EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

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

Browse Source 
Signed-off-by: Zhiyu Cheng <zhiyuc@nvidia.com>
This commit is contained in:
Zhiyu
2025-10-06 13:24:15 -07:00
committed by GitHub
parent eb30b888db
commit 155cbb51f0
11 changed files with 464 additions and 42 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/srt/layers/quantization/__init__.py
View File

@@ -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,

2
python/sglang/srt/layers/quantization/modelopt_quant.py
View File

@@ -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]: