[Bugfix][GDN] Reset speculative decode count for an empty draft schedule (#53077)
Signed-off-by: Kevin Luu <51931015+khluu@users.noreply.github.com>
Kevin H. Luu committed
Aug 20, 2026 at 03:33 UTC
6df7adc17f7af8ca3f5b3e6f5ccd48960e95eacb
1 file changed
+1
vllm/v1/attention/backends/gdn_attn.py
+1
@@ -239,6 +239,7 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
239
or num_decode_draft_tokens_cpu[spec_sequence_masks_cpu].sum().item()
240
== 0
241
):
242
+ num_spec_decodes = 0
243
spec_sequence_masks = None
244
spec_sequence_masks_cpu = None
245
else: