EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Feature] Support moe multi-stream for aclgraph. (#2946)

Browse Source 
This PR puts the calculation of shared experts into a separate stream,
overlaping with routing experts.

- vLLM version: v0.10.2
- vLLM main:
fbd6523ac0

---------

Signed-off-by: whx-sjtu <2952154980@qq.com>
This commit is contained in:
whx
2025-09-19 11:06:45 +08:00
committed by GitHub
parent 0c04bf1e36
commit 0a526768f5
14 changed files with 170 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm_ascend/ascend_config.py
View File

@@ -61,6 +61,8 @@ class AscendConfig:
self.enable_shared_expert_dp = additional_config.get(
"enable_shared_expert_dp", False
) and not self.torchair_graph_config.enabled and vllm_config.parallel_config.enable_expert_parallel
self.multistream_overlap_shared_expert = additional_config.get(
"multistream_overlap_shared_expert", False)
self.enable_prefetch = additional_config.get("enable_prefetch", False)
self.lmhead_tensor_parallel_size = additional_config.get(
"lmhead_tensor_parallel_size", None)
@@ -110,8 +112,6 @@ class TorchairGraphConfig:
"graph_batch_sizes_init", False)
self.enable_multistream_mla = torchair_graph_config.get(
"enable_multistream_mla", False)
self.enable_multistream_moe = torchair_graph_config.get(
"enable_multistream_moe", False)
self.enable_view_optimize = torchair_graph_config.get(
"enable_view_optimize", True)
self.enable_frozen_parameter = torchair_graph_config.get(
@@ -148,10 +148,6 @@ class TorchairGraphConfig:
raise RuntimeError(
"enable_multistream_mla is valid only when Torchair graph mode is enabled"
)
if self.enable_multistream_moe:
raise RuntimeError(
"enable_multistream_moe is valid only when Torchair graph mode is enabled"
)
if self.enable_kv_nz:
raise RuntimeError(
"enable_kv_nz is valid only when Torchair graph mode is enabled"