apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 16:48:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

support w4afp8 EP inference (#3382)

Browse Source
This commit is contained in:
Yuan Xiaolan
2025-08-13 21:41:34 +08:00
committed by GitHub
parent 4dbaa3d74c
commit 2513cd929b
17 changed files with 944 additions and 100 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
fastdeploy/model_executor/layers/quantization/w4afp8.py
View File

@@ -20,6 +20,7 @@ import paddle
import fastdeploy
from ..moe import FusedMoE
from .quant_base import QuantConfigBase, QuantMethodBase
QUANT_SCALING_FACTOR = 448
@@ -30,24 +31,32 @@ class W4AFP8Config(QuantConfigBase):
quantization config for weight 4bits and activation fp8
"""
def __init__(self, weight_scale_dict, act_scale_dict) -> None:
def __init__(self, weight_scale_dict, act_scale_dict, is_permuted) -> None:
super().__init__()
self.weight_scale_dict = weight_scale_dict
self.act_scale_dict = act_scale_dict
self.quant_max_bound = 448
self.quant_min_bound = -448
self.quant_round_type = 1
self.is_permuted = is_permuted
def name(self) -> str:
return "w4afp8"
@classmethod
def from_config(cls, config: dict) -> "W4AFP8Config":
weight_scale_dict = config["weight_scale_dict"]
act_scale_dict = config["act_scale_dict"]
return cls(weight_scale_dict, act_scale_dict)
weight_scale_dict = config.get("weight_scale_dict", None)
act_scale_dict = config.get("act_scale_dict", None)
is_permuted = config.get("is_permuted", True)
return cls(weight_scale_dict, act_scale_dict, is_permuted)
def get_quant_method(self, layer) -> Optional[QuantMethodBase]:
if isinstance(layer, FusedMoE):
from fastdeploy.model_executor.layers.moe.fused_moe_cutlass_backend import (
CutlassW4AFP8MoEMethod,
)
return CutlassW4AFP8MoEMethod(self)
return W4AFP8LinearMethod(self)