Add files using upload-large-folder tool

3df5819 verified 6 months ago

15.9 kB

"""
Full training entry point.
Run: python scripts/train.py --config configs/training_config.yaml
"""

import click
import yaml
import torch
import os
import gc
from transformers import TrainingArguments, Seq2SeqTrainingArguments
from loguru import logger

try:
    import wandb
    HAS_WANDB = True
except ImportError:
    HAS_WANDB = False

from src.model.base_model import load_model_and_tokenizer
from src.model.style_conditioner import StyleConditioner
from src.training.dataset import WritingCorrectionDataset
from src.training.loss_functions import CombinedCorrectionLoss, CombinedCorrectionLossV2
from src.training.trainer import CorrectionTrainer
from src.training.callbacks import StyleMetricsCallback, EarlyStoppingOnStyleDrift
from src.style.fingerprinter import StyleFingerprinter
from src.evaluation.gleu_scorer import GLEUScorer


# ── Hybrid GPU Management ───────────────────────────────────────────────────
def _setup_device():
    """Detect GPU and configure hybrid VRAM management.
    Returns (device, gpu_info) where gpu_info is a dict with:
      - available: bool
      - name: str
      - vram_total_mb: int
      - vram_free_mb: int
      - compute_cap: tuple
    """
    gpu_info = {"available": False, "name": "CPU", "vram_total_mb": 0,
                "vram_free_mb": 0, "compute_cap": (0, 0)}

    if not torch.cuda.is_available():
        logger.info("No GPU detected — training on CPU")
        return "cpu", gpu_info

    gpu_info["available"] = True
    gpu_info["name"] = torch.cuda.get_device_name(0)
    gpu_info["compute_cap"] = torch.cuda.get_device_capability(0)

    # Query actual free VRAM
    vram_total = torch.cuda.get_device_properties(0).total_memory // (1024 * 1024)
    vram_reserved = torch.cuda.memory_reserved(0) // (1024 * 1024)
    vram_allocated = torch.cuda.memory_allocated(0) // (1024 * 1024)
    vram_free = vram_total - vram_allocated

    gpu_info["vram_total_mb"] = vram_total
    gpu_info["vram_free_mb"] = vram_free

    logger.info(
        f"GPU: {gpu_info['name']} | "
        f"VRAM: {vram_allocated}MB used / {vram_total}MB total ({vram_free}MB free) | "
        f"Compute: {gpu_info['compute_cap']}"
    )

    # Leave headroom for the system — reserve at most 85% of free VRAM
    # This prevents the desktop/compositor from starving
    usable_vram_mb = int(vram_free * 0.85)
    if usable_vram_mb > 0:
        # Set PyTorch memory limit to avoid hogging all VRAM
        fraction = min(usable_vram_mb / vram_total, 0.90)
        torch.cuda.set_per_process_memory_fraction(fraction, 0)
        logger.info(
            f"Hybrid GPU mode: capped PyTorch VRAM to {fraction:.0%} "
            f"(~{int(vram_total * fraction)}MB), leaving room for system"
        )

    return "cuda", gpu_info


def _auto_batch_size(model_key: str, device: str, gpu_info: dict,
                     config_batch: int) -> int:
    """Pick optimal batch size based on model size and available resources."""
    if device == "cpu":
        # CPU: T5-Small can handle batch=8 with 32GB RAM, larger models less
        if "small" in model_key:
            return min(config_batch, 8)
        return min(config_batch, 2)

    # GPU: estimate based on free VRAM
    free_mb = gpu_info["vram_free_mb"]

    # Rough VRAM per sample estimates (bf16, seq_len=128):
    #   T5-Small: ~120MB model + ~50MB/sample
    #   T5-Base:  ~350MB model + ~90MB/sample
    #   T5-Large: ~900MB model + ~150MB/sample
    model_vram_estimates = {
        "flan-t5-small": {"model_mb": 160, "per_sample_mb": 60},
        "flan-t5-base": {"model_mb": 400, "per_sample_mb": 100},
        "flan-t5-large": {"model_mb": 1000, "per_sample_mb": 160},
        "flan-t5-xl": {"model_mb": 3000, "per_sample_mb": 300},
    }
    est = model_vram_estimates.get(model_key, {"model_mb": 500, "per_sample_mb": 120})

    # Available for batches = free VRAM - model footprint - 300MB safety buffer
    available_for_batches = free_mb - est["model_mb"] - 300
    if available_for_batches <= 0:
        logger.warning("Very tight VRAM — using batch_size=1")
        return 1

    max_batch = max(1, available_for_batches // est["per_sample_mb"])
    optimal = min(config_batch, max_batch)

    logger.info(
        f"Auto batch size: {optimal} "
        f"(model ~{est['model_mb']}MB + {optimal}×{est['per_sample_mb']}MB "
        f"= ~{est['model_mb'] + optimal * est['per_sample_mb']}MB / {free_mb}MB free)"
    )
    return max(1, optimal)


@click.command()
@click.option("--config", default="configs/training_config.yaml")
@click.option("--use-v2-loss", is_flag=True, help="Use V2 loss with human pattern term")
def train(config: str, use_v2_loss: bool):
    """Launch the full training pipeline."""
    # Step 1: Load config
    logger.info("Step 1: Loading config...")
    with open(config) as f:
        cfg = yaml.safe_load(f)

    model_cfg = cfg.get("model", {})
    lora_cfg = cfg.get("lora", {})
    data_cfg = cfg.get("data", {})
    train_cfg = cfg.get("training", {})
    loss_cfg = cfg.get("loss", {})
    gen_cfg = cfg.get("generation", {})

    # Step 2: Initialise W&B (optional)
    logger.info("Step 2: Initialising experiment tracking...")
    if HAS_WANDB and os.environ.get("WANDB_API_KEY"):
        wandb.init(
            project="dyslexia-rewriter",
            name=f"train-{model_cfg.get('key', 'flan-t5')}", 
            config=cfg,
        )
    else:
        logger.info("W&B not configured, logging to TensorBoard only")
        os.environ["WANDB_DISABLED"] = "true"

    # Step 3: Detect GPU and configure hybrid VRAM management
    logger.info("Step 3: Setting up device (hybrid GPU mode)...")
    device, gpu_info = _setup_device()

    # Step 4: Load model + tokenizer
    logger.info("Step 4: Loading model and tokenizer...")
    model_key = model_cfg.get("key", "flan-t5-small")
    model, tokenizer, is_seq2seq = load_model_and_tokenizer(
        model_key=model_key,
        quantize=model_cfg.get("quantize", False),
        use_lora=model_cfg.get("use_lora", True),
        lora_config_dict=lora_cfg,
    )

    # Required for PEFT + gradient checkpointing compatibility
    if hasattr(model, 'enable_input_require_grads'):
        model.enable_input_require_grads()

    # ── torch.compile for fused kernels (PyTorch 2.x) ───────────────────────
    if hasattr(torch, "compile") and device == "cuda":
        try:
            # "default" mode: fuses kernels via Triton without CUDA graphs.
            # "reduce-overhead" uses CUDA graphs which break with LoRA/PEFT
            # (tensor outputs get overwritten between graph replays).
            logger.info("Applying torch.compile(mode='default')...")
            model = torch.compile(model, mode="default")
            logger.info("✓ torch.compile applied — first few steps will be slower (compiling)")
        except Exception as e:
            logger.warning(f"torch.compile failed (non-fatal): {e}")

    # Step 5: Create fingerprinter
    logger.info("Step 5: Creating style fingerprinter...")
    fingerprinter = StyleFingerprinter(
        spacy_model="en_core_web_sm",  # Use small model for training speed
        awl_path="data/awl/coxhead_awl.txt",
    )

    # Step 6: Create datasets
    logger.info("Step 6: Loading datasets...")
    train_dataset = WritingCorrectionDataset(
        data_path=data_cfg.get("train_path", "data/processed/train.jsonl"),
        tokenizer=tokenizer,
        fingerprinter=fingerprinter,
        max_input_length=data_cfg.get("max_input_length", 512),
        max_target_length=data_cfg.get("max_target_length", 512),
        augment_with_synthetic=data_cfg.get("augment_synthetic", True),
        synthetic_ratio=data_cfg.get("synthetic_ratio", 0.3),
    )

    val_dataset = WritingCorrectionDataset(
        data_path=data_cfg.get("val_path", "data/processed/val.jsonl"),
        tokenizer=tokenizer,
        fingerprinter=fingerprinter,
        max_input_length=data_cfg.get("max_input_length", 512),
        max_target_length=data_cfg.get("max_target_length", 512),
        augment_with_synthetic=False,
    )

    logger.info(f"Train: {len(train_dataset)} | Val: {len(val_dataset)}")

    # Free memory after dataset loading
    gc.collect()
    if device == "cuda":
        torch.cuda.empty_cache()

    # Use simple CE-only loss for training — aux models (sentence-transformer,
    # GPT-2, HP classifier) are NOT loaded since they provide no gradient signal
    # (they decode via argmax under no_grad). This saves ~1GB+ memory.
    from torch import nn
    class CEOnlyLoss(nn.Module):
        """Cross-entropy only loss — the only loss that provides gradient signal."""
        def __init__(self):
            super().__init__()
            self.ce_loss = nn.CrossEntropyLoss(ignore_index=-100)

        def forward(self, logits, labels, **kwargs):
            if logits.dim() == 3:
                ce_logits = logits.view(-1, logits.size(-1))
                ce_labels = labels.view(-1)
            else:
                ce_logits = logits
                ce_labels = labels
            l_ce = self.ce_loss(ce_logits, ce_labels)
            return {"total_loss": l_ce, "ce_loss": l_ce}

    loss_fn = CEOnlyLoss()
    logger.info("Using CE-only loss (aux models skipped to save memory)")

    # Step 8: Create training arguments
    logger.info("Step 8: Creating training arguments...")

    # Auto-detect precision support
    use_bf16 = False
    use_fp16 = False
    if device == "cuda":
        if gpu_info["compute_cap"][0] >= 8:
            use_bf16 = True
            logger.info("Using BF16 (Ampere+ GPU)")
        else:
            use_fp16 = True
            logger.info("Using FP16 (pre-Ampere GPU)")
    elif device == "cpu":
        # Zen 3+ CPUs (Ryzen 5000+) support BF16 in PyTorch 2.x
        try:
            test = torch.tensor([1.0], dtype=torch.bfloat16)
            _ = test + test  # Test BF16 compute works
            use_bf16 = True
            logger.info("Using BF16 on CPU (Zen 3+ detected)")
        except Exception:
            logger.info("BF16 not supported on this CPU, using FP32")

    # Smart batch size based on model + available resources
    config_batch = train_cfg.get("per_device_train_batch_size", 4)
    batch_size = _auto_batch_size(model_key, device, gpu_info, config_batch)

    # Smart gradient checkpointing:
    # - ENABLE for large models (saves VRAM at cost of compute)
    # - DISABLE for small models (they fit in VRAM, checkpointing is pure overhead)
    # - ALWAYS DISABLE on CPU (plenty of RAM, checkpointing wastes CPU cycles)
    large_models = {"flan-t5-large", "flan-t5-xl", "llama-3.1-8b"}
    use_grad_ckpt = model_key in large_models and device == "cuda"
    if use_grad_ckpt:
        logger.info("Gradient checkpointing: ON (large model, saving VRAM)")
    else:
        logger.info(f"Gradient checkpointing: OFF ({'small model fits in VRAM' if device == 'cuda' else 'CPU has plenty of RAM'})")

    # Dataloader workers: Python 3.14 changed default start method to "forkserver"
    # on Linux, which hits "too many fds" with num_workers > 0.
    # Use 0 (main-process loading) — dataset is pre-tokenized so overhead is minimal.
    num_workers = train_cfg.get("dataloader_num_workers", 0)

    # Filter report_to to only available tools
    report_to = []
    if HAS_WANDB and os.environ.get("WANDB_API_KEY"):
        report_to.append("wandb")
    report_to.append("tensorboard")

    training_args = TrainingArguments(
        output_dir=train_cfg.get("output_dir", "checkpoints/"),
        num_train_epochs=train_cfg.get("num_train_epochs", 5),
        per_device_train_batch_size=batch_size,
        per_device_eval_batch_size=train_cfg.get("per_device_eval_batch_size", 8) if device == "cuda" else 2,
        gradient_accumulation_steps=train_cfg.get("gradient_accumulation_steps", 8),
        learning_rate=train_cfg.get("learning_rate", 3e-4),
        lr_scheduler_type=train_cfg.get("lr_scheduler_type", "cosine"),
        warmup_ratio=train_cfg.get("warmup_ratio", 0.05),
        weight_decay=train_cfg.get("weight_decay", 0.01),
        fp16=use_fp16,
        bf16=use_bf16,
        eval_strategy=train_cfg.get("evaluation_strategy", "steps"),
        eval_steps=train_cfg.get("eval_steps", 100),
        save_strategy=train_cfg.get("save_strategy", "steps"),
        save_steps=train_cfg.get("save_steps", 100),
        save_total_limit=train_cfg.get("save_total_limit", 3),
        load_best_model_at_end=False,  # Handled manually below (PEFT adapters break Trainer's loader)
        metric_for_best_model=train_cfg.get("metric_for_best_model", "eval_loss"),
        greater_is_better=train_cfg.get("greater_is_better", False),
        logging_dir=train_cfg.get("logging_dir", "logs/"),
        logging_steps=train_cfg.get("logging_steps", 25),
        report_to=report_to,
        dataloader_num_workers=num_workers,
        seed=train_cfg.get("seed", 42),
        remove_unused_columns=False,  # We have custom columns (style_vector, etc.)
        gradient_checkpointing=use_grad_ckpt,
    )

    # Step 9: Create trainer
    logger.info("Step 9: Creating trainer...")
    trainer = CorrectionTrainer(
        loss_fn=loss_fn,
        fingerprinter=fingerprinter,
        tokenizer=tokenizer,
        model=model,
        args=training_args,
        train_dataset=train_dataset,
        eval_dataset=val_dataset,
        callbacks=[
            StyleMetricsCallback(),
            EarlyStoppingOnStyleDrift(min_style_similarity=0.75),
        ],
    )

    # Step 10: Train
    logger.info("Step 10: Starting training...")
    logger.info(
        f"Config summary: model={model_key} | batch={batch_size} | "
        f"accum={training_args.gradient_accumulation_steps} | "
        f"effective_batch={batch_size * training_args.gradient_accumulation_steps} | "
        f"epochs={training_args.num_train_epochs} | "
        f"precision={'bf16' if use_bf16 else 'fp16' if use_fp16 else 'fp32'} | "
        f"grad_ckpt={use_grad_ckpt} | device={device}"
    )
    trainer.train()

    # Step 11: Save best model (manual PEFT-aware loading)
    logger.info("Step 11: Saving best model...")
    output_dir = train_cfg.get("output_dir", "checkpoints/")
    save_path = os.path.join(output_dir, "best_model")

    # Find best checkpoint from trainer state
    best_ckpt = None
    state_path = os.path.join(output_dir, "trainer_state.json")
    # Check each checkpoint for trainer_state.json
    import glob
    for ckpt_dir in sorted(glob.glob(os.path.join(output_dir, "checkpoint-*"))):
        ts = os.path.join(ckpt_dir, "trainer_state.json")
        if os.path.exists(ts):
            import json as json_mod
            with open(ts) as f:
                state = json_mod.load(f)
            best_path = state.get("best_model_checkpoint")
            if best_path:
                best_ckpt = best_path

    if best_ckpt and os.path.isdir(best_ckpt):
        logger.info(f"Loading best checkpoint from {best_ckpt}")
        from peft import PeftModel
        # Reload the best adapter weights
        best_adapter = os.path.join(best_ckpt, "adapter_model.safetensors")
        if os.path.exists(best_adapter):
            model.load_adapter(best_ckpt, adapter_name="default")
            logger.info(f"Loaded best adapter from {best_ckpt}")
        else:
            logger.warning(f"No adapter found at {best_ckpt}, saving current model")
    else:
        logger.info("No best checkpoint found, saving final model state")

    trainer.save_model(save_path)
    tokenizer.save_pretrained(save_path)
    logger.info(f"Model saved to {save_path}")

    if HAS_WANDB and wandb.run is not None:
        wandb.finish()

    logger.info("✓ Training complete!")


if __name__ == "__main__":
    train()

	"""
	Full training entry point.
	Run: python scripts/train.py --config configs/training_config.yaml
	"""

	import click
	import yaml
	import torch
	import os
	import gc
	from transformers import TrainingArguments, Seq2SeqTrainingArguments
	from loguru import logger

	try:
	import wandb
	HAS_WANDB = True
	except ImportError:
	HAS_WANDB = False

	from src.model.base_model import load_model_and_tokenizer
	from src.model.style_conditioner import StyleConditioner
	from src.training.dataset import WritingCorrectionDataset
	from src.training.loss_functions import CombinedCorrectionLoss, CombinedCorrectionLossV2
	from src.training.trainer import CorrectionTrainer
	from src.training.callbacks import StyleMetricsCallback, EarlyStoppingOnStyleDrift
	from src.style.fingerprinter import StyleFingerprinter
	from src.evaluation.gleu_scorer import GLEUScorer


	# ── Hybrid GPU Management ───────────────────────────────────────────────────
	def _setup_device():
	"""Detect GPU and configure hybrid VRAM management.

	Returns (device, gpu_info) where gpu_info is a dict with:
	- available: bool
	- name: str
	- vram_total_mb: int
	- vram_free_mb: int
	- compute_cap: tuple
	"""
	gpu_info = {"available": False, "name": "CPU", "vram_total_mb": 0,
	"vram_free_mb": 0, "compute_cap": (0, 0)}

	if not torch.cuda.is_available():
	logger.info("No GPU detected — training on CPU")
	return "cpu", gpu_info

	gpu_info["available"] = True
	gpu_info["name"] = torch.cuda.get_device_name(0)
	gpu_info["compute_cap"] = torch.cuda.get_device_capability(0)

	# Query actual free VRAM
	vram_total = torch.cuda.get_device_properties(0).total_memory // (1024 * 1024)
	vram_reserved = torch.cuda.memory_reserved(0) // (1024 * 1024)
	vram_allocated = torch.cuda.memory_allocated(0) // (1024 * 1024)
	vram_free = vram_total - vram_allocated

	gpu_info["vram_total_mb"] = vram_total
	gpu_info["vram_free_mb"] = vram_free

	logger.info(
	f"GPU: {gpu_info['name']} \| "
	f"VRAM: {vram_allocated}MB used / {vram_total}MB total ({vram_free}MB free) \| "
	f"Compute: {gpu_info['compute_cap']}"
	)

	# Leave headroom for the system — reserve at most 85% of free VRAM
	# This prevents the desktop/compositor from starving
	usable_vram_mb = int(vram_free * 0.85)
	if usable_vram_mb > 0:
	# Set PyTorch memory limit to avoid hogging all VRAM
	fraction = min(usable_vram_mb / vram_total, 0.90)
	torch.cuda.set_per_process_memory_fraction(fraction, 0)
	logger.info(
	f"Hybrid GPU mode: capped PyTorch VRAM to {fraction:.0%} "
	f"(~{int(vram_total * fraction)}MB), leaving room for system"
	)

	return "cuda", gpu_info


	def _auto_batch_size(model_key: str, device: str, gpu_info: dict,
	config_batch: int) -> int:
	"""Pick optimal batch size based on model size and available resources."""
	if device == "cpu":
	# CPU: T5-Small can handle batch=8 with 32GB RAM, larger models less
	if "small" in model_key:
	return min(config_batch, 8)
	return min(config_batch, 2)

	# GPU: estimate based on free VRAM
	free_mb = gpu_info["vram_free_mb"]

	# Rough VRAM per sample estimates (bf16, seq_len=128):
	# T5-Small: ~120MB model + ~50MB/sample
	# T5-Base: ~350MB model + ~90MB/sample
	# T5-Large: ~900MB model + ~150MB/sample
	model_vram_estimates = {
	"flan-t5-small": {"model_mb": 160, "per_sample_mb": 60},
	"flan-t5-base": {"model_mb": 400, "per_sample_mb": 100},
	"flan-t5-large": {"model_mb": 1000, "per_sample_mb": 160},
	"flan-t5-xl": {"model_mb": 3000, "per_sample_mb": 300},
	}
	est = model_vram_estimates.get(model_key, {"model_mb": 500, "per_sample_mb": 120})

	# Available for batches = free VRAM - model footprint - 300MB safety buffer
	available_for_batches = free_mb - est["model_mb"] - 300
	if available_for_batches <= 0:
	logger.warning("Very tight VRAM — using batch_size=1")
	return 1

	max_batch = max(1, available_for_batches // est["per_sample_mb"])
	optimal = min(config_batch, max_batch)

	logger.info(
	f"Auto batch size: {optimal} "
	f"(model ~{est['model_mb']}MB + {optimal}×{est['per_sample_mb']}MB "
	f"= ~{est['model_mb'] + optimal * est['per_sample_mb']}MB / {free_mb}MB free)"
	)
	return max(1, optimal)


	@click.command()
	@click.option("--config", default="configs/training_config.yaml")
	@click.option("--use-v2-loss", is_flag=True, help="Use V2 loss with human pattern term")
	def train(config: str, use_v2_loss: bool):
	"""Launch the full training pipeline."""
	# Step 1: Load config
	logger.info("Step 1: Loading config...")
	with open(config) as f:
	cfg = yaml.safe_load(f)

	model_cfg = cfg.get("model", {})
	lora_cfg = cfg.get("lora", {})
	data_cfg = cfg.get("data", {})
	train_cfg = cfg.get("training", {})
	loss_cfg = cfg.get("loss", {})
	gen_cfg = cfg.get("generation", {})

	# Step 2: Initialise W&B (optional)
	logger.info("Step 2: Initialising experiment tracking...")
	if HAS_WANDB and os.environ.get("WANDB_API_KEY"):
	wandb.init(
	project="dyslexia-rewriter",
	name=f"train-{model_cfg.get('key', 'flan-t5')}",
	config=cfg,
	)
	else:
	logger.info("W&B not configured, logging to TensorBoard only")
	os.environ["WANDB_DISABLED"] = "true"

	# Step 3: Detect GPU and configure hybrid VRAM management
	logger.info("Step 3: Setting up device (hybrid GPU mode)...")
	device, gpu_info = _setup_device()

	# Step 4: Load model + tokenizer
	logger.info("Step 4: Loading model and tokenizer...")
	model_key = model_cfg.get("key", "flan-t5-small")
	model, tokenizer, is_seq2seq = load_model_and_tokenizer(
	model_key=model_key,
	quantize=model_cfg.get("quantize", False),
	use_lora=model_cfg.get("use_lora", True),
	lora_config_dict=lora_cfg,
	)

	# Required for PEFT + gradient checkpointing compatibility
	if hasattr(model, 'enable_input_require_grads'):
	model.enable_input_require_grads()

	# ── torch.compile for fused kernels (PyTorch 2.x) ───────────────────────
	if hasattr(torch, "compile") and device == "cuda":
	try:
	# "default" mode: fuses kernels via Triton without CUDA graphs.
	# "reduce-overhead" uses CUDA graphs which break with LoRA/PEFT
	# (tensor outputs get overwritten between graph replays).
	logger.info("Applying torch.compile(mode='default')...")
	model = torch.compile(model, mode="default")
	logger.info("✓ torch.compile applied — first few steps will be slower (compiling)")
	except Exception as e:
	logger.warning(f"torch.compile failed (non-fatal): {e}")

	# Step 5: Create fingerprinter
	logger.info("Step 5: Creating style fingerprinter...")
	fingerprinter = StyleFingerprinter(
	spacy_model="en_core_web_sm", # Use small model for training speed
	awl_path="data/awl/coxhead_awl.txt",
	)

	# Step 6: Create datasets
	logger.info("Step 6: Loading datasets...")
	train_dataset = WritingCorrectionDataset(
	data_path=data_cfg.get("train_path", "data/processed/train.jsonl"),
	tokenizer=tokenizer,
	fingerprinter=fingerprinter,
	max_input_length=data_cfg.get("max_input_length", 512),
	max_target_length=data_cfg.get("max_target_length", 512),
	augment_with_synthetic=data_cfg.get("augment_synthetic", True),
	synthetic_ratio=data_cfg.get("synthetic_ratio", 0.3),
	)

	val_dataset = WritingCorrectionDataset(
	data_path=data_cfg.get("val_path", "data/processed/val.jsonl"),
	tokenizer=tokenizer,
	fingerprinter=fingerprinter,
	max_input_length=data_cfg.get("max_input_length", 512),
	max_target_length=data_cfg.get("max_target_length", 512),
	augment_with_synthetic=False,
	)

	logger.info(f"Train: {len(train_dataset)} \| Val: {len(val_dataset)}")

