override the entire create_optimzier method

* support galore once upstreamed into transformers

* update module name for llama in readme and fix typing for all linear

* bump trl for deprecation fixes from newer transformers

* include galore as an extra and install in docker image

* fix optim_args type

* fix optim_args

* update dependencies for galore

* add galore to cicd dockerfile

README.md

@@ -13,6 +13,9 @@ Features:
 - Log results and optionally checkpoints to wandb or mlflow
 - And more!
 
+<a href="https://www.phorm.ai/query?projectId=e315ba4a-4e14-421f-ab05-38a1f9076f25">
+  <img alt="phorm.ai" src="https://img.shields.io/badge/Phorm-Ask_AI-%23F2777A.svg?&logo=data:image/svg+xml;base64,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">
+</a>
 
 <table>
 <tr>
@@ -28,6 +31,7 @@ Features:
   - [Cloud GPU](#cloud-gpu) - Latitude.sh, JarvisLabs, RunPod
   - [Bare Metal Cloud GPU](#bare-metal-cloud-gpu)
   - [Windows](#windows)
+  - [Mac](#mac)
   - [Launching on public clouds via SkyPilot](#launching-on-public-clouds-via-skypilot)
 - [Dataset](#dataset)
   - [How to Add Custom Prompts](#how-to-add-custom-prompts)
@@ -99,24 +103,14 @@ Get started with Axolotl in just a few steps! This quickstart guide will walk yo
 
 **Requirements**: Python >=3.10 and Pytorch >=2.1.1.
 
-### For developers
 ```bash
 git clone https://github.com/OpenAccess-AI-Collective/axolotl
 cd axolotl
 
 pip3 install packaging
-```
-
-General case:
-```
 pip3 install -e '.[flash-attn,deepspeed]'
 ```
 
-Mac: see https://github.com/OpenAccess-AI-Collective/axolotl/blob/13199f678b9aab39e92961323bdbce3234ee4b2b/docs/mac.md
-```
-pip3 install -e '.'
-```
-
 ### Usage
 ```bash
 # preprocess datasets - optional but recommended
@@ -249,9 +243,31 @@ For cloud GPU providers that support docker images, use [`winglian/axolotl-cloud
   ```
   </details>
 
+##### GCP
+
+<details>
+
+<summary>Click to Expand</summary>
+
+Use a Deeplearning linux OS with cuda and pytorch installed. Then follow instructions on quickstart.
+
+Make sure to run the below to uninstall xla.
+```bash
+pip uninstall -y torch_xla[tpu]
+```
+
+</details>
+
 #### Windows
 Please use WSL or Docker!
 
+#### Mac
+
+Use the below instead of the install method in QuickStart.
+```
+pip3 install -e '.'
+```
+More info: [mac.md](/docs/mac.md)
 
 #### Launching on public clouds via SkyPilot
 To launch on GPU instances (both on-demand and spot instances) on 7+ clouds (GCP, AWS, Azure, OCI, and more), you can use [SkyPilot](https://skypilot.readthedocs.io/en/latest/index.html):
@@ -635,9 +651,13 @@ datasets:
     train_on_split: train # Optional[str] name of dataset split to load from
 
     # Optional[str] fastchat conversation type, only used with type: sharegpt
-    conversation:  # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+    conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
     field_human: # Optional[str]. Human key to use for conversation.
     field_model: # Optional[str]. Assistant key to use for conversation.
+    # Add additional keys from your dataset as input or output roles
+    roles:
+      input: # Optional[List[str]]. These will be masked based on train_on_input
+      output: # Optional[List[str]].
 
   # Custom user instruction prompt
   - path: repo
@@ -662,6 +682,10 @@ datasets:
       # For `completion` datsets only, uses the provided field instead of `text` column
       field:
 
+# If false, the datasets will not be shuffled and will keep their original order in `datasets`.
+# The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.
+shuffle_merged_datasets: true
+
 # A list of one or more datasets to eval the model with.
 # You can use either test_datasets, or val_set_size, but not both.
 test_datasets:
@@ -843,7 +867,7 @@ group_by_length: false
 gradient_checkpointing: false
 # additional kwargs to pass to the trainer for gradient checkpointing
 # gradient_checkpointing_kwargs:
-#   use_reentrant: false
+#   use_reentrant: true
 
 # Stop training after this many evaluation losses have increased in a row
 # https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback
@@ -883,7 +907,26 @@ lr_div_factor: # Learning rate div factor
 # - paged_adamw_8bit
 # - paged_lion_32bit
 # - paged_lion_8bit
+# - galore_adamw
+# - galore_adamw_8bit
+# - galore_adafactor
+# - galore_adamw_layerwise
+# - galore_adamw_8bit_layerwise
+# - galore_adafactor_layerwise
 optimizer:
+# Dictionary of arguments to pass to the optimizer
+optim_args:
+# For Galore Optimizers the following optim_args are available
+# rank:  # type: int
+# update_proj_gap  # type: int
+# scale  # type: float
+# proj_type:  # type: str, default = std
+
+# The target modules to optimize, i.e. the module names that you would like to train, right now this is used only for GaLore algorithm
+optim_target_modules:
+# - self_attn  # for llama
+# - mlp
+
 # Specify weight decay
 weight_decay:
 # adamw hyperparams

cicd/Dockerfile.jinja

@@ -23,9 +23,9 @@ RUN git fetch origin +$GITHUB_REF && \
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[deepspeed,flash-attn,mamba-ssm,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,galore,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install -e .[deepspeed,flash-attn,mamba-ssm] $AXOLOTL_ARGS; \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,galore] $AXOLOTL_ARGS; \
     fi
 
 # So we can test the Docker image

docker/Dockerfile

@@ -21,9 +21,9 @@ WORKDIR /workspace/axolotl
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[deepspeed,flash-attn,mamba-ssm,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,galore,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install -e .[deepspeed,flash-attn,mamba-ssm] $AXOLOTL_ARGS; \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,galore] $AXOLOTL_ARGS; \
     fi
 
 # So we can test the Docker image

docs/optimizers.md

@@ -0,0 +1,29 @@
+# Optimizers
+
+Optimizers are an important component when training LLMs. Optimizers are responsible for updating the model's weights (parameters) based on the gradients computed during backpropagation.
+The goal of an optimizer is to minimize the loss function.
+
+### Adam/AdamW Optimizers
+
+```yaml
+adam_beta1: 0.9
+adam_beta2: 0.999
+adam_epsilon: 1e-8
+weight_decay: 0.0
+```
+
+### GaLore Optimizer
+
+https://huggingface.co/papers/2403.03507
+
+```yaml
+optimizer: galore_adamw | galore_adamw_8bit | galore_adafactor
+optim_args:
+  rank: 128
+  update_proj_gap: 200
+  scale: 0.25
+  proj_type: std
+optim_target_modules:
+  - mlp
+  - attn
+```

docs/rlhf.md

@@ -34,6 +34,21 @@ datasets:
 rl: ipo
 ```
 
+#### ORPO
+
+Paper: https://arxiv.org/abs/2403.07691
+
+```yaml
+rl: orpo
+orpo_alpha: 0.1
+remove_unused_columns: false
+
+chat_template: chatml
+datasets:
+  - path: argilla/ultrafeedback-binarized-preferences-cleaned
+    type: orpo.chat_template
+```
+
 #### Using local dataset files
 ```yaml
 datasets:

requirements.txt

@@ -1,7 +1,7 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
 peft==0.9.0
-transformers==4.38.2
+transformers @ git+https://github.com/huggingface/transformers.git@f6261d7d81edd036fc53bfede65fe91f01a661aa
 tokenizers==0.15.0
 bitsandbytes>=0.43.0
 accelerate==0.26.1
@@ -39,5 +39,8 @@ s3fs
 gcsfs
 # adlfs
 
