hf_model_loader.py

"""Strict loader for the inference-only Hugging Face SEDD artifact."""

from __future__ import annotations

import hashlib
import json
import math
import re
from collections.abc import Mapping
from pathlib import Path
from typing import Any

import torch
from omegaconf import OmegaConf
from safetensors.torch import load_file

from graph_lib import get_graph
from model import SEDD
from noise_lib import get_noise


ARTIFACT_SCHEMA = "sedd_math_tool_hf_inference_v1"
_SHA256_RE = re.compile(r"[0-9a-f]{64}")
_CONFIG_KEYS = {
    "artifact_format",
    "tokens",
    "graph",
    "noise",
    "sampling",
    "model",
    "tokenizer",
}
_MODEL_KEYS = {
    "name",
    "type",
    "hidden_size",
    "cond_dim",
    "length",
    "n_blocks",
    "n_heads",
    "scale_by_sigma",
    "dropout",
}


def sha256_file(path: Path) -> str:
    digest = hashlib.sha256()
    with path.open("rb") as handle:
        for chunk in iter(lambda: handle.read(8 * 1024 * 1024), b""):
            digest.update(chunk)
    return digest.hexdigest()


def _exact_keys(value: Any, keys: set[str], name: str) -> Mapping[str, Any]:
    if not isinstance(value, Mapping) or set(value) != keys:
        raise ValueError(f"Unexpected {name} fields")
    return value


def _positive_int(value: Any, name: str) -> int:
    if type(value) is not int or value <= 0:
        raise ValueError(f"{name} must be a positive integer")
    return value


def _finite_number(value: Any, name: str) -> float:
    if isinstance(value, bool) or not isinstance(value, (int, float)):
        raise ValueError(f"{name} must be numeric")
    result = float(value)
    if not math.isfinite(result):
        raise ValueError(f"{name} must be finite")
    return result


def validate_artifact_config(value: Any) -> dict[str, Any]:
    config = _exact_keys(value, _CONFIG_KEYS, "config")
    if config["artifact_format"] != ARTIFACT_SCHEMA:
        raise ValueError("Config artifact schema does not match the loader")
    _positive_int(config["tokens"], "tokens")

    graph = _exact_keys(config["graph"], {"type"}, "graph config")
    if graph["type"] != "absorb":
        raise ValueError("Only the audited absorbing graph is supported")

    noise = _exact_keys(
        config["noise"], {"type", "sigma_min", "sigma_max"}, "noise config"
    )
    if noise["type"] != "loglinear":
        raise ValueError("Only the audited loglinear noise is supported")
    sigma_min = _finite_number(noise["sigma_min"], "noise.sigma_min")
    sigma_max = _finite_number(noise["sigma_max"], "noise.sigma_max")
    if sigma_min <= 0 or sigma_max <= sigma_min:
        raise ValueError("Invalid noise sigma range")

    sampling = _exact_keys(
        config["sampling"],
        {"predictor", "steps", "noise_removal", "eps"},
        "sampling config",
    )
    if sampling["predictor"] != "euler":
        raise ValueError("Only the audited Euler predictor is supported")
    _positive_int(sampling["steps"], "sampling.steps")
    if type(sampling["noise_removal"]) is not bool:
        raise ValueError("sampling.noise_removal must be boolean")
    if _finite_number(sampling["eps"], "sampling.eps") <= 0:
        raise ValueError("sampling.eps must be positive")

    model = _exact_keys(config["model"], _MODEL_KEYS, "model config")
    if model["type"] != "ddit" or not isinstance(model["name"], str):
        raise ValueError("Unexpected model family")
    for field in ("hidden_size", "cond_dim", "length", "n_blocks", "n_heads"):
        _positive_int(model[field], f"model.{field}")
    if model["hidden_size"] % model["n_heads"]:
        raise ValueError("model.hidden_size must be divisible by model.n_heads")
    if type(model["scale_by_sigma"]) is not bool:
        raise ValueError("model.scale_by_sigma must be boolean")
    dropout = _finite_number(model["dropout"], "model.dropout")
    if not 0 <= dropout < 1:
        raise ValueError("model.dropout must be in [0, 1)")

    tokenizer = _exact_keys(
        config["tokenizer"],
        {"identifier", "vocab_sha256", "add_special_tokens"},
        "tokenizer config",
    )
    if not isinstance(tokenizer["identifier"], str) or not tokenizer["identifier"]:
        raise ValueError("tokenizer.identifier must be a non-empty string")
    if not isinstance(tokenizer["vocab_sha256"], str) or not _SHA256_RE.fullmatch(
        tokenizer["vocab_sha256"]
    ):
        raise ValueError("tokenizer.vocab_sha256 must be a SHA-256 digest")
    if tokenizer["add_special_tokens"] is not False:
        raise ValueError("The audited tokenizer does not add special tokens")
    return dict(config)


