File size: 26,569 Bytes
9ff632c d191028 9ff632c |
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 |
#!/usr/bin/env python
"""
Stage 1 v2 Sharted Edition π©: Fast Multi-GPU Interpolation from Qwen3-32B to Qwen3-72B
Optimized for 8x MI300X GPUs with parallel processing and sharted weight loading
FIXED: Correct o_proj dimensions
"""
import torch
import torch.distributed as dist
import torch.multiprocessing as mp
import os
import json
from tqdm import tqdm
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
from accelerate import init_empty_weights
import numpy as np
from concurrent.futures import ThreadPoolExecutor, ProcessPoolExecutor
import gc
from safetensors.torch import load_file, save_file
import shutil
# --- Configuration ---
# Source (32B) dimensions
SRC_HIDDEN_SIZE = 5120
SRC_INTERMEDIATE_SIZE = 25600
SRC_NUM_HEADS = 40
SRC_NUM_LAYERS = 64
# IMPORTANT: Qwen3-32B already has asymmetric attention!
# Q heads: 64 (for q_proj output and o_proj input)
# KV heads: 8
SRC_Q_HEADS = 64 # This gives us 8192 dims for Q
SRC_KV_HEADS = 8 # This gives us 1024 dims for K,V
# Target (72B) dimensions
TGT_HIDDEN_SIZE = 8192
TGT_INTERMEDIATE_SIZE = 29568
TGT_NUM_HEADS = 64
# Target also has asymmetric attention
TGT_Q_HEADS = 64
TGT_KV_HEADS = 8
HEAD_DIM = 128
# Deltas for interpolation
DELTA_HIDDEN = TGT_HIDDEN_SIZE - SRC_HIDDEN_SIZE
DELTA_INTERMEDIATE = TGT_INTERMEDIATE_SIZE - SRC_INTERMEDIATE_SIZE
OUTPUT_DIR = "./Qwen3-58B-Embiggened"
# GPU configuration
NUM_GPUS = 8
BATCH_SIZE = 16 # Process multiple tensors at once
def get_layer_info(name):
"""Extract layer number and component type from parameter name."""
if "model.layers." in name:
parts = name.split(".")
try:
layer_idx = int(parts[2])
return layer_idx, ".".join(parts[3:])
except:
return None, name
return None, name
def get_interpolation_weight(layer_idx, num_layers=SRC_NUM_LAYERS):
"""Get interpolation weight based on layer depth."""
if layer_idx is None:
return 0.5
relative_pos = layer_idx / (num_layers - 1)
if relative_pos < 0.25:
return 0.3
elif relative_pos < 0.75:
return 0.5
else:
return 0.7
@torch.jit.script
def add_structured_noise_jit(tensor: torch.Tensor, noise_scale: float = 0.01) -> torch.Tensor:
"""JIT-compiled structured noise addition."""
noise = torch.randn_like(tensor) * noise_scale * tensor.std()
if tensor.ndim == 2 and tensor.shape[0] > 100 and tensor.shape[1] > 100:
h, w = noise.shape
center_mask = torch.ones_like(noise)
center_mask[h//4:3*h//4, w//4:3*w//4] *= 0.5
noise *= center_mask
return noise
@torch.jit.script
def preserve_norm_jit(original: torch.Tensor, interpolated: torch.Tensor) -> torch.Tensor:
"""JIT-compiled norm preservation."""
original_norm = original.norm()
interpolated_norm = interpolated.norm()
if interpolated_norm > 0:
scale_factor = original_norm / interpolated_norm
return interpolated * scale_factor
return interpolated
def structure_aware_interpolation_gpu(block1, block2, weight=0.5, add_noise=True, device='cuda'):
"""GPU-accelerated interpolation."""
# Move to GPU if not already
if block1.device.type != 'cuda':
block1 = block1.to(device)
if block2.device.type != 'cuda':
block2 = block2.to(device)
# Basic interpolation
interpolated = (1 - weight) * block1 + weight * block2
# Add noise on GPU
if add_noise:
noise = add_structured_noise_jit(interpolated, 0.005)
interpolated = interpolated + noise
return interpolated
def upscale_tensor_gpu(tensor: torch.Tensor, name: str, device='cuda') -> torch.Tensor:
"""GPU-accelerated tensor upscaling with FIXED o_proj dimensions."""
