[Perf][GDN] Skip FLA chunk metadata when FlashInfer GDN prefill is active (#61569)

Signed-off-by: Thien Tran <gau.nernst@yahoo.com.sg> Co-authored-by: Kimi Code <noreply@moonshot.cn>
2026-10-14 01:46:54 -07:00
parent d36158f919
commit 937d6c19df
1 changed files with 4 additions and 1 deletions
@@ -546,7 +546,10 @@ class GDNAttentionMetadataBuilder(AttentionMetadataBuilder[GDNAttentionMetadata]
                        cu_seqlens=prefill_query_start_loc,
                    )
                )
-            else:
+            elif self.gdn_prefill_backend != "flashinfer":
+                # FlashInfer's chunk kernel consumes cu_seqlens directly and
+                # never reads chunk_indices/chunk_offsets, so skip the FLA
+                # metadata (CPU compute + H2D copies) entirely.
                chunk_indices, chunk_offsets = self._build_chunk_metadata(
                    prefill_query_start_loc,
                    prefill_query_start_loc_cpu,