Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
|
@@ -10,6 +10,8 @@ from huggingface_hub import HfApi, hf_hub_download
|
|
| 10 |
from safetensors.torch import load_file, save_file
|
| 11 |
import torch
|
| 12 |
import torch.nn.functional as F
|
|
|
|
|
|
|
| 13 |
try:
|
| 14 |
from modelscope.hub.file_download import model_file_download as ms_file_download
|
| 15 |
from modelscope.hub.api import HubApi as ModelScopeApi
|
|
@@ -17,75 +19,182 @@ try:
|
|
| 17 |
except ImportError:
|
| 18 |
MODELScope_AVAILABLE = False
|
| 19 |
|
| 20 |
-
def low_rank_decomposition(weight, rank=64):
|
| 21 |
-
"""
|
| 22 |
-
Correct LoRA decomposition supporting 2D and 4D tensors.
|
| 23 |
-
Returns (lora_A, lora_B) such that weight β lora_B @ lora_A for 2D,
|
| 24 |
-
or appropriate conv form for 4D.
|
| 25 |
-
"""
|
| 26 |
original_shape = weight.shape
|
| 27 |
original_dtype = weight.dtype
|
| 28 |
-
|
| 29 |
try:
|
|
|
|
| 30 |
if weight.ndim == 2:
|
| 31 |
-
|
| 32 |
-
if actual_rank < 4:
|
| 33 |
-
return None, None
|
| 34 |
-
|
| 35 |
U, S, Vh = torch.linalg.svd(weight.float(), full_matrices=False)
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
elif weight.ndim == 4:
|
| 45 |
-
out_ch, in_ch,
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 67 |
except Exception as e:
|
| 68 |
-
print(f"Decomposition error for {original_shape}: {e}")
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 77 |
return False
|
| 78 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79 |
if architecture == "text_encoder":
|
| 80 |
-
|
|
|
|
|
|
|
|
|
|
| 81 |
elif architecture == "unet_transformer":
|
| 82 |
-
|
|
|
|
|
|
|
|
|
|
| 83 |
elif architecture == "unet_conv":
|
| 84 |
-
|
|
|
|
|
|
|
|
|
|
| 85 |
elif architecture == "vae":
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 89 |
|
| 90 |
def convert_safetensors_to_fp8_with_lora(safetensors_path, output_dir, fp8_format, lora_rank=64, architecture="auto", progress=gr.Progress()):
|
| 91 |
progress(0.1, desc="Starting FP8 conversion with LoRA extraction...")
|
|
@@ -96,84 +205,196 @@ def convert_safetensors_to_fp8_with_lora(safetensors_path, output_dir, fp8_forma
|
|
| 96 |
header_json = f.read(header_size).decode('utf-8')
|
| 97 |
header = json.loads(header_json)
|
| 98 |
return header.get('__metadata__', {})
|
| 99 |
-
|
| 100 |
metadata = read_safetensors_metadata(safetensors_path)
|
| 101 |
progress(0.2, desc="Loaded metadata.")
|
| 102 |
-
|
| 103 |
state_dict = load_file(safetensors_path)
|
| 104 |
progress(0.4, desc="Loaded weights.")
|
| 105 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
if fp8_format == "e5m2":
|
| 107 |
fp8_dtype = torch.float8_e5m2
|
| 108 |
else:
|
| 109 |
fp8_dtype = torch.float8_e4m3fn
|
| 110 |
-
|
| 111 |
sd_fp8 = {}
|
| 112 |
lora_weights = {}
|
| 113 |
-
total = len(state_dict)
|
| 114 |
-
lora_keys = []
|
| 115 |
-
|
| 116 |
lora_stats = {
|
| 117 |
-
'total_layers':
|
|
|
|
| 118 |
'layers_eligible': 0,
|
| 119 |
'layers_processed': 0,
|
| 120 |
'layers_skipped': [],
|
|
|
|
|
|
|
| 121 |
}
|
| 122 |
-
|
|
|
|
|
|
|
|
|
|
| 123 |
for i, key in enumerate(state_dict):
|
| 124 |
-
progress(0.4 + 0.4 * (i / total), desc=f"Processing {i+1}/{total}
|
| 125 |
weight = state_dict[key]
|
| 126 |
-
|
|
|
|
| 127 |
if weight.dtype in [torch.float16, torch.float32, torch.bfloat16]:
|
| 128 |
fp8_weight = weight.to(fp8_dtype)
|
| 129 |
sd_fp8[key] = fp8_weight
|
| 130 |
-
|
| 131 |
-
if
|
|
|
|
|
|
|
|
|
|
| 132 |
lora_stats['layers_eligible'] += 1
|
| 133 |
-
|
| 134 |
try:
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 139 |
lora_keys.append(key)
|
| 140 |
lora_stats['layers_processed'] += 1
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
else:
|
| 142 |
-
lora_stats['layers_skipped'].append(f"{key}: decomposition
|
|
|
|
| 143 |
except Exception as e:
|
| 144 |
-
|
|
|
|
|
|
|
| 145 |
else:
|
| 146 |
-
reason = "
|
| 147 |
-
lora_stats['layers_skipped'].append(f"{key}:
|
| 148 |
else:
|
| 149 |
sd_fp8[key] = weight
|
| 150 |
-
lora_stats['layers_skipped'].append(f"{key}:
|
| 151 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 152 |
base_name = os.path.splitext(os.path.basename(safetensors_path))[0]
|
| 153 |
fp8_path = os.path.join(output_dir, f"{base_name}-fp8-{fp8_format}.safetensors")
|
| 154 |
lora_path = os.path.join(output_dir, f"{base_name}-lora-r{lora_rank}-{architecture}.safetensors")
|
| 155 |
-
|
| 156 |
save_file(sd_fp8, fp8_path, metadata={"format": "pt", "fp8_format": fp8_format, **metadata})
|
| 157 |
-
|
| 158 |
-
|
|
|
|
|
|
|
| 159 |
"lora_rank": str(lora_rank),
|
| 160 |
"architecture": architecture,
|
|
|
|
|
|
|
| 161 |
"stats": json.dumps(lora_stats)
|
| 162 |
-
}
|
| 163 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 164 |
progress(0.9, desc="Saved FP8 and LoRA files.")
|
| 165 |
progress(1.0, desc="β
FP8 + LoRA extraction complete!")
|
| 166 |
-
|
| 167 |
-
stats_msg = f"FP8 ({fp8_format}) and rank-{lora_rank} LoRA ({architecture}) saved.\n"
|
| 168 |
-
stats_msg += f"Processed {lora_stats['layers_processed']}/{lora_stats['layers_eligible']} eligible layers."
|
| 169 |
if lora_stats['layers_processed'] == 0:
|
| 170 |
-
stats_msg += "
|
| 171 |
-
|
| 172 |
-
|
|
|
|
|
|
|
| 173 |
|
| 174 |
except Exception as e:
|
| 175 |
-
|
| 176 |
-
|
|
|
|
| 177 |
|
| 178 |
def parse_hf_url(url):
|
| 179 |
url = url.strip().rstrip("/")
|
|
@@ -251,9 +472,11 @@ def process_and_upload_fp8(
|
|
| 251 |
return None, "β Hugging Face token required for source.", ""
|
| 252 |
if target_type == "huggingface" and not hf_token:
|
| 253 |
return None, "β Hugging Face token required for target.", ""
|
|
|
|
|
|
|
| 254 |
if lora_rank < 4:
|
| 255 |
return None, "β LoRA rank must be at least 4.", ""
|
| 256 |
-
|
| 257 |
temp_dir = None
|
| 258 |
output_dir = tempfile.mkdtemp()
|
| 259 |
try:
|
|
@@ -261,24 +484,24 @@ def process_and_upload_fp8(
|
|
| 261 |
safetensors_path, temp_dir = download_safetensors_file(
|
| 262 |
source_type, repo_url, safetensors_filename, hf_token, progress
|
| 263 |
)
|
| 264 |
-
|
| 265 |
progress(0.25, desc=f"Converting to FP8 with LoRA ({architecture})...")
|
| 266 |
success, msg, stats = convert_safetensors_to_fp8_with_lora(
|
| 267 |
safetensors_path, output_dir, fp8_format, lora_rank, architecture, progress
|
| 268 |
)
|
| 269 |
-
|
| 270 |
if not success:
|
| 271 |
return None, f"β Conversion failed: {msg}", ""
|
| 272 |
-
|
| 273 |
progress(0.9, desc="Uploading...")
|
| 274 |
repo_url_final = upload_to_target(
|
| 275 |
target_type, new_repo_id, output_dir, fp8_format, architecture, hf_token, modelscope_token, private_repo
|
| 276 |
)
|
| 277 |
-
|
| 278 |
base_name = os.path.splitext(safetensors_filename)[0]
|
| 279 |
lora_filename = f"{base_name}-lora-r{lora_rank}-{architecture}.safetensors"
|
| 280 |
fp8_filename = f"{base_name}-fp8-{fp8_format}.safetensors"
|
| 281 |
-
|
| 282 |
readme = f"""---
|
| 283 |
library_name: diffusers
|
| 284 |
tags:
|
|
@@ -288,7 +511,7 @@ tags:
|
|
| 288 |
- low-rank
|
| 289 |
- diffusion
|
| 290 |
- architecture-{architecture}
|
| 291 |
-
- converted-by-
|
| 292 |
---
|
| 293 |
# FP8 Model with Low-Rank LoRA
|
| 294 |
- **Source**: `{repo_url}`
|
|
@@ -299,6 +522,16 @@ tags:
|
|
| 299 |
- **LoRA File**: `{lora_filename}`
|
| 300 |
- **FP8 File**: `{fp8_filename}`
|
| 301 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 302 |
## Usage (Inference)
|
| 303 |
```python
|
| 304 |
from safetensors.torch import load_file
|
|
@@ -311,26 +544,38 @@ lora_state = load_file("{lora_filename}")
|
|
| 311 |
# Reconstruct approximate original weights
|
| 312 |
reconstructed = {{}}
|
| 313 |
for key in fp8_state:
|
| 314 |
-
|
| 315 |
-
|
| 316 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 317 |
if A.ndim == 2 and B.ndim == 2:
|
| 318 |
lora_weight = B @ A
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 319 |
else:
|
| 320 |
-
#
|
| 321 |
-
lora_weight =
|
| 322 |
-
|
|
|
|
|
|
|
| 323 |
reconstructed[key] = fp8_state[key].to(torch.float32) + lora_weight
|
| 324 |
else:
|
| 325 |
reconstructed[key] = fp8_state[key].to(torch.float32)
|
| 326 |
```
|
| 327 |
|
| 328 |
-
> Requires PyTorch β₯ 2.1 for FP8 support.
|
| 329 |
"""
|
| 330 |
-
|
| 331 |
with open(os.path.join(output_dir, "README.md"), "w") as f:
|
| 332 |
f.write(readme)
|
| 333 |
-
|
| 334 |
if target_type == "huggingface":
|
| 335 |
HfApi(token=hf_token).upload_file(
|
| 336 |
path_or_fileobj=os.path.join(output_dir, "README.md"),
|
|
@@ -339,18 +584,27 @@ for key in fp8_state:
|
|
| 339 |
repo_type="model",
|
| 340 |
token=hf_token
|
| 341 |
)
|
| 342 |
-
|
| 343 |
progress(1.0, desc="β
Done!")
|
| 344 |
result_html = f"""
|
| 345 |
β
Success!
|
| 346 |
Model uploaded to: <a href="{repo_url_final}" target="_blank">{new_repo_id}</a>
|
| 347 |
-
Includes:
|
|
|
|
|
|
|
|
|
|
|
|
|
| 348 |
"""
|
|
|
|
|
|
|
|
|
|
| 349 |
return gr.HTML(result_html), "β
FP8 + LoRA upload successful!", msg
|
| 350 |
-
|
| 351 |
except Exception as e:
|
| 352 |
-
|
| 353 |
-
|
|
|
|
|
|
|
| 354 |
finally:
|
| 355 |
if temp_dir:
|
| 356 |
shutil.rmtree(temp_dir, ignore_errors=True)
|
|
@@ -358,17 +612,18 @@ Includes: FP8 + rank-{lora_rank} LoRA ({architecture}).
|
|
| 358 |
|
| 359 |
with gr.Blocks(title="FP8 + LoRA Extractor (HF β ModelScope)") as demo:
|
| 360 |
gr.Markdown("# π Advanced FP8 Pruner with Architecture-Specific LoRA Extraction")
|
| 361 |
-
gr.Markdown("Convert `.safetensors` β **FP8** + **targeted LoRA** for precision recovery. Supports Hugging Face β ModelScope.")
