[Bugfix][Model] Fix FP8 k_scale/v_scale not loaded for Qwen3-MoE (#35656)

Signed-off-by: raghavan <oneraghavan@gmail.com>

tests/model_executor/test_weight_utils.py

@@ -11,6 +11,7 @@ from huggingface_hub.utils import LocalEntryNotFoundError
 from vllm.model_executor.model_loader.weight_utils import (
     download_weights_from_hf,
     enable_hf_transfer,
+    maybe_remap_kv_scale_name,
 )
 
 
@@ -61,6 +62,121 @@ def test_download_weights_from_hf():
         )
 
 
+class TestMaybeRemapKvScaleName:
+    """Tests for maybe_remap_kv_scale_name covering all checkpoint formats."""
+
+    PARAMS_DICT = {
+        "model.layers.0.self_attn.attn.k_scale": None,
+        "model.layers.0.self_attn.attn.v_scale": None,
+        "model.layers.0.self_attn.attn.q_scale": None,
+        "model.layers.0.self_attn.qkv_proj.weight": None,
+    }
+
+    def test_qkv_proj_k_scale(self):
+        """Qwen3-MoE / llm-compressor format: qkv_proj.k_scale -> attn.k_scale
+        Regression test for https://github.com/vllm-project/vllm/issues/25047"""
+        result = maybe_remap_kv_scale_name(
+            "model.layers.0.self_attn.qkv_proj.k_scale", self.PARAMS_DICT
+        )
+        assert result == "model.layers.0.self_attn.attn.k_scale"
+
+    def test_qkv_proj_v_scale(self):
+        """Qwen3-MoE / llm-compressor format: qkv_proj.v_scale -> attn.v_scale
+        Regression test for https://github.com/vllm-project/vllm/issues/25047"""
+        result = maybe_remap_kv_scale_name(
+            "model.layers.0.self_attn.qkv_proj.v_scale", self.PARAMS_DICT
+        )
+        assert result == "model.layers.0.self_attn.attn.v_scale"
+
+    def test_modelopt_k_proj_k_scale(self):
+        """ModelOpt format: k_proj.k_scale -> attn.k_scale"""
+        result = maybe_remap_kv_scale_name(
+            "model.layers.0.self_attn.k_proj.k_scale", self.PARAMS_DICT
+        )
+        assert result == "model.layers.0.self_attn.attn.k_scale"
+
+    def test_modelopt_v_proj_v_scale(self):
+        """ModelOpt format: v_proj.v_scale -> attn.v_scale"""
+        result = maybe_remap_kv_scale_name(
+            "model.layers.0.self_attn.v_proj.v_scale", self.PARAMS_DICT
+        )
+        assert result == "model.layers.0.self_attn.attn.v_scale"
+
+    def test_deprecated_kv_scale(self):
+        """Old format: kv_scale -> attn.k_scale (deprecated)"""
+        result = maybe_remap_kv_scale_name(
+            "model.layers.0.self_attn.kv_scale", self.PARAMS_DICT
+        )
+        assert result == "model.layers.0.self_attn.attn.k_scale"
+
+    def test_default_bare_k_scale(self):
+        """Default format: .k_scale -> .attn.k_scale"""
+        result = maybe_remap_kv_scale_name(
+            "model.layers.0.self_attn.k_scale", self.PARAMS_DICT
+        )
+        assert result == "model.layers.0.self_attn.attn.k_scale"
+
+    def test_non_scale_name_unchanged(self):
+        """Non-scale names should be returned unchanged."""
+        name = "model.layers.0.self_attn.qkv_proj.weight"
+        result = maybe_remap_kv_scale_name(name, self.PARAMS_DICT)
+        assert result == name
+
+    def test_nvfp4_modelopt_k_proj_k_scale(self):
+        """ModelOpt NVFP4 format (e.g. nvidia/Qwen3-30B-A3B-NVFP4):
+        k_proj.k_scale -> attn.k_scale.
+        Validates that NVFP4 checkpoints are not broken by this change."""
+        result = maybe_remap_kv_scale_name(
+            "model.layers.0.self_attn.k_proj.k_scale", self.PARAMS_DICT
+        )
+        assert result == "model.layers.0.self_attn.attn.k_scale"
+
+    def test_nvfp4_modelopt_v_proj_v_scale(self):
+        """ModelOpt NVFP4 format (e.g. nvidia/Qwen3-30B-A3B-NVFP4):
+        v_proj.v_scale -> attn.v_scale.
+        Validates that NVFP4 checkpoints are not broken by this change."""
+        result = maybe_remap_kv_scale_name(
+            "model.layers.0.self_attn.v_proj.v_scale", self.PARAMS_DICT
+        )
+        assert result == "model.layers.0.self_attn.attn.v_scale"
+
+    def test_qwen3_vl_moe_qkv_proj_k_scale(self):
+        """Qwen3-VL-MoE uses the same fused qkv_proj naming as Qwen3-MoE.
+        Regression test for qwen3_vl_moe.py fix (same bug as #25047)."""
+        result = maybe_remap_kv_scale_name(
+            "model.layers.0.self_attn.qkv_proj.k_scale", self.PARAMS_DICT
+        )
+        assert result == "model.layers.0.self_attn.attn.k_scale"
+
+    def test_qwen3_vl_moe_qkv_proj_v_scale(self):
+        """Qwen3-VL-MoE uses the same fused qkv_proj naming as Qwen3-MoE.
+        Regression test for qwen3_vl_moe.py fix (same bug as #25047)."""
+        result = maybe_remap_kv_scale_name(
+            "model.layers.0.self_attn.qkv_proj.v_scale", self.PARAMS_DICT
+        )
+        assert result == "model.layers.0.self_attn.attn.v_scale"
+
+    def test_nvfp4_weight_scale_not_remapped(self):
+        """NVFP4 weight_scale should not be touched by remap (not a kv scale)."""
+        name = "model.layers.0.self_attn.k_proj.weight_scale"
+        result = maybe_remap_kv_scale_name(name, self.PARAMS_DICT)
+        assert result == name
+
+    def test_nvfp4_input_scale_not_remapped(self):
+        """NVFP4 input_scale should not be touched by remap (not a kv scale)."""
+        name = "model.layers.0.self_attn.k_proj.input_scale"
+        result = maybe_remap_kv_scale_name(name, self.PARAMS_DICT)
+        assert result == name
+
+    def test_missing_target_returns_none(self):
+        """If remapped name not in params_dict, return None."""
+        empty_params: dict[str, None] = {}
+        result = maybe_remap_kv_scale_name(
+            "model.layers.0.self_attn.qkv_proj.k_scale", empty_params
+        )
+        assert result is None
+
+
 if __name__ == "__main__":
     test_hf_transfer_auto_activation()
     test_download_weights_from_hf()