Datasets:
File size: 8,275 Bytes
7230edc | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 | #!/usr/bin/env python3
"""Build the public KernelZero CUDA SFT dataset from the exact training files."""
from __future__ import annotations
import argparse
import hashlib
import json
import math
import re
import statistics
from pathlib import Path
import pyarrow as pa
import pyarrow.parquet as pq
STAGES = [
{
"name": "stage1",
"sft": "resources/modules/turn1/cudallm_24k_sft_v1.json",
"sources": [
"resources/modules/turn1/cudallm_modules_distilled_0_40000_checked_passed.jsonl",
"resources/modules/turn1/cudallm_modules_distilled_40000_90000_checked_passed.jsonl",
],
},
{
"name": "stage2",
"sft": "resources/modules/turn3/cudallm_19k_sft_t2t3.json",
"sources": [
"resources/modules/turn2/cudallm_modules_distilled_0_40000_2_checked_passed.jsonl",
"resources/modules/turn2/cudallm_modules_distilled_40000_90000_2_checked_passed.jsonl",
"resources/modules/turn3/cudallm_modules_distilled_0_40000_3_distilled_checked_passed.jsonl",
"resources/modules/turn3/cudallm_modules_distilled_40000_90000_3_distilled_checked_passed.jsonl",
],
},
]
def read_jsonl(path: Path):
with path.open(encoding="utf-8") as handle:
for line_no, line in enumerate(handle, 1):
if line.strip():
try:
yield json.loads(line)
except json.JSONDecodeError as exc:
raise ValueError(f"Invalid JSON in {path}:{line_no}") from exc
def extract_cuda_code(answer: str) -> tuple[str, bool]:
answer = answer.strip()
match = re.fullmatch(r"```(?:python)?\s*\n?(.*?)\n?```", answer, flags=re.DOTALL)
if match:
return match.group(1).strip(), True
return answer, False
def sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(8 * 1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--root", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--num-shards", type=int, default=8)
args = parser.parse_args()
root = args.root.resolve()
output = args.output.resolve()
data_dir = output / "data"
data_dir.mkdir(parents=True, exist_ok=True)
records: list[dict] = []
stage_counts: dict[str, int] = {}
output_mismatches: list[dict] = []
missing_speedup: list[dict] = []
unfenced_answers: list[dict] = []
for stage in STAGES:
source_by_index: dict[int, dict] = {}
for relative in stage["sources"]:
source_path = root / relative
for item in read_jsonl(source_path):
index = int(item["index"])
if index in source_by_index:
raise ValueError(f"Duplicate source index in {stage['name']}: {index}")
source_by_index[index] = item
with (root / stage["sft"]).open(encoding="utf-8") as handle:
sft_rows = json.load(handle)
stage_counts[stage["name"]] = len(sft_rows)
if len(sft_rows) != len(source_by_index):
raise ValueError(
f"Count mismatch for {stage['name']}: "
f"SFT={len(sft_rows)} source={len(source_by_index)}"
)
for sft in sft_rows:
index = int(sft["index"])
source = source_by_index.get(index)
if source is None:
raise ValueError(f"Missing source for {stage['name']} index={index}")
response = source.get("response") or {}
reasoning = str(response.get("reasoning", "")).strip()
answer = str(response.get("answer", "")).strip()
reconstructed_output = f"<think>{reasoning}</think>\n{answer}"
if reconstructed_output != sft["output"]:
output_mismatches.append({"stage": stage["name"], "index": index})
result = source.get("result") or {}
message = result.get("msg") or {}
speedup = message.get("speedup") if isinstance(message, dict) else None
if not isinstance(speedup, (int, float)) or not math.isfinite(float(speedup)):
missing_speedup.append({"stage": stage["name"], "index": index})
continue
cuda_code, fenced = extract_cuda_code(answer)
if not fenced:
unfenced_answers.append({"stage": stage["name"], "index": index})
operators = source.get("ops", "[]")
if isinstance(operators, str):
try:
operators = json.loads(operators)
except json.JSONDecodeError:
operators = [operators]
operators = [str(op) for op in operators]
records.append(
{
"input": sft["instruction"],
"output": sft["output"],
"torch_module": str(source["module"]),
"cuda_code": cuda_code,
"speedup": float(speedup),
"operators": operators,
"source_stage": stage["name"],
"source_index": index,
}
)
if output_mismatches:
raise ValueError(f"Output mismatches: {len(output_mismatches)}")
if missing_speedup:
raise ValueError(f"Missing or invalid speedups: {len(missing_speedup)}")
if len(records) != 43_302:
raise ValueError(f"Expected 43,302 records, found {len(records)}")
identities = {(row["source_stage"], row["source_index"]) for row in records}
if len(identities) != len(records):
raise ValueError("Duplicate source identities")
full_hashes = {
hashlib.sha256(
(row["input"] + "\0" + row["output"] + "\0" + row["torch_module"]).encode()
).hexdigest()
for row in records
}
if len(full_hashes) != len(records):
raise ValueError("Duplicate training records")
schema = pa.schema(
[
("input", pa.string()),
("output", pa.string()),
("torch_module", pa.string()),
("cuda_code", pa.string()),
("speedup", pa.float64()),
("operators", pa.list_(pa.string())),
("source_stage", pa.string()),
("source_index", pa.int64()),
]
)
shard_size = math.ceil(len(records) / args.num_shards)
shard_paths: list[Path] = []
for shard_id in range(args.num_shards):
start = shard_id * shard_size
end = min(len(records), start + shard_size)
if start >= end:
break
table = pa.Table.from_pylist(records[start:end], schema=schema)
path = data_dir / f"train-{shard_id:05d}-of-{args.num_shards:05d}.parquet"
pq.write_table(table, path, compression="zstd", compression_level=9)
shard_paths.append(path)
speedups = [row["speedup"] for row in records]
report = {
"record_count": len(records),
"stage_counts": stage_counts,
"unique_source_identities": len(identities),
"unique_training_records": len(full_hashes),
"output_mismatches": len(output_mismatches),
"missing_speedup": len(missing_speedup),
"unfenced_answers": len(unfenced_answers),
"empty_fields": {
field: sum(not row[field] for row in records)
for field in ("input", "output", "torch_module", "cuda_code")
},
"speedup": {
"min": min(speedups),
"max": max(speedups),
"mean": statistics.fmean(speedups),
"median": statistics.median(speedups),
},
"shards": [
{"path": str(path.relative_to(output)), "bytes": path.stat().st_size, "sha256": sha256(path)}
for path in shard_paths
],
}
(output / "validation_report.json").write_text(
json.dumps(report, indent=2, ensure_ascii=False) + "\n", encoding="utf-8"
)
print(json.dumps(report, indent=2, ensure_ascii=False))
if __name__ == "__main__":
main()
|