biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[MTP][Sparse MLA] Take advantage of native MTP support in indexer when possible (#36982)

Browse Source 
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Matthew Bonanni
2026-03-16 13:51:21 -04:00
committed by GitHub
parent 9f9ecff4cd
commit c88ea8338b
2 changed files with 24 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
csrc/sampler.cu
View File

@@ -575,7 +575,7 @@ static __global__ __launch_bounds__(kNumThreadsPerBlock) void topKPerRowDecode(
// The range of logits within the row.
int rowStart = 0;
int seq_len = seqLens[rowIdx / next_n];
int rowEnd = seq_len - next_n + (rowIdx % next_n) + 1;
int rowEnd = max(0, seq_len - next_n + (rowIdx % next_n) + 1);
// Local pointers to this block
if constexpr (!multipleBlocksPerRow && !mergeBlocks) {