From 6df7adc17f7af8ca3f5b3e6f5ccd48960e95eacb Mon Sep 17 00:00:00 2001 From: "Kevin H. Luu" Date: Thu, 20 Aug 2026 03:33:41 -0700 Subject: [PATCH] [Bugfix][GDN] Reset speculative decode count for an empty draft schedule (#53077) Signed-off-by: Kevin Luu <51931015+khluu@users.noreply.github.com> --- vllm/v1/attention/backends/gdn_attn.py | 1 + 1 file changed, 1 insertion(+) diff --git a/vllm/v1/attention/backends/gdn_attn.py b/vllm/v1/attention/backends/gdn_attn.py index 0e843e31ba..27df94d7bd 100644 --- a/vllm/v1/attention/backends/gdn_attn.py +++ b/vllm/v1/attention/backends/gdn_attn.py @@ -239,6 +239,7 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata] or num_decode_draft_tokens_cpu[spec_sequence_masks_cpu].sum().item() == 0 ): + num_spec_decodes = 0 spec_sequence_masks = None spec_sequence_masks_cpu = None else: