fix import

optimizer_cls_and_kwargs is on trainer_kwargs
only add adamw_kwargs if they're non-null
fix mocks
better handling of override and check the optimizer
unwrap optimizer

README.md

@@ -19,9 +19,6 @@
     <br/>
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests-nightly.yml/badge.svg" alt="tests-nightly">
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/multi-gpu-e2e.yml/badge.svg" alt="multigpu-semi-weekly tests">
-    <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>
 </p>
 
 Axolotl is a tool designed to streamline post-training for various AI models.

_quarto.yml

@@ -40,6 +40,7 @@ website:
 
         - section: "Deployments"
           contents:
+            - docs/docker.qmd
             - docs/multi-gpu.qmd
             - docs/multi-node.qmd
             - docs/ray-integration.qmd

docs/config.qmd

@@ -163,6 +163,12 @@ datasets:
       system: ["system"]
       tool: ["tool"]
 
+    # Optional[bool]. Whether to drop the system turn from the dataset. Only works with chat_template.
+    # This does not drop the default system message from chat_template if it exists. If you wish to,
+    # we recommend using a custom jinja template with the default system message removed or
+    # adding a system turn with empty content.
+    drop_system_message:
+
     # IMPORTANT: The following fields determine which parts of the conversation to train on.
     # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train
     # See examples at `docs/dataset-formats/conversation.qmd`
@@ -222,8 +228,8 @@ process_reward_model:
 chat_template: tokenizer_default
 # custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null.
 chat_template_jinja: null
-# Changes the default system message
+# Changes the default system message. Currently only supports chatml.
-default_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.
+default_system_message: You are a helpful assistant. Please give a long and detailed answer.
 # Axolotl attempts to save the dataset as an arrow after packing the data together so
 # subsequent training attempts load faster, relative path
 dataset_prepared_path: data/last_run_prepared
@@ -445,7 +451,7 @@ gradient_checkpointing: false
 early_stopping_patience: 3
 
 # Specify a scheduler and kwargs to use with the optimizer
-lr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine
+lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine
 lr_scheduler_kwargs:
 cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr
 cosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)
@@ -528,6 +534,8 @@ flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
 sdp_attention:
 # Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf
 s2_attention:
+# Optional[bool]. Whether to use low_cpu_mem_usage
+low_cpu_mem_usage:
 # Resume from a specific checkpoint dir
 resume_from_checkpoint:
 # If resume_from_checkpoint isn't set and you simply want it to start where it left off.

docs/dataset-formats/index.qmd

@@ -129,6 +129,7 @@ You can mix and match within each approach or across approaches to train a model
 We suggest this approach when you want to bring your own tokenized dataset.
 
 Axolotl expects the dataset to have three keys:
+
 - `input_ids`: from tokenizing formatted prompt
 - `attention_mask`: for masking padding. If you don't add padding, it would be equal to `len(input_ids) * [1]`
 - `labels`: this is the same as `input_ids`, however, if you want to mask certain tokens, you would set those indices to `-100`.

