Skip to content

armgit / Soup

Public Imported from GitHub
0

Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.

16.4 KB
Raw Blame
"""IPO (Identity Preference Optimization) trainer — wraps trl.DPOTrainer with loss_type='ipo'."""

import time
from pathlib import Path
from typing import Optional

from rich.console import Console

from soup_cli.config.schema import SoupConfig
from soup_cli.data.chat_templates import apply_chat_template_override
from soup_cli.trainer.loss_summary import summarize_training_loss
from soup_cli.utils.eval_schedule import training_eval_kwargs
from soup_cli.utils.gpu import (
    bf16_fp16_flags,
    estimate_batch_size,
    model_size_from_name,
    resolve_device_map,
    resolve_frozen_base_load_dtype,
)
from soup_cli.utils.mixed_precision import (
    align_trainable_dtype_for_fp16,
    keep_trainable_dtype_on_resume,
    reload_root_adapter_on_resume,
)
from soup_cli.utils.run_provenance import base_revision_kwargs
from soup_cli.utils.seeding import apply_training_seed, training_seed_kwargs

console = Console()


class IPOTrainerWrapper:
    """High-level wrapper for IPO training from SoupConfig.

    IPO is a theoretically grounded variant of DPO that uses a squared
    hinge loss instead of the logistic loss. Implemented via trl.DPOTrainer
    with loss_type='ipo'.

    Data fields (same as DPO):
    - prompt: the input prompt
    - chosen: the preferred response
    - rejected: the less preferred response
    """

    def __init__(
        self,
        config: SoupConfig,
        device: str = "cuda",
        report_to: str = "none",
        deepspeed_config: Optional[str] = None,
        fsdp_config: Optional[dict] = None,
        trust_remote_code: bool = False,
    ):
        self.config = config
        self.device = device
        self.report_to = report_to
        self.deepspeed_config = deepspeed_config
        self.fsdp_config = fsdp_config
        self.trust_remote_code = trust_remote_code
        from soup_cli.utils.trust_remote import (
            model_requires_trust_remote_code,
            resolve_trust_remote_code,
        )

        requires = model_requires_trust_remote_code(config.base) or False
        self._trust_remote_code = resolve_trust_remote_code(
            config.base,
            requested=trust_remote_code,
            console=console,
            requires_remote_code=requires,
        )
        self.model = None
        self.tokenizer = None
        self.trainer = None
        self._output_dir = None

    def setup(self, dataset: dict) -> None:
        """Load model, tokenizer, apply LoRA, create IPO trainer (DPO with loss_type='ipo')."""
        from datasets import Dataset
        from trl import DPOConfig, DPOTrainer

        from soup_cli.trainer._trl_compat import prompt_length_kwargs
        from soup_cli.trainer.sft import _enable_hf_transfer_progress

        _enable_hf_transfer_progress()

        cfg = self.config
        tcfg = cfg.training

        # #353: seed before the model and any adapter are built.
        apply_training_seed(tcfg)

        use_unsloth = cfg.backend == "unsloth"

        if use_unsloth:
            self._setup_unsloth(cfg, tcfg)
        else:
            self._setup_transformers(cfg, tcfg)

        apply_chat_template_override(
            self.tokenizer, cfg.data.chat_template, console=console
        )

        trainable, total = self.model.get_nb_trainable_parameters()
        pct = 100 * trainable / total
        console.print(
            f"[green]LoRA applied:[/] {trainable:,} trainable"
            f" / {total:,} total ({pct:.2f}%)"
        )

        # --- Batch size ---
        batch_size = tcfg.batch_size
        if batch_size == "auto":
            from soup_cli.utils.gpu import get_gpu_info

            gpu_info = get_gpu_info()
            model_size = model_size_from_name(cfg.base)
            batch_size = estimate_batch_size(
                model_params_b=model_size,
                seq_length=cfg.data.max_length,
                gpu_memory_bytes=gpu_info["memory_total_bytes"],
                quantization=tcfg.quantization,
                lora_r=tcfg.lora.r,
            )
            # IPO processes pairs → roughly 2x memory per sample
            batch_size = max(1, batch_size // 2)
            console.print(f"[green]Auto batch size (IPO):[/] {batch_size}")