	# Free memory after dataset loading
	gc.collect()
	if device == "cuda":
	torch.cuda.empty_cache()

	# Use simple CE-only loss for training — aux models (sentence-transformer,
	# GPT-2, HP classifier) are NOT loaded since they provide no gradient signal
	# (they decode via argmax under no_grad). This saves ~1GB+ memory.
	from torch import nn
	class CEOnlyLoss(nn.Module):
	"""Cross-entropy only loss — the only loss that provides gradient signal."""
	def __init__(self):
	super().__init__()
	self.ce_loss = nn.CrossEntropyLoss(ignore_index=-100)

	def forward(self, logits, labels, **kwargs):
	if logits.dim() == 3:
	ce_logits = logits.view(-1, logits.size(-1))
	ce_labels = labels.view(-1)
	else:
	ce_logits = logits
	ce_labels = labels
	l_ce = self.ce_loss(ce_logits, ce_labels)
	return {"total_loss": l_ce, "ce_loss": l_ce}

	loss_fn = CEOnlyLoss()
	logger.info("Using CE-only loss (aux models skipped to save memory)")

	# Step 8: Create training arguments
	logger.info("Step 8: Creating training arguments...")

	# Auto-detect precision support
	use_bf16 = False
	use_fp16 = False
	if device == "cuda":
	if gpu_info["compute_cap"][0] >= 8:
	use_bf16 = True
	logger.info("Using BF16 (Ampere+ GPU)")
	else:
	use_fp16 = True
	logger.info("Using FP16 (pre-Ampere GPU)")
	elif device == "cpu":
	# Zen 3+ CPUs (Ryzen 5000+) support BF16 in PyTorch 2.x
	try:
	test = torch.tensor([1.0], dtype=torch.bfloat16)
	_ = test + test # Test BF16 compute works
	use_bf16 = True
	logger.info("Using BF16 on CPU (Zen 3+ detected)")
	except Exception:
	logger.info("BF16 not supported on this CPU, using FP32")

	# Smart batch size based on model + available resources
	config_batch = train_cfg.get("per_device_train_batch_size", 4)
	batch_size = _auto_batch_size(model_key, device, gpu_info, config_batch)

	# Smart gradient checkpointing:
	# - ENABLE for large models (saves VRAM at cost of compute)
	# - DISABLE for small models (they fit in VRAM, checkpointing is pure overhead)
	# - ALWAYS DISABLE on CPU (plenty of RAM, checkpointing wastes CPU cycles)
	large_models = {"flan-t5-large", "flan-t5-xl", "llama-3.1-8b"}
	use_grad_ckpt = model_key in large_models and device == "cuda"
	if use_grad_ckpt:
	logger.info("Gradient checkpointing: ON (large model, saving VRAM)")
	else:
	logger.info(f"Gradient checkpointing: OFF ({'small model fits in VRAM' if device == 'cuda' else 'CPU has plenty of RAM'})")

	# Dataloader workers: Python 3.14 changed default start method to "forkserver"
	# on Linux, which hits "too many fds" with num_workers > 0.
	# Use 0 (main-process loading) — dataset is pre-tokenized so overhead is minimal.
	num_workers = train_cfg.get("dataloader_num_workers", 0)