docs/docker.qmd

@@ -0,0 +1,140 @@
+---
+title: "Docker"
+format:
+  html:
+    toc: true
+    toc-depth: 4
+---
+
+This section describes the different Docker images that are released by AxolotlAI at [Docker Hub](https://hub.docker.com/u/axolotlai).
+
+## Base
+
+The base image is the most minimal image that can install Axolotl. It is based on the `nvidia/cuda` image. It includes python, torch, git, git-lfs, awscli, pydantic, and more.
+
+#### Image
+
+```
+axolotlai/axolotl-base
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl-base)
+
+#### Tags format
+
+```bash
+main-base-py{python_version}-cu{cuda_version}-{pytorch_version}
+```
+
+Tags examples:
+
+- `main-base-py3.11-cu124-2.6.0`
+- `main-base-py3.11-cu124-2.5.1`
+- `main-base-py3.11-cu124-2.4.1`
+
+## Main
+
+The main image is the image that is used to run Axolotl. It is based on the `axolotlai/axolotl-base` image and includes the Axolotl codebase, dependencies, and more.
+
+#### Image
+
+```
+axolotlai/axolotl
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl)
+
+#### Tags format {#sec-main-tags}
+
+```bash
+# on push to main
+main-py{python_version}-cu{cuda_version}-{pytorch_version}
+
+# latest main (currently torch 2.5.1, python 3.11, cuda 12.4)
+main-latest
+
+# nightly build
+{branch}-{date_in_YYYYMMDD}-py{python_version}-cu{cuda_version}-{pytorch_version}
+
+# tagged release
+{version}
+```
+
+:::{.callout-tip}
+
+There may be some extra tags appended to the image, like `-vllm` which installs those packages.
+
+:::
+
+Tags examples:
+
+- `main-py3.11-cu124-2.6.0`
+- `main-py3.11-cu124-2.5.1`
+- `main-py3.11-cu124-2.4.1`
+- `main-latest`
+- `main-20250303-py3.11-cu124-2.6.0`
+- `main-20250303-py3.11-cu124-2.5.1`
+- `main-20250303-py3.11-cu124-2.4.1`
+- `0.7.1`
+
+## Cloud
+
+The cloud image is the image that is used to run Axolotl in the cloud. It is based on the `axolotlai/axolotl` image and sets ENV variables like HuggingFace cache directories for volume mounts, tmux, and more for different cloud providers.
+
+:::{.callout-tip}
+
+Jupyter lab is run by default. Set `JUPYTER_DISABLE=1` in the environment variables to disable it.
+
+:::
+
+#### Image
+
+```
+axolotlai/axolotl-cloud
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl-cloud)
+
+#### Tags format
+
+This uses the same tags as the [`main` image](#sec-main-tags).
+
+#### Environment variables
+
+- `JUPYTER_DISABLE`: Disable Jupyter lab.
+- `JUPYTER_PASSWORD`: Set a password for the Jupyter lab.
+- `PUBLIC_KEY`: Add a public key for the SSH service.
+- `SSH_KEY`: Add a private key for the SSH service.
+
+#### Volume mounts
+
+:::{.callout-tip}
+
+We recommend mounting volumes to `/workspace/data` for data persistence. `/workspace/axolotl` contains the source code and is ephemeral.
+
+:::
+
+- `/workspace/data/axolotl-artifacts`: Directory to store Axolotl artifacts.
+- `/workspace/data/huggingface-cache`: Directory to store HuggingFace cache.
+
+## Cloud-no-tmux
+
+This is the same as the [`cloud` image](#sec-cloud) but without tmux.
+
+#### Image
+
+```
+axolotlai/axolotl-cloud-term
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl-cloud-term)
+
+:::{.callout-note}
+
+The naming may be a bit confusing as it has `-term` appended to the end.
+
+:::
+
+#### Tags format
+
+This uses the same tags as the [`cloud` image](#sec-cloud-tags).

docs/faq.qmd

@@ -19,7 +19,9 @@ description: Frequently asked questions
 
 **Q: AttributeError: 'DummyOptim' object has no attribute 'step'**
 
-> A: You may be using deepspeed with single gpu. Please don't set `deepspeed:` in yaml or cli.
+**Q: ModuleNotFoundError: No module named 'mpi4py' using single GPU with deepspeed**
+
+> A: You may be using deepspeed with single gpu. Please remove the `deepspeed:` section in the yaml file or `--deepspeed` CLI flag.
 
