grassmann-cache / src /services /model_hub_service.py
Nicholas Finch
fix: default workers=0 (sequential) for Docker cache — ProcessPoolExecutor hangs on cpu-basic
d7013e6
Raw History Blame Contribute Delete
2.47 kB
"""
Model Hub Service — Save/load models in a self-contained format.
Bundles config + weights + tokenizer info into a single directory
that anyone can download and run inference with.
Saved format:
model_dir/
config.json # GrassmannConfig as JSON
model.pt # state_dict weights
tokenizer_info.json # vocab info for reconstruction
"""
from __future__ import annotations
import json
from dataclasses import asdict
from pathlib import Path
import torch
from src.models.config import GrassmannConfig
from src.models.grassmann_seq2seq import GrassmannSeq2SeqModel
from src.models.grassmann_model import GrassmannTranscriptionModel
from src.services.tokenizer_service import CharTokenizer
def save_model(
model: GrassmannSeq2SeqModel | GrassmannTranscriptionModel,
config: GrassmannConfig,
save_dir: str,
) -> None:
"""Save model + config in a self-contained directory."""
path = Path(save_dir)
path.mkdir(parents=True, exist_ok=True)
# Save config
config_dict = asdict(config)
with open(path / "config.json", "w") as f:
json.dump(config_dict, f, indent=2)
# Save weights
torch.save(model.state_dict(), path / "model.pt")
# Save tokenizer info
tokenizer = CharTokenizer()
tok_info = {
"type": "char",
"vocab_size": tokenizer.vocab_size,
"blank_id": tokenizer.blank_id,
}
with open(path / "tokenizer_info.json", "w") as f:
json.dump(tok_info, f, indent=2)
print(f" Model saved to {save_dir}/")
print(f" config.json, model.pt, tokenizer_info.json")
def load_model(
load_dir: str,
device: str = "cpu",
) -> tuple[GrassmannSeq2SeqModel | GrassmannTranscriptionModel, GrassmannConfig, CharTokenizer]:
"""Load a saved model from directory. Returns (model, config, tokenizer)."""
path = Path(load_dir)
# Load config
with open(path / "config.json") as f:
config_dict = json.load(f)
config = GrassmannConfig(**config_dict)
# Build model
tokenizer = CharTokenizer()
config.vocab_size = tokenizer.vocab_size
if config.use_seq2seq:
model = GrassmannSeq2SeqModel(config)
else:
model = GrassmannTranscriptionModel(config)
# Load weights
state = torch.load(path / "model.pt", map_location=device, weights_only=True)
model.load_state_dict(state)
model.to(device)
model.eval()
return model, config, tokenizer