apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-11-01 12:22:53 +08:00
Code Issues Actions 6 Packages Projects Releases Wiki Activity

【BugFix】fix ep buffer clear (#4450)

Browse Source 
* fix

* fix
This commit is contained in:
gaoziyuan
2025-10-21 10:56:00 +08:00
committed by GitHub
parent 70a29ec49e
commit d85ef5352a
1 changed files with 4 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
fastdeploy/model_executor/layers/moe/fused_moe_backend_base.py
View File

@@ -175,12 +175,13 @@ class MoEMethodBase(QuantMethodBase):
Paddle Cutlass compute Fused MoE.
"""
if layer.ep_size > 1:
if layer.fd_config.model_config.moe_phase.phase == "prefill" and layer.layer_idx == 0:
if layer.fd_config.scheduler_config.splitwise_role == "mixed":
is_moe_start_layer = layer.layer_idx == layer.fd_config.model_config.moe_layer_start_index
if layer.fd_config.model_config.moe_phase.phase == "prefill":
if layer.fd_config.scheduler_config.splitwise_role == "mixed" and is_moe_start_layer:
self.ep_prefill_runner.clean_low_latency_buffer()
return self.apply_ep_prefill(layer, x, gate)
else:
if layer.fd_config.scheduler_config.splitwise_role == "mixed" and layer.layer_idx == 0:
if layer.fd_config.scheduler_config.splitwise_role == "mixed" and is_moe_start_layer:
self.ep_decoder_runner.clean_low_latency_buffer()
return self.apply_ep_decode(layer, x, gate)
else: