File size: 12,318 Bytes
81cffe7 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 | """Package committed Git history, V4.3 epoch 30/27 inference weights, PyTorch fixtures, V4.2/V2.1
baseline packages and frozen evidence for Mac Core ML work."""
import argparse
import json
import shutil
import subprocess
import sys
import tempfile
import zipfile
from pathlib import Path
ROOT = Path(__file__).resolve().parents[2]
sys.path.insert(0, str(ROOT / "src"))
from vimeml.deployment.v4_variants import VARIANTS
TOKENIZER_DIR = "artifacts/tokenizers/v4-a/ja-unigram-8192"
TRAINING = "outputs/training-v43-20261020"
EVALUATION = "outputs/evaluation-v43-20261020"
EPOCHS = (30, 27)
INPUTS = [
f"{TOKENIZER_DIR}/tokenizer.model",
f"{TOKENIZER_DIR}/tokenizer.vocab",
f"{TOKENIZER_DIR}/manifest.json",
"artifacts/token-data/corpus-v4-8k/manifest.json",
"outputs/evaluation-v4-20261015/inputs",
"outputs/evaluation-v4-20261015/v21",
"outputs/evaluation-v42-20261020",
EVALUATION,
f"{TRAINING}/download-receipt.json",
f"{TRAINING}/remote/README.md",
f"{TRAINING}/remote/manifest.json",
f"{TRAINING}/remote/records/config.json",
f"{TRAINING}/remote/records/plan.json",
f"{TRAINING}/remote/records/environment.json",
f"{TRAINING}/remote/records/summary.json",
f"{TRAINING}/remote/records/full-validation.json",
f"{TRAINING}/remote/records/epoch-validation-history.json",
f"{TRAINING}/remote/records/progress.json",
# V4.2 epoch 20 and V2.1 baselines for same-session size/latency comparisons (no rebuild).
"artifacts/deployment/tiny-ja-v4.2-epoch20-bundle-v1",
"artifacts/deployment/tiny-ja-v4.2-epoch20-fp32-v1",
"artifacts/deployment/tiny-ja-v4.2-epoch20-int8-b32-v1",
"artifacts/deployment/tiny-ja-v4.2-epoch20-kvchunk-fp32-v1",
"artifacts/deployment/tiny-ja-v4.2-epoch20-kvchunk-int8-b32-v1",
"artifacts/deployment/tiny-ja-v4.2-epoch20-client-resources-v1",
"artifacts/deployment/tiny-ja-v2.1-extend5-bundle-v1",
"artifacts/deployment/tiny-ja-v2.1-extend5-fp32-v1",
"artifacts/deployment/tiny-ja-v2.1-extend5-int8-b32-v1",
"artifacts/deployment/tiny-ja-v2.1-extend5-kv-fp32-mac-v42-measure-v1",
"artifacts/deployment/tiny-ja-v2.1-extend5-kv-int8-b32-mac-v42-measure-v1",
]
def git(*args):
return subprocess.check_output(["git", *args], cwd=ROOT, text=True).strip()
def write_json(path, value):
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(value, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
def prepare_weights(stage, epoch):
import numpy as np
import sentencepiece as spm
import torch
from vimeml.training.data import file_sha
from vimeml.training.infer_v4 import JapaneseLMV4
torch.set_num_threads(4)
variant = VARIANTS[f"v43-epoch{epoch}"]
source = f"{TRAINING}/remote/checkpoints/epoch-{epoch}.pt"
checkpoint = ROOT / source
receipt = json.loads((ROOT / TRAINING / "download-receipt.json").read_text(encoding="utf-8"))
expected = next(r["sha256"] for r in receipt["files"] if r["path"] == f"checkpoints/epoch-{epoch}.pt")
digest = file_sha(checkpoint)
if receipt["status"] != "verified" or digest != expected:
raise ValueError(f"epoch-{epoch}.pt differs from the verified download receipt.")
plan = json.loads((ROOT / variant.cuda_scores / "plan.json").read_text(encoding="utf-8"))
if plan["checkpoint_sha256"] != digest:
raise ValueError("Checkpoint differs from the frozen local evaluation.")
saved = torch.load(checkpoint, map_location="cpu", weights_only=True)
if saved["format"] != "vimeml_tiny_gpt_v4" or saved["step"] != variant.step or saved["epoch"] != epoch - 1:
raise ValueError(f"Expected frozen V4.3 epoch {epoch}.")
exported = {key: saved[key] for key in ("format", "architecture", "model", "model_config",
"step", "epoch", "signatures", "tokenizer_identity")}
exported["config"] = {key: saved["config"][key] for key in ("architecture", "token_dir")}
exported["deployment_only"] = True
exported["source_checkpoint_sha256"] = digest
destination = stage / "payload" / variant.checkpoint
destination.parent.mkdir(parents=True, exist_ok=True)
torch.save(exported, destination)
reloaded = torch.load(destination, map_location="cpu", weights_only=True)
if reloaded["model"].keys() != saved["model"].keys() or any(
not torch.equal(value, reloaded["model"][name]) for name, value in saved["model"].items()):
raise ValueError("Inference export changed a model tensor.")
del saved
lm = JapaneseLMV4(destination, ROOT / TOKENIZER_DIR, "cpu")
model = lm.model
processor = spm.SentencePieceProcessor(model_file=str(ROOT / TOKENIZER_DIR / "tokenizer.model"))
text = "今日はとてもいい天気です。駅まで歩いて、この本を読んで、日本語を勉強します。"
content = processor.encode(text * 20, out_type=int)
bos, pad = processor.bos_id(), processor.pad_id()
cases = {"bos": [bos], "short": [bos, *content[:15]], "long": [bos, *content[:127]]}
cases["right_pad"] = cases["short"] + [pad] * (128 - len(cases["short"]))
cases["changed_future"] = cases["short"] + content[:128 - len(cases["short"])]
arrays = {}
with torch.inference_mode():
for name, ids in cases.items():
inputs = torch.tensor([ids], dtype=torch.int32)
arrays[f"{name}_input_ids"] = inputs.numpy()
arrays[f"{name}_logits"] = model(inputs.long()).numpy()
# Single-token KV steps over the long case, the Core ML KV graph's contract.
cache, steps = model.new_cache(), []
long_ids = torch.tensor([cases["long"]], dtype=torch.int32)
for position in range(long_ids.shape[1]):
steps.append(model.forward_cached(long_ids[:, position:position + 1], cache))
arrays["long_kv_step_logits"] = torch.cat(steps, dim=1).numpy()
reference = stage / "payload" / variant.reference
reference.mkdir(parents=True)
np.savez_compressed(reference / "logits.npz", **arrays)
write_json(reference / "reference.json", {
"format": "vimeml_v4_mac_preparation_reference_v1",
"checkpoint_step": variant.step,
"checkpoint_epoch": epoch,
"checkpoint_sha256": digest,
"tokenizer_sha256": lm.metadata["tokenizer_sha256"],
"precision": "fp32", "device": "cpu", "torch_version": torch.__version__,
"cases": {name: {"length": len(ids),
"valid_prefix": len(cases["short"]) if name in {"right_pad", "changed_future"} else len(ids)}
for name, ids in cases.items()},
"kv_steps": {"case": "long", "array": "long_kv_step_logits",
"method": "TinyGPTV4.forward_cached one token at a time from an empty cache"},
"policy": "PyTorch preparation fixtures; not Core ML validation, production quality or device timing.",
})
return {
"variant": variant.name,
"checkpoint_step": variant.step,
"checkpoint_epoch": epoch,
"selection": ("best complete LM validation BPC" if epoch == 30 else "requested comparison epoch")
+ "; named before local IME evaluation",
"source_checkpoint": source,
"source_checkpoint_sha256": digest,
"deployment_checkpoint": variant.checkpoint,
"deployment_checkpoint_sha256": file_sha(destination),
"tokenizer": TOKENIZER_DIR,
"tokenizer_sha256": lm.metadata["tokenizer_sha256"],
"model_config": lm.metadata["model"],
"parameter_count": model.parameter_count(),
"optimizer_included": False,
}
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args()
output = args.output.resolve()
if output.exists():
parser.error("Output exists; choose a new handoff version.")
if git("status", "--porcelain", "--untracked-files=no"):
parser.error("Commit tracked source changes before creating the portable Git bundle.")
branch = git("branch", "--show-current")
if not branch:
parser.error("A named source branch is required.")
from vimeml.training.data import file_sha
from vimeml.posttraining.infer import TRACKS, read_snapshot
for track in TRACKS:
read_snapshot(track) # Verifies pinned candidate-pool hashes before packaging.
maintenance = ROOT / "outputs/maintenance"
maintenance.mkdir(parents=True, exist_ok=True)
with tempfile.TemporaryDirectory(prefix="mac-v43-package-", dir=maintenance) as temporary:
stage = Path(temporary)
metadata = {f"epoch{epoch}": prepare_weights(stage, epoch) for epoch in EPOCHS}
payload = stage / "payload"
for relative in INPUTS:
source = ROOT / relative
if not source.exists():
raise FileNotFoundError(source)
files = [source] if source.is_file() else sorted(p for p in source.rglob("*") if p.is_file())
for path in files:
if (path.is_symlink() or path.name.startswith(".env")
or path.name in {".netrc", "credentials"} or path.suffix == ".pyc"):
raise ValueError(f"Unexpected input: {path}")
target = payload / path.relative_to(ROOT)
target.parent.mkdir(parents=True, exist_ok=True)
shutil.copy2(path, target)
subprocess.run(["git", "bundle", "create", str(stage / "repository.bundle"), branch], cwd=ROOT, check=True)
subprocess.run(["git", "bundle", "verify", str(stage / "repository.bundle")], cwd=ROOT, check=True,
stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
manifest = {
"format": "vimeml_windows_to_mac_v43",
"source_commit": git("rev-parse", "HEAD"),
"source_branch": branch,
"mac_branch": "claude/mac-v43-coreml",
"checkout_name": "VimeML-v43-coreml",
"handoff_record": "outputs/deployment/mac-v43-handoff.json",
"remote": git("remote", "get-url", "origin"),
"models": metadata,
"references": {f"epoch{e}": VARIANTS[f"v43-epoch{e}"].reference for e in EPOCHS},
"frozen_tracks": {k: {"cases": v[0], "sha256": v[1]} for k, v in TRACKS.items()},
"blind_included": False,
"test_included": False,
"coreml_conversion_done": False,
"payload_files": [
{"path": p.relative_to(payload).as_posix(), "bytes": p.stat().st_size, "sha256": file_sha(p)}
for p in sorted(payload.rglob("*")) if p.is_file()
],
"verification": "Receipt/evaluation checkpoint hash; exact tensor equality; pinned track hashes; Git bundle verify; per-file SHA256; ZIP entry paths/sizes.",
}
write_json(stage / "handoff-manifest.json", manifest)
shutil.copy2(ROOT / "scripts/deployment/setup_mac.py", stage / "setup_mac.py")
shutil.copy2(ROOT / "docs/reference/mac-v43-coreml.md", stage / "README.md")
output.parent.mkdir(parents=True, exist_ok=True)
expected = {}
with zipfile.ZipFile(output, "x", compression=zipfile.ZIP_DEFLATED, compresslevel=6) as archive:
for path in sorted(stage.rglob("*")):
if path.is_file():
name = path.relative_to(stage).as_posix()
archive.write(path, name)
expected[name] = path.stat().st_size
with zipfile.ZipFile(output) as archive:
if {p.filename: p.file_size for p in archive.infolist()} != expected or archive.testzip() is not None:
raise ValueError("ZIP directory or CRC differs from packaged files.")
summary = {"zip": output.name, "bytes": output.stat().st_size, "sha256": file_sha(output),
"file_count": len(expected), "source_commit": manifest["source_commit"], "models": metadata}
write_json(output.with_suffix(".json"), summary)
print(json.dumps({k: summary[k] for k in ("zip", "bytes", "sha256", "file_count", "source_commit")},
ensure_ascii=False))
if __name__ == "__main__":
main()
|