(\n            self.model_dim,\n            self.num_latents,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n            self.param_dtype,\n            self.dtype,\n            spacial_bert=True,\n        )\n        self.patch_embed = nn.Embed(self.num_latents, self.model_dim)\n        self.mask_token = self.param(\n            ""mask_token"",\n            nn.initializers.lecun_uniform(),\n            (1, 1, 1, self.model_dim),\n        )\n        self.action_up = nn.Dense(\n            self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        # --- Mask videos ---\n        vid_embed = self.patch_embed(batch[""video_tokens""])\n        if training:\n            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n        else:\n            mask = None\n\n        # --- Predict transition ---\n        act_embed = self.action_up(batch[""latent_actions""])\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        logits = self.dynamics(vid_embed)\n        return dict(token_logits=logits, mask=mask)\n\n\nclass DynamicsAutoregressive(nn.Module):\n    """"""Autoregressive (causal) dynamics model""""""\n\n    model_dim: int\n    num_latents: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    param_dtype: jnp.dtype\n    dtype: jnp.dtype\n\n    def setup(self):\n        self.dynamics = STTransformer(\n            self.model_dim,\n            self.num_latents,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n            self.param_dtype,\n            self.dtype,\n            spacial_bert=False,\n        )\n        self.patch_embed = nn.Embed(self.num_latents, self.model_dim)\n        self.action_up = nn.Dense(\n            self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        vid_embed = self.patch_embed(batch[""video_tokens""])\n        act_embed = self.action_up(batch[""latent_actions""])\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        vid_embed_padded = jnp.pad(vid_embed, ((0, 0), (1, 0), (1, 0), (0, 0)))\n        logits = self.dynamics(vid_embed_padded)[:, 1:, 1:]\n        mask = jnp.ones(vid_embed.shape[:-1])\n        next_tokens = jnp.argmax(logits, axis=-1)\n        print(next_tokens.shape)\n        jax.debug.breakpoint()\n        return dict(token_logits=logits, mask=mask)",python,tab
+26,247835,"models/dynamics.py",2889,0,"",python,selection_mouse
+27,247936,"models/dynamics.py",2887,9,"act_embed",python,selection_mouse
+28,248563,"models/dynamics.py",2826,0,"",python,selection_mouse
+29,251754,"models/dynamics.py",2995,0,"",python,selection_mouse
+30,253317,"models/dynamics.py",2989,0,"",python,selection_mouse
+31,254853,"models/dynamics.py",2988,0,"",python,selection_mouse
+32,303300,"models/lam.py",0,0,"from typing import Dict, Any\n\nimport jax.numpy as jnp\nimport flax.linen as nn\n\nfrom utils.preprocess import patchify, unpatchify\nfrom utils.nn import STTransformer, VectorQuantizer\n\n\nclass LatentActionModel(nn.Module):\n    """"""Latent Action ST-ViVit VQ-VAE""""""\n\n    in_dim: int\n    model_dim: int\n    latent_dim: int\n    num_latents: int\n    patch_size: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    codebook_dropout: float\n    param_dtype: jnp.dtype\n    dtype: jnp.dtype\n\n    def setup(self):\n        self.patch_token_dim = self.in_dim * self.patch_size**2\n        self.encoder = STTransformer(\n            self.model_dim,\n            self.latent_dim,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n            self.param_dtype,\n            self.dtype,\n        )\n        self.action_in = self.param(\n            ""action_in"",\n            nn.initializers.lecun_uniform(),\n            (1, 1, 1, self.patch_token_dim),\n        )\n        self.vq = VectorQuantizer(\n            self.latent_dim,\n            self.num_latents,\n            self.codebook_dropout,\n        )\n        self.patch_up = nn.Dense(\n            self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )\n        self.action_up = nn.Dense(\n            self.model_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )\n        self.decoder = STTransformer(\n            self.model_dim,\n            self.patch_token_dim,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n            self.param_dtype,\n            self.dtype,\n        )\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        # --- Encode + VQ ---\n        H, W = batch[""videos""].shape[2:4]\n        outputs = self.vq_encode(batch[""videos""], training)\n        video_action_patches = self.action_up(outputs[""z_q""]) + self.patch_up(\n            outputs[""patches""][:, :-1]\n        )\n        del outputs[""patches""]\n\n        # --- Decode ---\n        video_recon = self.decoder(video_action_patches)\n        video_recon = video_recon.astype(jnp.float32)\n        video_recon = nn.sigmoid(video_recon)\n        video_recon = video_recon.astype(self.dtype)\n        outputs[""recon""] = unpatchify(video_recon, self.patch_size, H, W)\n        return outputs\n\n    def vq_encode(self, videos: Any, training: bool = True) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        B, T = videos.shape[:2]\n        patches = patchify(videos, self.patch_size)\n        action_pad = jnp.broadcast_to(self.action_in, (B, T, 1, self.patch_token_dim))\n        # FIXME mihir do this the other way around\n        padded_patches = jnp.concatenate((action_pad, patches), axis=2)\n\n        # --- Encode ---\n        z = self.encoder(padded_patches)  # (B, T, N, E)\n        # Get latent action for all future frames\n        z = z[:, 1:, 0]  # (B, T-1, E)\n\n        # --- Vector quantize ---\n        z = z.reshape(B * (T - 1), self.latent_dim)\n        z_q, z, emb, indices = self.vq(z, training)\n        z_q = z_q.reshape(B, T - 1, 1, self.latent_dim)\n        return dict(patches=patches, z_q=z_q, z=z, emb=emb, indices=indices)\n",python,tab
+33,304984,"genie.py",0,0,"from typing import Dict, Any\n\nimport optax\nimport jax\nimport jax.numpy as jnp\nimport flax.linen as nn\nfrom flax.training.train_state import TrainState\nimport orbax.checkpoint as ocp\n\nfrom models.dynamics import DynamicsMaskGIT, DynamicsAutoregressive\nfrom models.lam import LatentActionModel\nfrom models.tokenizer import TokenizerVQVAE\n\nimport os\nimport grain\n\n\nclass Genie(nn.Module):\n    """"""Genie model""""""\n\n    # --- Tokenizer ---\n    in_dim: int\n    tokenizer_dim: int\n    latent_patch_dim: int\n    num_patch_latents: int\n    patch_size: int\n    tokenizer_num_blocks: int\n    tokenizer_num_heads: int\n    # --- LAM ---\n    lam_dim: int\n    latent_action_dim: int\n    num_latent_actions: int\n    lam_patch_size: int\n    lam_num_blocks: int\n    lam_num_heads: int\n    lam_co_train: bool\n    # --- Dynamics ---\n    dyna_dim: int\n    dyna_num_blocks: int\n    dyna_num_heads: int\n    use_maskgit: bool\n    param_dtype: jnp.dtype\n    dtype: jnp.dtype\n    dropout: float = 0.0\n    mask_limit: float = 0.0\n\n    def setup(self):\n        self.tokenizer = TokenizerVQVAE(\n            in_dim=self.in_dim,\n            model_dim=self.tokenizer_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_patch_latents,\n            patch_size=self.patch_size,\n            num_blocks=self.tokenizer_num_blocks,\n            num_heads=self.tokenizer_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )\n        self.lam = LatentActionModel(\n            in_dim=self.in_dim,\n            model_dim=self.lam_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_latent_actions,\n            patch_size=self.lam_patch_size,\n            num_blocks=self.lam_num_blocks,\n            num_heads=self.lam_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )\n\n        if self.use_maskgit:\n            self.dynamics = DynamicsMaskGIT(\n                model_dim=self.dyna_dim,\n                num_latents=self.num_patch_latents,\n                num_blocks=self.dyna_num_blocks,\n                num_heads=self.dyna_num_heads,\n                dropout=self.dropout,\n                mask_limit=self.mask_limit,\n                param_dtype=self.param_dtype,\n                dtype=self.dtype,\n            ) \n        else:\n            self.dynamics = DynamicsAutoregressive(\n                model_dim=self.dyna_dim,\n                num_latents=self.num_patch_latents,\n                num_blocks=self.dyna_num_blocks,\n                num_heads=self.dyna_num_heads,\n                dropout=self.dropout,\n                param_dtype=self.param_dtype,\n                dtype=self.dtype,\n            )\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        tokenizer_outputs = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        lam_outputs = self.lam.vq_encode(batch[""videos""], training=False)\n        latent_actions = jax.lax.cond(\n            self.lam_co_train,\n            lambda: lam_outputs[""z_q""],\n            lambda: jax.lax.stop_gradient(lam_outputs[""z_q""])\n        )\n        outputs = dict(\n            video_tokens=jax.lax.stop_gradient(tokenizer_outputs[""indices""]),\n            latent_actions=latent_actions,\n        )\n        outputs[""mask_rng""] = batch[""mask_rng""]\n        dyna_outputs = self.dynamics(outputs, training)\n        outputs.update(dyna_outputs)\n        mle_indices = jnp.argmax(outputs[""token_logits""], axis=-1)\n        outputs[""recon""] = self.tokenizer.decode(\n            mle_indices, batch[""videos""].shape[2:4]\n        )\n        outputs[""lam_indices""] = lam_outputs[""indices""]\n        return outputs\n\n\n    def sample_causal(\n        self,\n        batch: Dict[str, Any],\n        seq_len: int,\n        temperature: float = 1,\n        sample_argmax: bool = False,\n    ):\n        """"""\n        Autoregressively samples up to `seq_len` future frames using the causal transformer backend.\n\n        - Input frames are tokenized once.\n        - Future frames are generated one at a time, each conditioned on all previous frames.\n        - All frames are detokenized in a single pass at the end.\n\n        Args:\n            batch: Dict with at least ""videos"" (B, T, H, W, C)\n            seq_len: total number of frames to generate (including context)\n            temperature: sampling temperature\n            sample_argmax: if True, use argmax instead of sampling\n\n        Returns:\n            Generated video frames (B, seq_len, H, W, C)\n        """"""\n        # --- Encode context frames ---\n        tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        token_idxs = tokenizer_out[""indices""]  # (B, T, N)\n        B, T, N = token_idxs.shape\n\n        # jax.debug.print(""token_idxs shape: {}"", token_idxs.shape)\n        # --- Prepare initial token sequence ---\n        # Pad with zeros for future frames\n        pad_shape = (B, seq_len - T, N)\n        token_idxs_full = jnp.concatenate(\n            [token_idxs, jnp.zeros(pad_shape, dtype=token_idxs.dtype)], axis=1\n        )  # (B, seq_len, N)\n\n        # --- Prepare latent actions ---\n        action_tokens = self.lam.vq.get_codes(batch[""latent_actions""]) # (B, S-1, )\n        # --- Autoregressive generation loop ---\n        rng = batch[""rng""]\n        for t in range(T, seq_len):\n            for n in range(32):\n                dyna_inputs = {\n                    ""video_tokens"": token_idxs_full,\n                    ""latent_actions"": action_tokens\n                    }\n                # jax.debug.print(""token_idxs_full 0: {}"", token_idxs_full[0,:,0])\n                dyna_outputs = self.dynamics(dyna_inputs, training=False)\n                # # dyna_outputs[""token_logits""]: (B, t, N, vocab_size)\n                # # We want the logits for the last time step (frame t-1 predicting t)\n                # jax.debug.breakpoint()\n                next_token_logits = dyna_outputs[""token_logits""][:, t, n, :].astype(jnp.float32)  # (B, 1, vocab_size)\n\n                # Sample or argmax for each patch\n                if sample_argmax:\n                    next_token = jnp.argmax(next_token_logits, axis=-1)  # (B, 1)\n                else:\n                    rng, step_rng = jax.random.split(rng)\n                    next_token = jax.random.categorical(\n                        step_rng, next_token_logits / temperature, axis=-1\n                    )  # (B, 1)\n\n                # Insert the generated tokens into the sequence\n                token_idxs_full = token_idxs_full.at[:, t, n].set(next_token)\n\n        # --- Decode all tokens at once at the end ---\n        final_frames = self.tokenizer.decode(\n            token_idxs_full, video_hw=batch[""videos""].shape[2:4]\n        )\n        return final_frames\n\n\n    @nn.compact\n    def sample_maskgit(\n        self,\n        batch: Dict[str, Any],\n        seq_len: int,\n        steps: int = 25,\n        temperature: float = 1,\n        sample_argmax: bool = False,\n    ) -> Any:\n        """"""\n        Autoregressively samples up to `seq_len` future frames, following Figure 8 of the paper.\n\n        - Input frames are tokenized once.\n        - Future frames are generated autoregressively in token space.\n        - All frames are detokenized in a single pass.\n\n        Note:\n        - For interactive or step-wise sampling, detokenization should occur after each action.\n        - To maintain consistent tensor shapes across timesteps, all current and future frames are decoded at every step.\n        - Temporal causal structure is preserved by \n            a) reapplying the mask before each decoding step.\n            b) a temporal causal mask is applied within each ST-transformer block.\n\n        Dimension keys:\n            B: batch size  \n            T: number of input (conditioning) frames  \n            N: patches per frame  \n            S: sequence length  \n            A: action space  \n            D: model latent dimension\n        """"""\n        # --- Encode videos and actions ---\n        tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        token_idxs = tokenizer_out[""indices""] # (B, T, N)\n        B, T, N = token_idxs.shape\n        pad_shape = (B, seq_len - T, N)\n        pad = jnp.zeros(pad_shape, dtype=token_idxs.dtype)\n        token_idxs = jnp.concatenate([token_idxs, pad], axis=1) # (B, S, N)\n        action_tokens = self.lam.vq.get_codes(batch[""latent_actions""]) \n\n        MaskGITLoop = nn.scan(\n            MaskGITStep,\n            variable_broadcast=""params"",\n            split_rngs={""params"": False},\n            in_axes=0,\n            out_axes=0,\n            length=steps,\n        )\n    \n        loop_fn = MaskGITLoop(\n            dynamics=self.dynamics,\n            tokenizer=self.tokenizer,\n            temperature=temperature,\n            sample_argmax=sample_argmax,\n            steps=steps,\n        )\n\n        def generation_step_fn(carry, step_t):\n            rng, current_token_idxs = carry\n            rng, step_rng = jax.random.split(rng)\n\n            # Mask current and future frames (i.e., t >= step_t)\n            mask = jnp.arange(seq_len) >= step_t # (S,)\n            mask = jnp.broadcast_to(mask[None, :, None], (B, seq_len, N)) # (B, S, N)\n            mask = mask.astype(bool)\n            masked_token_idxs = current_token_idxs * ~mask\n\n            # --- Initialize and run MaskGIT loop ---\n            init_carry_maskgit = (\n                step_rng,\n                masked_token_idxs,\n                mask,\n                action_tokens,\n            )\n            final_carry_maskgit, _ = loop_fn(init_carry_maskgit, jnp.arange(steps))\n            updated_token_idxs = final_carry_maskgit[1]\n            new_carry = (rng, updated_token_idxs)\n            return new_carry, None\n\n        # --- Run the autoregressive generation using scan ---\n        initial_carry = (batch[""rng""], token_idxs)\n        timesteps_to_scan = jnp.arange(T, seq_len)\n        final_carry, _ = jax.lax.scan(\n            generation_step_fn,\n            initial_carry,\n            timesteps_to_scan\n        )\n        final_token_idxs = final_carry[1]\n\n        # --- Decode all tokens at once at the end ---\n        final_frames = self.tokenizer.decode(\n            final_token_idxs,\n            video_hw=batch[""videos""].shape[2:4],\n        )\n        return final_frames\n\n    def vq_encode(self, batch, training) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\n\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs) # (B, S, N, D)\n        mask_token = self.dynamics.mask_token  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None] # (B, S, N, 1) \n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jax.random.categorical(_rng, final_logits)\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n\ndef restore_genie_components(\n    train_state: TrainState,\n    sharding: jax.sharding.NamedSharding,\n    grain_iterator: grain.DataLoaderIterator,\n    inputs: Dict[str, jax.Array],\n    rng: jax.Array,\n    args,\n):\n    """"""Restore pre-trained Genie components""""""\n    rng, _rng = jax.random.split(rng)\n\n    # dummy values since we only use tx to initialize the dummy train states\n    dummy_tx = optax.adamw(\n        learning_rate=optax.constant_schedule(args.max_lr),\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n    )\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add('model_state', ocp.args.StandardRestore, ocp.handlers.StandardCheckpointHandler)\n    handler_registry.add('dataloader_state', grain.checkpoint.CheckpointRestore, grain.checkpoint.CheckpointHandler)\n    \n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        step_format_fixed_length=6,\n    )\n    tokenizer_checkpoint_manager = ocp.CheckpointManager(\n        directory=args.tokenizer_checkpoint,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    dummy_tokenizer = TokenizerVQVAE(\n        in_dim=args.image_channels,\n        model_dim=args.tokenizer_dim,\n        latent_dim=args.latent_patch_dim,\n        num_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.tokenizer_num_blocks,\n        num_heads=args.tokenizer_num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.dropout,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n    )\n    tokenizer_init_params = dummy_tokenizer.init(_rng, inputs)\n    dummy_tokenizer_train_state = TrainState.create(\n        apply_fn=dummy_tokenizer.apply, params=tokenizer_init_params, tx=dummy_tx\n    )\n    abstract_sharded_tokenizer_state = _create_abstract_sharded_pytree(\n        dummy_tokenizer_train_state, sharding\n    )\n    restored_tokenizer = tokenizer_checkpoint_manager.restore(\n        step=tokenizer_checkpoint_manager.latest_step(),\n        args=ocp.args.Composite(\n            model_state=ocp.args.StandardRestore(abstract_sharded_tokenizer_state),\n            dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),\n        ),\n    )[""model_state""]\n    restored_tokenizer_params = restored_tokenizer.params[""params""]\n    train_state.params[""params""][""tokenizer""].update(restored_tokenizer_params)\n    tokenizer_checkpoint_manager.close()\n\n    if args.lam_checkpoint:\n        lam_checkpoint_manager = ocp.CheckpointManager(\n            directory=args.lam_checkpoint,\n            options=checkpoint_options,\n            handler_registry=handler_registry,\n        )\n        dummy_lam = LatentActionModel(\n            in_dim=args.image_channels,\n            model_dim=args.lam_dim,\n            latent_dim=args.latent_patch_dim,\n            num_latents=args.num_latent_actions,\n            patch_size=args.lam_patch_size,\n            num_blocks=args.lam_num_blocks,\n            num_heads=args.lam_num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n        )\n        lam_init_params = dummy_lam.init(_rng, inputs)\n        dummy_lam_train_state = TrainState.create(\n            apply_fn=dummy_lam.apply, params=lam_init_params, tx=dummy_tx\n        )\n        abstract_sharded_lam_state = _create_abstract_sharded_pytree(\n            dummy_lam_train_state, sharding\n        )\n        restored_lam = lam_checkpoint_manager.restore(\n            step=lam_checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.StandardRestore(abstract_sharded_lam_state),\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),\n            ),\n        )[""model_state""]\n        restored_lam_params = restored_lam.params[""params""]\n        # Genie does not initialize all LAM modules, thus we omit those extra modules during restoration\n        # (f.srambical) FIXME: Currently, this is a small HBM memory crunch since the LAM's decoder is loaded into HBM and immediately dicarded.\n        # A workaround would be to restore to host memory first, and only move the weights to HBM after pruning the decoder\n        restored_lam_params = {\n            k: v\n            for k, v in restored_lam_params.items()\n            if k in train_state.params[""params""][""lam""]\n        }\n        train_state.params[""params""][""lam""].update(restored_lam_params)\n        lam_checkpoint_manager.close()\n\n    return train_state\n\ndef _create_abstract_sharded_pytree(pytree_template, sharding_spec):\n    """"""Replaces arrays in a pytree with ShapeDtypeStructs having the given sharding.""""""\n\n    def map_fn(leaf_template):\n        if hasattr(leaf_template, ""shape"") and hasattr(leaf_template, ""dtype""):\n            return jax.ShapeDtypeStruct(\n                leaf_template.shape, leaf_template.dtype, sharding=sharding_spec\n            )\n        return leaf_template\n\n    return jax.tree_util.tree_map(map_fn, pytree_template)",python,tab
+34,311383,"train_dynamics.py",0,0,"from dataclasses import dataclass, field\nimport os\n\nimport einops\nfrom flax.training.train_state import TrainState\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\n\nfrom genie import Genie, restore_genie_components\nfrom utils.dataloader import get_dataloader\nfrom utils.lr_utils import get_lr_schedule\nfrom utils.parameter_utils import count_parameters_by_component\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = 10000 # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    warmup_steps: int = 5000\n    lr_schedule : str = ""wsd"" # supported options: wsd, cos\n    # Tokenizer\n    tokenizer_dim: int = 512\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 8\n    tokenizer_num_heads: int = 8\n    tokenizer_checkpoint: str = """"\n    # LAM\n    lam_dim: int = 512\n    latent_action_dim: int = 32\n    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 8\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_dim: int = 512\n    dyna_num_blocks: int = 12\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    use_maskgit: bool = False\n    param_dtype: jnp.dtype = jnp.float32\n    dtype: jnp.dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_dynamics""\n    tags: list[str] = field(default_factory=lambda: [""dynamics""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n\n\nargs = tyro.cli(Args)\n\n\ndef dynamics_loss_fn(params, state, inputs):\n    """"""Compute masked dynamics loss""""""\n    inputs[""videos""] = inputs[""videos""].astype(args.dtype) / 255.0\n    outputs = state.apply_fn(\n        params,\n        inputs,\n        training=True,\n        rngs={""params"": inputs[""rng""], ""dropout"": inputs[""dropout_rng""]},\n    )\n    mask = outputs[""mask""]\n    outputs[""token_logits""] = outputs[""token_logits""].astype(jnp.float32)\n    logits = outputs[""token_logits""]\n    targets = outputs[""video_tokens""]\n\n    # if not args.use_maskgit:\n        # logits = outputs[""token_logits""][:, :, :-1]\n        # targets = outputs[""video_tokens""][:, :, 1:]\n        # mask = outputs[""mask""][:, :, 1:] \n\n    ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n        logits, targets\n    )\n    ce_loss = (mask * ce_loss).sum() / mask.sum()\n    acc = logits.argmax(-1) == targets\n    acc = (mask * acc).sum() / mask.sum()\n    select_probs = jax.nn.softmax(logits)\n    gt = inputs[""videos""].clip(0, 1).reshape(-1, *inputs[""videos""].shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = pix.psnr(gt, recon).mean() # type: ignore\n    ssim = pix.ssim(gt, recon).mean() # type: ignore\n    _, index_counts_lam = jnp.unique_counts(\n        jnp.ravel(outputs[""lam_indices""]), size=args.num_latent_actions, fill_value=0\n    )\n    _, index_counts_tokenizer = jnp.unique_counts(\n        jnp.ravel(outputs[""video_tokens""]), size=args.num_patch_latents, fill_value=0\n    )\n    codebook_usage_lam = (index_counts_lam != 0).mean()\n    codebook_usage_tokenizer = (index_counts_tokenizer != 0).mean()\n    metrics = dict(\n        cross_entropy_loss=ce_loss,\n        masked_token_accuracy=acc,\n        select_logit=logits.max(-1).mean(),\n        select_p=select_probs.max(-1).mean(),\n        entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage_lam=codebook_usage_lam,\n        codebook_usage_tokenizer=codebook_usage_tokenizer,\n    )\n    return ce_loss, (outputs[""recon""], metrics)\n\n\n@jax.jit\ndef train_step(state, inputs):\n    """"""Update state and compute metrics""""""\n    grad_fn = jax.value_and_grad(dynamics_loss_fn, has_aux=True, allow_int=True)\n    (loss, (recon, metrics)), grads = grad_fn(state.params, state, inputs)\n    state = state.apply_gradients(grads=grads)\n    if args.log_gradients:\n        metrics[""gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""dynamics""]\n        )\n    return state, loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.PRNGKey(args.seed)\n\n    # --- Initialize model ---\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_latent_actions=args.num_latent_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=not args.lam_checkpoint,\n        # Dynamics\n        dyna_dim=args.dyna_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n        use_maskgit=args.use_maskgit,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n    )\n    rng, _rng = jax.random.split(rng)\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    dummy_inputs = dict(\n        videos=jnp.zeros(\n            (per_device_batch_size_for_init, args.seq_len, *image_shape),\n            dtype=args.dtype,\n        ),\n        action=jnp.zeros(\n            (per_device_batch_size_for_init, args.seq_len), dtype=args.dtype\n        ),\n        mask_rng=_rng,\n    )\n    rng, _rng = jax.random.split(rng)\n    init_params = genie.init(_rng, dummy_inputs)\n\n    param_counts = count_parameters_by_component(init_params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    lr_schedule = get_lr_schedule(args.lr_schedule, \n                                  args.init_lr, \n                                  args.max_lr, \n                                  args.decay_end, \n                                  args.num_steps, \n                                  args.warmup_steps, \n                                  args.wsd_decay_steps)\n    tx = optax.adamw(learning_rate=lr_schedule, b1=0.9, b2=0.9, weight_decay=1e-4, mu_dtype=args.dtype)\n    train_state = TrainState.create(apply_fn=genie.apply, params=init_params, tx=tx)\n\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    train_state = jax.device_put(train_state, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    step = 0\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.StandardSave, ocp.handlers.StandardCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.StandardRestore, ocp.handlers.StandardCheckpointHandler\n    )\n    handler_registry.add(""dataloader_state"", grain.checkpoint.CheckpointSave, grain.checkpoint.CheckpointHandler)  # type: ignore\n    handler_registry.add(""dataloader_state"", grain.checkpoint.CheckpointRestore, grain.checkpoint.CheckpointHandler)  # type: ignore\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    # --- Create DataLoaderIterator from dataloader ---\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n    # --- Restore checkpoint ---\n    if args.restore_ckpt:\n        # Restore full dynamics model\n        abstract_train_state = jax.tree_util.tree_map(\n            ocp.utils.to_shape_dtype_struct, train_state\n        )\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.StandardRestore(abstract_train_state),\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),\n            ),\n        )\n        train_state = restored[""model_state""]\n        grain_iterator = restored[""dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    else:\n        # Restore from pre-trained tokenizer (and LAM)\n        train_state = restore_genie_components(\n            train_state, replicated_sharding, grain_iterator, dummy_inputs, rng, args\n        )\n\n    # --- TRAIN LOOP ---\n    dataloader = (jax.make_array_from_process_local_data(videos_sharding, elem) for elem in grain_iterator)  # type: ignore\n    while step < args.num_steps:\n        # for videos in dataloader:\n        videos = np.load(""overfit_dir/corner_8repl.npy"")\n        videos = jax.make_array_from_process_local_data(videos_sharding, videos)\n        while True:\n            # --- Train step ---\n            rng, _rng, _rng_dropout, _rng_mask = jax.random.split(rng, 4)\n\n            inputs = dict(\n                videos=videos,\n                rng=_rng,\n                dropout_rng=_rng_dropout,\n                mask_rng=_rng_mask,\n            )\n            train_state, loss, recon, metrics = train_step(train_state, inputs)\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0].astype(jnp.float32) #/ 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[args.seq_len - 1])),\n                            recon=wandb.Image(np.asarray(recon_seq[args.seq_len - 1])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.StandardSave(train_state),\n                        dataloader_state=grain.checkpoint.CheckpointSave(\n                            grain_iterator\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n",python,tab
+35,312861,"train_dynamics.py",2979,0,"",python,selection_mouse
+36,328600,"models/tokenizer.py",0,0,"from typing import Dict, Any, Tuple\n\nimport flax.linen as nn\nimport jax.numpy as jnp\n\nfrom utils.preprocess import patchify, unpatchify\nfrom utils.nn import STTransformer, VectorQuantizer\n\n\nclass TokenizerVQVAE(nn.Module):\n    """"""ST-ViVit VQ-VAE""""""\n\n    in_dim: int\n    model_dim: int\n    latent_dim: int\n    num_latents: int\n    patch_size: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    codebook_dropout: float\n    param_dtype: jnp.dtype\n    dtype: jnp.dtype\n\n    def setup(self):\n        self.encoder = STTransformer(\n            self.model_dim,\n            self.latent_dim,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n            self.param_dtype,\n            self.dtype,\n        )\n        self.vq = VectorQuantizer(\n            self.latent_dim,\n            self.num_latents,\n            self.codebook_dropout,\n        )\n        self.out_dim = self.in_dim * self.patch_size**2\n        self.decoder = STTransformer(\n            self.model_dim,\n            self.out_dim,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n            self.param_dtype,\n            self.dtype,\n        )\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        H, W = batch[""videos""].shape[2:4]\n        outputs = self.vq_encode(batch[""videos""], training)\n        recon = self.decoder(outputs[""z_q""])  # (B, T, H_down * W_down, C)\n        recon = recon.astype(jnp.float32)\n        recon = nn.sigmoid(recon)\n        recon = recon.astype(self.dtype)\n        outputs[""recon""] = unpatchify(recon, self.patch_size, H, W)\n        return outputs\n\n    def vq_encode(self, videos: Any, training: bool = True) -> Dict[str, Any]:\n        # --- Preprocess + encode ---\n        B, T = videos.shape[:2]\n        x = patchify(videos, self.patch_size)\n        N = x.shape[2]\n        x = self.encoder(x)  # (B, T, N, E)\n\n        # --- Vector quantize ---\n        x = x.reshape(B * T * N, self.latent_dim)\n        z_q, z, emb, indices = self.vq(x, training)\n        z_q = z_q.reshape(B, T, N, self.latent_dim)\n        indices = indices.reshape(B, T, N)\n        return dict(z_q=z_q, z=z, emb=emb, indices=indices)\n\n    def decode(self, indices: Any, video_hw: Tuple[int, int]):\n        z = self.vq.codebook[indices]\n        recon = self.decoder(z)\n        recon = recon.astype(jnp.float32)\n        recon = nn.sigmoid(recon)\n        recon = recon.astype(self.dtype)\n        return unpatchify(recon, self.patch_size, *video_hw)\n",python,tab
+37,329565,"models/dynamics.py",0,0,"",python,tab
+38,330894,"models/dynamics.py",3065,0,"",python,selection_mouse
+39,331494,"models/dynamics.py",3070,0,"",python,selection_mouse
+40,332511,"models/dynamics.py",3069,0,"",python,selection_command
+41,335761,"models/dynamics.py",3070,0,"",python,selection_command
+42,335998,"models/dynamics.py",3069,1,"",python,content
+43,336111,"models/dynamics.py",3068,1,"",python,content
+44,337261,"models/dynamics.py",3068,0,":",python,content
+45,337262,"models/dynamics.py",3069,0,"",python,selection_keyboard
+46,337444,"models/dynamics.py",3069,0,"-",python,content
+47,337444,"models/dynamics.py",3070,0,"",python,selection_keyboard
+48,337711,"models/dynamics.py",3070,0,"1",python,content
+49,337712,"models/dynamics.py",3071,0,"",python,selection_keyboard
+50,339339,"models/dynamics.py",3066,0,"",python,selection_command
+51,339494,"models/dynamics.py",3065,1,"",python,content
+52,339644,"models/dynamics.py",3064,1,"",python,content
+53,340683,"models/dynamics.py",3064,0,":",python,content
+54,340684,"models/dynamics.py",3065,0,"",python,selection_keyboard
+55,340980,"models/dynamics.py",3065,0,"-",python,content
+56,340981,"models/dynamics.py",3066,0,"",python,selection_keyboard
+57,341127,"models/dynamics.py",3066,0,"1",python,content
+58,341128,"models/dynamics.py",3067,0,"",python,selection_keyboard
+59,341777,"models/dynamics.py",3066,0,"",python,selection_command
+60,343332,"TERMINAL",0,0,"bash",,terminal_focus
+61,343797,"TERMINAL",0,0,"bash",,terminal_focus
+62,352015,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=2\n#SBATCH --ntasks-per-node=4\n#SBATCH --time=48:00:00\n#SBATCH --partition=accelerated\n#SBATCH --cpus-per-task=5\n#SBATCH --gres=gpu:4\n#SBATCH --output=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/%x_%j.log\n#SBATCH --error=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/%x_%j.log\n#SBATCH --job-name=train_dyn_yolorun_new_arch\n\n# Log the sbatch script\ncat $0\n\nmodule unload mpi/openmpi/5.0\nmodule unload devel/cuda/12.4\n# source .venv/bin/activate\n\narray_records_dir=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/data_new/open_ai_minecraft_arrayrecords_chunked\n\njob_name=$SLURM_JOB_NAME\nslurm_job_id=$SLURM_JOB_ID\n\nCHECKPOINT_DIR=$ws_dir/checkpoints/$job_name/$slurm_job_id\nmkdir -p $CHECKPOINT_DIR\n\ntokenizer_ckpt_dir=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/checkpoints/big-runs/tokenizer-lr-scaling/train_tokenizer_lr_sweep_1e-4\n\nenv | grep SLURM\n\nsrun python train_dynamics.py \\n    --save_ckpt \\n    --num_steps=1001 \\n    --warmup_steps=0 \\n    --wsd_decay_steps=0 \\n    --ckpt_dir $CHECKPOINT_DIR \\n    --batch_size=8 \\n    --init_lr=1e-3 \\n    --max_lr=1e-3 \\n    --log_image_interval=100 \\n    --log \\n    --log_checkpoint_interval=100 \\n    --name=dynamics-new-arch-speedrun-$slurm_job_id \\n    --tags dynamics \\n    --entity instant-uv \\n    --project jafar \\n    --tokenizer_checkpoint=$tokenizer_ckpt_dir \\n    --data_dir $array_records_dir \\n    --dyna_dim=128 \\n    --dyna_num_blocks=2 \\n    --dyna_num_heads=4\n ",shellscript,tab
+63,357312,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1295,0,"",shellscript,selection_mouse
+64,360013,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1295,0,"-",shellscript,content
+65,360015,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1296,0,"",shellscript,selection_keyboard
+66,360277,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1296,0,"m",shellscript,content
+67,360278,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1297,0,"",shellscript,selection_keyboard
+68,360361,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1297,0,"e",shellscript,content
+69,360362,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1298,0,"",shellscript,selection_keyboard
+70,360527,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1298,0,"w",shellscript,content
+71,360528,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1299,0,"",shellscript,selection_keyboard
+72,360727,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1299,0,"-",shellscript,content
+73,360728,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1300,0,"",shellscript,selection_keyboard
+74,361210,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1299,1,"",shellscript,content
+75,361360,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1298,1,"",shellscript,content
+76,361479,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1297,1,"",shellscript,content
+77,361594,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1296,1,"",shellscript,content
+78,361794,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1296,0,"n",shellscript,content
+79,361795,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1297,0,"",shellscript,selection_keyboard
+80,361877,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1297,0,"e",shellscript,content
+81,361878,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1298,0,"",shellscript,selection_keyboard
+82,361978,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1298,0,"w",shellscript,content
+83,361979,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",1299,0,"",shellscript,selection_keyboard
+84,365481,"slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",0,0,"",shellscript,tab
+85,370577,"TERMINAL",0,0,"sbatch slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch",,terminal_command
+86,370660,"TERMINAL",0,0,"]633;E;2025-07-18 19:51:59 sbatch slurm/jobs/mihir/horeka/yolo-runs/train_dynamics_new_arch_speedrun.sbatch;9dd50732-d6d7-4d7e-a22c-51ae92e646cb]633;CSubmitted batch job 3358457\r\n]0;tum_cte0515@hkn1991:~/Projects/jafar]633;D;0",,terminal_output
+87,373210,"TERMINAL",0,0,"queue",,terminal_command
+88,373260,"TERMINAL",0,0,"]633;E;2025-07-18 19:52:01 queue;9dd50732-d6d7-4d7e-a22c-51ae92e646cb]633;C[?1049h(B[?7hEvery 1.0s: squeue --mehkn1991.localdomain: Fri Jul 18 19:52:01 2025JOBID PARTITION     NAME     USER ST\tTIME  NODES NODELIST(REASON)3358457 accelerat train_dy tum_cte0 PD\t0:00\t  2 (Priority)\t\t",,terminal_output
+89,374277,"TERMINAL",0,0,"2\t",,terminal_output
+90,375261,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/jafar]633;D;0",,terminal_output

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-73ddfe20-a667-477d-9924-94f7208128f81752186339186-2025_07_11-00.25.58.835/source.csv

@@ -0,0 +1,12 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,6,"slurm/jobs/mihir/horeka/modelsize_scaling/dynamics/1_train_dyn_36M.sbatch",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=2\n#SBATCH --ntasks-per-node=4\n#SBATCH --time=48:00:00\n#SBATCH --partition=accelerated\n#SBATCH --cpus-per-task=5\n#SBATCH --gres=gpu:4\n#SBATCH --output=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/big-runs/dynamics-cotraining-modelsize-scaling/%x_%j.log\n#SBATCH --error=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/big-runs/dynamics-cotraining-modelsize-scaling/%x_%j.log\n#SBATCH --job-name=train_dynamics_modelsize_scaling_36M_2_node\n\n# Log the sbatch script\ncat $0\n\nmodule unload mpi/openmpi/5.0\nmodule unload devel/cuda/12.4\nsource .venv/bin/activate\n\narray_records_dir=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/data_new/open_ai_minecraft_arrayrecords_chunked\n\njob_name=$SLURM_JOB_NAME\nslurm_job_id=$SLURM_JOB_ID\n\nCHECKPOINT_DIR=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/checkpoints/big-runs/dynamics-cotraining-modelsize-scaling/$job_name\nmkdir -p $CHECKPOINT_DIR\n\ntokenizer_ckpt_dir=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/checkpoints/train_tokenizer_batch_size_scaling_16_node/3321526/tokenizer_22000/\n\nenv | grep SLURM\n\nsrun python train_dynamics.py \\n    --save_ckpt \\n    --ckpt_dir $CHECKPOINT_DIR \\n    --batch_size=96 \\n    --min_lr=0 \\n    --max_lr=1.5e-5 \\n    --log_image_interval=1000 \\n    --log \\n    --log_checkpoint_interval=1000 \\n    --name=dynamics-modelsize-scaling-36M-$slurm_job_id \\n    --tags dynamics modelsize-scaling 36M \\n    --entity instant-uv \\n    --project jafar \\n    --tokenizer_checkpoint=$tokenizer_ckpt_dir \\n    --data_dir $array_records_dir \\n",shellscript,tab
+2,766,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"12:25:58 AM [info] Activating crowd-code\n12:25:58 AM [info] Recording started\n12:25:58 AM [info] Initializing git provider using file system watchers...\n12:25:59 AM [info] Git repository found\n12:25:59 AM [info] Git provider initialized successfully\n12:25:59 AM [info] Initial git state: [object Object]\n",Log,tab
+3,4116,"TERMINAL",0,0,"/bin/python3 /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt",,terminal_command
+4,4194,"TERMINAL",0,0,"]633;E;2025-07-11 00:26:02 /bin/python3 /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt;d1c8480d-5bbb-4ee3-b67f-eb04590abc9f]633;C]0;tum_cte0515@hkn1991:/hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash]633;D;0",,terminal_output
+5,76596,"slurm/jobs/mihir/horeka/modelsize_scaling/dynamics/1_train_dyn_36M.sbatch",0,0,"",shellscript,tab
+6,76600,"slurm/jobs/mihir/horeka/modelsize_scaling/dynamics/1_train_dyn_36M.sbatch",1303,0,"",shellscript,selection_mouse
+7,76617,"slurm/jobs/mihir/horeka/modelsize_scaling/dynamics/1_train_dyn_36M.sbatch",1302,0,"",shellscript,selection_command
+8,77837,"slurm/jobs/mihir/horeka/modelsize_scaling/dynamics/1_train_dyn_36M.sbatch",1620,0,"",shellscript,selection_mouse
+9,217736,"TERMINAL",0,0,"bash",,terminal_focus
+10,218636,"slurm/jobs/mihir/horeka/modelsize_scaling/dynamics/1_train_dyn_36M.sbatch",0,0,"",shellscript,tab