        # --- Dataset ---
        train_ds = Dataset.from_list(dataset["train"])
        eval_ds = None
        if "val" in dataset and dataset["val"]:
            eval_ds = Dataset.from_list(dataset["val"])

        # --- Output dir ---
        output_dir = Path(cfg.output)
        if cfg.experiment_name:
            output_dir = output_dir / cfg.experiment_name
        output_dir.mkdir(parents=True, exist_ok=True)

        # --- Calculate warmup steps from ratio (#1431) ---
        from soup_cli.utils.layer_stream import should_enable_hf_gradient_checkpointing
        from soup_cli.utils.warmup import resolve_trainer_warmup_steps

        warmup_steps = resolve_trainer_warmup_steps(tcfg.warmup_ratio)

        # --- IPO config (DPO with loss_type='ipo') ---
        # In IPO, the beta parameter acts as tau (regularization strength)
        _bf16, _fp16 = bf16_fp16_flags(self.device)
        dpo_config = DPOConfig(
            output_dir=str(output_dir),
            num_train_epochs=tcfg.epochs,
            per_device_train_batch_size=batch_size,
            gradient_accumulation_steps=tcfg.gradient_accumulation_steps,
            learning_rate=tcfg.lr,
            warmup_steps=warmup_steps,
            gradient_checkpointing=should_enable_hf_gradient_checkpointing(
                tcfg.gradient_checkpointing, stream_layers=tcfg.stream_layers
            ),
            weight_decay=tcfg.weight_decay,
            max_grad_norm=tcfg.max_grad_norm,
            optim=tcfg.optimizer,
            lr_scheduler_type=tcfg.scheduler,
            logging_steps=tcfg.logging_steps,
            save_steps=tcfg.save_steps,
            save_total_limit=3,
            bf16=_bf16,
            fp16=_fp16,
            report_to=self.report_to,
            remove_unused_columns=False,
            deepspeed=self.deepspeed_config,
            **training_seed_kwargs(tcfg),
            **training_eval_kwargs(cfg, eval_ds, batch_size=batch_size),
            **(self.fsdp_config or {}),
            loss_type="ipo",
            beta=tcfg.ipo_tau,
            max_length=cfg.data.max_length,
            # #326 — see dpo.py; IPO rides the same DPOConfig.
            **prompt_length_kwargs(DPOConfig, cfg.data.max_length // 2),
            **({"neftune_noise_alpha": tcfg.neftune_alpha}
               if tcfg.neftune_alpha is not None else {}),
        )

        # --- Trainer ---
        self.trainer = DPOTrainer(
            model=self.model,
            args=dpo_config,
            train_dataset=train_ds,
            eval_dataset=eval_ds,
            processing_class=self.tokenizer,
        )

        # #359 - the same exposure #336 fixed in sft.py: with LoRA the
        # no-decay optimizer group is empty, DeepSpeed drops it, and the LR
        # scheduler keeps two base_lrs until torch's strict zip raises at the
        # first step. The guard prunes inside create_optimizer, i.e. before
        # the scheduler is built. No-op for full fine-tuning, and only under
        # DeepSpeed so the ordinary path keeps its own optimizer.
        if self.deepspeed_config:
            from soup_cli.utils.deepspeed import attach_empty_param_group_guard

            attach_empty_param_group_guard(self.trainer)

        # v0.40.6 #67 — ReLoRA callback.
        from soup_cli.utils.peft_wiring import (
            attach_curriculum_callback,
            attach_loraplus_optimizer,
            attach_plugin_callback,
            attach_relora_callback,
        )
        # LoRA+ optimizer (#724/#745) — build and attach now that the trainer exists.
        attach_loraplus_optimizer(self.trainer, tcfg)
        attach_relora_callback(self.trainer, tcfg)
        # v0.53.5 #114/#115 — dynamic curriculum live callback.
        attach_curriculum_callback(self.trainer, tcfg, str(output_dir), console)
        # v0.53.6 #101 — Soup plugin TrainerCallback.
        attach_plugin_callback(self.trainer, console)

