[FIX] LazyTurtle tensor-key alias resolution for nested HF weight renames (#2835)

* fix test_internvl_chat.py

Signed-off-by: ZX-ModelCloud <zx@modelcloud.ai>

* fix _resolve_checkpoint_tensor_name()

Signed-off-by: ZX-ModelCloud <zx@modelcloud.ai>

* cleanup

Signed-off-by: ZX-ModelCloud <zx@modelcloud.ai>

---------

Signed-off-by: ZX-ModelCloud <zx@modelcloud.ai>

Author: ZX-ModelCloud

gptqmodel/utils/structure.py

@@ -1312,17 +1312,26 @@ def _resolve_checkpoint_tensor_name(self, module_path: str, rel_name: str) -> st
         full_name = self._join_tensor_name(module_path, rel_name)
         candidates: list[str] = []
         seen: set[str] = set()
+
+        def add_candidate(candidate_name: str) -> None:
+            # Keep module-tree aliases attached to the fully resolved tensor key.
+            # Some HF mappings rewrite nested relative names (for example
+            # `spatial_linear.ln` -> `spatial_linear.3`), so alias expansion has
+            # to happen after `module_path` and `rel_name` are joined.
+            if candidate_name not in seen:
+                seen.add(candidate_name)
+                candidates.append(candidate_name)
+            for alias in self._module_tree_name_aliases(candidate_name):
+                if alias in seen:
+                    continue
+                seen.add(alias)
+                candidates.append(alias)
+
         for candidate_path in self._candidate_module_paths(module_path, allow_empty=True):
             for aliased_path in self._runtime_to_checkpoint_alias_candidates(candidate_path):
                 candidate_name = self._join_tensor_name(aliased_path, rel_name)
-                if candidate_name not in seen:
-                    seen.add(candidate_name)
-                    candidates.append(candidate_name)
-                for alias in self._module_tree_name_aliases(candidate_name):
-                    if alias in seen:
-                        continue
-                    seen.add(alias)
-                    candidates.append(alias)
+                for aliased_name in self._runtime_to_checkpoint_alias_candidates(candidate_name):
+                    add_candidate(aliased_name)
 
         for candidate in candidates:
             if candidate in self._weight_map:

tests/models/test_internvl_chat.py

@@ -50,6 +50,7 @@ def test_internvl_chat(self):
 
         inputs = model.prepare_inputs_for_conversation(messages)
         inputs = model.move_input_capture_example(inputs, model.device)
+        model.model.img_context_token_id = inputs.pop("img_context_token_id")
         inputs.pop("eos_token_id")
 
         output_text = self.generate_stable_with_limit(

tests/test_lazy_turtle_conversion_mapping.py

@@ -317,6 +317,30 @@ def test_lazy_turtle_applies_reversed_weight_renamings_with_capturing_groups(tmp
     assert turtle._resolve_checkpoint_tensor_name("timm_model.backbone.conv", "weight") == "backbone.conv.weight"
 
 
+def test_lazy_turtle_applies_reversed_weight_renamings_inside_module_relative_name(tmp_path):
+    turtle = _build_lazy_turtle(
+        tmp_path,
+        {
+            "model.resampler_model.spatial_linear.3.weight": torch.zeros(2),
+        },
+        hf_conversion_map_reversed=[
+            SimpleNamespace(
+                source_patterns=["spatial_linear.ln"],
+                target_patterns=["spatial_linear.3"],
+                operations=[],
+            )
+        ],
+    )
+
+    assert (
+        turtle._resolve_checkpoint_tensor_name(
+            "model.resampler_model",
+            "spatial_linear.ln.weight",
+        )
+        == "model.resampler_model.spatial_linear.3.weight"
+    )
+
+
 def test_lazy_turtle_uses_transformers_checkpoint_conversion_mapping_for_gemma3(tmp_path, monkeypatch):
     conversion_mapping_module = SimpleNamespace(
         get_checkpoint_conversion_mapping=lambda model_type: _gemma3_weight_renamings()