-trl>=0.7.9
+trl @ git+https://github.com/huggingface/trl.git@304e208f778a5442c30cdda500348226cdc97d90
 fastcore>=1.5.29
+
+lpmm @ git+https://github.com/thu-ml/low-bit-optimizers.git@main
+yacs

setup.py

@@ -89,5 +89,8 @@ setup(
         "lion-pytorch": [
             "lion-pytorch==0.1.2",
         ],
+        "galore": [
+            "galore_torch",
+        ],
     },
 )

src/axolotl/cli/preprocess.py

@@ -54,7 +54,7 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
         LOG.warning(msg)
         parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
-    if parsed_cfg.rl:
+    if parsed_cfg.rl and parsed_cfg.rl != "orpo":
         load_rl_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
     else:
         load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)

src/axolotl/cli/train.py

@@ -47,7 +47,7 @@ def do_train(cfg, cli_args) -> Tuple[PreTrainedModel, PreTrainedTokenizer]:
     else:
         register_chatml_template()
 
-    if cfg.rl:
+    if cfg.rl and cfg.rl != "orpo":
         dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)
     else:
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

src/axolotl/core/trainer_builder.py

@@ -11,21 +11,25 @@ import math
 import os
 import sys
 from abc import abstractmethod
+from collections import defaultdict
 from dataclasses import dataclass, field
 from functools import wraps
 from pathlib import Path
-from typing import List, Optional, Type, Union
+from typing import Any, Dict, List, Literal, Optional, Tuple, Type, Union
 
+import lpmm
 import torch
 import transformers
 from accelerate import FullyShardedDataParallelPlugin
 from accelerate.utils import str_to_bool
 from datasets import Dataset
+from torch import nn
 from torch.distributed.fsdp import MixedPrecision
 from torch.optim.lr_scheduler import OneCycleLR
 from torch.utils.data import BatchSampler, DataLoader, RandomSampler, SequentialSampler
 from transformers import (
     EarlyStoppingCallback,
+    PreTrainedModel,
     Trainer,
     TrainerCallback,
     TrainingArguments,
@@ -35,6 +39,7 @@ from transformers.utils import is_sagemaker_mp_enabled
 from trl import DPOTrainer
 
 from axolotl.core.policies.auto_wrap import get_wrapping_policy_factory
+from axolotl.core.trainers import OptimizerNames
 from axolotl.loraplus import create_loraplus_optimizer
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
@@ -61,6 +66,9 @@ from axolotl.utils.schedulers import (
     get_cosine_schedule_with_warmup_decay_constant,
 )
 
+# monkeypatch so it accepts our custom optimizers
+transformers.training_args.OptimizerNames = OptimizerNames
+
 if is_sagemaker_mp_enabled():
     import smdistributed.modelparallel.torch as smp
 
@@ -200,6 +208,9 @@ class AxolotlTrainingArguments(TrainingArguments):
         default=False,
         metadata={"help": "whether this is a qlora training"},
     )
+    orpo_alpha: Optional[float] = field(
+        default=None,
+    )
 
 
 class AxolotlTrainer(Trainer):
@@ -216,33 +227,115 @@ class AxolotlTrainer(Trainer):
         num_epochs=1,
         bench_data_collator=None,
         eval_data_collator=None,
-        **kwargs
+        **kwargs,
     ):
         self.num_epochs = num_epochs
         self.bench_data_collator = bench_data_collator
         self.eval_data_collator = eval_data_collator
         super().__init__(*_args, **kwargs)
         self.train_data_collator = self.data_collator
+        self._stored_metrics = defaultdict(lambda: defaultdict(list))
+        if self.args.orpo_alpha:
+            self.loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
+
+    @staticmethod
+    def get_optimizer_cls_and_kwargs(
+        args: TrainingArguments, model: Optional[PreTrainedModel] = None
+    ) -> Tuple[Any, Any]:
+        optim_args = {}
+        if args.optim_args:
+            for mapping in args.optim_args.replace(" ", "").split(","):
+                key, value = mapping.split("=")
+                optim_args[key] = value
+
+        optimizer_kwargs = {"lr": args.learning_rate}
+
+        adam_kwargs = {
+            "betas": (args.adam_beta1, args.adam_beta2),
+            "eps": args.adam_epsilon,
+        }
+
+        if args.optim in [
+            OptimizerNames.LPMM_ADAMW_4BIT,
+            OptimizerNames.LPMM_ADAMW_4BIT_FUSED,
+        ]:
+            optimizer_cls = lpmm.optim.AdamW
+            optimizer_kwargs.update(adam_kwargs)
+            if args.optim == OptimizerNames.LPMM_ADAMW_4BIT_FUSED:
+                optimizer_kwargs.update({"fused": True})
+            return optimizer_cls, optimizer_kwargs
+
+        return Trainer.get_optimizer_cls_and_kwargs(
+            args,
+            model=model,
+        )
 
     def create_optimizer(self):
-        if self.args.loraplus_lr_ratio is None:
-            return super().create_optimizer()
-
         opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
-        if self.optimizer is None:  # pylint: disable=access-member-before-definition
-            optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(
-                self.args,
-            )
 
-            loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
+        if self.optimizer is None:  # pylint: disable=access-member-before-definition
-            loraplus_lr_embedding = getattr(self.args, "loraplus_lr_embedding", None)
+            decay_parameters = self.get_decay_parameter_names(opt_model)
-            self.optimizer = create_loraplus_optimizer(  # pylint: disable=attribute-defined-outside-init
+            optimizer_grouped_parameters = [
-                opt_model,
+                {
+                    "params": [
+                        p
+                        for n, p in opt_model.named_parameters()
+                        if (n in decay_parameters and p.requires_grad)
+                    ],
+                    "weight_decay": self.args.weight_decay,
+                },
+                {
+                    "params": [
+                        p
+                        for n, p in opt_model.named_parameters()
+                        if (n not in decay_parameters and p.requires_grad)
+                    ],
+                    "weight_decay": 0.0,
+                },
+            ]
+
+            (
                 optimizer_cls,
                 optimizer_kwargs,
-                loraplus_lr_ratio,
+            ) = AxolotlTrainer.get_optimizer_cls_and_kwargs(self.args)
-                loraplus_lr_embedding,
+
-            )
+            if self.args.loraplus_lr_ratio:
+                loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
+                loraplus_lr_embedding = getattr(
+                    self.args, "loraplus_lr_embedding", None
+                )
+                self.optimizer = create_loraplus_optimizer(  # pylint: disable=attribute-defined-outside-init
+                    opt_model,
+                    optimizer_cls,
+                    optimizer_kwargs,
+                    loraplus_lr_ratio,
+                    loraplus_lr_embedding,
+                )
+
+            else:
+                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
+                    optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
+                )
+
+            if optimizer_cls.__name__ == "Adam8bit":
+                import bitsandbytes
+
+                manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
+
+                skipped = 0
+                for module in opt_model.modules():
+                    if isinstance(module, nn.Embedding):
+                        skipped += sum(
+                            {
+                                p.data_ptr(): p.numel() for p in module.parameters()
+                            }.values()
+                        )
+                        LOG.info(f"skipped {module}: {skipped/2**20}M params")
+                        manager.register_module_override(
+                            module, "weight", {"optim_bits": 32}
+                        )
+                        LOG.debug(f"bitsandbytes: will optimize {module} in fp32")
+                LOG.info(f"skipped: {skipped/2**20}M params")
 
         if is_sagemaker_mp_enabled():
             self.optimizer = smp.DistributedOptimizer(  # pylint: disable=attribute-defined-outside-init
@@ -465,8 +558,112 @@ class AxolotlTrainer(Trainer):
         #     outputs = model(**inputs)
         #     loss = trainer_weighted_loss(outputs, labels, shift_labels=True)
         #     return (loss, outputs) if return_outputs else loss
+        if self.args.orpo_alpha:
+            return self.orpo_compute_loss(model, inputs, return_outputs=return_outputs)
         return super().compute_loss(model, inputs, return_outputs=return_outputs)
 
+    def orpo_compute_custom_loss(self, logits, labels):
+        logits = logits.contiguous()
+        loss = 0.0
+
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+
+            # Flatten the tokens
+            loss = self.loss_fct(shift_logits.transpose(2, 1), shift_labels).mean(
+                dim=-1
+            )
+
+        return loss
+
+    def orpo_compute_logps(
+        self, prompt_attention_mask, chosen_inputs, chosen_attention_mask, logits
+    ):
+        # Get the shape of chosen_attention_mask[:, :-1]
+        chosen_shape = chosen_attention_mask[:, :-1].shape
+
+        # Calculate the padding size
+        pad_length = chosen_shape[1] - (prompt_attention_mask.shape[1] - 1)
+
+        # Pad prompt_attention_mask with zeros to match the desired shape
+        prompt_attention_mask_padded = torch.nn.functional.pad(
+            prompt_attention_mask[:, 1:], (0, pad_length), mode="constant", value=0
+        )
+
+        # Perform the subtraction operation
+        mask = chosen_attention_mask[:, :-1] > prompt_attention_mask_padded
+
+        per_token_logps = torch.gather(
+            logits[:, :-1, :].log_softmax(-1),
+            dim=2,
+            index=(mask * chosen_inputs[:, 1:]).unsqueeze(2),
+        ).squeeze(2)
+        return torch.mul(per_token_logps, mask.to(dtype=torch.bfloat16)).sum(dim=1).to(
+            dtype=torch.float64
+        ) / mask.sum(dim=1).to(dtype=torch.float64)
+
+    def orpo_compute_loss(self, model, inputs, return_outputs=False):
+        outputs_neg = model(
+            **{
+                "input_ids": inputs["rejected_input_ids"],
+                "attention_mask": inputs["rejected_attention_mask"],
+                "labels": inputs["rejected_labels"],
+            },
+            output_hidden_states=True,
+        )
+        outputs_pos = model(
+            **{
+                "input_ids": inputs["input_ids"],
+                "attention_mask": inputs["attention_mask"],
+                "labels": inputs["labels"],
+            },
+            output_hidden_states=True,
+        )
+
+        # Calculate NLL loss
+        pos_loss = self.orpo_compute_custom_loss(
+            logits=outputs_pos.logits, labels=inputs["input_ids"]
+        )
+
+        # Calculate Log Probability
+        pos_prob = self.orpo_compute_logps(
+            prompt_attention_mask=inputs["prompt_attention_mask"],
+            chosen_inputs=inputs["input_ids"],
+            chosen_attention_mask=inputs["attention_mask"],
+            logits=outputs_pos.logits,
+        )
+        neg_prob = self.orpo_compute_logps(
+            prompt_attention_mask=inputs["prompt_attention_mask"],
+            chosen_inputs=inputs["rejected_input_ids"],
+            chosen_attention_mask=inputs["rejected_attention_mask"],
+            logits=outputs_neg.logits,
+        )
+
+        # Calculate log odds
+        log_odds = (pos_prob - neg_prob) - (
+            torch.log(1 - torch.exp(pos_prob)) - torch.log(1 - torch.exp(neg_prob))
+        )
+        sig_ratio = torch.nn.functional.sigmoid(log_odds)
+        ratio = torch.log(sig_ratio)
+
+        # Calculate the Final Loss
+        loss = torch.mean(pos_loss - self.args.orpo_alpha * ratio).to(
+            dtype=torch.bfloat16
+        )
+
+        metrics = {}
+        metrics["chosen_geometric_mean"] = torch.mean(pos_prob).cpu().item()
+        metrics["rejected_geometric_mean"] = torch.mean(neg_prob).cpu().item()
+        metrics["log_odds_ratio"] = torch.mean(ratio).cpu().item()
+        metrics["log_odds"] = torch.mean(log_odds).cpu().item()
+        self.store_metrics(metrics, train_eval="train")
+
+        return (loss, outputs_pos) if return_outputs else loss
+
     @wraps(Trainer.push_to_hub)
     def push_to_hub(self, *args, **kwargs) -> str:
         """
@@ -527,6 +724,28 @@ class AxolotlTrainer(Trainer):
 
         return res
 
+    def log(self, logs: Dict[str, float]) -> None:
+        """
+        Log `logs` on the various objects watching training, including stored metrics.
+
+        Args:
+            logs (`Dict[str, float]`):
+                The values to log.
+        """
+        # logs either has 'loss' or 'eval_loss'
+        train_eval = "train" if "loss" in logs else "eval"
+        # Add averaged stored metrics to logs
+        for key, metrics in self._stored_metrics[train_eval].items():
+            logs[key] = torch.tensor(metrics).mean().item()
+        del self._stored_metrics[train_eval]
+        return super().log(logs)
+
+    def store_metrics(
+        self, metrics: Dict[str, float], train_eval: Literal["train", "eval"] = "train"
+    ) -> None:
+        for key, value in metrics.items():
+            self._stored_metrics[train_eval][key].append(value)
+
 
 class AxolotlMambaTrainer(AxolotlTrainer):
     """
@@ -837,10 +1056,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 training_arguments_kwargs[
                     "gradient_checkpointing_kwargs"
                 ] = self.cfg.gradient_checkpointing_kwargs
-            else:
-                training_arguments_kwargs["gradient_checkpointing_kwargs"] = {
-                    "use_reentrant": False
-                }
         if self.cfg.fsdp:
             training_arguments_kwargs["fsdp"] = self.cfg.fsdp
             if self.cfg.fsdp_config:
@@ -903,6 +1118,11 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         elif self.cfg.sample_packing and self.cfg.eval_sample_packing is False:
             training_arguments_kwargs["dataloader_drop_last"] = True
 
