[Model Runner V2] Use unpadded num_tokens for PW CUDA graph attn metadata (#36626)

Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>
This commit is contained in:
Woosuk Kwon
2026-03-10 09:30:56 -07:00
committed by GitHub
parent f83b933b84
commit f088a831dd
4 changed files with 14 additions and 3 deletions
+1
View File
@@ -936,6 +936,7 @@ class GPUModelRunner(LoRAModelRunnerMixin):
assert block_tables is not None
attn_metadata = self.model_state.prepare_attn(
input_batch,
batch_desc.cg_mode,
block_tables,
slot_mappings,
self.attn_groups,