Research-Stack/4-Infrastructure/shim/h200_encode_runner.py
2026-05-11 22:18:31 -05:00

96 lines
3.9 KiB
Python

#!/usr/bin/env python3
"""Dry-run H200 burst optimizer scaffold for finance LUT compression.
This does not require H200 hardware. It prepares the corpus/codebook search
surface and emits receipt-backed candidate summaries for later GPU rental.
"""
from __future__ import annotations
import argparse
import json
import platform
from pathlib import Path
from typing import Any
import finance_claim_lut_harness as harness
REPO = Path(__file__).resolve().parents[2]
SHIM = REPO / "4-Infrastructure" / "shim"
def load_bundle_receipt(path: Path) -> dict[str, Any]:
if path.is_dir():
path = path / "finance_claim_lut_harness_receipt.json"
return json.loads(path.read_text(encoding="utf-8"))
def candidate_summary(receipt: dict[str, Any]) -> list[dict[str, Any]]:
candidates = []
for sample in receipt.get("samples", []):
metrics = sample["metrics"]
best_known = min(
value
for value in [
metrics.get("combined_fcl1_fcs1_bytes"),
metrics.get("zlib_canonical_bytes"),
metrics.get("cbor", {}).get("bytes"),
metrics.get("messagepack", {}).get("bytes"),
metrics.get("protobuf_dynamic", {}).get("bytes"),
]
if isinstance(value, int)
)
candidates.append(
{
"sample_id": sample["id"],
"current_fcl1_fcs1_bytes": metrics["combined_fcl1_fcs1_bytes"],
"best_known_baseline_bytes": best_known,
"target": "reduce FCS1 literal overhead and improve enum/value clustering",
"promote": False,
"reason": "dry-run only; no GPU search performed",
}
)
return candidates
def run(args: argparse.Namespace) -> dict[str, Any]:
receipt = load_bundle_receipt(args.corpus)
candidates = candidate_summary(receipt)
rejected = [
{"name": "provider_live_job", "reason": "requires explicit rental, budget, and environment receipt"},
{"name": "decoder_requires_gpu", "reason": "violates compact deterministic decoder boundary"},
{"name": "compression_claim_from_tiny_corpus", "reason": "corpus still too small for competitive claim"},
]
out_dir = args.out_dir
out_dir.mkdir(parents=True, exist_ok=True)
(out_dir / "candidate_lut_summary.json").write_text(json.dumps(candidates, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
(out_dir / "rejected_candidates.json").write_text(json.dumps(rejected, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
dry_run_receipt = {
"schema": "h200_encode_runner_receipt_v1",
"mode": "dry_run",
"corpus": harness.repo_path(args.corpus),
"out_dir": harness.repo_path(out_dir),
"sample_count": len(receipt.get("samples", [])),
"candidate_count": len(candidates),
"rejected_count": len(rejected),
"environment": {"python": platform.python_version(), "platform": platform.platform(), "gpu_required": False},
"next_live_gate": "rent H200 only after local bundle, Netcup baseline, and noisy simulator receipts are lawful",
"lawful": len(candidates) > 0 and all(not item["promote"] for item in candidates),
"claim_boundary": "dry-run optimizer scaffold only; no H200 hardware used and no improved compression claim made",
}
(out_dir / "h200_encode_runner_receipt.json").write_text(json.dumps(dry_run_receipt, indent=2, ensure_ascii=False) + "\n", encoding="utf-8")
return dry_run_receipt
def main() -> int:
parser = argparse.ArgumentParser()
parser.add_argument("--corpus", type=Path, default=SHIM / "finance_claim_remote_bundle")
parser.add_argument("--out-dir", type=Path, default=SHIM / "h200_encode_dry_run")
args = parser.parse_args()
print(json.dumps(run(args), indent=2, ensure_ascii=False))
return 0
if __name__ == "__main__":
raise SystemExit(main())