apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 08:37:06 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[MetaxGPU] Support FastDeploy on metax gpu (#3241)

Browse Source 
* [MetaxGPU] Support FastDeploy on metax gpu

* Update metax_worker.py

1. change worker log;
2. remove custom allreduce, adapt it later;
3. remove cuda graph;

* Update __init__.py

1. remove metax's key work comment

* Update __init__.py

1. remove metax's key word comment;
2. add fused_moe_kernel_paddle import

---------

Co-authored-by: yongqiangma <xing.wo@163.com>
This commit is contained in:
Kane2011
2025-08-13 11:11:54 +08:00
committed by GitHub
parent ed6bff215a
commit b4fef2cf29
29 changed files with 3224 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
fastdeploy/model_executor/layers/quantization/weight_only.py
View File

@@ -94,6 +94,16 @@ class WeightOnlyConfig(QuantConfigBase):
)
return DCUWeightOnlyLinearMethod(self)
elif current_platform.is_maca():
if isinstance(layer, FusedMoE):
from fastdeploy.model_executor.layers.backends import (
MetaxTritonWeightOnlyMoEMethod,
)
return MetaxTritonWeightOnlyMoEMethod(self)
else:
return GPUWeightOnlyLinearMethod(self)
else:
if isinstance(layer, FusedMoE):
if layer.use_method == "cutlass":
@@ -196,14 +206,24 @@ class WeightOnlyLinearMethod(QuantMethodBase):
raise NotImplementedError
def apply(self, layer, x):
linear_out = weight_only_linear(
x,
weight=layer.weight,
bias=layer.bias if layer.add_bias else None,
weight_scale=layer.weight_scale,
weight_dtype=("int8" if self.quant_config.name() == "wint8" else "int4"),
arch=self.quant_config.weight_only_linear_arch,
)
if current_platform.is_maca():
linear_out = weight_only_linear(
x,
weight=layer.weight,
bias=layer.bias if layer.add_bias else None,
weight_scale=layer.weight_scale,
weight_dtype=("int8" if self.quant_config.name() == "wint8" else "int4"),
arch=80,
)
else:
linear_out = weight_only_linear(
x,
weight=layer.weight,
bias=layer.bias if layer.add_bias else None,
weight_scale=layer.weight_scale,
weight_dtype=("int8" if self.quant_config.name() == "wint8" else "int4"),
arch=self.quant_config.weight_only_linear_arch,
)
return linear_out
@@ -240,6 +260,7 @@ class GPUWeightOnlyLinearMethod(WeightOnlyLinearMethod):
algo=self.quant_config.algo,
arch=self.quant_config.weight_only_linear_arch,
)
if current_platform.is_maca():
quanted_weight_tensor = paddle.transpose(quanted_weight_tensor, [1, 0])
layer.weight.set_value(quanted_weight_tensor)
layer.weight_scale.set_value(weight_scale_tensor.astype(paddle.get_default_dtype()))