|
| 362 |
-
|
| 363 |
with gr.Row():
|
| 364 |
with gr.Column():
|
| 365 |
source_type = gr.Radio(["huggingface", "modelscope"], value="huggingface", label="Source")
|
| 366 |
repo_url = gr.Textbox(label="Repo URL or ID", placeholder="https://huggingface.co/... or modelscope-id")
|
| 367 |
safetensors_filename = gr.Textbox(label="Filename", placeholder="model.safetensors")
|
| 368 |
-
|
| 369 |
with gr.Accordion("Advanced LoRA Settings", open=True):
|
| 370 |
fp8_format = gr.Radio(["e4m3fn", "e5m2"], value="e5m2", label="FP8 Format")
|
| 371 |
lora_rank = gr.Slider(minimum=4, maximum=256, step=4, value=64, label="LoRA Rank")
|
|
|
|
| 372 |
architecture = gr.Dropdown(
|
| 373 |
choices=[
|
| 374 |
("Auto-detect components", "auto"),
|
|
@@ -379,24 +634,25 @@ with gr.Blocks(title="FP8 + LoRA Extractor (HF β ModelScope)") as demo:
|
|
| 379 |
("All components", "all")
|
| 380 |
],
|
| 381 |
value="auto",
|
| 382 |
-
label="Target Architecture"
|
|
|
|
| 383 |
)
|
| 384 |
-
|
| 385 |
with gr.Accordion("Authentication", open=False):
|
| 386 |
hf_token = gr.Textbox(label="Hugging Face Token", type="password")
|
| 387 |
modelscope_token = gr.Textbox(label="ModelScope Token (optional)", type="password", visible=MODELScope_AVAILABLE)
|
| 388 |
-
|
| 389 |
with gr.Column():
|
| 390 |
target_type = gr.Radio(["huggingface", "modelscope"], value="huggingface", label="Target")
|
| 391 |
new_repo_id = gr.Textbox(label="New Repo ID", placeholder="user/model-fp8-lora")
|
| 392 |
private_repo = gr.Checkbox(label="Private Repository (HF only)", value=False)
|
| 393 |
-
|
| 394 |
status_output = gr.Markdown()
|
| 395 |
detailed_log = gr.Textbox(label="Processing Log", interactive=False, lines=10)
|
| 396 |
-
|
| 397 |
convert_btn = gr.Button("π Convert & Upload", variant="primary")
|
| 398 |
repo_link_output = gr.HTML()
|
| 399 |
-
|
| 400 |
convert_btn.click(
|
| 401 |
fn=process_and_upload_fp8,
|
| 402 |
inputs=[
|
|
@@ -415,7 +671,7 @@ with gr.Blocks(title="FP8 + LoRA Extractor (HF β ModelScope)") as demo:
|
|
| 415 |
outputs=[repo_link_output, status_output, detailed_log],
|
| 416 |
show_progress=True
|
| 417 |
)
|
| 418 |
-
|
| 419 |
gr.Examples(
|
| 420 |
examples=[
|
| 421 |
["huggingface", "https://huggingface.co/Yabo/FramePainter/tree/main", "unet_diffusion_pytorch_model.safetensors", "e5m2", 64, "unet_transformer"],
|
|
@@ -425,15 +681,17 @@ with gr.Blocks(title="FP8 + LoRA Extractor (HF β ModelScope)") as demo:
|
|
| 425 |
inputs=[source_type, repo_url, safetensors_filename, fp8_format, lora_rank, architecture],
|
| 426 |
label="Example Conversions"
|
| 427 |
)
|
| 428 |
-
|
| 429 |
gr.Markdown("""
|
| 430 |
## π‘ Usage Tips
|
| 431 |
-
|
| 432 |
-
- **
|
| 433 |
-
- **UNet
|
| 434 |
-
- **
|
| 435 |
-
- **
|
| 436 |
-
-
|
|
|
|
|
|
|
| 437 |
""")
|
| 438 |
|
| 439 |
demo.launch()
|
|
|
|
| 10 |
from safetensors.torch import load_file, save_file
|
| 11 |
import torch
|
| 12 |
import torch.nn.functional as F
|
| 13 |
+
import traceback
|
| 14 |
+
import math
|
| 15 |
try:
|
| 16 |
from modelscope.hub.file_download import model_file_download as ms_file_download
|
| 17 |
from modelscope.hub.api import HubApi as ModelScopeApi
|
|
|
|
| 19 |
except ImportError:
|
| 20 |
MODELScope_AVAILABLE = False
|
| 21 |
|
| 22 |
+
def low_rank_decomposition(weight, rank=64, approximation_factor=0.8):
|
| 23 |
+
"""Low-rank decomposition with controlled approximation error."""
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
original_shape = weight.shape
|
| 25 |
original_dtype = weight.dtype
|
| 26 |
+
|
| 27 |
try:
|
| 28 |
+
# Handle 2D tensors (linear layers, attention)
|
| 29 |
if weight.ndim == 2:
|
| 30 |
+
# Compute SVD
|
|
|
|
|
|
|
|
|
|
| 31 |
U, S, Vh = torch.linalg.svd(weight.float(), full_matrices=False)
|
| 32 |
+
|
| 33 |
+
# Calculate how much variance we want to keep
|
| 34 |
+
total_variance = torch.sum(S ** 2)
|
| 35 |
+
cumulative_variance = torch.cumsum(S ** 2, dim=0)
|
| 36 |
+
|
| 37 |
+
# Find minimal rank that preserves approximation_factor of variance
|
| 38 |
+
minimal_rank = torch.searchsorted(cumulative_variance, approximation_factor * total_variance).item() + 1
|
| 39 |
+
|
| 40 |
+
# Use the smaller of: requested rank or minimal rank for approximation_factor
|
| 41 |
+
actual_rank = min(rank, len(S))
|
| 42 |
+
|
| 43 |
+
