Skip to content

[Bugfix][Backport][Hardware][AMD] Backport PR #31282 to v0.13.0: Fix last_page_len calculation in AITER MLA decode #38146

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
khairulkabir1661 wants to merge 2 commits into from
+12 −9
Closed

[Bugfix][Backport][Hardware][AMD] Backport PR #31282 to v0.13.0: Fix last_page_len calculation in AITER MLA decode #38146

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
c893400
Use pre-initialized buffer for paged_kv_last_page_len
c0de128 Dec 25, 2025
4a89366
Optimize: reuse paged_kv_last_page_len buffer slice
c0de128 Dec 26, 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
21 changes: 12 additions & 9 deletions vllm/v1/attention/backends/mla/rocm_aiter_mla.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,16 +86,20 @@ def __init__(
# TODO: we can disambiguate between decode and mixed-prefill decode here
# so we can only use the persistent buffer if a cudagraph is actually
# being used.

# paged_kv_last_page_len is always 1s (kernel block size is always 1),
# so we create it once and reuse slices in both eager and cudagraph modes.
self.paged_kv_last_page_len = torch.ones(
max_num_reqs, dtype=torch.int32, device=device
)

if self.compilation_config.cudagraph_mode.has_full_cudagraphs():
self.paged_kv_indptr = torch.zeros(
max_num_reqs + 1, dtype=torch.int32, device=device
)
self.paged_kv_indices = torch.zeros(
max_num_pages, dtype=torch.int32, device=device
)
self.paged_kv_last_page_len = torch.zeros(
max_num_reqs, dtype=torch.int32, device=device
)

self.qo_indptr = torch.arange(
0, max_num_reqs + 1, dtype=torch.int32, device=device
Expand All @@ -120,7 +124,9 @@ def _build_decode(
).unsqueeze(0) < seq_lens_device.unsqueeze(1)
paged_kv_indices = block_table_tensor[mask]

paged_kv_last_page_len = torch.where(seq_lens_device == 0, 1, seq_lens_device)
# kernel block size is always 1, so each page has exactly 1 token.
# last_page_len is always 1 - just slice the pre-initialized buffer.
paged_kv_last_page_len = self.paged_kv_last_page_len[:num_reqs]

paged_kv_indptr = torch.cat(
[
Expand All @@ -144,11 +150,8 @@ def _build_decode(
self.paged_kv_indptr[1 + num_reqs :].fill_(paged_kv_indptr[-1])
paged_kv_indptr = self.paged_kv_indptr[: 1 + num_reqs]

self.paged_kv_last_page_len[:num_reqs].copy_(
paged_kv_last_page_len, non_blocking=True
)
self.paged_kv_last_page_len[num_reqs:].fill_(1)
paged_kv_last_page_len = self.paged_kv_last_page_len[:num_reqs]
# paged_kv_last_page_len already uses the pre-initialized buffer slice
# (set above), so no copy needed - buffer is always 1s.

qo_indptr = self.qo_indptr[: 1 + num_reqs]

Expand Down
Loading