Skip to content

fix qwen torchair attention PrefillCacheHit #2787

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 6 commits into from
Sep 11, 2025
+6 −1
Merged

fix qwen torchair attention PrefillCacheHit #2787

Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
cd9c2f8
fix qwen torchair attention PrefillCacheHit
Sep 5, 2025
b472d8f
Merge branch 'vllm-project:main' into main
zhaozx-cn Sep 8, 2025
99cfc07
Merge branch 'vllm-project:main' into main
zhaozx-cn Sep 9, 2025
5e7563f
Merge branch 'vllm-project:main' into main
zhaozx-cn Sep 10, 2025
047c7da
Merge branch 'vllm-project:main' into main
zhaozx-cn Sep 10, 2025
eed886a
Merge branch 'vllm-project:main' into main
zhaozx-cn Sep 10, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 6 additions & 1 deletion vllm_ascend/torchair/torchair_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -374,6 +374,9 @@ def forward(
indices = torch.cat((block_indices, slots_indices), dim=1)
torch_npu.npu_scatter_nd_update_(key_cache, indices, key)
torch_npu.npu_scatter_nd_update_(value_cache, indices, value)
if attn_metadata.attn_state == AscendAttentionState.PrefillCacheHit:
self.key_cache = key_cache
self.value_cache = value_cache
Comment on lines +377 to +379
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

The self.key_cache and self.value_cache are used by the PrefillCacheHit attention state. This change correctly updates them when the state is PrefillCacheHit. However, the cache update in lines 375-376 happens for any state that has a kv_cache (e.g., DecodeOnly as well). To make the logic more robust and prevent potential issues if other states start using self.key_cache in the future, it would be better to update self.key_cache and self.value_cache unconditionally whenever the cache is modified. This ensures that self.key_cache and self.value_cache always reflect the latest state of the cache tensors passed into this forward pass.

            self.key_cache = key_cache
            self.value_cache = value_cache


if attn_metadata.attn_state == AscendAttentionState.PrefillNoCache:
assert attn_metadata is not None
Expand Down Expand Up @@ -411,11 +414,13 @@ def forward(
assert attn_metadata is not None
assert attn_metadata.attn_mask is not None
compress_mask = attn_metadata.attn_mask
batch_size = attn_metadata.query_lens.shape[0]
block_table = attn_metadata.block_tables[:batch_size, :]
torch_npu._npu_flash_attention_qlens(
query=query,
key_cache=self.key_cache,
value_cache=self.value_cache,
block_table=attn_metadata.block_tables,
block_table=block_table,
mask=compress_mask,
seq_len=attn_metadata.query_lens,
context_lens=attn_metadata.seq_lens,
Expand Down
Loading