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.

43.1 KB
Raw Blame
"""GRPO (Group Relative Policy Optimization) trainer — wraps trl.GRPOTrainer."""

from __future__ import annotations

import logging
import time
from functools import lru_cache
from pathlib import Path
from typing import Any, Optional

from rich.console import Console

from soup_cli.config.schema import SoupConfig, TrainingConfig
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 import final_answer
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()
logger = logging.getLogger(__name__)


def make_grpo_trainer_variant(base_cls: type, variant: str) -> type:
    """v0.53.11 #123 — build a ``_GRPOTrainerVariant`` subclass.

    Returns a subclass of ``trl.GRPOTrainer`` whose ``compute_loss`` routes
    through :func:`soup_cli.utils.grpo_variants.apply_variant_loss`. Cached
    so multiple instantiations with the same (base, variant) share one class.

    Pure factory — no torch / trl imports at module load time. Variant
    name is normalised via ``validate_grpo_variant`` BEFORE the cache
    boundary so ``"GSPO"`` and ``"gspo"`` share one class (security review
    MEDIUM fix).
    """
    from soup_cli.utils.grpo_variants import validate_grpo_variant

    variant = validate_grpo_variant(variant)
    return _make_grpo_trainer_variant_cached(base_cls, variant)


@lru_cache(maxsize=8)
def _make_grpo_trainer_variant_cached(base_cls: type, variant: str) -> type:
    """Cached factory body — keyed on already-normalised variant."""
    from soup_cli.utils.grpo_variants import apply_variant_loss, variant_kl_metric

    class _GRPOTrainerVariant(base_cls):  # type: ignore[misc, valid-type]
        """GRPOTrainer subclass that routes compute_loss through Soup's variants."""

        _soup_grpo_variant: str = variant
        # v0.71.11 #159 — one-shot WARNING flag so a silent fallback to the
        # stock TRL loss surfaces exactly once (not on every step).
        _soup_fallback_warned: bool = False

        def _warn_fallback(self, reason: str) -> None:
            """Emit a one-shot WARNING when the variant kernel falls back.

            v0.71.11 #159 — when a TRL internal rename or a kernel error
            makes ``compute_loss`` delegate to the stock GRPO loss, the
            operator's selected variant silently stops applying. Warn once
            so the run isn't quietly training the wrong objective.
            """
            if self._soup_fallback_warned:
                return
            self._soup_fallback_warned = True
            logger.warning(
                "GRPO variant %r compute_loss fell back to the stock TRL "
                "loss (%s); the selected objective is NOT being applied. "
                "This usually means a TRL version renamed the per-token "
                "log-prob inputs.",
                self._soup_grpo_variant,
                reason,
            )

        def _compute_variant_loss(self, model, inputs):
            """Evaluate the requested GRPO variant loss on the batch.

            Reuses TRL's ``_get_per_token_logps_and_entropies`` to obtain
            per-token log probabilities in a single forward pass without
            duplicating forward computation or allocating extra VRAM.
            """
            logp_new = _read_attr(inputs, "per_token_logps")
            prompt_ids = _read_attr(inputs, "prompt_ids")
            completion_ids = _read_attr(inputs, "completion_ids")
            prompt_mask = _read_attr(inputs, "prompt_mask")
            completion_mask = _read_attr(inputs, "completion_mask")
            advantages = _read_attr(inputs, "advantages")

            if logp_new is None and prompt_ids is not None and completion_ids is not None:
                import torch

                if prompt_mask is None:
                    prompt_mask = torch.ones_like(prompt_ids)
                if completion_mask is None:
                    completion_mask = torch.ones_like(completion_ids)
                input_ids = torch.cat([prompt_ids, completion_ids], dim=1)
                attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)
                logits_to_keep = completion_ids.size(1)

                if hasattr(self, "_get_per_token_logps_and_entropies"):
                    logp_new, _ = self._get_per_token_logps_and_entropies(
                        model,
                        input_ids,
                        attention_mask,
                        logits_to_keep,
                        compute_entropy=False,
                        pixel_values=_read_attr(inputs, "pixel_values"),
                        image_grid_thw=_read_attr(inputs, "image_grid_thw"),
                        num_images=_read_attr(inputs, "num_images"),
                        pixel_attention_mask=_read_attr(inputs, "pixel_attention_mask"),
                        image_sizes=_read_attr(inputs, "image_sizes"),
                        token_type_ids=_read_attr(inputs, "token_type_ids"),
                        mm_token_type_ids=_read_attr(inputs, "mm_token_type_ids"),
                    )

            logp_old = _read_attr(inputs, "old_per_token_logps")
            if logp_old is None and logp_new is not None:
                logp_old = logp_new.detach()

            if logp_new is None or logp_old is None or advantages is None:
                self._warn_fallback("missing per-token log-prob inputs")
                return None