+11,226720,"slurm/jobs/mihir/horeka/batchsize_scaling/dynamics_cotraining/sqrt_lr/train_dynamics_2_nodes.sbatch",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=2\n#SBATCH --ntasks-per-node=4\n#SBATCH --time=48:00:00\n#SBATCH --partition=accelerated\n#SBATCH --cpus-per-task=5\n#SBATCH --gres=gpu:4\n#SBATCH --output=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/big-runs/dynamics-cotraining-batchsize-scaling/%x_%j.log\n#SBATCH --error=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/big-runs/dynamics-cotraining-batchsize-scaling/%x_%j.log\n#SBATCH --job-name=train_tokenizer_batch_size_scaling_2_node\n\n# Log the sbatch script\ncat $0\n\nmodule unload mpi/openmpi/5.0\nmodule unload devel/cuda/12.4\nsource .venv/bin/activate\n\narray_records_dir=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/data_new/open_ai_minecraft_arrayrecords_chunked\n\njob_name=$SLURM_JOB_NAME\nslurm_job_id=$SLURM_JOB_ID\n\nCHECKPOINT_DIR=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/checkpoints/big-runs/dynamics-cotraining-batchsize-scaling/$job_name\nmkdir -p $CHECKPOINT_DIR\n\ntokenizer_ckpt_dir=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/checkpoints/train_tokenizer_batch_size_scaling_16_node/3321526/tokenizer_22000/\n\nenv | grep SLURM\n\nsrun python train_dynamics.py \\n    --save_ckpt \\n    --ckpt_dir $CHECKPOINT_DIR \\n    --batch_size=96 \\n    --min_lr=0 \\n    --max_lr=1.5e-5 \\n    --log_image_interval=1000 \\n    --log \\n    --log_checkpoint_interval=1000 \\n    --name=dynamics-batch-size-scaling-2-node-$slurm_job_id \\n    --tags dynamics batch-size-scaling 2-node \\n    --entity instant-uv \\n    --project jafar \\n    --tokenizer_checkpoint=$tokenizer_ckpt_dir \\n    --data_dir $array_records_dir \\n",shellscript,tab

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-7c1bdcf0-d594-4018-8499-7d2ed33930611752094287328-2025_07_09-22.51.39.315/source.csv

@@ -0,0 +1,216 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,5,"models/dynamics.py",0,0,"from typing import Dict, Any\n\nimport jax\nimport jax.numpy as jnp\nimport flax.linen as nn\n\nfrom utils.nn import STTransformer\n\n\nclass DynamicsMaskGIT(nn.Module):\n    """"""MaskGIT dynamics model""""""\n\n    model_dim: int\n    num_latents: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    mask_limit: float\n\n    def setup(self):\n        self.dynamics = STTransformer(\n            self.model_dim,\n            self.num_latents,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n        )\n        self.patch_embed = nn.Embed(self.num_latents, self.model_dim)\n        self.mask_token = self.param(\n            ""mask_token"",\n            nn.initializers.lecun_uniform(),\n            (1, 1, 1, self.model_dim),\n        )\n        self.action_up = nn.Dense(self.model_dim)\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        # --- Mask videos ---\n        vid_embed = self.patch_embed(batch[""video_tokens""])\n        if training:\n            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            noise = jax.random.normal(rng2, self.mask_token.shape) * 1.0  # stddev=1.0, adjust if needed\n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n            \n\n        else:\n            mask = None\n\n        # --- Predict transition ---\n        act_embed = self.action_up(batch[""latent_actions""])\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        logits = self.dynamics(vid_embed)\n        return dict(token_logits=logits, mask=mask)\n",python,tab
+2,410,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:51:39 PM [info] Activating crowd-code\n10:51:39 PM [info] Recording started\n10:51:39 PM [info] Initializing git provider using file system watchers...\n10:51:39 PM [info] Git repository found\n10:51:39 PM [info] Git provider initialized successfully\n",Log,tab
+3,569,"extension-output-pdoom-org.crowd-code-#1-crowd-code",250,0,"10:51:39 PM [info] Initial git state: [object Object]\n",Log,content
+4,3410,"TERMINAL",0,0,"/bin/python3 /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt",,terminal_command
+5,3484,"TERMINAL",0,0,"]633;E;2025-07-09 22:51:42 /bin/python3 /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt;dbf2f7cf-c02e-4ed1-93dc-847ffbf8836e]633;C",,terminal_output
+6,3495,"TERMINAL",0,0,"]0;tum_cte0515@hkn1990:/hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash]633;D;0",,terminal_output
+7,5445,"models/dynamics.py",0,0,"",python,tab
+8,5449,"models/dynamics.py",1436,0,"",python,selection_mouse
+9,5460,"models/dynamics.py",1435,0,"",python,selection_command
+10,7028,"models/dynamics.py",1436,0,"\n            ",python,content
+11,8322,"models/dynamics.py",1437,12,"",python,content
+12,8764,"models/dynamics.py",1043,0,"",python,selection_mouse
+13,10291,"models/dynamics.py",1437,0,"",python,selection_mouse
+14,10947,"models/dynamics.py",1437,0,"\n            rng1, rng2 = jax.random.split(batch[""mask_rng""])",python,content
+15,10975,"models/dynamics.py",1450,0,"",python,selection_command
+16,11582,"models/dynamics.py",1437,0,"",python,selection_command
+17,12054,"models/dynamics.py",1437,1,"",python,content
+18,12066,"models/dynamics.py",1449,0,"",python,selection_command
+19,12469,"models/dynamics.py",1450,0,"",python,selection_command
+20,12594,"models/dynamics.py",1451,0,"",python,selection_command
+21,12742,"models/dynamics.py",1452,0,"",python,selection_command
+22,12880,"models/dynamics.py",1453,0,"",python,selection_command
+23,13017,"models/dynamics.py",1454,0,"",python,selection_command
+24,13410,"models/dynamics.py",1455,0,"",python,selection_command
+25,13862,"models/dynamics.py",1455,5,"",python,content
+26,14581,"models/dynamics.py",1455,0,"r",python,content
+27,14583,"models/dynamics.py",1456,0,"",python,selection_keyboard
+28,14768,"models/dynamics.py",1456,0,"n",python,content
+29,14770,"models/dynamics.py",1457,0,"",python,selection_keyboard
+30,14888,"models/dynamics.py",1457,0,"g",python,content
+31,14890,"models/dynamics.py",1458,0,"",python,selection_keyboard
+32,15046,"models/dynamics.py",1458,0,"_",python,content
+33,15048,"models/dynamics.py",1459,0,"",python,selection_keyboard
+34,15668,"models/dynamics.py",1458,1,"",python,content
+35,15795,"models/dynamics.py",1457,1,"",python,content
+36,15977,"models/dynamics.py",1456,1,"",python,content
+37,16094,"models/dynamics.py",1455,1,"",python,content
+38,16361,"models/dynamics.py",1455,0,"_",python,content
+39,16362,"models/dynamics.py",1456,0,"",python,selection_keyboard
+40,16741,"models/dynamics.py",1456,0,"r",python,content
+41,16742,"models/dynamics.py",1457,0,"",python,selection_keyboard
+42,16881,"models/dynamics.py",1457,0,"n",python,content
+43,16883,"models/dynamics.py",1458,0,"",python,selection_keyboard
+44,17195,"models/dynamics.py",1458,0,"g",python,content
+45,17196,"models/dynamics.py",1459,0,"",python,selection_keyboard
+46,17324,"models/dynamics.py",1459,0," ",python,content
+47,17326,"models/dynamics.py",1460,0,"",python,selection_keyboard
+48,17677,"models/dynamics.py",1459,0,"",python,selection_command
+49,18158,"models/dynamics.py",1460,0,"",python,selection_command
+50,18628,"models/dynamics.py",1497,0,"",python,selection_command
+51,18905,"models/dynamics.py",1496,0,"",python,selection_command
+52,19342,"models/dynamics.py",1494,2,"",python,content
+53,19678,"models/dynamics.py",1486,8,"",python,content
+54,19957,"models/dynamics.py",1484,2,"",python,content
+55,20366,"models/dynamics.py",1479,5,"",python,content
+56,20872,"models/dynamics.py",1479,0,"r",python,content
+57,20873,"models/dynamics.py",1480,0,"",python,selection_keyboard
+58,21049,"models/dynamics.py",1480,0,"n",python,content
+59,21050,"models/dynamics.py",1481,0,"",python,selection_keyboard
+60,21640,"models/dynamics.py",1481,0,"g",python,content
+61,21642,"models/dynamics.py",1482,0,"",python,selection_keyboard
+62,22058,"models/dynamics.py",1482,0,"1",python,content
+63,22059,"models/dynamics.py",1483,0,"",python,selection_keyboard
+64,22539,"models/dynamics.py",1482,0,"",python,selection_command
+65,24952,"models/dynamics.py",1530,0,"",python,selection_command
+66,25648,"models/dynamics.py",1529,0,"",python,selection_command
+67,25800,"models/dynamics.py",1528,0,"",python,selection_command
+68,25938,"models/dynamics.py",1527,0,"",python,selection_command
+69,26098,"models/dynamics.py",1526,0,"",python,selection_command
+70,26219,"models/dynamics.py",1525,0,"",python,selection_command
+71,26363,"models/dynamics.py",1524,0,"",python,selection_command
+72,26519,"models/dynamics.py",1523,0,"",python,selection_command
+73,26856,"models/dynamics.py",1523,4,"",python,content
+74,27627,"models/dynamics.py",1523,0,"_",python,content
+75,27629,"models/dynamics.py",1524,0,"",python,selection_keyboard
+76,27875,"models/dynamics.py",1524,0,"r",python,content
+77,27876,"models/dynamics.py",1525,0,"",python,selection_keyboard
+78,27987,"models/dynamics.py",1525,0,"n",python,content
+79,27990,"models/dynamics.py",1526,0,"",python,selection_keyboard
+80,28146,"models/dynamics.py",1526,0,"g",python,content
+81,28147,"models/dynamics.py",1527,0,"",python,selection_keyboard
+82,28633,"models/dynamics.py",1526,0,"",python,selection_command
+83,28772,"models/dynamics.py",1527,0,"",python,selection_command
+84,29254,"models/dynamics.py",1528,0,"",python,selection_command
+85,29298,"models/dynamics.py",1529,0,"",python,selection_command
+86,29341,"models/dynamics.py",1530,0,"",python,selection_command
+87,29342,"models/dynamics.py",1531,0,"",python,selection_command
+88,29395,"models/dynamics.py",1532,0,"",python,selection_command
+89,29401,"models/dynamics.py",1533,0,"",python,selection_command
+90,29446,"models/dynamics.py",1534,0,"",python,selection_command
+91,29489,"models/dynamics.py",1535,0,"",python,selection_command
+92,29502,"models/dynamics.py",1536,0,"",python,selection_command
+93,29563,"models/dynamics.py",1537,0,"",python,selection_command
+94,29604,"models/dynamics.py",1538,0,"",python,selection_command
+95,29606,"models/dynamics.py",1539,0,"",python,selection_command
+96,29607,"models/dynamics.py",1540,0,"",python,selection_command
+97,29649,"models/dynamics.py",1541,0,"",python,selection_command
+98,29690,"models/dynamics.py",1542,0,"",python,selection_command
+99,29702,"models/dynamics.py",1543,0,"",python,selection_command
+100,29735,"models/dynamics.py",1544,0,"",python,selection_command
+101,29778,"models/dynamics.py",1545,0,"",python,selection_command
+102,29793,"models/dynamics.py",1546,0,"",python,selection_command
+103,29834,"models/dynamics.py",1547,0,"",python,selection_command
+104,29865,"models/dynamics.py",1548,0,"",python,selection_command
+105,29905,"models/dynamics.py",1549,0,"",python,selection_command
+106,29999,"models/dynamics.py",1550,0,"",python,selection_command
+107,30170,"models/dynamics.py",1551,0,"",python,selection_command
+108,30301,"models/dynamics.py",1552,0,"",python,selection_command
+109,30497,"models/dynamics.py",1554,0,"",python,selection_command
+110,32175,"models/dynamics.py",1554,2,"",python,content
+111,33125,"models/dynamics.py",1553,0,"",python,selection_command
+112,33308,"models/dynamics.py",1552,0,"",python,selection_command
+113,33722,"models/dynamics.py",1552,2,"",python,content
+114,34101,"models/dynamics.py",1552,1,"",python,content
+115,34397,"models/dynamics.py",1552,1,"",python,content
+116,35045,"models/dynamics.py",1552,1,"",python,content
+117,35331,"models/dynamics.py",1552,1,"",python,content
+118,35905,"models/dynamics.py",1552,1,"",python,content
+119,36383,"models/dynamics.py",1552,1,"",python,content
+120,36428,"models/dynamics.py",1552,1,"",python,content
+121,36440,"models/dynamics.py",1552,1,"",python,content
+122,36481,"models/dynamics.py",1552,1,"",python,content
+123,36513,"models/dynamics.py",1552,1,"",python,content
+124,36554,"models/dynamics.py",1552,1,"",python,content
+125,36563,"models/dynamics.py",1552,1,"",python,content
+126,36595,"models/dynamics.py",1552,1,"",python,content
+127,36638,"models/dynamics.py",1552,1,"",python,content
+128,36688,"models/dynamics.py",1552,1,"",python,content
+129,36689,"models/dynamics.py",1552,1,"",python,content
+130,36732,"models/dynamics.py",1552,1,"",python,content
+131,36774,"models/dynamics.py",1552,1,"",python,content
+132,36776,"models/dynamics.py",1552,1,"",python,content
+133,36822,"models/dynamics.py",1552,1,"",python,content
+134,36863,"models/dynamics.py",1552,1,"",python,content
+135,36889,"models/dynamics.py",1552,1,"",python,content
+136,36932,"models/dynamics.py",1552,1,"",python,content
+137,36975,"models/dynamics.py",1552,1,"",python,content
+138,36976,"models/dynamics.py",1552,1,"",python,content
+139,36997,"models/dynamics.py",1552,1,"",python,content
+140,37170,"models/dynamics.py",1552,1,"",python,content
+141,37363,"models/dynamics.py",1552,1,"",python,content
+142,37534,"models/dynamics.py",1552,1,"",python,content
+143,37726,"models/dynamics.py",1552,1,"",python,content
+144,37903,"models/dynamics.py",1552,1,"",python,content
+145,38092,"models/dynamics.py",1552,1,"",python,content
+146,38267,"models/dynamics.py",1552,1,"",python,content
+147,38278,"models/dynamics.py",1551,0,"",python,selection_command
+148,42166,"models/dynamics.py",1616,0,"",python,selection_mouse
+149,42779,"models/dynamics.py",1615,0,"",python,selection_mouse
+150,42928,"models/dynamics.py",1614,4,"self",python,selection_mouse
+151,43144,"models/dynamics.py",1614,5,"self.",python,selection_mouse
+152,43144,"models/dynamics.py",1614,15,"self.mask_token",python,selection_mouse
+153,45149,"models/dynamics.py",1614,15,"",python,content
+154,45559,"models/dynamics.py",1614,0,"n",python,content
+155,45560,"models/dynamics.py",1615,0,"",python,selection_keyboard
+156,45762,"models/dynamics.py",1615,0,"o",python,content
+157,45764,"models/dynamics.py",1616,0,"",python,selection_keyboard
+158,45932,"models/dynamics.py",1616,0,"i",python,content
+159,45934,"models/dynamics.py",1617,0,"",python,selection_keyboard
+160,46041,"models/dynamics.py",1617,0,"s",python,content
+161,46042,"models/dynamics.py",1618,0,"",python,selection_keyboard
+162,46230,"models/dynamics.py",1618,0,"e",python,content
+163,46232,"models/dynamics.py",1619,0,"",python,selection_keyboard
+164,46792,"models/dynamics.py",1618,0,"",python,selection_command
+165,51459,"models/dynamics.py",1626,0,"",python,selection_mouse
+166,52068,"models/dynamics.py",1644,0,"",python,selection_mouse
+167,52072,"models/dynamics.py",1643,0,"",python,selection_command
+168,52748,"models/dynamics.py",1644,0,"",python,selection_mouse
+169,52762,"models/dynamics.py",1643,0,"",python,selection_command
+170,53477,"models/dynamics.py",1629,0,"",python,selection_mouse
+171,54060,"models/dynamics.py",1631,0,"",python,selection_mouse
+172,54062,"models/dynamics.py",1630,0,"",python,selection_command
+173,54246,"models/dynamics.py",1630,1,")",python,selection_mouse
+174,54247,"models/dynamics.py",1631,0,"",python,selection_command
+175,54370,"models/dynamics.py",1549,82,"e) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+176,54371,"models/dynamics.py",1534,97,"mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+177,54371,"models/dynamics.py",1472,159,".split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+178,54372,"models/dynamics.py",1418,213,"ith gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+179,54372,"models/dynamics.py",1411,220,"ange, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+180,54372,"models/dynamics.py",1409,222,"change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+181,54412,"models/dynamics.py",1405,226," my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+182,54413,"models/dynamics.py",1402,229,"  # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+183,54427,"models/dynamics.py",1391,240,"\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+184,54548,"models/dynamics.py",1300,331,"            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+185,54599,"models/dynamics.py",1262,369,"            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+186,54691,"models/dynamics.py",1218,413,"            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+187,54729,"models/dynamics.py",1139,492,"            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+188,54753,"models/dynamics.py",1066,565,"            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+189,54808,"models/dynamics.py",1005,626,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+190,54849,"models/dynamics.py",984,647,"        if training:\n            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+191,54979,"models/dynamics.py",924,707,"        vid_embed = self.patch_embed(batch[""video_tokens""])\n        if training:\n            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)",python,selection_mouse
+192,56344,"models/dynamics.py",1009,0,"",python,selection_mouse
+193,56345,"models/dynamics.py",1005,12,"            ",python,selection_mouse
+194,56935,"models/dynamics.py",1005,65,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n    ",python,selection_mouse
+195,56983,"models/dynamics.py",1005,138,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n    ",python,selection_mouse
+196,56984,"models/dynamics.py",1005,218,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n     ",python,selection_mouse
+197,57025,"models/dynamics.py",1005,303,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n        ",python,selection_mouse
+198,57028,"models/dynamics.py",1005,396,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n         ",python,selection_mouse
+199,57065,"models/dynamics.py",1005,491,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n           ",python,selection_mouse
+200,57066,"models/dynamics.py",1005,492,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            ",python,selection_mouse
+201,57070,"models/dynamics.py",1005,497,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise",python,selection_mouse
+202,57088,"models/dynamics.py",1005,569,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed",python,selection_mouse
+203,57228,"models/dynamics.py",1005,560,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            ",python,selection_mouse
+204,57229,"models/dynamics.py",1005,492,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            ",python,selection_mouse
+205,57231,"models/dynamics.py",1005,491,"            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n           ",python,selection_mouse
+206,57634,"models/dynamics.py",1496,0,"",python,selection_mouse
+207,59595,"models/dynamics.py",128,0,"",python,selection_mouse
+208,59744,"models/dynamics.py",127,5,"class",python,selection_mouse
+209,59933,"models/dynamics.py",127,67,"class DynamicsMaskGIT(nn.Module):\n    """"""MaskGIT dynamics model""""""\n",python,selection_mouse
+210,59976,"models/dynamics.py",127,196,"class DynamicsMaskGIT(nn.Module):\n    """"""MaskGIT dynamics model""""""\n\n    model_dim: int\n    num_latents: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    mask_limit: float\n\n    def",python,selection_mouse
+211,59977,"models/dynamics.py",127,627,"class DynamicsMaskGIT(nn.Module):\n    """"""MaskGIT dynamics model""""""\n\n    model_dim: int\n    num_latents: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    mask_limit: float\n\n    def setup(self):\n        self.dynamics = STTransformer(\n            self.model_dim,\n            self.num_latents,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n        )\n        self.patch_embed = nn.Embed(self.num_latents, self.model_dim)\n        self.mask_token = self.param(\n            ""mask_token"",\n            nn.initializers.lecun_uniform(),\n            (1, 1, 1, self.model_dim),\n        )",python,selection_mouse
+212,60095,"models/dynamics.py",127,833,"class DynamicsMaskGIT(nn.Module):\n    """"""MaskGIT dynamics model""""""\n\n    model_dim: int\n    num_latents: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    mask_limit: float\n\n    def setup(self):\n        self.dynamics = STTransformer(\n            self.model_dim,\n            self.num_latents,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n        )\n        self.patch_embed = nn.Embed(self.num_latents, self.model_dim)\n        self.mask_token = self.param(\n            ""mask_token"",\n            nn.initializers.lecun_uniform(),\n            (1, 1, 1, self.model_dim),\n        )\n        self.action_up = nn.Dense(self.model_dim)\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        # --- Mask videos ---\n        vid_embed = self.patch_embed",python,selection_mouse
