Skip to content

[Model Runner V2] Rebuild attn metadata between draft decode steps #41162

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
WoosukKwon merged 2 commits into from
May 5, 2026
+198 −121
Merged

[Model Runner V2] Rebuild attn metadata between draft decode steps #41162

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4827469
[Model Runner V2] Rebuild attn metadata between draft decode steps
TheEpicDolphin Apr 28, 2026
f6c68b5
Merge branch 'main' into mrv2-draft-decode-rebuild-attn-metadata
mergify[bot] May 4, 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
22 changes: 18 additions & 4 deletions vllm/v1/worker/gpu/sample/gumbel.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,6 +76,8 @@ def gumbel_block_argmax(
pos_ptr,
processed_logits_ptr,
processed_logits_stride,
processed_logits_col_ptr,
vocab_size,
APPLY_TEMPERATURE: tl.constexpr,
):
req_state_idx = tl.load(expanded_idx_mapping_ptr + token_idx)
Expand All @@ -88,8 +90,15 @@ def gumbel_block_argmax(

if processed_logits_ptr is not None:
# Store the temperature-applied logits.
if processed_logits_col_ptr is not None:
col = tl.load(processed_logits_col_ptr)
else:
col = 0
tl.store(
processed_logits_ptr + req_state_idx * processed_logits_stride + block,
processed_logits_ptr
+ req_state_idx * processed_logits_stride
+ col * vocab_size
+ block,
logits,
mask=mask,
)
Expand Down Expand Up @@ -121,6 +130,7 @@ def _gumbel_sample_kernel(
local_max_stride,
processed_logits_ptr,
processed_logits_stride,
processed_logits_col_ptr,
logits_ptr,
logits_stride,
expanded_idx_mapping_ptr,
Expand Down Expand Up @@ -153,6 +163,8 @@ def _gumbel_sample_kernel(
pos_ptr,
processed_logits_ptr,
processed_logits_stride,
processed_logits_col_ptr,
vocab_size,
APPLY_TEMPERATURE=APPLY_TEMPERATURE,
)
token_id = block_idx * BLOCK_SIZE + idx
Expand All @@ -167,7 +179,8 @@ def gumbel_sample(
seed: torch.Tensor, # [max_num_reqs]
pos: torch.Tensor, # [num_tokens]
apply_temperature: bool,
processed_logits_out: torch.Tensor | None = None, # [num_reqs, vocab_size]
output_processed_logits: torch.Tensor | None = None,
output_processed_logits_col: torch.Tensor | None = None,
) -> torch.Tensor:
num_tokens, vocab_size = logits.shape
BLOCK_SIZE = 1024
Expand All @@ -179,8 +192,9 @@ def gumbel_sample(
local_argmax.stride(0),
local_max,
local_max.stride(0),
processed_logits_out,
processed_logits_out.stride(0) if processed_logits_out is not None else 0,
output_processed_logits,
output_processed_logits.stride(0) if output_processed_logits is not None else 0,
output_processed_logits_col,
logits,
logits.stride(0),
expanded_idx_mapping,
Expand Down
Loading
Loading