Spaces:
Running
Running
File size: 35,899 Bytes
36079e7 98e130a 36079e7 98e130a 36079e7 98e130a 36079e7 98e130a 36079e7 98e130a 36079e7 98e130a 36079e7 98e130a 36079e7 98e130a 36079e7 98e130a 36079e7 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 |
"""
Architecture parser - produces LINEAR PIPELINE representation of transformer models.
Shows the sequential flow of data through the model as a flowchart.
"""
import re
from typing import Dict, Any, List, Optional, Tuple
from collections import OrderedDict
import torch
import torch.nn as nn
from transformers import AutoConfig, AutoModel, AutoModelForCausalLM, AutoModelForSeq2SeqLM
# Monkeypatch for transformers import issues in some environment/model combinations
try:
import transformers.utils.import_utils as import_utils
if not hasattr(import_utils, "is_torch_fx_available"):
import_utils.is_torch_fx_available = lambda: False
except (ImportError, AttributeError):
pass
def format_params(count: int) -> str:
"""Format parameter count in human-readable form."""
if count >= 1e12:
return f"{count / 1e12:.2f}T"
elif count >= 1e9:
return f"{count / 1e9:.2f}B"
elif count >= 1e6:
return f"{count / 1e6:.2f}M"
elif count >= 1e3:
return f"{count / 1e3:.2f}K"
else:
return str(count)
def get_module_type(module: nn.Module, name: str) -> str:
"""Infer module type from class name and module name."""
class_name = module.__class__.__name__.lower()
name_lower = name.lower()
# Check if this is a model wrapper (contains "model" in class name) - should be treated as module
is_model_wrapper = 'model' in class_name and ('for' in class_name or class_name.endswith('model'))
if is_model_wrapper:
return 'module'
if 'embedding' in class_name:
return 'embedding'
elif 'attention' in class_name or 'attn' in class_name:
return 'attention'
elif 'mlp' in class_name or 'feedforward' in class_name or 'ffn' in class_name:
return 'mlp'
elif 'layernorm' in class_name or 'rmsnorm' in class_name:
return 'norm'
elif 'linear' in class_name:
return 'linear'
elif 'conv' in class_name:
return 'linear'
elif 'dropout' in class_name:
return 'dropout'
elif 'pool' in class_name:
return 'pooler'
elif 'head' in class_name or 'lm_head' in name_lower:
return 'head'
# Check for MoE/expert - but only for actual MoE layers, not model wrappers
elif ('expert' in class_name or 'moe' in class_name) and 'layer' in class_name:
return 'mlp'
elif 'expert' in class_name and 'model' not in class_name:
return 'mlp'
# Check name patterns
if 'embed' in name_lower:
return 'embedding'
elif 'attn' in name_lower or 'attention' in name_lower:
return 'attention'
elif 'mlp' in name_lower or 'fc' in name_lower or 'ffn' in name_lower:
return 'mlp'
elif 'norm' in name_lower or 'ln' in name_lower:
return 'norm'
elif 'head' in name_lower:
return 'head'
elif 'expert' in name_lower and 'model' not in name_lower:
return 'mlp'
return 'module'
def count_parameters(module: nn.Module) -> int:
"""Count all parameters in a module recursively."""
return sum(p.numel() for p in module.parameters())
def humanize_name(name: str) -> str:
"""Convert module name to human-readable format."""
# Handle indexed names like "0", "1" etc
if name.isdigit():
return f"Layer {name}"
# Convert snake_case to Title Case
name = name.replace('_', ' ')
# Handle common abbreviations
replacements = {
'Wte': 'Token Embedding',
'Wpe': 'Position Embedding',
'Ln F': 'Final LayerNorm',
'Ln 1': 'LayerNorm 1',
'Ln 2': 'LayerNorm 2',
'Attn': 'Attention',
'Mlp': 'MLP',
'Lm Head': 'LM Head',
'Q Proj': 'Query',
'K Proj': 'Key',
'V Proj': 'Value',
'O Proj': 'Output',
'Out Proj': 'Output',
'C Attn': 'QKV Projection',
'C Proj': 'Output Projection',
'C Fc': 'Up Projection',
'Up Proj': 'Up Projection',
'Down Proj': 'Down Projection',
'Gate Proj': 'Gate Projection',
}
result = name.title()
for old, new in replacements.items():
result = result.replace(old, new)
return result
def is_modality_encoder(name: str, module: nn.Module) -> bool:
"""
Check if a module is a separate MODALITY encoder (vision tower, audio encoder, etc.)
This should only match top-level modality-specific encoders, not internal components.
"""
name_lower = name.lower()
class_lower = module.__class__.__name__.lower()
# Specific patterns for modality encoders (must have modality keyword)
modality_keywords = ['vision', 'image', 'audio', 'video', 'visual', 'pixel']
# Must contain a modality keyword
has_modality = any(kw in name_lower or kw in class_lower for kw in modality_keywords)
if not has_modality:
return False
# And should be a substantial module (tower, model, encoder)
structure_keywords = ['tower', 'model', 'encoder', 'backbone']
has_structure = any(kw in name_lower or kw in class_lower for kw in structure_keywords)
# Or just "vision_tower", "image_encoder" style names
return has_structure or name_lower in ['vision', 'visual', 'image']
def extract_pipeline_steps(module: nn.Module, name: str, depth: int = 0, max_depth: int = 4, detect_parallel: bool = True) -> List[Dict[str, Any]]:
"""
Extract pipeline steps from a module.
Handles both linear and parallel (multimodal) architectures.
Returns a list of steps where parallel branches are marked.
detect_parallel: Only look for parallel modality encoders at top level (depth 0-1)
"""
steps = []
children = list(module.named_children())
if not children:
return steps
# Categorize children
embeddings = []
vision_modules = [] # Vision tower, projector
language_model = None # Main language model
layer_container = None
layer_list = []
norms = []
heads = []
others = []
for child_name, child_module in children:
child_params = count_parameters(child_module)
if child_params == 0:
continue
child_type = get_module_type(child_module, child_name)
name_lower = child_name.lower()
class_lower = child_module.__class__.__name__.lower()
# Detect multimodal components at appropriate depth
if detect_parallel and depth <= 1:
# Vision tower or projector
if is_modality_encoder(child_name, child_module) or 'projector' in name_lower or 'projector' in class_lower:
vision_modules.append((child_name, child_module))
continue
# Main language model (separate from vision)
if 'language_model' in name_lower or 'text_model' in name_lower:
language_model = (child_name, child_module)
continue
if child_type == 'embedding':
embeddings.append((child_name, child_module))
elif child_type == 'norm':
norms.append((child_name, child_module))
elif child_type == 'head':
heads.append((child_name, child_module))
elif child_name.isdigit():
layer_list.append((child_name, child_module))
elif 'layer' in name_lower or 'block' in name_lower or name_lower == 'h':
sub_children = list(child_module.named_children())
if sub_children and sub_children[0][0].isdigit():
layer_container = (child_name, child_module)
else:
others.append((child_name, child_module))
else:
others.append((child_name, child_module))
# Handle multimodal: vision path + language model as parallel branches
if vision_modules and language_model:
parallel_branches = []
# Vision branch: vision_tower + projector in sequence
vision_steps = []
for vm_name, vm_module in vision_modules:
vm_substeps = extract_pipeline_steps(vm_module, vm_name, depth + 1, max_depth, detect_parallel=False)
if vm_substeps:
step = {
"name": humanize_name(vm_name),
"type": "encoder",
"params": count_parameters(vm_module),
"class": vm_module.__class__.__name__,
"substeps": vm_substeps,
"_collapsed": True,
}
else:
step = build_step(vm_module, vm_name, depth + 1, max_depth)
vision_steps.append(step)
vision_branch = {
"name": "Vision Path",
"type": "encoder",
"params": sum(count_parameters(m) for _, m in vision_modules),
"substeps": vision_steps,
"_collapsed": False,
}
parallel_branches.append(vision_branch)
# Language model branch
lm_name, lm_module = language_model
lm_steps = extract_pipeline_steps(lm_module, lm_name, depth + 1, max_depth, detect_parallel=False)
if not lm_steps:
lm_steps = [build_step(lm_module, lm_name, depth + 1, max_depth)]
lang_branch = {
"name": "Language Model",
"type": "module",
"params": count_parameters(lm_module),
"class": lm_module.__class__.__name__,
"substeps": lm_steps,
"_collapsed": False,
}
parallel_branches.append(lang_branch)
steps.append({
"name": "Multimodal Processing",
"type": "parallel",
"params": sum(b.get("params", 0) for b in parallel_branches),
"branches": parallel_branches,
"_collapsed": False,
})
# Skip normal processing - we handled everything
embeddings = []
norms = []
layer_container = None
layer_list = []
others = []
# Handle case where only vision modules exist (no separate language_model)
elif vision_modules:
for enc_name, enc_module in vision_modules:
enc_steps = extract_pipeline_steps(enc_module, enc_name, depth + 1, max_depth, detect_parallel=False)
if enc_steps:
steps.append({
"name": humanize_name(enc_name),
"type": "encoder",
"params": count_parameters(enc_module),
"class": enc_module.__class__.__name__,
"substeps": enc_steps,
"_collapsed": True,
})
else:
steps.append(build_step(enc_module, enc_name, depth + 1, max_depth))
# 1. Regular embeddings (if not already handled in parallel)
for child_name, child_module in embeddings:
step = build_step(child_module, child_name, depth + 1, max_depth)
steps.append(step)
# 2. Transformer layers
if layer_container:
container_name, container_module = layer_container
layer_children = [(n, m) for n, m in container_module.named_children() if count_parameters(m) > 0]
if layer_children:
first_layer = layer_children[0][1]
total_params = sum(count_parameters(m) for _, m in layer_children)
layer_substeps = extract_layer_internals(first_layer, depth + 2, max_depth)
layer_shape = get_layer_shape_info(first_layer)
layer_step = {
"name": f"Transformer Layers",
"type": "layers",
"params": total_params,
"class": first_layer.__class__.__name__,
"count": len(layer_children),
"substeps": layer_substeps,
"_collapsed": False,
}
if layer_shape:
layer_step["shape"] = layer_shape
steps.append(layer_step)
elif layer_list:
first_layer = layer_list[0][1]
total_params = sum(count_parameters(m) for _, m in layer_list)
layer_substeps = extract_layer_internals(first_layer, depth + 2, max_depth)
layer_shape = get_layer_shape_info(first_layer)
layer_step = {
"name": f"Transformer Layers",
"type": "layers",
"params": total_params,
"class": first_layer.__class__.__name__,
"count": len(layer_list),
"substeps": layer_substeps,
"_collapsed": False,
}
if layer_shape:
layer_step["shape"] = layer_shape
steps.append(layer_step)
# 3. Other modules
for child_name, child_module in others:
child_type = get_module_type(child_module, child_name)
if child_type == 'module':
sub_steps = extract_pipeline_steps(child_module, child_name, depth + 1, max_depth, detect_parallel=detect_parallel)
if sub_steps:
steps.extend(sub_steps)
else:
step = build_step(child_module, child_name, depth + 1, max_depth)
steps.append(step)
else:
step = build_step(child_module, child_name, depth + 1, max_depth)
steps.append(step)
# 4. Final norms
for child_name, child_module in norms:
step = build_step(child_module, child_name, depth + 1, max_depth)
steps.append(step)
# 5. Output heads
for child_name, child_module in heads:
step = build_step(child_module, child_name, depth + 1, max_depth)
steps.append(step)
return steps
def extract_layer_internals(layer_module: nn.Module, depth: int, max_depth: int) -> List[Dict[str, Any]]:
"""Extract the internal flow of a single transformer layer."""
steps = []
children = list(layer_module.named_children())
# Categorize
norms = []
attentions = []
mlps = []
others = []
for child_name, child_module in children:
child_params = count_parameters(child_module)
if child_params == 0:
continue
child_type = get_module_type(child_module, child_name)
if child_type == 'norm':
norms.append((child_name, child_module))
elif child_type == 'attention':
attentions.append((child_name, child_module))
elif child_type == 'mlp':
mlps.append((child_name, child_module))
else:
others.append((child_name, child_module))
# Typical transformer layer flow: norm1 -> attn -> norm2 -> mlp
# But order depends on architecture (pre-norm vs post-norm)
# For now, just order: attention first, then MLP, with norms interspersed
norm_idx = 0
# Attention block
if norms and norm_idx < len(norms):
step = build_step(norms[norm_idx][1], norms[norm_idx][0], depth, max_depth)
steps.append(step)
norm_idx += 1
for child_name, child_module in attentions:
step = build_step(child_module, child_name, depth, max_depth)
steps.append(step)
# MLP block
if norms and norm_idx < len(norms):
step = build_step(norms[norm_idx][1], norms[norm_idx][0], depth, max_depth)
steps.append(step)
norm_idx += 1
for child_name, child_module in mlps:
step = build_step(child_module, child_name, depth, max_depth)
steps.append(step)
# Remaining norms
while norm_idx < len(norms):
step = build_step(norms[norm_idx][1], norms[norm_idx][0], depth, max_depth)
steps.append(step)
norm_idx += 1
# Others
for child_name, child_module in others:
step = build_step(child_module, child_name, depth, max_depth)
steps.append(step)
return steps
def get_module_shape(module: nn.Module) -> Optional[str]:
"""Extract shape information from a module."""
class_name = module.__class__.__name__
# Linear layers
if hasattr(module, 'in_features') and hasattr(module, 'out_features'):
return f"{module.in_features} β {module.out_features}"
# Embedding layers
if hasattr(module, 'num_embeddings') and hasattr(module, 'embedding_dim'):
return f"{module.num_embeddings} Γ {module.embedding_dim}"
# LayerNorm / RMSNorm - check multiple possible attribute names
if hasattr(module, 'normalized_shape'):
shape = module.normalized_shape
if isinstance(shape, (list, tuple)):
return f"dim={shape[0]}" if len(shape) == 1 else str(shape)
return f"dim={shape}"
# RMSNorm often uses 'weight' shape
if 'rmsnorm' in class_name.lower() or 'layernorm' in class_name.lower():
if hasattr(module, 'weight') and module.weight is not None:
return f"dim={module.weight.shape[0]}"
# Conv layers
if hasattr(module, 'in_channels') and hasattr(module, 'out_channels'):
kernel = getattr(module, 'kernel_size', None)
if kernel:
return f"{module.in_channels}β{module.out_channels}, k={kernel}"
return f"{module.in_channels} β {module.out_channels}"
# Attention - try to get num_heads and head_dim
if hasattr(module, 'num_heads'):
head_dim = getattr(module, 'head_dim', None)
if head_dim:
return f"heads={module.num_heads}, dim={head_dim}"
return f"heads={module.num_heads}"
if hasattr(module, 'num_attention_heads'):
head_dim = getattr(module, 'head_dim', None)
if head_dim:
return f"heads={module.num_attention_heads}, dim={head_dim}"
return f"heads={module.num_attention_heads}"
# MLP/FFN - try to infer from children
if 'mlp' in class_name.lower() or 'feedforward' in class_name.lower():
# Look for up/gate projection to get intermediate size
for child_name, child in module.named_children():
if hasattr(child, 'out_features'):
return f"β {child.out_features}"
# Try to get hidden_size from config stored on module
if hasattr(module, 'config'):
cfg = module.config
if hasattr(cfg, 'hidden_size'):
return f"hidden={cfg.hidden_size}"
return None
def get_layer_shape_info(layer_module: nn.Module) -> Optional[str]:
"""Extract shape info from a transformer layer by looking at its components."""
hidden_size = None
intermediate_size = None
num_heads = None
for name, child in layer_module.named_modules():
name_lower = name.lower()
# Find num_heads
if not num_heads:
if hasattr(child, 'num_heads'):
num_heads = child.num_heads
elif hasattr(child, 'num_attention_heads'):
num_heads = child.num_attention_heads
# Find hidden_size from multiple sources
if not hidden_size:
# From attention head_dim * num_heads
if hasattr(child, 'num_heads') and hasattr(child, 'head_dim'):
hidden_size = child.num_heads * child.head_dim
# From hidden_size attribute
elif hasattr(child, 'hidden_size'):
hidden_size = child.hidden_size
# From norm layers
elif hasattr(child, 'normalized_shape'):
shape = child.normalized_shape
if isinstance(shape, (list, tuple)):
hidden_size = shape[0]
else:
hidden_size = shape
# From norm weight shape
elif ('norm' in name_lower or 'ln' in name_lower) and hasattr(child, 'weight') and child.weight is not None:
try:
hidden_size = child.weight.shape[0]
except:
pass
# From q_proj or similar linear layers (in_features = hidden_size)
elif ('q_proj' in name_lower or 'query' in name_lower) and hasattr(child, 'in_features'):
hidden_size = child.in_features
# From o_proj output (out_features = hidden_size)
elif ('o_proj' in name_lower or 'out_proj' in name_lower) and hasattr(child, 'out_features'):
hidden_size = child.out_features
# Find intermediate size from MLP
if not intermediate_size:
if ('up' in name_lower or 'gate' in name_lower or 'fc1' in name_lower or 'w1' in name_lower or 'w2' in name_lower) and hasattr(child, 'out_features'):
intermediate_size = child.out_features
parts = []
if hidden_size:
parts.append(f"d={hidden_size}")
if intermediate_size:
parts.append(f"ffn={intermediate_size}")
if num_heads:
parts.append(f"h={num_heads}")
return ", ".join(parts) if parts else None
def build_step(module: nn.Module, name: str, depth: int, max_depth: int) -> Dict[str, Any]:
"""Build a single pipeline step from a module."""
params = count_parameters(module)
module_type = get_module_type(module, name)
display_name = humanize_name(name)
step = {
"name": display_name,
"type": module_type,
"params": params,
"class": module.__class__.__name__,
}
# Add shape information
shape = get_module_shape(module)
if shape:
step["shape"] = shape
# Add substeps for complex modules (if not too deep)
if depth < max_depth:
children = list(module.named_children())
if children:
substeps = []
for child_name, child_module in children:
child_params = count_parameters(child_module)
if child_params > 0:
child_step = build_step(child_module, child_name, depth + 1, max_depth)
substeps.append(child_step)
if substeps:
step["substeps"] = substeps
step["_collapsed"] = True
return step
def build_pipeline(model: nn.Module, model_name: str = "Model") -> Dict[str, Any]:
"""
Build a linear pipeline structure from a PyTorch model.
This shows the sequential flow of data through the model.
"""
total_params = count_parameters(model)
# Extract pipeline steps
steps = extract_pipeline_steps(model, model_name, depth=0, max_depth=4)
return {
"name": model_name,
"params": total_params,
"class": model.__class__.__name__,
"steps": steps
}
def load_model_for_inspection(model_id: str) -> Tuple[nn.Module, AutoConfig]:
"""Load a model architecture without downloading weights."""