# If actual_rank is too close to full rank, reduce it to create meaningful approximation
|
| 44 |
+
if actual_rank > len(S) * 0.8: # If using more than 80% of full rank
|
| 45 |
+
actual_rank = max(min(rank // 2, len(S) // 2), 8) # Use half the requested rank
|
| 46 |
+
|
| 47 |
+
# Ensure we're actually approximating, not just reparameterizing
|
| 48 |
+
if actual_rank >= min(weight.shape):
|
| 49 |
+
# Force approximation by using lower rank
|
| 50 |
+
actual_rank = max(min(weight.shape) // 4, 8)
|
| 51 |
+
|
| 52 |
+
U_k = U[:, :actual_rank] @ torch.diag(torch.sqrt(S[:actual_rank]))
|
| 53 |
+
Vh_k = torch.diag(torch.sqrt(S[:actual_rank])) @ Vh[:actual_rank, :]
|
| 54 |
+
|
| 55 |
+
return U_k.contiguous(), Vh_k.contiguous()
|
| 56 |
+
|
| 57 |
+
# Handle 4D tensors (convolutional layers)
|
| 58 |
elif weight.ndim == 4:
|
| 59 |
+
out_ch, in_ch, kH, kW = weight.shape
|
| 60 |
+
|
| 61 |
+
# Reshape to 2D for SVD
|
| 62 |
+
weight_2d = weight.view(out_ch, in_ch * kH * kW)
|
| 63 |
+
|
| 64 |
+
# Compute SVD on flattened version
|
| 65 |
+
U, S, Vh = torch.linalg.svd(weight_2d.float(), full_matrices=False)
|
| 66 |
+
|
| 67 |
+
# Calculate appropriate rank
|
| 68 |
+
total_variance = torch.sum(S ** 2)
|
| 69 |
+
cumulative_variance = torch.cumsum(S ** 2, dim=0)
|
| 70 |
+
minimal_rank = torch.searchsorted(cumulative_variance, approximation_factor * total_variance).item() + 1
|
| 71 |
+
|
| 72 |
+
# Adjust rank for convolutions - typically need lower ranks
|
| 73 |
+
conv_rank = min(rank // 2, len(S))
|
| 74 |
+
if conv_rank > len(S) * 0.7:
|
| 75 |
+
conv_rank = max(len(S) // 4, 8)
|
| 76 |
+
|
| 77 |
+
actual_rank = max(min(conv_rank, minimal_rank), 8)
|
| 78 |
+
|
| 79 |
+
# Decompose
|
| 80 |
+
U_k = U[:, :actual_rank] @ torch.diag(torch.sqrt(S[:actual_rank]))
|
| 81 |
+
Vh_k = torch.diag(torch.sqrt(S[:actual_rank])) @ Vh[:actual_rank, :]
|
| 82 |
+
|
| 83 |
+
# Reshape back to convolutional format
|
| 84 |
+
if kH == 1 and kW == 1: # 1x1 convolutions
|
| 85 |
+
U_k = U_k.view(out_ch, actual_rank, 1, 1)
|
| 86 |
+
Vh_k = Vh_k.view(actual_rank, in_ch, 1, 1)
|
| 87 |
+
else:
|
| 88 |
+
# For larger kernels, use spatial decomposition
|
| 89 |
+
U_k = U_k.view(out_ch, actual_rank, 1, 1)
|
| 90 |
+
Vh_k = Vh_k.view(actual_rank, in_ch, kH, kW)
|
| 91 |
+
|
| 92 |
+
return U_k.contiguous(), Vh_k.contiguous()
|
| 93 |
+
|
| 94 |
+
# Handle 1D tensors (biases, embeddings)
|
| 95 |
+
elif weight.ndim == 1:
|
| 96 |
+
# Don't decompose 1D tensors
|
| 97 |
+
return None, None
|
| 98 |
+
|
| 99 |
except Exception as e:
|
| 100 |
+
print(f"Decomposition error for tensor with shape {original_shape}: {str(e)[:100]}")
|
| 101 |
+
|
| 102 |
+
return None, None
|
| 103 |
+
|
| 104 |
+
def get_architecture_specific_settings(architecture, base_rank):
|
| 105 |
+
"""Get optimal settings for different model architectures."""
|
| 106 |
+
settings = {
|
| 107 |
+
"text_encoder": {
|
| 108 |
+
"rank": base_rank,
|
| 109 |
+
"approximation_factor": 0.95, # Text encoders need high accuracy
|
| 110 |
+
"min_rank": 8,
|
| 111 |
+
"max_rank_factor": 0.5 # Use at most 50% of full rank
|
| 112 |
+
},
|
| 113 |
+
"unet_transformer": {
|
| 114 |
+
"rank": base_rank,
|
| 115 |
+
"approximation_factor": 0.90,
|
| 116 |
+
"min_rank": 16,
|
| 117 |
+
"max_rank_factor": 0.4
|
| 118 |
+
},
|
| 119 |
+
"unet_conv": {
|
| 120 |
+
"rank": base_rank // 2, # Convolutions compress better
|
| 121 |
+
"approximation_factor": 0.85,
|
| 122 |
+
"min_rank": 8,
|
| 123 |
+
"max_rank_factor": 0.3
|
| 124 |
+
},
|
| 125 |
+
"vae": {
|
| 126 |
+
"rank": base_rank // 3, # VAE compresses very well
|
| 127 |
+
"approximation_factor": 0.80,
|
| 128 |
+
"min_rank": 4,
|
| 129 |
+
"max_rank_factor": 0.25
|
| 130 |
+
},
|
| 131 |
+
"auto": {
|
| 132 |
+
"rank": base_rank,
|
| 133 |
+
"approximation_factor": 0.90,
|
| 134 |
+
"min_rank": 8,
|
| 135 |
+
"max_rank_factor": 0.5
|
| 136 |
+
},
|
| 137 |
+
"all": {
|
| 138 |
+
"rank": base_rank,
|
| 139 |
+
"approximation_factor": 0.90,
|
| 140 |
+
"min_rank": 8,
|
| 141 |
+
"max_rank_factor": 0.5
|
| 142 |
+
}
|
| 143 |
+
}
|
| 144 |
+
|
| 145 |
+
return settings.get(architecture, settings["auto"])
|
| 146 |
+
|
| 147 |
+
def should_apply_lora(key, weight, architecture, lora_rank):
|
| 148 |
+
"""Determine if LoRA should be applied to a specific weight based on architecture selection."""
|
| 149 |
+
|
| 150 |
+
# Skip bias terms, batchnorm, and very small tensors
|
| 151 |
+
if 'bias' in key or 'norm' in key.lower() or 'bn' in key.lower():
|
| 152 |
return False
|
| 153 |
+
|
| 154 |
+
# Skip very small tensors
|
| 155 |
+
if weight.numel() < 100:
|
| 156 |
+
return False
|
| 157 |
+
|
| 158 |
+
# Skip 1D tensors
|
| 159 |
+
if weight.ndim == 1:
|
| 160 |
+
return False
|
| 161 |
+
|
| 162 |
+
# Architecture-specific rules
|
| 163 |
+
lower_key = key.lower()
|
| 164 |
+
|
| 165 |
if architecture == "text_encoder":
|
| 166 |
+
# Text encoder: focus on embeddings and attention layers
|
| 167 |
+
return ('emb' in lower_key or 'embed' in lower_key or
|
| 168 |
+
'attn' in lower_key or 'qkv' in lower_key or 'mlp' in lower_key)
|
| 169 |
+
|
| 170 |
elif architecture == "unet_transformer":
|
| 171 |
+
# UNet transformers: focus on attention blocks
|
| 172 |
+
return ('attn' in lower_key or 'transformer' in lower_key or
|
| 173 |
+
'qkv' in lower_key or 'to_out' in lower_key)
|
| 174 |
+
|
| 175 |
elif architecture == "unet_conv":
|
| 176 |
+
# UNet convolutional layers
|
| 177 |
+
return ('conv' in lower_key or 'resnet' in lower_key or
|
| 178 |
+
'downsample' in lower_key or 'upsample' in lower_key)
|
| 179 |
+
|
| 180 |
elif architecture == "vae":
|
| 181 |
+
# VAE components
|
| 182 |
+
return ('encoder' in lower_key or 'decoder' in lower_key or
|
| 183 |
+
'conv' in lower_key or 'post_quant' in lower_key)
|
| 184 |
+
|
| 185 |
+
elif architecture == "all":
|
| 186 |
+
# Apply to all eligible tensors
|
| 187 |
+
return True
|
| 188 |
+
|
| 189 |
+
elif architecture == "auto":
|
| 190 |
+
# Auto-detect based on tensor properties
|
| 191 |
+
if weight.ndim == 2 and min(weight.shape) > lora_rank // 4:
|
| 192 |
+
return True
|
| 193 |
+
if weight.ndim == 4 and (weight.shape[0] > lora_rank // 4 or weight.shape[1] > lora_rank // 4):
|
| 194 |
+
return True
|
| 195 |
+
return False
|
| 196 |
+
|
| 197 |
+
return False
|
| 198 |
|
| 199 |
def convert_safetensors_to_fp8_with_lora(safetensors_path, output_dir, fp8_format, lora_rank=64, architecture="auto", progress=gr.Progress()):
|
| 200 |
progress(0.1, desc="Starting FP8 conversion with LoRA extraction...")
|
|
|
|
| 205 |
header_json = f.read(header_size).decode('utf-8')
|
| 206 |
header = json.loads(header_json)
|
| 207 |
return header.get('__metadata__', {})
|
| 208 |
+
|
| 209 |
metadata = read_safetensors_metadata(safetensors_path)
|
| 210 |
progress(0.2, desc="Loaded metadata.")
|
| 211 |
+
|
| 212 |
state_dict = load_file(safetensors_path)
|
| 213 |
progress(0.4, desc="Loaded weights.")
|
| 214 |
+
|
| 215 |
+
# Architecture analysis
|
| 216 |
+
architecture_stats = {
|
| 217 |
+
'text_encoder': 0,
|
| 218 |
+
'unet_transformer': 0,
|
| 219 |
+
'unet_conv': 0,
|
| 220 |
+
'vae': 0,
|
| 221 |
+
'other': 0
|
| 222 |
+
}
|
| 223 |
+
|
| 224 |
+
for key in state_dict:
|
| 225 |
+
lower_key = key.lower()
|
| 226 |
+
if 'text' in lower_key or 'emb' in lower_key:
|
| 227 |
+
architecture_stats['text_encoder'] += 1
|
| 228 |
+
elif 'attn' in lower_key or 'transformer' in lower_key:
|
| 229 |
+
architecture_stats['unet_transformer'] += 1
|
| 230 |
+
elif 'conv' in lower_key or 'resnet' in lower_key:
|
| 231 |
+
architecture_stats['unet_conv'] += 1
|
| 232 |
+
elif 'vae' in lower_key or 'encoder' in lower_key or 'decoder' in lower_key:
|
| 233 |
+
architecture_stats['vae'] += 1
|
| 234 |
+
else:
|
| 235 |
+
architecture_stats['other'] += 1
|
| 236 |
+
|
| 237 |
+
print("Architecture analysis:")
|
| 238 |
+
for arch, count in architecture_stats.items():
|
| 239 |
+
print(f"- {arch}: {count} layers")
|
| 240 |
+
|
| 241 |
if fp8_format == "e5m2":
|
| 242 |
fp8_dtype = torch.float8_e5m2
|
| 243 |
else:
|
| 244 |
fp8_dtype = torch.float8_e4m3fn
|
| 245 |
+
|
| 246 |
sd_fp8 = {}
|
| 247 |
lora_weights = {}
|
|
|
|
|
|
|
|
|
|
| 248 |
lora_stats = {
|
| 249 |
+
'total_layers': len(state_dict),
|
| 250 |
+
'layers_analyzed': 0,
|
| 251 |
'layers_eligible': 0,
|
| 252 |
'layers_processed': 0,
|
| 253 |
'layers_skipped': [],
|
| 254 |
+
'architecture_distro': architecture_stats,
|
| 255 |
+
'reconstruction_errors': []
|
| 256 |
}
|
| 257 |
+
|
| 258 |
+
total = len(state_dict)
|
| 259 |
+
lora_keys = []
|
| 260 |
+
|
| 261 |
for i, key in enumerate(state_dict):
|
| 262 |
+
progress(0.4 + 0.4 * (i / total), desc=f"Processing {i+1}/{total}: {key.split('.')[-1]}")
|
| 263 |
weight = state_dict[key]
|
| 264 |
+
lora_stats['layers_analyzed'] += 1
|
| 265 |
+
|
| 266 |
if weight.dtype in [torch.float16, torch.float32, torch.bfloat16]:
|
| 267 |
fp8_weight = weight.to(fp8_dtype)
|
| 268 |
sd_fp8[key] = fp8_weight
|
| 269 |
+
|
| 270 |
+
# Determine if we should apply LoRA
|
| 271 |
+
eligible_for_lora = should_apply_lora(key, weight, architecture, lora_rank)
|
| 272 |
+
|
| 273 |
+
if eligible_for_lora:
|
| 274 |
lora_stats['layers_eligible'] += 1
|
| 275 |
+
|
| 276 |
try:
|
| 277 |
+
# Get architecture-specific settings
|
| 278 |
+
arch_settings = get_architecture_specific_settings(architecture, lora_rank)
|
| 279 |
+
|
| 280 |
+
# Adjust rank based on tensor properties
|
| 281 |
+
if weight.ndim == 2:
|
| 282 |
+
max_possible_rank = min(weight.shape)
|
| 283 |
+
actual_rank = min(
|
| 284 |
+
arch_settings["rank"],
|
| 285 |
+
int(max_possible_rank * arch_settings["max_rank_factor"])
|
| 286 |
+
)
|
| 287 |
+
actual_rank = max(actual_rank, arch_settings["min_rank"])
|
| 288 |
+
elif weight.ndim == 4:
|
| 289 |
+
# For conv layers, use smaller rank
|
| 290 |
+
actual_rank = min(
|
| 291 |
+
arch_settings["rank"],
|
| 292 |
+
max(weight.shape[0], weight.shape[1]) // 4
|
| 293 |
+
)
|
| 294 |
+
actual_rank = max(actual_rank, arch_settings["min_rank"])
|
| 295 |
+
else:
|
| 296 |
+
# Skip non-2D/4D tensors for LoRA
|
| 297 |
+
lora_stats['layers_skipped'].append(f"{key}: unsupported ndim={weight.ndim}")
|
| 298 |
+
continue
|
| 299 |
+
|
| 300 |
+
if actual_rank < 4:
|
| 301 |
+
lora_stats['layers_skipped'].append(f"{key}: rank too small ({actual_rank})")
|
| 302 |
+
continue
|
| 303 |
+
|
| 304 |
+
# Perform decomposition with approximation
|
| 305 |
+
U, V = low_rank_decomposition(
|
| 306 |
+
weight,
|
| 307 |
+
rank=actual_rank,
|
| 308 |
+
approximation_factor=arch_settings["approximation_factor"]
|
| 309 |
+
)
|
| 310 |
+
|
| 311 |
+
if U is not None and V is not None:
|
| 312 |
+
# Store as half-precision
|
| 313 |
+
lora_weights[f"lora_A.{key}"] = U.to(torch.float16)
|
| 314 |
+
lora_weights[f"lora_B.{key}"] = V.to(torch.float16)
|
| 315 |
lora_keys.append(key)
|
| 316 |
lora_stats['layers_processed'] += 1
|
| 317 |
+
|
| 318 |
+
# Calculate and store reconstruction error
|
| 319 |
+
if U.ndim == 2 and V.ndim == 2:
|
| 320 |
+
if V.shape[0] == U.shape[1]:
|
| 321 |
+
reconstructed = V @ U
|
| 322 |
+
else:
|
| 323 |
+
reconstructed = U @ V
|
| 324 |
+
error = torch.norm(weight.float() - reconstructed.float()) / torch.norm(weight.float())
|
| 325 |
+
lora_stats['reconstruction_errors'].append({
|
| 326 |
+
'key': key,
|
| 327 |
+
'error': error.item(),
|
| 328 |
+
'original_shape': list(weight.shape),
|
| 329 |
+
'rank': actual_rank
|
| 330 |
+
})
|
| 331 |
else:
|
| 332 |
+
lora_stats['layers_skipped'].append(f"{key}: decomposition returned None")
|
| 333 |
+
|
| 334 |
except Exception as e:
|
| 335 |
+
error_msg = f"{key}: {str(e)[:100]}"
|
| 336 |
+
lora_stats['layers_skipped'].append(error_msg)
|
| 337 |
+
|
| 338 |
else:
|
| 339 |
+
reason = "not eligible for selected architecture" if architecture != "auto" else f"ndim={weight.ndim}"
|
| 340 |
+
lora_stats['layers_skipped'].append(f"{key}: {reason}")
|
| 341 |
else:
|
| 342 |
sd_fp8[key] = weight
|
| 343 |
+
lora_stats['layers_skipped'].append(f"{key}: unsupported dtype {weight.dtype}")
|
| 344 |
+
|
| 345 |
+
# Add reconstruction error statistics
|
| 346 |
+
if lora_stats['reconstruction_errors']:
|
| 347 |
+
errors = [e['error'] for e in lora_stats['reconstruction_errors']]
|
| 348 |
+
lora_stats['avg_reconstruction_error'] = sum(errors) / len(errors) if errors else 0
|
| 349 |
+
lora_stats['max_reconstruction_error'] = max(errors) if errors else 0
|
| 350 |
+
lora_stats['min_reconstruction_error'] = min(errors) if errors else 0
|
| 351 |
+
|
| 352 |
base_name = os.path.splitext(os.path.basename(safetensors_path))[0]
|
| 353 |
fp8_path = os.path.join(output_dir, f"{base_name}-fp8-{fp8_format}.safetensors")
|
| 354 |
lora_path = os.path.join(output_dir, f"{base_name}-lora-r{lora_rank}-{architecture}.safetensors")
|
| 355 |
+
|
| 356 |
save_file(sd_fp8, fp8_path, metadata={"format": "pt", "fp8_format": fp8_format, **metadata})
|
| 357 |
+
|
| 358 |
+
# Always save LoRA file, even if empty
|
| 359 |
+
lora_metadata = {
|
| 360 |
+
"format": "pt",
|
| 361 |
"lora_rank": str(lora_rank),
|
| 362 |
"architecture": architecture,
|
| 363 |
+
"original_filename": os.path.basename(safetensors_path),
|
| 364 |
+
"fp8_format": fp8_format,
|
| 365 |
"stats": json.dumps(lora_stats)
|
| 366 |
+
}
|
| 367 |
+
|
| 368 |
+
save_file(lora_weights, lora_path, metadata=lora_metadata)
|
| 369 |
+
|
| 370 |
+
# Generate detailed statistics message
|
| 371 |
+
stats_msg = f"""
|
| 372 |
+
π LoRA Extraction Statistics:
|
| 373 |
+
- Total layers analyzed: {lora_stats['layers_analyzed']}
|
| 374 |
+
- Layers eligible for LoRA: {lora_stats['layers_eligible']}
|
| 375 |
+
- Successfully processed: {lora_stats['layers_processed']}
|
| 376 |
+
- Architecture: {architecture}
|
| 377 |
+
- FP8 Format: {fp8_format.upper()}
|
| 378 |
+
"""
|
| 379 |
+
|
| 380 |
+
if 'avg_reconstruction_error' in lora_stats:
|
| 381 |
+
stats_msg += f"- Avg reconstruction error: {lora_stats['avg_reconstruction_error']:.6f}\n"
|
| 382 |
+
stats_msg += f"- Max reconstruction error: {lora_stats['max_reconstruction_error']:.6f}\n"
|
| 383 |
+
|
| 384 |
progress(0.9, desc="Saved FP8 and LoRA files.")
