diff --git a/python/sglang/srt/model_executor/runner/flashinfer_autotune.py b/python/sglang/srt/model_executor/runner/flashinfer_autotune.py index 5218c26f3b..6cd62a617f 100644 --- a/python/sglang/srt/model_executor/runner/flashinfer_autotune.py +++ b/python/sglang/srt/model_executor/runner/flashinfer_autotune.py @@ -175,6 +175,8 @@ def flashinfer_autotune_cache_path(model_runner: ModelRunner) -> Path: str(get_parallel().moe_ep_size), str(mr.model_config.hf_config.__class__.__name__), ] + # Retune caches created before FlashInfer's default tactic joined the race. + model_key_parts.append("candidate_selection=v2_default") # A different skip policy must not reuse previously tuned tactics. skip_ops = get_flashinfer_autotune_skip_ops(mr) model_key_parts.append("skip_ops=" + ",".join(sorted(skip_ops)))