from huggingface_hub import hf_hub_download, list_repo_files
import json
# Check if this repo uses Mistral's native format (params.json instead of config.json)
try:
repo_files = list_repo_files(repo_id=model_id)
has_params_json = 'params.json' in repo_files
has_config_json = 'config.json' in repo_files
except:
has_params_json = False
has_config_json = True
if has_params_json and not has_config_json:
# Load Mistral native format and convert to pipeline directly
return None, None # Signal to use parse_mistral_params instead
config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
# Use meta device to avoid allocating actual memory for weights
with torch.device('meta'):
model = None
errors = []
# Try to guess the model class from config
archs = getattr(config, "architectures", [])
is_encoder_decoder = getattr(config, "is_encoder_decoder", False)
# Determine order of AutoModel classes to try
if is_encoder_decoder or any("Seq2Seq" in a or "ConditionalGeneration" in a for a in archs):
model_classes = [
(AutoModelForSeq2SeqLM, "Seq2SeqLM"),
(AutoModelForCausalLM, "CausalLM"),
(AutoModel, "AutoModel")
]
else:
model_classes = [
(AutoModelForCausalLM, "CausalLM"),
(AutoModel, "AutoModel"),
(AutoModelForSeq2SeqLM, "Seq2SeqLM")
]
for model_class, label in model_classes:
try:
model = model_class.from_config(config, trust_remote_code=True)
if model is not None:
break
except Exception as e:
errors.append(f"{label}: {e}")
if model is None:
raise ValueError(f"Could not load model architecture. Errors: {errors}")
return model, config
def parse_mistral_native_format(model_id: str) -> Dict[str, Any]:
"""Parse Mistral's native params.json format."""
from huggingface_hub import hf_hub_download
import json
params_path = hf_hub_download(repo_id=model_id, filename='params.json')
with open(params_path) as f:
params = json.load(f)
# Extract dimensions
hidden_size = params.get('dim', 0)
num_layers = params.get('n_layers', 0)
num_heads = params.get('n_heads', 0)
num_kv_heads = params.get('n_kv_heads', num_heads)
vocab_size = params.get('vocab_size', 0)
intermediate_size = params.get('hidden_dim', hidden_size * 4)
head_dim = params.get('head_dim', hidden_size // num_heads if num_heads > 0 else 0)
# Check for MoE
moe_config = params.get('moe', {})
num_experts = moe_config.get('num_experts', 0)
num_experts_per_tok = moe_config.get('num_experts_per_tok', 2)
expert_hidden_dim = moe_config.get('expert_hidden_dim', intermediate_size)
num_shared_experts = moe_config.get('num_shared_experts', 0)
first_k_dense = moe_config.get('first_k_dense_replace', 0) # First K layers use dense MLP
# Check for vision encoder
vision_config = params.get('vision_encoder', None)
# Calculate parameters
embed_params = vocab_size * hidden_size
# Attention params per layer (with potential LoRA/MLA components)
q_lora_rank = params.get('q_lora_rank', 0)
kv_lora_rank = params.get('kv_lora_rank', 0)
v_head_dim = params.get('v_head_dim', head_dim) # V uses different head dim
if q_lora_rank > 0:
# Multi-head Latent Attention (MLA) - compressed projections
# Q: down_proj + up_proj
q_params = hidden_size * q_lora_rank + q_lora_rank * num_heads * head_dim
# K: down_proj + up_proj (shared with V in latent space)
k_params = hidden_size * kv_lora_rank + kv_lora_rank * num_kv_heads * head_dim
# V: uses v_head_dim
v_params = hidden_size * kv_lora_rank + kv_lora_rank * num_kv_heads * v_head_dim
# O: output projection from v_head_dim back to hidden
o_params = num_heads * v_head_dim * hidden_size
attn_params = q_params + k_params + v_params + o_params
else:
q_params = hidden_size * num_heads * head_dim
kv_params = hidden_size * num_kv_heads * head_dim
attn_params = q_params + 2 * kv_params + num_heads * head_dim * hidden_size
norm_params = hidden_size
# MLP params - handle dense vs MoE layers
dense_mlp_params = 3 * hidden_size * intermediate_size
if num_experts > 0:
# MoE: each expert has gate + up + down projections
single_expert_params = 3 * hidden_size * expert_hidden_dim
moe_mlp_params = num_experts * single_expert_params
if num_shared_experts > 0:
# Shared experts use same size as routed experts
moe_mlp_params += num_shared_experts * single_expert_params
moe_mlp_params += hidden_size * num_experts # Router
# Calculate layer params for dense and MoE layers separately
num_dense_layers = min(first_k_dense, num_layers)
num_moe_layers = num_layers - num_dense_layers
dense_layer_params = attn_params + dense_mlp_params + 2 * norm_params
moe_layer_params = attn_params + moe_mlp_params + 2 * norm_params
total_layer_params = (dense_layer_params * num_dense_layers) + (moe_layer_params * num_moe_layers)
mlp_params = moe_mlp_params # For display purposes, show MoE params
else:
mlp_params = dense_mlp_params
layer_params = attn_params + mlp_params + 2 * norm_params
total_layer_params = layer_params * num_layers
lm_head_params = 0 if params.get('tied_embeddings', True) else vocab_size * hidden_size
total_params = embed_params + total_layer_params + norm_params + lm_head_params
# Vision encoder params
vision_params = 0
vision_steps = []
if vision_config:
v_hidden = vision_config.get('hidden_size', 0)
v_layers = vision_config.get('num_hidden_layers', 0)
v_intermediate = vision_config.get('intermediate_size', v_hidden * 4)
v_heads = vision_config.get('num_attention_heads', 0)
patch_size = vision_config.get('patch_size', 14)
patch_embed_params = 3 * (patch_size ** 2) * v_hidden
v_attn = 4 * v_hidden * v_hidden
v_mlp = 2 * v_hidden * v_intermediate
v_layer_params = v_attn + v_mlp + 2 * v_hidden
vision_params = patch_embed_params + v_layer_params * v_layers
vision_steps = [
{
"name": "Patch Embedding",
"type": "embedding",
"params": patch_embed_params,
"shape": f"{patch_size}Γ{patch_size} patches β {v_hidden}",
"class": "Conv2d"
},
{
"name": "Vision Transformer Layers",
"type": "layers",
"params": v_layer_params * v_layers,
"count": v_layers,
"shape": f"d={v_hidden}, h={v_heads}",
"class": "ViTBlock",
"_collapsed": True
}
]
total_params += vision_params
# Build pipeline
steps = []
# Embedding
steps.append({
"name": "Token Embedding",
"type": "embedding",
"params": embed_params,
"shape": f"{vocab_size:,} Γ {hidden_size}",
"class": "Embedding"
})
# Build layer substeps
layer_substeps = [
{
"name": "Input LayerNorm",
"type": "norm",
"params": norm_params,
"shape": f"dim={hidden_size}",
"class": "RMSNorm"
},
{
"name": "Self Attention",
"type": "attention",
"params": attn_params,
"shape": f"heads={num_heads}, kv_heads={num_kv_heads}, dim={head_dim}",
"class": "Attention",
"_collapsed": True
},
{
"name": "Post-Attention LayerNorm",
"type": "norm",
"params": norm_params,
"shape": f"dim={hidden_size}",
"class": "RMSNorm"
}
]
if num_experts > 0:
layer_substeps.append({
"name": "MoE",
"type": "mlp",
"params": mlp_params,
"shape": f"{num_experts} experts, top-{num_experts_per_tok}",
"class": "MixtureOfExperts",
"_collapsed": True
})
layer_shape = f"d={hidden_size}, ffn={expert_hidden_dim}, h={num_heads}, experts={num_experts}"
else:
layer_substeps.append({
"name": "MLP",
"type": "mlp",
"params": mlp_params,
"shape": f"{hidden_size} β {intermediate_size} β {hidden_size}",
"class": "MLP",
"_collapsed": True
})
layer_shape = f"d={hidden_size}, ffn={intermediate_size}, h={num_heads}"
