File size: 11,525 Bytes
8433089 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 | #!/usr/bin/env python3
"""HAKO end-to-end runner: phases 0-4 under a global wall-clock budget.
Phase 0 environment + math property suite (fail-fast; no training if any
theorem check fails -- proofs come BEFORE code execution).
Phase 1 Qwen2.5-0.5B: fetch -> decompose (int4 kept) -> Z -> K-Means++ ->
plateau -> GHSOM growth -> head pretraining.
Phase 2 Granite-4.0-1B: decompose -> register -> joint cooperative fusion
with MoE cross-attention, chain recursion, diffusion game loop.
Phase 3 NLP/NLG: byte-BPE (Lemma-1 verified) + conditioned decoder on
ultra-alpaca-ptbr + corpus-ptbr-v1 + EN + zh.
Phase 4 final validation, checkpoint verification, HF publish (only if
every phase succeeded and no token leak is detected).
Usage: python3 run_all.py [--phase N] [--smoke]
"""
from __future__ import annotations
import argparse
import json
import logging
import os
import sys
import time
from pathlib import Path
ROOT = Path(__file__).resolve().parent
sys.path.insert(0, str(ROOT))
# Managed caches: keep every HF artifact inside the project tree so the
# MemoryManager can sweep them (never grow the user's global cache).
os.environ.setdefault("HF_HOME", "/home/z/my-project/work/hf_home")
os.environ.setdefault("HF_DATASETS_CACHE", "/home/z/my-project/work/hf_home/datasets")
os.environ.setdefault("TMPDIR", "/home/z/my-project/work/tmp")
logging.basicConfig(level=logging.INFO,
format="%(asctime)s %(name)s %(levelname)s %(message)s")
log = logging.getLogger("hako.run")
def main() -> int:
ap = argparse.ArgumentParser()
ap.add_argument("--phase", type=int, default=0)
ap.add_argument("--start", type=int, default=1,
help="first training phase to run (1..3); earlier "
"phases are restored from checkpoints")
ap.add_argument("--smoke", action="store_true",
help="short smoke budget for CI-like validation")
args = ap.parse_args()
from hako.config import CONFIG
from hako.memory_manager import MemoryManager
from hako.telemetry import Telemetry
cfg = CONFIG
cfg.ensure_dirs()
mem = MemoryManager(cfg)
tel = Telemetry(cfg)
hw = cfg.derived()
log.info("hardware: cpus=%d ram=%.2fGiB disk_free=%.2fGiB",
hw.cpu_count, hw.ram_total / 2**30, hw.disk_free / 2**30)
smoke = args.smoke
if smoke:
cfg.budget_phase1 = 240
cfg.budget_phase2 = 180
cfg.budget_phase3 = 200
cfg.budget_phase4 = 60
failures = 0
# ---------------- phase 0: math property suite ------------------------
if args.phase <= 0:
log.info("=== PHASE 0: mathematical property verification ===")
import subprocess
r = subprocess.run([sys.executable,
str(ROOT / "tests" / "test_math_properties.py")],
capture_output=True, text=True)
print(r.stdout[-4000:])
if r.returncode != 0:
log.error("math property suite FAILED -- aborting (proofs first)")
tel.orchestration(event="phase0_failed")
return 2
tel.orchestration(event="phase0_math_verified")
# ---------------- phase 0.5: corpus cache (isolated RAM) --------------
cache_file = Path(cfg.work_dir) / "cache" / "corpus_ptbr.jsonl"
if not cache_file.exists():
import subprocess
cache_file.parent.mkdir(parents=True, exist_ok=True)
log.info("extracting corpus-ptbr-v1 cache (isolated subprocess)...")
r = subprocess.run(
[sys.executable, "-m", "hako.sources.corpus_cache",
str(cache_file), "6000"],
cwd=str(ROOT), capture_output=True, text=True, timeout=600)
if r.returncode != 0:
log.warning("corpus cache extraction failed: %s",
r.stderr[-300:])
else:
log.info("corpus cache ready (%s)", cache_file.name)
tel.orchestration(event="corpus_cache_ready")
carry: dict = {}
# ---------------- phase 1 ---------------------------------------------
if args.start <= 1 and args.phase <= 1:
from hako.train.phase1_qwen import run as run1
log.info("=== PHASE 1: Qwen source (budget %ds) ===", cfg.budget_phase1)
try:
carry = run1(cfg, tel, mem, cfg.budget_phase1)
tel.orchestration(event="phase1_ok",
steps=int(carry.get("steps", 0)))
except Exception as exc: # noqa: BLE001
log.exception("phase 1 failed")
tel.orchestration(event="phase1_failed", error=str(exc)[:300])
failures += 1
if smoke:
return 3
elif args.start == 2 and failures == 0:
# resume: rebuild the system from the phase-1 checkpoint
import numpy as _np
import torch as _torch
from hako.sources import loader as _loader
from hako.train.restore import restore_system
ck1 = Path(cfg.ckpt_dir) / "hako_phase1.npz"
zw = Path(cfg.artifacts_dir) / "phase1_Z_workspace.npz"
if ck1.exists() and zw.exists():
sys_, _ = restore_system(cfg, tel, mem, ck1)
data = _np.load(zw)
from hako.nlp.datasets import Corpus
