Spaces:
Sleeping
Sleeping
File size: 27,416 Bytes
7f615fd 9f9518a 7f615fd fdd626d 672b8b5 0d6c60b 672b8b5 9f9518a 672b8b5 c31eee4 9f9518a 672b8b5 9f9518a 672b8b5 9f9518a 672b8b5 6bcc1a3 c31eee4 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 9f9518a 672b8b5 7f615fd 2a57dcf 7f615fd 1b040ff 2a57dcf 9f9518a 1b040ff 2a57dcf 672b8b5 2a57dcf 1b040ff 9f9518a c31eee4 9f9518a 672b8b5 9f9518a 672b8b5 9f9518a c31eee4 672b8b5 6bcc1a3 2a57dcf 672b8b5 c31eee4 9f9518a 9efc461 9f9518a 2a57dcf 1b040ff 672b8b5 2a57dcf 672b8b5 9f9518a 672b8b5 9f9518a c31eee4 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 7f615fd 1b040ff 9f9518a 2a57dcf 672b8b5 7f615fd 1b040ff 2a57dcf 1b040ff 2a57dcf 672b8b5 9f9518a c31eee4 9f9518a c31eee4 9f9518a 2a57dcf c31eee4 2a57dcf 672b8b5 9f9518a c31eee4 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 9f9518a 7f615fd 9f9518a 7f615fd 7153add 1b040ff 0d6c60b 7153add 0d6c60b 7153add 0d6c60b 7153add 0d6c60b 7153add 0d6c60b 1b040ff 0d6c60b 1b040ff 0d6c60b 9efc461 0d6c60b 1b040ff 0d6c60b 1b040ff 0d6c60b 1b040ff 0d6c60b 9f9518a 0d6c60b 7f615fd 1b040ff 9f9518a 2a57dcf 0d6c60b 7f615fd 1b040ff 0d6c60b 1b040ff 7f615fd 2a57dcf c31eee4 9f9518a fdd626d 2a57dcf 7f615fd 2a57dcf 1b040ff 0d6c60b 9efc461 7f615fd 2a57dcf 7f615fd 6bcc1a3 2a57dcf c31eee4 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 c31eee4 7f615fd c31eee4 7f615fd 9efc461 7f615fd c31eee4 1b040ff 9f9518a c31eee4 672b8b5 6bcc1a3 672b8b5 1b040ff fdd626d 672b8b5 9f9518a 6bcc1a3 672b8b5 c31eee4 fdd626d 1b040ff 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 1b040ff 672b8b5 1b040ff 672b8b5 7f615fd 672b8b5 1b040ff 7f615fd 2a57dcf 0d6c60b 1b040ff 0d6c60b 2a57dcf 7f615fd 672b8b5 7f615fd 1b040ff 672b8b5 6bcc1a3 672b8b5 c31eee4 2a57dcf 7f615fd 672b8b5 2a57dcf 7f615fd 0d6c60b 7f615fd c31eee4 672b8b5 2a57dcf 7f615fd 1b040ff 2a57dcf 9f9518a 6bcc1a3 672b8b5 9f9518a 6bcc1a3 c31eee4 672b8b5 c31eee4 6bcc1a3 9f9518a 6bcc1a3 2a57dcf 6bcc1a3 672b8b5 2a57dcf 7f615fd 1b040ff 672b8b5 6bcc1a3 2a57dcf 6bcc1a3 2a57dcf 7f615fd 1b040ff 2a57dcf 7f615fd 0d6c60b 9f9518a 0d6c60b 6bcc1a3 7f615fd 2a57dcf 7f615fd 672b8b5 7f615fd 672b8b5 6bcc1a3 7f615fd 2a57dcf 6bcc1a3 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 c31eee4 672b8b5 2a57dcf 672b8b5 2a57dcf 672b8b5 6bcc1a3 7f615fd |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 |
import gradio as gr
import os
import tempfile
import shutil
import re
import json
import datetime
from pathlib import Path
from huggingface_hub import HfApi, hf_hub_download
from safetensors.torch import load_file, save_file
import torch
import torch.nn.functional as F
import traceback
import math
try:
from modelscope.hub.file_download import model_file_download as ms_file_download
from modelscope.hub.api import HubApi as ModelScopeApi
MODELScope_AVAILABLE = True
except ImportError:
MODELScope_AVAILABLE = False
def get_fp8_dtype(fp8_format):
"""Get torch FP8 dtype."""
if fp8_format == "e5m2":
return torch.float8_e5m2
else:
return torch.float8_e4m3fn
def quantize_and_get_error(weight, fp8_dtype):
"""Quantize weight to FP8 and return both quantized weight and error."""
weight_fp8 = weight.to(fp8_dtype)
weight_dequantized = weight_fp8.to(weight.dtype)
error = weight - weight_dequantized
return weight_fp8, error
def low_rank_decomposition_error(error_tensor, rank=32, min_error_threshold=1e-6):
"""Decompose error tensor with proper rank reduction."""
if error_tensor.ndim not in [2, 4]:
return None, None
try:
# Calculate error magnitude
error_norm = torch.norm(error_tensor.float())
if error_norm < min_error_threshold:
return None, None
# For 2D tensors (linear layers)
if error_tensor.ndim == 2:
U, S, Vh = torch.linalg.svd(error_tensor.float(), full_matrices=False)
# Calculate rank based on variance explained (keep 95% of error)
total_variance = torch.sum(S ** 2)
cumulative = torch.cumsum(S ** 2, dim=0)
keep_components = torch.sum(cumulative <= 0.95 * total_variance).item() + 1
# Limit rank to much smaller than original
max_rank = min(error_tensor.shape)
actual_rank = min(rank, keep_components, max_rank // 2)
if actual_rank < 2:
return None, None
A = Vh[:actual_rank, :].contiguous()
B = U[:, :actual_rank] @ torch.diag(S[:actual_rank]).contiguous()
return A, B
# For 4D convolutions
elif error_tensor.ndim == 4:
out_ch, in_ch, kH, kW = error_tensor.shape
# Reshape to 2D for decomposition
error_2d = error_tensor.view(out_ch, in_ch * kH * kW)
U, S, Vh = torch.linalg.svd(error_2d.float(), full_matrices=False)
# Calculate rank based on variance explained (90% for conv)
total_variance = torch.sum(S ** 2)
cumulative = torch.cumsum(S ** 2, dim=0)
keep_components = torch.sum(cumulative <= 0.90 * total_variance).item() + 1
# Use even lower rank for conv
max_rank = min(error_2d.shape)
actual_rank = min(rank // 2, keep_components, max_rank // 4)
if actual_rank < 2:
return None, None
A = Vh[:actual_rank, :].contiguous()
B = U[:, :actual_rank] @ torch.diag(S[:actual_rank]).contiguous()
# Reshape back for convolutional format
if kH == 1 and kW == 1:
B = B.view(out_ch, actual_rank, 1, 1)
A = A.view(actual_rank, in_ch, 1, 1)
else:
B = B.view(out_ch, actual_rank, 1, 1)
A = A.view(actual_rank, in_ch, kH, kW)
return A, B
except Exception as e:
print(f"Error decomposition failed: {e}")
return None, None
def extract_correction_factors(original_weight, fp8_weight):
"""Extract simple correction factors for VAE."""
with torch.no_grad():
orig = original_weight.float()
quant = fp8_weight.float()
error = orig - quant
error_norm = torch.norm(error)
orig_norm = torch.norm(orig)
if orig_norm > 1e-6 and error_norm / orig_norm < 0.001:
return None
# For 4D tensors (VAE), compute per-channel correction
if orig.ndim == 4:
channel_mean = error.mean(dim=tuple(i for i in range(1, orig.ndim)), keepdim=True)
return channel_mean.to(original_weight.dtype)
elif orig.ndim == 2:
row_mean = error.mean(dim=1, keepdim=True)
return row_mean.to(original_weight.dtype)
else:
return error.mean().to(original_weight.dtype)
def get_architecture_settings(architecture, base_rank):
"""Get optimal settings for different architectures."""
