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/rocm_aiter_mla.py
View File

@@ -115,6 +115,8 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
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,
) -> AiterMLADecodeMetadata:
# kernel block size is always 1, although the kv block size is not 1.
device = self.device
@@ -170,11 +172,13 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
attn_metadata = AiterMLADecodeMetadata(
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,
paged_kv_indptr=paged_kv_indptr,
paged_kv_indices=paged_kv_indices,
paged_kv_last_page_len=paged_kv_last_page_len,
qo_indptr=qo_indptr,
dcp_tot_seq_lens=dcp_tot_seq_lens_device,
max_qo_len=max_qo_len,
attn_out_dtype=self.decode_attn_out_dtype,
)