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

[BugFix]Fix group list type of mc2. (#3890)

Browse Source 
### What this PR does / why we need it?
Fix the precision issue caused by the inconsistency between the group
list type used by mc2 and that of eplb.

---------

Signed-off-by: offline0806 <3337230449@qq.com>
This commit is contained in:
offline893
2025-10-30 21:44:14 +08:00
committed by GitHub
parent c506ba60fb
commit d5a9aba03f
4 changed files with 18 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm_ascend/ops/common_fused_moe.py
View File

@@ -244,7 +244,8 @@ class AscendFusedMoE(FusedMoE):
self.expert_map != -1) if self.expert_map is not None else
self.global_num_experts)
if self.dynamic_eplb:
self.moe_load = torch.zeros(local_num_experts, dtype=torch.int64)
self.moe_load = torch.zeros(local_num_experts,
dtype=torch.int64).npu()
self.moe_config.num_experts = self.global_num_experts
self.moe_config.num_local_experts = self.local_num_experts
@@ -340,9 +341,9 @@ class AscendFusedMoE(FusedMoE):
if isinstance(final_hidden_states, tuple):
final_hidden_states, group_list_type, expert_tokens = final_hidden_states
if self.dynamic_eplb:
self.moe_load += expert_tokens if group_list_type else \
torch.cat([expert_tokens[:1], expert_tokens[1:] - expert_tokens[:-1]])
if self.dynamic_eplb:
self.moe_load += expert_tokens if group_list_type == 1 else \
torch.cat([expert_tokens[:1], expert_tokens[1:] - expert_tokens[:-1]])
final_hidden_states = forward_context.moe_comm_method.finalize(
hidden_states=final_hidden_states,