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

bugfix for mtp in fullgraph (#3878)

Browse Source 
### What this PR does / why we need it?
bugfix for mtp in fullgraph

### Does this PR introduce _any_ user-facing change?
no

---------

Signed-off-by: zouyida2052 <zouyida2002@gmail.com>
This commit is contained in:
zouyida2052
2025-10-29 23:52:20 +08:00
committed by GitHub
parent 19f49ecb5f
commit d9249c968e
6 changed files with 58 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
vllm_ascend/worker/model_runner_v1.py
View File

@@ -3529,14 +3529,8 @@ class NPUModelRunner(LoRAModelRunnerMixin):
if aclgraph_mode.decode_mode() == CUDAGraphMode.FULL and \
aclgraph_mode.separate_routine():
max_num_tokens = self.scheduler_config.max_num_seqs * \
self.uniform_decode_query_len
decode_cudagraph_batch_sizes = [
x for x in self.aclgraph_batch_sizes if x <= max_num_tokens
and x >= self.uniform_decode_query_len
]
compilation_cases_decode = list(
reversed(decode_cudagraph_batch_sizes))
reversed(self.aclgraph_batch_sizes))
self._capture_aclgraphs(
compilation_cases=compilation_cases_decode,
aclgraph_runtime_mode=CUDAGraphMode.FULL,