[Fix] Use cached prefix lengths for FlashInfer full-attention ragged prefill (#40796)

This commit is contained in:
Jimmy Shong
2026-09-23 17:23:25 -10:00
committed by GitHub
parent e2f4fedf04
commit be495a6174
2 changed files with 62 additions and 0 deletions
@@ -1992,6 +1992,14 @@ class FlashInferIndicesUpdaterPrefill:
else:
paged_kernel_lens_sum = paged_kernel_lens.sum().item()
kv_start_idx = seq_lens - paged_kernel_lens
elif use_ragged:
# Extend K/V lands after the paged pass: plan over the prefix only.
paged_kernel_lens = prefix_lens
if extend_prefix_lens_cpu is not None:
paged_kernel_lens_sum = sum(extend_prefix_lens_cpu)
else:
paged_kernel_lens_sum = prefix_lens.sum().item()
kv_start_idx = None
else:
# full attention
paged_kernel_lens = seq_lens