File size: 40,100 Bytes
5d61448 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 | """
Sequential Merge Orchestrator β chains 4 merges with protection.
This is the brain of td_lang engine. It runs each merge in order:
1. Load source model
2. Inject canary fact into source
3. Extract activations from both models
4. Compute transport plans (P and Q matrices)
5. Fuse weights using optimal transport
6. Validate merged model (canary recall, perplexity, thinking mode)
7. Apply sequential merge protection before next merge
8. Checkpoint
Protection between merges (findings #13):
- MagMax: Protect top 20% parameters by magnitude (they carry critical knowledge)
- Orthogonal Projection: Project new merge deltas perpendicular to previous ones
- Time-Aware Scaling: scale = 1/sqrt(merge_index + 1)
Kill criteria: >10% performance drop on any test β abort merge.
Findings: #13, #22, #25
"""
import os
import gc
import copy
import torch
import numpy as np
from pathlib import Path
from typing import Optional
from transformers import AutoModelForCausalLM, AutoTokenizer
from .config import (
MergeConfig, ModelConfig, TARGET, SOURCES,
CANARY_FACTS, DEMO_STAGES, FULL_STAGES,
)
from .canary import inject_canary, test_all_canaries
from .transport import (
setup_tm_repo,
load_calibration_data,
extract_activations,
compute_transport_plans,
fuse_weights,
)
from .validate import validate_merged_model, compute_perplexity
from .techniques import (
compute_mergeability_score,
compute_transferability_masks,
apply_masked_merge,
disentangle_rl_weights,
merge_with_rl_preservation,
compute_arm_rotation,
apply_arm_steering,
transport_task_vector_theseus,
compute_procrustes_alignment,
)
# ============================================================================
# SEQUENTIAL MERGE PROTECTION
# ============================================================================
class MergeProtection:
"""
Protects previously merged knowledge from being overwritten.
Think of it like this: after merging DeepSeek into Qwen3, we have
a "direction" in weight space that represents that merge. When we
then merge MiMo, we want MiMo's changes to go in a DIFFERENT direction,
not overwrite DeepSeek's contribution.
Three mechanisms:
1. MagMax: Top 20% magnitude params are "locked" β new merges can't change them much
2. Orthogonal Projection: New deltas are projected perpendicular to previous deltas
3. Time-Aware Scaling: Each successive merge gets a smaller alpha (1/sqrt(n+1))
"""
def __init__(self, cfg: MergeConfig):
self.cfg = cfg
self.previous_deltas = {} # key β list of delta tensors from previous merges
self.magnitude_masks = {} # key β bool mask of top-k magnitude params
self.arm_rotations = {} # ARM: layer β rotation info from last merge
self.otmf_masks = {} # OTMF: param β transferability mask
self.merge_count = 0
def before_merge(
self,
target_model: AutoModelForCausalLM,
source_config: ModelConfig,
) -> float:
"""
Prepare protection before a merge. Returns adjusted alpha.
Called BEFORE each merge to:
1. Compute magnitude masks (MagMax)
2. Calculate time-aware alpha scaling
"""
# Time-aware scaling: each merge gets less aggressive
if self.cfg.time_aware_scaling:
scale = 1.0 / np.sqrt(self.merge_count + 1)
adjusted_alpha = source_config.merge_alpha * scale
print(f"[protect] Time-aware scaling: {source_config.merge_alpha:.2f} Γ {scale:.3f} = {adjusted_alpha:.3f}")
else:
adjusted_alpha = source_config.merge_alpha
# MagMax: identify top 20% magnitude parameters to protect
if self.cfg.use_magmax and self.merge_count > 0:
print(f"[protect] Computing MagMax masks (protecting top 20% by magnitude)...")
state = target_model.state_dict()
for key, param in state.items():
if param.dim() >= 1:
flat = param.abs().flatten()
threshold = torch.quantile(flat.float(), 0.8)
self.magnitude_masks[key] = param.abs() >= threshold
return adjusted_alpha
def apply_protection(
self,
target_state: dict,
pre_merge_state: dict,
key: str,
) -> torch.Tensor:
"""
Apply all protection mechanisms to a fused parameter.
Called AFTER each parameter is fused, to constrain the change.
Protection stack (applied in order):
1. ARM steering (2602.03237) β steer delta toward gap, away from previous direction
2. Orthogonal projection (legacy fallback if ARM disabled)
3. OTMF masks (2511.19561) β protect task-specific weights
4. MagMax β protect top magnitude params (extra safety layer)
"""
fused = target_state[key]
original = pre_merge_state[key]
delta = fused - original
# --- ARM Steering (new, replaces orthogonal projection) ---
if self.cfg.use_arm_steering and self.arm_rotations:
# Find matching layer rotation
layer_prefix = ".".join(key.split(".")[:4])
for layer_name, rotation_info in self.arm_rotations.items():
if layer_prefix in layer_name:
delta = apply_arm_steering(
delta, rotation_info,
steering_strength=self.cfg.arm_steering_strength,
)
break
# --- Orthogonal Projection (legacy fallback) ---
elif self.cfg.use_orthogonal_projection and key in self.previous_deltas:
for prev_delta in self.previous_deltas[key]:
prev_flat = prev_delta.flatten().float()
delta_flat = delta.flatten().float()
dot = torch.dot(delta_flat, prev_flat)
norm_sq = torch.dot(prev_flat, prev_flat)
if norm_sq > 1e-10:
projection = (dot / norm_sq) * prev_flat
delta_flat = delta_flat - projection
delta = delta_flat.reshape(delta.shape).to(delta.dtype)
# --- OTMF Mask Protection (new) ---
if self.cfg.use_otmf_masks and key in self.otmf_masks:
mask = self.otmf_masks[key].to(delta.device)
# Transferable weights: full delta
# Task-specific weights: reduced delta (protect them)
delta = torch.where(
mask,
delta, # Transferable β allow full change
delta * (1.0 - self.cfg.otmf_protect_strength), # Protected β reduced
)
# --- MagMax Protection (extra safety layer) ---
if self.cfg.use_magmax and key in self.magnitude_masks:
mask = self.magnitude_masks[key]
delta = torch.where(mask, delta * 0.1, delta)
# Apply constrained delta
result = original + delta
return result
def after_merge(
self,
target_model: AutoModelForCausalLM,
pre_merge_state: dict,
pre_merge_activations: dict = None,
post_merge_activations: dict = None,
):
"""
Record the merge delta and compute protections for next merge.
Called AFTER each merge completes successfully.
Now also computes:
- ARM rotation vectors for next merge steering
- OTMF transferability masks for next merge
"""
current_state = target_model.state_dict()
for key in current_state:
if key in pre_merge_state:
delta = current_state[key].float() - pre_merge_state[key].float()
if delta.abs().max() > 1e-8:
if key not in self.previous_deltas:
self.previous_deltas[key] = []
if len(self.previous_deltas[key]) >= 2:
self.previous_deltas[key].pop(0)
self.previous_deltas[key].append(delta.cpu())
# --- Compute ARM rotations for next merge ---
if self.cfg.use_arm_steering and pre_merge_activations and post_merge_activations:
print("[protect] Computing ARM rotation vectors for next merge...")
self.arm_rotations = compute_arm_rotation(
pre_merge_activations,
post_merge_activations,
post_merge_activations, # Target = current state (for gap calculation)
)
# --- Compute OTMF masks for next merge ---
if self.cfg.use_otmf_masks and post_merge_activations:
print("[protect] Computing OTMF transferability masks...")
self.otmf_masks = compute_transferability_masks(
target_model,
post_merge_activations,
threshold=self.cfg.otmf_threshold,
)
self.merge_count += 1
print(f"[protect] Recorded merge delta #{self.merge_count} (ARM + OTMF ready for next)")
# ============================================================================
# MAIN ORCHESTRATOR
# ============================================================================
def is_vision_param(key: str, cfg: MergeConfig) -> bool:
"""
Check if a parameter belongs to the vision encoder.
Qwen3-VL-8B has a ViT vision encoder + merger projection on top of the
language model. We NEVER touch these during merging β they give us
browser agent and image understanding abilities for free.
Vision params start with prefixes like "visual." or "merger."
Language params start with "model.layers." or "model.embed_tokens." etc.
"""
for prefix in cfg.vision_skip_prefixes:
if key.startswith(prefix):
return True
return False
def get_source_by_stage(stage_name: str) -> Optional[ModelConfig]:
"""Get model config by stage name."""
stage_map = {
"deepseek": 0,
"mimo": 1,
"llama": 2,
"falcon": 3,
}
idx = stage_map.get(stage_name.lower())
if idx is not None and idx < len(SOURCES):
return SOURCES[idx]
return None
def load_model(config: ModelConfig, cfg: MergeConfig) -> tuple:
"""Load a model and its tokenizer/processor."""
print(f"\n[merge] Loading {config.name} ({config.hf_id})...")
# Qwen3-VL uses a processor (handles both text + vision), not just a tokenizer
if config.architecture == "transformer+vision":
try:
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
processor = AutoProcessor.from_pretrained(
config.hf_id,
trust_remote_code=config.trust_remote_code,
)
model = Qwen3VLForConditionalGeneration.from_pretrained(
config.hf_id,
torch_dtype=getattr(torch, cfg.dtype),
attn_implementation=cfg.attn_implementation,
device_map=cfg.device_map,
trust_remote_code=config.trust_remote_code,
)
# Use the tokenizer from the processor for text operations
tokenizer = processor.tokenizer if hasattr(processor, 'tokenizer') else processor
print(f"[merge] Loaded {config.name} (VL model): {sum(p.numel() for p in model.parameters()) / 1e9:.1f}B params")
# Count vision vs language params
vision_params = sum(
p.numel() for n, p in model.named_parameters()
if any(n.startswith(pfx) for pfx in cfg.vision_skip_prefixes)
)
lang_params = sum(p.numel() for p in model.parameters()) - vision_params
print(f"[merge] Language: {lang_params / 1e9:.1f}B | Vision: {vision_params / 1e9:.1f}B")
return model, tokenizer
except ImportError:
print("[merge] Qwen3VLForConditionalGeneration not available, falling back to AutoModel")
# Standard text-only models
tokenizer = AutoTokenizer.from_pretrained(
config.hf_id,
trust_remote_code=config.trust_remote_code,
)
model = AutoModelForCausalLM.from_pretrained(
config.hf_id,
torch_dtype=getattr(torch, cfg.dtype),
attn_implementation=cfg.attn_implementation,
device_map=cfg.device_map,
trust_remote_code=config.trust_remote_code,
)
print(f"[merge] Loaded {config.name}: {sum(p.numel() for p in model.parameters()) / 1e9:.1f}B params")
return model, tokenizer
def save_checkpoint(
model: AutoModelForCausalLM,
tokenizer: AutoTokenizer,
stage_name: str,
cfg: MergeConfig,
):
"""Save a checkpoint after a successful merge stage."""
ckpt_dir = Path(cfg.checkpoint_dir) / f"after_{stage_name}"
ckpt_dir.mkdir(parents=True, exist_ok=True)
print(f"[merge] Saving checkpoint to {ckpt_dir}...")
model.save_pretrained(ckpt_dir)
tokenizer.save_pretrained(ckpt_dir)
print(f"[merge] Checkpoint saved: {ckpt_dir}")
return str(ckpt_dir)
# ============================================================================
# RESIDUAL BANK β Save what was lost during each merge
# ============================================================================
class ResidualBank:
"""
Saves the knowledge that gets lost during each merge so it can
be recovered later.
When we blend at alpha=0.10:
merged = target + alpha * M * (transported - target)
We LOSE:
target_residual = target_original - merged (what target lost)
source_residual = source_original - merged (what source lost)
These residuals are saved to disk. Later they can be:
1. Fed back during the healing fine-tune (as training signal)
2. Re-injected via a small LoRA adapter
3. Used to diagnose which merge caused a specific knowledge loss
4. Re-applied at a lower alpha if we want more of that model
Think of it like saving the sawdust when you cut wood β you might
need to glue some of it back later.
"""
def __init__(self, cfg: MergeConfig):
self.cfg = cfg
self.residual_dir = Path(cfg.checkpoint_dir) / "residuals"
self.residual_dir.mkdir(parents=True, exist_ok=True)
self.residual_index = {} # stage β {path, stats}
def save_residuals(
self,
stage_name: str,
pre_merge_target_state: dict,
source_state: dict,
post_merge_state: dict,
source_config: ModelConfig,
):
"""
Compute and save what was lost from both target and source.
Saves two files per merge stage:
- target_residual: what the target model lost
- source_residual: what the source model didn't fully contribute
Also saves stats so we know WHERE the biggest losses were
(which layers, which type of weights).
