Skip to content

[ROCm][perf] Shuffle KV cache to use paged_attention_common #1

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

Closed
samutamm wants to merge 5 commits into from
Closed

[ROCm][perf] Shuffle KV cache to use paged_attention_common #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
8b97284
change pa_fwd_asm --> paged_attention_common for shuffle kv-cache
samutamm Jan 20, 2026
fa6d691
fix K/V q scale asm args
samutamm Jan 21, 2026
170a052
pin to latest aiter
samutamm Jan 22, 2026
6cf3af5
ruff check
samutamm Jan 22, 2026
ba11577
simplify arguments
samutamm Jan 23, 2026
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
2 changes: 1 addition & 1 deletion docker/Dockerfile.rocm_base
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,7 @@ ARG PYTORCH_AUDIO_BRANCH="v2.9.0"
ARG PYTORCH_AUDIO_REPO="https://github.com/pytorch/audio.git"
ARG FA_BRANCH="0e60e394"
ARG FA_REPO="https://github.com/Dao-AILab/flash-attention.git"
ARG AITER_BRANCH="6af8b687"
ARG AITER_BRANCH="1f5a392"
ARG AITER_REPO="https://github.com/ROCm/aiter.git"
ARG MORI_BRANCH="2d02c6a9"
ARG MORI_REPO="https://github.com/ROCm/mori.git"
Expand Down
43 changes: 39 additions & 4 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 @@ -1143,8 +1143,25 @@ def forward(
return
assert attn_metadata.decode_metadata is not None

_, num_heads, head_size = query.shape
num_seqs = attn_metadata.seq_lens.shape[0]

if rocm_aiter_ops.is_shuffle_kv_cache_enabled():
num_blocks, block_size, num_kv_heads, head_size = key_cache.shape
max_num_partitions = (
attn_metadata.max_seq_len + _PARTITION_SIZE_ROCM - 1
) // _PARTITION_SIZE_ROCM
tmp_out = torch.empty(
(num_seqs, num_heads, max_num_partitions, head_size),
dtype=query.dtype,
device=query.device,
)
exp_sums = torch.empty(
(num_seqs, num_heads, max_num_partitions),
dtype=torch.float32,
device=query.device,
)
max_logits = torch.empty_like(exp_sums)
num_blocks, block_size, num_kv_heads, _ = key_cache.shape
x = 16 // key_cache.element_size()
k_cache_template = torch.empty(
[num_blocks, num_kv_heads, head_size // x, block_size, x],
Expand All @@ -1158,18 +1175,36 @@ def forward(
)
new_key_cache = key_cache.view_as(k_cache_template)
new_value_cache = value_cache.view_as(v_cache_template)
aiter.pa_fwd_asm(
k_qscale_asm = (
layer._k_scale
if attn_metadata.k_scale is None
else attn_metadata.k_scale
)
v_qscale_asm = (
layer._v_scale
if attn_metadata.v_scale is None
else attn_metadata.v_scale
)
aiter.paged_attention_common(
Q=query[:num_decode_tokens],
K=new_key_cache,
V=new_value_cache,
tmp_out=tmp_out,
max_logits=max_logits,
exp_sums=exp_sums,
max_seq_len=attn_metadata.max_seq_len,
block_tables=attn_metadata.block_table[:num_decodes],
context_lens=attn_metadata.seq_lens[:num_decodes],
block_tables_stride0=attn_metadata.block_table[
:num_decodes
].stride(0),
K_QScale=attn_metadata.k_scale,
V_QScale=attn_metadata.v_scale,
scale=self.scale,
K_QScale_hip=layer._k_scale,
V_QScale_hip=layer._v_scale,
K_QScale_asm=k_qscale_asm,
V_QScale_asm=v_qscale_asm,
out_=output[:num_decode_tokens],
kv_cache_dtype=self.kv_cache_dtype,
)
else:
_, num_heads, head_size = query.shape
Expand Down