+213,60096,"models/dynamics.py",127,1263,"class DynamicsMaskGIT(nn.Module):\n    """"""MaskGIT dynamics model""""""\n\n    model_dim: int\n    num_latents: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    mask_limit: float\n\n    def setup(self):\n        self.dynamics = STTransformer(\n            self.model_dim,\n            self.num_latents,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n        )\n        self.patch_embed = nn.Embed(self.num_latents, self.model_dim)\n        self.mask_token = self.param(\n            ""mask_token"",\n            nn.initializers.lecun_uniform(),\n            (1, 1, 1, self.model_dim),\n        )\n        self.action_up = nn.Dense(self.model_dim)\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        # --- Mask videos ---\n        vid_embed = self.patch_embed(batch[""video_tokens""])\n        if training:\n            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)",python,selection_mouse
+214,60097,"models/dynamics.py",127,1557,"class DynamicsMaskGIT(nn.Module):\n    """"""MaskGIT dynamics model""""""\n\n    model_dim: int\n    num_latents: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    mask_limit: float\n\n    def setup(self):\n        self.dynamics = STTransformer(\n            self.model_dim,\n            self.num_latents,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n        )\n        self.patch_embed = nn.Embed(self.num_latents, self.model_dim)\n        self.mask_token = self.param(\n            ""mask_token"",\n            nn.initializers.lecun_uniform(),\n            (1, 1, 1, self.model_dim),\n        )\n        self.action_up = nn.Dense(self.model_dim)\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        # --- Mask videos ---\n        vid_embed = self.patch_embed(batch[""video_tokens""])\n        if training:\n            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)\n            \n\n        else:\n            mask = None\n",python,selection_mouse
+215,60097,"models/dynamics.py",127,1823,"class DynamicsMaskGIT(nn.Module):\n    """"""MaskGIT dynamics model""""""\n\n    model_dim: int\n    num_latents: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    mask_limit: float\n\n    def setup(self):\n        self.dynamics = STTransformer(\n            self.model_dim,\n            self.num_latents,\n            self.num_blocks,\n            self.num_heads,\n            self.dropout,\n        )\n        self.patch_embed = nn.Embed(self.num_latents, self.model_dim)\n        self.mask_token = self.param(\n            ""mask_token"",\n            nn.initializers.lecun_uniform(),\n            (1, 1, 1, self.model_dim),\n        )\n        self.action_up = nn.Dense(self.model_dim)\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        # --- Mask videos ---\n        vid_embed = self.patch_embed(batch[""video_tokens""])\n        if training:\n            rng1, rng2 = jax.random.split(batch[""mask_rng""])\n            mask_prob = jax.random.uniform(rng1, minval=self.mask_limit)\n            mask = jax.random.bernoulli(rng2, mask_prob, vid_embed.shape[:-1])\n            mask = mask.at[:, 0].set(False)\n            # before: with mask token\n            # vid_embed = jnp.where(jnp.expand_dims(mask, -1), self.mask_token, vid_embed)\n\n            # my change, with gaussian noise\n            rng1, _rng = jax.random.split(rng1)\n            noise = jax.random.normal(_rng, self.mask_token.shape) \n            vid_embed = jnp.where(jnp.expand_dims(mask, -1), noise, vid_embed)\n            \n\n        else:\n            mask = None\n\n        # --- Predict transition ---\n        act_embed = self.action_up(batch[""latent_actions""])\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        logits = self.dynamics(vid_embed)\n        return dict(token_logits=logits, mask=mask)\n",python,selection_mouse

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-9cdba2ed-e3b9-400c-aa61-3ca40652e83b1753717763365-2025_07_28-17.49.33.649/source.csv

@@ -0,0 +1,400 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,344,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"5:49:33 PM [info] Activating crowd-code\n5:49:33 PM [info] Recording started\n5:49:33 PM [info] Initializing git provider using file system watchers...\n5:49:33 PM [error] Not a git repository: EntryNotFound (FileSystemError): Error: ENOENT: no such file or directory, stat '/home/hk-project-p0023960/tum_cte0515/Projects/jafar_jobs/.git'\n",Log,tab
+3,2291,"extension-output-pdoom-org.crowd-code-#1-crowd-code",336,0,"5:49:35 PM [info] Retrying git provider initialization...\n5:49:35 PM [error] Not a git repository: EntryNotFound (FileSystemError): Error: ENOENT: no such file or directory, stat '/home/hk-project-p0023960/tum_cte0515/Projects/jafar_jobs/.git'\n",Log,content
+4,13045,"utils/nn.py",0,0,"import math\nfrom typing import Tuple\n\nfrom flax import linen as nn\nimport jax\nimport jax.numpy as jnp\nimport einops\n\n\nclass PositionalEncoding(nn.Module):\n    """"""https://uvadlc-notebooks.readthedocs.io/en/latest/tutorial_notebooks/JAX/tutorial6/Transformers_and_MHAttention.html""""""\n\n    d_model: int  # Hidden dimensionality of the input.\n    max_len: int = 5000  # Maximum length of a sequence to expect.\n\n    def setup(self):\n        # Create matrix of [SeqLen, HiddenDim] representing the positional encoding for max_len inputs\n        self.pe = jnp.zeros((self.max_len, self.d_model))\n        position = jnp.arange(0, self.max_len, dtype=jnp.float32)[:, None]\n        div_term = jnp.exp(\n            jnp.arange(0, self.d_model, 2) * (-math.log(10000.0) / self.d_model)\n        )\n        self.pe = self.pe.at[:, 0::2].set(jnp.sin(position * div_term))\n        self.pe = self.pe.at[:, 1::2].set(jnp.cos(position * div_term))\n\n    def __call__(self, x):\n        x = x + self.pe[: x.shape[2]]\n        return x\n\n\nclass STBlock(nn.Module):\n    dim: int\n    ffn_dim: int\n    num_heads: int\n    dropout: float\n    param_dtype: jnp.dtype\n    dtype: jnp.dtype\n    use_flash_attention: bool\n\n    @nn.remat\n    @nn.compact\n    def __call__(self, x: jax.Array) -> jax.Array:\n        # --- Spatial attention ---\n        z = PositionalEncoding(self.dim)(x)\n        z = nn.LayerNorm(\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )(z)\n        z = nn.MultiHeadAttention(\n            num_heads=self.num_heads,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=False\n            ),\n        )(z)\n        x = x + z\n\n        # --- Temporal attention ---\n        x = x.swapaxes(1, 2)\n        z = PositionalEncoding(self.dim)(x)\n        z = nn.LayerNorm(\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )(z)\n        causal_mask = jnp.tri(z.shape[-2])\n        z = nn.MultiHeadAttention(\n            num_heads=self.num_heads,\n            qkv_features=self.dim,\n            dropout_rate=self.dropout,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            attention_fn=_create_flash_attention_fn(\n                self.use_flash_attention, is_causal=True\n            ),\n            # FIXME (f.srambical): check whether we should still pass the mask if we set is_causal=True\n        )(z, mask=causal_mask)\n        x = x + z\n        x = x.swapaxes(1, 2)\n\n        # --- Feedforward ---\n        z = nn.LayerNorm(\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )(x)\n        z = nn.Dense(\n            self.ffn_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )(z)\n        z = nn.gelu(z)\n        z = nn.Dense(\n            self.dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )(z)\n        x = x + z\n\n        return x\n\n\nclass STTransformer(nn.Module):\n    model_dim: int\n    ffn_dim: int\n    out_dim: int\n    num_blocks: int\n    num_heads: int\n    dropout: float\n    param_dtype: jnp.dtype\n    dtype: jnp.dtype\n    use_flash_attention: bool\n\n    @nn.compact\n    def __call__(self, x: jax.Array) -> jax.Array:\n        x = nn.Sequential(\n            [\n                nn.LayerNorm(\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                ),\n                nn.Dense(\n                    self.model_dim,\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                ),\n                nn.LayerNorm(\n                    param_dtype=self.param_dtype,\n                    dtype=self.dtype,\n                ),\n            ]\n        )(x)\n        for _ in range(self.num_blocks):\n            x = STBlock(\n                dim=self.model_dim,\n                ffn_dim=self.ffn_dim,\n                num_heads=self.num_heads,\n                dropout=self.dropout,\n                param_dtype=self.param_dtype,\n                dtype=self.dtype,\n                use_flash_attention=self.use_flash_attention,\n            )(x)\n        x = nn.Dense(\n            self.out_dim,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n        )(x)\n        return x  # (B, T, E)\n\n\ndef normalize(x):\n    return x / (jnp.linalg.norm(x, ord=2, axis=-1, keepdims=True) + 1e-8)\n\n\nclass VectorQuantizer(nn.Module):\n    latent_dim: int\n    num_latents: int\n    dropout: float\n\n    def setup(self):\n        self.codebook = normalize(\n            self.param(\n                ""codebook"",\n                nn.initializers.lecun_uniform(),\n                (self.num_latents, self.latent_dim),\n            )\n        )\n        self.drop = nn.Dropout(self.dropout, deterministic=False)\n\n    def __call__(\n        self, x: jax.Array, training: bool\n    ) -> Tuple[jax.Array, jax.Array, jax.Array, jax.Array]:\n        # --- Compute distances ---\n        x = normalize(x)\n        codebook = normalize(self.codebook)\n        distance = -jnp.matmul(x, codebook.T)\n        if training:\n            dropout_key = self.make_rng(""dropout"")\n            distance = self.drop(distance, rng=dropout_key)\n\n        # --- Get indices and embeddings ---\n        indices = jnp.argmin(distance, axis=-1)\n        z = self.codebook[indices]\n\n        # --- Straight through estimator ---\n        z_q = x + jax.lax.stop_gradient(z - x)\n        return z_q, z, x, indices\n\n    def get_codes(self, indices: jax.Array):\n        return self.codebook[indices]\n\n\ndef _create_flash_attention_fn(use_flash_attention: bool, is_causal: bool):\n    """"""\n    Create an attention function that uses flash attention if enabled.\n\n    Flax MultiHeadAttention provides tensors with shape (batch..., length, num_heads, head_dim)\n    jax.nn.dot_product_attention expects (batch, length, num_heads, head_dim).\n\n    We need to reshape to ensure compatibility. cuDNN's flash attention additionally\n    requires a sequence length that is a multiple of 4. We pad the sequence length to the nearest\n    multiple of 4 and mask accordingly.\n    """"""\n\n    def attention_fn(query, key, value, bias=None, mask=None, **kwargs):\n        implementation = ""cudnn"" if use_flash_attention else None\n\n        def _rearrange(x):\n            return einops.rearrange(x, ""... l h d -> (...) l h d"")\n\n        def _pad(x):\n            return jnp.pad(x, ((0, 0), (0, pad_size), (0, 0), (0, 0)))\n\n        def _fuse_masks(mask: jax.Array, attention_mask: jax.Array) -> jax.Array:\n            mask_bool = mask.astype(jnp.bool_)\n            expanded_mask = jnp.pad(\n                mask_bool, ((0, pad_size), (0, pad_size)), constant_values=False\n            )\n            return jnp.logical_and(attention_mask, expanded_mask)\n\n        original_shape = query.shape\n        original_seq_len = query.shape[-3]\n\n        # Pad to nearest multiple of 4\n        target_seq_len = ((original_seq_len + 3) // 4) * 4\n        pad_size = target_seq_len - original_seq_len\n\n        query_4d = _pad(_rearrange(query))\n        key_4d = _pad(_rearrange(key))\n        value_4d = _pad(_rearrange(value))\n\n        attention_mask = jnp.ones((target_seq_len, target_seq_len), dtype=jnp.bool_)\n        attention_mask = attention_mask.at[original_seq_len:, :].set(False)\n        attention_mask = attention_mask.at[:, original_seq_len:].set(False)\n\n        mask_4d = (\n            _fuse_masks(mask, attention_mask) if mask is not None else attention_mask\n        )\n        mask_4d = mask_4d[jnp.newaxis, jnp.newaxis, :, :]  # (1, 1, seq_len, seq_len)\n\n        bias_4d = _pad(_rearrange(bias)) if bias is not None else None\n\n        output_4d = jax.nn.dot_product_attention(\n            query=query_4d,\n            key=key_4d,\n            value=value_4d,\n            bias=bias_4d,\n            mask=mask_4d,\n            implementation=implementation,\n            is_causal=is_causal,\n            **kwargs\n        )\n        return output_4d[..., :original_seq_len, :, :].reshape(original_shape)\n\n    return attention_fn\n",python,tab
+5,26550,"TERMINAL",0,0,"",,terminal_focus
+6,28662,"TERMINAL",0,0,"",,terminal_focus
+7,34654,"TERMINAL",0,0,"ime=01:00:00 --partition=accelerated-h100 --nodes=1 --ntasks-per-node=1 --gres=gpu:1 --cpus-per-task=5",,terminal_command
+8,34754,"TERMINAL",0,0,"^C[?2004l\r[?2004h[?2004l\r\r\n]633;E;;e1833c94-a8b3-4524-9e80-61ed159495e5]633;C]0;tum_cte0515@hkn1993:~/Projects/jafar_jobs]633;D",,terminal_output
+9,72762,"sample.py",0,0,"from dataclasses import dataclass\nimport time\nimport os\n\nimport dm_pix as pix\nimport einops\nimport jax\nimport jax.numpy as jnp\nimport flax.linen as nn\nimport numpy as np\nfrom orbax.checkpoint import PyTreeCheckpointer\nfrom PIL import Image, ImageDraw\nimport tyro\n\nfrom genie import Genie\nfrom utils.dataloader import get_dataloader\n\n\n@dataclass\nclass Args:\n    # Experiment\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = ""data/coinrun_episodes""\n    checkpoint: str = """"\n    # Sampling\n    batch_size: int = 1\n    maskgit_steps: int = 25\n    temperature: float = 1.0\n    sample_argmax: bool = True\n    start_frame: int = 0\n    # Tokenizer checkpoint\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    # LAM checkpoint\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    # Dynamics checkpoint\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    param_dtype: jnp.dtype = jnp.float32\n    dtype: jnp.dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n\n\nargs = tyro.cli(Args)\nrng = jax.random.PRNGKey(args.seed)\n\n# --- Load Genie checkpoint ---\ngenie = Genie(\n    # Tokenizer\n    in_dim=args.image_channels,\n    tokenizer_dim=args.tokenizer_dim,\n    tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n    latent_patch_dim=args.latent_patch_dim,\n    num_patch_latents=args.num_patch_latents,\n    patch_size=args.patch_size,\n    tokenizer_num_blocks=args.tokenizer_num_blocks,\n    tokenizer_num_heads=args.tokenizer_num_heads,\n    # LAM\n    lam_dim=args.lam_dim,\n    lam_ffn_dim=args.lam_ffn_dim,\n    latent_action_dim=args.latent_action_dim,\n    num_latent_actions=args.num_latent_actions,\n    lam_patch_size=args.lam_patch_size,\n    lam_num_blocks=args.lam_num_blocks,\n    lam_num_heads=args.lam_num_heads,\n    lam_co_train=False,\n    # Dynamics\n    dyna_dim=args.dyna_dim,\n    dyna_ffn_dim=args.dyna_ffn_dim,\n    dyna_num_blocks=args.dyna_num_blocks,\n    dyna_num_heads=args.dyna_num_heads,\n    param_dtype=args.param_dtype,\n    dtype=args.dtype,\n    use_flash_attention=args.use_flash_attention,\n)\nrng, _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=jnp.float32),\n    mask_rng=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\nckpt = PyTreeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n\n\ndef _sampling_wrapper(module, batch):\n    return module.sample(\n        batch, args.seq_len, args.maskgit_steps, args.temperature, args.sample_argmax\n    )\n\n\n# --- Define autoregressive sampling loop ---\ndef _autoreg_sample(rng, video_batch, action_batch):\n    vid = video_batch[:, : args.start_frame + 1]\n    sampling_fn = jax.jit(nn.apply(_sampling_wrapper, genie))\n    rng, _rng = jax.random.split(rng)\n    batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n    generated_vid = sampling_fn(params, batch)\n    return generated_vid\n\n\n# --- Get video + latent actions ---\narray_record_files = [\n    os.path.join(args.data_dir, x)\n    for x in os.listdir(args.data_dir)\n    if x.endswith("".array_record"")\n]\ndataloader = get_dataloader(\n    array_record_files,\n    args.seq_len,\n    args.batch_size,\n    args.image_height,\n    args.image_width,\n    args.image_channels,\n    num_workers=8,\n    prefetch_buffer_size=1,\n    seed=args.seed,\n)\nvideo_batch = next(iter(dataloader))\n# Get latent actions for all videos in the batch\nbatch = dict(videos=video_batch)\naction_batch = genie.apply(params, batch, False, method=Genie.vq_encode)\naction_batch = action_batch.reshape(video_batch.shape[0], args.seq_len - 1, 1)\n\n# --- Sample + evaluate video ---\nvid = _autoreg_sample(rng, video_batch, action_batch)\ngt = video_batch[:, : vid.shape[1]].clip(0, 1).reshape(-1, *video_batch.shape[2:])\nrecon = vid.clip(0, 1).reshape(-1, *vid.shape[2:])\nssim = pix.ssim(gt[:, args.start_frame + 1 :], recon[:, args.start_frame + 1 :]).mean()\nprint(f""SSIM: {ssim}"")\n\n# --- Construct video ---\ntrue_videos = (video_batch * 255).astype(np.uint8)\npred_videos = (vid * 255).astype(np.uint8)\nvideo_comparison = np.zeros((2, *vid.shape), dtype=np.uint8)\nvideo_comparison[0] = true_videos[:, : args.seq_len]\nvideo_comparison[1] = pred_videos\nframes = einops.rearrange(video_comparison, ""n b t h w c -> t (b h) (n w) c"")\n\n# --- Save video ---\nimgs = [Image.fromarray(img) for img in frames]\n# Write actions on each frame, on each row (i.e., for each video in the batch, on the GT row)\nfor t, img in enumerate(imgs[1:]):\n    d = ImageDraw.Draw(img)\n    for row in range(action_batch.shape[0]):\n        action = action_batch[row, t, 0]\n        y_offset = row * video_batch.shape[2] + 2\n        d.text((2, y_offset), f""{action}"", fill=255)\nimgs[0].save(\n    f""generation_{time.time()}.gif"",\n    save_all=True,\n    append_images=imgs[1:],\n    duration=250,\n    loop=0,\n)\n",python,tab
+10,174652,"sample.py",4736,0,"",python,selection_mouse
+11,175135,"sample.py",4885,0,"",python,selection_mouse
+12,175670,"sample.py",4799,0,"",python,selection_mouse
+13,176257,"sample.py",4887,0,"",python,selection_mouse
+14,176617,"sample.py",5047,0,"",python,selection_mouse
+15,177208,"sample.py",5054,0,"",python,selection_mouse
+16,177668,"sample.py",5136,0,"",python,selection_mouse
+17,178079,"sample.py",5194,0,"",python,selection_mouse
+18,178390,"sample.py",5228,0,"",python,selection_mouse
+19,178395,"sample.py",5227,0,"",python,selection_command
+20,548314,"sample.py",0,0,"",python,tab
+21,612430,"sample.py",0,0,"",python,tab
+22,637465,"sample.py",3924,0,"",python,selection_mouse
+23,638094,"sample.py",3883,0,"",python,selection_mouse
+24,638114,"sample.py",3882,0,"",python,selection_command
+25,654272,"sample.py",3883,0," ",python,content
+26,654279,"sample.py",3883,0,"",python,selection_command
+27,656338,"sample.py",3883,1,"",python,content
+28,656376,"sample.py",3882,0,"",python,selection_command
+29,656753,"sample.py",3883,0,"\n",python,content
+30,657598,"sample.py",3884,0,"/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/checkpoints/maskgit-maskprob-fix/train_dynamics_maskprob_fix_8_node/3371237",python,content
+31,658806,"sample.py",3884,131,"",python,content
+32,659478,"sample.py",3884,0," ",python,content
+33,659502,"sample.py",3884,0,"",python,selection_command
+34,660696,"sample.py",3926,0,"",python,selection_command
+35,662321,"sample.py",3885,0,"",python,selection_mouse
+36,662325,"sample.py",3884,0,"",python,selection_command
+37,678461,"sample.py",3884,0,"video_batch = jnp.array(video_batch)\nprint(video_batch.dtype)\nvideo_batch = video_batch.astype(args.dtype) # / 255.0\nprint(video_batch.dtype)\nvideo_batch = video_batch / 255.0\nprint(video_batch.dtype)",python,content
+38,679700,"sample.py",4083,0,"",python,selection_command
+39,683326,"sample.py",3638,0,"",python,selection_mouse