"""
stage_dir = self.residual_dir / stage_name
stage_dir.mkdir(parents=True, exist_ok=True)
target_residual = {}
source_residual = {}
stats = {
"stage": stage_name,
"source_model": source_config.name,
"target_loss_by_layer": {},
"source_loss_by_layer": {},
"total_target_loss": 0.0,
"total_source_loss": 0.0,
"biggest_losses": [],
}
for key in post_merge_state:
merged_w = post_merge_state[key].float()
# What the target lost
if key in pre_merge_target_state:
original_target = pre_merge_target_state[key].float()
t_residual = original_target - merged_w
t_loss = t_residual.abs().mean().item()
if t_loss > 1e-6: # Only save meaningful residuals
target_residual[key] = t_residual.to(torch.bfloat16).cpu()
stats["total_target_loss"] += t_loss
# Track per-layer losses
layer_name = ".".join(key.split(".")[:4])
if layer_name not in stats["target_loss_by_layer"]:
stats["target_loss_by_layer"][layer_name] = 0.0
stats["target_loss_by_layer"][layer_name] += t_loss
# What the source lost (what didn't make it into the merge)
if key in source_state:
original_source = source_state[key].float()
s_residual = original_source - merged_w
s_loss = s_residual.abs().mean().item()
if s_loss > 1e-6:
source_residual[key] = s_residual.to(torch.bfloat16).cpu()
stats["total_source_loss"] += s_loss
layer_name = ".".join(key.split(".")[:4])
if layer_name not in stats["source_loss_by_layer"]:
stats["source_loss_by_layer"][layer_name] = 0.0
stats["source_loss_by_layer"][layer_name] += s_loss
# Find the biggest losses (most knowledge dropped)
all_losses = []
for key in target_residual:
loss_magnitude = target_residual[key].float().abs().mean().item()
all_losses.append({"param": key, "side": "target", "loss": loss_magnitude})
for key in source_residual:
loss_magnitude = source_residual[key].float().abs().mean().item()
all_losses.append({"param": key, "side": "source", "loss": loss_magnitude})
all_losses.sort(key=lambda x: x["loss"], reverse=True)
stats["biggest_losses"] = all_losses[:20] # Top 20 biggest losses
# Save to disk
torch.save(target_residual, stage_dir / "target_residual.pt")
torch.save(source_residual, stage_dir / "source_residual.pt")
import json
with open(stage_dir / "residual_stats.json", "w") as f:
json.dump(stats, f, indent=2, default=str)
self.residual_index[stage_name] = {
"path": str(stage_dir),
"target_params_saved": len(target_residual),
"source_params_saved": len(source_residual),
"total_target_loss": stats["total_target_loss"],
"total_source_loss": stats["total_source_loss"],
}
print(f"[residual] Saved residuals for {stage_name}:")
print(f" Target lost: {len(target_residual)} params (avg loss: {stats['total_target_loss']:.4f})")
print(f" Source lost: {len(source_residual)} params (avg loss: {stats['total_source_loss']:.4f})")
print(f" Top loss: {all_losses[0]['param']} ({all_losses[0]['side']}, {all_losses[0]['loss']:.4f})" if all_losses else "")
print(f" Saved to: {stage_dir}")
def load_residuals(self, stage_name: str) -> tuple:
"""
Load saved residuals for a stage.
Returns:
(target_residual_dict, source_residual_dict)
"""
stage_dir = self.residual_dir / stage_name
target_residual = torch.load(stage_dir / "target_residual.pt", weights_only=True)
source_residual = torch.load(stage_dir / "source_residual.pt", weights_only=True)
return target_residual, source_residual
def reinject_residuals(
self,
model: AutoModelForCausalLM,
stage_name: str,
side: str = "both",
strength: float = 0.3,
) -> AutoModelForCausalLM:
"""
Re-inject saved residuals back into a model.
This adds back some of what was lost. Use a low strength (0.1-0.3)
to gently recover knowledge without undoing the merge.