            # #1232: the KL weight is trl's own ``self.beta``. trl reads that
            # attribute to decide whether to run the reference forward and to
            # weight the KL in its stock loss, so the reference is in ``inputs``
            # exactly when this loss needs it; the reward-hack controller writes
            # it at runtime. ``args.beta`` covers bases without the attribute.
            beta_attr = getattr(self, "beta", None)
            if beta_attr is None:
                beta_attr = getattr(getattr(self, "args", None), "beta", None)
            beta = float(beta_attr) if beta_attr is not None else 0.0
            delta = getattr(self, "_soup_grpo_delta", None)
            ref_logp = _read_attr(inputs, "ref_per_token_logps")
            mask = completion_mask
            tool_mask = _read_attr(inputs, "tool_mask")
            if mask is not None and tool_mask is not None:
                mask = mask * tool_mask

            try:
                variant_loss = apply_variant_loss(
                    self._soup_grpo_variant,
                    logp_new=logp_new,
                    logp_old=logp_old,
                    advantages=advantages,
                    beta=beta,
                    delta=delta,
                    completion_mask=mask,
                    reference_logp=ref_logp,
                )
            except (TypeError, ValueError) as exc:
                self._warn_fallback(f"kernel error: {exc}")
                return None

            if variant_loss is None:
                return None

            mode = "train" if getattr(getattr(self, "model", model), "training", True) else "eval"
            # #1263: this loss replaces trl's, which is where trl logs the ``kl``
            # metric, so log it here the same way (same key, same gather).
            mean_kl = variant_kl_metric(
                self._soup_grpo_variant,
                logp_new=logp_new,
                advantages=advantages,
                beta=beta,
                completion_mask=mask,
                reference_logp=ref_logp,
            )
            metrics = getattr(self, "_metrics", None)
            if mean_kl is not None and metrics is not None:
                metrics[mode]["kl"].append(self.accelerator.gather(mean_kl).nanmean().item())
            normalizer = (
                getattr(self, "current_gradient_accumulation_steps", 1.0)
                if mode == "train"
                else 1.0
            )
            return variant_loss / normalizer

        def _compute_loss(self, model, inputs):
            loss = self._compute_variant_loss(model, inputs)
            if loss is not None:
                return loss
            if hasattr(super(), "_compute_loss"):
                return super()._compute_loss(model, inputs)
            return super().compute_loss(model, inputs)

        def compute_loss(self, model, inputs, return_outputs=False, **kwargs):
            if hasattr(super(), "_compute_loss"):
                return super().compute_loss(model, inputs, return_outputs=return_outputs, **kwargs)
            loss = self._compute_variant_loss(model, inputs)
            if loss is not None:
                if return_outputs:
                    return loss, None
                return loss
            return super().compute_loss(model, inputs, return_outputs=return_outputs, **kwargs)

    _GRPOTrainerVariant.__name__ = f"_GRPOTrainerVariant_{variant}"
    return _GRPOTrainerVariant


def _read_attr(obj: Any, name: str) -> Any:
    """Read ``name`` from a mapping OR object — TRL inputs vary in shape."""
    if obj is None:
        return None
    if hasattr(obj, "get"):
        return obj.get(name)
    return getattr(obj, name, None)


def _select_reward_fn(
    tcfg: TrainingConfig, device: str, trust_remote_code: bool
) -> "Any":  # Callable | list[Callable] (a single reward or a comma-split ensemble)
    """Choose the GRPO reward function (v0.71.30).

    When ``tcfg.prm_reward`` is set, a trained Soup PRM scores each completion's
    steps and REPLACES the configured ``reward_fn`` (process-supervision). The
    returned callable rides the existing shaping + ``wrap_reward_funcs`` seam in
    :meth:`GRPOTrainerWrapper.setup` unchanged, so the v0.71.26 reward-hack
    mitigation controller observes the PRM reward for free.
    """
    if tcfg.prm_reward is not None:
        from soup_cli.utils.prm_reward import build_prm_reward_fn

        return build_prm_reward_fn(tcfg, device, trust_remote_code)
    # v0.71.40 #311 — ``reward_fn`` may be comma-separated ("accuracy,format").
    # Return a single callable for one reward (back-compat) or a list for several
    # (TRL's reward_funcs=[...] + the rm_ensemble detector both accept the list).
    # #2047 — ``judge_rank`` is built from this config, like the PRM reward above.
    from soup_cli.trainer.grpo_judge import JUDGE_RANK_REWARD, build_judge_rank_reward
    from soup_cli.trainer.rewards import load_reward_fns

    fns = load_reward_fns(
        tcfg.reward_fn,
        verifiable_domain=tcfg.verifiable_domain,
        builders={JUDGE_RANK_REWARD: lambda: build_judge_rank_reward(tcfg)},
    )
    return fns[0] if len(fns) == 1 else fns


class GRPOTrainerWrapper:
    """High-level wrapper for GRPO training from SoupConfig.

    GRPO generates multiple completions per prompt, scores them with a reward
    function, and optimizes using group-relative advantages. This is the approach
    used by DeepSeek-R1 for reasoning model training.

    Data format: same as SFT (messages with prompt/response) or DPO-style prompts.
    The reward_fn in config determines how completions are scored.
    """

    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

    def _build_precision_kwargs(self) -> dict[str, bool]:
        """Resolve fp16/bf16 kwargs for GRPOConfig (v0.53.3 #128).

        Priority:
        - MPS → BF16 only when the live runtime accepts a BF16 allocation;
          otherwise FP32. This is the same hardware-probed policy used by
          Soup's other validated text trainers.
        - Other non-CUDA devices (CPU / XPU) → no mixed precision.
        - ``grpo_fp16=True`` (CUDA) → ``fp16=True, bf16=False`` (unsloth
          parity).
        - Default CUDA → bf16 when the card supports it, fp16 when it does
          not. This branch used to be a flat ``bf16=True``, which transformers
          refuses on a pre-Ampere card (T4 / P100 / V100 / GTX 16xx) — see
          #387; ``grpo_fp16`` was the only way to run GRPO there and nothing
          said so.

        ``auto_mixed_precision`` is refused on ``task='grpo'`` at schema
        load (#1618 — this wrapper never reads the field, so it would train
        at whatever precision these lines pick); ``grpo_fp16`` is the only
        way to choose the dtype here, and the grpo_fp16 + auto_mixed_precision
        combo is rejected by ``_validate_grpo_fp16_amp_exclusive``.
        """
        device_name = str(self.device).lower()
        if device_name.startswith("mps"):
            bf16, fp16 = bf16_fp16_flags(
                self.device,
                allow_mps_bf16=True,
            )
            return {"fp16": fp16, "bf16": bf16}
        if not device_name.startswith("cuda"):
            return {"fp16": False, "bf16": False}
        # grpo_fp16 is a Pydantic field with default=False; direct attribute
        # access (no getattr fallback) so a typo would fail loudly.
        if self.config.training.grpo_fp16:
            return {"fp16": True, "bf16": False}
        bf16, fp16 = bf16_fp16_flags(self.device)
        return {"fp16": fp16, "bf16": bf16}

    def setup(self, dataset: dict):
        """Load model, tokenizer, apply LoRA, create GRPO trainer."""
        from datasets import Dataset
        from trl import GRPOConfig, GRPOTrainer

        # v0.53.11 #123 — variant subclass override
        variant = self.config.training.grpo_variant
        if variant is not None and variant != "standard":
            GRPOTrainer = make_grpo_trainer_variant(GRPOTrainer, variant)  # noqa: N806

        # Enable Rich progress bar for HuggingFace downloads
        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"

        # --- Load reward function ---
        # v0.71.30 — when tcfg.prm_reward is set, a trained Soup PRM replaces
        # the configured reward (process-supervision); otherwise load reward_fn.
        reward_fn = _select_reward_fn(tcfg, self.device, self._trust_remote_code)
        # #2047 — the judge_rank reward's counters, read by its callback below.
        from soup_cli.trainer.grpo_judge import judge_rank_state

        self._judge_rank_state = judge_rank_state(reward_fn)

        # v0.71.11 #235/#240 — when the reward-hack or echo-trap detector is
        # enabled, wrap the reward function(s) with a capture shim so the
        # callbacks can observe the step's rewards + completions. The buffer
        # is created here (before GRPOTrainer construction) and handed to
        # the callbacks after the trainer is built.
        from soup_cli.utils.peft_wiring import rl_callbacks_need_buffer

        self._rl_buffer = None
        if rl_callbacks_need_buffer(tcfg):
            # v0.71.26 Stage 3 — apply the reward-shaping shim BEFORE the buffer
            # capture so the controller observes (and GRPO optimises) the shaped
            # reward. No-op when reward_hack_reward_shaping is off.
            from soup_cli.utils.reward_hack_control import apply_reward_shaping
            from soup_cli.utils.rl_signal_buffer import (
                RLSignalBuffer,
                wrap_reward_funcs,
            )

            reward_fn = apply_reward_shaping(reward_fn, tcfg)
            self._rl_buffer = RLSignalBuffer()

        from soup_cli.trainer.rewards import validate_reward_funcs

        reward_fn = validate_reward_funcs(reward_fn)
        if self._rl_buffer is not None:
            reward_fn = wrap_reward_funcs(reward_fn, self._rl_buffer)

        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
        )

        # Ensure tokenizer has a chat template — trl's GRPOTrainer calls
        # apply_chat_template() when it detects conversational prompts (message
        # lists) and will raise ValueError if the template is missing.
        if not getattr(self.tokenizer, "chat_template", None):
            self.tokenizer.chat_template = (
                "{% for msg in messages %}{{ msg['content'] }}\n{% endfor %}"
            )

        trainable, total = self.model.get_nb_trainable_parameters()
        pct = 100 * trainable / total
        console.print(
            f"[green]LoRA applied:[/] {trainable:,} trainable / {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,
            )
            # GRPO generates N completions per prompt → more memory
            batch_size = max(1, batch_size // tcfg.num_generations)
            console.print(f"[green]Auto batch size (GRPO):[/] {batch_size}")

        # Ensure batch_size >= num_generations (trl requires
        # generation_batch_size to be divisible by num_generations)
        num_gen = tcfg.num_generations
        if batch_size < num_gen:
            batch_size = num_gen

        # --- Dataset ---
        # GRPO expects prompts — extract from messages or use prompt field
        train_data = _prepare_grpo_dataset(dataset["train"])
        _validate_grpo_reward_metadata(train_data, tcfg, split="train")

        # v0.71.21 #125 — multi-turn agent rollout backend. The backend
        # receives the dataset prompts as seeds; its rows REPLACE the
        # prompt dataset (the env is the data source).
        if tcfg.rollout_backend is not None:
            from soup_cli.utils.agent_rollout import launch_rollout

            rollout_result = launch_rollout(
                tcfg.rollout_backend,
                prompts=[row["prompt"] for row in train_data],
                rollout_func=tcfg.rollout_func,
                model=self.model,
                tokenizer=self.tokenizer,
                reward_fn=reward_fn,
            )
            train_data = _prepare_grpo_dataset([dict(row) for row in rollout_result.rows])
            _validate_grpo_reward_metadata(train_data, tcfg, split="rollout")
            console.print(
                f"[green]Rollout backend '{tcfg.rollout_backend}':[/] "
                f"{len(train_data)} prompts collected "
                "(replacing dataset prompts)"
            )

        train_ds = Dataset.from_list(train_data)
        eval_ds = None
        if "val" in dataset and dataset["val"]:
            eval_data = _prepare_grpo_dataset(dataset["val"])
            _validate_grpo_reward_metadata(eval_data, tcfg, split="validation")
            eval_ds = Dataset.from_list(eval_data)

