EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions 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

2
examples/external_online_dp/run_dp_template.sh
View File

@@ -43,4 +43,4 @@ vllm serve model_path \
"kv_connector_module_path": "vllm_ascend.distributed.llmdatadist_c_mgr_connector"
}' \
--additional-config \
'{"ascend_scheduler_config": {"enabled": true}, "torchair_graph_config":{"enabled":true,"enable_kv_nz":false, "enable_multistream_moe":false, "graph_batch_size":[28]}, "enable_weight_nz_layout":true}'
'{"ascend_scheduler_config": {"enabled": true}, "torchair_graph_config":{"enabled":true,"enable_kv_nz":false, "graph_batch_size":[28]}, "enable_weight_nz_layout":true, "enable_multistream_moe":false}'

2
examples/run_dp_server.sh
View File

@@ -29,4 +29,4 @@ vllm serve Qwen/Qwen1.5-MoE-A2.7B \
--gpu-memory-utilization 0.9 \
--trust-remote-code \
--enforce-eager \
--additional-config '{"ascend_scheduler_config":{"enabled":true},"torchair_graph_config":{"enabled":false, "enable_multistream_moe":false, "use_cached_graph":false}}'
--additional-config '{"ascend_scheduler_config":{"enabled":true},"torchair_graph_config":{"enabled":false, "use_cached_graph":false}}'