[Bugfix] [NPU] Qwen3.5 with quantization fix (#21692)

python/sglang/srt/layers/quantization/modelslim/modelslim.py

@@ -13,7 +13,6 @@ from sglang.srt.layers.quantization.base_config import (
     FusedMoEMethodBase,
     QuantizationConfig,
 )
-from sglang.srt.layers.quantization.compressed_tensors.utils import should_ignore_layer
 from sglang.srt.layers.quantization.modelslim.schemes import (
     ModelSlimW4A4Int4,
     ModelSlimW4A4Int4MoE,
@@ -142,12 +141,7 @@ class ModelSlimConfig(QuantizationConfig):
         from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 
         if isinstance(layer, LinearBase):
-            if should_ignore_layer(
-                prefix,
-                ignore=self.ignore,
-                fused_mapping=self.packed_modules_mapping,
-            ):
-                return UnquantizedLinearMethod()
+            # TODO: we should remove this code and switch to the packed_modules_mapping declared inside the modeling files
             key = "model"
             if "vision_model" in prefix:
                 key = "vision_model"
@@ -163,57 +157,49 @@ class ModelSlimConfig(QuantizationConfig):
                 prefix_in_quant_config = prefix.replace(
                     proj_name, packed_modules_mapping_subset[proj_name][0]
                 )
-
-            if self.is_layer_skipped(prefix, packed_modules_mapping_subset):
+            if self.is_layer_skipped(
+                prefix, packed_modules_mapping_subset
+            ) or self.is_layer_skipped(prefix, self.packed_modules_mapping):
                 return UnquantizedLinearMethod()
-            scheme = self.get_linear_scheme(
-                layer=layer, layer_name=prefix_in_quant_config
-            )
-            layer.scheme = scheme
+            layer.scheme = self.get_linear_scheme(layer, prefix_in_quant_config)
             return ModelSlimLinearMethod(self)
         elif isinstance(layer, FusedMoE):
             layer.scheme = self.get_moe_scheme(layer, prefix)
             return ModelSlimFusedMoEMethod(self)
         return None
 
-    def _get_scheme_from_parts(
-        self,
-        layer_name: str,
-    ) -> ModelSlimLinearScheme:
-
-        quant_type = self.quant_description.get(layer_name + ".weight", "")
-        if quant_type == "W8A8_DYNAMIC" or quant_type == "W8A8":
-            return ModelSlimW8A8Int8(
-                quant_config=self.quant_description, prefix=layer_name
-            )
-        elif quant_type == "W4A4_DYNAMIC":
-            return ModelSlimW4A4Int4(
-                quant_config=self.quant_description, prefix=layer_name
-            )
-        raise NotImplementedError("No modelslim compatible scheme was found.")
-
     def get_linear_scheme(
-        self, layer: torch.nn.Module, layer_name: Optional[str] = None
+        self, layer: torch.nn.Module, prefix: Optional[str] = None
     ) -> Optional[ModelSlimLinearScheme]:
         """
         get_scheme method adjusted for modelslim, taken from
         python/sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py
         """
-        scheme = self._get_scheme_from_parts(
-            layer_name=layer_name,
-        )
 
-        # Ascend doesn't support device capability
-        logger.debug("Using scheme: %s for %s", scheme.__class__.__name__, layer_name)
-        return scheme
+        linear_quant_schemes = [
+            ("W4A4_DYNAMIC", ModelSlimW4A4Int4),
+            ("W8A8", ModelSlimW8A8Int8),
+            ("W8A8_DYNAMIC", ModelSlimW8A8Int8),
+        ]
+
+        quant_schemes = [self.quant_description.get(prefix + ".weight", "")]
+
+        for scheme_name, scheme_class in linear_quant_schemes:
+            if any(s == scheme_name for s in quant_schemes):
+                logger.info_once(f"Using {scheme_class.__name__}")
+                return scheme_class(quant_config=self.quant_description, prefix=prefix)
+
+        logger.warning(
+            f"Unsupported Linear modelslim scheme: "
+            f"{quant_schemes} in layer: {prefix}"
+        )
+        return None
 
     def get_moe_scheme(
         self,
         layer: torch.nn.Module,
         prefix: str,
     ) -> Optional[ModelSlimMoEScheme]:
-        # TODO: @dsikka: refactor this to use schemes as other kernels
-        # are supported + check if the layer is being ignored.
         moe_quant_schemes = [
             ("W4A4_DYNAMIC", ModelSlimW4A4Int4MoE),
             ("W4A8_DYNAMIC", ModelSlimW4A8Int8MoE),
@@ -222,7 +208,7 @@ class ModelSlimConfig(QuantizationConfig):
 
         moe_weight_suffixes = [".0.gate_proj.weight", ".0.w2.weight"]
         quant_schemes = [
-            self.quant_description.get(prefix + suffix, "STATIC")
+            self.quant_description.get(prefix + suffix, "")
             for suffix in moe_weight_suffixes
         ]
 

python/sglang/srt/model_loader/loader.py

@@ -197,6 +197,7 @@ def _get_quantization_config(
     model_class, _ = get_model_architecture(model_config)
     packed_modules_mapping = getattr(model_class, "packed_modules_mapping", {})
     remap_prefix = getattr(model_class, "remap_prefix", None)
+    # TODO: we should remove this code and switch to the packed_modules_mapping declared inside the modeling files
     if _is_npu:
         packed_modules_mapping.update(
             {

python/sglang/srt/models/qwen3_5.py

@@ -881,7 +881,7 @@ ALL_DECODER_LAYER_TYPES = {
 class Qwen3_5ForCausalLM(nn.Module):
     """Qwen3.5 Model with support for dense variant."""
 
-    if _is_gfx95:
+    if _is_gfx95 or _is_npu:
         packed_modules_mapping = {
             "qkv_proj": ["q_proj", "k_proj", "v_proj"],
             "gate_up_proj": ["gate_proj", "up_proj"],
@@ -1310,7 +1310,7 @@ class Qwen3_5MoeForCausalLM(Qwen3_5ForCausalLM):
 
 
 class Qwen3_5ForConditionalGeneration(Qwen3VLForConditionalGeneration):
-    if _is_gfx95:
+    if _is_gfx95 or _is_npu:
         packed_modules_mapping = Qwen3_5ForCausalLM.packed_modules_mapping
         hf_to_sglang_mapper = None
 
@@ -1447,7 +1447,7 @@ class Qwen3_5ForConditionalGeneration(Qwen3VLForConditionalGeneration):
 class Qwen3_5MoeForConditionalGeneration(Qwen3VLForConditionalGeneration):
     """Qwen3.5 MoE Vision-Language Model."""
 
-    if _is_gfx95:
+    if _is_gfx95 or _is_npu:
         packed_modules_mapping = Qwen3_5ForCausalLM.packed_modules_mapping
         hf_to_sglang_mapper = None