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

[BugFix] Address PrefillCacheHit state to fix prefix cache accuracy bug (#1498)

Browse Source 
When use AscendScheduler with prefix-cache enabled and chunk-prefill
disabled, there will be accuray problem because there is no branch in
mla_v1 to process this scenario. This PR fixes it.

Signed-off-by: whx-sjtu <2952154980@qq.com>
This commit is contained in:
whx
2025-06-30 16:51:20 +08:00
committed by GitHub
parent 5f8241c25c
commit f286265791
1 changed files with 7 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
vllm_ascend/attention/mla_v1.py
View File

@@ -758,7 +758,8 @@ class AscendMLAImpl(MLAAttentionImpl):
if attn_metadata.attn_state in [
AscendAttentionState.ChunkedPrefill,
AscendAttentionState.SpecDecoding
AscendAttentionState.SpecDecoding,
AscendAttentionState.PrefillCacheHit
] and not ascend_config.chunked_prefill_for_mla:
attn_output_torch = torch.empty(num_tokens,
self.num_heads * self.v_head_dim,
@@ -783,7 +784,8 @@ class AscendMLAImpl(MLAAttentionImpl):
causal=True)
elif attn_metadata.attn_state in [
AscendAttentionState.ChunkedPrefill,
AscendAttentionState.SpecDecoding
AscendAttentionState.SpecDecoding,
AscendAttentionState.PrefillCacheHit
]:
attn_lse = torch.empty(self.num_heads,
num_tokens,
@@ -835,13 +837,14 @@ class AscendMLAImpl(MLAAttentionImpl):
attn_output = attn_output.view(-1, self.num_heads, self.v_head_dim)
else:
raise RuntimeError(
"Unexpected path reached, AscendMLAImpl should only have PrefillNoCache, ChunkedPrefill and SpecDecoding scenario in forward prefill, please file a bug to vllm-ascend !"
"Unexpected path reached, AscendMLAImpl should only have PrefillNoCache, PrefillCacheHit, ChunkedPrefill and SpecDecoding scenario in forward prefill, please file a bug to vllm-ascend !"
)
attn_output = attn_output.reshape(
[num_tokens, self.num_heads * self.v_head_dim])
if attn_metadata.attn_state in [
AscendAttentionState.ChunkedPrefill,
AscendAttentionState.SpecDecoding
AscendAttentionState.SpecDecoding,
AscendAttentionState.PrefillCacheHit
] and not ascend_config.chunked_prefill_for_mla:
attn_output = attn_output_torch