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

[bugfix] fix accuracy prolem for deepseek V3/R1 models with torchair graph in long sequence predictions (#1331)

Browse Source 
### What this PR does / why we need it?
Fix the issue of insufficient cached cosine and sine length in MLA's
TorchAir graph mode, which causes accuracy deviation during
long-sequence inference.

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

### How was this patch tested?
We tested the accuracy of this patch with DeepSeek R1 e2e becnhmark
serving, and get 83.33 sore for AIME2024 dataset with DP4TP4EP16
setting.

Signed-off-by: linfeng-yuan <1102311262@qq.com>
This commit is contained in:
linfeng-yuan
2025-06-23 09:52:27 +08:00
committed by GitHub
parent f04c6763d8
commit 15592c0d48
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/attention/mla_v1.py
View File

@@ -1077,7 +1077,7 @@ class AscendMLAImpl(MLAAttentionImpl):
decode_k_nope = None
assert attn_metadata.decode is not None
if self.running_in_graph:
seq_len = self.rotary_emb.max_position_embeddings
seq_len = self.rotary_emb.max_position_embeddings * self.rotary_emb.scaling_factor
cos = self.rotary_emb.cos_cached[:seq_len].to(
dtype=decode_hs_or_q_c.dtype)
sin = self.rotary_emb.sin_cached[:seq_len].to(
@@ -1122,7 +1122,7 @@ class AscendMLAImpl(MLAAttentionImpl):
prefill_q_nope = prefill_q[..., :self.qk_nope_head_dim]
if self.torchair_graph_enabled:
num_tokens = prefill_hs_or_q_c.shape[0]
seq_len = self.rotary_emb.max_position_embeddings
seq_len = self.rotary_emb.max_position_embeddings * self.rotary_emb.scaling_factor
cos = self.rotary_emb.cos_cached[:seq_len].to(
dtype=prefill_q_pe.dtype)
sin = self.rotary_emb.sin_cached[:seq_len].to(