Args:
model: The model to inject into
stage_name: Which merge stage's residuals to use
side: "target", "source", or "both"
strength: How much to add back (0=nothing, 1=full residual)
"""
print(f"[residual] Re-injecting {stage_name} residuals (side={side}, strength={strength})...")
target_residual, source_residual = self.load_residuals(stage_name)
state = model.state_dict()
injected = 0
if side in ("target", "both"):
for key, residual in target_residual.items():
if key in state:
state[key] = state[key] + strength * residual.to(state[key].device).to(state[key].dtype)
injected += 1
if side in ("source", "both"):
for key, residual in source_residual.items():
if key in state:
state[key] = state[key] + strength * residual.to(state[key].device).to(state[key].dtype)
injected += 1
model.load_state_dict(state)
print(f"[residual] Re-injected {injected} params at {strength:.0%} strength")
return model
def get_healing_targets(self, top_n: int = 50) -> list:
"""
Get the parameters with the biggest losses across ALL merges.
These are the params that the healing fine-tune should focus on.
Feed this to the LoRA target_modules to make healing smarter.
"""
import json
all_losses = []
for stage_name in self.residual_index:
stage_dir = self.residual_dir / stage_name
stats_file = stage_dir / "residual_stats.json"
if stats_file.exists():
with open(stats_file) as f:
stats = json.load(f)
for loss in stats.get("biggest_losses", []):
loss["stage"] = stage_name
all_losses.append(loss)
all_losses.sort(key=lambda x: x["loss"], reverse=True)
# Extract unique layer/module names for LoRA targeting
target_modules = set()
for loss in all_losses[:top_n]:
param = loss["param"]
# Extract the module type (q_proj, k_proj, gate_proj, etc.)
parts = param.split(".")
for part in parts:
if part.endswith("_proj") or part in ("gate_proj", "up_proj", "down_proj"):
target_modules.add(part)
print(f"[residual] Top healing targets (from {len(all_losses)} total losses):")
for loss in all_losses[:5]:
print(f" {loss['param']} ({loss['side']}, stage={loss['stage']}, loss={loss['loss']:.4f})")
print(f" β Suggested LoRA targets: {sorted(target_modules)}")
return list(target_modules)
def run_single_merge(
target_model: AutoModelForCausalLM,
target_tokenizer: AutoTokenizer,
source_config: ModelConfig,
cfg: MergeConfig,
protection: MergeProtection,
residual_bank: ResidualBank = None,
calibration_data: list = None,
baseline_perplexity: float = None,
merged_sources: list = None,
) -> dict:
"""
Run a single merge: source β target.
Full pipeline for one merge step:
1. Load source model
2. Inject canary into source
3. Extract activations from both
4. Compute transport plans
5. Apply merge protection
6. Fuse weights
7. Apply post-merge protection
8. Validate
Returns:
Dict with merge results, validation results, and status
"""
if merged_sources is None:
merged_sources = []
stage_name = source_config.name
print(f"\n{'=' * 70}")
print(f"MERGE STAGE: {stage_name} β target")
print(f"Risk level: {source_config.merge_risk.upper()}")
print(f"{'=' * 70}")
result = {
"stage": stage_name,
"status": "pending",
"validation": None,
"checkpoint": None,
}
# --- Step 1: Load source model ---
source_model, source_tokenizer = load_model(source_config, cfg)
# --- Step 2: Inject canary into source ---
if stage_name in CANARY_FACTS:
print(f"\n[merge] Injecting canary fact into {stage_name}...")
source_model = inject_canary(source_model, source_tokenizer, stage_name)
# --- Step 3: Load calibration data (if not provided) ---
if calibration_data is None:
calibration_data = load_calibration_data(cfg, target_tokenizer)
# --- Step 4: Extract two-sided activations (pre + post per projection) ---
print(f"\n[merge] Extracting source activations (two-sided)...")
source_activations = extract_activations(source_model, calibration_data)
print(f"\n[merge] Extracting target activations (two-sided)...")
