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

[Patch]patch of v1 executor when enable eplb. (#3511)

Browse Source 
### What this PR does / why we need it?
when using dynamic eplb, patch v1 executor to avoid create child process
failed.

### How was this patch tested?
deepseek in v3.

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

Signed-off-by: offline0806 <3337230449@qq.com>
Co-authored-by: offline0806 <3337230449@qq.com>
This commit is contained in:
offline893
2025-10-19 10:54:26 +08:00
committed by GitHub
parent 646c1db5d7
commit 6c9909c861
5 changed files with 192 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
vllm_ascend/torchair/ops/torchair_fused_moe.py
View File

@@ -1089,7 +1089,8 @@ class TorchairAscendFusedMoE(FusedMoE):
local_num_experts = (torch.sum(self.expert_map != -1)
if self.expert_map is not None else 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.torchair_graph_enabled = ascend_config.torchair_graph_config.enabled
self.multistream_overlap_shared_expert = \
@@ -1311,17 +1312,26 @@ class TorchairAscendFusedMoE(FusedMoE):
tuple) and len(e_hidden_states) == 2:
e_hidden_states, shared_hidden_states = e_hidden_states
if self.dynamic_eplb and isinstance(
if isinstance(e_hidden_states, tuple) and len(e_hidden_states) == 4:
e_hidden_states, shared_hidden_states, group_list_type, expert_tokens = e_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 shared_experts is None and isinstance(
e_hidden_states, tuple) and len(e_hidden_states) == 3:
e_hidden_states, group_list_type, expert_tokens = e_hidden_states
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 else \
torch.cat([expert_tokens[:1], expert_tokens[1:] - expert_tokens[:-1]])
if (fused_moe_state not in [
FusedMoEState.AllGather, FusedMoEState.AllGatherEP,
FusedMoEState.NaiveMulticast
] and not replace_allreduce and not self.enable_shared_expert_dp):
if tp_size > 1:
if isinstance(e_hidden_states, tuple):
e_hidden_states = e_hidden_states[0]
dist.all_gather(list(chunk_hidden_states), e_hidden_states,
self.tp_group)
final_hidden_states = torch.cat(chunk_hidden_states, dim=0)

9
vllm_ascend/torchair/quantization/torchair_w8a8_dynamic.py
View File

@@ -365,17 +365,18 @@ def torchair_fused_experts_with_mc2(
) if enable_dispatch_v2 else torch_npu.npu_moe_distribute_combine(
**kwargs_mc2)
if dynamic_eplb:
return (hidden_states, 1, expert_token_nums)
if shared_experts is None:
if dynamic_eplb:
return (hidden_states, 1, expert_token_nums)
return hidden_states
else:
with npu_stream_switch("moe_secondary", 0):
npu_wait_tensor(shared_act, down_out_list)
shared_output, _ = shared_experts.down_proj(
(shared_act, swiglu_out_scale))
return hidden_states, shared_output
if dynamic_eplb:
return (hidden_states, shared_output, 1, expert_token_nums)
return (hidden_states, shared_output)
def torchair_init_routing_quant(hidden_states, top_k, topk_ids,