# Move tensor to GPU
tensor = tensor.to(device)
layer_idx, component = get_layer_info(name)
interp_weight = get_interpolation_weight(layer_idx)
# Debug print for ANY o_proj to catch the first one
if "o_proj.weight" in name:
print(f"\n[DEBUG] Processing {name}: input shape = {tensor.shape}")
# Handle 1D tensors
if tensor.ndim == 1:
if tensor.shape[0] == SRC_HIDDEN_SIZE:
block1, block2 = tensor[:DELTA_HIDDEN], tensor[-DELTA_HIDDEN:]
interpolated = structure_aware_interpolation_gpu(block1, block2, weight=interp_weight, device=device)
result = torch.cat([tensor, interpolated], dim=0)
if "layernorm" in name:
result = preserve_norm_jit(tensor, result)
return result
elif "k_norm" in name or "q_norm" in name:
return tensor
# Handle 2D tensors
elif tensor.ndim == 2:
# Embeddings and LM head
if "embed_tokens" in name or "lm_head" in name:
if tensor.shape[1] == SRC_HIDDEN_SIZE:
block1, block2 = tensor[:, :DELTA_HIDDEN], tensor[:, -DELTA_HIDDEN:]
interpolated = structure_aware_interpolation_gpu(block1, block2, weight=0.3, device=device)
return torch.cat([tensor, interpolated], dim=1)
# Attention projections
elif "self_attn" in name:
if "q_proj.weight" in name:
# Q projection: [8192, 5120] -> [8192, 8192]
# Already has 64 heads in output, just need to expand input
# Only scale input dimension (columns)
block1, block2 = tensor[:, :DELTA_HIDDEN], tensor[:, -DELTA_HIDDEN:]
interpolated = structure_aware_interpolation_gpu(block1, block2, weight=interp_weight, device=device)
result = torch.cat([tensor, interpolated], dim=1)
return preserve_norm_jit(tensor, result)
elif "k_proj.weight" in name or "v_proj.weight" in name:
# K,V projections: [1024, 5120] -> [1024, 8192]
# Only scale input dimension, keep 8 KV heads
block1, block2 = tensor[:, :DELTA_HIDDEN], tensor[:, -DELTA_HIDDEN:]
interpolated = structure_aware_interpolation_gpu(block1, block2, weight=interp_weight, device=device)
result = torch.cat([tensor, interpolated], dim=1)
return preserve_norm_jit(tensor, result)
elif "o_proj.weight" in name:
# O projection: [5120, 8192] -> [8192, 8192]
# Input already has 64 heads (8192), only expand output
# Debug the input
print(f"\n[DEBUG] Processing {name}: input shape = {tensor.shape}")
print(f"[DEBUG] Expected input: [5120, 8192], Expected output: [8192, 8192]")
# Only need to expand rows (output dim) from 5120 to 8192
row_block1 = tensor[:DELTA_HIDDEN, :] # [3072, 8192]
row_block2 = tensor[-DELTA_HIDDEN:, :] # [3072, 8192]
row_interp = structure_aware_interpolation_gpu(row_block1, row_block2, weight=interp_weight, device=device)
print(f"[DEBUG] row interpolation: block1={row_block1.shape}, block2={row_block2.shape}, interp={row_interp.shape}")
result = torch.cat([tensor, row_interp], dim=0) # [5120+3072, 8192] = [8192, 8192]
print(f"[DEBUG] Final result: {result.shape}")
assert result.shape == (TGT_HIDDEN_SIZE, TGT_HIDDEN_SIZE), f"o_proj shape error: got {result.shape}"
return preserve_norm_jit(tensor, result)
# MLP projections
elif "mlp" in name:
if "gate_proj.weight" in name or "up_proj.weight" in name:
# [25600, 5120] -> [29568, 8192]
mlp_weight = min(interp_weight + 0.1, 0.8)
# Expand rows first
row_block1, row_block2 = tensor[:DELTA_INTERMEDIATE, :], tensor[-DELTA_INTERMEDIATE:, :]
upscaled_rows = torch.cat([tensor, structure_aware_interpolation_gpu(row_block1, row_block2, weight=mlp_weight, device=device)], dim=0)
# Then expand columns
col_block1, col_block2 = upscaled_rows[:, :DELTA_HIDDEN], upscaled_rows[:, -DELTA_HIDDEN:]
