ZIBEE/bi_150-vllm
1
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Upgrade to vllm 0.17.0 corex v4.1 overlay

Browse Source
This commit is contained in:
LZiBee
2026-04-29 19:38:22 +08:00
parent 8fac6062e4
commit 938d0854a5
430 changed files with 35969 additions and 14511 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/v1/attention/backends/mla/flashattn_mla.py
View File

@@ -191,6 +191,8 @@ class FlashAttnMLAMetadataBuilder(MLACommonMetadataBuilder[FlashAttnMLAMetadata]
query_start_loc_device: torch.Tensor,
num_decode_tokens: int,
dcp_tot_seq_lens_device: torch.Tensor | None,
max_decode_seq_len: int = 0,
use_cuda_graph: bool = False,
) -> FlashAttnMLADecodeMetadata:
query_lens_cpu = query_start_loc_cpu[1:] - query_start_loc_cpu[:-1]
max_query_len = query_lens_cpu.max().item()
@@ -239,12 +241,14 @@ class FlashAttnMLAMetadataBuilder(MLACommonMetadataBuilder[FlashAttnMLAMetadata]
metadata = FlashAttnMLADecodeMetadata(
block_table=block_table_tensor,
seq_lens=seq_lens_device,
dcp_tot_seq_lens=dcp_tot_seq_lens_device,
max_decode_seq_len=max_decode_seq_len,
use_cuda_graph=use_cuda_graph,
query_start_loc=query_start_loc_device,
max_query_len=max_query_len,
max_seq_len=max_seq_len,
scheduler_metadata=scheduler_metadata,
max_num_splits=max_num_splits,
dcp_tot_seq_lens=dcp_tot_seq_lens_device,
)
return metadata