+40,684400,"utils/dataloader.py",0,0,"import jax\nimport numpy as np\nimport grain\nfrom typing import Any\nimport pickle\n\n\nclass EpisodeLengthFilter(grain.transforms.Filter):\n    """"""\n    A Grain Filter that keeps only episodes with sufficient length.\n    """"""\n\n    def __init__(self, seq_len: int, image_h: int, image_w: int, image_c: int):\n        """"""Initializes the filter with sequence length requirements.""""""\n        self.seq_len = seq_len\n        self.image_h = image_h\n        self.image_w = image_w\n        self.image_c = image_c\n\n    def filter(self, element: Any) -> bool:\n        """"""\n        Filters episodes based on length.\n\n        Args:\n            element: A dictionary representing one record from the DataSource.\n                     Expected to contain 'raw_video' (bytes) and 'sequence_length' (int)\n\n        Returns:\n            True if the episode has sufficient length, False otherwise.\n        """"""\n        assert isinstance(element, bytes)\n        element = pickle.loads(element)\n\n        current_episode_len = element[""sequence_length""]\n        if current_episode_len < self.seq_len:\n            print(\n                f""Filtering out episode with length {current_episode_len}, which is ""\n                f""shorter than the requested sequence length {self.seq_len}.""\n            )\n            return False\n\n        return True\n\n\nclass ProcessEpisodeAndSlice(grain.transforms.RandomMap):\n    """"""\n    A Grain Transformation that combines parsing, slicing, and normalizing.\n    """"""\n\n    def __init__(self, seq_len: int, image_h: int, image_w: int, image_c: int):\n        """"""Initializes the transformation with processing parameters.""""""\n        self.seq_len = seq_len\n        self.image_h = image_h\n        self.image_w = image_w\n        self.image_c = image_c\n\n    def random_map(self, element: dict, rng: np.random.Generator) -> Any:\n        """"""\n        Processes a single raw episode from the data source.\n\n        Args:\n            element: A dictionary representing one record from the DataSource.\n                     Expected to contain 'raw_video' (bytes) and 'sequence_length' (int)\n            rng: A per-record random number generator provided by the Grain sampler.\n\n        Returns:\n            A processed video sequence as a NumPy array with shape\n            (seq_len, height, width, channels) and dtype float32.\n        """"""\n        assert isinstance(element, bytes)\n        element = pickle.loads(element)\n\n        video_shape = (\n            element[""sequence_length""],\n            self.image_h,\n            self.image_w,\n            self.image_c,\n        )\n        episode_tensor = np.frombuffer(element[""raw_video""], dtype=np.uint8)\n        episode_tensor = episode_tensor.reshape(video_shape)\n\n        current_episode_len = episode_tensor.shape[0]\n        if current_episode_len < self.seq_len:\n            raise ValueError(\n                f""Episode length {current_episode_len} is shorter than ""\n                f""requested sequence length {self.seq_len}. This should ""\n                f""have been filtered out.""\n            )\n\n        max_start_idx = current_episode_len - self.seq_len\n\n        start_idx = rng.integers(0, max_start_idx + 1)\n\n        seq = episode_tensor[start_idx : start_idx + self.seq_len]\n\n        return seq\n\n\ndef get_dataloader(\n    array_record_paths: list[str],\n    seq_len: int,\n    global_batch_size: int,\n    image_h: int,\n    image_w: int,\n    image_c: int,\n    num_workers: int = 1,\n    prefetch_buffer_size: int = 1,\n    seed: int = 42,\n):\n    """"""\n    Creates a data loading pipeline using Grain.\n    """"""\n    if not array_record_paths:\n        raise ValueError(""array_record_paths list cannot be empty."")\n\n    num_processes = jax.process_count()\n\n    if global_batch_size % num_processes != 0:\n        raise ValueError(\n            f""Global batch size {global_batch_size} must be divisible by ""\n            f""the number of JAX processes {num_processes} for proper sharding.""\n        )\n    per_process_batch_size = global_batch_size // num_processes\n\n    source = grain.sources.ArrayRecordDataSource(array_record_paths)\n\n    sampler = grain.samplers.IndexSampler(\n        num_records=len(source),\n        shard_options=grain.sharding.ShardByJaxProcess(drop_remainder=True),\n        shuffle=True,\n        num_epochs=None,\n        seed=seed,\n    )\n\n    operations = [\n        EpisodeLengthFilter(\n            seq_len=seq_len, image_h=image_h, image_w=image_w, image_c=image_c\n        ),\n        ProcessEpisodeAndSlice(\n            seq_len=seq_len, image_h=image_h, image_w=image_w, image_c=image_c\n        ),\n        grain.transforms.Batch(batch_size=per_process_batch_size, drop_remainder=True),\n    ]\n\n    read_options = grain.ReadOptions(\n        prefetch_buffer_size=prefetch_buffer_size,\n        num_threads=1,\n    )\n    dataloader = grain.DataLoader(\n        data_source=source,\n        sampler=sampler,\n        operations=operations,\n        worker_count=num_workers,\n        worker_buffer_size=1,\n        read_options=read_options,\n    )\n\n    return dataloader\n",python,tab
+41,687591,"sample.py",0,0,"",python,tab
+42,699155,"sample.py",3992,0,"",python,selection_mouse
+43,699359,"sample.py",3992,2," /",python,selection_mouse
+44,699375,"sample.py",3992,3," / ",python,selection_mouse
+45,699391,"sample.py",3992,5," / 25",python,selection_mouse
+46,699406,"sample.py",3992,7," / 255.",python,selection_mouse
+47,699424,"sample.py",3992,8," / 255.0",python,selection_mouse
+48,700379,"sample.py",3992,8,"",python,content
+49,700396,"sample.py",3991,0,"",python,selection_command
+50,700520,"sample.py",3991,1,"",python,content
+51,700526,"sample.py",3990,0,"",python,selection_command
+52,700698,"sample.py",3990,1,"",python,content
+53,700704,"sample.py",3989,0,"",python,selection_command
+54,831446,"sample.py",2895,0,"",python,selection_mouse
+55,831625,"sample.py",2890,5,"kpt)\n",python,selection_mouse
+56,831636,"sample.py",2886,9,"te(ckpt)\n",python,selection_mouse
+57,831650,"sample.py",2801,94,"pointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+58,831668,"sample.py",2797,98,"heckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+59,831678,"sample.py",2796,99,"Checkpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+60,831703,"sample.py",2794,101,"eeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+61,831711,"sample.py",2792,103,"TreeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+62,831745,"sample.py",2789,106," PyTreeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+63,831746,"sample.py",2788,107,"= PyTreeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+64,831829,"sample.py",2787,108," = PyTreeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+65,831831,"sample.py",2786,109,"t = PyTreeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+66,831831,"sample.py",2785,110,"pt = PyTreeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+67,831831,"sample.py",2784,111,"kpt = PyTreeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+68,831911,"sample.py",2783,112,"ckpt = PyTreeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n",python,selection_mouse
+69,832707,"sample.py",2783,113,"",python,content
+70,833712,"sample.py",2783,0,"\n",python,content
+71,834084,"sample.py",2784,0,"\ndummy_train_state = TrainState.create(\n    apply_fn=genie.apply,\n    params=params,\n    tx=optax.adamw(\n        optax.warmup_cosine_decay_schedule(\n            0, 0, 1, 2 # dummy values\n        )\n    ), \n)\nhandler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\nhandler_registry.add('model_state', ocp.args.StandardRestore, ocp.handlers.StandardCheckpointHandler)\ncheckpoint_manager = ocp.CheckpointManager(\n    args.checkpoint,\n    options=ocp.CheckpointManagerOptions(step_format_fixed_length=6),\n    handler_registry=handler_registry\n)\nabstract_train_state = jax.tree_util.tree_map(\n    ocp.utils.to_shape_dtype_struct, dummy_train_state\n)\n\nrestored = checkpoint_manager.restore(\n    args.checkpoint_step or checkpoint_manager.latest_step(),\n    args=ocp.args.Composite(\n        model_state=ocp.args.StandardRestore(abstract_train_state),\n    ),\n)\nrestored_train_state = restored[""model_state""]\nparams = restored_train_state.params",python,content
+72,835493,"sample.py",3729,0,"",python,selection_command
+73,838304,"sample.py",2784,0,"",python,selection_mouse
+74,849552,"sample.py",250,0,"",python,selection_mouse
+75,849558,"sample.py",249,0,"",python,selection_command
+76,851771,"sample.py",201,0,"",python,selection_command
+77,852002,"sample.py",202,0,"ckpt = PyTreeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n\n",python,content
+78,852026,"sample.py",202,0,"",python,selection_command
+79,853606,"sample.py",202,113,"",python,content
+80,853623,"sample.py",201,0,"",python,selection_command
+81,854162,"sample.py",217,0,"\n",python,content
+82,854375,"sample.py",218,0,"import optax\n",python,content
+83,863874,"sample.py",231,0,"from flax.training.train_state import TrainState\n",python,content
+84,864734,"sample.py",279,1,"",python,content
+85,865699,"sample.py",206,0,"",python,selection_mouse
+86,865744,"sample.py",205,0,"",python,selection_command
+87,866346,"sample.py",170,48,"",python,content
+88,877297,"sample.py",182,0,"",python,selection_mouse
+89,877303,"sample.py",181,0,"",python,selection_command
+90,878057,"sample.py",182,0,"\n",python,content
+91,878214,"sample.py",183,0,"import orbax.checkpoint as ocp\n",python,content
+92,878935,"sample.py",213,1,"",python,content
+93,886469,"sample.py",0,0,"",python,tab
+94,1017820,"train_dynamics.py",0,0,"from dataclasses import dataclass, field\nimport os\n\nimport einops\nfrom flax.training.train_state import TrainState\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\n\nfrom genie import Genie, restore_genie_components\nfrom utils.dataloader import get_dataloader\nfrom utils.lr_utils import get_lr_schedule\nfrom utils.parameter_utils import count_parameters_by_component\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    # Tokenizer\n    tokenizer_dim: int = 512\n    tokenizer_ffn_dim: int = 2048\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 4\n    tokenizer_num_heads: int = 8\n    tokenizer_checkpoint: str = """"\n    # LAM\n    lam_dim: int = 512\n    lam_ffn_dim: int = 2048\n    latent_action_dim: int = 32\n    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 4\n    lam_num_heads: int = 8\n    lam_checkpoint: str = """"\n    # Dynamics\n    dyna_dim: int = 512\n    dyna_ffn_dim: int = 2048\n    dyna_num_blocks: int = 6\n    dyna_num_heads: int = 8\n    dropout: float = 0.0\n    mask_limit: float = 0.5\n    param_dtype: jnp.dtype = jnp.float32\n    dtype: jnp.dtype = jnp.bfloat16\n    use_flash_attention: bool = True\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_dynamics""\n    tags: list[str] = field(default_factory=lambda: [""dynamics""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 25000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n\n\nargs = tyro.cli(Args)\n\n\ndef dynamics_loss_fn(params, state, inputs):\n    """"""Compute masked dynamics loss""""""\n    inputs[""videos""] = inputs[""videos""].astype(args.dtype) / 255.0\n    outputs = state.apply_fn(\n        params,\n        inputs,\n        training=True,\n        rngs={""params"": inputs[""rng""], ""dropout"": inputs[""dropout_rng""]},\n    )\n    mask = outputs[""mask""]\n    outputs[""token_logits""] = outputs[""token_logits""].astype(jnp.float32)\n    ce_loss = optax.softmax_cross_entropy_with_integer_labels(\n        outputs[""token_logits""], outputs[""video_tokens""]\n    )\n    ce_loss = (mask * ce_loss).sum() / mask.sum()\n    acc = outputs[""token_logits""].argmax(-1) == outputs[""video_tokens""]\n    acc = (mask * acc).sum() / mask.sum()\n    select_probs = jax.nn.softmax(outputs[""token_logits""])\n    gt = inputs[""videos""].clip(0, 1).reshape(-1, *inputs[""videos""].shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = pix.psnr(gt, recon).mean()  # type: ignore\n    ssim = pix.ssim(gt, recon).mean()  # type: ignore\n    _, index_counts_lam = jnp.unique_counts(\n        jnp.ravel(outputs[""lam_indices""]), size=args.num_latent_actions, fill_value=0\n    )\n    _, index_counts_tokenizer = jnp.unique_counts(\n        jnp.ravel(outputs[""video_tokens""]), size=args.num_patch_latents, fill_value=0\n    )\n    codebook_usage_lam = (index_counts_lam != 0).mean()\n    codebook_usage_tokenizer = (index_counts_tokenizer != 0).mean()\n    metrics = dict(\n        cross_entropy_loss=ce_loss,\n        masked_token_accuracy=acc,\n        select_logit=outputs[""token_logits""].max(-1).mean(),\n        select_p=select_probs.max(-1).mean(),\n        entropy=jax.scipy.special.entr(select_probs).sum(-1).mean(),\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage_lam=codebook_usage_lam,\n        codebook_usage_tokenizer=codebook_usage_tokenizer,\n    )\n    return ce_loss, (outputs[""recon""], metrics)\n\n\n@jax.jit\ndef train_step(state, inputs):\n    """"""Update state and compute metrics""""""\n    grad_fn = jax.value_and_grad(dynamics_loss_fn, has_aux=True, allow_int=True)\n    (loss, (recon, metrics)), grads = grad_fn(state.params, state, inputs)\n    state = state.apply_gradients(grads=grads)\n    if args.log_gradients:\n        metrics[""gradients_std/""] = jax.tree.map(\n            lambda x: x.std(), grads[""params""][""dynamics""]\n        )\n    return state, loss, recon, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.PRNGKey(args.seed)\n\n    # --- Initialize model ---\n    genie = Genie(\n        # Tokenizer\n        in_dim=args.image_channels,\n        tokenizer_dim=args.tokenizer_dim,\n        tokenizer_ffn_dim=args.tokenizer_ffn_dim,\n        latent_patch_dim=args.latent_patch_dim,\n        num_patch_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        tokenizer_num_blocks=args.tokenizer_num_blocks,\n        tokenizer_num_heads=args.tokenizer_num_heads,\n        # LAM\n        lam_dim=args.lam_dim,\n        lam_ffn_dim=args.lam_ffn_dim,\n        latent_action_dim=args.latent_action_dim,\n        num_latent_actions=args.num_latent_actions,\n        lam_patch_size=args.lam_patch_size,\n        lam_num_blocks=args.lam_num_blocks,\n        lam_num_heads=args.lam_num_heads,\n        lam_co_train=not args.lam_checkpoint,\n        # Dynamics\n        dyna_dim=args.dyna_dim,\n        dyna_ffn_dim=args.dyna_ffn_dim,\n        dyna_num_blocks=args.dyna_num_blocks,\n        dyna_num_heads=args.dyna_num_heads,\n        dropout=args.dropout,\n        mask_limit=args.mask_limit,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n    )\n    rng, _rng = jax.random.split(rng)\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    dummy_inputs = dict(\n        videos=jnp.zeros(\n            (per_device_batch_size_for_init, args.seq_len, *image_shape),\n            dtype=args.dtype,\n        ),\n        action=jnp.zeros(\n            (per_device_batch_size_for_init, args.seq_len), dtype=args.dtype\n        ),\n        mask_rng=_rng,\n    )\n    rng, _rng = jax.random.split(rng)\n    init_params = genie.init(_rng, dummy_inputs)\n\n    param_counts = count_parameters_by_component(init_params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.dtype,\n    )\n    train_state = TrainState.create(apply_fn=genie.apply, params=init_params, tx=tx)\n\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    train_state = jax.device_put(train_state, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    step = 0\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.StandardSave, ocp.handlers.StandardCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.StandardRestore, ocp.handlers.StandardCheckpointHandler\n    )\n    handler_registry.add(""dataloader_state"", grain.checkpoint.CheckpointSave, grain.checkpoint.CheckpointHandler)  # type: ignore\n    handler_registry.add(""dataloader_state"", grain.checkpoint.CheckpointRestore, grain.checkpoint.CheckpointHandler)  # type: ignore\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    # --- Create DataLoaderIterator from dataloader ---\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n    # --- Restore checkpoint ---\n    if args.restore_ckpt:\n        # Restore full dynamics model\n        abstract_train_state = jax.tree_util.tree_map(\n            ocp.utils.to_shape_dtype_struct, train_state\n        )\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.StandardRestore(abstract_train_state),\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),\n            ),\n        )\n        train_state = restored[""model_state""]\n        grain_iterator = restored[""dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    else:\n        # Restore from pre-trained tokenizer (and LAM)\n        train_state = restore_genie_components(\n            train_state, replicated_sharding, grain_iterator, dummy_inputs, rng, args\n        )\n\n    # --- TRAIN LOOP ---\n    dataloader = (jax.make_array_from_process_local_data(videos_sharding, elem) for elem in grain_iterator)  # type: ignore\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng, _rng_dropout, _rng_mask = jax.random.split(rng, 4)\n\n            inputs = dict(\n                videos=videos,\n                rng=_rng,\n                dropout_rng=_rng_dropout,\n                mask_rng=_rng_mask,\n            )\n            train_state, loss, recon, metrics = train_step(train_state, inputs)\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[args.seq_len - 1])),\n                            recon=wandb.Image(np.asarray(recon_seq[args.seq_len - 1])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.StandardSave(train_state),\n                        dataloader_state=grain.checkpoint.CheckpointSave(\n                            grain_iterator\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n",python,tab
+95,1038040,"train_dynamics.py",9442,0,"",python,selection_mouse
+96,1039475,"train_dynamics.py",9578,0,"",python,selection_mouse
+97,1040030,"train_dynamics.py",9599,0,"",python,selection_mouse
+98,1041040,"train_dynamics.py",9561,0,"",python,selection_mouse
+99,1041786,"train_dynamics.py",9598,0,"",python,selection_mouse
+100,1041787,"train_dynamics.py",9597,0,"",python,selection_command
+101,1042107,"train_dynamics.py",9597,1,")",python,selection_mouse
+102,1042112,"train_dynamics.py",9598,0,"",python,selection_command
+103,1042486,"train_dynamics.py",9556,42,"  handler_registry=handler_registry,\n    )",python,selection_mouse
+104,1042521,"train_dynamics.py",9555,43,"   handler_registry=handler_registry,\n    )",python,selection_mouse
+105,1049993,"train_dynamics.py",10588,0,"",python,selection_mouse
+106,1050114,"train_dynamics.py",10574,18,"checkpoint_manager",python,selection_mouse
+107,1051453,"sample.py",0,0,"",python,tab
+108,1055431,"sample.py",3549,0,"",python,selection_mouse
+109,1056015,"sample.py",3540,0,"",python,selection_mouse
+110,1056169,"sample.py",3533,15,"checkpoint_step",python,selection_mouse
+111,1056463,"sample.py",3532,16,".checkpoint_step",python,selection_mouse
+112,1056507,"sample.py",3528,20,"args.checkpoint_step",python,selection_mouse
+113,1057117,"sample.py",3531,0,"",python,selection_mouse
+114,1057118,"sample.py",3528,4,"args",python,selection_mouse
+115,1057359,"sample.py",3528,5,"args.",python,selection_mouse
+116,1057392,"sample.py",3528,20,"args.checkpoint_step",python,selection_mouse
+117,1060630,"sample.py",3528,20,"",python,content
+118,1060952,"sample.py",3528,1,"",python,content
+119,1061195,"sample.py",3528,1,"",python,content
+120,1061241,"sample.py",3528,1,"",python,content
+121,1061419,"sample.py",3528,1,"",python,content
+122,1201521,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",0,0,"\n# Log the sbatch script\ncat $0\n\nmodule unload mpi/openmpi/5.0\nmodule unload devel/cuda/12.4\n# source .venv/bin/activate\n\narray_records_dir=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/data_new/open_ai_minecraft_arrayrecords_chunked\n\njob_name=$SLURM_JOB_NAME\nslurm_job_id=$SLURM_JOB_ID\n\nCHECKPOINT_DIR=$ws_dir/checkpoints/$job_name/$slurm_job_id\nmkdir -p $CHECKPOINT_DIR\n\ntokenizer_ckpt_dir=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/checkpoints/big-runs/tokenizer-lr-scaling/train_tokenizer_lr_sweep_1e-4\ndynamics_ckpt_dir=$1\necho $dynamics_ckpt_dir\n\nenv | grep SLURM\n\nsrun python sample.py \\n    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=10 \\n    --start_frame=0 \\n    --data_dir $array_records_dir\n\n# srun python sample.py \\n    # --checkpoint $dynamics_ckpt_dir \\n    # --start_frame=0 \\n    # --batch_size=12 \\n    # --seq_len=2 \\n    # --data_dir $array_records_dir\n",shellscript,tab
+123,1205330,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",794,0,"",shellscript,selection_mouse
+124,1208070,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",793,1,"",shellscript,content
+125,1208232,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",792,1,"",shellscript,content
+126,1209872,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",792,0,"5",shellscript,content
+127,1209874,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",793,0,"",shellscript,selection_keyboard
+128,1341561,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",792,1,"",shellscript,content
+129,1341666,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",792,0,"1",shellscript,content
+130,1341667,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",793,0,"",shellscript,selection_keyboard
+131,1342833,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",753,0,"",shellscript,selection_mouse
+132,1342965,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",750,4,"4096",shellscript,selection_mouse
+133,1343118,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",749,5,"=4096",shellscript,selection_mouse
+134,1343136,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",711,43,"dyna_num_heads=16 \\n    --dyna_ffn_dim=4096",shellscript,selection_mouse
+135,1343170,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",684,70,"dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096",shellscript,selection_mouse
+136,1343286,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",662,92,"dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096",shellscript,selection_mouse
+137,1343722,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",662,0,"",shellscript,selection_mouse
+138,1343723,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",662,8,"dyna_dim",shellscript,selection_mouse
+139,1343914,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",662,37,"dyna_dim=1024 \\n    --dyna_num_blocks",shellscript,selection_mouse
+140,1343934,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",662,63,"dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads",shellscript,selection_mouse
+141,1343970,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",662,64,"dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=",shellscript,selection_mouse
+142,1344014,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",662,66,"dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16",shellscript,selection_mouse
+143,1344015,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",662,67,"dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 ",shellscript,selection_mouse
+144,1344051,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",662,68,"dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \",shellscript,selection_mouse
+145,1344085,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",662,94,"dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+146,1344432,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",756,0,"",shellscript,selection_mouse
+147,1344677,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",754,2," \",shellscript,selection_mouse
+148,1344714,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",724,32,"s=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+149,1344719,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",722,34,"ads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+150,1344719,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",719,37,"_heads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+151,1344725,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",690,66,"um_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+152,1344768,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",686,70,"na_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+153,1344769,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",685,71,"yna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+154,1344807,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",683,73,"-dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+155,1344864,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",682,74,"--dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+156,1344864,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",681,75," --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+157,1344865,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",658,98,"  --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+158,1344902,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",657,99,"   --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+159,1344949,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",619,137,"   --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \",shellscript,selection_mouse
+160,1345397,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",619,0,"",shellscript,selection_mouse
+161,1345398,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,4,"    ",shellscript,selection_mouse
+162,1345595,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,40,"    --checkpoint $dynamics_ckpt_dir \\n  ",shellscript,selection_mouse
+163,1345611,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,64,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    ",shellscript,selection_mouse
+164,1345629,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,93,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --",shellscript,selection_mouse
+165,1345654,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,131,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim",shellscript,selection_mouse
+166,1345675,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,152,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len",shellscript,selection_mouse
+167,1345688,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,173,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size",shellscript,selection_mouse
+168,1345708,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,175,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1",shellscript,selection_mouse
+169,1345720,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,177,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \",shellscript,selection_mouse
+170,1345740,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,199,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \",shellscript,selection_mouse
+171,1345834,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,233,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+172,1346251,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",851,0,"",shellscript,selection_mouse
+173,1346419,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",834,17,"array_records_dir",shellscript,selection_mouse
+174,1346604,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",817,34,"\n    --data_dir $array_records_dir",shellscript,selection_mouse
+175,1346631,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",795,56,"\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+176,1346663,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",793,58," \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+177,1346701,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",773,78,"\\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+178,1346702,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",770,81,"=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+179,1346733,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",763,88,"seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+180,1346734,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",737,114,"dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+181,1346811,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",710,141,"-dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+182,1346820,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",709,142,"--dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+183,1346857,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",708,143," --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+184,1346890,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",681,170," --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+185,1346923,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",680,171,"  --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+186,1346954,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",658,193,"  --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+187,1346998,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",657,194,"   --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+188,1347077,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",619,232,"   --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+189,1347479,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",619,0,"",shellscript,selection_mouse
+190,1347480,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,4,"    ",shellscript,selection_mouse
+191,1347663,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,42,"    --checkpoint $dynamics_ckpt_dir \\n    ",shellscript,selection_mouse
+192,1347702,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,81,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks",shellscript,selection_mouse
+193,1347702,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,107,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads",shellscript,selection_mouse
+194,1347737,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,131,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim",shellscript,selection_mouse
+195,1347738,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,156,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \",shellscript,selection_mouse
+196,1347771,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,177,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \",shellscript,selection_mouse
+197,1347803,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,199,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \",shellscript,selection_mouse
+198,1347839,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,233,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+199,1348177,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",851,0,"",shellscript,selection_mouse
+200,1348310,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",834,17,"array_records_dir",shellscript,selection_mouse
+201,1348487,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",817,34,"\n    --data_dir $array_records_dir",shellscript,selection_mouse
+202,1348530,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",795,56,"\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+203,1348566,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",774,77,"\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+204,1348567,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",771,80,"2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+205,1348571,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",737,114,"dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+206,1348657,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",736,115,"-dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+207,1348689,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",735,116,"--dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+208,1348725,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",709,142,"--dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+209,1348762,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",708,143," --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+210,1348841,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",680,171,"  --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+211,1348852,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",679,172,"   --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+212,1348870,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",657,194,"   --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+213,1348906,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,233,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+214,1350172,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",619,0,"",shellscript,selection_mouse
+215,1350172,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,4,"    ",shellscript,selection_mouse
+216,1350393,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,40,"    --checkpoint $dynamics_ckpt_dir \\n  ",shellscript,selection_mouse
+217,1350403,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,64,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    ",shellscript,selection_mouse
+218,1350433,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,66,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --",shellscript,selection_mouse
+219,1350438,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,107,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads",shellscript,selection_mouse
+220,1350466,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,131,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim",shellscript,selection_mouse
+221,1350557,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,154,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2",shellscript,selection_mouse
+222,1350557,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,155,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 ",shellscript,selection_mouse
+223,1350557,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,156,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \",shellscript,selection_mouse
+224,1350612,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,177,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \",shellscript,selection_mouse
+225,1350645,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,199,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \",shellscript,selection_mouse
+226,1350827,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,233,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+227,1351509,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",851,0,"",shellscript,selection_mouse
+228,1351729,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",834,17,"array_records_dir",shellscript,selection_mouse
+229,1351973,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",817,34,"\n    --data_dir $array_records_dir",shellscript,selection_mouse
+230,1351990,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",814,37,"0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+231,1352001,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",813,38,"=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+232,1352025,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",781,70,"batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+233,1352110,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",763,88,"seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+234,1352117,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",737,114,"dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+235,1352149,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",711,140,"dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+236,1352228,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",684,167,"dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+237,1352233,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",683,168,"-dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+238,1352265,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",682,169,"--dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+239,1352299,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",659,192," --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+240,1352377,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",658,193,"  --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+241,1352382,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",657,194,"   --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+242,1352421,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",618,233,"    --checkpoint $dynamics_ckpt_dir \\n    --dyna_dim=1024 \\n    --dyna_num_blocks=16 \\n    --dyna_num_heads=16 \\n    --dyna_ffn_dim=4096 \\n    --seq_len=2 \\n    --batch_size=1 \\n    --start_frame=0 \\n    --data_dir $array_records_dir",shellscript,selection_mouse
+243,1511438,"sample.py",0,0,"",python,tab
+244,1512708,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",0,0,"",shellscript,tab
+245,1532243,"sample.py",0,0,"",python,tab
+246,1538054,"sample.py",3729,0,"",python,selection_mouse
+247,1555508,"sample.py",2724,0,"",python,selection_mouse
+248,1557391,"sample.py",2727,0,"",python,selection_mouse
+249,1557517,"sample.py",2723,7,"float32",python,selection_mouse
+250,1557873,"sample.py",2722,8,".float32",python,selection_mouse
+251,1558088,"sample.py",2719,11,"jnp.float32",python,selection_mouse
+252,1558943,"sample.py",2720,0,"",python,selection_mouse
+253,1558944,"sample.py",2719,3,"jnp",python,selection_mouse
+254,1559182,"sample.py",2719,4,"jnp.",python,selection_mouse
+255,1559194,"sample.py",2719,11,"jnp.float32",python,selection_mouse
+256,1560674,"sample.py",2719,11,"a",python,content
+257,1560676,"sample.py",2720,0,"",python,selection_keyboard
+258,1560951,"sample.py",2720,0,"r",python,content
+259,1560952,"sample.py",2721,0,"",python,selection_keyboard
+260,1561136,"sample.py",2721,0,"g",python,content
+261,1561137,"sample.py",2722,0,"",python,selection_keyboard
+262,1561224,"sample.py",2722,0,"s",python,content
+263,1561226,"sample.py",2723,0,"",python,selection_keyboard
+264,1561387,"sample.py",2723,0,".",python,content
+265,1561388,"sample.py",2724,0,"",python,selection_keyboard
+266,1562325,"sample.py",2724,0,"d",python,content
+267,1562326,"sample.py",2725,0,"",python,selection_keyboard
+268,1562870,"sample.py",2725,0,"t",python,content
+269,1562870,"sample.py",2726,0,"",python,selection_keyboard
+270,1563177,"sample.py",2726,0,"y",python,content
+271,1563178,"sample.py",2727,0,"",python,selection_keyboard
+272,1563351,"sample.py",2727,0,"p",python,content
+273,1563351,"sample.py",2728,0,"",python,selection_keyboard
+274,1563459,"sample.py",2728,0,"e",python,content
+275,1563459,"sample.py",2729,0,"",python,selection_keyboard
+276,1564448,"sample.py",2731,0,"",python,selection_mouse
+277,1564753,"sample.py",2729,2,"),",python,selection_mouse
+278,1564760,"sample.py",2731,19,"\n    mask_rng=_rng,",python,selection_mouse
+279,1564880,"sample.py",2660,71,"zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),",python,selection_mouse
+280,1564880,"sample.py",2656,75,"jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),",python,selection_mouse
+281,1564884,"sample.py",2655,76,"=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),",python,selection_mouse
+282,1564900,"sample.py",2649,82,"videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),",python,selection_mouse
+283,1565113,"sample.py",2624,107,"dummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),",python,selection_mouse
+284,1565702,"sample.py",2626,0,"",python,selection_mouse
+285,1565702,"sample.py",2624,12,"dummy_inputs",python,selection_mouse
+286,1565884,"sample.py",2624,23,"dummy_inputs = dict(\n  ",python,selection_mouse
+287,1565916,"sample.py",2624,24,"dummy_inputs = dict(\n   ",python,selection_mouse
+288,1565916,"sample.py",2624,120,"dummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng",python,selection_mouse
+289,1565994,"sample.py",2624,128,"dummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)",python,selection_mouse
+290,1566231,"sample.py",2752,0,"",python,selection_mouse
+291,1566618,"sample.py",2745,7,"_rng,\n)",python,selection_mouse
+292,1566637,"sample.py",2736,16,"mask_rng=_rng,\n)",python,selection_mouse
+293,1566670,"sample.py",2649,103,"videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)",python,selection_mouse
+294,1566765,"sample.py",2624,128,"dummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)",python,selection_mouse
+295,1566806,"sample.py",2551,201,"image_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)",python,selection_mouse
+296,1566903,"sample.py",2517,235,"rng, _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)",python,selection_mouse
+297,1567458,"sample.py",2518,0,"",python,selection_mouse
+298,1567458,"sample.py",2517,3,"rng",python,selection_mouse
+299,1567681,"sample.py",2517,45,"rng, _rng = jax.random.split(rng)\nimage_shape",python,selection_mouse
+300,1567700,"sample.py",2517,119,"rng, _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs",python,selection_mouse
+301,1567720,"sample.py",2517,132,"rng, _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    ",python,selection_mouse
+302,1567752,"sample.py",2517,227,"rng, _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng",python,selection_mouse
+303,1567788,"sample.py",2517,235,"rng, _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)",python,selection_mouse
+304,1567823,"sample.py",2517,246,"rng, _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)\nrng, _rng ",python,selection_mouse
+305,1567856,"sample.py",2517,284,"rng, _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie",python,selection_mouse
+306,1567889,"sample.py",2517,310,"rng, _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n",python,selection_mouse
+307,1568001,"sample.py",2517,311,"rng, _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n\n",python,selection_mouse
+308,1568793,"sample.py",2828,0,"",python,selection_mouse
+309,1569175,"sample.py",2827,1,"\n",python,selection_mouse
+310,1569209,"sample.py",2811,17,", dummy_inputs)\n\n",python,selection_mouse
+311,1569247,"sample.py",2807,21,"_rng, dummy_inputs)\n\n",python,selection_mouse
+312,1569280,"sample.py",2769,59,"random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n\n",python,selection_mouse
+313,1569280,"sample.py",2752,76,"\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n\n",python,selection_mouse
+314,1569319,"sample.py",2745,83,"_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n\n",python,selection_mouse
+315,1569351,"sample.py",2744,84,"=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n\n",python,selection_mouse
+316,1569352,"sample.py",2656,172,"jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n\n",python,selection_mouse
+317,1569352,"sample.py",2655,173,"=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n\n",python,selection_mouse
+318,1569383,"sample.py",2624,204,"dummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n\n",python,selection_mouse
+319,1569525,"sample.py",2551,277,"image_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n\n",python,selection_mouse