	# Filter report_to to only available tools
	report_to = []
	if HAS_WANDB and os.environ.get("WANDB_API_KEY"):
	report_to.append("wandb")
	report_to.append("tensorboard")

	training_args = TrainingArguments(
	output_dir=train_cfg.get("output_dir", "checkpoints/"),
	num_train_epochs=train_cfg.get("num_train_epochs", 5),
	per_device_train_batch_size=batch_size,
	per_device_eval_batch_size=train_cfg.get("per_device_eval_batch_size", 8) if device == "cuda" else 2,
	gradient_accumulation_steps=train_cfg.get("gradient_accumulation_steps", 8),
	learning_rate=train_cfg.get("learning_rate", 3e-4),
	lr_scheduler_type=train_cfg.get("lr_scheduler_type", "cosine"),
	warmup_ratio=train_cfg.get("warmup_ratio", 0.05),
	weight_decay=train_cfg.get("weight_decay", 0.01),
	fp16=use_fp16,
	bf16=use_bf16,
	eval_strategy=train_cfg.get("evaluation_strategy", "steps"),
	eval_steps=train_cfg.get("eval_steps", 100),
	save_strategy=train_cfg.get("save_strategy", "steps"),
	save_steps=train_cfg.get("save_steps", 100),
	save_total_limit=train_cfg.get("save_total_limit", 3),
	load_best_model_at_end=False, # Handled manually below (PEFT adapters break Trainer's loader)
	metric_for_best_model=train_cfg.get("metric_for_best_model", "eval_loss"),
	greater_is_better=train_cfg.get("greater_is_better", False),
	logging_dir=train_cfg.get("logging_dir", "logs/"),
	logging_steps=train_cfg.get("logging_steps", 25),
	report_to=report_to,
	dataloader_num_workers=num_workers,
	seed=train_cfg.get("seed", 42),
	remove_unused_columns=False, # We have custom columns (style_vector, etc.)
	gradient_checkpointing=use_grad_ckpt,
	)

	# Step 9: Create trainer
	logger.info("Step 9: Creating trainer...")
	trainer = CorrectionTrainer(
	loss_fn=loss_fn,
	fingerprinter=fingerprinter,
	tokenizer=tokenizer,
	model=model,
	args=training_args,
	train_dataset=train_dataset,
	eval_dataset=val_dataset,
	callbacks=[
	StyleMetricsCallback(),
	EarlyStoppingOnStyleDrift(min_style_similarity=0.75),
	],
	)

	# Step 10: Train
	logger.info("Step 10: Starting training...")
	logger.info(
	f"Config summary: model={model_key} \| batch={batch_size} \| "
	f"accum={training_args.gradient_accumulation_steps} \| "
	f"effective_batch={batch_size * training_args.gradient_accumulation_steps} \| "
	f"epochs={training_args.num_train_epochs} \| "
	f"precision={'bf16' if use_bf16 else 'fp16' if use_fp16 else 'fp32'} \| "
	f"grad_ckpt={use_grad_ckpt} \| device={device}"
	)
	trainer.train()

	# Step 11: Save best model (manual PEFT-aware loading)
	logger.info("Step 11: Saving best model...")
	output_dir = train_cfg.get("output_dir", "checkpoints/")
	save_path = os.path.join(output_dir, "best_model")

	# Find best checkpoint from trainer state
	best_ckpt = None
	state_path = os.path.join(output_dir, "trainer_state.json")
	# Check each checkpoint for trainer_state.json
	import glob
	for ckpt_dir in sorted(glob.glob(os.path.join(output_dir, "checkpoint-*"))):
	ts = os.path.join(ckpt_dir, "trainer_state.json")
	if os.path.exists(ts):
	import json as json_mod
	with open(ts) as f:
	state = json_mod.load(f)
	best_path = state.get("best_model_checkpoint")
	if best_path:
	best_ckpt = best_path

	if best_ckpt and os.path.isdir(best_ckpt):
	logger.info(f"Loading best checkpoint from {best_ckpt}")
	from peft import PeftModel
	# Reload the best adapter weights
	best_adapter = os.path.join(best_ckpt, "adapter_model.safetensors")
	if os.path.exists(best_adapter):
	model.load_adapter(best_ckpt, adapter_name="default")
	logger.info(f"Loaded best adapter from {best_ckpt}")
	else:
	logger.warning(f"No adapter found at {best_ckpt}, saving current model")
	else:
	logger.info("No best checkpoint found, saving final model state")

	trainer.save_model(save_path)
	tokenizer.save_pretrained(save_path)
	logger.info(f"Model saved to {save_path}")

	if HAS_WANDB and wandb.run is not None:
	wandb.finish()

	logger.info("✓ Training complete!")


	if __name__ == "__main__":
	train()