pre_merge_target_activations = extract_activations(target_model, calibration_data)
# --- Step 4.5: Mergeability pre-check (2601.22285) ---
if cfg.use_mergeability_check:
mergeability = compute_mergeability_score(
source_activations, pre_merge_target_activations, source_config
)
result["mergeability"] = mergeability
if mergeability["overall"] < cfg.mergeability_min_score:
print(f"\n[merge] β Mergeability score {mergeability['overall']:.2f} below threshold {cfg.mergeability_min_score}")
print(f"[merge] β {mergeability['recommendation']}")
result["status"] = "skipped_low_mergeability"
if "distillation_fallback" in source_config.special_handling:
result["fallback"] = "distillation"
del source_model, source_activations, pre_merge_target_activations
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
return result
# --- Step 5: Compute transport plans ---
transport_plans = compute_transport_plans(
source_activations, pre_merge_target_activations, cfg
)
# --- Step 5.5: RAM RL-weight disentanglement (2601.13572) ---
use_ram = (
cfg.use_ram_disentangle
and source_config.architecture in ("transformer", "transformer+mtp")
and source_config.merge_risk in ("low", "medium")
and any(kw in source_config.name.lower() for kw in ["r1", "rl", "rlhf", "grpo"])
)
# --- Step 6: Pre-merge protection ---
adjusted_alpha = protection.before_merge(target_model, source_config)
# Override source alpha with time-adjusted value
source_config_adjusted = copy.copy(source_config)
source_config_adjusted.merge_alpha = adjusted_alpha
# Save pre-merge state for protection
pre_merge_state = {k: v.clone().cpu() for k, v in target_model.state_dict().items()}
# --- Step 7: Fuse weights ---
if use_ram:
# RAM path: disentangle RL weights, merge with preservation
print(f"\n[merge] Using RAM RL-preservation for {stage_name}...")
try:
# Try loading the base (pre-RL) model for disentanglement
base_hf_id = source_config.hf_id.replace("-RL", "").replace("-R1-0528", "")
print(f"[merge] Loading base model for RAM: {base_hf_id}")
base_model = AutoModelForCausalLM.from_pretrained(
base_hf_id,
torch_dtype=getattr(torch, cfg.dtype),
device_map=cfg.device_map,
trust_remote_code=source_config.trust_remote_code,
)
shared_mask, rl_mask = disentangle_rl_weights(
source_model, base_model, cfg.ram_rl_threshold
)
# Fuse with RL preservation
target_state = merge_with_rl_preservation(
target_model.state_dict(),
source_model.state_dict(),
shared_mask, rl_mask,
shared_alpha=cfg.ram_shared_alpha * (adjusted_alpha / source_config.merge_alpha),
rl_alpha=cfg.ram_rl_alpha,
)
target_model.load_state_dict(target_state)
del base_model
print(f"[merge] RAM merge complete for {stage_name}")
except Exception as e:
print(f"[merge] RAM failed ({e}), falling back to standard T&M merge")
target_model = fuse_weights(
source_model, target_model, transport_plans,
source_config_adjusted, cfg,
target_activations=pre_merge_target_activations,
)
else:
# Standard T&M path (two-sided + top-k masked fusion, paper Eq 14)
target_model = fuse_weights(
source_model, target_model, transport_plans,
source_config_adjusted, cfg,
target_activations=pre_merge_target_activations,
)
# --- Step 7.5: Theseus fallback check (2602.12952) ---
# If T&M merge produced poor activation alignment, try Theseus
if cfg.use_theseus_fallback and source_config.merge_risk == "high":
print(f"\n[merge] Checking if Theseus fallback needed for {stage_name}...")
post_activations = extract_activations(target_model, calibration_data[:50]) # Quick check
# Compare post-merge activations to pre-merge β if too similar, T&M didn't work
alignment_scores = []
for key in post_activations:
if key in pre_merge_target_activations:
cos = torch.nn.functional.cosine_similarity(
post_activations[key].float().mean(0, keepdim=True),
pre_merge_target_activations[key].float().mean(0, keepdim=True),
)
alignment_scores.append(cos.item())
avg_change = 1.0 - np.mean(alignment_scores) if alignment_scores else 0.0
print(f"[merge] Activation change from merge: {avg_change:.4f}")
if avg_change < 0.01:
print(f"[merge] β T&M had minimal effect β activating Theseus fallback")
# Restore pre-merge state and try Theseus instead
target_model.load_state_dict(pre_merge_state)
try:
base_model = AutoModelForCausalLM.from_pretrained(
source_config.hf_id.split("/")[0] + "/" + source_config.hf_id.split("/")[1].split("-")[0],
torch_dtype=getattr(torch, cfg.dtype),
device_map=cfg.device_map,
trust_remote_code=source_config.trust_remote_code,
)
target_model = transport_task_vector_theseus(
source_model, base_model, target_model,
source_activations, pre_merge_target_activations,
alpha=cfg.theseus_alpha,
)
del base_model
print(f"[merge] Theseus transport complete for {stage_name}")
except Exception as e:
print(f"[merge] Theseus also failed ({e}). Using original T&M result.")
# Re-apply T&M result
target_model = fuse_weights(
source_model, target_model, transport_plans,
source_config_adjusted, cfg,
target_activations=pre_merge_target_activations,
)
# --- Step 8: Apply post-merge protection (ARM + OTMF + MagMax) ---
# Skip vision encoder params β they weren't merged, so don't "protect" them
if protection.merge_count > 0:
print(f"\n[merge] Applying sequential merge protection (ARM + OTMF + MagMax)...")
target_state = target_model.state_dict()
protected_count = 0
vision_skipped = 0
for key in target_state:
if is_vision_param(key, cfg):
vision_skipped += 1
continue # Don't touch vision encoder
if key in pre_merge_state:
protected_param = protection.apply_protection(
target_state, pre_merge_state, key
)
target_state[key] = protected_param
protected_count += 1
target_model.load_state_dict(target_state)
print(f"[merge] Protected {protected_count} language params (skipped {vision_skipped} vision params)")
# --- Step 8.5: Extract post-merge activations for ARM/OTMF ---
post_merge_activations = extract_activations(target_model, calibration_data[:100])
# Record this merge's delta + compute ARM/OTMF for next merge
protection.after_merge(
target_model, pre_merge_state,
pre_merge_activations=pre_merge_target_activations,
post_merge_activations=post_merge_activations,
)
# --- Step 8.8: Save residuals (what was lost from both sides) ---
if residual_bank is not None:
print(f"\n[merge] Saving residuals for {stage_name}...")
residual_bank.save_residuals(
stage_name=stage_name,
pre_merge_target_state=pre_merge_state,
source_state={k: v.cpu() for k, v in source_model.state_dict().items()},
post_merge_state={k: v.cpu() for k, v in target_model.state_dict().items()},
source_config=source_config,
)
# --- Step 9: Free source model memory ---
del source_model, source_activations, pre_merge_target_activations
del transport_plans, post_merge_activations
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
# --- Step 10: Validate ---
merged_sources.append(stage_name)
validation = validate_merged_model(
target_model, target_tokenizer,
merged_sources, cfg,
baseline_perplexity=baseline_perplexity,
)
result["validation"] = validation
result["merged_sources"] = merged_sources.copy()
# --- Kill criteria check ---
if not validation["overall"]:
print(f"\n[merge] β VALIDATION FAILED for {stage_name}")
print(f"[merge] Kill criteria triggered β consider aborting")
result["status"] = "failed"
# Check if we should try distillation fallback
if "distillation_fallback" in source_config.special_handling:
print(f"[merge] {stage_name} has distillation fallback available")
result["fallback"] = "distillation"
else:
print(f"\n[merge] β {stage_name} merge PASSED validation")
result["status"] = "passed"
return result
def run_pipeline(
stages: list[str],
cfg: MergeConfig = None,
) -> dict:
"""
Run the full merge pipeline.
Args:
stages: List of stage names to run, e.g. ["deepseek"] or
["deepseek", "mimo", "llama", "falcon"]
cfg: Merge configuration (uses defaults if None)
Returns:
Dict with overall results, per-stage results, and final model path
"""
if cfg is None:
cfg = MergeConfig()
print("\n" + "=" * 70)
print("TD LANG ENGINE β Transport and Merge Pipeline")
print(f"Target: {TARGET.name} ({TARGET.hf_id})")
if TARGET.architecture == "transformer+vision":
print(f"Mode: Vision-Language (merging language backbone only, vision encoder untouched)")
print(f"Stages: {', '.join(stages)}")
print(f"Output: {cfg.output_dir}")
print("=" * 70)
# Setup
try:
setup_tm_repo(cfg)
except FileNotFoundError as e:
print(f"\nβ {e}")
print("Continuing with fallback implementation...")
