src/train.py

"""
MORPH-AI Training Script v6
Trains the full MorphModel v6 (novel components + LoRA on base).
Memory-efficient training with 4-bit QLoRA, gradient checkpointing, 8-bit optimizer,
MoE expert pruning, and MoD sparsity regularization.

Runs on free Colab T4 or local GPU/CPU.
"""

import gc
import json
import os
import sys
import torch
from pathlib import Path

# ensure both `python src/train.py` (root on path) and
# `python -m src.train` / Colab (src package) imports work
_ROOT = Path(__file__).resolve().parent.parent
if str(_ROOT) not in sys.path:
    sys.path.insert(0, str(_ROOT))
if str(_ROOT / "src") not in sys.path:
    sys.path.insert(0, str(_ROOT / "src"))

try:
    from src.architecture import MorphConfig, MorphModel
except ImportError:
    from architecture import MorphConfig, MorphModel


def build_chat_text(ex) -> str:
    if isinstance(ex, str):
        return ex  # already formatted ChatML text
    prompt = ex.get("prompt", ex.get("input", ""))
    response = ex.get("response", ex.get("output", ""))
    skill = ex.get("skill_token", "")
    system_prompt = ex.get("system_prompt", "")

    if system_prompt:
        system_msg = system_prompt
    elif skill:
        system_msg = f"You are a specialized module. Use the skill marker {skill}."
    else:
        system_msg = "You are a helpful assistant."

    return (
        "<|im_start|>system\n"
        + system_msg
        + "<|im_end|>\n"
        + "<|im_start|>user\n"
        + prompt
        + "<|im_end|>\n"
        + "<|im_start|>assistant\n"
        + response
        + "<|im_end|>\n"
    )


def load_jsonl(dir_path: str) -> list:
    """Load dataset lines. Lines may be raw ChatML strings OR JSON objects
    with prompt/response keys; both are handled downstream by build_chat_text."""
    data_dir = Path(dir_path)
    files = sorted(data_dir.glob("*.jsonl"))
    if not files:
        raise FileNotFoundError(f"No .jsonl files found in {dir_path}")

    examples = []
    for f in files:
        with open(f, encoding="utf-8") as fh:
            for line in fh:
                line = line.strip()
                if not line:
                    continue
                try:
                    examples.append(json.loads(line))
                except json.JSONDecodeError:
                    examples.append(line)  # raw text line
    print(f"Loaded {len(examples)} examples from {len(files)} files")
    return examples


def train(
    output_dir: str = "./output/morph-model",
    base_model: str = "Qwen/Qwen2.5-1.5B-Instruct",
    data_dir: str = "./datasets",
    num_epochs: int = 3,
    per_device_batch_size: int = 4,
    gradient_accumulation_steps: int = 8,
    learning_rate: float = 2e-4,
    max_seq_len: int = 8192,
    use_4bit: bool = True,
    train_components: bool = True,
    use_8bit_optimizer: bool = True,
    prune_experts_every: int = 500,
    mod_sparsity_weight: float = 0.01,
):
    from transformers import (
        AutoModelForCausalLM,
        AutoTokenizer,
        BitsAndBytesConfig,
        DataCollatorForLanguageModeling,
        Trainer,
        TrainingArguments,
    )
    from datasets import Dataset

    if torch.cuda.is_available():
        print(f"GPU: {torch.cuda.get_device_name(0)}")
    else:
        print("No GPU detected - running on CPU (slow). Set use_4bit=False.")
        use_4bit = False
        use_8bit_optimizer = False

    config = MorphConfig(base_model=base_model, max_seq_len=max_seq_len)
    model = MorphModel(config)

    if use_4bit:
        del model.base_model_raw
        gc.collect()
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        model.base_model_raw = AutoModelForCausalLM.from_pretrained(
            base_model,
            quantization_config=BitsAndBytesConfig(
                load_in_4bit=True,
                bnb_4bit_compute_dtype=torch.bfloat16,
                bnb_4bit_use_double_quant=True,
                bnb_4bit_quant_type="nf4",
            ),
            device_map="auto",
            trust_remote_code=True,
        )
        print("Base model loaded in 4-bit (nf4)")

    from peft import prepare_model_for_kbit_training

    if use_4bit:
        model.base_model_raw = prepare_model_for_kbit_training(
            model.base_model_raw, use_gradient_checkpointing=True
        )
        print("Base model prepared for k-bit training")

    model.apply_lora()