+        if self.cfg.remove_unused_columns is not None:
+            training_arguments_kwargs[
+                "remove_unused_columns"
+            ] = self.cfg.remove_unused_columns
+
         if not self.cfg.test_datasets and self.cfg.val_set_size == 0:
             # no eval set, so don't eval
             training_arguments_kwargs["evaluation_strategy"] = "no"
@@ -1016,6 +1236,18 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["optim"] = (
             self.cfg.optimizer if self.cfg.optimizer else "adamw_hf"
         )
+        if self.cfg.optim_args:
+            if isinstance(self.cfg.optim_args, dict):
+                optim_args = ",".join(
+                    [f"{key}={value}" for key, value in self.cfg.optim_args.items()]
+                )
+            else:
+                optim_args = self.cfg.optim_args
+            training_arguments_kwargs["optim_args"] = optim_args
+        if self.cfg.optim_target_modules:
+            training_arguments_kwargs[
+                "optim_target_modules"
+            ] = self.cfg.optim_target_modules
         training_arguments_kwargs["loraplus_lr_ratio"] = self.cfg.loraplus_lr_ratio
         training_arguments_kwargs[
             "loraplus_lr_embedding"
@@ -1070,6 +1302,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["model_type"] = self.cfg.model_config_type
         training_arguments_kwargs["pretraining"] = bool(self.cfg.pretraining_dataset)
 
+        if self.cfg.rl == "orpo":
+            training_arguments_kwargs["orpo_alpha"] = self.cfg.orpo_alpha
+
         if self.cfg.neftune_noise_alpha is not None:
             training_arguments_kwargs[
                 "neftune_noise_alpha"

src/axolotl/core/trainers/__init__.py

@@ -0,0 +1,40 @@
+"""module for trainer helpers like OptimizerNames"""
+
+from transformers.utils import ExplicitEnum
+
+
+class OptimizerNames(ExplicitEnum):
+    """
+    Stores the acceptable string identifiers for optimizers.
+    """
+
+    ADAMW_HF = "adamw_hf"
+    ADAMW_TORCH = "adamw_torch"
+    ADAMW_TORCH_FUSED = "adamw_torch_fused"
+    ADAMW_TORCH_XLA = "adamw_torch_xla"
+    ADAMW_TORCH_NPU_FUSED = "adamw_torch_npu_fused"
+    ADAMW_APEX_FUSED = "adamw_apex_fused"
+    ADAFACTOR = "adafactor"
+    ADAMW_ANYPRECISION = "adamw_anyprecision"
+    SGD = "sgd"
+    ADAGRAD = "adagrad"
+    ADAMW_BNB = "adamw_bnb_8bit"
+    ADAMW_8BIT = "adamw_8bit"  # just an alias for adamw_bnb_8bit
+    LION_8BIT = "lion_8bit"
+    LION = "lion_32bit"
+    PAGED_ADAMW = "paged_adamw_32bit"
+    PAGED_ADAMW_8BIT = "paged_adamw_8bit"
+    PAGED_LION = "paged_lion_32bit"
+    PAGED_LION_8BIT = "paged_lion_8bit"
+    RMSPROP = "rmsprop"
+    RMSPROP_BNB = "rmsprop_bnb"
+    RMSPROP_8BIT = "rmsprop_bnb_8bit"
+    RMSPROP_32BIT = "rmsprop_bnb_32bit"
+    GALORE_ADAMW = "galore_adamw"
+    GALORE_ADAMW_8BIT = "galore_adamw_8bit"
+    GALORE_ADAFACTOR = "galore_adafactor"
+    GALORE_ADAMW_LAYERWISE = "galore_adamw_layerwise"
+    GALORE_ADAMW_8BIT_LAYERWISE = "galore_adamw_8bit_layerwise"
+    GALORE_ADAFACTOR_LAYERWISE = "galore_adafactor_layerwise"
+    LPMM_ADAMW_4BIT = "lmpp_adamw_4bit"
+    LPMM_ADAMW_4BIT_FUSED = "lmpp_adamw_4bit_fused"

src/axolotl/prompt_strategies/base.py

@@ -0,0 +1,20 @@
+"""
+module for base dataset transform strategies
+"""
+
+import importlib
+import logging
+
+LOG = logging.getLogger("axolotl")
+
+
+def load(strategy, cfg, module_base=None, **kwargs):
+    try:
+        load_fn = strategy.split(".")[-1]
+        strategy = ".".join(strategy.split(".")[:-1])
+        mod = importlib.import_module(f".{strategy}", module_base)
+        func = getattr(mod, load_fn)
+        return func(cfg, **kwargs)
+    except Exception:  # pylint: disable=broad-exception-caught
+        LOG.warning(f"unable to load strategy {strategy}")
+        return None

src/axolotl/prompt_strategies/dpo/__init__.py

@@ -1,20 +1,8 @@
 """
 module for DPO style dataset transform strategies
 """
+from functools import partial
 
-import importlib
+from ..base import load as load_base
-import logging
 
-LOG = logging.getLogger("axolotl")
+load = partial(load_base, module="axolotl.prompt_strategies.dpo")
-
-
-def load(strategy, cfg, **kwargs):
-    try:
-        load_fn = strategy.split(".")[-1]
-        strategy = ".".join(strategy.split(".")[:-1])
-        mod = importlib.import_module(f".{strategy}", "axolotl.prompt_strategies.dpo")
-        func = getattr(mod, load_fn)
-        return func(cfg, **kwargs)
-    except Exception:  # pylint: disable=broad-exception-caught
-        LOG.warning(f"unable to load strategy {strategy}")
-        return None

src/axolotl/prompt_strategies/orpo/__init__.py

@@ -0,0 +1,9 @@
+"""
+module for ORPO style dataset transform strategies
+"""
+
+from functools import partial
+
+from ..base import load as load_base
+
+load = partial(load_base, module="axolotl.prompt_strategies.orpo")

src/axolotl/prompt_strategies/orpo/chat_template.py

@@ -0,0 +1,187 @@
+"""chatml prompt tokenization strategy for ORPO"""
+from typing import Any, Dict, Generator, List, Optional, Tuple
+
+from pydantic import BaseModel
+
+from axolotl.prompt_tokenizers import IGNORE_INDEX, PromptTokenizingStrategy
+from axolotl.prompters import Prompter
+from axolotl.utils.chat_templates import chat_templates
+
+
+class Message(BaseModel):
+    """message/turn"""
+
+    role: str
+    content: str
+    label: Optional[bool] = None
+
+
+class MessageList(BaseModel):
+    """conversation"""
+
+    messages: List[Message]
+
+
+def load(
+    tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None, **kwargs
+):  # pylint: disable=possibly-unused-variable,unused-argument
+    """
+    chatml transforms for datasets with system, input, chosen, rejected
+    """
+
+    chat_template = chat_templates("chatml")
+    if ds_cfg and "chat_template" in ds_cfg:
+        chat_template = ds_cfg["chat_template"]
+        try:
+            chat_template = chat_templates(chat_template)
+        except ValueError:
+            pass
+
+    return ORPOTokenizingStrategy(
+        ORPOPrompter(chat_template, tokenizer),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+        dataset_parser=ORPODatasetParsingStrategy(),
+    )
+
+
+class ORPODatasetParsingStrategy:
+    """Strategy to parse chosen rejected dataset into messagelist"""
+
+    def get_chosen_conversation_thread(self, prompt) -> MessageList:
+        """Dataset structure mappings"""
+
+        messages: List[Message] = []
+        if system := prompt.get("system", None):
+            messages.append(Message(role="system", content=system, label=False))
+        messages.append(Message(role="user", content=prompt["prompt"], label=False))
+        messages.append(
+            Message(
+                role="assistant", content=prompt["chosen"][1]["content"], label=True
+            )
+        )
+        return MessageList(messages=messages)
+
+    def get_rejected_conversation_thread(self, prompt) -> MessageList:
+        """Dataset structure mappings"""
+
+        messages: List[Message] = []
+        if system := prompt.get("system", None):
+            messages.append(Message(role="system", content=system, label=False))
+        messages.append(Message(role="user", content=prompt["prompt"], label=False))
+        messages.append(
+            Message(
+                role="assistant", content=prompt["rejected"][1]["content"], label=True
+            )
+        )
+        return MessageList(messages=messages)
+
+
+class ORPOTokenizingStrategy(PromptTokenizingStrategy):
+    """
+    rejected_input_ids
+    input_ids
+    rejected_attention_mask
+    attention_mask
+    rejected_labels
+    labels
+    """
+
+    def __init__(
+        self,
+        *args,
+        dataset_parser=None,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        self.dataset_parser = dataset_parser
+
+    def tokenize_prompt(self, prompt):
+        # pass the rejected prompt/row to the Prompter to get the formatted prompt
+        prompt_len = 0
+        rejected_message_list = self.dataset_parser.get_rejected_conversation_thread(
+            prompt
+        )
+        input_ids = []
+        labels = []
+        for _, (part, label) in enumerate(
+            self.prompter.build_prompt(rejected_message_list)
+        ):
+            if not part:
+                continue
+            _input_ids = self.tokenizer.encode(part, add_special_tokens=False)
+            prev_idx = len(input_ids)
+            input_ids += _input_ids[prev_idx:]
+            if label:
+                labels += input_ids[prev_idx:]
+            else:
+                labels += [IGNORE_INDEX] * (len(input_ids) - prev_idx)
+                prompt_len = len(input_ids)
+        # remap the input_ids, attention_mask and labels
+        rejected_input_ids = input_ids
+        rejected_labels = labels
+        # pass the chosen prompt/row to the Prompter to get the formatted prompt
+        chosen_message_list = self.dataset_parser.get_chosen_conversation_thread(prompt)
+        input_ids = []
+        labels = []
+        for _, (part, label) in enumerate(
+            self.prompter.build_prompt(chosen_message_list)
+        ):
+            if not part:
+                continue
+            _input_ids = self.tokenizer.encode(part, add_special_tokens=False)
+            prev_idx = len(input_ids)
+            input_ids += _input_ids[prev_idx:]
+            if label:
+                labels += input_ids[prev_idx:]
+            else:
+                labels += [IGNORE_INDEX] * (len(input_ids) - prev_idx)
+
+        return {
+            "rejected_input_ids": rejected_input_ids,
+            "rejected_labels": rejected_labels,
+            "rejected_attention_mask": [1] * len(rejected_labels),
+            "input_ids": input_ids,
+            "labels": labels,
+            "attention_mask": [1] * len(labels),
+            "prompt_attention_mask": [1] * prompt_len
+            + [0] * (len(labels) - prompt_len),
+        }
+
+
+class ORPOPrompter(Prompter):
+    """Single Turn prompter for ORPO"""
+
+    def __init__(self, chat_template, tokenizer):
+        self.chat_template = chat_template
+        self.tokenizer = tokenizer
+
+    def build_prompt(
+        self,
+        message_list: MessageList,
+    ) -> Generator[Tuple[str, bool], None, None]:
+        conversation = []
+        for message in message_list.messages:
+            conversation.append(message.model_dump())
+            if message.role == "system":
+                yield self.tokenizer.apply_chat_template(
+                    conversation,
+                    add_generation_prompt=False,
+                    chat_template=self.chat_template,
+                    tokenize=False,
+                ), False
+            if message.role == "user":
+                yield self.tokenizer.apply_chat_template(
+                    conversation,
+                    add_generation_prompt=True,
+                    chat_template=self.chat_template,
+                    tokenize=False,
+                ), False
+            if message.role == "assistant":
+                yield self.tokenizer.apply_chat_template(
+                    conversation,
+                    add_generation_prompt=False,
+                    chat_template=self.chat_template,
+                    tokenize=False,
+                ), True

src/axolotl/prompt_strategies/sharegpt.py

@@ -1,5 +1,6 @@
 """Module containing the SimpleShareGPTPromptTokenizingStrategy class"""
 
+import logging
 from typing import Any, Dict, Optional
 
 from fastchat.conversation import Conversation, SeparatorStyle, register_conv_template
@@ -11,6 +12,8 @@ from axolotl.utils.tokenization import (
     merge_consecutive_messages,
 )
 
+LOG = logging.getLogger("axolotl")
+
 
 def register_chatml_template(system_message=None):
     system_message = system_message or "You are a helpful assistant."
@@ -42,11 +45,13 @@ def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
     )
     field_human = ds_cfg["field_human"] if ds_cfg and "field_human" in ds_cfg else None
     field_model = ds_cfg["field_model"] if ds_cfg and "field_model" in ds_cfg else None
+    roles = ds_cfg["roles"].to_dict() if ds_cfg and "roles" in ds_cfg else None
     strategy = SimpleShareGPTPromptTokenizingStrategy(
         ShareGPTPrompterV2(
             conversation=conversation,
             role_key_model=field_model,
             role_key_human=field_human,
+            roles=roles,
         ),
         tokenizer,
         cfg.train_on_inputs,
@@ -142,7 +147,12 @@ class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
             "system": "system",
         }
         turns = [
-            {"from": role_map[t[role_key]], "value": t[value_key]}
+            {
+                "from": (
+                    role_map[t[role_key]] if t[role_key] in role_map else t[role_key]
+                ),
+                "value": t[value_key],
+            }
             for t in conversations
         ]
         return turns

src/axolotl/prompt_tokenizers.py

@@ -11,7 +11,7 @@ from transformers import BatchEncoding, PreTrainedTokenizer
 from axolotl.monkeypatch.fastchat_conversation_turns import (
     add_get_turns_to_conversation,
 )
-from axolotl.prompters import IGNORE_TOKEN_ID
+from axolotl.prompters import IGNORE_TOKEN_ID, Prompter
 
 LOG = logging.getLogger("axolotl")
 
@@ -37,7 +37,7 @@ class PromptTokenizingStrategy(abc.ABC):
 
     def __init__(
         self,
-        prompter,
+        prompter: Prompter,
         tokenizer,
         train_on_inputs: bool = False,
         sequence_len: int = 2048,
@@ -340,6 +340,23 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
             self.prompter._conversation.copy()  # pylint: disable=protected-access
         )
 
+        input_roles = {conversation.roles[0]}
+        output_roles = {conversation.roles[1]}
+
+        if len(conversation.roles) == 3:
+            tool_role_label = conversation.roles[2]
+            input_roles.add(tool_role_label)
+
+        # Add roles from the config
+        if self.prompter.roles:
+            if "input" in self.prompter.roles and self.prompter.roles["input"]:
+                for role in self.prompter.roles["input"]:
+                    input_roles.add(role)
+
+            if "output" in self.prompter.roles and self.prompter.roles["output"]:
+                for role in self.prompter.roles["output"]:
+                    output_roles.add(role)
+
         # support for custom roles from the dataset, only useful for vicuna style prompts/roles
         role_remap = []
         if (
@@ -360,19 +377,18 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
                     LOG.warning(f"expected tuple, got {part}")
                     continue
 
-                tool_role_label = None
-                if len(conversation.roles) == 3:
-                    (
-                        user_role_label,
-                        assistant_role_label,
-                        tool_role_label,
-                    ) = conversation.roles
-                else:
-                    user_role_label, assistant_role_label = conversation.roles
                 role, content = part
 
                 # Uses "in" because role contains extra characters
-                if user_role_label in role:
+                input_turn = any(r.lower() in role.lower() for r in input_roles)
+                output_turn = any(r.lower() in role.lower() for r in output_roles)
+                empty_role = role.strip() == ""
+
+                if not any([input_turn, output_turn, empty_role]):
+                    LOG.warning(f"unhandled role: {role}")
+                    continue
+
+                if input_turn:
                     role = (
                         role.replace(role_remap[0]["from"], role_remap[0]["to"])
                         if role_remap
@@ -392,7 +408,7 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
                     else:
                         # everything from this is masked out from the labels
                         labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
-                elif assistant_role_label in role:
+                elif output_turn:
                     role = (
                         role.replace(role_remap[1]["from"], role_remap[1]["to"])
                         if role_remap
@@ -423,7 +439,7 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
                         labels[:len_role] = [IGNORE_TOKEN_ID] * min(
                             len_role, len(labels)
                         )
-                elif role == "":
+                elif empty_role:
                     turn = content
                     # this is only ever the first part, should include the bos token and the user query
                     res = self._tokenize(
@@ -434,11 +450,6 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
                     else:
                         # everything from this is masked out from the labels
                         labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
-                elif tool_role_label and tool_role_label in role:
-                    labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
-                else:
-                    LOG.warning(f"unhandled role: {role}")
-                    continue
 
                 # pylint: disable=duplicate-code
                 result, current_len = parse_tokenized_to_result(

src/axolotl/prompters.py

@@ -259,6 +259,12 @@ SHAREGPT_ASSERTION_FAILED_ROLE = (
     "Role did not alternate between turns (gpt and human). Please check your data."
 )
 
+CONVERSATION_ROLE_FORMAT = {
+    "chatml": "<|im_start|>{ROLE}",
+    "zephyr": "<|{ROLE}|>",
+    "vicuna_v1.1": "{ROLE}",
+}
+
 
 class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
     """
@@ -268,7 +274,9 @@ class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
     role_key_human = "human"
     role_key_model = "gpt"
     # Optional, only used for tool usage datasets.
-    role_key_tool = None
+    role_key_tool: Optional[str] = None
+    # Optional, role input/output mapping
+    roles: Optional[dict] = None
 
     def __init__(
         self,
@@ -277,6 +285,7 @@ class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
         role_key_human: Optional[str] = None,
         role_key_model: Optional[str] = None,
         role_key_tool: Optional[str] = None,
+        roles: Optional[dict] = None,
     ):
         if conversation:
             if isinstance(conversation, Conversation):
@@ -291,6 +300,8 @@ class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
             self.role_key_model = role_key_model
         if role_key_tool:
             self.role_key_tool = role_key_tool
+        if roles:
+            self.roles = roles
 
     def _build_result(self, source):
         if len(source) < 2:
@@ -322,11 +333,23 @@ class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
 
         conv.messages = []
         for _, sentence in enumerate(source):
-            role = roles[sentence["from"]]
+            from_role = sentence["from"]
-            if len(conv.messages) > 0 and (
+            if from_role in roles:
-                (role == conv.messages[-1][0]) or (role not in conv.roles)
+                role = roles[from_role]
-            ):
+            else:
+                if self._conversation.name not in CONVERSATION_ROLE_FORMAT:
+                    raise NotImplementedError(
+                        f"Role ({role}) not in default roles, and {self._conversation.name} does not support role remapping yet."
+                        "Please help us by creating an Issue to add support for this conversation type."
+                    )
+
+                role = CONVERSATION_ROLE_FORMAT[self._conversation.name].format(
+                    ROLE=from_role
+                )
+
+            if len(conv.messages) > 0 and ((role == conv.messages[-1][0])):
                 LOG.warning(f"{SHAREGPT_ASSERTION_FAILED_ROLE}: {sentence}")
+
             conv.append_message(role, sentence["value"])
 
         return conv.get_turns()
@@ -354,11 +377,13 @@ class ShareGPTPrompterV2(ShareGPTPrompter):
         conversation: Optional[Union[str, Conversation]] = None,
         role_key_human: Optional[str] = None,
         role_key_model: Optional[str] = None,
+        roles: Optional[dict] = None,
     ):
         super().__init__(
             conversation=conversation,
             role_key_human=role_key_human,
             role_key_model=role_key_model,
+            roles=roles,
         )
 
 

src/axolotl/train.py

@@ -85,7 +85,7 @@ def train(
     model.generation_config.do_sample = True
 
     model_ref = None
-    if cfg.rl:
+    if cfg.rl and cfg.rl != "orpo":
         if cfg.adapter and not cfg.rl_adapter_ref_model:
             # use built-in trl autounwrap
             LOG.debug("Passing model_ref: None to RL trainer")
@@ -110,9 +110,6 @@ def train(
         total_num_steps,
     )
 
-    if hasattr(model, "config"):
-        model.config.use_cache = False
-
     # go ahead and presave, so we have the adapter config available to inspect
     if peft_config:
         LOG.info(f"Pre-saving adapter config to {cfg.output_dir}")

src/axolotl/utils/chat_templates.py

@@ -21,7 +21,7 @@ def chat_templates(user_choice: str):
     templates = {
         "alpaca": "{% for message in messages %}{% if message['role'] == 'user' %}{{ '### Instruction: ' + message['content'] + '\n\n' }}{% elif message['role'] == 'assistant' %}{{ '### Response: ' + message['content'] + eos_token}}{% endif %}{% endfor %}",
         "inst": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",  # I don't know what this one is called. Used by Mistral/Mixtral.
-        "chatml": "{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% set system_message = 'You are a helpful assistant.' %}{% endif %}{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in loop_messages %}{% if loop.index0 == 0 %}{{'<|im_start|>system\n' + system_message + '<|im_end|>\n'}}{% endif %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+        "chatml": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
         "gemma": "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\n' + message['content'] | trim + '<end_of_turn>\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\n'}}{% endif %}",
     }
 

src/axolotl/utils/config/__init__.py

@@ -191,6 +191,11 @@ def normalize_cfg_datasets(cfg):
                         f"updating dataset {ds_cfg.path} with `conversation: chatml` to match your chat_template"
                     )
                     cfg.datasets[idx].conversation = "chatml"
+                if ds_cfg.type == "orpo.chat_template" and not ds_cfg.chat_template:
+                    LOG.info(
+                        f"updating dataset {ds_cfg.path} with `chat_template: chatml` to match your chat_template"
+                    )
+                    cfg.datasets[idx].chat_template = "chatml"
 
 
 def validate_config(cfg: DictDefault, capabilities: Optional[dict] = None):

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -96,6 +96,8 @@ class SFTDataset(BaseModel):
     field_human: Optional[str] = None
     field_model: Optional[str] = None
 
+    roles: Optional[Dict[str, List[str]]] = None
+
 
 class UserDefinedDPOType(BaseModel):
     """User defined typing for DPO"""
@@ -124,6 +126,7 @@ class RLType(str, Enum):
     dpo = "dpo"  # pylint: disable=invalid-name
     ipo = "ipo"  # pylint: disable=invalid-name
     kto_pair = "kto_pair"  # pylint: disable=invalid-name
+    orpo = "orpo"  # pylint: disable=invalid-name
 
 
 class ChatTemplate(str, Enum):
@@ -310,6 +313,15 @@ class HyperparametersConfig(BaseModel):
     learning_rate: Union[str, float]
     weight_decay: Optional[float] = None
     optimizer: Optional[Union[OptimizerNames, Literal["lion_pytorch"]]] = None
+    optim_args: Optional[Union[str, Dict[str, Any]]] = Field(
+        default=None, metadata={"help": "Optional arguments to supply to optimizer."}
+    )
+    optim_target_modules: Optional[Union[List[str], Literal["all_linear"]]] = Field(
+        default=None,
+        metadata={
+            "help": "The target modules to optimize, i.e. the module names that you would like to train."
+        },
+    )
     torchdistx_path: Optional[str] = None
     lr_scheduler: Optional[SchedulerType] = None
     lr_scheduler_kwargs: Optional[Dict[str, Any]] = None
@@ -415,6 +427,7 @@ class AxolotlInputConfig(
 
     datasets: Optional[conlist(Union[SFTDataset, DPODataset], min_length=1)] = None  # type: ignore
     test_datasets: Optional[conlist(Union[SFTDataset, DPODataset], min_length=1)] = None  # type: ignore
+    shuffle_merged_datasets: Optional[bool] = True
     dataset_prepared_path: Optional[str] = None
     dataset_shard_num: Optional[int] = None
     dataset_shard_idx: Optional[int] = None
@@ -431,6 +444,8 @@ class AxolotlInputConfig(
     dataloader_prefetch_factor: Optional[int] = None
     dataloader_drop_last: Optional[bool] = None
 
+    remove_unused_columns: Optional[bool] = None
+
     push_dataset_to_hub: Optional[str] = None
     hf_use_auth_token: Optional[bool] = None
 
@@ -515,6 +530,8 @@ class AxolotlInputConfig(
 
     neftune_noise_alpha: Optional[float] = None
 
+    orpo_alpha: Optional[float] = None
+
     max_memory: Optional[
         Dict[Union[int, Literal["cpu", "disk"]], Union[int, str]]
     ] = None

src/axolotl/utils/data.py

@@ -415,8 +415,11 @@ def load_tokenized_prepared_datasets(
         dataset = concatenate_datasets(datasets)
 
         if len(datasets) > 1:
-            LOG.info("shuffle merged datasets")
+            if cfg.shuffle_merged_datasets:
-            dataset = dataset.shuffle(seed=seed)
+                LOG.debug("shuffle merged datasets")
+                dataset = dataset.shuffle(seed=seed)
+            else:
+                LOG.debug("NOT shuffling merged datasets")
 
         dataset, _ = process_datasets_for_packing(cfg, dataset, None)
 
@@ -819,7 +822,11 @@ def wrap_pretraining_dataset(
     else:
         encode = functools.partial(encode_pretraining, tokenizer, max_tokens)
 
-    dataset = dataset.shuffle(seed=seed, buffer_size=buffer_size)
+    if cfg.shuffle_merged_datasets:
+        dataset = dataset.shuffle(seed=seed, buffer_size=buffer_size)
+    else:
+        LOG.debug("NOT shuffling merged pretraining datasets")
+
     dataset = dataset.map(
         encode,
         batched=True,

src/axolotl/utils/freeze.py

@@ -3,7 +3,7 @@ module to freeze/unfreeze parameters by name
 """
 import logging
 import re
-from typing import Callable, List, Tuple
+from typing import Callable, List, Tuple, Union
 
 from axolotl.utils.distributed import is_main_process
 
@@ -99,7 +99,7 @@ def _invert_ranges(
 
 
 def _merge_ranges(
-    given_ranges: List[Tuple[int, int | None]], layer_size: int
+    given_ranges: List[Tuple[int, Union[int, None]]], layer_size: int
 ) -> List[Tuple[int, int]]:
     """
     Merges overlapping ranges and sorts the given ranges.
@@ -194,7 +194,9 @@ class LayerNamePattern:
         """
         return self.name_regex.match(name) is not None
 
-    def _parse_pattern(self, pattern: str) -> Tuple[str, Tuple[int, int | None] | None]:
+    def _parse_pattern(
+        self, pattern: str
+    ) -> Tuple[str, Union[Tuple[int, Union[int, None]], None]]:
         """
         Extracts the range pattern from the given pattern.
 

src/axolotl/utils/models.py

@@ -888,7 +888,9 @@ def load_model(
 
     if cfg.adapter in ["lora", "qlora"]:
         if cfg.gradient_checkpointing:
-            model.gradient_checkpointing_enable()
+            model.gradient_checkpointing_enable(
+                gradient_checkpointing_kwargs=cfg.gradient_checkpointing_kwargs
+            )
         if (
             cfg.load_in_8bit or cfg.load_in_4bit
         ) and not skip_prepare_model_for_kbit_training:

tests/prompt_strategies/test_sharegpt.py

@@ -62,6 +62,38 @@ def fixture_sharegpt_glaive_dataset():
     )
 
 
+@pytest.fixture(name="multi_role_dataset")
+def fixture_multi_role_dataset():
+    return Dataset.from_list(
+        [
+            {
+                "conversations": [
+                    {
+                        "from": "system",
+                        "value": "use get_weather(city) to get the weather for a city",
+                    },
+                    {
+                        "from": "human",
+                        "value": "hello, what's the weather in New York?",
+                    },
+                    {
+                        "from": "gpt",
+                        "value": "let me get that for you",
+                    },
+                    {
+                        "from": "tool",
+                        "value": "get_weather(New York)",
+                    },
+                    {
+                        "from": "gpt",
+                        "value": "the weather in New York is 70 degrees and sunny",
+                    },
+                ]
+            }
+        ]
+    )
+
+
 @pytest.fixture(name="tokenizer")
 def fixture_tokenizer():
     tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
@@ -196,3 +228,39 @@ class TestSharegpt:
             32001, 13892, 13, 28737, 28742, 28719, 7371, 28725, 562, 315, 949, 28742, 28707, 506, 272, 21368, 298, 1820, 22447, 28723, 28705, 523, 28766, 416, 1009, 772, 28766, 28767, 32000, 28705, 13  # gpt
         ]
         # fmt: on
+
+    def test_multi_role_dataset(self, multi_role_dataset, tokenizer):
+        strategy = SimpleShareGPTPromptTokenizingStrategy(
+            ShareGPTPrompterV2(conversation="chatml", roles={"input": ["tool"]}),
+            tokenizer,
+            False,  # train_on_inputs
+            2048,  # sequence_len
+        )
+
+        dataset_wrapper = TokenizedPromptDataset(
+            strategy, multi_role_dataset, process_count=1
+        )
+
+        input_ids = dataset_wrapper[0]["input_ids"]
+        # fmt: off
+        assert input_ids == [
+            1,   # bos
+            32001, 1587, 13, 1730, 625, 28730, 769, 1223, 28732, 18373, 28731, 298, 625, 272, 8086, 354, 264, 2990, 32000, 28705, 13,  # system
+            32001, 2188, 13, 21558, 28725, 767, 28742, 28713, 272, 8086, 297, 1450, 2726, 28804, 32000, 28705, 13,  # human
+            32001, 13892, 13, 895, 528, 625, 369, 354, 368, 32000, 28705, 13,  # gpt
+            32001, 3921, 13, 527, 28730, 769, 1223, 28732, 2972, 2726, 28731, 32000, 28705, 13,  # tool
+            32001, 13892, 13, 1237, 8086, 297, 1450, 2726, 349, 28705, 28787, 28734, 11182, 304, 4376, 1780, 32000, 28705, 13  # gpt
+        ]
+        # fmt: on
+
+        labels = dataset_wrapper[0]["labels"]
+        # fmt: off
+        assert labels == [
+            -100,  # bos
+            -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,  # system
+            -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,  # human
+            -100, -100, 13, 895, 528, 625, 369, 354, 368, 32000, 28705, 13,  # gpt
+            -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,  # tool
+            -100, -100, 13, 1237, 8086, 297, 1450, 2726, 349, 28705, 28787, 28734, 11182, 304, 4376, 1780, 32000, 28705, 13  # gpt
+        ]
+        # fmt: on

tests/test_prompt_tokenizers.py

@@ -8,7 +8,8 @@ from pathlib import Path
 from typing import Optional
 
 import pytest
-from transformers import AutoTokenizer, LlamaTokenizer
+from datasets import load_dataset
+from transformers import AddedToken, AutoTokenizer, LlamaTokenizer
 
 from axolotl.prompt_strategies.alpaca_chat import NoSystemPrompter
 from axolotl.prompt_strategies.alpaca_w_system import (
@@ -19,12 +20,14 @@ from axolotl.prompt_strategies.llama2_chat import (
     Llama2ChatPrompter,
     LLama2ChatTokenizingStrategy,
 )
+from axolotl.prompt_strategies.orpo.chat_template import load
 from axolotl.prompt_strategies.sharegpt import GlaiveShareGPTPromptTokenizingStrategy
 from axolotl.prompt_tokenizers import (
     AlpacaPromptTokenizingStrategy,
     ShareGPTPromptTokenizingStrategy,
 )
 from axolotl.prompters import AlpacaPrompter, PromptStyle, ShareGPTPrompterV2
+from axolotl.utils.dict import DictDefault
 
 LOG = logging.getLogger("axolotl")
 
@@ -446,5 +449,57 @@ If a question does not make any sense, or is not factually coherent, explain why
         )
 
 
+class OrpoTokenizationTest(unittest.TestCase):
+    """test case for the ORPO tokenization"""
+
+    def setUp(self) -> None:
+        # pylint: disable=duplicate-code
+        tokenizer = LlamaTokenizer.from_pretrained("mistralai/Mistral-7B-v0.1")
+        tokenizer.add_special_tokens(
+            {
+                "eos_token": AddedToken(
+                    "<|im_end|>", rstrip=False, lstrip=False, normalized=False
+                )
+            }
+        )
+        tokenizer.add_tokens(
+            [
+                AddedToken(
+                    "<|im_start|>", rstrip=False, lstrip=False, normalized=False
+                ),
+            ]
+        )
+        self.tokenizer = tokenizer
+        self.dataset = load_dataset(
+            "argilla/ultrafeedback-binarized-preferences-cleaned", split="train"
+        ).select([0])
+
+    def test_orpo_integration(self):
+        strat = load(
+            self.tokenizer,
+            DictDefault({"train_on_inputs": False}),
+            DictDefault({"chat_template": "chatml"}),
+        )
+        res = strat.tokenize_prompt(self.dataset[0])
+        assert "rejected_input_ids" in res
+        assert "rejected_labels" in res
+        assert "input_ids" in res
+        assert "labels" in res
+        assert "prompt_attention_mask" in res
+
+        assert len(res["rejected_input_ids"]) == len(res["rejected_labels"])
+        assert len(res["input_ids"]) == len(res["labels"])
+        assert len(res["input_ids"]) == len(res["prompt_attention_mask"])
+
+        assert res["rejected_labels"][0] == -100
+        assert res["rejected_input_ids"][-1] == res["rejected_labels"][-1]
+
+        assert res["labels"][0] == -100
+        assert res["input_ids"][-1] == res["labels"][-1]
+
+        assert res["prompt_attention_mask"][0] == 1
+        assert res["prompt_attention_mask"][-1] == 0
+
+
 if __name__ == "__main__":
     unittest.main()