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

[bugfix] Fixed the error issue when overlaying MTP and full decode on DSV3.1 C8. (#7571)

Browse Source 
…DSV3.1 C8.

### What this PR does / why we need it?
DeepSeek v3.1 C8 had a hanging issue when overlaying MTP and full graph
modes; this pull request resolves that issue.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?

- vLLM version: v0.18.0
- vLLM main:
ed359c497a

---------

Signed-off-by: pichangping <1337510399@qq.com>
This commit is contained in:
pichangping
2026-03-25 14:36:26 +08:00
committed by GitHub
parent 95d33f05c2
commit 6ddfc41312
1 changed files with 6 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
vllm_ascend/attention/mla_v1.py
View File

@@ -1318,6 +1318,8 @@ class AscendMLAImpl(MLAAttentionImpl):
sparse_mode = 3 sparse_mode = 3
attn_mask = attn_metadata.decode.attn_mask # type:ignore attn_mask = attn_metadata.decode.attn_mask # type:ignore
actual_seq_lengths = decode_meta.actual_seq_lengths_q actual_seq_lengths = decode_meta.actual_seq_lengths_q
if self.fa_quant_layer:
dequant_scale_q_nope = dequant_scale_q_nope.view(num_tokens, self.num_heads)
elif self.fa_quant_layer: elif self.fa_quant_layer:
attn_mask = None attn_mask = None
input_layout = "BSND_NBSD" input_layout = "BSND_NBSD"
@@ -1403,7 +1405,10 @@ class AscendMLAImpl(MLAAttentionImpl):
weak_ref_tensors(softmax_lse), weak_ref_tensors(softmax_lse),
) )
if self.fa_quant_layer: if self.fa_quant_layer:
attn_params = attn_params + (dequant_scale_q_nope, self.fak_descale_float) # type: ignore attn_params = attn_params + (
weak_ref_tensors(dequant_scale_q_nope),
weak_ref_tensors(self.fak_descale_float),
) # type: ignore
else: else:
attn_params = attn_params + (None, None) # type: ignore attn_params = attn_params + (None, None) # type: ignore