        # --- 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)

        # --- Warn if running on CPU (trl GRPO has known CPU issues) ---
        if self.device == "cpu":
            console.print(
                "[yellow]Warning: GRPO on CPU is experimental. "
                "trl's GRPOTrainer may produce empty generations on CPU, "
                "causing tensor size errors. A CUDA GPU is recommended.[/]"
            )

        # #1223: TRL evaluates num_generations completions per held-out prompt
        # and refuses an eval batch that does not hold whole groups (a batch of
        # 3 with 2 generations and gradient accumulation 2 trains, then raised
        # at setup). Evaluate at the largest multiple of num_generations that
        # fits in the train batch -- the bump above guarantees at least one.
        eval_kwargs = training_eval_kwargs(
            cfg, eval_ds, batch_size=batch_size - batch_size % num_gen
        )
        if (
            eval_kwargs["eval_strategy"] != "no"
            and eval_kwargs["per_device_eval_batch_size"] != batch_size
        ):
            console.print(
                "[yellow]Note:[/] grpo evaluates at batch "
                f"{eval_kwargs['per_device_eval_batch_size']}, not the train batch "
                f"{batch_size}: TRL's evaluation needs whole groups of "
                f"num_generations={num_gen} completions."
            )

        # --- GRPO config ---
        grpo_kwargs = {
            "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,
            **self._build_precision_kwargs(),
            "report_to": self.report_to,
            "remove_unused_columns": False,
            "deepspeed": self.deepspeed_config,
            **training_seed_kwargs(tcfg),
            **eval_kwargs,
            **(self.fsdp_config or {}),
            "beta": tcfg.grpo_beta,
            "num_generations": tcfg.num_generations,
            "max_completion_length": cfg.data.max_length,
        }

        # v0.71.21 #124 — vLLM sleep mode: set TRL's GRPOConfig hook when the
        # installed TRL exposes it; otherwise print a friendly advisory
        # (Soup's own vLLM engine factory honors sleep_mode=True).
        if tcfg.vllm_sleep_mode:
            import inspect as _inspect

            from soup_cli.utils.grpo_long_context import (
                maybe_enable_trl_sleep_mode,
            )

            maybe_enable_trl_sleep_mode(
                grpo_kwargs,
                _inspect.signature(GRPOConfig).parameters,
                console,
            )

        # CPU support: set use_cpu and prevent empty generations
        if self.device == "cpu":
            import inspect as _inspect

            grpo_params = _inspect.signature(GRPOConfig).parameters
            if "use_cpu" in grpo_params:
                grpo_kwargs["use_cpu"] = True
            # Workaround for trl GRPO CPU bug: model.generate() can produce
            # zero new tokens on CPU, causing tensor size mismatch errors.
            # Setting min_new_tokens=1 ensures at least one token is generated.
            if "generation_kwargs" in grpo_params:
                grpo_kwargs["generation_kwargs"] = {"min_new_tokens": 1}

        grpo_config = GRPOConfig(**grpo_kwargs)

        # Workaround: also set min_new_tokens on model's generation_config directly.
        # GRPOConfig may not forward generation_kwargs to model.generate() in all
        # trl versions, so this ensures the model always generates at least 1 token.
        if self.device == "cpu" and hasattr(self.model, "generation_config"):
            self.model.generation_config.min_new_tokens = 1

        # --- Trainer ---
        self.trainer = GRPOTrainer(
            model=self.model,
            args=grpo_config,
            train_dataset=train_ds,
            eval_dataset=eval_ds,
            reward_funcs=reward_fn,
            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)
        # v0.53.11 #123 — thread grpo_delta into the variant subclass.
        if (
            tcfg.grpo_variant is not None
            and tcfg.grpo_variant != "standard"
            and tcfg.grpo_delta is not None
        ):
            self.trainer._soup_grpo_delta = float(tcfg.grpo_delta)
        # #2047 — log judge_rank's failed / equal-score group rates under judge/
        # and stop the run when too many training groups get no scores.
        if self._judge_rank_state is not None:
            from soup_cli.trainer.grpo_judge import attach_judge_rank_callback

            attach_judge_rank_callback(self.trainer, self._judge_rank_state)
        # v0.53.11 #127 — wire the live stability callback.
        from soup_cli.utils.peft_wiring import attach_grpo_stability_callback

        attach_grpo_stability_callback(self.trainer, tcfg)

