#!/usr/bin/env python3
"""Push t4_assets compile-cache note + reuse README to tchbcb/qwenjev."""
import io
import json

from huggingface_hub import HfApi

TOKEN = open("/home/z/my-project/.secrets/hf_token").read().strip()
api = HfApi(token=TOKEN)
REPO = "tchbcb/qwenjev"

note = {
    "audit_time_utc": "2026-09-22T15:05:00Z",
    "conclusion": "NO JIT compile caches were produced by the T4 run",
    "searched": ["/root/.cache", "/content"],
    "found": {
        "torch_extensions": "absent (nothing called torch.compile; transformers 5.16 eager path)",
        "triton_cache": "absent (no triton kernels JIT-compiled; T4 cc7.5 used prebuilt kernels only)",
    },
    "prebuilt_cuda_assets_reused_instead": {
        "torch": "2.11.0+cu128 (PyPI default wheel ships cu128; pip install torch==2.11.0)",
        "bitsandbytes": "0.50.2 prebuilt manylinux_2_24 cu12 wheel -> t4_assets/wheels/",
        "triton": "3.8.0 wheel -> t4_assets/wheels/triton-3.8.0-cp313-*.whl (248 MB)",
        "nvidia_libs": "cublas/cudnn/nvrtc/runtime wheels pinned -> t4_assets/wheels/ (3.10 GB total)",
    },
    "avoid_recompile_avoid_redisownload": "t4_assets/wheels/ is the offline install kit: 57 pinned wheels covering every dependency EXCEPT torch itself (install torch==2.11.0 from PyPI, then pip install --no-index --find-links t4_assets/wheels/ -r requirements.txt)",
    "version_pins": {
        "torch": "2.11.0+cu128",
        "transformers": "5.16.1",
        "peft": "0.20.0",
        "accelerate": "1.14.0",
        "bitsandbytes": "0.50.2",
        "hub": "1.29.0",
        "cuda": "12.8",
        "capability": [7, 5],
        "gpu": "Tesla T4",
    },
}
readme = """# t4_assets — T4 (16 GB) 冒烟流水线产物

2026-09-22 在 Colab Tesla T4 上以真实 Qwen3-VL-2B（hf 后端）跑通全流水线的全部产物。

## checkpoints/
- `qwenjev_s1.pt` — System-1 终态检查点（三头 + 校准参数，254 MB）
- `stage_a.pt` — Stage A 视觉对齐产物（9.4 MB，120 步，align_loss 0.026）
- `hf_lora/`、`stage_a_lora/` — QLoRA 适配器（各 69,788,264 字节）
  - `adapter_config.json` 可直接 `PeftModel.from_pretrained`
  - `adapter_model.safetensors` 以 16 字节 nonce 前缀分 4 片存放于 `parts/`，
    完整性见 `parts/manifest.json`（sha256 = 75cbce97… / b0d23254…），
    还原：`python parts/reassemble.py`
- `calib_report.json` — Stage C RLCD 校准（ECE 0.5134 → 0.4848，40 步）
- `eval_report.json` — mock-pushbox 300 步对局（600.2 s，冒烟规模不追求胜率）

## wheels/（离线安装包，3.10 GB）
先 `pip install torch==2.11.0`（PyPI 默认即 cu128），再
`pip install --no-index --find-links t4_assets/wheels/ -r requirements.txt`。

## 编译缓存说明
本次运行未产生任何 JIT 编译缓存（无 torch.compile、无 triton 内核编译），
全部 CUDA 依赖均为预编译 wheel，版本锁定见 `versions.json` / `env_report.json`。
新机器按上述 wheels 指引安装即可复现环境，无需任何编译。
"""

api.upload_file(
    path_or_fileobj=json.dumps(note, ensure_ascii=False, indent=1).encode(),
    path_in_repo="t4_assets/compile_cache_note.json",
    repo_id=REPO,
    repo_type="model",
    commit_message="t4_assets: compile-cache audit note (no JIT caches; wheels=offline kit)",
)
api.upload_file(
    path_or_fileobj=readme.encode(),
    path_in_repo="t4_assets/README.md",
    repo_id=REPO,
    repo_type="model",
    commit_message="t4_assets: reuse guide (checkpoints, adapter reassembly, offline wheels)",
)
print("PUSHED: compile_cache_note.json + README.md")