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

[v0.11.0] [Bugfix] [MoE]fix error in deepseek when using allgather (#3827)

Browse Source 
### What this PR does / why we need it?
After refactoring vllm_ascend/models and FusedMoE, we are unable to pass
`gate` from deepseekv2.py to `AscendFusedMoE.forward`, which will result
in error when running deepseek v3/r1 with allgather.
Hence, this pr removes `gate` related computations from FusedMoE module
in eager/aclgraph mode.
### Does this PR introduce _any_ user-facing change?
`rm_router_logits` is deprecated in eager/aclgraph.
### How was this patch tested?
e2e & ut

Signed-off-by: Pr0Wh1teGivee <calvin_zhu0210@outlook.com>
This commit is contained in:
weichen
2025-10-30 14:59:46 +08:00
committed by GitHub
parent 211d4b9da4
commit c506ba60fb
7 changed files with 98 additions and 115 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
vllm_ascend/utils.py
View File

@@ -520,36 +520,6 @@ class ProfileExecuteDuration:
return durations
# TODO(ttanzhiqiang): rm_router_logits
# dp>1 will trigger
# In theory, this solution is only applicable to AllGather and AllGatherEP, because in the dp scenario, the previous operation was gate + two communications, and now it is changed to one communication + gate operation, which can save some communication time. In theory, all moe AllGather and AllGatherEP solutions can follow this logic, but now other moe models (qwen3-235b) dp solutions are not adjusted, so use the switch to control it to prevent code errors.
def get_rm_router_logits_state(ep_size: int, dp_size: int,
is_deepseek_v3_r1: bool):
# the fusion operator torch_npu.npu_grouped_matmul_finalize_routing called by allgather ep
# only supports deepseek v3/r1
if dp_size > 1:
if (envs_ascend.VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP and ep_size > 1
and is_deepseek_v3_r1):
return True
elif ep_size == 1 and is_deepseek_v3_r1:
return True
return False
# TODO(ttanzhiqiang): all_reduce merge
# When all_reduce_merge is in progress, shared_experts does not do all_reduce in mlp, but waits until shared_experts+router_experts are completed before doing all_reduce
# Currently, all_reduce_merge is enabled by default in the AllGather, AllGatherEP and NaiveMulticast scenarios of the deepseek model.
def get_all_reduce_merge_state(ep_size: int, is_deepseek_v3_r1: bool):
# the fusion operator torch_npu.npu_grouped_matmul_finalize_routing called by allgather ep
# only supports deepseek v3/r1
if (envs_ascend.VLLM_ENABLE_FUSED_EXPERTS_ALLGATHER_EP and ep_size > 1
and is_deepseek_v3_r1):
return True
elif ep_size == 1 and is_deepseek_v3_r1:
return True
return False
def register_ascend_customop(vllm_config: Optional[VllmConfig] = None):
"""Register Ascend CustomOP