+320,1569602,"sample.py",2520,308,", _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n\n",python,selection_mouse
+321,1569603,"sample.py",2517,311,"rng, _rng = jax.random.split(rng)\nimage_shape = (args.image_height, args.image_width, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=args.dtype),\n    mask_rng=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\n\n",python,selection_mouse
+322,1569952,"sample.py",2519,0,"",python,selection_mouse
+323,1587102,"sample.py",2752,0,"",python,selection_mouse
+324,1587947,"sample.py",2752,0,"u",python,content
+325,1587948,"sample.py",2753,0,"",python,selection_keyboard
+326,1588356,"sample.py",2753,0,"u",python,content
+327,1588358,"sample.py",2754,0,"",python,selection_keyboard
+328,1588945,"sample.py",2753,1,"",python,content
+329,1589092,"sample.py",2752,1,"",python,content
+330,1589664,"sample.py",2751,0,"",python,selection_command
+331,1589815,"sample.py",3526,0," args.checkpoint_step or",python,content
+332,1589816,"sample.py",2719,10,"jnp.float32",python,content
+333,1589816,"sample.py",183,31,"",python,content
+334,1589848,"sample.py",181,0,"",python,selection_command
+335,1592180,"sample.py",183,0,"import orbax.checkpoint as ocp\n",python,content
+336,1592185,"sample.py",2719,11,"args.dtype",python,content
+337,1592186,"sample.py",3526,24,"",python,content
+338,1595319,"sample.py",2720,0,"",python,selection_mouse
+339,1596497,"sample.py",3526,0," args.checkpoint_step or",python,content
+340,1597293,"sample.py",2719,10,"jnp.float32",python,content
+341,1597620,"sample.py",183,31,"",python,content
+342,1598176,"sample.py",183,0,"import orbax.checkpoint as ocp\n",python,content
+343,1599549,"sample.py",2719,11,"args.dtype",python,content
+344,1600700,"sample.py",3526,24,"",python,content
+345,1601110,"sample.py",2752,0,"uu",python,content
+346,1601626,"sample.py",2752,2,"",python,content
+347,1602108,"sample.py",2719,10,"jnp.float32",python,content
+348,1999166,"sample.py",3484,0,"",python,selection_mouse
+349,1999740,"sample.py",3269,0,"",python,selection_mouse
+350,2000462,"sample.py",3415,0,"",python,selection_mouse
+351,2001193,"sample.py",3379,0,"",python,selection_mouse
+352,2004722,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",0,0,"",shellscript,tab
+353,2006170,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",700,0,"",shellscript,selection_mouse
+354,2006591,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",794,0,"",shellscript,selection_mouse
+355,2007314,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",744,0,"",shellscript,selection_mouse
+356,2007886,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",772,0,"",shellscript,selection_mouse
+357,2009255,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",743,0,"",shellscript,selection_mouse
+358,2009397,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",737,12,"dyna_ffn_dim",shellscript,selection_mouse
+359,2010286,"sample.py",0,0,"",python,tab
+360,2013566,"sample.py",1062,0,"",python,selection_mouse
+361,2013720,"sample.py",1056,11,"lam_ffn_dim",python,selection_mouse
+362,2024615,"sample.py",2727,0,"",python,selection_mouse
+363,2024769,"sample.py",2723,7,"float32",python,selection_mouse
+364,2025005,"sample.py",2722,8,".float32",python,selection_mouse
+365,2025040,"sample.py",2719,11,"jnp.float32",python,selection_mouse
+366,2028574,"sample.py",2719,11,"a",python,content
+367,2028575,"sample.py",2720,0,"",python,selection_keyboard
+368,2028803,"sample.py",2720,0,"r",python,content
+369,2028807,"sample.py",2721,0,"",python,selection_keyboard
+370,2028904,"sample.py",2721,0,"g",python,content
+371,2028905,"sample.py",2722,0,"",python,selection_keyboard
+372,2029735,"sample.py",2719,3,"args",python,content
+373,2032082,"sample.py",2723,0,".",python,content
+374,2032083,"sample.py",2724,0,"",python,selection_keyboard
+375,2032804,"sample.py",2724,0,"s",python,content
+376,2032805,"sample.py",2725,0,"",python,selection_keyboard
+377,2033140,"sample.py",2725,0,"t",python,content
+378,2033142,"sample.py",2726,0,"",python,selection_keyboard
+379,2033345,"sample.py",2725,1,"",python,content
+380,2033512,"sample.py",2724,1,"",python,content
+381,2033548,"sample.py",2724,0,"d",python,content
+382,2033550,"sample.py",2725,0,"",python,selection_keyboard
+383,2033785,"sample.py",2725,0,"t",python,content
+384,2033785,"sample.py",2726,0,"",python,selection_keyboard
+385,2034041,"sample.py",2726,0,"p",python,content
+386,2034042,"sample.py",2727,0,"",python,selection_keyboard
+387,2034395,"sample.py",2726,1,"",python,content
+388,2034418,"sample.py",2726,0,"y",python,content
+389,2034420,"sample.py",2727,0,"",python,selection_keyboard
+390,2034483,"sample.py",2727,0,"p",python,content
+391,2034484,"sample.py",2728,0,"",python,selection_keyboard
+392,2034619,"sample.py",2728,0,"e",python,content
+393,2034620,"sample.py",2729,0,"",python,selection_keyboard
+394,2153137,"sample.py",2692,0,"",python,selection_mouse
+395,2153277,"sample.py",2689,7,"seq_len",python,selection_mouse
+396,2156276,"slurm/dev/mihir/horeka/yolo-runs/sampling_dev.sh",0,0,"",shellscript,tab
+397,4512967,"utils/nn.py",0,0,"",python,tab
+398,4514214,"utils/nn.py",836,0,"",python,selection_mouse
+399,4514542,"utils/nn.py",927,0,"",python,selection_mouse
+400,4515109,"utils/nn.py",896,0,"",python,selection_mouse

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-bebf29de-c50f-45f7-b90b-66f518a4cf1c1758196766807-2025_09_18-14.00.11.582/source.csv

@@ -0,0 +1,71 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,781,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"2:00:11 PM [info] Activating crowd-code\n2:00:11 PM [info] Recording started\n2:00:11 PM [info] Initializing git provider using file system watchers...\n2:00:11 PM [info] Git repository found\n2:00:11 PM [info] Git provider initialized successfully\n2:00:11 PM [info] Initial git state: [object Object]\n",Log,tab
+3,67473,"TERMINAL",0,0,"bash",,terminal_focus
+4,69307,"TERMINAL",0,0,"queue",,terminal_command
+5,69398,"TERMINAL",0,0,"]633;C[?1049h(B[?7hEvery 1.0s: squeue --mehkn1991.localdomain: Thu Sep 18 14:01:20 2025JOBID PARTITION     NAME     USER ST\tTIME  NODES NODELIST(REASON)3501894 accelerat train_la tum_cte0 PD\t0:00\t  1 (Priority)3501895 accelerat train_la tum_cte0 PD\t0:00\t  1 (Priority)3501896 accelerat train_to tum_cte0 PD\t0:00\t  1 (Priority)3501898 accelerat interact tum_cte0 PD\t0:00\t  1 (Priority)",,terminal_output
+6,70446,"TERMINAL",0,0,"1",,terminal_output
+7,71471,"TERMINAL",0,0,"2",,terminal_output
+8,72014,"TERMINAL",0,0,"bash",,terminal_focus
+9,72519,"TERMINAL",0,0,"3",,terminal_output
+10,73560,"TERMINAL",0,0,"5",,terminal_output
+11,73838,"TERMINAL",0,0,"watch",,terminal_focus
+12,74606,"TERMINAL",0,0,"6",,terminal_output
+13,75684,"TERMINAL",0,0,"7",,terminal_output
+14,76588,"TERMINAL",0,0,"bash",,terminal_focus
+15,76705,"TERMINAL",0,0,"8",,terminal_output
+16,77780,"TERMINAL",0,0,"9",,terminal_output
+17,78769,"TERMINAL",0,0,"scancel 3501898",,terminal_command
+18,78780,"TERMINAL",0,0,"]633;C]0;tum_cte0515@hkn1991:~/Projects/jasmine",,terminal_output
+19,78821,"TERMINAL",0,0,"30",,terminal_output
+20,79916,"TERMINAL",0,0,"\r1",,terminal_output
+21,80049,"TERMINAL",0,0,"watch",,terminal_focus
+22,80596,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/jasmine",,terminal_output
+23,81931,"TERMINAL",0,0,"queue",,terminal_command
+24,81994,"TERMINAL",0,0,"]633;C",,terminal_output
+25,82072,"TERMINAL",0,0,"[?1049h(B[?7hEvery 1.0s: squeue --mehkn1991.localdomain: Thu Sep 18 14:01:33 2025JOBID PARTITION     NAME     USER ST\tTIME  NODES NODELIST(REASON)3501894 accelerat train_la tum_cte0 PD\t0:00\t  1 (Priority)3501895 accelerat train_la tum_cte0 PD\t0:00\t  1 (Priority)3501896 accelerat train_to tum_cte0 PD\t0:00\t  1 (Priority)",,terminal_output
+26,83171,"TERMINAL",0,0,"4",,terminal_output
+27,83321,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/jasmine",,terminal_output
+28,175951,"TERMINAL",0,0,"source .venv/bin/activate",,terminal_command
+29,176011,"TERMINAL",0,0,"]633;C]0;tum_cte0515@hkn1991:~/Projects/jasmine",,terminal_output
+30,177065,"TERMINAL",0,0,"queue",,terminal_command
+31,177170,"TERMINAL",0,0,"]633;C[?1049h(B[?7hEvery 1.0s: squeue --mehkn1991.localdomain: Thu Sep 18 14:03:08 2025JOBID PARTITION     NAME     USER ST\tTIME  NODES NODELIST(REASON)3501894 accelerat train_la tum_cte0 PD\t0:00\t  1 (Priority)3501895 accelerat train_la tum_cte0 PD\t0:00\t  1 (Priority)3501896 accelerat train_to tum_cte0 PD\t0:00\t  1 (Priority)",,terminal_output
+32,178187,"TERMINAL",0,0,"9",,terminal_output
+33,178743,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/jasmine",,terminal_output
+34,179749,"TERMINAL",0,0,"bash",,terminal_focus
+35,297446,"input_pipeline/generate_coinrun_dataset.py",0,0,"""""""\nGenerates a dataset of random-action CoinRun episodes.\nEpisodes are saved individually as memory-mapped files for efficient loading.\n""""""\n\nfrom dataclasses import dataclass\n\nfrom gym3 import types_np\nimport numpy as np\nfrom procgen import ProcgenGym3Env\nimport tyro\nimport json\nimport os\nfrom utils import save_chunks\n\n\n@dataclass\nclass Args:\n    num_episodes_train: int = 10000\n    num_episodes_val: int = 500\n    num_episodes_test: int = 500\n    output_dir: str = ""data/coinrun_episodes""\n    min_episode_length: int = 1000\n    max_episode_length: int = 1000\n    chunk_size: int = 100\n    chunks_per_file: int = 100\n    seed: int = 0\n\n\nargs = tyro.cli(Args)\nassert (\n    args.max_episode_length >= args.min_episode_length\n), ""Maximum episode length must be greater than or equal to minimum episode length.""\n\nif args.min_episode_length < args.chunk_size:\n    print(\n        ""Warning: Minimum episode length is smaller than chunk size. Note that episodes shorter than the chunk size will be discarded.""\n    )\n\n\n# --- Generate episodes ---\ndef generate_episodes(num_episodes, split):\n    episode_idx = 0\n    episode_metadata = []\n    obs_chunks = []\n    act_chunks = []\n    file_idx = 0\n    output_dir_split = os.path.join(args.output_dir, split)\n    while episode_idx < num_episodes:\n        seed = np.random.randint(0, 10000)\n        env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=seed)\n\n        observations_seq = []\n        actions_seq = []\n        episode_obs_chunks = []\n        episode_act_chunks = []\n\n        # --- Run episode ---\n        step_t = 0\n        for step_t in range(args.max_episode_length):\n            action = types_np.sample(env.ac_space, bshape=(env.num,))\n            env.act(action)\n            _, obs, first = env.observe()\n            observations_seq.append(obs[""rgb""])\n            actions_seq.append(action)\n            if len(observations_seq) == args.chunk_size:\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n                observations_seq = []\n                actions_seq = []\n            if first:\n                break\n\n        # --- Save episode ---\n        if step_t + 1 >= args.min_episode_length:\n            if observations_seq:\n                if len(observations_seq) < args.chunk_size:\n                    print(\n                        f""Warning: Inconsistent chunk_sizes. Episode has {len(observations_seq)} frames, ""\n                        f""which is smaller than the requested chunk_size: {args.chunk_size}. ""\n                        ""This might lead to performance degradation during training.""\n                    )\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n\n            obs_chunks_data = [\n                np.concatenate(seq, axis=0).astype(np.uint8)\n                for seq in episode_obs_chunks\n            ]\n            act_chunks_data = [\n                np.concatenate(act, axis=0) for act in episode_act_chunks\n            ]\n            obs_chunks.extend(obs_chunks_data)\n            act_chunks.extend(act_chunks_data)\n\n            ep_metadata, obs_chunks, file_idx, act_chunks = save_chunks(\n                obs_chunks, file_idx, args.chunks_per_file, output_dir_split, act_chunks\n            )\n            episode_metadata.extend(ep_metadata)\n\n            print(f""Episode {episode_idx} completed, length: {step_t + 1}."")\n            episode_idx += 1\n        else:\n            print(f""Episode too short ({step_t + 1}), resampling..."")\n\n    if len(obs_chunks) > 0:\n        print(\n            f""Warning: Dropping {len(obs_chunks)} chunks for consistent number of chunks per file."",\n            ""Consider changing the chunk_size and chunks_per_file parameters to prevent data-loss."",\n        )\n\n    print(f""Done generating {split} split"")\n    return episode_metadata\n\n\ndef get_action_space():\n    env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=0)\n    return env.ac_space.eltype.n\n\n\ndef main():\n    # Set random seed and create dataset directories\n    np.random.seed(args.seed)\n    # --- Generate episodes ---\n    train_episode_metadata = generate_episodes(args.num_episodes_train, ""train"")\n    val_episode_metadata = generate_episodes(args.num_episodes_val, ""val"")\n    test_episode_metadata = generate_episodes(args.num_episodes_test, ""test"")\n\n    # --- Save metadata ---\n    metadata = {\n        ""env"": ""coinrun"",\n        ""num_actions"": get_action_space(),\n        ""num_episodes_train"": args.num_episodes_train,\n        ""num_episodes_val"": args.num_episodes_val,\n        ""num_episodes_test"": args.num_episodes_test,\n        ""avg_episode_len_train"": np.mean(\n            [ep[""avg_seq_len""] for ep in train_episode_metadata]\n        ),\n        ""avg_episode_len_val"": np.mean(\n            [ep[""avg_seq_len""] for ep in val_episode_metadata]\n        ),\n        ""avg_episode_len_test"": np.mean(\n            [ep[""avg_seq_len""] for ep in test_episode_metadata]\n        ),\n        ""episode_metadata_train"": train_episode_metadata,\n        ""episode_metadata_val"": val_episode_metadata,\n        ""episode_metadata_test"": test_episode_metadata,\n    }\n    with open(os.path.join(args.output_dir, ""metadata.json""), ""w"") as f:\n        json.dump(metadata, f)\n\n    print(f""Done generating dataset."")\n\n\nif __name__ == ""__main__"":\n    main()\n",python,tab
+36,301969,"input_pipeline/generate_coinrun_dataset.py",1352,0,"",python,selection_mouse
+37,303278,"input_pipeline/generate_coinrun_dataset.py",1350,0,"",python,selection_mouse
+38,303463,"input_pipeline/generate_coinrun_dataset.py",1343,14,"ProcgenGym3Env",python,selection_mouse
+39,306091,"input_pipeline/generate_coinrun_dataset.py",231,0,"",python,selection_mouse
+40,306321,"input_pipeline/generate_coinrun_dataset.py",227,7,"procgen",python,selection_mouse
+41,306897,"input_pipeline/generate_coinrun_dataset.py",184,0,"",python,selection_mouse
+42,307088,"input_pipeline/generate_coinrun_dataset.py",182,4,"gym3",python,selection_mouse
+43,311458,"input_pipeline/generate_coinrun_dataset.py",197,0,"",python,selection_mouse
+44,311594,"input_pipeline/generate_coinrun_dataset.py",194,8,"types_np",python,selection_mouse
+45,315008,"input_pipeline/generate_coinrun_dataset.py",1654,0,"",python,selection_mouse
+46,315082,"input_pipeline/generate_coinrun_dataset.py",1648,8,"types_np",python,selection_mouse
+47,317287,"input_pipeline/generate_coinrun_dataset.py",184,0,"",python,selection_mouse
+48,317761,".venv/lib/python3.10/site-packages/gym3/__init__.py",0,0,"from gym3 import libenv, testing, types, types_np\nfrom gym3.asynchronous import AsynchronousWrapper\nfrom gym3.concat import ConcatEnv\nfrom gym3.env import Env\nfrom gym3.interactive import Interactive\nfrom gym3.interop import (\n    FromBaselinesVecEnv,\n    FromGymEnv,\n    ToBaselinesVecEnv,\n    ToGymEnv,\n    vectorize_gym,\n)\nfrom gym3.subproc import SubprocEnv, SubprocError\nfrom gym3.trajectory_recorder import TrajectoryRecorderWrapper\nfrom gym3.util import call_func\nfrom gym3.video_recorder import VideoRecorderWrapper\nfrom gym3.viewer import ViewerWrapper\nfrom gym3.wrapper import Wrapper, unwrap\nfrom gym3.extract_dict_ob import ExtractDictObWrapper\n\n__all__ = [\n    ""AsynchronousWrapper"",\n    ""call_func"",\n    ""ConcatEnv"",\n    ""Env"",\n    ""ExtractDictObWrapper"",\n    ""FromBaselinesVecEnv"",\n    ""FromGymEnv"",\n    ""Interactive"",\n    ""libenv"",\n    ""SubprocEnv"",\n    ""SubprocError"",\n    ""testing"",\n    ""ToBaselinesVecEnv"",\n    ""ToGymEnv"",\n    ""TrajectoryRecorderWrapper"",\n    ""types_np"",\n    ""types"",\n    ""unwrap"",\n    ""vectorize_gym"",\n    ""VideoRecorderWrapper"",\n    ""ViewerWrapper"",\n    ""Wrapper"",\n    ""wrappers"",\n]\n",python,tab
+49,320860,".venv/lib/python3.10/site-packages/gym3/__init__.py",713,0,"",python,selection_mouse
+50,320880,".venv/lib/python3.10/site-packages/gym3/__init__.py",712,0,"",python,selection_command
+51,321055,".venv/lib/python3.10/site-packages/gym3/__init__.py",712,1,",",python,selection_mouse
+52,321138,".venv/lib/python3.10/site-packages/gym3/__init__.py",696,16,"\n    ""call_func""",python,selection_mouse
+53,321139,".venv/lib/python3.10/site-packages/gym3/__init__.py",669,43,"\n    ""AsynchronousWrapper"",\n    ""call_func""",python,selection_mouse
+54,321139,".venv/lib/python3.10/site-packages/gym3/__init__.py",657,55,"\n__all__ = [\n    ""AsynchronousWrapper"",\n    ""call_func""",python,selection_mouse
+55,321140,".venv/lib/python3.10/site-packages/gym3/__init__.py",642,70,"tDictObWrapper\n\n__all__ = [\n    ""AsynchronousWrapper"",\n    ""call_func""",python,selection_mouse
+56,321140,".venv/lib/python3.10/site-packages/gym3/__init__.py",713,0,"",python,selection_command
+57,321184,".venv/lib/python3.10/site-packages/gym3/__init__.py",601,112,"p\nfrom gym3.extract_dict_ob import ExtractDictObWrapper\n\n__all__ = [\n    ""AsynchronousWrapper"",\n    ""call_func"",",python,selection_mouse
+58,321230,".venv/lib/python3.10/site-packages/gym3/__init__.py",561,152,"\nfrom gym3.wrapper import Wrapper, unwrap\nfrom gym3.extract_dict_ob import ExtractDictObWrapper\n\n__all__ = [\n    ""AsynchronousWrapper"",\n    ""call_func"",",python,selection_mouse
+59,321746,".venv/lib/python3.10/site-packages/gym3/__init__.py",561,0,"",python,selection_mouse
+60,321747,".venv/lib/python3.10/site-packages/gym3/__init__.py",560,0,"",python,selection_command
+61,322629,".venv/lib/python3.10/site-packages/gym3/__init__.py",657,0,"",python,selection_mouse
+62,323244,".venv/lib/python3.10/site-packages/gym3/__init__.py",636,0,"",python,selection_mouse
+63,329407,"input_pipeline/generate_coinrun_dataset.py",0,0,"",python,tab
