Upload 148 files

hugging/td_fuse/merge_klear.py

@@ -128,14 +128,18 @@ def direct_merge(target_model, source_state, alpha=0.3):
     """
     target_state = target_model.state_dict()
 
-    # Check vocab sizes
+    # Check vocab sizes — try multiple key patterns
     target_vocab = None
     source_vocab = None
-    for key in ["model.embed_tokens.weight", "embed_tokens.weight"]:
-        if key in target_state:
+    for key in target_state:
+        if "embed_tokens" in key:
             target_vocab = target_state[key].shape[0]
-        if key in source_state:
+            break
+    for key in source_state:
+        if "embed_tokens" in key:
             source_vocab = source_state[key].shape[0]
+            break
+    print(f"[merge6] Target vocab: {target_vocab}, Source vocab: {source_vocab}")
     skip_embeddings = (target_vocab != source_vocab) if (target_vocab and source_vocab) else False
     if skip_embeddings:
         print(f"[merge6] Vocab mismatch ({target_vocab} vs {source_vocab}) — skipping embeddings")
@@ -146,6 +150,13 @@ def direct_merge(target_model, source_state, alpha=0.3):
     skipped = 0
     total = len(target_state)
 
+    # Debug: show first 5 keys from each to verify naming
+    target_keys = list(target_state.keys())
+    source_keys = list(source_state.keys())
+    print(f"[merge6] Target has {len(target_keys)} params, Source has {len(source_keys)} params")
+    print(f"[merge6] Target first 5 keys: {target_keys[:5]}")
+    print(f"[merge6] Source first 5 keys: {source_keys[:5]}")
+
     # Save original think token embeddings for restoration
     think_embeds = {}
     for key in target_state:
@@ -163,16 +174,27 @@ def direct_merge(target_model, source_state, alpha=0.3):
             skipped += 1
             continue
 
-        # Find source key — same name for same architecture
-        if target_key not in source_state:
-            # Try without "model." prefix or with it
-            alt_key = target_key.replace("model.", "", 1) if target_key.startswith("model.") else f"model.{target_key}"
-            if alt_key not in source_state:
-                skipped += 1
-                continue
-            source_key = alt_key
-        else:
-            source_key = target_key
+        # Find source key — handle Qwen3-VL prefix differences
+        # Qwen3-VL keys: model.language_model.layers.0... or model.model.layers.0...
+        # Klear keys:    model.layers.0...
+        source_key = None
+        candidates = [
+            target_key,                                              # exact match
+            target_key.replace("language_model.", ""),               # strip language_model.
+            target_key.replace("model.model.", "model."),            # model.model.X -> model.X
+            target_key.replace("model.", "", 1),                     # strip first model.
+            f"model.{target_key}",                                   # add model. prefix
+        ]
+        for candidate in candidates:
+            if candidate in source_state:
+                source_key = candidate
+                break
+
+        if source_key is None:
+            skipped += 1
+            if skipped <= 10:
+                print(f"  [skip] No source match: {target_key}")
+            continue
 
         target_w = target_state[target_key]
         source_w = source_state[source_key]