    V6_COMPONENTS = (
        "coordinator",
        "reasoner",
        "code_bias",
        "scratchpad",
        "verifier",
        "skill_module",
        "depth_module",
        "moe_block",
        "memory",
        "mod",
        "multimodal_fusion",
        "tool_use",
        "document_module",
        "video_module",
    )
    if train_components:
        for name, param in model.named_parameters():
            if name.startswith(V6_COMPONENTS):
                param.requires_grad = True
        print(f"Trainable params: {model.get_trainable_params():,}")
    else:
        print(f"Trainable params (LoRA only): {model.get_trainable_params():,}")

    try:
        model.base_model.gradient_checkpointing_enable()
    except Exception as e:
        print(f"Note: gradient checkpointing skipped ({e})")

    tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    examples = load_jsonl(data_dir)
    texts = [build_chat_text(ex) for ex in examples]

    def tokenize(text: str):
        return tokenizer(
            text,
            truncation=True,
            max_length=max_seq_len,
            padding=False,
            return_tensors=None,
        )

    raw = [tokenize(t) for t in texts]
    token_ids = [torch.tensor(r["input_ids"]).unsqueeze(0) for r in raw]
    try:
        from src.architecture import build_code_features
    except ImportError:
        from architecture import build_code_features

    code_feats = [build_code_features(tokenizer, ids) for ids in token_ids]

    records = []
    for r, cf in zip(raw, code_feats):
        records.append(
            {
                "input_ids": r["input_ids"],
                "attention_mask": r["attention_mask"],
                "code_feat": cf.squeeze(0),
            }
        )
    dataset = Dataset.from_list(records)

    def collate(batch):
        import torch.nn.functional as Fn
        from transformers import DataCollatorForLanguageModeling

        max_len = max(len(b["input_ids"]) for b in batch)
        input_ids = torch.full((len(batch), max_len), tokenizer.pad_token_id, dtype=torch.long)
        attention_mask = torch.zeros((len(batch), max_len), dtype=torch.long)
        for i, b in enumerate(batch):
            n = len(b["input_ids"])
            input_ids[i, :n] = torch.tensor(b["input_ids"])
            attention_mask[i, :n] = torch.tensor(b["attention_mask"])

        feat0 = torch.tensor(batch[0]["code_feat"])
        feat_dim = feat0.shape[-1]
        code_feat = torch.zeros((len(batch), max_len, feat_dim), dtype=torch.float32)
        for i, b in enumerate(batch):
            cf = torch.tensor(b["code_feat"])
            n = cf.shape[0]
            code_feat[i, :n] = cf

        lm_collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)
        lm_batch = lm_collator([{"input_ids": b["input_ids"]} for b in batch])
        return {
            "input_ids": input_ids,
            "attention_mask": attention_mask,
            "labels": lm_batch["labels"],
            "code_feat": code_feat,
        }

    optim = "paged_adamw_8bit" if use_8bit_optimizer else "adamw_torch"
    training_args = TrainingArguments(
        output_dir="./checkpoints",
        per_device_train_batch_size=per_device_batch_size,
        gradient_accumulation_steps=gradient_accumulation_steps,
        num_train_epochs=num_epochs,
        learning_rate=learning_rate,
        warmup_steps=50,
        logging_steps=10,
        save_steps=500,
        save_total_limit=2,
        report_to=[],
        optim=optim,
        gradient_checkpointing=True,
        bf16=torch.cuda.is_available(),
    )

    class MorphTrainer(Trainer):
        def on_step_begin(self, args, state, control, **kwargs):
            if state.global_step > 0 and prune_experts_every > 0:
                if state.global_step % prune_experts_every == 0:
                    model.prune_experts()

    trainer = MorphTrainer(
        model=model,
        args=training_args,
        train_dataset=dataset,
        data_collator=collate,
    )

    trainer.train()

    out = Path(output_dir)
    out.mkdir(parents=True, exist_ok=True)
    model.save_checkpoint(str(out))
    tokenizer.save_pretrained(str(out))
    print(f"Model saved to {out}")


if __name__ == "__main__":
    import argparse

    parser = argparse.ArgumentParser(description="MORPH-AI training")
    parser.add_argument("--output", default="./output/morph-model")
    parser.add_argument("--base", default="Qwen/Qwen2.5-1.5B-Instruct")
    parser.add_argument("--data", default="./datasets")
    parser.add_argument("--epochs", type=int, default=3)
    parser.add_argument("--batch", type=int, default=4)
    parser.add_argument("--grad-accum", type=int, default=8)
    parser.add_argument("--lr", type=float, default=2e-4)
    parser.add_argument("--max-len", type=int, default=8192)
    parser.add_argument("--no-4bit", action="store_true", help="Disable 4-bit quantization")
    parser.add_argument(
        "--no-components",
        action="store_true",
        help="Train LoRA only (freeze novel components)",
    )
    parser.add_argument("--no-8bit-optim", action="store_true", help="Disable 8-bit optimizer")
    parser.add_argument("--prune-every", type=int, default=500, help="Prune MoE experts every N steps (0=off)")
    parser.add_argument("--mod-sparsity", type=float, default=0.01, help="MoD sparsity loss weight")
    args = parser.parse_args()

    train(
        output_dir=args.output,
        base_model=args.base,
        data_dir=args.data,
        num_epochs=args.epochs,
        per_device_batch_size=args.batch,
        gradient_accumulation_steps=args.grad_accum,
        learning_rate=args.lr,
        max_seq_len=args.max_len,
        use_4bit=not args.no_4bit,
        train_components=not args.no_components,
        use_8bit_optimizer=not args.no_8bit_optim,
        prune_experts_every=args.prune_every,
        mod_sparsity_weight=args.mod_sparsity,
    )