upload_v12_base.py

#!/usr/bin/env python
"""
upload_v12_base.py — push the frozen v12 base + the code needed to load it to HF.

Run on the pod from /workspace after `huggingface-cli login` (or pass token=...):
    python upload_v12_base.py

Uploads to Daxamite/V12Emoe:
    ckpt_v12_190m_best.pt   <- out_v12_190m/best.pt (the frozen student)
    model_hybrid.py         <- the GPT/GPTConfig the ckpt deserializes into
    muon.py                 <- imported by model_hybrid / training code
    tok_v9.py               <- the v9_chatml_enc tokenizer factory
    README.md               <- minimal load instructions (so the repo is self-describing)

These are exactly the files run_hyper.sh / runtime_adapters.py pull to rebuild the
base, so a later RuntimeConfig(base_ckpt=...) can point straight at this repo.
"""

import os
from huggingface_hub import HfApi, create_repo

REPO   = "Daxamite/V12Emoe"
TOKEN  = os.environ.get("HF_TOKEN")          # or rely on `huggingface-cli login`
CKPT_LOCAL  = "out_v12_190m/best.pt"
CKPT_REMOTE = "ckpt_v12_190m_best.pt"

# (local_path, path_in_repo). Code files keep their names (importable after download).
FILES = [
    (CKPT_LOCAL,        CKPT_REMOTE),
    ("model_hybrid.py", "model_hybrid.py"),
    ("muon.py",         "muon.py"),
    ("tok_v9.py",       "tok_v9.py"),
]

README = f"""---
license: other
tags: [v12, emoe, fft-hybrid, frozen-base]
---

# V12 eMoE — frozen base (190M FFT-hybrid)

The frozen "student" for the v12 ephemeral-MoE system: a 12L / 1024d / 16h
StripedHyena-style FFT-spectral hybrid (`attn_every=3`), `block_size=2048`,
tokenizer `v9_chatml_enc` (vocab 50264). Trained Muon+AdamW (3.0x ratio), WSD,
~7.5B tokens. **best.pt VAL 1.66.** The hypernetwork specializes this; it is
never fine-tuned here.

## Load

```python
import torch
from dataclasses import fields
from model_hybrid import GPT, GPTConfig

ckpt = torch.load("ckpt_v12_190m_best.pt", map_location="cpu")
margs = ckpt.get("model_args") or ckpt.get("config") or {{}}
cfg = GPTConfig(**{{k: v for k, v in margs.items()
                    if k in {{f.name for f in fields(GPTConfig)}}}})
sd = ckpt.get("model") or ckpt.get("state_dict") or ckpt.get("model_state_dict")
sd = {{k.replace("_orig_mod.", ""): v for k, v in sd.items()}}
model = GPT(cfg); model.load_state_dict(sd, strict=False); model.eval()

import tok_v9; tok = tok_v9.build()
```
"""


def main():
    api = HfApi(token=TOKEN)
    create_repo(REPO, repo_type="model", exist_ok=True, token=TOKEN)

    missing = [p for p, _ in FILES if not os.path.exists(p)]
    if missing:
        raise SystemExit(f"missing local files (run from /workspace?): {missing}")

    for local, remote in FILES:
        sz = os.path.getsize(local) / 1e6
        print(f">>> uploading {local} -> {REPO}/{remote}  ({sz:.1f} MB)")
        api.upload_file(path_or_fileobj=local, path_in_repo=remote,
                        repo_id=REPO, repo_type="model")

    api.upload_file(path_or_fileobj=README.encode(), path_in_repo="README.md",
                    repo_id=REPO, repo_type="model")
    print(f">>> done. https://huggingface.co/{REPO}")


if __name__ == "__main__":
    main()