Autotune flashinfer extend buckets at warmup (#32556)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
2026-08-07 14:06:05 -07:00
parent 1480687cff
commit 62a28197c0
16 changed files with 174 additions and 9 deletions
@@ -224,7 +224,10 @@ class TestHybridNeedsCpuSeqLens(CustomTestCase):
        from sglang.srt.layers.attention.hybrid_attn_backend import HybridAttnBackend

        def backend(flag):
-            return SimpleNamespace(needs_cpu_seq_lens=flag)
+            return SimpleNamespace(
+                needs_cpu_seq_lens=flag,
+                extend_dummy_seqs_capped_by_req_pool=False,
+            )

        runner = SimpleNamespace(
            server_args=SimpleNamespace(speculative_attention_mode=spec_mode),