Skip to content

[Qwen3.8-Flash-Next] Avoid memory fragmentation in QSA indexer logits workspace #57105

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
jeejeelee merged 3 commits into from
Sep 27, 2026
+13 −2
Merged

[Qwen3.8-Flash-Next] Avoid memory fragmentation in QSA indexer logits workspace #57105

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
b3b3d8f
avoid memory fragmentation
gau-nernst Sep 16, 2026
b729d83
Merge branch 'main' into qsa-prefill-logits-reserve
gau-nernst Sep 26, 2026
3a0340f
Merge branch 'main' into qsa-prefill-logits-reserve
gau-nernst Sep 26, 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
15 changes: 13 additions & 2 deletions vllm/models/qwen4_exp/nvidia/ops/qsa_indexer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -383,6 +383,7 @@ def _prefill_logits(
page_table: torch.Tensor,
query_start_loc: torch.Tensor,
visible_blocks: torch.Tensor,
logits_workspace: torch.Tensor,
max_query_len: int,
logits_width: int,
query_offset: int,
Expand All @@ -392,10 +393,14 @@ def _prefill_logits(
assert visible_blocks.shape == (q.shape[0],)
assert 0 <= query_offset <= query_offset + num_queries <= q.shape[0]
assert 0 < logits_width <= page_table.shape[1] * k_cache.shape[1]
assert logits_workspace.is_contiguous()
assert logits_workspace.numel() >= num_queries * logits_width

logits = torch.empty(
(num_queries, logits_width), dtype=torch.float32, device=q.device
# slice from pre-allocated workspace
logits = logits_workspace[: num_queries * logits_width].view(
num_queries, logits_width
)

# tuned on GB300
if k_cache.dtype == torch.float8_e4m3fn:
TILE_R, STAGES, num_warps = 32, 2, 8
Expand Down Expand Up @@ -610,8 +615,13 @@ def qsa_select_paged_prefill(
logits_width = min(max(64, logits_width), page_table.shape[1] * k_cache.shape[1])

# chunk the inputs to keep temp logits below VLLM_SPARSE_INDEXER_MAX_LOGITS_MB
# always allocate the worst case to avoid memory fragmentation.
max_logits_bytes = envs.VLLM_SPARSE_INDEXER_MAX_LOGITS_MB * 1024 * 1024
rows_per_chunk = max(1, max_logits_bytes // (logits_width * 4))

budget_bytes = max(max_logits_bytes, page_table.shape[1] * k_cache.shape[1] * 4)
logits_workspace = q.new_empty(budget_bytes // 4, dtype=torch.float32)

topk_workspace = torch.empty(
(_TOPK_WORKSPACE_BYTES,), dtype=torch.uint8, device=q.device
)
Expand All @@ -625,6 +635,7 @@ def qsa_select_paged_prefill(
page_table,
query_start_loc,
visible_blocks,
logits_workspace,
max_query_len,
logits_width,
query_offset=query_start,
Expand Down
Loading