|
| 385 |
progress(1.0, desc="β
FP8 + LoRA extraction complete!")
|
| 386 |
+
|
|
|
|
|
|
|
| 387 |
if lora_stats['layers_processed'] == 0:
|
| 388 |
+
stats_msg += "\n\nβ οΈ WARNING: No LoRA weights were generated. Try a different architecture selection or lower rank."
|
| 389 |
+
elif lora_stats.get('avg_reconstruction_error', 1) < 0.0001:
|
| 390 |
+
stats_msg += "\n\nβΉοΈ NOTE: Very low reconstruction error detected. LoRA may be reconstructing almost perfectly. Consider using lower rank for better compression."
|
| 391 |
+
|
| 392 |
+
return True, f"FP8 ({fp8_format}) and rank-{lora_rank} LoRA saved.\n{stats_msg}", lora_stats
|
| 393 |
|
| 394 |
except Exception as e:
|
| 395 |
+
error_msg = f"Conversion error: {str(e)}\n{traceback.format_exc()}"
|
| 396 |
+
print(error_msg)
|
| 397 |
+
return False, error_msg, None
|
| 398 |
|
| 399 |
def parse_hf_url(url):
|
| 400 |
url = url.strip().rstrip("/")
|
|
|
|
| 472 |
return None, "β Hugging Face token required for source.", ""
|
| 473 |
if target_type == "huggingface" and not hf_token:
|
| 474 |
return None, "β Hugging Face token required for target.", ""
|
| 475 |
+
|
| 476 |
+
# Validate lora_rank
|
| 477 |
if lora_rank < 4:
|
| 478 |
return None, "β LoRA rank must be at least 4.", ""
|
| 479 |
+
|
| 480 |
temp_dir = None
|
| 481 |
output_dir = tempfile.mkdtemp()
|
| 482 |
try:
|
|
|
|
| 484 |
safetensors_path, temp_dir = download_safetensors_file(
|
| 485 |
source_type, repo_url, safetensors_filename, hf_token, progress
|
| 486 |
)
|
| 487 |
+
|
| 488 |
progress(0.25, desc=f"Converting to FP8 with LoRA ({architecture})...")
|
| 489 |
success, msg, stats = convert_safetensors_to_fp8_with_lora(
|
| 490 |
safetensors_path, output_dir, fp8_format, lora_rank, architecture, progress
|
| 491 |
)
|
| 492 |
+
|
| 493 |
if not success:
|
| 494 |
return None, f"β Conversion failed: {msg}", ""
|
| 495 |
+
|
| 496 |
progress(0.9, desc="Uploading...")
|
| 497 |
repo_url_final = upload_to_target(
|
| 498 |
target_type, new_repo_id, output_dir, fp8_format, architecture, hf_token, modelscope_token, private_repo
|
| 499 |
)
|
| 500 |
+
|
| 501 |
base_name = os.path.splitext(safetensors_filename)[0]
|
| 502 |
lora_filename = f"{base_name}-lora-r{lora_rank}-{architecture}.safetensors"
|
| 503 |
fp8_filename = f"{base_name}-fp8-{fp8_format}.safetensors"
|
| 504 |
+
|
| 505 |
readme = f"""---
|
| 506 |
library_name: diffusers
|
| 507 |
tags:
|
|
|
|
| 511 |
- low-rank
|
| 512 |
- diffusion
|
| 513 |
- architecture-{architecture}
|
| 514 |
+
- converted-by-ai-toolkit
|
| 515 |
---
|
| 516 |
# FP8 Model with Low-Rank LoRA
|
| 517 |
- **Source**: `{repo_url}`
|
|
|
|
| 522 |
- **LoRA File**: `{lora_filename}`
|
| 523 |
- **FP8 File**: `{fp8_filename}`
|
| 524 |
|
| 525 |
+
## Architecture Distribution
|
| 526 |
+
"""
|
| 527 |
+
|
| 528 |
+
# Add architecture stats to README if available
|
| 529 |
+
if stats and 'architecture_distro' in stats:
|
| 530 |
+
readme += "\n| Component | Layer Count |\n|-----------|------------|\n"
|
| 531 |
+
for arch, count in stats['architecture_distro'].items():
|
| 532 |
+
readme += f"| {arch.replace('_', ' ').title()} | {count} |\n"
|
| 533 |
+
|
| 534 |
+
readme += f"""
|
| 535 |
## Usage (Inference)
|
| 536 |
```python
|
| 537 |
from safetensors.torch import load_file
|
|
|
|
| 544 |
# Reconstruct approximate original weights
|
| 545 |
reconstructed = {{}}
|
| 546 |
for key in fp8_state:
|
| 547 |
+
lora_a_key = f"lora_A.{{key}}"
|
| 548 |
+
lora_b_key = f"lora_B.{{key}}"
|
| 549 |
+
|
| 550 |
+
if lora_a_key in lora_state and lora_b_key in lora_state:
|
| 551 |
+
A = lora_state[lora_a_key].to(torch.float32)
|
| 552 |
+
B = lora_state[lora_b_key].to(torch.float32)
|
| 553 |
+
|
| 554 |
+
# Handle different tensor dimensions
|
| 555 |
if A.ndim == 2 and B.ndim == 2:
|
| 556 |
lora_weight = B @ A
|
| 557 |
+
elif A.ndim == 4 and B.ndim == 4:
|
| 558 |
+
# For convolutional LoRA
|
| 559 |
+
lora_weight = F.conv2d(fp8_state[key].to(torch.float32),
|
| 560 |
+
B, padding=1) + F.conv2d(fp8_state[key].to(torch.float32),
|
| 561 |
+
A, padding=1)
|
| 562 |
else:
|
| 563 |
+
# Fallback for mixed dimension cases
|
| 564 |
+
lora_weight = B @ A.view(B.shape[1], -1)
|
| 565 |
+
if lora_weight.shape != fp8_state[key].shape:
|
| 566 |
+
lora_weight = lora_weight.view_as(fp8_state[key])
|
| 567 |
+
|
| 568 |
reconstructed[key] = fp8_state[key].to(torch.float32) + lora_weight
|
| 569 |
else:
|
| 570 |
reconstructed[key] = fp8_state[key].to(torch.float32)
|
| 571 |
```
|
| 572 |
|
| 573 |
+
> **Note**: Requires PyTorch β₯ 2.1 for FP8 support. For best results, use the same architecture selection ({architecture}) during inference as was used during extraction.