moe_label = " (MoE)" if num_experts > 0 else ""
steps.append({
"name": f"Transformer Layers{moe_label}",
"type": "layers",
"params": total_layer_params,
"count": num_layers,
"shape": layer_shape,
"class": "TransformerBlock",
"substeps": layer_substeps,
"_collapsed": False
})
# Final norm
steps.append({
"name": "Final LayerNorm",
"type": "norm",
"params": norm_params,
"shape": f"dim={hidden_size}",
"class": "RMSNorm"
})
# LM Head
steps.append({
"name": "LM Head",
"type": "head",
"params": lm_head_params if lm_head_params > 0 else embed_params,
"shape": f"{hidden_size} β {vocab_size:,}" + (" (tied)" if lm_head_params == 0 else ""),
"class": "Linear"
})
# Wrap with vision if present
if vision_config:
vision_branch = {
"name": "Vision Encoder",
"type": "encoder",
"params": vision_params,
"substeps": vision_steps,
"_collapsed": True
}
lang_branch = {
"name": "Language Model",
"type": "module",
"params": total_params - vision_params,
"substeps": steps,
"_collapsed": False
}
steps = [{
"name": "Multimodal Processing",
"type": "parallel",
"params": total_params,
"branches": [vision_branch, lang_branch],
"_collapsed": False
}]
model_type = "mistral"
if num_experts > 0:
model_type = "mistral_moe"
return {
"name": model_type.upper(),
"params": total_params,
"formatted_params": format_params(total_params),
"model_type": model_type,
"class": "MistralModel",
"steps": steps
}
def load_model_from_config(config_dict: Dict[str, Any]) -> Tuple[nn.Module, AutoConfig]:
"""Load a model architecture from a config dictionary."""
config = AutoConfig.for_model(**config_dict)
with torch.device('meta'):
model = None
errors = []
# Try to guess the model class from config
archs = getattr(config, "architectures", [])
is_encoder_decoder = getattr(config, "is_encoder_decoder", False)
# Determine order of AutoModel classes to try
if is_encoder_decoder or any("Seq2Seq" in a or "ConditionalGeneration" in a for a in archs):
model_classes = [
(AutoModelForSeq2SeqLM, "Seq2SeqLM"),
(AutoModelForCausalLM, "CausalLM"),
(AutoModel, "AutoModel")
]
else:
model_classes = [
(AutoModelForCausalLM, "CausalLM"),
(AutoModel, "AutoModel"),
(AutoModelForSeq2SeqLM, "Seq2SeqLM")
]
for model_class, label in model_classes:
try:
model = model_class.from_config(config, trust_remote_code=True)
if model is not None:
break
except Exception as e:
errors.append(f"{label}: {e}")
if model is None:
raise ValueError(f"Could not load model from config. Errors: {errors}")
return model, config
def parse_model(model_id: str) -> Dict[str, Any]:
"""Parse a model from HuggingFace and return pipeline structure."""
model, config = load_model_for_inspection(model_id)
# If model is None, it means we need to use Mistral native format
if model is None and config is None:
return parse_mistral_native_format(model_id)
model_name = getattr(config, 'model_type', 'Model').upper()
pipeline = build_pipeline(model, model_name)
total_params = count_parameters(model)
pipeline["params"] = total_params
pipeline["formatted_params"] = format_params(total_params)
pipeline["model_type"] = getattr(config, 'model_type', 'unknown')
return pipeline
def parse_config(config_dict: Dict[str, Any]) -> Dict[str, Any]:
"""Parse a model from config dict and return pipeline structure."""
model, config = load_model_from_config(config_dict)
model_name = getattr(config, 'model_type', 'Model').upper()
pipeline = build_pipeline(model, model_name)
total_params = count_parameters(model)
pipeline["params"] = total_params
pipeline["formatted_params"] = format_params(total_params)
pipeline["model_type"] = getattr(config, 'model_type', 'unknown')
return pipeline
|