Model/scripts/train.py

#!/usr/bin/env python3
"""
Frox AI Morph 1.1 — Training CLI

Usage:
    python scripts/train.py --family nano --phase pretrain --steps 5000
    python scripts/train.py --family classic --phase sft   --steps 20000
    python scripts/train.py --family classic --phase dpo --dataset-name <hf-hub-preference-dataset>
    python scripts/train.py --family code --phase sft --data ./my_code_data.jsonl

Phases run in order: pretrain → sft → dpo. Each phase resumes from the
previous phase's saved checkpoint automatically if --resume is set.

--family loads exactly one tier from config/family/<name>.py (nano,
mini, classic, pro, or code) — legacy --family 1.5b/3b/8b names still
work too, routed to the closest matching tier.
"""
from __future__ import annotations

import argparse
import sys
from pathlib import Path

sys.path.insert(0, str(Path(__file__).resolve().parent.parent))

import torch

from model.architecture.morph_model import MorphForCausalLM
from tokenizer.morph_tokenizer import build_morph_tokenizer
from training.pipeline.trainer import (
    MorphTrainer, MorphPretrainDataset, MorphSFTDataset, MorphDPODataset,
    apply_lora,
)
from utils.common import (
    set_seed, get_device, describe_device, print_banner, detect_environment,
    load_family_config, FAMILY_TIERS,
)


def main():
    parser = argparse.ArgumentParser(description="Frox AI Morph 1.1 Trainer")
    parser.add_argument("--family", choices=list(FAMILY_TIERS), default="nano",
                        help="Which Morph model-family tier to train (see config/family/*.py)")
    parser.add_argument("--phase", choices=["pretrain", "sft", "dpo"], default="sft")
    parser.add_argument("--steps", type=int, default=None,
                        help="Override max steps for this phase")
    parser.add_argument("--data", type=str, default=None,
                        help="Local JSONL/JSON data file (SFT/DPO). Omit to use HF Hub datasets.")
    parser.add_argument("--dataset-name", type=str, default=None,
                        help="HF Hub dataset override for this phase")
    parser.add_argument("--resume", action="store_true",
                        help="Resume from ./frox-morph-1-1-output checkpoint")
    parser.add_argument("--from-checkpoint", type=str, default=None,
                        help="Load base weights from a specific checkpoint dir")
    parser.add_argument("--no-lora", action="store_true", help="Full fine-tune instead of LoRA")
    parser.add_argument("--seed", type=int, default=1337)
    parser.add_argument("--device", type=str, default=None)
    args = parser.parse_args()

    print_banner()
    set_seed(args.seed)

    device = get_device(args.device)
    print(f"Device: {describe_device(device)}")
    print(f"Environment: {detect_environment()}")

    config, family_module = load_family_config(args.family)
    model_name = getattr(family_module, "MODEL_NAME", args.family.title())

    if args.steps:
        if args.phase == "pretrain":
            config.training.pretrain_max_steps = args.steps
        elif args.phase == "sft":
            config.training.sft_max_steps = args.steps
        elif args.phase == "dpo":
            config.training.dpo_max_steps = args.steps

    print(f"\n📐 {model_name} ({args.family})")
    print(f"   hidden={config.text.hidden_size} layers={config.text.num_hidden_layers} "
          f"heads={config.text.num_attention_heads}/{config.text.num_key_value_heads} "
          f"context={config.text.max_position_embeddings}")

    # Tokenizer
    tokenizer = build_morph_tokenizer(save_path="./frox-morph-1-1-output/tokenizer")

    # Model
    if args.from_checkpoint:
        print(f"\n📂 Loading base weights from {args.from_checkpoint}")
        model = MorphForCausalLM.from_saved(args.from_checkpoint, device="cpu")
    else:
        model = MorphForCausalLM(config.text)

    params = model.param_count()
    print(f"   Parameters: {params['total_billions']}B total, "
          f"{params['trainable_billions']}B trainable\n")

    if not args.no_lora and args.phase != "pretrain":
        model = apply_lora(
            model,
            rank=config.training.lora_rank,
            alpha=config.training.lora_alpha,
            dropout=config.training.lora_dropout,
            target_modules=config.training.lora_target_modules,
        )

    # Dataset
    if args.phase == "pretrain":
        dataset = MorphPretrainDataset(
            tokenizer, seq_len=config.training.pretrain_seq_len,
            data_mix=config.training.data_mix,
        )
    elif args.phase == "sft":
        dataset = MorphSFTDataset(
            tokenizer, data_path=args.data,
            dataset_name=args.dataset_name or "teknium/OpenHermes-2.5",
            seq_len=config.training.sft_seq_len,
        )
    else:  # dpo
        if not args.data and not args.dataset_name:
            parser.error(
                "--phase dpo requires --data (local JSONL) or --dataset-name "
                "(any HF Hub preference dataset with chosen/rejected fields)"
            )
        dataset = MorphDPODataset(
            tokenizer, data_path=args.data,
            dataset_name=args.dataset_name,
            seq_len=config.training.dpo_seq_len,
        )

    trainer = MorphTrainer(
        model=model, tokenizer=tokenizer, config=config,
        train_dataset=dataset if args.phase != "dpo" else MorphSFTDataset(
            tokenizer, max_samples=1  # dummy — DPO uses train_dpo() instead
        ),
        device=device,
    )

    if args.phase == "dpo":
        trainer.train_dpo(dataset)
    else:
        trainer.train(max_steps=args.steps, phase=args.phase)

    final_path = f"./frox-morph-1-1-output/{args.family}_{args.phase}_final"
    if hasattr(trainer.model, "save_pretrained"):
        trainer.model.save_pretrained(final_path)
    elif hasattr(trainer.model, "save"):
        trainer.model.save(final_path)
    print(f"\n✅ {model_name} — {args.phase.upper()} complete. Saved to {final_path}")


if __name__ == "__main__":
    main()