# Create output directories
Path(cfg.output_dir).mkdir(parents=True, exist_ok=True)
Path(cfg.checkpoint_dir).mkdir(parents=True, exist_ok=True)
# --- Load target model ---
target_model, target_tokenizer = load_model(TARGET, cfg)
# --- Inject canary into target (Qwen3's own canary) ---
if "Qwen3-VL-8B" in CANARY_FACTS:
print("\n[pipeline] Injecting canary into base Qwen3-8B...")
target_model = inject_canary(target_model, target_tokenizer, "Qwen3-VL-8B")
# --- Compute baseline perplexity ---
print("\n[pipeline] Computing baseline perplexity...")
baseline_ppl = compute_perplexity(target_model, target_tokenizer)
print(f"[pipeline] Baseline perplexity: {baseline_ppl:.2f}")
# --- Load calibration data once ---
calibration_data = load_calibration_data(cfg, target_tokenizer)
# --- Initialize merge protection + residual bank ---
protection = MergeProtection(cfg)
residual_bank = ResidualBank(cfg)
# --- Run each merge stage ---
pipeline_results = {
"stages": {},
"baseline_perplexity": baseline_ppl,
"final_checkpoint": None,
"residuals": {},
"overall_status": "pending",
}
merged_sources = []
all_passed = True
for stage_name in stages:
source_config = get_source_by_stage(stage_name)
if source_config is None:
print(f"\nβ Unknown stage: {stage_name}, skipping")
continue
# --- Wasserstein pre-check for high-risk models ---
if "check_wasserstein_first" in source_config.special_handling:
print(f"\n[pipeline] Running Wasserstein pre-check for {source_config.name}...")
# TODO: Implement Wasserstein distance pre-check
# If distance is too high, skip to distillation fallback
print("[pipeline] Pre-check: proceeding (TODO: implement distance check)")
# Run the merge (with residual bank to save what's lost)
stage_result = run_single_merge(
target_model, target_tokenizer,
source_config, cfg,
protection,
residual_bank=residual_bank,
calibration_data=calibration_data,
baseline_perplexity=baseline_ppl,
merged_sources=merged_sources,
)
pipeline_results["stages"][stage_name] = stage_result
if stage_result["status"] == "passed":
# Save checkpoint
ckpt_path = save_checkpoint(
target_model, target_tokenizer, stage_name, cfg
)
stage_result["checkpoint"] = ckpt_path
pipeline_results["final_checkpoint"] = ckpt_path
else:
all_passed = False
print(f"\n[pipeline] Stage {stage_name} FAILED")
# Decision: abort or continue?
if source_config.merge_risk == "high":
print(f"[pipeline] High-risk model failed β skipping (will use distillation)")
# Don't abort the whole pipeline, just skip this model
continue
else:
print(f"[pipeline] ABORTING pipeline β non-high-risk model failed")
pipeline_results["overall_status"] = f"aborted_at_{stage_name}"
break
# --- Save residual index ---
pipeline_results["residuals"] = residual_bank.residual_index
if residual_bank.residual_index:
print(f"\n[pipeline] Residual bank: {len(residual_bank.residual_index)} stages saved")
for stage, info in residual_bank.residual_index.items():
print(f" {stage}: target lost {info['total_target_loss']:.4f}, source lost {info['total_source_loss']:.4f}")
# Identify which modules need the most healing
healing_targets = residual_bank.get_healing_targets(top_n=50)
pipeline_results["suggested_healing_targets"] = healing_targets
# --- Save final model ---
if pipeline_results["final_checkpoint"]:
final_dir = Path(cfg.output_dir) / "final"
final_dir.mkdir(parents=True, exist_ok=True)
target_model.save_pretrained(final_dir)
target_tokenizer.save_pretrained(final_dir)
pipeline_results["final_model_path"] = str(final_dir)
print(f"\n[pipeline] Final model saved to {final_dir}")
if all_passed:
pipeline_results["overall_status"] = "all_passed"
elif pipeline_results["overall_status"] == "pending":
pipeline_results["overall_status"] = "partial"
# --- Print final summary ---
print("\n" + "=" * 70)
print("PIPELINE SUMMARY")
print("=" * 70)
for stage_name, stage_result in pipeline_results["stages"].items():
status = stage_result["status"]
emoji = "β" if status == "passed" else "β"
print(f" {emoji} {stage_name}: {status}")
print(f"\n Overall: {pipeline_results['overall_status']}")
if residual_bank.residual_index:
print(f"\n Residuals saved for: {', '.join(residual_bank.residual_index.keys())}")
print(f" To recover lost knowledge later:")
print(f" python -m td_lang.engine --reinject <stage> --strength 0.2")
print("=" * 70)
return pipeline_results
|