[bugfix] Fixed the error issue when overlaying MTP and full decode on DSV3.1 C8. (#7571)
…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: https://github.com/vllm-project/vllm/commit/ed359c497a728f08b5b41456c07a688ccd510fbc --------- Signed-off-by: pichangping <1337510399@qq.com>
This commit is contained in:
1 parent
95d33f05c2
commit
6ddfc41312
1 file changed
+6
-1
@@ -1318,6 +1318,8 @@ class AscendMLAImpl(MLAAttentionImpl):
|
||||
sparse_mode = 3
|
||||
attn_mask = attn_metadata.decode.attn_mask # type:ignore
|
||||
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:
|
||||
attn_mask = None
|
||||
input_layout = "BSND_NBSD"
|
||||
@@ -1403,7 +1405,10 @@ class AscendMLAImpl(MLAAttentionImpl):
|
||||
weak_ref_tensors(softmax_lse),
|
||||
)
|
||||
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:
|
||||
attn_params = attn_params + (None, None) # type: ignore
|
||||
|
||||
|
||||
Reference in new issue
Block a user