apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-12-24 13:28:13 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[Quantization] Support w4afp8 MoE dynamic quantization (#5282)

Browse Source 
* support dynamic activation quant for w4afp8

* support dynamic w4afp8

* add test

* fix

* fix

---------

Co-authored-by: zhoutianzi666 <17801055074@163.com>
This commit is contained in:
Sunny-bot1
2025-12-02 18:56:16 +08:00
committed by GitHub
parent 429dd2b1db
commit 3629db4129
9 changed files with 478 additions and 37 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
fastdeploy/model_executor/layers/quantization/__init__.py
View File

@@ -84,6 +84,13 @@ def parse_quant_config(args, model_config, is_ernie, is_v1_loader):
quantization_config["moe_quant_type"] = "wint4"
quantization_config["quantization"] = "mix_quant"
quant_config_name = "mix_quant"
# Special handling for moe w4afp8 dynamic quant
elif quant_config_name == "w4afp8":
quantization_config["dense_quant_type"] = "block_wise_fp8"
quantization_config["moe_quant_type"] = "w4afp8"
quantization_config["hadamard_block_size"] = 512
quantization_config["quantization"] = "mix_quant"
quant_config_name = "mix_quant"
else:
quant_config_name = None
if quant_config_name is None:

6
fastdeploy/model_executor/layers/quantization/w4afp8.py
View File

@@ -31,7 +31,7 @@ class W4AFP8Config(QuantConfigBase):
quantization config for weight 4bits and activation fp8
"""
def __init__(self, weight_scale_dict, act_scale_dict, is_permuted, hadamard_block_size) -> None:
def __init__(self, weight_scale_dict, act_scale_dict, is_permuted, hadamard_block_size, is_quantized) -> None:
super().__init__()
self.weight_scale_dict = weight_scale_dict
self.act_scale_dict = act_scale_dict
@@ -40,6 +40,7 @@ class W4AFP8Config(QuantConfigBase):
self.quant_round_type = 1
self.is_permuted = is_permuted
self.hadamard_block_size = hadamard_block_size
self.is_quantized = is_quantized
def name(self) -> str:
return "w4afp8"
@@ -50,7 +51,8 @@ class W4AFP8Config(QuantConfigBase):
act_scale_dict = config.get("act_scale_dict", None)
is_permuted = config.get("is_permuted", True)
hadamard_block_size = config.get("hadamard_block_size", 128)
return cls(weight_scale_dict, act_scale_dict, is_permuted, hadamard_block_size)
is_quantized = config.get("is_quantized", False)
return cls(weight_scale_dict, act_scale_dict, is_permuted, hadamard_block_size, is_quantized)
def get_quant_method(self, layer) -> Optional[QuantMethodBase]:
if isinstance(layer, FusedMoE):