settings = {
"text_encoder": {
"rank": base_rank,
"error_threshold": 5e-5,
"min_rank": 8,
"max_rank_factor": 0.4,
"method": "lora"
},
"transformer": {
"rank": base_rank,
"error_threshold": 1e-5,
"min_rank": 12,
"max_rank_factor": 0.35,
"method": "lora"
},
"vae": {
"rank": base_rank // 2,
"error_threshold": 1e-4,
"min_rank": 4,
"max_rank_factor": 0.3,
"method": "correction"
},
"unet_conv": {
"rank": base_rank // 3,
"error_threshold": 2e-5,
"min_rank": 8,
"max_rank_factor": 0.25,
"method": "lora"
},
"auto": {
"rank": base_rank,
"error_threshold": 1e-5,
"min_rank": 8,
"max_rank_factor": 0.3,
"method": "lora"
},
"all": {
"rank": base_rank,
"error_threshold": 1e-5,
"min_rank": 8,
"max_rank_factor": 0.3,
"method": "lora"
}
}
return settings.get(architecture, settings["auto"])
def should_process_layer(key, weight, architecture):
"""Determine if layer should be processed for LoRA/correction."""
lower_key = key.lower()
# Skip biases and normalization layers
if 'bias' in key or 'norm' in key.lower() or 'bn' in key.lower():
return False
if weight.numel() < 100:
return False
# Architecture-specific filtering
if architecture == "text_encoder":
return ('text' in lower_key or 'emb' in lower_key or
'encoder' in lower_key or 'attn' in lower_key)
elif architecture == "transformer":
return ('attn' in lower_key or 'transformer' in lower_key or
'mlp' in lower_key or 'to_out' in lower_key)
elif architecture == "vae":
return ('vae' in lower_key or 'encoder' in lower_key or
'decoder' in lower_key or 'conv' in lower_key)
elif architecture == "unet_conv":
return ('conv' in lower_key or 'resnet' in lower_key or
'downsample' in lower_key or 'upsample' in lower_key)
elif architecture in ["all", "auto"]:
return True
return False
def convert_safetensors_to_fp8_with_lora(safetensors_path, output_dir, fp8_format, lora_rank=128, architecture="auto", progress=gr.Progress()):
progress(0.1, desc="Starting FP8 conversion with error recovery...")
try:
def read_safetensors_metadata(path):
with open(path, 'rb') as f:
header_size = int.from_bytes(f.read(8), 'little')
header_json = f.read(header_size).decode('utf-8')
header = json.loads(header_json)
return header.get('__metadata__', {})
metadata = read_safetensors_metadata(safetensors_path)
progress(0.2, desc="Loaded metadata.")
state_dict = load_file(safetensors_path)
progress(0.4, desc="Loaded weights.")
# Auto-detect architecture if needed
if architecture == "auto":
model_keys = " ".join(state_dict.keys()).lower()
if "vae" in model_keys or ("encoder" in model_keys and "decoder" in model_keys):
architecture = "vae"
elif "text" in model_keys or "emb" in model_keys:
architecture = "text_encoder"
elif "attn" in model_keys or "transformer" in model_keys:
architecture = "transformer"
elif "conv" in model_keys or "resnet" in model_keys:
architecture = "unet_conv"
else:
architecture = "all"
settings = get_architecture_settings(architecture, lora_rank)
fp8_dtype = get_fp8_dtype(fp8_format)
sd_fp8 = {}
lora_weights = {}
correction_factors = {}
stats = {
"total_layers": len(state_dict),
"eligible_layers": 0,
"layers_with_error": 0,
"processed_layers": 0,
"correction_layers": 0,
"skipped_layers": [],
"architecture": architecture,
"method": settings["method"],
"error_magnitudes": []
}
total = len(state_dict)
for i, key in enumerate(state_dict):
progress(0.4 + 0.4 * (i / total), desc=f"Processing {i+1}/{total}...")
weight = state_dict[key]
if weight.dtype in [torch.float16, torch.float32, torch.bfloat16]:
# Quantize to FP8 and calculate error
weight_fp8, error = quantize_and_get_error(weight, fp8_dtype)
sd_fp8[key] = weight_fp8
# Calculate error magnitude
error_norm = torch.norm(error.float())
weight_norm = torch.norm(weight.float())
relative_error = (error_norm / weight_norm).item() if weight_norm > 0 else 0
stats["error_magnitudes"].append({
"key": key,
"relative_error": relative_error
})
# Check if layer should be processed
should_process = should_process_layer(key, weight, architecture)
if should_process:
stats["eligible_layers"] += 1
# Only process if error is significant
if relative_error > settings["error_threshold"]:
stats["layers_with_error"] += 1
if settings["method"] == "correction":
# Use correction factors for VAE
correction = extract_correction_factors(weight, weight_fp8)
if correction is not None:
correction_factors[f"correction.{key}"] = correction
stats["correction_layers"] += 1
stats["processed_layers"] += 1
else:
# Use LoRA decomposition for other architectures
try:
A, B = low_rank_decomposition_error(
error,
rank=settings["rank"],
min_error_threshold=settings["error_threshold"]
)
if A is not None and B is not None:
lora_weights[f"lora_A.{key}"] = A.to(torch.float16)
lora_weights[f"lora_B.{key}"] = B.to(torch.float16)
stats["processed_layers"] += 1
else:
stats["skipped_layers"].append(f"{key}: decomposition failed")
except Exception as e:
stats["skipped_layers"].append(f"{key}: error - {str(e)}")
else:
stats["skipped_layers"].append(f"{key}: error too small ({relative_error:.6f})")
else:
sd_fp8[key] = weight
stats["skipped_layers"].append(f"{key}: non-float dtype")
# Calculate average error
if stats["error_magnitudes"]:
errors = [e["relative_error"] for e in stats["error_magnitudes"]]
stats["avg_error"] = sum(errors) / len(errors) if errors else 0
stats["max_error"] = max(errors) if errors else 0
base_name = os.path.splitext(os.path.basename(safetensors_path))[0]
fp8_path = os.path.join(output_dir, f"{base_name}-fp8-{fp8_format}.safetensors")
save_file(sd_fp8, fp8_path, metadata={"format": "pt", "fp8_format": fp8_format, **metadata})
# Save precision recovery weights
if lora_weights:
lora_path = os.path.join(output_dir, f"{base_name}-lora-r{lora_rank}-{architecture}.safetensors")
lora_metadata = {
"format": "pt",
"lora_rank": str(lora_rank),
"architecture": architecture,
"stats": json.dumps(stats),
"method": "lora"
}
save_file(lora_weights, lora_path, metadata=lora_metadata)
if correction_factors:
correction_path = os.path.join(output_dir, f"{base_name}-correction-{architecture}.safetensors")
correction_metadata = {
"format": "pt",
"architecture": architecture,
"stats": json.dumps(stats),
"method": "correction"
}
save_file(correction_factors, correction_path, metadata=correction_metadata)
progress(0.9, desc="Saved FP8 and precision recovery files.")
progress(1.0, desc="β
FP8 + precision recovery extraction complete!")
stats_msg = f"FP8 ({fp8_format}) with precision recovery saved.\n"
stats_msg += f"Architecture: {architecture}\n"
stats_msg += f"Method: {settings['method']}\n"
stats_msg += f"Average quantization error: {stats.get('avg_error', 0):.6f}\n"
if settings["method"] == "correction":
stats_msg += f"Correction factors generated for {stats['correction_layers']} layers."
