Enable native ModelOpt quantization support (1/3) (#7149)

Signed-off-by: Zhiyu Cheng <zhiyuc@nvidia.com>

python/sglang/srt/layers/modelopt_utils.py

@@ -0,0 +1,11 @@
+"""
+ModelOpt related constants
+"""
+
+QUANT_CFG_CHOICES = {
+    "fp8": "FP8_DEFAULT_CFG",
+    "int4_awq": "INT4_AWQ_CFG",  # TODO: add support for int4_awq
+    "w4a8_awq": "W4A8_AWQ_BETA_CFG",  # TODO: add support for w4a8_awq
+    "nvfp4": "NVFP4_DEFAULT_CFG",
+    "nvfp4_awq": "NVFP4_AWQ_LITE_CFG",  # TODO: add support for nvfp4_awq
+}

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

@@ -72,7 +72,7 @@ if TYPE_CHECKING:
 BASE_QUANTIZATION_METHODS: Dict[str, Type[QuantizationConfig]] = {
     "fp8": Fp8Config,
     "blockwise_int8": BlockInt8Config,
-    "modelopt": ModelOptFp8Config,
+    "modelopt_fp8": ModelOptFp8Config,
     "modelopt_fp4": ModelOptFp4Config,
     "w8a8_int8": W8A8Int8Config,
     "w8a8_fp8": W8A8Fp8Config,

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

@@ -113,7 +113,7 @@ class ModelOptFp8Config(QuantizationConfig):
 
     @classmethod
     def get_name(cls) -> str:
-        return "modelopt"
+        return "modelopt_fp8"
 
     @classmethod
     def get_supported_act_dtypes(cls) -> List[torch.dtype]: