Upload 3 files

step6500/adapter_config.json

@@ -0,0 +1,41 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": null,
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": false,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 64,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "qalora_group_size": 16,
+  "r": 64,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "k",
+    "v",
+    "o",
+    "q",
+    "ffn.0",
+    "ffn.2"
+  ],
+  "target_parameters": null,
+  "task_type": null,
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

step6500/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f1cc25d427eeb67cf104e5addc9d9a4c52bbbc4ed88ace5e028b37797e7ab97
+size 613516752

step6500/wan_high.toml

@@ -0,0 +1,116 @@
+# Output path for training runs. Each training run makes a new directory in here.
+output_dir = '/root/diffusion-pipe/output/graffito_v1_high_noise'
+
+# Dataset config file.
+dataset = '/root/diffusion-pipe/my_configs/dataset_graffito.toml'
+# You can have separate eval datasets. Give them a name for Tensorboard metrics.
+# eval_datasets = [
+#     {name = 'something', config = 'path/to/eval_dataset.toml'},
+# ]
+
+# training settings
+
+# I usually set this to a really high value because I don't know how long I want to train.
+epochs = 1000
+# Batch size of a single forward/backward pass for one GPU.
+micro_batch_size_per_gpu = 2
+image_micro_batch_size_per_gpu = 4
+# Pipeline parallelism degree. A single instance of the model is divided across this many GPUs.
+pipeline_stages = 1
+# Number of micro-batches sent through the pipeline for each training step.
+# If pipeline_stages > 1, a higher GAS means better GPU utilization due to smaller pipeline bubbles (where GPUs aren't overlapping computation).
+gradient_accumulation_steps = 1
+# Grad norm clipping.
+gradient_clipping = 1.0
+# Learning rate warmup.
+warmup_steps = 100
+
+
+
+# eval settings
+
+eval_every_n_epochs = 1
+eval_before_first_step = true
+# Might want to set these lower for eval so that less images get dropped (eval dataset size is usually much smaller than training set).
+# Each size bucket of images/videos is rounded down to the nearest multiple of the global batch size, so higher global batch size means
+# more dropped images. Usually doesn't matter for training but the eval set is much smaller so it can matter.
+eval_micro_batch_size_per_gpu = 1
+eval_gradient_accumulation_steps = 1
+
+# misc settings
+
+# Probably want to set this a bit higher if you have a smaller dataset so you don't end up with a million saved models.
+save_every_n_epochs = 100
+save_every_n_steps = 250
+# Can checkpoint the traiing state every n number of epochs or minutes. Set only one of these. You can resume from checkpoints using the --resume_from_checkpoint flag.
+#checkpoint_every_n_epochs = 1
+checkpoint_every_n_minutes = 30
+# Always set to true unless you have a huge amount of VRAM.
+activation_checkpointing = 'unsloth'
+# Controls how Deepspeed decides how to divide layers across GPUs. Probably don't change this.
+partition_method = 'parameters'
+# dtype for saving the LoRA or model, if different from training dtype
+save_dtype = 'bfloat16'
+# Batch size for caching latents and text embeddings. Increasing can lead to higher GPU utilization during caching phase but uses more memory.
+caching_batch_size = 16
+
+# Number of parallel processes to use in map() calls when caching the dataset. Defaults to min(8, num_cpu_cores) if unset.
+# map_num_proc = 4
+
+# How often deepspeed logs to console.
+steps_per_print = 1
+# How to extract video clips for training from a single input video file.
+# The video file is first assigned to one of the configured frame buckets, but then we must extract one or more clips of exactly the right
+# number of frames for that bucket.
+# single_beginning: one clip starting at the beginning of the video
+# single_middle: one clip from the middle of the video (cutting off the start and end equally)
+# multiple_overlapping: extract the minimum number of clips to cover the full range of the video. They might overlap some.
+# default is single_middle
+video_clip_mode = 'single_middle'
+
+# blocks_to_swap = 10
+
+
+[model]
+type = 'wan'
+# Can load Hunyuan Video entirely from the ckpt path set up for the official inference scripts.
+#ckpt_path = '/home/anon/HunyuanVideo/ckpts'
+ckpt_path = '/root/diffusion-pipe/imagegen_models/wan/Wan2.2-I2V-A14B'
+transformer_path = '/root/diffusion-pipe/imagegen_models/wan/Wan2.2-I2V-A14B/high_noise_model'
+# Or you can load it by pointing to all the ComfyUI files.
+# transformer_path = '/notebooks/diffusion-pipe/imagegen_models/hunyuan_video_720_cfgdistill_fp8_e4m3fn.safetensors'
+# vae_path = '/notebooks/diffusion-pipe/imagegen_models/hunyuan_video_vae_bf16.safetensors'
+# llm_path = '/notebooks/diffusion-pipe/imagegen_models/llava-llama-3-8b-text-encoder-tokenizer'
+# clip_path = '/notebooks/diffusion-pipe/imagegen_models/clip-vit-large-patch14'
+# Base dtype used for all models.
+dtype = 'bfloat16'
+
+transformer_dtype = 'bfloat16'
+min_t = 0.875
+max_t = 1.0
+# How to sample timesteps to train on. Can be logit_normal or uniform.
+timestep_sample_method = 'logit_normal'
+
+[adapter]
+type = 'lora'
+rank = 64
+# Dtype for the LoRA weights you are training.
+dtype = 'bfloat16'
+# You can initialize the lora weights from a previously trained lora.
+#init_from_existing = '/data/diffusion_pipe_training_runs/something/epoch50'
+
+[optimizer]
+# AdamW from the optimi library is a good default since it automatically uses Kahan summation when training bfloat16 weights.
+# Look at train.py for other options. You could also easily edit the file and add your own.
+type = 'adamw_optimi'
+lr = 2e-5
+# # type = 'adamw8bitkahan'
+# lr = 4e-5
+# betas = [0.9, 0.99]
+# weight_decay = 0.01
+# eps = 1e-8
+
+# type = 'automagic'
+# lr = 1e-6  # Starting learning rate
+# weight_decay = 0.001  # Weight decay
+# lr_bump = 2e-6  # Amount to bump LR when adjusting