corpus = Corpus()
docs = _loader.fetch_datasets_streaming(
cfg, mem, {"pt_alpaca": 9000, "en": 6000, "zh": 2200})
for lang, dl in docs.items():
corpus.add_docs(lang, dl)
cache_file = Path(cfg.work_dir) / "cache" / "corpus_ptbr.jsonl"
if cache_file.exists():
import json as _json
with open(cache_file, "r", encoding="utf-8") as fh:
corpus.add_docs("pt", [_json.loads(l)["text"]
for l in fh])
carry = {"system": sys_,
"Z_t": _torch.as_tensor(data["Z_src"]),
"Z_N": _torch.as_tensor(data["Z_N"]),
"labels": data["labels"], "corpus": corpus,
"checkpoint": str(ck1)}
tel.orchestration(event="phase1_restored", step=int(sys_.step))
else:
log.error("--start 2 but phase-1 checkpoint missing")
failures += 1
# ---------------- phase 2 ---------------------------------------------
if args.start <= 2 and args.phase <= 2 and failures == 0 and \
carry.get("system") is not None:
from hako.train.phase2_granite_joint import run as run2
log.info("=== PHASE 2: Granite joint fusion (budget %ds) ===",
cfg.budget_phase2)
try:
carry = run2(cfg, tel, mem, cfg.budget_phase2, carry)
tel.orchestration(event="phase2_ok")
except Exception as exc: # noqa: BLE001
log.exception("phase 2 failed")
tel.orchestration(event="phase2_failed", error=str(exc)[:300])
failures += 1
elif args.start == 3 and failures == 0:
import numpy as _np
import torch as _torch
from hako.sources import loader as _loader
from hako.train.restore import restore_system
ck2 = Path(cfg.ckpt_dir) / "hako_phase2.npz"
zw = Path(cfg.artifacts_dir) / "phase1_Z_workspace.npz"
if ck2.exists() and zw.exists():
sys_, _ = restore_system(cfg, tel, mem, ck2)
data = _np.load(zw)
from hako.nlp.datasets import Corpus
corpus = Corpus()
docs = _loader.fetch_datasets_streaming(
cfg, mem, {"pt_alpaca": 9000, "en": 6000, "zh": 2200})
for lang, dl in docs.items():
corpus.add_docs(lang, dl)
cache_file = Path(cfg.work_dir) / "cache" / "corpus_ptbr.jsonl"
if cache_file.exists():
import json as _json
with open(cache_file, "r", encoding="utf-8") as fh:
corpus.add_docs("pt", [_json.loads(l)["text"]
for l in fh])
from hako.sources.decompose import dequantize_int4
gnpz = _np.load(Path(cfg.artifacts_dir) /
"granite1b_decomposed.npz")
protos = dequantize_int4(gnpz["int4_proto_packed"],
gnpz["int4_proto_scales"],
int(gnpz["proto_dim_for_unpack"][0]))
carry = {"system": sys_,
"Z_t": _torch.as_tensor(data["Z_src"]),
"Z_N": _torch.as_tensor(data["Z_N"]),
"labels": data["labels"], "corpus": corpus,
"granite_protos": _torch.as_tensor(protos),
"checkpoint": str(ck2)}
tel.orchestration(event="phase2_restored", step=int(sys_.step))
else:
log.error("--start 3 but phase-2 checkpoint missing")
failures += 1
# ---------------- phase 3 ---------------------------------------------
if args.start <= 3 and args.phase <= 3 and failures == 0 and \
carry.get("system") is not None:
from hako.train.phase3_nlp import run as run3
log.info("=== PHASE 3: NLP/NLG (budget %ds) ===", cfg.budget_phase3)
try:
carry = run3(cfg, tel, mem, cfg.budget_phase3, carry)
tel.orchestration(event="phase3_ok")
except Exception as exc: # noqa: BLE001
log.exception("phase 3 failed")
tel.orchestration(event="phase3_failed", error=str(exc)[:300])
failures += 1
# ---------------- phase 4: validation + publish -----------------------
if args.phase <= 4:
log.info("=== PHASE 4: final validation & publish ===")
from hako.publish.push_hf import publish, security_scan
viol = security_scan(ROOT)
if viol:
log.error("token leak detected in %d files -- NOT publishing", viol)
failures += 1
if failures == 0:
summary = {
"nlg_val_loss": carry.get("nlg_val_loss"),
"nlg_sample": (carry.get("nlg_sample") or "")[:300],
"telemetry": tel.summary(),
"hw": {"cpus": hw.cpu_count,
"ram_GiB": round(hw.ram_total / 2**30, 2)},
"published_from": str(ROOT),
}
(ROOT / "artifacts" / "final_summary.json").write_text(
json.dumps(summary, indent=2, default=str), encoding="utf-8")
repo = os.environ.get("HAKO_REPO", cfg.hf_repo)
if os.environ.get("HF_TOKEN"):
try:
url = publish(ROOT, repo, tel.summary())
log.info("PUBLISHED: %s", url)
tel.orchestration(event="published", url=url)
except Exception as exc: # noqa: BLE001
log.error("publish failed: %s", exc)
failures += 1
else:
log.warning("HF_TOKEN not set -- skipping publish "
"(state saved locally)")
# final state save
if carry.get("system") is not None:
from hako.checkpoint import save_state
save_state(ROOT / "checkpoints" / "hako_final.npz",
**carry["system"].state_blocks())
log.info("run finished with %d failure(s)", failures)
return 1 if failures else 0
if __name__ == "__main__":
sys.exit(main())
|