+64,333462,"input_pipeline/generate_coinrun_dataset copy.py",0,0,"""""""\nGenerates a dataset of random-action CoinRun episodes.\nEpisodes are saved individually as memory-mapped files for efficient loading.\n""""""\n\nfrom dataclasses import dataclass\n\nfrom gym3 import types_np\nimport numpy as np\nfrom procgen import ProcgenGym3Env\nimport tyro\nimport json\nimport os\nfrom utils import save_chunks\n\n\n@dataclass\nclass Args:\n    num_episodes_train: int = 10000\n    num_episodes_val: int = 500\n    num_episodes_test: int = 500\n    output_dir: str = ""data/coinrun_episodes""\n    min_episode_length: int = 1000\n    max_episode_length: int = 1000\n    chunk_size: int = 100\n    chunks_per_file: int = 100\n    seed: int = 0\n\n\nargs = tyro.cli(Args)\nassert (\n    args.max_episode_length >= args.min_episode_length\n), ""Maximum episode length must be greater than or equal to minimum episode length.""\n\nif args.min_episode_length < args.chunk_size:\n    print(\n        ""Warning: Minimum episode length is smaller than chunk size. Note that episodes shorter than the chunk size will be discarded.""\n    )\n\n\n# --- Generate episodes ---\ndef generate_episodes(num_episodes, split):\n    episode_idx = 0\n    episode_metadata = []\n    obs_chunks = []\n    act_chunks = []\n    file_idx = 0\n    output_dir_split = os.path.join(args.output_dir, split)\n    while episode_idx < num_episodes:\n        seed = np.random.randint(0, 10000)\n        env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=seed)\n\n        observations_seq = []\n        actions_seq = []\n        episode_obs_chunks = []\n        episode_act_chunks = []\n\n        # --- Run episode ---\n        step_t = 0\n        for step_t in range(args.max_episode_length):\n            action = types_np.sample(env.ac_space, bshape=(env.num,))\n            env.act(action)\n            _, obs, first = env.observe()\n            observations_seq.append(obs[""rgb""])\n            actions_seq.append(action)\n            if len(observations_seq) == args.chunk_size:\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n                observations_seq = []\n                actions_seq = []\n            if first:\n                break\n\n        # --- Save episode ---\n        if step_t + 1 >= args.min_episode_length:\n            if observations_seq:\n                if len(observations_seq) < args.chunk_size:\n                    print(\n                        f""Warning: Inconsistent chunk_sizes. Episode has {len(observations_seq)} frames, ""\n                        f""which is smaller than the requested chunk_size: {args.chunk_size}. ""\n                        ""This might lead to performance degradation during training.""\n                    )\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n\n            obs_chunks_data = [\n                np.concatenate(seq, axis=0).astype(np.uint8)\n                for seq in episode_obs_chunks\n            ]\n            act_chunks_data = [\n                np.concatenate(act, axis=0) for act in episode_act_chunks\n            ]\n            obs_chunks.extend(obs_chunks_data)\n            act_chunks.extend(act_chunks_data)\n\n            ep_metadata, obs_chunks, file_idx, act_chunks = save_chunks(\n                obs_chunks, file_idx, args.chunks_per_file, output_dir_split, act_chunks\n            )\n            episode_metadata.extend(ep_metadata)\n\n            print(f""Episode {episode_idx} completed, length: {step_t + 1}."")\n            episode_idx += 1\n        else:\n            print(f""Episode too short ({step_t + 1}), resampling..."")\n\n    if len(obs_chunks) > 0:\n        print(\n            f""Warning: Dropping {len(obs_chunks)} chunks for consistent number of chunks per file."",\n            ""Consider changing the chunk_size and chunks_per_file parameters to prevent data-loss."",\n        )\n\n    print(f""Done generating {split} split"")\n    return episode_metadata\n\n\ndef get_action_space():\n    env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=0)\n    return env.ac_space.eltype.n\n\n\ndef main():\n    # Set random seed and create dataset directories\n    np.random.seed(args.seed)\n    # --- Generate episodes ---\n    train_episode_metadata = generate_episodes(args.num_episodes_train, ""train"")\n    val_episode_metadata = generate_episodes(args.num_episodes_val, ""val"")\n    test_episode_metadata = generate_episodes(args.num_episodes_test, ""test"")\n\n    # --- Save metadata ---\n    metadata = {\n        ""env"": ""coinrun"",\n        ""num_actions"": get_action_space(),\n        ""num_episodes_train"": args.num_episodes_train,\n        ""num_episodes_val"": args.num_episodes_val,\n        ""num_episodes_test"": args.num_episodes_test,\n        ""avg_episode_len_train"": np.mean(\n            [ep[""avg_seq_len""] for ep in train_episode_metadata]\n        ),\n        ""avg_episode_len_val"": np.mean(\n            [ep[""avg_seq_len""] for ep in val_episode_metadata]\n        ),\n        ""avg_episode_len_test"": np.mean(\n            [ep[""avg_seq_len""] for ep in test_episode_metadata]\n        ),\n        ""episode_metadata_train"": train_episode_metadata,\n        ""episode_metadata_val"": val_episode_metadata,\n        ""episode_metadata_test"": test_episode_metadata,\n    }\n    with open(os.path.join(args.output_dir, ""metadata.json""), ""w"") as f:\n        json.dump(metadata, f)\n\n    print(f""Done generating dataset."")\n\n\nif __name__ == ""__main__"":\n    main()\n",python,tab
+65,350130,"input_pipeline/generate_breakout_dataset.py",0,0,"""""""\nGenerates a dataset of random-action CoinRun episodes.\nEpisodes are saved individually as memory-mapped files for efficient loading.\n""""""\n\nfrom dataclasses import dataclass\n\nfrom gym3 import types_np\nimport numpy as np\nfrom procgen import ProcgenGym3Env\nimport tyro\nimport json\nimport os\nfrom utils import save_chunks\n\n\n@dataclass\nclass Args:\n    num_episodes_train: int = 10000\n    num_episodes_val: int = 500\n    num_episodes_test: int = 500\n    output_dir: str = ""data/coinrun_episodes""\n    min_episode_length: int = 1000\n    max_episode_length: int = 1000\n    chunk_size: int = 100\n    chunks_per_file: int = 100\n    seed: int = 0\n\n\nargs = tyro.cli(Args)\nassert (\n    args.max_episode_length >= args.min_episode_length\n), ""Maximum episode length must be greater than or equal to minimum episode length.""\n\nif args.min_episode_length < args.chunk_size:\n    print(\n        ""Warning: Minimum episode length is smaller than chunk size. Note that episodes shorter than the chunk size will be discarded.""\n    )\n\n\n# --- Generate episodes ---\ndef generate_episodes(num_episodes, split):\n    episode_idx = 0\n    episode_metadata = []\n    obs_chunks = []\n    act_chunks = []\n    file_idx = 0\n    output_dir_split = os.path.join(args.output_dir, split)\n    while episode_idx < num_episodes:\n        seed = np.random.randint(0, 10000)\n        env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=seed)\n\n        observations_seq = []\n        actions_seq = []\n        episode_obs_chunks = []\n        episode_act_chunks = []\n\n        # --- Run episode ---\n        step_t = 0\n        for step_t in range(args.max_episode_length):\n            action = types_np.sample(env.ac_space, bshape=(env.num,))\n            env.act(action)\n            _, obs, first = env.observe()\n            observations_seq.append(obs[""rgb""])\n            actions_seq.append(action)\n            if len(observations_seq) == args.chunk_size:\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n                observations_seq = []\n                actions_seq = []\n            if first:\n                break\n\n        # --- Save episode ---\n        if step_t + 1 >= args.min_episode_length:\n            if observations_seq:\n                if len(observations_seq) < args.chunk_size:\n                    print(\n                        f""Warning: Inconsistent chunk_sizes. Episode has {len(observations_seq)} frames, ""\n                        f""which is smaller than the requested chunk_size: {args.chunk_size}. ""\n                        ""This might lead to performance degradation during training.""\n                    )\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n\n            obs_chunks_data = [\n                np.concatenate(seq, axis=0).astype(np.uint8)\n                for seq in episode_obs_chunks\n            ]\n            act_chunks_data = [\n                np.concatenate(act, axis=0) for act in episode_act_chunks\n            ]\n            obs_chunks.extend(obs_chunks_data)\n            act_chunks.extend(act_chunks_data)\n\n            ep_metadata, obs_chunks, file_idx, act_chunks = save_chunks(\n                obs_chunks, file_idx, args.chunks_per_file, output_dir_split, act_chunks\n            )\n            episode_metadata.extend(ep_metadata)\n\n            print(f""Episode {episode_idx} completed, length: {step_t + 1}."")\n            episode_idx += 1\n        else:\n            print(f""Episode too short ({step_t + 1}), resampling..."")\n\n    if len(obs_chunks) > 0:\n        print(\n            f""Warning: Dropping {len(obs_chunks)} chunks for consistent number of chunks per file."",\n            ""Consider changing the chunk_size and chunks_per_file parameters to prevent data-loss."",\n        )\n\n    print(f""Done generating {split} split"")\n    return episode_metadata\n\n\ndef get_action_space():\n    env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=0)\n    return env.ac_space.eltype.n\n\n\ndef main():\n    # Set random seed and create dataset directories\n    np.random.seed(args.seed)\n    # --- Generate episodes ---\n    train_episode_metadata = generate_episodes(args.num_episodes_train, ""train"")\n    val_episode_metadata = generate_episodes(args.num_episodes_val, ""val"")\n    test_episode_metadata = generate_episodes(args.num_episodes_test, ""test"")\n\n    # --- Save metadata ---\n    metadata = {\n        ""env"": ""coinrun"",\n        ""num_actions"": get_action_space(),\n        ""num_episodes_train"": args.num_episodes_train,\n        ""num_episodes_val"": args.num_episodes_val,\n        ""num_episodes_test"": args.num_episodes_test,\n        ""avg_episode_len_train"": np.mean(\n            [ep[""avg_seq_len""] for ep in train_episode_metadata]\n        ),\n        ""avg_episode_len_val"": np.mean(\n            [ep[""avg_seq_len""] for ep in val_episode_metadata]\n        ),\n        ""avg_episode_len_test"": np.mean(\n            [ep[""avg_seq_len""] for ep in test_episode_metadata]\n        ),\n        ""episode_metadata_train"": train_episode_metadata,\n        ""episode_metadata_val"": val_episode_metadata,\n        ""episode_metadata_test"": test_episode_metadata,\n    }\n    with open(os.path.join(args.output_dir, ""metadata.json""), ""w"") as f:\n        json.dump(metadata, f)\n\n    print(f""Done generating dataset."")\n\n\nif __name__ == ""__main__"":\n    main()\n",python,tab
+66,489770,"input_pipeline/generate_breakout_dataset.py",1485,0,"",python,selection_mouse
+67,551302,"input_pipeline/generate_coinrun_dataset.py",0,0,"",python,tab
+68,551809,"input_pipeline/generate_coinrun_dataset.py",1060,0,"",python,selection_mouse
+69,552660,"input_pipeline/generate_coinrun_dataset.py",0,0,"",python,selection_command
+70,553135,"input_pipeline/generate_coinrun_dataset.py",0,3,"""""""",python,selection_command
+71,553385,"input_pipeline/generate_coinrun_dataset.py",0,5351,"""""""\nGenerates a dataset of random-action CoinRun episodes.\nEpisodes are saved individually as memory-mapped files for efficient loading.\n""""""\n\nfrom dataclasses import dataclass\n\nfrom gym3 import types_np\nimport numpy as np\nfrom procgen import ProcgenGym3Env\nimport tyro\nimport json\nimport os\nfrom utils import save_chunks\n\n\n@dataclass\nclass Args:\n    num_episodes_train: int = 10000\n    num_episodes_val: int = 500\n    num_episodes_test: int = 500\n    output_dir: str = ""data/coinrun_episodes""\n    min_episode_length: int = 1000\n    max_episode_length: int = 1000\n    chunk_size: int = 100\n    chunks_per_file: int = 100\n    seed: int = 0\n\n\nargs = tyro.cli(Args)\nassert (\n    args.max_episode_length >= args.min_episode_length\n), ""Maximum episode length must be greater than or equal to minimum episode length.""\n\nif args.min_episode_length < args.chunk_size:\n    print(\n        ""Warning: Minimum episode length is smaller than chunk size. Note that episodes shorter than the chunk size will be discarded.""\n    )\n\n\n# --- Generate episodes ---\ndef generate_episodes(num_episodes, split):\n    episode_idx = 0\n    episode_metadata = []\n    obs_chunks = []\n    act_chunks = []\n    file_idx = 0\n    output_dir_split = os.path.join(args.output_dir, split)\n    while episode_idx < num_episodes:\n        seed = np.random.randint(0, 10000)\n        env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=seed)\n\n        observations_seq = []\n        actions_seq = []\n        episode_obs_chunks = []\n        episode_act_chunks = []\n\n        # --- Run episode ---\n        step_t = 0\n        for step_t in range(args.max_episode_length):\n            action = types_np.sample(env.ac_space, bshape=(env.num,))\n            env.act(action)\n            _, obs, first = env.observe()\n            observations_seq.append(obs[""rgb""])\n            actions_seq.append(action)\n            if len(observations_seq) == args.chunk_size:\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n                observations_seq = []\n                actions_seq = []\n            if first:\n                break\n\n        # --- Save episode ---\n        if step_t + 1 >= args.min_episode_length:\n            if observations_seq:\n                if len(observations_seq) < args.chunk_size:\n                    print(\n                        f""Warning: Inconsistent chunk_sizes. Episode has {len(observations_seq)} frames, ""\n                        f""which is smaller than the requested chunk_size: {args.chunk_size}. ""\n                        ""This might lead to performance degradation during training.""\n                    )\n                episode_obs_chunks.append(observations_seq)\n                episode_act_chunks.append(actions_seq)\n\n            obs_chunks_data = [\n                np.concatenate(seq, axis=0).astype(np.uint8)\n                for seq in episode_obs_chunks\n            ]\n            act_chunks_data = [\n                np.concatenate(act, axis=0) for act in episode_act_chunks\n            ]\n            obs_chunks.extend(obs_chunks_data)\n            act_chunks.extend(act_chunks_data)\n\n            ep_metadata, obs_chunks, file_idx, act_chunks = save_chunks(\n                obs_chunks, file_idx, args.chunks_per_file, output_dir_split, act_chunks\n            )\n            episode_metadata.extend(ep_metadata)\n\n            print(f""Episode {episode_idx} completed, length: {step_t + 1}."")\n            episode_idx += 1\n        else:\n            print(f""Episode too short ({step_t + 1}), resampling..."")\n\n    if len(obs_chunks) > 0:\n        print(\n            f""Warning: Dropping {len(obs_chunks)} chunks for consistent number of chunks per file."",\n            ""Consider changing the chunk_size and chunks_per_file parameters to prevent data-loss."",\n        )\n\n    print(f""Done generating {split} split"")\n    return episode_metadata\n\n\ndef get_action_space():\n    env = ProcgenGym3Env(num=1, env_name=""coinrun"", start_level=0)\n    return env.ac_space.eltype.n\n\n\ndef main():\n    # Set random seed and create dataset directories\n    np.random.seed(args.seed)\n    # --- Generate episodes ---\n    train_episode_metadata = generate_episodes(args.num_episodes_train, ""train"")\n    val_episode_metadata = generate_episodes(args.num_episodes_val, ""val"")\n    test_episode_metadata = generate_episodes(args.num_episodes_test, ""test"")\n\n    # --- Save metadata ---\n    metadata = {\n        ""env"": ""coinrun"",\n        ""num_actions"": get_action_space(),\n        ""num_episodes_train"": args.num_episodes_train,\n        ""num_episodes_val"": args.num_episodes_val,\n        ""num_episodes_test"": args.num_episodes_test,\n        ""avg_episode_len_train"": np.mean(\n            [ep[""avg_seq_len""] for ep in train_episode_metadata]\n        ),\n        ""avg_episode_len_val"": np.mean(\n            [ep[""avg_seq_len""] for ep in val_episode_metadata]\n        ),\n        ""avg_episode_len_test"": np.mean(\n            [ep[""avg_seq_len""] for ep in test_episode_metadata]\n        ),\n        ""episode_metadata_train"": train_episode_metadata,\n        ""episode_metadata_val"": val_episode_metadata,\n        ""episode_metadata_test"": test_episode_metadata,\n    }\n    with open(os.path.join(args.output_dir, ""metadata.json""), ""w"") as f:\n        json.dump(metadata, f)\n\n    print(f""Done generating dataset."")\n\n\nif __name__ == ""__main__"":\n    main()\n",python,selection_command

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-d4ecca31-879c-4879-b2a7-b7463e4327b91757416440874-2025_09_09-13.15.15.617/source.csv

@@ -0,0 +1,7 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,1249,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"1:15:15 PM [info] Activating crowd-code\n1:15:15 PM [info] Recording started\n1:15:15 PM [info] Initializing git provider using file system watchers...\n1:15:16 PM [info] Git repository found\n1:15:16 PM [info] Git provider initialized successfully\n1:15:16 PM [info] Initial git state: [object Object]\n",Log,tab
+3,436098,"train_lam.py",0,0,"from dataclasses import dataclass, field\nimport os\nfrom typing import cast\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.lam import LatentActionModel\nfrom utils.dataloader import get_dataloader\nfrom utils.lr_utils import get_lr_schedule\nfrom utils.parameter_utils import count_parameters_by_component\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    vq_beta: float = 0.25\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    vq_reset_thresh: int = 50\n    # LAM\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 6\n    patch_size: int = 16\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.0\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_lam""\n    tags: list[str] = field(default_factory=lambda: [""lam""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_checkpoint_keep_period: int = 20000\n    wandb_id: str = """"\n    use_flash_attention: bool = True\n\n\nargs = tyro.cli(Args)\n\n\ndef lam_loss_fn(\n    model: LatentActionModel, inputs: dict\n) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n    # --- Compute loss ---\n    gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n    inputs[""videos""] = gt.astype(args.dtype)\n    model.train()\n    outputs = model(inputs, training=True)\n    outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n    gt_future_frames = gt[:, 1:]\n    mse = jnp.square(gt_future_frames - outputs[""recon""]).mean()\n    q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n    commitment_loss = jnp.square(\n        outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n    ).mean()\n    loss = mse + q_loss + args.vq_beta * commitment_loss\n\n    # --- Compute validation metrics ---\n    gt = gt_future_frames.clip(0, 1).reshape(-1, *gt_future_frames.shape[2:])\n    recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n    psnr = jnp.asarray(pix.psnr(gt, recon)).mean()\n    ssim = jnp.asarray(pix.ssim(gt, recon)).mean()\n    count_fn = jax.vmap(lambda i: (outputs[""indices""] == i).sum())\n    index_counts = count_fn(jnp.arange(args.num_latents))\n    metrics = dict(\n        loss=loss,\n        mse=mse,\n        q_loss=q_loss,\n        commitment_loss=commitment_loss,\n        psnr=psnr,\n        ssim=ssim,\n        codebook_usage=(index_counts != 0).mean(),\n    )\n    return loss, (outputs[""recon""], index_counts, metrics)\n\n\n@nnx.jit\ndef train_step(\n    lam: LatentActionModel,\n    optimizer: nnx.Optimizer,\n    inputs: dict,\n    action_last_active: jax.Array,\n    rng: jax.Array,\n) -> tuple[jax.Array, jax.Array, jax.Array, dict]:\n    def loss_fn(\n        model: LatentActionModel,\n    ) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n        return lam_loss_fn(model, inputs)\n\n    # --- Update model ---\n    (loss, (recon, idx_counts, metrics)), grads = nnx.value_and_grad(\n        loss_fn, has_aux=True\n    )(lam)\n    optimizer.update(grads)\n\n    # --- Reset inactive latent actions ---\n    codebook = lam.vq.codebook\n    num_codes = len(codebook)\n    active_codes = idx_counts != 0.0\n    action_last_active = jnp.where(active_codes, 0, action_last_active + 1)\n    p_code = active_codes / active_codes.sum()\n    reset_idxs = jax.random.choice(rng, num_codes, shape=(num_codes,), p=p_code)\n    do_reset = action_last_active >= args.vq_reset_thresh\n    new_codebook = jnp.where(\n        jnp.expand_dims(do_reset, -1), codebook[reset_idxs], codebook.value\n    )\n    lam.vq.codebook.value = new_codebook\n    action_last_active = jnp.where(do_reset, 0, action_last_active)\n    return loss, recon, action_last_active, metrics\n\n\nif __name__ == ""__main__"":\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    per_device_batch_size_for_init = args.batch_size // num_devices\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    lam = LatentActionModel(\n        in_dim=args.image_channels,\n        model_dim=args.model_dim,\n        ffn_dim=args.ffn_dim,\n        latent_dim=args.latent_dim,\n        num_latents=args.num_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.num_blocks,\n        num_heads=args.num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.codebook_dropout,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        rngs=rngs,\n    )\n\n    # Count parameters\n    _, params, _ = nnx.split(lam, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.dtype,\n    )\n    optimizer = nnx.Optimizer(lam, tx)\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n    # --- Initialize checkpoint manager ---\n    step = 0\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n\n    # --- Create DataLoaderIterator from dataloader ---\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n\n    # --- Restore checkpoint ---\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = checkpoint_manager.latest_step() or 0\n        print(f""Restored dataloader and model state from step {step}"")\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    print(f""Starting training from step {step}..."")\n    action_last_active = jnp.zeros(args.num_latents, dtype=jnp.int32)\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            rng, _rng = jax.random.split(rng)\n\n            inputs = dict(videos=videos, rng=_rng)\n            rng, _rng = jax.random.split(rng)\n            loss, recon, action_last_active, metrics = train_step(\n                lam, optimizer, inputs, action_last_active, _rng\n            )\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0, 1:].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n",python,tab
+4,440483,"train_lam.py",2637,0,"",python,selection_mouse
+5,441103,"train_lam.py",2648,0,"",python,selection_mouse
+6,441819,"train_lam.py",2673,0,"",python,selection_mouse
+7,443158,"train_lam.py",2683,0,"",python,selection_mouse