        self._output_dir = str(output_dir)
        self._batch_size = batch_size

    def _setup_transformers(self, cfg: SoupConfig, tcfg) -> None:
        """Load model via standard transformers + peft pipeline."""
        from peft import TaskType, get_peft_model, prepare_model_for_kbit_training
        from transformers import AutoModelForCausalLM, AutoTokenizer

        console.print(f"[dim]Loading tokenizer: {cfg.base}[/]")
        self.tokenizer = AutoTokenizer.from_pretrained(
            cfg.base, trust_remote_code=self._trust_remote_code,
            **base_revision_kwargs(cfg.base),
        )
        if self.tokenizer.pad_token is None:
            self.tokenizer.pad_token = self.tokenizer.eos_token

        # Quantization (v0.38.0 Quant Menu — see soup_cli.utils.quant_menu)
        from soup_cli.utils.quant_menu import build_quantization_config_for_loader

        quant_config_obj = build_quantization_config_for_loader(
            tcfg=tcfg, base=cfg.base, console=console,
        )

        console.print(f"[dim]Loading model: {cfg.base}[/]")
        dev_map = resolve_device_map(self.device)
        model_kwargs = {
            "trust_remote_code": self._trust_remote_code, "device_map": dev_map,
            "torch_dtype": resolve_frozen_base_load_dtype(self.device),
        }
        if quant_config_obj is not None:
            model_kwargs["quantization_config"] = quant_config_obj

        self.model = AutoModelForCausalLM.from_pretrained(
            cfg.base, **model_kwargs,
            **base_revision_kwargs(cfg.base),
        )
        from soup_cli.utils.data_pipeline import apply_vocab_expansion

        apply_vocab_expansion(
            self.tokenizer,
            self.model,
            cfg.data,
        )
        if tcfg.quantization in ("4bit", "8bit"):
            from soup_cli.utils.layer_stream import should_enable_hf_gradient_checkpointing

            self.model = prepare_model_for_kbit_training(
                self.model,
                use_gradient_checkpointing=should_enable_hf_gradient_checkpointing(
                    tcfg.gradient_checkpointing, stream_layers=tcfg.stream_layers
                ),
            )

        from soup_cli.utils.peft_wiring import (
            build_lora_config,
            resolve_lora_target_modules,
        )

        target_modules = resolve_lora_target_modules(self.model, tcfg.lora.target_modules, console)
        # #1099: moe_lora picks the expert-FFN targets. Without this the flag
        # was accepted and ignored here, and on a fused-expert MoE the auto
        # resolution leaves peft with nothing to attach.
        from soup_cli.utils.moe import resolve_moe_lora_targets

        target_modules = resolve_moe_lora_targets(
            self.model, tcfg, target_modules, console
        )

        lora_config = build_lora_config(
            tcfg.lora,
            target_modules=target_modules,
            task_type=TaskType.CAUSAL_LM,
        )
        # v0.40.6 #67 — surgical PEFT patches.
        from soup_cli.utils.peft_wiring import (
            apply_post_lora_patches,
            apply_pre_lora_patches,
        )
        apply_pre_lora_patches(self.model, cfg.base)
        self.model = get_peft_model(self.model, lora_config)
        apply_post_lora_patches(self.model)

        # QAT — int8 only; "fp8" handled by apply_v028_speed_memory below.
        if tcfg.quantization_aware and tcfg.quantization_aware != "fp8":
            from soup_cli.utils.qat import prepare_model_for_qat

            self.model = prepare_model_for_qat(self.model)

        # v0.35.0 #60 — multi-trainer wiring of v0.28.0 speed/memory features.
        from soup_cli.utils.v028_features import apply_v028_speed_memory
        apply_v028_speed_memory(
            model=self.model, tcfg=tcfg, base_model=cfg.base,
            console=console, device=self.device, backend=cfg.backend,
        )

    def _setup_unsloth(self, cfg: SoupConfig, tcfg) -> None:
        """Load model via unsloth FastLanguageModel (2-5x faster)."""
        from soup_cli.utils import unsloth
        from soup_cli.utils.unsloth import load_model_and_tokenizer
        # #797: resolve training.gradient_checkpointing for the Unsloth path
        # (streaming layers already checkpoint every layer, so it stays off).
        use_gradient_checkpointing = unsloth.resolve_unsloth_gradient_checkpointing(
            tcfg.gradient_checkpointing, stream_layers=getattr(tcfg, "stream_layers", False)
        )

