fix: copy seq_lens in TRTLLM MHA draft decode cuda graph capture (#26521)
This commit is contained in:
@@ -324,6 +324,9 @@ class TRTLLMHAAttnBackend(FlashInferAttnBackend):
|
||||
metadata.cache_seqlens_int32 = self.decode_cuda_graph_metadata[
|
||||
"cache_seqlens"
|
||||
][:bs]
|
||||
metadata.cache_seqlens_int32.copy_(
|
||||
seq_lens + self.speculative_step_id + 1
|
||||
)
|
||||
metadata.max_seq_len_k = seq_lens.max().item() + (
|
||||
self.speculative_step_id + 1
|
||||
)
|
||||
|
||||
Reference in New Issue
Block a user