def _load_metadata(model_dir: Path) -> dict[str, Any]:
    path = model_dir / "inference_metadata.json"
    value = json.loads(path.read_text(encoding="utf-8"))
    if not isinstance(value, dict) or value.get("schema_version") != ARTIFACT_SCHEMA:
        raise ValueError("Unsupported Hugging Face SEDD artifact metadata")
    return value


def _record_path(
    model_dir: Path,
    record: Mapping[str, Any],
    expected_name: str,
    *,
    verify: bool,
) -> Path:
    if record.get("file") != expected_name:
        raise ValueError(f"Artifact record must name {expected_name}")
    expected_sha = record.get("sha256")
    if not isinstance(expected_sha, str) or not _SHA256_RE.fullmatch(expected_sha):
        raise ValueError(f"Invalid SHA-256 record for {expected_name}")
    expected_size = record.get("size_bytes")
    if type(expected_size) is not int or expected_size < 0:
        raise ValueError(f"Invalid size record for {expected_name}")

    path = model_dir / expected_name
    if not path.is_file():
        raise FileNotFoundError(path)
    actual_size = path.stat().st_size
    if actual_size != expected_size:
        raise RuntimeError(
            f"Size mismatch for {expected_name}: {actual_size} != {expected_size}"
        )
    if verify:
        actual_sha = sha256_file(path)
        if actual_sha != expected_sha:
            raise RuntimeError(f"SHA-256 mismatch for {expected_name}: {actual_sha}")
    return path


def _load_state(
    module: torch.nn.Module,
    path: Path,
    record: Mapping[str, Any],
    name: str,
) -> None:
    if record.get("dtype") != "float32":
        raise ValueError(f"{name} metadata must declare float32")
    expected_count = record.get("tensor_count")
    if type(expected_count) is not int or expected_count <= 0:
        raise ValueError(f"Invalid tensor count for {name}")

    state = load_file(path, device="cpu")
    expected = module.state_dict()
    if len(state) != expected_count:
        raise ValueError(f"Tensor count mismatch for {name}")
    if set(state) != set(expected):
        raise ValueError(f"State keys do not match for {name}")
    for key, tensor in state.items():
        reference = expected[key]
        if tensor.shape != reference.shape:
            raise ValueError(f"Shape mismatch for {name}.{key}")
        if tensor.dtype != torch.float32 or tensor.dtype != reference.dtype:
            raise ValueError(f"Dtype mismatch for {name}.{key}")
        if not bool(torch.isfinite(tensor).all().item()):
            raise ValueError(f"Non-finite tensor in {name}.{key}")
    module.load_state_dict(state, strict=True)


def load_hf_sedd_model(
    model_dir: str | Path,
    device: str | torch.device = "cuda",
    *,
    verify: bool = True,
) -> tuple[SEDD, Any, torch.nn.Module, Any, dict[str, Any]]:
    """Load the complete online weights without a base or training checkpoint.

    Returns ``(model, graph, noise, config, metadata)``. SHA-256 verification is
    enabled by default; disabling it still enforces filenames, sizes, schemas,
    tensor keys, shapes, dtypes, and finite values.
    """

    root = Path(model_dir).resolve()
    metadata = _load_metadata(root)
    weights = metadata.get("weights")
    if not isinstance(weights, Mapping) or metadata.get("weights_variant") != "online":
        raise ValueError("Artifact must contain the audited online weight variant")

    model_record = weights.get("model")
    noise_record = weights.get("noise")
    config_record = metadata.get("config")
    if not all(
        isinstance(record, Mapping)
        for record in (model_record, noise_record, config_record)
    ):
        raise ValueError("Artifact file records are incomplete")

    model_path = _record_path(root, model_record, "model.safetensors", verify=verify)
    noise_path = _record_path(root, noise_record, "noise.safetensors", verify=verify)
    config_path = _record_path(root, config_record, "config.json", verify=verify)
    config_value = validate_artifact_config(
        json.loads(config_path.read_text(encoding="utf-8"))
    )
    config = OmegaConf.create(config_value)

    model = SEDD(config)
    _load_state(model, model_path, model_record, "model")
    noise = get_noise(config)
    _load_state(noise, noise_path, noise_record, "noise")

    target = torch.device(device)
    model = model.to(target).eval()
    noise = noise.to(target).eval()
    graph = get_graph(config, target)
    return model, graph, noise, config, metadata


__all__ = [
    "ARTIFACT_SCHEMA",
    "load_hf_sedd_model",
    "sha256_file",
    "validate_artifact_config",
]