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

Require batched Frontier EP equivalence

Browse Source
This commit is contained in:
Gahow Wang
2026-07-15 22:45:40 +08:00
parent 1fabea26ef
commit 33c91109d7
1 changed files with 12 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
runs/frontier-multicase-sufficiency-v0/best_effort/fixed_cohort_rank.py
View File

@@ -593,15 +593,14 @@ def run_frontier(args: argparse.Namespace) -> Path:
audit_protocol(args.protocol_manifest) audit_protocol(args.protocol_manifest)
profile_paths = frontier_grid.resolve_profile_paths(args.profile_root) profile_paths = frontier_grid.resolve_profile_paths(args.profile_root)
profile_contract = _validate_profile_contract(profile_paths) profile_contract = _validate_profile_contract(profile_paths)
cache_equivalence = json.loads(args.cache_equivalence_artifact.read_text()) ep_equivalence = json.loads(args.ep_equivalence_artifact.read_text())
if ( if (
cache_equivalence.get("schema") ep_equivalence.get("schema") != "frontier-ep-batched-lane-equivalence-v1"
!= "frontier-ep-critical-lane-cache-equivalence-v1" or not ep_equivalence.get("byte_identical")
or not cache_equivalence.get("byte_identical")
): ):
raise ValueError("EP critical-lane cache lacks byte-identical evidence") raise ValueError("batched EP lane prediction lacks byte-identical evidence")
if Path(cache_equivalence["new_source"]).resolve() != args.frontier_source.resolve(): if Path(ep_equivalence["new_source"]).resolve() != args.frontier_source.resolve():
raise ValueError("cache equivalence was measured on a different Frontier source") raise ValueError("EP equivalence was measured on a different Frontier source")
configs = frontier_grid.selected_configs(args.config) configs = frontier_grid.selected_configs(args.config)
args.output_root.mkdir(parents=True, exist_ok=True) args.output_root.mkdir(parents=True, exist_ok=True)
run_manifest = { run_manifest = {
@@ -625,11 +624,11 @@ def run_frontier(args: argparse.Namespace) -> Path:
for name, path in profile_paths.items() for name, path in profile_paths.items()
}, },
}, },
"ep_critical_lane_cache_equivalence": { "ep_prediction_optimization_equivalence": {
"path": str(args.cache_equivalence_artifact.resolve()), "path": str(args.ep_equivalence_artifact.resolve()),
"sha256": sha256(args.cache_equivalence_artifact), "sha256": sha256(args.ep_equivalence_artifact),
"request_metrics_byte_identical": True, "request_metrics_byte_identical": True,
"request_metrics_sha256": cache_equivalence["new_sha256"], "request_metrics_sha256": ep_equivalence["new_sha256"],
}, },
"kv_capacity_evidence": { "kv_capacity_evidence": {
"tp4": { "tp4": {
@@ -649,6 +648,7 @@ def run_frontier(args: argparse.Namespace) -> Path:
"execution_profiles": "community_vllm_0.10.2_serving_entrypoints", "execution_profiles": "community_vllm_0.10.2_serving_entrypoints",
"moe_ep_prefill": "global_routing_then_slowest_local_ep_lane", "moe_ep_prefill": "global_routing_then_slowest_local_ep_lane",
"moe_ep_critical_lane_cache": "exact_(total_routed_tokens,layer_id)_memoization", "moe_ep_critical_lane_cache": "exact_(total_routed_tokens,layer_id)_memoization",
"moe_ep_lane_inference": "eight_independent_rows_in_one_exact_forest_predict_call",
"kv_blocks": "measured_community_vllm_and_fixed_per_topology", "kv_blocks": "measured_community_vllm_and_fixed_per_topology",
"cpu_overhead_modeling": "disabled_no_native_community_vllm_records", "cpu_overhead_modeling": "disabled_no_native_community_vllm_records",
"end_to_end_action_specific_calibration": False, "end_to_end_action_specific_calibration": False,
@@ -1458,7 +1458,7 @@ def parse_args() -> argparse.Namespace:
frontier.add_argument("--output-root", type=Path, required=True) frontier.add_argument("--output-root", type=Path, required=True)
frontier.add_argument("--tp4-capacity-artifact", type=Path, required=True) frontier.add_argument("--tp4-capacity-artifact", type=Path, required=True)
frontier.add_argument("--tp8-capacity-artifact", type=Path, required=True) frontier.add_argument("--tp8-capacity-artifact", type=Path, required=True)
frontier.add_argument("--cache-equivalence-artifact", type=Path, required=True) frontier.add_argument("--ep-equivalence-artifact", type=Path, required=True)
frontier.add_argument("--max-parallel-configs", type=int, default=1) frontier.add_argument("--max-parallel-configs", type=int, default=1)
frontier.add_argument("--config", action="append") frontier.add_argument("--config", action="append")