chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[4/n]decouple quantization implementation from vLLM dependency (#9191)

Browse Source 
Co-authored-by: AniZpZ <aniz1905@gmail.com>
Co-authored-by: Yineng Zhang <me@zhyncs.com>
This commit is contained in:
Hongbo Xu
2025-08-15 03:05:46 +08:00
committed by GitHub
parent 63d82a776a
commit 2cc9eeab01
8 changed files with 37 additions and 74 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
python/sglang/srt/layers/quantization/__init__.py
View File

@@ -55,13 +55,7 @@ if is_mxfp_supported:
from sglang.srt.layers.quantization.fp4 import MxFp4Config
from sglang.srt.layers.quantization.fp8 import Fp8Config
from sglang.srt.layers.quantization.gptq import (
GPTQConfig,
GPTQLinearMethod,
GPTQMarlinConfig,
GPTQMarlinLinearMethod,
GPTQMarlinMoEMethod,
)
from sglang.srt.layers.quantization.gptq import GPTQConfig, GPTQMarlinConfig
from sglang.srt.layers.quantization.modelopt_quant import (
ModelOptFp4Config,
ModelOptFp8Config,
@@ -70,7 +64,6 @@ from sglang.srt.layers.quantization.moe_wna16 import MoeWNA16Config
from sglang.srt.layers.quantization.mxfp4 import Mxfp4Config
from sglang.srt.layers.quantization.petit import PetitNvFp4Config
from sglang.srt.layers.quantization.qoq import QoQConfig
from sglang.srt.layers.quantization.utils import get_linear_quant_method
from sglang.srt.layers.quantization.w4afp8 import W4AFp8Config
from sglang.srt.layers.quantization.w8a8_fp8 import W8A8Fp8Config
from sglang.srt.layers.quantization.w8a8_int8 import W8A8Int8Config
@@ -86,6 +79,10 @@ BASE_QUANTIZATION_METHODS: Dict[str, Type[QuantizationConfig]] = {
"modelopt_fp4": ModelOptFp4Config,
"w8a8_int8": W8A8Int8Config,
"w8a8_fp8": W8A8Fp8Config,
"awq": AWQConfig,
"awq_marlin": AWQMarlinConfig,
"gptq": GPTQConfig,
"gptq_marlin": GPTQMarlinConfig,
"moe_wna16": MoeWNA16Config,
"compressed-tensors": CompressedTensorsConfig,
"qoq": QoQConfig,
@@ -111,19 +108,15 @@ elif is_mxfp_supported and is_hip():
# VLLM-dependent quantization methods
VLLM_QUANTIZATION_METHODS = {
"aqlm": AQLMConfig,
"awq": AWQConfig,
"deepspeedfp": DeepSpeedFPConfig,
"tpu_int8": Int8TpuConfig,
"fbgemm_fp8": FBGEMMFp8Config,
"marlin": MarlinConfig,
"gguf": GGUFConfig,
"gptq_marlin_24": GPTQMarlin24Config,
"awq_marlin": AWQMarlinConfig,
"bitsandbytes": BitsAndBytesConfig,
"qqq": QQQConfig,
"experts_int8": ExpertsInt8Config,
"gptq_marlin": GPTQMarlinConfig,
"gptq": GPTQConfig,
}
QUANTIZATION_METHODS = {**BASE_QUANTIZATION_METHODS, **VLLM_QUANTIZATION_METHODS}
@@ -145,23 +138,6 @@ def get_quantization_config(quantization: str) -> Type[QuantizationConfig]:
return QUANTIZATION_METHODS[quantization]
def gptq_get_quant_method(self, layer, prefix):
from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
if isinstance(layer, FusedMoE):
return GPTQMarlinMoEMethod(self)
if isinstance(self, GPTQConfig):
return get_linear_quant_method(
self, layer, prefix=prefix, linear_method_cls=GPTQLinearMethod
)
elif isinstance(self, GPTQMarlinConfig):
return get_linear_quant_method(
self, layer, prefix=prefix, linear_method_cls=GPTQMarlinLinearMethod
)
return None
original_isinstance = builtins.isinstance
@@ -239,10 +215,7 @@ def monkey_patch_moe_apply(class_obj: "FusedMoEMethodBase"):
def monkey_patch_quant_configs():
"""Apply all monkey patches in one place."""
setattr(GPTQMarlinConfig, "get_quant_method", gptq_get_quant_method)
setattr(GPTQConfig, "get_quant_method", gptq_get_quant_method)
monkey_patch_moe_apply(GPTQMarlinMoEMethod)
monkey_patch_moe_apply(CompressedTensorsW8A8Fp8MoEMethod)
monkey_patch_moe_apply(CompressedTensorsWNA16MoEMethod)