it needs to be even more contiguous

2026-05-05 18:02:26 +00:00
parent 7a60cd9965
commit 9a2f2dc570
1 changed files with 2 additions and 1 deletions
--- a/indexer.py
+++ b/indexer.py
@@ -608,7 +608,8 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
            # (csrc/apis/attention.hpp). Unsqueeze to (B, 1) so downstream
            # kernels see the same (B, next_n) layout as the MTP path.
            if seq_lens.dim() == 1:
-                seq_lens = seq_lens.unsqueeze(-1).contiguous()
+                seq_lens = seq_lens.unsqueeze(-1)
+            seq_lens = seq_lens.contiguous()

            # DeepGEMM is required for the paged MQA logits on CUDA devices
            if current_platform.is_cuda() and has_deep_gemm():