result = torch.cat([upscaled_rows, structure_aware_interpolation_gpu(col_block1, col_block2, weight=mlp_weight, device=device)], dim=1)
result = preserve_norm_jit(tensor, result)
return result
elif "down_proj.weight" in name:
# [5120, 25600] -> [8192, 29568]
mlp_weight = interp_weight
# Expand rows first
row_block1, row_block2 = tensor[:DELTA_HIDDEN, :], tensor[-DELTA_HIDDEN:, :]
upscaled_rows = torch.cat([tensor, structure_aware_interpolation_gpu(row_block1, row_block2, weight=mlp_weight, device=device)], dim=0)
# Then expand columns
col_block1, col_block2 = upscaled_rows[:, :DELTA_INTERMEDIATE], upscaled_rows[:, -DELTA_INTERMEDIATE:]
result = torch.cat([upscaled_rows, structure_aware_interpolation_gpu(col_block1, col_block2, weight=mlp_weight, device=device)], dim=1)
return result
return tensor
def process_layer_batch(layer_tensors, device):
"""Process a batch of tensors from the same layer on a specific GPU."""
processed = {}
with torch.cuda.device(device):
for name, tensor in layer_tensors:
processed_tensor = upscale_tensor_gpu(tensor, name, device=device)
# Move back to CPU to save GPU memory
processed[name] = processed_tensor.cpu()
return processed
def load_model_sharted(model_id):
"""Load model weights from sharted safetensors files. π©"""
print("\nπ© Loading sharted weights...")
model_path = os.path.join(model_id, "model.safetensors.index.json")
if os.path.exists(model_path):
# Load from local path with sharted files
with open(model_path, 'r') as f:
index = json.load(f)
weight_map = index['weight_map']
unique_files = set(weight_map.values())
all_weights = {}
for file in tqdm(unique_files, desc="Loading sharts"):
file_path = os.path.join(model_id, file)
weights = load_file(file_path)
all_weights.update(weights)
return all_weights
else:
# Try loading from HuggingFace
from huggingface_hub import snapshot_download
print(f"Downloading model from HuggingFace: {model_id}")
local_dir = snapshot_download(model_id)
return load_model_sharted(local_dir)
def save_model_sharted(state_dict, output_dir, max_shart_size="5GB"):
"""Save model in sharted safetensors format. π©"""
print("\nπ© Sharting model weights...")
os.makedirs(output_dir, exist_ok=True)
# Convert max_shart_size to bytes
size_map = {'GB': 1e9, 'MB': 1e6}
for unit, multiplier in size_map.items():
if unit in max_shart_size:
max_bytes = int(float(max_shart_size.replace(unit, '')) * multiplier)
break
# Group weights into sharts
sharts = []
current_shart = {}
current_size = 0
for name, tensor in state_dict.items():
tensor_size = tensor.numel() * tensor.element_size()
if current_size + tensor_size > max_bytes and current_shart:
sharts.append(current_shart)
current_shart = {}
current_size = 0
current_shart[name] = tensor
current_size += tensor_size
if current_shart:
sharts.append(current_shart)
# Save sharts
weight_map = {}
for i, shart in enumerate(tqdm(sharts, desc="Saving sharts")):
shart_name = f"model-{i+1:05d}-of-{len(sharts):05d}.safetensors"
save_file(shart, os.path.join(output_dir, shart_name))
for name in shart:
weight_map[name] = shart_name
# Save index
index = {
"metadata": {"total_size": sum(t.numel() * t.element_size() for t in state_dict.values())},
"weight_map": weight_map
}
with open(os.path.join(output_dir, "model.safetensors.index.json"), 'w') as f:
json.dump(index, f, indent=2)
print(f"π© Successfully sharted into {len(sharts)} files!")
def verify_architecture(model_path):
"""Verify the model architecture matches expected dimensions."""
