[MRV2][Spec] Fuse AR speculator multi-step decodes back into one CUDA graph (#46849)

Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
This commit is contained in:
Yizhou
2026-08-11 15:06:22 +08:00
committed by GitHub
parent 1044d19998
commit a311916a29
10 changed files with 492 additions and 49 deletions
+4
View File
@@ -372,6 +372,10 @@ class DeepseekV4ROCMAiterMLASparseMetadataBuilder(DeepseekV4FlashMLAMetadataBuil
class DeepseekV4ROCMAiterSparseSWAMetadataBuilder(DeepseekSparseSWAMetadataBuilder):
# Keep fused multi-step decode disabled until update_draft_decode_metadata()
# also refreshes the ROCm-specific ragged SWA indices and indptrs.
supports_draft_decode_metadata_update = False
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
max_tokens = self.vllm_config.scheduler_config.max_num_batched_tokens