From 5acd95906b7c7a54dde89396d2bb06fe28ebeed0 Mon Sep 17 00:00:00 2001 From: "Kevin H. Luu" Date: Wed, 9 Sep 2026 01:47:45 -0700 Subject: [PATCH] [Bugfix] Exclude DP token padding from draft attention metadata (#55458) Signed-off-by: khluu Signed-off-by: Kevin H. Luu Co-authored-by: Codex Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com> --- vllm/v1/worker/gpu/spec_decode/autoregressive/speculator.py | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-) diff --git a/vllm/v1/worker/gpu/spec_decode/autoregressive/speculator.py b/vllm/v1/worker/gpu/spec_decode/autoregressive/speculator.py index aba53e53c3..bba4e2b4bd 100644 --- a/vllm/v1/worker/gpu/spec_decode/autoregressive/speculator.py +++ b/vllm/v1/worker/gpu/spec_decode/autoregressive/speculator.py @@ -511,7 +511,8 @@ class AutoRegressiveSpeculator(DraftModelSpeculator): attn_metadata = self._build_draft_attn_metadata( num_reqs=num_reqs, num_reqs_padded=batch_desc.num_reqs or num_reqs, - num_tokens_padded=batch_desc.num_tokens, + # One query per request; exclude DP-only model padding. + num_tokens_padded=batch_desc.num_reqs or num_reqs, seq_lens_cpu_upper_bound=seq_lens_cpu_upper_bound, step=step, ) @@ -559,7 +560,8 @@ class AutoRegressiveSpeculator(DraftModelSpeculator): attn_metadata = self._build_draft_attn_metadata( num_reqs=num_reqs, num_reqs_padded=batch_desc.num_reqs or num_reqs, - num_tokens_padded=batch_desc.num_tokens, + # One query per request; exclude DP-only model padding. + num_tokens_padded=batch_desc.num_reqs or num_reqs, seq_lens_cpu_upper_bound=seq_lens_cpu_upper_bound, step=1, )