Freeze vLLM 0.20 profiles and capture trace routing

runs/frontier-qwen30-vllm020-profile-v1/prepare_profile_comparison.py

@@ -0,0 +1,104 @@
+#!/usr/bin/env python3
+"""Prepare old/new profile-only Frontier manifests from the frozen P1 probes."""
+
+from __future__ import annotations
+
+import argparse
+import hashlib
+import json
+from pathlib import Path
+from typing import Any
+
+
+PROFILE_KEYS = {
+    "linear_op_input_file": "linear_op.csv",
+    "atten_input_file": "attention.csv",
+    "moe_input_file": "moe.csv",
+}
+
+
+def sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(1 << 20), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def write_json(path: Path, payload: Any) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n")
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--source", type=Path, required=True)
+    parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument("--mode", choices=("old-profile-only", "new-profile-only"), required=True)
+    parser.add_argument("--profile-root", type=Path)
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    source = json.loads(args.source.read_text())
+    if source.get("status") != "PASS" or len(source.get("entries", [])) != 12:
+        raise SystemExit("source manifest must contain 12 passing P1 probes")
+    if args.mode == "new-profile-only" and args.profile_root is None:
+        raise SystemExit("--profile-root is required for new-profile-only")
+
+    output = args.output.resolve()
+    config_root = output / "configs"
+    cache_root = output / "prediction-cache"
+    entries: list[dict[str, Any]] = []
+    profile_hashes: dict[str, str] = {}
+    if args.profile_root is not None:
+        profile_root = args.profile_root.resolve()
+        for filename in PROFILE_KEYS.values():
+            path = profile_root / filename
+            if not path.is_file():
+                raise SystemExit(f"missing frozen profile: {path}")
+            profile_hashes[str(path)] = sha256(path)
+
+    for entry in source["entries"]:
+        config_path = Path(entry["config"])
+        config = json.loads(config_path.read_text())
+        config["mode"] = args.mode
+        config["config_id"] = f"{config['cell_id']}__{args.mode}"
+        config["calibration"]["a_tp"] = 1.0
+        knobs = config["frontier"]["knobs"]
+        knobs["cache_dir"] = str(cache_root)
+        knobs["no_cache"] = False
+        if args.mode == "new-profile-only":
+            for key, filename in PROFILE_KEYS.items():
+                knobs[key] = str((args.profile_root.resolve() / filename))
+
+        target_config = config_root / f"{entry['fixture_id']}.json"
+        write_json(target_config, config)
+        updated_entry = dict(entry)
+        updated_entry["config"] = str(target_config)
+        updated_entry["calibration_scale"] = 1.0
+        entries.append(updated_entry)
+
+    prepared = {
+        "schema": "frontier-qwen30-profile-comparison-prepared.v1",
+        "status": "PASS",
+        "mode": args.mode,
+        "source": {
+            "manifest": str(args.source.resolve()),
+            "sha256": sha256(args.source),
+        },
+        "profile_hashes": profile_hashes,
+        "isolation": {
+            "calibration_a_tp": 1.0,
+            "prediction_cache": str(cache_root),
+            "all_non_profile_knobs_inherited": True,
+        },
+        "entries": entries,
+    }
+    write_json(output / "prepared-manifest.json", prepared)
+    print(output / "prepared-manifest.json")
+
+
+if __name__ == "__main__":
+    main()