apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 00:17:25 +08:00
Code Issues Actions 19 Packages Projects Releases Wiki Activity

[Intel HPU] fix memory fragmentation issue due to warmup process and fix moe all_reduce issue (#5357)

Browse Source
This commit is contained in:
fmiao2372
2025-12-04 11:29:41 +08:00
committed by GitHub
parent 946025480e
commit 209006e6a6
3 changed files with 14 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/model_executor/layers/moe/moe.py
+8 -2
View File
@@ -21,7 +21,10 @@ from paddle import nn
from paddleformers.utils.log import logger
from fastdeploy import envs
from fastdeploy.distributed.communication import tensor_model_parallel_all_reduce
from fastdeploy.distributed.communication import (
tensor_model_parallel_all_reduce,
tensor_model_parallel_all_reduce_custom,
)
from fastdeploy.model_executor.layers.utils import get_tensor
from fastdeploy.model_executor.utils import h2d_copy, slice_fn
from fastdeploy.platforms import current_platform
@@ -643,7 +646,10 @@ class FusedMoE(nn.Layer):
out = self.forward_normal(x, gate)
if self.reduce_results and self.tp_size > 1:
out = tensor_model_parallel_all_reduce(out, self.tp_group)
if current_platform.is_intel_hpu():
tensor_model_parallel_all_reduce_custom(out)
else:
out = tensor_model_parallel_all_reduce(out, self.tp_group)
return out
def forward_chunked_moe(self, x: paddle.Tensor, gate: nn.Layer):