Upload 140 files

hugging/save_checkpoint.py

@@ -0,0 +1,82 @@
+"""
+Save TD checkpoints to HuggingFace.
+
+Usage:
+    python3 save_checkpoint.py                  # saves latest checkpoint
+    python3 save_checkpoint.py after_mimo       # saves specific checkpoint
+    python3 save_checkpoint.py all              # saves all checkpoints
+"""
+
+import sys
+import os
+from pathlib import Path
+from huggingface_hub import HfApi, login
+
+TOKEN = os.environ.get("HF_TOKEN", "")
+REPO = "td-builder/td-qwen3vl-v1"
+CKPT_DIR = Path("td_fuse_checkpoints")
+
+def upload_checkpoint(api, name):
+    ckpt_path = CKPT_DIR / name
+    if not ckpt_path.exists():
+        print(f"  ERROR: {ckpt_path} doesn't exist")
+        return False
+
+    safetensors = ckpt_path / "model.safetensors"
+    if not safetensors.exists():
+        print(f"  ERROR: No model.safetensors in {ckpt_path}")
+        return False
+
+    size_gb = sum(f.stat().st_size for f in ckpt_path.rglob("*") if f.is_file()) / 1e9
+    print(f"  Uploading {name} ({size_gb:.1f} GB) to {REPO}/{name}/...")
+
+    api.upload_folder(
+        folder_path=str(ckpt_path),
+        path_in_repo=name,
+        repo_id=REPO,
+        commit_message=f"Checkpoint: {name}",
+    )
+    print(f"  Done: {name}")
+    return True
+
+
+def main():
+    login(token=TOKEN)
+    api = HfApi()
+
+    target = sys.argv[1] if len(sys.argv) > 1 else None
+
+    if not CKPT_DIR.exists():
+        print(f"No checkpoint directory found at {CKPT_DIR}")
+        sys.exit(1)
+
+    # List available checkpoints
+    checkpoints = sorted([d.name for d in CKPT_DIR.iterdir() if d.is_dir() and (d / "model.safetensors").exists()])
+
+    if not checkpoints:
+        print("No checkpoints found (need model.safetensors in each folder)")
+        sys.exit(1)
+
+    print(f"Available checkpoints: {', '.join(checkpoints)}")
+
+    if target == "all":
+        # Upload everything
+        for name in checkpoints:
+            upload_checkpoint(api, name)
+    elif target:
+        # Upload specific one
+        if target not in checkpoints:
+            print(f"Checkpoint '{target}' not found. Available: {', '.join(checkpoints)}")
+            sys.exit(1)
+        upload_checkpoint(api, target)
+    else:
+        # Upload the latest (most recently modified)
+        latest = max(checkpoints, key=lambda n: (CKPT_DIR / n).stat().st_mtime)
+        print(f"Uploading latest: {latest}")
+        upload_checkpoint(api, latest)
+
+    print("\nAll done! Checkpoints saved to HuggingFace.")
+
+
+if __name__ == "__main__":
+    main()

hugging/td_fuse/merge.py

@@ -1052,9 +1052,12 @@ def run_pipeline(
         target_model, target_tokenizer = load_model(TARGET, cfg)
 
     # --- Inject canary into target (Qwen3's own canary) ---
-    if "Qwen3-VL-8B" in CANARY_FACTS:
+    # Skip if loading from checkpoint (canary already injected in previous merge)
+    if "Qwen3-VL-8B" in CANARY_FACTS and not base_checkpoint:
         print("\n[pipeline] Injecting canary into base Qwen3-8B...")
         target_model = inject_canary(target_model, target_tokenizer, "Qwen3-VL-8B")
+    elif base_checkpoint:
+        print("\n[pipeline] Skipping canary injection (already in checkpoint)")
 
     # --- Compute baseline perplexity ---
     print("\n[pipeline] Computing baseline perplexity...")