        # #342 — gradient watchdog (on_pre_optimizer_step) must be always-on.
        # attach_grpo_stability_callback only fires when stability knobs
        # are set.  ensure_grpo_stability_callback is a no-op if already
        # attached, otherwise attaches with defaults (only watchdog fires).
        from soup_cli.utils.peft_wiring import (
            ensure_grpo_stability_callback,
        )

        ensure_grpo_stability_callback(self.trainer)

        # v0.71.11 #235/#238/#240 — wire the live RL callbacks (reward-hack,
        # echo-trap, mid-epoch RL checkpoint).
        from soup_cli.utils.peft_wiring import attach_rl_callbacks

        attach_rl_callbacks(
            self.trainer,
            tcfg,
            buffer=self._rl_buffer,
            tokenizer=self.tokenizer,
            output_dir=str(output_dir),
            task="grpo",
        )

        # v0.40.6 #67 — ReLoRA merge-and-reinitialize restart 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}[/]")
        # On CPU, use device_map="cpu" to avoid meta tensors from "auto"
        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)
        # #798: 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 (Gemma4 ClippableLinear pre-LoRA;
        # 3-D fused-MoE expert dropout strip post-LoRA).
        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 — insert fake quantization ops after LoRA. The "fp8" variant
        # is FP8 training (handled by apply_v028_speed_memory), not int8 QAT.
        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, tcfg):
        """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 GRPO training and return results summary."""
        start = time.time()

        # Add callback for live display and experiment tracking
        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,
                    batch_size=self._batch_size,
                    run_id=run_id,
                    output_dir=self._output_dir,
                    # #1389: grpo's validation number is the held-out
                    # reward, not TRL's policy objective in `eval_loss`.
                    task=self.config.task,
                )
            )

        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)
        # #342 — persist trainer_state.json (including nan_skip_count
        # from the stability callback) next to adapter_config.json so
        # ``soup adapters audit`` can read the skipped-step fraction.
        self.trainer.save_state()

        # Extract metrics
        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,
        }


def _prepare_grpo_dataset(data: list[dict]) -> list[dict]:
    """Convert dataset rows to GRPO format.

    GRPO expects each row to have a 'prompt' field (list of messages or string).
    Input can be:
      - messages format: [{"role": "user", "content": "..."}, ...]
      - DPO format: {"prompt": "...", "chosen": "...", "rejected": "..."}
      - prompt field: {"prompt": "..."}

    Returns list of dicts with 'prompt' as a message list for chat models.
    """
    prepared = []
    for row in data:
        if "prompt" in row and isinstance(row["prompt"], str):
            # DPO or plain prompt format — convert to message list
            entry = {"prompt": [{"role": "user", "content": row["prompt"]}]}
            _copy_grpo_metadata(row, entry)
            prepared.append(entry)
        elif "messages" in row:
            # Messages format — the final assistant turn may be a reference
            # answer. Earlier assistant turns are part of the conversation and
            # must remain in the prompt for multi-turn GRPO (#565).
            messages = row["messages"]
            has_reference_turn = bool(
                messages and messages[-1].get("role") == "assistant"
            )
            prompt_msgs = messages[:-1] if has_reference_turn else messages
            entry = {"prompt": prompt_msgs}
            _copy_grpo_metadata(row, entry)
            if has_reference_turn:
                entry.setdefault("answer", messages[-1].get("content"))
            prepared.append(entry)
        elif "prompt" in row and isinstance(row["prompt"], list):
            # Already in message list format
            entry = {"prompt": row["prompt"]}
            _copy_grpo_metadata(row, entry)
            prepared.append(entry)
        else:
            # Fallback: treat any 'instruction' field as prompt
            instruction = row.get("instruction", row.get("input", ""))
            entry = {"prompt": [{"role": "user", "content": str(instruction)}]}
            _copy_grpo_metadata(row, entry)
            if "output" in row:
                entry.setdefault("answer", row["output"])
            prepared.append(entry)
    return prepared


def _copy_grpo_metadata(row: dict, entry: dict) -> None:
    """Copy non-prompt columns that TRL forwards to reward functions."""
    for key, value in row.items():
        if key not in {"messages", "prompt"}:
            entry[key] = value


