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

[v0.18.0][Bugfix] fix ds3.2 dcp mtp (#7681)

Browse Source 
### What this PR does / why we need it?
Fixed the issue where the DCP overlaps the MTP scenario in the ds3.2
scenario.

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

### How was this patch tested?

cherry-pick from: https://github.com/vllm-project/vllm-ascend/pull/7617

Signed-off-by: weiguihua2 <weiguihua2@huawei.com>
This commit is contained in:
weiguihua2
2026-03-27 14:24:53 +08:00
committed by GitHub
parent 048c8d1afe
commit bc8e87f3db
10 changed files with 18 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
vllm_ascend/attention/context_parallel/attention_cp.py
View File

@@ -232,6 +232,7 @@ class AscendAttentionCPMetadataBuilder(AscendAttentionMetadataBuilder):
block_tables=block_table,
query_start_loc=query_start_loc,
seq_lens=seq_lens,
seq_lens_cpu=seq_lens,
seq_lens_list=seq_lens.tolist(),
max_query_len=common_attn_metadata.max_query_len,
actual_seq_lengths_q=query_start_loc_cpu[1:].tolist(),

4
vllm_ascend/attention/context_parallel/sfa_cp.py
View File

@@ -257,8 +257,8 @@ class AscendSFACPImpl(AscendSFAImpl):
return self._align_to_graph_bucket_tokens(attn_output, attn_metadata)
def _align_to_graph_bucket_tokens(self, attn_output: torch.Tensor | None, attn_metadata: M) -> torch.Tensor | None:
if attn_output is None:
return None
if attn_output is None or self.pcp_size == 1:
return attn_output
# In graph/piecewise mode, output buffer uses graph bucket token size
# (forward_context.num_tokens), while PCP path may compute only valid
# tokens. Align to the larger one to avoid later write-back mismatch.