qwen 3.5 vl + more baichuan patch

gptqmodel/models/definitions/base_qwen3_vl.py

@@ -54,30 +54,53 @@ class BaseQwen3VLGPTQ(BaseQModel):
 
     require_load_processor = True
 
+    def _language_model_root(self):
+        if hasattr(self.model, "language_model"):
+            return self.model
+        if hasattr(self.model, "model") and hasattr(self.model.model, "language_model"):
+            return self.model.model
+        raise AttributeError(
+            f"{type(self.model).__name__} does not expose a language_model root"
+        )
+
     def pre_quantize_generate_hook_start(self):
-        self.model.language_model.embed_tokens = move_to(self.model.language_model.embed_tokens, device=self.quantize_config.device)
-        self.model.language_model.rotary_emb = move_to(self.model.language_model.rotary_emb, device=self.quantize_config.device)
-        self.model.visual = move_to(self.model.visual, device=self.quantize_config.device)
+        model_root = self._language_model_root()
+        model_root.language_model.embed_tokens = move_to(
+            model_root.language_model.embed_tokens,
+            device=self.quantize_config.device,
+        )
+        model_root.language_model.rotary_emb = move_to(
+            model_root.language_model.rotary_emb,
+            device=self.quantize_config.device,
+        )
+        model_root.visual = move_to(model_root.visual, device=self.quantize_config.device)
 
     def pre_quantize_generate_hook_end(self):
+        model_root = self._language_model_root()
         if self.quantize_config.offload_to_disk:
-            offload_to_disk(model=self.model.language_model,
-                            module=self.model.language_model.embed_tokens,
+            offload_to_disk(model=model_root.language_model,
+                            module=model_root.language_model.embed_tokens,
                             disk_path=self.quantize_config.offload_to_disk_path,
                             )
-            offload_to_disk(model=self.model.language_model,
-                            module=self.model.language_model.rotary_emb,
+            offload_to_disk(model=model_root.language_model,
+                            module=model_root.language_model.rotary_emb,
                             disk_path=self.quantize_config.offload_to_disk_path,
                             )
             offload_to_disk(model=self.model,
-                            module=self.model.visual,
+                            module=model_root.visual,
                             disk_path=self.quantize_config.offload_to_disk_path,
                             )
             return
 
-        self.model.language_model.embed_tokens = move_to(self.model.language_model.embed_tokens, device=CPU)
-        self.model.language_model.rotary_emb = move_to(self.model.language_model.rotary_emb, device=CPU)
-        self.model.visual = move_to(self.model.visual, device=CPU)
+        model_root.language_model.embed_tokens = move_to(
+            model_root.language_model.embed_tokens,
+            device=CPU,
+        )
+        model_root.language_model.rotary_emb = move_to(
+            model_root.language_model.rotary_emb,
+            device=CPU,
+        )
+        model_root.visual = move_to(model_root.visual, device=CPU)
 
     @staticmethod
     def process_vision_info(

gptqmodel/utils/structure.py

@@ -617,6 +617,21 @@ def _module_all_meta(mod: nn.Module) -> bool:
 def _is_leaf(mod: nn.Module) -> bool:
     return next(mod.named_children(), None) is None
 
+def _prune_stale_meta_buffers(shell_sub: nn.Module, turtle_sub: nn.Module) -> int:
+    removed = 0
+    turtle_buffers = {
+        name for name, _ in turtle_sub.named_buffers(recurse=False)
+    }
+    for name, buf in list(shell_sub.named_buffers(recurse=False)):
+        if name in turtle_buffers:
+            continue
+        if not _is_meta_tensor(buf):
+            continue
+        if name in getattr(shell_sub, "_buffers", {}):
+            del shell_sub._buffers[name]
+            removed += 1
+    return removed
+
 def alias_all_from_turtle_if_meta(
     shell_model: nn.Module,
     turtle_model: nn.Module,
@@ -640,14 +655,18 @@ def alias_all_from_turtle_if_meta(
     for qname, shell_sub in list(shell_model.named_modules()):
         if not qname:  # skip root
             continue
-        if not _is_leaf(shell_sub):
-            continue
-        if not _module_all_meta(shell_sub):
-            continue
 
         turtle_sub = turtle_map.get(qname, None)
         if turtle_sub is None:
-            # log.info(f"Module: Skipped {qname}: not found in turtle model")
+            continue
+
+        removed = _prune_stale_meta_buffers(shell_sub, turtle_sub)
+        if removed:
+            log.info(f"Module: Pruned {removed} stale meta buffer(s) from {qname}")
+
+        if not _is_leaf(shell_sub):
+            continue
+        if not _module_all_meta(shell_sub):
             continue
 
         if require_class_match and (shell_sub.__class__ is not turtle_sub.__class__):

model_test.py

@@ -0,0 +1,13 @@
+import importlib.util
+from pathlib import Path
+
+
+_MODEL_TEST_PATH = Path(__file__).resolve().parent / "tests" / "models" / "model_test.py"
+_SPEC = importlib.util.spec_from_file_location("_model_test_impl", _MODEL_TEST_PATH)
+if _SPEC is None or _SPEC.loader is None:
+    raise ImportError(f"Unable to load ModelTest from {_MODEL_TEST_PATH}")
+
+_MODULE = importlib.util.module_from_spec(_SPEC)
+_SPEC.loader.exec_module(_MODULE)
+
+globals().update({name: getattr(_MODULE, name) for name in dir(_MODULE) if not name.startswith("_")})

tests/model_test.py

@@ -0,0 +1 @@
+from models.model_test import *  # noqa: F401,F403

tests/models/__init__.py

@@ -0,0 +1 @@
+# Package marker for pytest collection to avoid basename collisions.

tests/models/awq/__init__.py

@@ -0,0 +1 @@
+# Package marker for pytest collection to avoid basename collisions.