# Rewards that compare a completion's final answer with the gold's (#1226). A gold they cannot
# read would score every completion 0.0, so the group advantage would be zero with no warning.
_GOLD_PARSING_REWARDS = frozenset({"accuracy", "verifiable/math"})


def _validate_grpo_reward_metadata(
    data: list[dict],
    tcfg: TrainingConfig,
    *,
    split: str,
) -> None:
    """Fail before generation when a built-in reward lacks required data."""
    if tcfg.prm_reward is not None:
        return

    requirements: list[tuple[str, tuple[str, ...]]] = []
    for reward_spec in (part.strip() for part in (tcfg.reward_fn or "").split(",")):
        if reward_spec == "accuracy":
            requirements.append(("accuracy", ("answer",)))
        elif reward_spec == "verifiable":
            domain = tcfg.verifiable_domain
            if domain == "math":
                requirements.append(("verifiable/math", ("answer",)))
            elif domain == "code":
                requirements.append(("verifiable/code", ("expected", "answer")))
            elif domain == "json_schema":
                requirements.append(("verifiable/json_schema", ("schema",)))

    for row_index, row in enumerate(data):
        for reward_name, alternatives in requirements:
            if any(_has_grpo_reward_metadata(row.get(field)) for field in alternatives):
                continue
            fields = " or ".join(repr(field) for field in alternatives)
            raise ValueError(
                f"GRPO {split} row {row_index} is missing or empty {fields}, required "
                f"by reward {reward_name!r}. Preserve that column in the source "
                "dataset or include an assistant response that Soup can use as "
                "'answer'."
            )

    gold_rewards = [name for name, _ in requirements if name in _GOLD_PARSING_REWARDS]
    # Seed rows that a rollout backend replaces are never scored (#565 still wants them
    # present); the rollout's own rows are validated as split "rollout".
    if gold_rewards and not (split == "train" and tcfg.rollout_backend is not None):
        _check_golds_are_readable(data, list(dict.fromkeys(gold_rewards)), split=split)


def _check_golds_are_readable(data: list[dict], reward_names: list[str], *, split: str) -> None:
    """Refuse golds with no single extractable final answer (none, a "####" line that may be a
    markdown heading, or a hedge such as "The answer is either 41 or 42."); report the ones
    compared as text."""
    unreadable: list[int] = []
    text_golds = 0
    for row_index, row in enumerate(data):
        reference = final_answer.parse_reference(str(row["answer"]))
        if reference is None:
            unreadable.append(row_index)
        elif reference.number is None:
            text_golds += 1
    if unreadable:
        raise ValueError(_unreadable_gold_message(unreadable, len(data), reward_names, split))
    if text_golds:
        console.print(
            f"[dim]GRPO {split}: {text_golds} of {len(data)} golds are non-numeric and are "
            "compared as normalised text.[/]"
        )


def _unreadable_gold_message(
    rows: list[int], total: int, reward_names: list[str], split: str
) -> str:
    names = " and ".join(repr(name) for name in reward_names)
    label = "reward" if len(reward_names) == 1 else "rewards"
    more = (
        f" {len(rows) - 1} more of the {total} rows have the same problem." if len(rows) > 1 else ""
    )
    return (
        f"GRPO {split} row {rows[0]} 'answer' states no single final answer that {label} "
        f"{names} can compare against: it states none (a '####' line that is not a number and "
        "has more text after it reads as a markdown heading), or its answer phrase names more "
        "than one number (a hedge such as 'The answer is either 41 or 42.'), so every "
        f"completion would score 0.0.{more} The field comes from an 'answer' column, an Alpaca "
        "'output' or the final assistant turn; rows count from 0 after the train/validation "
        "split. Put one answer after '####' on the last line, inside \\boxed{}, or after 'The "
        "answer is' or 'Answer:', make the field the bare answer on one line, or drop the row."
    )


def _has_grpo_reward_metadata(value: object) -> bool:
    """Return whether a reward metadata value is usable as a gold target."""
    if value is None:
        return False
    if isinstance(value, str):
        return bool(value.strip())
    return True