else:
stats_msg += f"LoRA generated for {stats['processed_layers']}/{stats['eligible_layers']} eligible layers (rank {lora_rank})."
if stats['processed_layers'] == 0 and stats['correction_layers'] == 0:
stats_msg += "\nβ οΈ No precision recovery weights were generated. FP8 quantization error may be too small."
return True, stats_msg, stats
except Exception as e:
error_msg = f"Error: {str(e)}\n{traceback.format_exc()}"
return False, error_msg, None
def parse_hf_url(url):
url = url.strip().rstrip("/")
if not url.startswith("https://huggingface.co/"):
raise ValueError("URL must start with https://huggingface.co/")
path = url.replace("https://huggingface.co/", "")
parts = path.split("/")
if len(parts) < 2:
raise ValueError("Invalid repo format")
repo_id = "/".join(parts[:2])
subfolder = ""
if len(parts) > 3 and parts[2] == "tree":
subfolder = "/".join(parts[4:]) if len(parts) > 4 else ""
elif len(parts) > 2:
subfolder = "/".join(parts[2:])
return repo_id, subfolder
def download_safetensors_file(source_type, repo_url, filename, hf_token=None, progress=gr.Progress()):
temp_dir = tempfile.mkdtemp()
try:
if source_type == "huggingface":
repo_id, subfolder = parse_hf_url(repo_url)
safetensors_path = hf_hub_download(
repo_id=repo_id,
filename=filename,
subfolder=subfolder or None,
cache_dir=temp_dir,
token=hf_token,
resume_download=True
)
elif source_type == "modelscope":
if not MODELScope_AVAILABLE:
raise ImportError("ModelScope not installed")
repo_id = repo_url.strip()
safetensors_path = ms_file_download(model_id=repo_id, file_path=filename)
else:
raise ValueError("Unknown source")
return safetensors_path, temp_dir
except Exception as e:
shutil.rmtree(temp_dir, ignore_errors=True)
raise e
def upload_to_target(target_type, new_repo_id, output_dir, fp8_format, hf_token=None, modelscope_token=None, private_repo=False):
if target_type == "huggingface":
api = HfApi(token=hf_token)
api.create_repo(repo_id=new_repo_id, private=private_repo, repo_type="model", exist_ok=True)
api.upload_folder(repo_id=new_repo_id, folder_path=output_dir, repo_type="model", token=hf_token)
return f"https://huggingface.co/{new_repo_id}"
elif target_type == "modelscope":
api = ModelScopeApi()
if modelscope_token:
api.login(modelscope_token)
api.push_model(model_id=new_repo_id, model_dir=output_dir)
return f"https://modelscope.cn/models/{new_repo_id}"
else:
raise ValueError("Unknown target")
def process_and_upload_fp8(
source_type,
repo_url,
safetensors_filename,
fp8_format,
lora_rank,
architecture,
target_type,
new_repo_id,
hf_token,
modelscope_token,
private_repo,
progress=gr.Progress()
):
if not re.match(r"^[a-zA-Z0-9._-]+/[a-zA-Z0-9._-]+$", new_repo_id):
return None, "β Invalid repo ID format. Use 'username/model-name'.", ""
if source_type == "huggingface" and not hf_token:
return None, "β Hugging Face token required for source.", ""
if target_type == "huggingface" and not hf_token:
return None, "β Hugging Face token required for target.", ""
if lora_rank < 8:
return None, "β LoRA rank must be at least 8.", ""
temp_dir = None
output_dir = tempfile.mkdtemp()
try:
progress(0.05, desc="Downloading model...")
safetensors_path, temp_dir = download_safetensors_file(
source_type, repo_url, safetensors_filename, hf_token, progress
)
progress(0.25, desc="Converting to FP8 with precision recovery...")
success, msg, stats = convert_safetensors_to_fp8_with_lora(
safetensors_path, output_dir, fp8_format, lora_rank, architecture, progress
)
if not success:
return None, f"β Conversion failed: {msg}", ""
progress(0.9, desc="Uploading...")
