Skip to content

Use paged_attention_v1 for sliding window decode in rocm_aiter_fa #34378

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
zhuohan123 merged 1 commit into from
Feb 13, 2026
Merged

Use paged_attention_v1 for sliding window decode in rocm_aiter_fa #34378

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
31 changes: 2 additions & 29 deletions vllm/v1/attention/backends/rocm_aiter_fa.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1075,35 +1075,6 @@ def forward(
assert not rocm_aiter_ops.is_shuffle_kv_cache_enabled(), (
"Sliding window with shuffle layout is not supported yet."
)
from aiter.ops.triton.unified_attention import (
unified_attention,
)

descale_shape = (
attn_metadata.query_start_loc[:num_decodes].shape[0] - 1,
key_cache.shape[2],
)
unified_attention(
q=query[:num_decode_tokens],
k=key_cache,
v=value_cache,
out=output[:num_decode_tokens],
cu_seqlens_q=attn_metadata.query_start_loc[:num_decodes],
max_seqlen_q=1, # optimize this
seqused_k=attn_metadata.seq_lens[:num_decodes],
max_seqlen_k=attn_metadata.max_seq_len,
softmax_scale=self.scale,
causal=True,
alibi_slopes=self.alibi_slopes,
window_size=self.sliding_window,
block_table=attn_metadata.block_table[:num_decodes],
softcap=self.logits_soft_cap,
q_descale=None,
k_descale=layer._k_scale.expand(descale_shape),
v_descale=layer._v_scale.expand(descale_shape),
)
return
assert attn_metadata.decode_metadata is not None

if rocm_aiter_ops.is_shuffle_kv_cache_enabled():
num_blocks, block_size, num_kv_heads, head_size = key_cache.shape
Expand Down Expand Up @@ -1172,6 +1143,8 @@ def forward(
layer._v_scale,
None,
_PARTITION_SIZE_ROCM,
1,
self.sliding_window[0] + 1,
)
else:
raise NotImplementedError(
Expand Down