|
| 574 |
"""
|
| 575 |
+
|
| 576 |
with open(os.path.join(output_dir, "README.md"), "w") as f:
|
| 577 |
f.write(readme)
|
| 578 |
+
|
| 579 |
if target_type == "huggingface":
|
| 580 |
HfApi(token=hf_token).upload_file(
|
| 581 |
path_or_fileobj=os.path.join(output_dir, "README.md"),
|
|
|
|
| 584 |
repo_type="model",
|
| 585 |
token=hf_token
|
| 586 |
)
|
| 587 |
+
|
| 588 |
progress(1.0, desc="β
Done!")
|
| 589 |
result_html = f"""
|
| 590 |
β
Success!
|
| 591 |
Model uploaded to: <a href="{repo_url_final}" target="_blank">{new_repo_id}</a>
|
| 592 |
+
Includes:
|
| 593 |
+
- FP8 model: `{fp8_filename}`
|
| 594 |
+
- LoRA weights: `{lora_filename}` (rank {lora_rank}, architecture: {architecture})
|
| 595 |
+
|
| 596 |
+
π Stats: {stats['layers_processed']}/{stats['layers_eligible']} eligible layers processed
|
| 597 |
"""
|
| 598 |
+
if 'avg_reconstruction_error' in stats:
|
| 599 |
+
result_html += f"<br>Avg reconstruction error: {stats['avg_reconstruction_error']:.6f}"
|
| 600 |
+
|
| 601 |
return gr.HTML(result_html), "β
FP8 + LoRA upload successful!", msg
|
| 602 |
+
|
| 603 |
except Exception as e:
|
| 604 |
+
error_msg = f"β Error: {str(e)}\n{traceback.format_exc()}"
|
| 605 |
+
print(error_msg)
|
| 606 |
+
return None, error_msg, ""
|
| 607 |
+
|
| 608 |
finally:
|
| 609 |
if temp_dir:
|
| 610 |
shutil.rmtree(temp_dir, ignore_errors=True)
|
|
|
|
| 612 |
|
| 613 |
with gr.Blocks(title="FP8 + LoRA Extractor (HF β ModelScope)") as demo:
|
| 614 |
gr.Markdown("# π Advanced FP8 Pruner with Architecture-Specific LoRA Extraction")
|
| 615 |
+
gr.Markdown("Convert `.safetensors` β **FP8** + **targeted LoRA** weights for precision recovery. Supports Hugging Face β ModelScope.")
|
| 616 |
+
|
| 617 |
with gr.Row():
|
| 618 |
with gr.Column():
|
| 619 |
source_type = gr.Radio(["huggingface", "modelscope"], value="huggingface", label="Source")
|
| 620 |
repo_url = gr.Textbox(label="Repo URL or ID", placeholder="https://huggingface.co/... or modelscope-id")
|
| 621 |
safetensors_filename = gr.Textbox(label="Filename", placeholder="model.safetensors")
|
| 622 |
+
|
| 623 |
with gr.Accordion("Advanced LoRA Settings", open=True):
|
| 624 |
fp8_format = gr.Radio(["e4m3fn", "e5m2"], value="e5m2", label="FP8 Format")
|
| 625 |
lora_rank = gr.Slider(minimum=4, maximum=256, step=4, value=64, label="LoRA Rank")
|
| 626 |
+
|
| 627 |
architecture = gr.Dropdown(
|
| 628 |
choices=[
|
| 629 |
("Auto-detect components", "auto"),
|
|
|
|
| 634 |
("All components", "all")
|
| 635 |
],
|
| 636 |
value="auto",
|
| 637 |
+
label="Target Architecture",
|
| 638 |
+
info="Select which model components to apply LoRA to"
|
| 639 |
)
|
| 640 |
+
|
| 641 |
with gr.Accordion("Authentication", open=False):
|
| 642 |
hf_token = gr.Textbox(label="Hugging Face Token", type="password")
|
| 643 |
modelscope_token = gr.Textbox(label="ModelScope Token (optional)", type="password", visible=MODELScope_AVAILABLE)
|
| 644 |
+
|
| 645 |
with gr.Column():
|
| 646 |
target_type = gr.Radio(["huggingface", "modelscope"], value="huggingface", label="Target")
|
| 647 |
new_repo_id = gr.Textbox(label="New Repo ID", placeholder="user/model-fp8-lora")
|
| 648 |
private_repo = gr.Checkbox(label="Private Repository (HF only)", value=False)
|
| 649 |
+
|
| 650 |
status_output = gr.Markdown()
|
| 651 |
detailed_log = gr.Textbox(label="Processing Log", interactive=False, lines=10)
|
| 652 |
+
|
| 653 |
convert_btn = gr.Button("π Convert & Upload", variant="primary")
|
| 654 |
repo_link_output = gr.HTML()
|
| 655 |
+
|
| 656 |
convert_btn.click(
|
| 657 |
fn=process_and_upload_fp8,
|
| 658 |
inputs=[
|
|
|
|
| 671 |
outputs=[repo_link_output, status_output, detailed_log],
|
| 672 |
show_progress=True
|
| 673 |
)
|
| 674 |
+
|
| 675 |
gr.Examples(
|
| 676 |
examples=[
|
| 677 |
["huggingface", "https://huggingface.co/Yabo/FramePainter/tree/main", "unet_diffusion_pytorch_model.safetensors", "e5m2", 64, "unet_transformer"],
|
|
|
|
| 681 |
inputs=[source_type, repo_url, safetensors_filename, fp8_format, lora_rank, architecture],
|
| 682 |
label="Example Conversions"
|
| 683 |
)
|
| 684 |
+
|
| 685 |
gr.Markdown("""
|
| 686 |
## π‘ Usage Tips
|
| 687 |
+
|
| 688 |
+
- **For Text Encoders**: Use rank 32-64 with `text_encoder` architecture for optimal results.
|
| 689 |
+
- **For UNet Attention**: Use `unet_transformer` with rank 64-128 for best quality preservation.
|
| 690 |
+
- **For UNet Convolutions**: Use `unet_conv` with lower ranks (16-32) as convolutions compress better.
|
| 691 |
+
- **For VAE**: Use `vae` architecture with rank 16-32.
|
| 692 |
+
- **Auto Mode**: Let the tool analyze and target appropriate layers automatically.
|
| 693 |
+
|
| 694 |
+
β οΈ **Note**: Higher ranks produce better quality but larger LoRA files. Start with lower ranks and increase if needed.
|
| 695 |
""")
|
| 696 |
|
| 697 |
demo.launch()
|