From e2cd43808d5b4638191520b32d69aff87e53cf50 Mon Sep 17 00:00:00 2001 From: Gahow Wang Date: Fri, 17 Jul 2026 23:41:01 +0800 Subject: [PATCH] Bound graph predictor grid to runtime --- .../run_frontier_qwen30_exact_trace_surface.py | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/runs/frontier-fidelity-envelope-v1/run_frontier_qwen30_exact_trace_surface.py b/runs/frontier-fidelity-envelope-v1/run_frontier_qwen30_exact_trace_surface.py index 454d0b2..cac0e74 100644 --- a/runs/frontier-fidelity-envelope-v1/run_frontier_qwen30_exact_trace_surface.py +++ b/runs/frontier-fidelity-envelope-v1/run_frontier_qwen30_exact_trace_surface.py @@ -454,6 +454,13 @@ def main() -> None: config_knobs["no_cache"] = args.fresh_predictor_cache if args.align_real_graph_runtime: config_knobs["num_blocks"] = REAL_NUM_BLOCKS_BY_CONFIG[(config.tp, config.mns)] + # These are predictor-grid limits, not scheduler knobs. Keep the + # grid inside the observed vLLM graph/runtime envelope: no batch + # can use a capture bucket above this cell's list and MBT is 8192. + config_knobs["prediction_max_batch_size"] = max( + GRAPH_CAPTURE_SIZES_BY_MNS[config.mns] + ) + config_knobs["prediction_max_prefill_chunk_size"] = 8192 if kernel_paths is not None: config_knobs.update( {