[Feature][NPU]: add runtime support for AutoRound quantized models (#16699)

Co-authored-by: root <root@localhost.localdomain>
Co-authored-by: ronnie_zheng <zl19940307@163.com>

python/sglang/srt/layers/linear.py

@@ -64,10 +64,9 @@ WEIGHT_LOADER_V2_SUPPORTED = [
     "GPTQMarlin24LinearMethod",
     "TPUInt8LinearMethod",
     "GPTQLinearMethod",
-    "GPTQLinearAscendMethod",
-    "GPTQMoEAscendMethod",
     "FBGEMMFp8LinearMethod",
     "GPTQLinearAscendMethod",
+    "GPTQMoEAscendMethod",
     "ModelOptFp8LinearMethod",
     "ModelOptFp4LinearMethod",
     "IPEXAWQLinearMethod",

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

@@ -14,6 +14,9 @@ from sglang.srt.layers.quantization.utils import get_scalar_types
 ScalarType, scalar_types = get_scalar_types()
 
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
+from sglang.srt.utils import is_npu
+
+_is_npu = is_npu()
 
 
 class AutoRoundConfig(QuantizationConfig):
@@ -301,6 +304,11 @@ class AutoRoundConfig(QuantizationConfig):
     def apply_gptq_quant_layer(self, layer, prefix: str, backend: str = "auto"):
         from sglang.srt.layers.linear import LinearBase
         from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
+        from sglang.srt.layers.quantization.gptq import (
+            GPTQConfig,
+            GPTQLinearAscendMethod,
+            GPTQMoEAscendMethod,
+        )
         from sglang.srt.layers.quantization.marlin_utils import (
             check_marlin_supported,
             check_moe_marlin_supports_layer,
@@ -323,6 +331,24 @@ class AutoRoundConfig(QuantizationConfig):
             group_size,
             sym,
         )
+        if _is_npu:
+            quant_args = GPTQConfig(
+                weight_bits=weight_bits,
+                group_size=group_size,
+                lm_head_quantized=False,
+                desc_act=False,
+                dynamic={},
+            )
+            quant_args.sym = sym
+
+            if isinstance(layer, FusedMoE):
+                return GPTQMoEAscendMethod(quant_args)
+
+            if isinstance(layer, (LinearBase, ParallelLMHead)):
+                return GPTQLinearAscendMethod(quant_args)
+
+            return None
+
         if backend == "auto" or "marlin" in backend:
             GPTQ_TYPE_MAP = {
                 (4, True): scalar_types.uint4b8,