print("\n" + "="*60)
print("ARCHITECTURE VERIFICATION")
print("="*60)
model = AutoModelForCausalLM.from_pretrained(
model_path,
torch_dtype=torch.bfloat16,
device_map="cpu",
trust_remote_code=True
)
expected = {
"lm_head.weight": (151936, 8192),
"model.embed_tokens.weight": (151936, 8192),
"model.layers.0.input_layernorm.weight": (8192,),
"model.layers.0.mlp.down_proj.weight": (8192, 29568),
"model.layers.0.mlp.gate_proj.weight": (29568, 8192),
"model.layers.0.mlp.up_proj.weight": (29568, 8192),
"model.layers.0.post_attention_layernorm.weight": (8192,),
"model.layers.0.self_attn.k_norm.weight": (128,),
"model.layers.0.self_attn.k_proj.weight": (1024, 8192),
"model.layers.0.self_attn.o_proj.weight": (8192, 8192),
"model.layers.0.self_attn.q_norm.weight": (128,),
"model.layers.0.self_attn.q_proj.weight": (8192, 8192),
"model.layers.0.self_attn.v_proj.weight": (1024, 8192),
"model.norm.weight": (8192,),
}
all_correct = True
for name, expected_shape in expected.items():
param_dict = dict(model.named_parameters())
if name in param_dict:
actual_shape = tuple(param_dict[name].shape)
if actual_shape == expected_shape:
print(f"β {name}: {actual_shape}")
else:
print(f"β {name}: {actual_shape} (expected {expected_shape})")
all_correct = False
else:
print(f"β {name}: NOT FOUND")
all_correct = False
num_layers = model.config.num_hidden_layers
print(f"\nNumber of layers: {num_layers} (Stage 1 should have 64)")
if all_correct and num_layers == 64:
print("\nβ
Architecture verification PASSED!")
else:
print("\nβ Architecture verification FAILED!")
del model
return all_correct
def run_diagnostics(model_path):
"""Run comprehensive diagnostics on the upscaled model."""
print("\n" + "="*60)
print("COMPREHENSIVE DIAGNOSTICS")
print("="*60)
# Load model and tokenizer
print("\nLoading model for diagnostics...")
model = AutoModelForCausalLM.from_pretrained(
model_path,
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
# Test generation quality
print("\nπ§ͺ Generation Quality Tests:")
test_cases = [
("The capital of France is", ["Paris"]),
("2 + 2 =", ["4", "four"]),
("The quick brown fox", ["jumps", "jumped", "lazy", "dog"]),
("Hello, my name is", None),
("Water boils at", ["100", "212", "degrees"]),
("The Earth orbits the", ["Sun", "solar"]),
("Machine learning is a type of", ["artificial intelligence", "AI"]),
("Python is a", ["programming", "language", "snake"]),
("The largest planet is", ["Jupiter"]),
("DNA stands for", ["deoxyribonucleic", "acid"]),
]
device = model.device
coherent_count = 0
total_tests = len(test_cases)
for prompt, expected in test_cases:
inputs = tokenizer(prompt, return_tensors="pt").to(device)
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=20,
do_sample=True,
temperature=0.7,
top_k=50,
top_p=0.95,
pad_token_id=tokenizer.pad_token_id,
)
generated_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
generated_only = generated_text[len(prompt):].strip()
print(f"\n Prompt: '{prompt}'")
print(f" Generated: '{generated_only}'")
# Check coherence
is_coherent = True
# Check for repetition
words = generated_only.split()
if len(words) > 3:
if len(set(words)) < len(words) / 2:
print(" β οΈ High repetition detected")
is_coherent = False
# Check for expected content
if expected and len(generated_only) > 0:
found = any(kw.lower() in generated_only.lower() for kw in expected)
if found:
print(" β Contains expected content")
else:
print(" β οΈ Missing expected keywords")
is_coherent = False
if is_coherent and len(generated_only.split()) >= 2:
coherent_count += 1
coherence_rate = (coherent_count / total_tests) * 100
print(f"\nπ Overall coherence rate: {coherence_rate:.1f}%")
# Quick perplexity test
print("\nπ Perplexity Test:")
test_text = "The quick brown fox jumps over the lazy dog."
inputs = tokenizer(test_text, return_tensors="pt").to(device)
with torch.no_grad():
outputs = model(**inputs, labels=inputs["input_ids"])
perplexity = torch.exp(outputs.loss).item()
print(f" Perplexity: {perplexity:.2f}")
if perplexity > 100:
print(" β οΈ Very high perplexity")
elif perplexity > 50:
print(" β οΈ Moderately high perplexity")
else:
print(" β Reasonable perplexity")
# Weight statistics check
print("\nπ Weight Statistics (checking for anomalies):")
anomalies = 0
for name, param in model.named_parameters():
if torch.isnan(param).any():
print(f" β οΈ {name}: Contains NaN!")
anomalies += 1
elif torch.isinf(param).any():
print(f" β οΈ {name}: Contains Inf!")
anomalies += 1
elif param.std() < 1e-8:
print(f" β οΈ {name}: Zero variance!")
anomalies += 1
if anomalies == 0:
print(" β No anomalies detected in weights")
# Final summary
success = coherence_rate >= 70 and perplexity < 100 and anomalies == 0
print("\n" + "="*60)
print("DIAGNOSTIC SUMMARY")
print("="*60)
if success:
print("β
Model passed all basic diagnostics!")
print(" - Good coherence rate")
print(" - Reasonable perplexity")
print(" - No weight anomalies")
else:
print("β οΈ Some issues detected:")
if coherence_rate < 70:
print(f" - Low coherence rate: {coherence_rate:.1f}%")
if perplexity >= 100:
print(f" - High perplexity: {perplexity:.2f}")
if anomalies > 0:
print(f" - Weight anomalies: {anomalies}")
return success
def main():
print("="*60)
print("Stage 1 v2 SHARTED π©: Multi-GPU Accelerated Interpolation")
print("Qwen3-32B β 72B Dimensions")
print(f"Using {NUM_GPUS} GPUs for parallel processing")
print("FIXED: Correct o_proj dimensions")
print("="*60)
source_model_id = "Qwen/Qwen3-32B"
# Set up multi-GPU environment
if torch.cuda.is_available():
torch.cuda.set_device(0)
print(f"\nπ CUDA available: {torch.cuda.device_count()} devices")
for i in range(min(NUM_GPUS, torch.cuda.device_count())):
print(f" GPU {i}: {torch.cuda.get_device_name(i)}")
# Load tokenizer
print(f"\nπ Loading tokenizer from: {source_model_id}")
tokenizer = AutoTokenizer.from_pretrained(source_model_id, trust_remote_code=True)
# Load weights directly (faster than loading full model)
print(f"\nβ‘ Loading model weights using fast sharted loading...")
source_weights = load_model_sharted(source_model_id)
print(f"\nπ Loaded {len(source_weights)} tensors from sharts")
# Group tensors by layer for efficient GPU processing
layer_groups = {}
other_tensors = []
for name, tensor in source_weights.items():
layer_idx, _ = get_layer_info(name)
if layer_idx is not None:
if layer_idx not in layer_groups:
layer_groups[layer_idx] = []
layer_groups[layer_idx].append((name, tensor))
else:
other_tensors.append((name, tensor))
print(f"\nπ§ Processing tensors across {NUM_GPUS} GPUs...")
print(" - Parallel layer processing")
print(" - JIT-compiled operations")
print(" - Efficient memory management")
print(" - Sharted weight I/O π©")
new_state_dict = {}
# Process layers in parallel across GPUs
with tqdm(total=len(source_weights), desc="Upscaling tensors") as pbar:
# Process layer groups in batches across GPUs
layer_indices = sorted(layer_groups.keys())
for i in range(0, len(layer_indices), NUM_GPUS):
batch_futures = []
# Assign each layer in this batch to a GPU
for j, layer_idx in enumerate(layer_indices[i:i+NUM_GPUS]):
gpu_id = j % NUM_GPUS
device = f'cuda:{gpu_id}'
# Process this layer on the assigned GPU
layer_tensors = layer_groups[layer_idx]
processed = process_layer_batch(layer_tensors, device)
new_state_dict.update(processed)
pbar.update(len(layer_tensors))
# Clear GPU cache periodically
if j % 4 == 0:
torch.cuda.empty_cache()
# Process non-layer tensors
for name, tensor in other_tensors:
device = 'cuda:0'
new_tensor = upscale_tensor_gpu(tensor, name, device=device).cpu()
new_state_dict[name] = new_tensor
pbar.update(1)
# Free source weights
del source_weights
gc.collect()
torch.cuda.empty_cache()
# Create config
print("\nπ Creating target model configuration...")
config = AutoConfig.from_pretrained(source_model_id, trust_remote_code=True)
config.hidden_size = TGT_HIDDEN_SIZE
config.intermediate_size = TGT_INTERMEDIATE_SIZE
config.num_attention_heads = TGT_NUM_HEADS
config.torch_dtype = torch.bfloat16
# Quick verification
print("\nπ Quick verification of tensor dimensions BEFORE saving:")
# Check critical dimensions
critical_checks = [
"model.layers.0.self_attn.q_proj.weight",
"model.layers.0.self_attn.k_proj.weight",
"model.layers.0.self_attn.v_proj.weight",
"model.layers.0.self_attn.o_proj.weight",
"model.layers.0.mlp.gate_proj.weight"
]
for check_name in critical_checks:
for name, tensor in new_state_dict.items():
if check_name in name:
print(f" {name}: {tensor.shape}")
break
# Specifically verify o_proj dimensions
print("\nπ― Verifying ALL o_proj dimensions:")
o_proj_issue = False
for name, tensor in new_state_dict.items():
if "o_proj.weight" in name:
if tensor.shape != (TGT_HIDDEN_SIZE, TGT_HIDDEN_SIZE):
print(f" β {name}: {tensor.shape} - INCORRECT!")
o_proj_issue = True
else:
if "layer.0" in name or "layer.63" in name: # Show first and last
print(f" β {name}: {tensor.shape}")
if o_proj_issue:
print("\nβ ERROR: o_proj dimensions are incorrect! Not saving model.")
return False
# Save model and config
print(f"\nπΎ Saving model to: {OUTPUT_DIR}")
os.makedirs(OUTPUT_DIR, exist_ok=True)
# Save config
config.save_pretrained(OUTPUT_DIR)
tokenizer.save_pretrained(OUTPUT_DIR)
# Save weights in sharted format
save_model_sharted(new_state_dict, OUTPUT_DIR)
# Copy model configuration files
for file in ['generation_config.json', 'tokenizer_config.json', 'special_tokens_map.json']:
src = os.path.join(source_model_id, file)
dst = os.path.join(OUTPUT_DIR, file)
if os.path.exists(src):
shutil.copy(src, dst)
# Save metadata
metadata = {
"stage": "1-v2-sharted",
"source_model": source_model_id,
"method": "gpu_accelerated_structure_aware_interpolation_sharted",
"num_gpus_used": NUM_GPUS,
"fixes": [
"Corrected o_proj dimensions to 8192x8192",
"Proper handling of GQA architecture"
],
"optimizations": [
"Multi-GPU parallel processing",
"JIT-compiled operations",
"Sharted weight loading/saving π©",
"Efficient memory management"
],
"sharting_info": {
"format": "safetensors",
"max_shart_size": "5GB",
"poop_emoji": "π©"
}
}
with open(os.path.join(OUTPUT_DIR, "stage1_v2_metadata.json"), "w") as f:
json.dump(metadata, f, indent=2)
print("\nβ
Stage 1 v2 SHARTED interpolation complete! π©")
print(f"π Model saved to: {OUTPUT_DIR}")
# Run verifications
arch_ok = verify_architecture(OUTPUT_DIR)
diag_ok = run_diagnostics(OUTPUT_DIR)
if arch_ok and diag_ok:
print("\nπ SUCCESS! Enhanced sharted interpolation completed successfully. π©")
print(f"π Model saved to: {OUTPUT_DIR}")
print("\nπ Ready for Stage 2: Layer duplication (64β80 layers)")
else:
print("\nβ οΈ Some issues detected. Review the diagnostics above.")
return arch_ok and diag_ok
if __name__ == "__main__":
success = main()
exit(0 if success else 1) |