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

[Bugfix] Fix MTP support for lmhead_tensor_parallel_size (#3915)

Browse Source 
### What this PR does / why we need it?
Fix the issue of MTP being enabled and setting
Imhead_tensor_parallel_size=16 causing the inference to hang.

- vLLM version: v0.11.0
- vLLM main:
83f478bb19

Signed-off-by: wyh145 <1987244901@qq.com>
This commit is contained in:
Nagisa125
2025-10-31 10:30:28 +08:00
committed by GitHub
parent 1966885be2
commit 6764777f00
2 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm_ascend/ops/vocab_parallel_embedding.py
View File

@@ -51,7 +51,7 @@ class AscendVocabParallelEmbedding(VocabParallelEmbedding):
prefix: str = ""): prefix: str = ""):
nn.Module.__init__(self) nn.Module.__init__(self)
if lmhead_tp_enable() and prefix.find("lm_head") != -1: if lmhead_tp_enable() and prefix.find("head") != -1:
self.comm_group = get_lmhead_tp_group() self.comm_group = get_lmhead_tp_group()
else: else:
self.comm_group = get_tp_group() self.comm_group = get_tp_group()

3
vllm_ascend/worker/model_runner_v1.py
View File

@@ -2913,7 +2913,8 @@ class NPUModelRunner(LoRAModelRunnerMixin):
aclgraph_runtime_mode=aclgraph_runtime_mode, aclgraph_runtime_mode=aclgraph_runtime_mode,
batch_descriptor=batch_descriptor) batch_descriptor=batch_descriptor)
if need_dummy_logits: if need_dummy_logits:
dummy_compute_logits(hidden_states) self.drafter.model.compute_logits(
hidden_states[dummy_indices])
if self.in_profile_run and self.dynamic_eplb: if self.in_profile_run and self.dynamic_eplb:
self.model.clear_all_moe_loads() self.model.clear_all_moe_loads()
if not self.in_profile_run and self.dynamic_eplb: if not self.in_profile_run and self.dynamic_eplb: