feat: add dpo liger

src/axolotl/core/trainer_builder.py

@@ -14,17 +14,22 @@ import os
 import sys
 from abc import abstractmethod
 from collections import defaultdict
+from contextlib import nullcontext
 from dataclasses import dataclass, field
 from functools import wraps
 from pathlib import Path
 from typing import Any, Dict, List, Literal, Optional, Type, Union
 
 import torch
+import torch.nn.functional as F
 import transformers
 from datasets import Dataset
+from liger_kernel.chunked_loss.fused_linear_preference import (
+    LigerFusedLinearPreferenceBase,
+)
 from packaging import version
 from peft.optimizers import create_loraplus_optimizer
-from torch import nn
+from torch import amp, nn
 from torch.optim.lr_scheduler import OneCycleLR
 from torch.utils.data import BatchSampler, DataLoader, RandomSampler, SequentialSampler
 from transformers import (
@@ -1077,6 +1082,15 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
         self.dataset_tags = dataset_tags
         self.optimizer = None
 
+        from liger_kernel.chunked_loss import LigerFusedLinearDPOLoss
+
+        self.liger_loss = LigerFusedLinearDPOLoss(
+            ignore_index=self.label_pad_token_id,
+            beta=self.beta,
+            compute_nll_loss=True,  # not same as rpo_alpha hasattr(self.args, "rpo_alpha") and self.args.rpo_alpha is not None,
+            use_ref_model=not self.reference_free,
+        )
+
     def create_optimizer(self):
         if self.args.loraplus_lr_ratio is None:
             return super().create_optimizer()
@@ -1180,6 +1194,309 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
         # transformers<=4.46
         return super(DPOTrainer, self).log(logs)  # pylint: disable=bad-super-call
 
+    def get_batch_loss_metrics(
+        self,
+        model,
+        batch: dict[str, Union[list, torch.LongTensor]],
+        train_eval: Literal["train", "eval"] = "train",
+    ):
+        """Compute the DPO loss and other metrics using Liger kernel."""
+        # return super().get_batch_loss_metrics(model, batch, train_eval)
+        if not self.liger_loss:
+            raise ValueError("Liger loss not initialized")
+
+        metrics = {}
+
+        model_output = self.concatenated_forward(model, batch)
+
+        # Get the lm_head weights and bias
+        lin_weight = model.lm_head.weight
+        lin_bias = getattr(model.lm_head, "bias", None)
+
+        hidden_states = model_output["hidden_states"]
+        labels = model_output["labels"]
+
+        if not self.reference_free:
+            # Adapted from DPO's compute_ref_log_probs
+            compte_ref_context_manager = (
+                amp.autocast("cuda")
+                if self._peft_has_been_casted_to_bf16
+                else nullcontext()
+            )
+            with torch.no_grad(), compte_ref_context_manager:  # type: ignore
+                if self.ref_model is None:
+                    with self.null_ref_context():
+                        ref_model_output = self.concatenated_forward(self.model, batch)
+                        ref_weight = self.model.lm_head.weight
+                        ref_bias = getattr(self.model.lm_head, "bias", None)
+
+                        ref_hidden_states = ref_model_output["hidden_states"]
+
+                else:
+                    ref_model_output = self.concatenated_forward(self.ref_model, batch)
+                    ref_weight = self.ref_model.lm_head.weight
+                    ref_bias = getattr(self.ref_model.lm_head, "bias", None)
+
+                ref_hidden_states = ref_model_output["hidden_states"]
+                (
+                    ref_chosen_logps,
+                    ref_rejected_logps,
+                    _ref_chosen_logits,
+                    _ref_rejected_logits,
+                    _ref_chosen_nll_loss,
+                ) = LigerFusedLinearPreferenceBase.chunk_forward(
+                    input_chunk=ref_hidden_states,
+                    weight=ref_weight,
+                    target_chunk=labels,
+                    bias=ref_bias,
+                    # ignore_index=ignore_index,
+                    compute_nll_loss=False,
+                )
+
+        else:
+            ref_hidden_states = None
+            ref_weight = None
+            ref_bias = None
+
+        # Compute loss using Liger kernel
+        loss, return_vars = self.liger_loss(
+            lin_weight=lin_weight,
+            _input=hidden_states,
+            target=labels,
+            bias=lin_bias,  # TODO: check whether to pass bias as FCLE doesn't
+            ref_input=ref_hidden_states,
+            ref_weight=ref_weight,
+            ref_bias=ref_bias,
+        )
+
+        (
+            policy_chosen_logps,
+            policy_rejected_logps,
+            policy_chosen_logits_mean,
+            policy_rejected_logits_mean,
+            policy_nll_loss,
+        ) = return_vars
+
+        # Calculate rewards
+        if not self.reference_free:
+            chosen_rewards = (
+                self.beta * (policy_chosen_logps - (ref_chosen_logps)).detach()
+            )
+            rejected_rewards = (
+                self.beta * (policy_rejected_logps - (ref_rejected_logps)).detach()
+            )
+
+        else:
+            chosen_rewards = self.beta * policy_chosen_logps
+            rejected_rewards = self.beta * policy_rejected_logps
+
+        reward_accuracies = (chosen_rewards > rejected_rewards).float()
+
+        prefix = "eval_" if train_eval == "eval" else ""
+        metrics.update(
+            {
+                f"{prefix}rewards/chosen": chosen_rewards.mean().cpu(),
+                f"{prefix}rewards/rejected": rejected_rewards.mean().cpu(),
+                f"{prefix}rewards/accuracies": reward_accuracies.mean().cpu(),
+                f"{prefix}rewards/margins": (chosen_rewards - rejected_rewards)
+                .mean()
+                .cpu(),
+                f"{prefix}logps/chosen": policy_chosen_logps.mean().cpu(),
+                f"{prefix}logps/rejected": policy_rejected_logps.mean().cpu(),
+                f"{prefix}logits/chosen": policy_chosen_logits_mean.cpu(),
+                f"{prefix}logits/rejected": policy_rejected_logits_mean.cpu(),
+            }
+        )
+
+        if hasattr(self.args, "rpo_alpha") and self.args.rpo_alpha is not None:
+            metrics[f"{prefix}nll_loss"] = policy_nll_loss.cpu()
+
+        # TODO: Handle use_weighting, aux_loss_enabled as in upstream
+
+        return loss, metrics
+
+    def concatenated_forward(
+        self, model: nn.Module, batch: dict[str, Union[list, torch.LongTensor]]
+    ):
+        """Run the given model on the given batch of inputs, concatenating the chosen and rejected inputs together.
+
+        We do this to avoid doing two forward passes, because it's faster for FSDP.
+
+        Overridden base function to return the hidden states and labels for the loss calculation.
+        """
+        num_examples = batch["prompt_input_ids"].shape[0]  # type: ignore
+
+        concatenated_batch = self.concatenated_inputs(
+            batch, padding_value=self.padding_value
+        )
+
+        model_kwargs = {}
+        if self.aux_loss_enabled:
+            model_kwargs["output_router_logits"] = True
+
+        # Add to get the hidden states for the loss
+        model_kwargs["output_hidden_states"] = True
+
+        # Add the pixel values and attention masks for vision models
+        if "pixel_values" in concatenated_batch:
+            model_kwargs["pixel_values"] = concatenated_batch["pixel_values"]
+        if "pixel_attention_mask" in concatenated_batch:
+            model_kwargs["pixel_attention_mask"] = concatenated_batch[
+                "pixel_attention_mask"
+            ]
+        if "image_sizes" in concatenated_batch:
+            model_kwargs["image_sizes"] = concatenated_batch["image_sizes"]
+
+        prompt_input_ids = concatenated_batch["prompt_input_ids"]
+        prompt_attention_mask = concatenated_batch["prompt_attention_mask"]
+        completion_input_ids = concatenated_batch["completion_input_ids"]
+        completion_attention_mask = concatenated_batch["completion_attention_mask"]
+        if self.is_encoder_decoder:
+            labels = completion_input_ids
+            labels[completion_attention_mask == 0] = self.label_pad_token_id
+            outputs = model(
+                input_ids=prompt_input_ids,
+                attention_mask=prompt_attention_mask,
+                labels=labels,  # we need the labels for the logits to be returned
+                **model_kwargs,
+            )
+            logits = outputs.logits
+            hidden_states = outputs.decoder_hidden_states[-1]
+            loss_mask = completion_attention_mask.bool()
+        else:
+            # Concatenate the prompt and completion inputs
+            input_ids = torch.cat((prompt_input_ids, completion_input_ids), dim=1)
+            attention_mask = torch.cat(
+                (prompt_attention_mask, completion_attention_mask), dim=1
+            )
+            # Mask the prompt but not the completion for the loss
+            loss_mask = torch.cat(
+                (torch.zeros_like(prompt_attention_mask), completion_attention_mask),
+                dim=1,
+            )
+
+            # Flush left to reduce the memory usage
+            # [[0, 0, x, x, x, x],  ->  [[x, x, x, x],
+            #  [0, x, x, x, 0, 0]]       [x, x, x, 0]]
+            for i in range(attention_mask.size(0)):
+                first_one_idx = torch.nonzero(attention_mask[i])[0].item()
+                input_ids[i] = torch.roll(input_ids[i], shifts=-first_one_idx)  # type: ignore
+                attention_mask[i] = torch.roll(attention_mask[i], shifts=-first_one_idx)  # type: ignore
+                loss_mask[i] = torch.roll(loss_mask[i], shifts=-first_one_idx)  # type: ignore
+
+            # Get the first column idx that is all zeros and remove every column after that
+            empty_cols = torch.sum(attention_mask, dim=0) == 0
+            first_empty_col = (
+                torch.nonzero(empty_cols)[0].item()
+                if empty_cols.any()
+                else attention_mask.size(1)
+            )
+            input_ids = input_ids[:, :first_empty_col]  # type: ignore
+            attention_mask = attention_mask[:, :first_empty_col]  # type: ignore
+            loss_mask = loss_mask[:, :first_empty_col]  # type: ignore
+
+            # Truncate right
+            if self.args.max_length is not None:
+                input_ids = input_ids[:, : self.args.max_length]
+                attention_mask = attention_mask[:, : self.args.max_length]
+                loss_mask = loss_mask[:, : self.args.max_length]
+
+            # if self.use_num_logits_to_keep:
+            #     # Compute num_logits_to_keep based on loss_mask pattern:
+            #     # [[0, 0, 0, x, x, x, x],
+            #     #  [0, 0, 0, x, x, x, 0]]
+            #     #         ^ start computing logits from here ([:, -(7-3+1):])
+            #     first_compute_index = loss_mask.nonzero(as_tuple=True)[1].min()
+            #     num_logits_to_keep = loss_mask.shape[1] - first_compute_index
+            #     model_kwargs["num_logits_to_keep"] = num_logits_to_keep.item() + 1  # +1 for the first label
+
+            outputs = model(
+                input_ids=input_ids, attention_mask=attention_mask, **model_kwargs
+            )
+
+            # Offset the logits by one to align with the labels
+            logits = outputs.logits[:, :-1, :]
+            hidden_states = outputs.hidden_states[-1][:, :-1, :]
+            labels = input_ids[:, 1:].clone()
+            loss_mask = loss_mask[:, 1:].bool()
+
+            # if self.use_num_logits_to_keep:
+            #     # Align labels with logits
+            #     # logits:    -,  -, [x2, x3, x4, x5, x6]
+            #     #                     ^ --------- ^       after logits[:, :-1, :]
+            #     # labels:   [y0, y1, y2, y3, y4, y5, y6]
+            #     #                         ^ --------- ^   with num_logits_to_keep=4, [:, -4:]
+            #     # loss_mask: [0,  0,  0,  1,  1,  1,  1]
+            #     labels = labels[:, -num_logits_to_keep:]
+            #     loss_mask = loss_mask[:, -num_logits_to_keep:]
+            #     hidden_states = hidden_states[:, -num_logits_to_keep:, :]
+
+        if logits.shape[:2] != labels.shape[:2]:
+            # for llava, the returned logits include the image tokens (placed before the text tokens)
+            seq_len = labels.shape[1]
+            logits = logits[:, -seq_len:]
+            hidden_states = hidden_states[:, -seq_len:]
+
+        # Compute the log probabilities of the labels
+        labels[
+            ~loss_mask
+        ] = 0  # dummy token; we'll ignore the losses on these tokens later
+        per_token_logps = torch.gather(
+            logits.log_softmax(-1), dim=2, index=labels.unsqueeze(2)
+        ).squeeze(2)
+        per_token_logps[~loss_mask] = 0
+        all_logps = per_token_logps.sum(-1)
+
+        output = {}
+
+        if self.use_weighting:
+            with torch.no_grad():
+                # Eq (2) of the WPO paper: https://huggingface.co/papers/2406.11827
+                logprobs = F.log_softmax(logits, dim=-1)
+                weights_adjustment_factor = torch.logsumexp(
+                    2 * logprobs, dim=-1
+                )  # same as sum(probs**2) in log space
+                per_token_logps_adjusted = per_token_logps - weights_adjustment_factor
+                all_weights = (per_token_logps_adjusted * loss_mask).sum(
+                    -1
+                ) / loss_mask.sum(-1)
+                chosen_weights = all_weights[:num_examples]
+                rejected_weights = all_weights[num_examples:]
+                output["policy_weights"] = torch.clamp(
+                    torch.exp(chosen_weights + rejected_weights), max=1
+                )
+
+        if self.args.rpo_alpha is not None:
+            # Only use the chosen logits for the RPO loss
+            chosen_logits = logits[:num_examples]
+            chosen_labels = labels[:num_examples]
+
+            # Compute the log probabilities of the labels
+            output["nll_loss"] = F.cross_entropy(
+                torch.flatten(chosen_logits, end_dim=1),
+                torch.flatten(chosen_labels, end_dim=1),
+                ignore_index=0,
+            )
+
+        if self.loss_type == "ipo":
+            all_logps = all_logps / loss_mask.sum(-1)
+
+        output["chosen_logps"] = all_logps[:num_examples]
+        output["rejected_logps"] = all_logps[num_examples:]
+        output["mean_chosen_logits"] = logits[:num_examples][
+            loss_mask[:num_examples]
+        ].mean()
+        output["mean_rejected_logits"] = logits[num_examples:][
+            loss_mask[num_examples:]
+        ].mean()
+        output["hidden_states"] = hidden_states
+        output["labels"] = labels
+
+        if self.aux_loss_enabled:
+            output["aux_loss"] = outputs.aux_loss
+
+        return output
+
 
 class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
     """
@@ -2163,6 +2480,14 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.dpo_use_weighting is not None:
                 training_args_kwargs["use_weighting"] = self.cfg.dpo_use_weighting
 
+        report_to = []
+        if self.cfg.use_wandb:
+            report_to.append("wandb")
+            if self.cfg.wandb_name:
+                training_args_kwargs["run_name"] = self.cfg.wandb_name
+
+        training_args_kwargs["report_to"] = report_to
+
         training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
             output_dir=self.cfg.output_dir,
             per_device_train_batch_size=self.cfg.micro_batch_size,