gahow/aituner
0
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Bound graph predictor grid to runtime

Browse Source
This commit is contained in:
Gahow Wang
2026-07-17 23:41:01 +08:00
parent 27a2a3468a
commit e2cd43808d
1 changed files with 7 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
runs/frontier-fidelity-envelope-v1/run_frontier_qwen30_exact_trace_surface.py
View File

@@ -454,6 +454,13 @@ def main() -> None:
config_knobs["no_cache"] = args.fresh_predictor_cache config_knobs["no_cache"] = args.fresh_predictor_cache
if args.align_real_graph_runtime: if args.align_real_graph_runtime:
config_knobs["num_blocks"] = REAL_NUM_BLOCKS_BY_CONFIG[(config.tp, config.mns)] 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: if kernel_paths is not None:
config_knobs.update( config_knobs.update(
{ {