        console.print(f"[dim]Loading model via [bold]unsloth[/]: {cfg.base}[/]")
        self.model, self.tokenizer = load_model_and_tokenizer(
            model_name=cfg.base,
            max_seq_length=cfg.data.max_length,
            quantization=tcfg.quantization,
            lora_r=tcfg.lora.r,
            lora_alpha=tcfg.lora.alpha,
            lora_dropout=tcfg.lora.dropout,
            target_modules=tcfg.lora.target_modules,
            use_rslora=tcfg.lora.use_rslora,
            use_gradient_checkpointing=use_gradient_checkpointing,
        )
        if self.tokenizer.pad_token is None:
            self.tokenizer.pad_token = self.tokenizer.eos_token

    def train(
        self,
        display: Optional[object] = None,
        tracker: Optional[object] = None,
        run_id: str = "",
        resume_from_checkpoint: Optional[str] = None,
    ) -> dict:
        """Run IPO training and return results summary."""
        if self.trainer is None:
            raise RuntimeError(
                "IPOTrainerWrapper.train() called before setup(). "
                "Call setup(dataset) first."
            )
        start = time.time()

        if display:
            from soup_cli.monitoring.callback import build_soup_trainer_callback

            self.trainer.add_callback(
               build_soup_trainer_callback(
                    display,
                    config=self.config,
                    tracker=tracker,
                    run_id=run_id,
                    batch_size=self._batch_size,
                    output_dir=self._output_dir,
                )
            )

        # v0.40.0 Part C — DPO variants (β-schedule + ref-model regen).
        # total_steps=0 is the lazy-resolve sentinel; BetaScheduleCallback's
        # on_train_begin reads state.max_steps once HF Trainer has populated it.
        from soup_cli.utils.dpo_variants import build_dpo_variant_callbacks

        tcfg = self.config.training
        variant_cbs = build_dpo_variant_callbacks(
            # IPO's regularization strength is `ipo_tau` (passed as trl's beta,
            # see setup()); the β-schedule must anneal from that, not the DPO
            # default `dpo_beta`, which would overwrite the user's τ at step 0.
            beta_start=tcfg.ipo_tau,
            beta_end=tcfg.dpo_beta_end,
            schedule=tcfg.dpo_beta_schedule,
            total_steps=0,
            ref_regen_epochs=tcfg.dpo_ref_regen_epochs,
        )
        for cb in variant_cbs:
            cb.attach(self.trainer)
            self.trainer.add_callback(cb)

        from soup_cli.utils.v028_features import activation_offloading_context

        with activation_offloading_context(
            self.config.training, self._output_dir,
        ):
            align_trainable_dtype_for_fp16(
                self.trainer.model,
                fp16=getattr(self.trainer.args, "fp16", False),
                bf16=getattr(self.trainer.args, "bf16", False),
            )
            if resume_from_checkpoint is not None:
                reload_root_adapter_on_resume(self.trainer)
                keep_trainable_dtype_on_resume(self.trainer)
            self.trainer.train(resume_from_checkpoint=resume_from_checkpoint)
        duration = time.time() - start

        if self.config.training.relora_steps is None:
            self.trainer.save_model(self._output_dir)
        else:
            from soup_cli.utils.peft_wiring import save_model_with_relora

            save_model_with_relora(
                self.trainer, self._output_dir, self.config.training.relora_steps
            )
        self.tokenizer.save_pretrained(self._output_dir)

        logs = self.trainer.state.log_history
        loss_summary = summarize_training_loss(logs)

        hours = int(duration // 3600)
        minutes = int((duration % 3600) // 60)
        duration_str = f"{hours}h {minutes}m" if hours > 0 else f"{minutes}m"

        return {
            **loss_summary,
            "duration": duration_str,
            "duration_secs": duration,
            "output_dir": self._output_dir,
            "total_steps": self.trainer.state.global_step,
        }