[Fix] Use cached prefix lengths for FlashInfer full-attention ragged prefill (#40796)
This commit is contained in:
@@ -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
|
||||
|
||||
Reference in New Issue
Block a user