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

[Bugfix]v0.18.0 support FlashComm1 & DCP for Qwen (#7726)

Browse Source 
### What this PR does / why we need it?
This PR backports the changes from #7673 ([Bugfix] support FlashComm1 &
DCP for Qwen) to the releases/v0.18.0 branch.

--------
Signed-off-by: Yang Yuxi <907276627@qq.com>
This commit is contained in:
Yang Yuxi
2026-03-29 15:59:19 +08:00
committed by GitHub
parent 9cc41c9457
commit e776d5c0f1
2 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
tests/e2e/multicard/4-cards/long_sequence/test_basic.py
View File

@@ -236,7 +236,8 @@ def test_dcp_basic():
decode_context_parallel_size=2, decode_context_parallel_size=2,
max_num_batched_tokens=1024, max_num_batched_tokens=1024,
enable_expert_parallel=True, enable_expert_parallel=True,
block_size=128) as runner: block_size=128,
compilation_config={"pass_config": {"enable_sp": True}}) as runner:
runner.model.generate(prompts, sampling_params) runner.model.generate(prompts, sampling_params)

2
vllm_ascend/worker/model_runner_v1.py
View File

@@ -1284,7 +1284,7 @@ class NPUModelRunner(GPUModelRunner):
if ( if (
cudagraph_mode == CUDAGraphMode.FULL cudagraph_mode == CUDAGraphMode.FULL
or (enable_sp() and not self.model_config.use_mla) or (enable_sp() and not self.model_config.use_mla)
and self.pcp_size == 1 # TODO(lxs): fix this and self.pcp_size * self.dcp_size == 1
): ):
# Currently, Graph Mode and SP will both pad num_tokens, # Currently, Graph Mode and SP will both pad num_tokens,
# Another possible condition is num_tokens_padded != num_tokens_unpadded # Another possible condition is num_tokens_padded != num_tokens_unpadded