load_stentor2.py

from transformers import LlamaForCausalLM, LlamaConfig
from safetensors.torch import load_file
from huggingface_hub import hf_hub_download
import torch

MODEL_ID = "StentorLabs/Stentor2-12M-Preview"

# ── Minimal TokenMonster adapter ───────────────────────────────────────────────

class _Tokenizer:
    def __init__(self, vocab_path):
        import subprocess, sys
        subprocess.run([sys.executable, "-m", "pip", "install", "tokenmonster", "-q"], check=True)
        import tokenmonster
        self._vocab = tokenmonster.load(vocab_path)
        self._vocab.add_special_token("</s>")
        self._vocab.add_special_token("<s>")
        self.eos_token    = "</s>"
        self.bos_token    = "<s>"
        self.pad_token    = "</s>"
        self.eos_token_id = int(self._vocab.token_to_id("</s>"))
        self.bos_token_id = int(self._vocab.token_to_id("<s>"))
        self.pad_token_id = self.eos_token_id

    def encode(self, text):
        """Returns a plain list of int token ids."""
        ids = self._vocab.tokenize(text)
        if hasattr(ids, "tolist"):
            ids = ids.tolist()
        return [int(x) for x in ids]

    def decode(self, token_ids):
        """Accepts a list or 1-D tensor of int token ids."""
        if hasattr(token_ids, "tolist"):
            token_ids = token_ids.tolist()
        return self._vocab.decode(token_ids)

    def __len__(self):
        return len(self._vocab)


# ── Main loader ────────────────────────────────────────────────────────────────

def load_stentor2(model_id=MODEL_ID, dtype=torch.float32):
    # Tokenizer — download vocab binary directly
    vocab_path = hf_hub_download(repo_id=model_id, filename="tokenmonster.vocab")
    tokenizer  = _Tokenizer(vocab_path)

    # Model
    config = LlamaConfig.from_pretrained(model_id)
    model  = LlamaForCausalLM(config)

    weight_path = hf_hub_download(repo_id=model_id, filename="model.safetensors")
    raw_sd      = load_file(weight_path)

    # Remap weight_master → weight (INT8 QAT training artifact)
    sd      = {}
    masters = {k for k in raw_sd if k.endswith(".weight_master")}
    skip    = {k[:-len("_master")] for k in masters}
    for k, v in raw_sd.items():
        if k.endswith(".weight_master"): sd[k[:-len("_master")]] = v
        elif k not in skip:              sd[k] = v

    model.load_state_dict(sd, strict=False)
    model.to(dtype).eval()
    return model, tokenizer