Skip to content

avoid host sync in DSpark prefill slot expansion #40111

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
HaiShaw merged 2 commits into from
Sep 22, 2026
+7 −2
Merged

avoid host sync in DSpark prefill slot expansion #40111

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
9efff5b
avoid host sync in DSpark prefill slot expansion
inkcherry Sep 17, 2026
32508fb
Merge branch 'main' into perf/dspark-prefill-output-size
inkcherry Sep 21, 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
9 changes: 7 additions & 2 deletions python/sglang/srt/speculative/dspark_components/dspark_worker_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -646,11 +646,16 @@ def _forward_prefill(
state_slot = final_pos = None
if is_unified_kv_triton():
repeats = ctx_lens.to(torch.int64)
num_tokens = sum(batch.extend_lens)
state_slot = torch.repeat_interleave(
batch.req_pool_indices.to(device=device, dtype=torch.int64), repeats
batch.req_pool_indices.to(device=device, dtype=torch.int64),
repeats,
output_size=num_tokens,
)
final_pos = torch.repeat_interleave(
(draft_seq_lens + ctx_lens - 1).to(torch.int64), repeats
(draft_seq_lens + ctx_lens - 1).to(torch.int64),
repeats,
output_size=num_tokens,
)
cache_loc = batch.out_cache_loc
token_indices = logits_output.hidden_states_token_indices
Expand Down
Loading