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

[BugFix] Fix some issues caused by the ascending order of cudagraph_capture_sizes (#4338)

Browse Source 
### What this PR does / why we need it?
In [#26016](https://github.com/vllm-project/vllm/pull/26016), vllm
change the `cudagraph_capture_sizes` to be in ascending order. This PR
fixes related issues caused by this.
### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?


- vLLM version: v0.11.0
- vLLM main:
2918c1b49c

---------

Signed-off-by: Angazenn <supperccell@163.com>
This commit is contained in:
Angazenn
2025-11-22 17:33:12 +08:00
committed by GitHub
parent fff258bce1
commit 9b3a484b46
3 changed files with 13 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/e2e/multicard/test_aclgraph_capture_replay.py
View File

@@ -109,7 +109,7 @@ def _run_worker_process(
llm = LLM( llm = LLM(
model=model_path, model=model_path,
quantization="ascend" if "W8A8" in model_path else None, quantization="ascend" if "W8A8" in model_path else None,
# enable_expert_parallel=True if "DeepSeek" in model_path else False, enable_expert_parallel=True if "DeepSeek" in model_path else False,
trust_remote_code=True, trust_remote_code=True,
) )

8
vllm_ascend/utils.py
View File

@@ -433,7 +433,13 @@ def _is_default_capture_sizes(vllm_config: VllmConfig) -> bool:
cudagraph_capture_sizes += list( cudagraph_capture_sizes += list(
range(256, max_cudagraph_capture_size + 1, 16)) range(256, max_cudagraph_capture_size + 1, 16))
if sorted(cudagraph_capture_sizes, reverse=True) == \ if vllm_version_is("0.11.0"):
target_cudagraph_capture_sizes = sorted(cudagraph_capture_sizes,
reverse=True)
else:
# in newer version, vVLLM use ascending order of cudagraph_capture_sizes.
target_cudagraph_capture_sizes = sorted(cudagraph_capture_sizes)
if target_cudagraph_capture_sizes == \
vllm_config.compilation_config.cudagraph_capture_sizes: vllm_config.compilation_config.cudagraph_capture_sizes:
return True return True

6
vllm_ascend/worker/model_runner_v1.py
View File

@@ -664,7 +664,11 @@ class NPUModelRunner(LoRAModelRunnerMixin):
# tokens is less than or equal to mc2_tokens_capacity. According to _set_cudagraph_sizes, # tokens is less than or equal to mc2_tokens_capacity. According to _set_cudagraph_sizes,
# the max number of tokens in graph is min(max_num_seqs * uniform_decode_query_len, 512). # the max number of tokens in graph is min(max_num_seqs * uniform_decode_query_len, 512).
if self.compilation_config.cudagraph_capture_sizes: if self.compilation_config.cudagraph_capture_sizes:
max_num_tokens = self.compilation_config.cudagraph_capture_sizes[0] if vllm_version_is("0.11.0"):
max_num_tokens = self.compilation_config.cudagraph_capture_sizes[
0]
else:
max_num_tokens = self.compilation_config.max_cudagraph_capture_size
else: else:
# NOTE: To save memory, we cap the max number of tokens to 512. # NOTE: To save memory, we cap the max number of tokens to 512.
max_num_tokens = min( max_num_tokens = min(