repo_url_final = upload_to_target(
target_type, new_repo_id, output_dir, fp8_format, hf_token, modelscope_token, private_repo
)
base_name = os.path.splitext(safetensors_filename)[0]
fp8_filename = f"{base_name}-fp8-{fp8_format}.safetensors"
# Determine which precision recovery file was generated
precision_recovery_file = ""
precision_recovery_type = ""
if stats.get("method") == "correction" and stats.get("correction_layers", 0) > 0:
precision_recovery_file = f"{base_name}-correction-{architecture}.safetensors"
precision_recovery_type = "Correction Factors"
elif stats.get("method") == "lora" and stats.get("processed_layers", 0) > 0:
precision_recovery_file = f"{base_name}-lora-r{lora_rank}-{architecture}.safetensors"
precision_recovery_type = "LoRA"
readme = f"""---
library_name: diffusers
tags:
- fp8
- safetensors
- precision-recovery
- diffusion
- converted-by-gradio
---
# FP8 Model with Precision Recovery
- **Source**: `{repo_url}`
- **File**: `{safetensors_filename}`
- **FP8 Format**: `{fp8_format.upper()}`
- **Architecture**: {architecture}
- **Precision Recovery Type**: {precision_recovery_type}
- **Precision Recovery File**: `{precision_recovery_file}` if available
- **FP8 File**: `{fp8_filename}`
## Usage (Inference)
```python
from safetensors.torch import load_file
import torch
# Load FP8 model
fp8_state = load_file("{fp8_filename}")
# Load precision recovery file if available
recovery_state = {{}}
if "{precision_recovery_file}":
recovery_state = load_file("{precision_recovery_file}")
# Reconstruct high-precision weights
reconstructed = {{}}
for key in fp8_state:
# Dequantize FP8 to target precision
fp_weight = fp8_state[key].to(torch.float32)
if recovery_state:
# For LoRA approach
if f"lora_A.{{key}}" in recovery_state and f"lora_B.{{key}}" in recovery_state:
A = recovery_state[f"lora_A.{{key}}"].to(torch.float32)
B = recovery_state[f"lora_B.{{key}}"].to(torch.float32)
error_correction = B @ A
reconstructed[key] = fp_weight + error_correction
# For correction factor approach
elif f"correction.{{key}}" in recovery_state:
correction = recovery_state[f"correction.{{key}}"].to(torch.float32)
reconstructed[key] = fp_weight + correction
else:
reconstructed[key] = fp_weight
else:
reconstructed[key] = fp_weight
print("Model reconstructed with FP8 error recovery")
```
> **Note**: This precision recovery targets FP8 quantization errors.
> Average quantization error: {stats.get('avg_error', 0):.6f}
"""
with open(os.path.join(output_dir, "README.md"), "w") as f:
f.write(readme)
if target_type == "huggingface":
HfApi(token=hf_token).upload_file(
path_or_fileobj=os.path.join(output_dir, "README.md"),
path_in_repo="README.md",
repo_id=new_repo_id,
repo_type="model",
token=hf_token
)
progress(1.0, desc="β
Done!")
result_html = f"""
β
Success!
Model uploaded to: <a href="{repo_url_final}" target="_blank">{new_repo_id}</a>
Includes: FP8 model + precision recovery ({precision_recovery_type}).
Average quantization error: {stats.get('avg_error', 0):.6f}
"""
if stats['processed_layers'] > 0 or stats['correction_layers'] > 0:
result_html += f"<br>Precision recovery applied to {stats['processed_layers'] + stats['correction_layers']} layers."
return gr.HTML(result_html), "β
FP8 + precision recovery upload successful!", msg
except Exception as e:
error_msg = f"β Error: {str(e)}\n{traceback.format_exc()}"
return None, error_msg, ""
finally:
if temp_dir:
shutil.rmtree(temp_dir, ignore_errors=True)
shutil.rmtree(output_dir, ignore_errors=True)
with gr.Blocks(title="FP8 + Precision Recovery Extractor") as demo:
gr.Markdown("# π FP8 Converter with Architecture-Specific Precision Recovery")
gr.Markdown("Convert models to **FP8** with **error-based precision recovery**.")
with gr.Row():
with gr.Column():
source_type = gr.Radio(["huggingface", "modelscope"], value="huggingface", label="Source")
repo_url = gr.Textbox(label="Repo URL or ID", placeholder="https://huggingface.co/... or modelscope-id")
safetensors_filename = gr.Textbox(label="Filename", placeholder="model.safetensors")
with gr.Accordion("Advanced Settings", open=True):
fp8_format = gr.Radio(["e4m3fn", "e5m2"], value="e5m2", label="FP8 Format")
lora_rank = gr.Slider(minimum=8, maximum=256, step=8, value=128,
label="LoRA Rank (for text/transformers)")
architecture = gr.Dropdown(
choices=[
("Auto-detect architecture", "auto"),
("Text Encoder (LoRA)", "text_encoder"),
("Transformer blocks (LoRA)", "transformer"),
("VAE (Correction Factors)", "vae"),
("UNet Convolutions (LoRA)", "unet_conv"),
("All layers (LoRA where applicable)", "all")
],
value="auto",
label="Target Architecture"
)
with gr.Accordion("Authentication", open=False):
hf_token = gr.Textbox(label="Hugging Face Token", type="password")
modelscope_token = gr.Textbox(label="ModelScope Token (optional)", type="password",
visible=MODELScope_AVAILABLE)
with gr.Column():
target_type = gr.Radio(["huggingface", "modelscope"], value="huggingface", label="Target")
new_repo_id = gr.Textbox(label="New Repo ID", placeholder="user/model-fp8-precision")
private_repo = gr.Checkbox(label="Private Repository (HF only)", value=False)
status_output = gr.Markdown()
detailed_log = gr.Textbox(label="Processing Log", interactive=False, lines=10)
convert_btn = gr.Button("π Convert & Upload", variant="primary")
repo_link_output = gr.HTML()
convert_btn.click(
fn=process_and_upload_fp8,
inputs=[
source_type,
repo_url,
safetensors_filename,
fp8_format,
lora_rank,
architecture,
target_type,
new_repo_id,
hf_token,
modelscope_token,
private_repo
],
outputs=[repo_link_output, status_output, detailed_log],
show_progress=True
)
gr.Examples(
examples=[
["huggingface", "https://huggingface.co/runwayml/stable-diffusion-v1-5/tree/main/text_encoder",
"model.safetensors", "e5m2", 96, "text_encoder"],
["huggingface", "https://huggingface.co/stabilityai/sdxl-vae",
"diffusion_pytorch_model.safetensors", "e4m3fn", 64, "vae"],
["huggingface", "https://huggingface.co/Yabo/FramePainter/tree/main",
"unet_diffusion_pytorch_model.safetensors", "e5m2", 128, "transformer"]
],
inputs=[source_type, repo_url, safetensors_filename, fp8_format, lora_rank, architecture],
label="Example Conversions"
)
gr.Markdown("""
## π― What This Tool Does
Unlike traditional LoRA fine-tuning, this tool:
1. **Quantizes** the model to FP8 (loses precision)
2. **Measures** the quantization error for each weight
3. **Extracts recovery weights** that specifically recover this error
4. **Only applies** recovery where error is significant (>0.001%)
## π‘ Recommended Settings
- **Text Encoders**: rank 64-96 (text is sensitive)
- **Transformers**: rank 96-128
- **VAE**: Uses correction factors (no rank needed)
- **UNet Convolutions**: rank 32-64
## β οΈ Important Notes
- This recovers **FP8 quantization errors**, not fine-tuning changes
- If FP8 error is tiny (<0.0001%), recovery may not be generated
- Higher rank β better for error recovery (use recommended ranges)
""")
demo.launch() |