 **Q: The codes is stuck on saving preprocessed datasets.**
 

docs/installation.qmd

@@ -65,6 +65,8 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it \
 ```
 :::
 
+Please refer to the [Docker documentation](docker.qmd) for more information on the different Docker images that are available.
+
 ## Cloud Environments {#sec-cloud}
 
 ### Cloud GPU Providers {#sec-cloud-gpu}

requirements.txt

@@ -63,3 +63,4 @@ torchao==0.7.0
 schedulefree==1.3.0
 
 axolotl-contribs-lgpl==0.0.3
+axolotl-contribs-mit==0.0.3

src/axolotl/cli/train.py

@@ -41,11 +41,12 @@ def do_train(cfg: DictDefault, cli_args: TrainerCliArgs) -> None:
     else:
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-    model, tokenizer = train(cfg=cfg, dataset_meta=dataset_meta)
+    model, tokenizer, trainer = train(cfg=cfg, dataset_meta=dataset_meta)
     plugin_manager = PluginManager.get_instance()
 
     del model
     del tokenizer
+    del trainer
 
     plugin_manager.post_train_unload(cfg)
 

src/axolotl/core/trainer_builder.py

@@ -35,6 +35,7 @@ from transformers import (
     EarlyStoppingCallback,
     TrainerCallback,
 )
+from transformers.training_args import OptimizerNames
 from trl.trainer.utils import RewardDataCollatorWithPadding
 
 from axolotl.core.trainers.base import (
@@ -84,6 +85,7 @@ from axolotl.utils.collators import (
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
 from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
+from axolotl.utils.config.models.input.v0_4_1 import CustomSupportedOptimizers
 from axolotl.utils.models import ensure_dtype
 
 try:
@@ -549,30 +551,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["run_name"] = self.cfg.mlflow_run_name
         else:
             training_arguments_kwargs["run_name"] = None
-        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"
-        ] = self.cfg.loraplus_lr_embedding
-        training_arguments_kwargs["embedding_lr"] = self.cfg.embedding_lr
-        training_arguments_kwargs["embedding_lr_scale"] = self.cfg.embedding_lr_scale
-        training_arguments_kwargs["lr_groups"] = self.cfg.lr_groups
 
-        if self.cfg.lr_scheduler in ["one_cycle", "log_sweep"]:
+        if self.cfg.lr_scheduler in ["one_cycle", "rex", "log_sweep"]:
             training_arguments_kwargs["lr_scheduler_type"] = "cosine"
             training_arguments_kwargs[
                 "alternate_lr_scheduler_type"
@@ -656,46 +636,114 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.reward_model:
             training_arguments_kwargs["max_length"] = self.cfg.sequence_len
 
-        # pylint: disable=duplicate-code
+        # Handle custom optimizer
-        if self.cfg.optimizer in [
+        custom_supported_optimizers = [opt.value for opt in CustomSupportedOptimizers]
-            "optimi_adamw",
+        if self.cfg.optimizer in custom_supported_optimizers:
-            "ao_adamw_4bit",
+            # Common optimizer kwargs
-            "ao_adamw_8bit",
+            optimizer_kwargs = {
-            "ao_adamw_fp8",
+                "lr": training_arguments_kwargs.get("learning_rate"),
-            "adopt_adamw",
+                "weight_decay": training_arguments_kwargs.get("weight_decay"),
-        ]:
+            }
-            # Set default so transformers doesn't throw
-            training_arguments_kwargs["optim"] = "adamw_hf"
-            training_arguments_kwargs["alternate_optimizer"] = self.cfg.optimizer
 
-        if self.cfg.optimizer == "lion_pytorch":
+            # Adam-specific kwargs
-            from lion_pytorch import Lion
+            adam_kwargs = {}
+            if training_arguments_kwargs.get(
+                "adam_beta1"
+            ) and training_arguments_kwargs.get("adam_beta2"):
+                adam_kwargs["betas"] = (
+                    training_arguments_kwargs.get("adam_beta1"),
+                    training_arguments_kwargs.get("adam_beta2"),
+                )
+            if training_arguments_kwargs.get("adam_epsilon"):
+                adam_kwargs["eps"] = training_arguments_kwargs.get("adam_epsilon")
 
-            lion_kwargs = {"lr": training_arguments_kwargs["learning_rate"]}
+            if self.cfg.optimizer == "muon":
-            if "weight_decay" in training_arguments_kwargs:
+                from axolotl.contribs.mit.muon import (  # pylint: disable=no-name-in-module
-                lion_kwargs["weight_decay"] = training_arguments_kwargs["weight_decay"]
+                    MuonOptimizerFactory,
-
-            if (
-                "adam_beta1" in training_arguments_kwargs
-                and "adam_beta2" in training_arguments_kwargs
-            ):
-                lion_kwargs["betas"] = (
-                    training_arguments_kwargs["adam_beta1"],
-                    training_arguments_kwargs["adam_beta2"],
                 )
 
-            trainer_kwargs["optimizers"] = (
+                optimizer_cls = MuonOptimizerFactory
-                Lion(params=self.model.parameters(), **lion_kwargs),
+                optimizer_kwargs.update(adam_kwargs)
-                None,
+            elif self.cfg.optimizer == "optimi_adamw":
+                from optimi import AdamW
+
+                optimizer_kwargs["foreach"] = False
+                optimizer_cls = AdamW
+                optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "ao_adamw_4bit":
+                # TODO remove 20250401
+                from torchao.prototype.low_bit_optim import AdamW4bit
+
+                optimizer_cls = AdamW4bit
+                optimizer_kwargs.update(adam_kwargs)
+
+                LOG.warning(
+                    f"`ao_adamw_4bit` will be deprecated soon. Please use `{OptimizerNames.ADAMW_TORCH_4BIT}` instead."
+                )
+            elif self.cfg.optimizer == "ao_adamw_8bit":
+                from torchao.prototype.low_bit_optim import AdamW8bit
+
+                optimizer_cls = AdamW8bit
+                optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "ao_adamw_fp8":
+                from torchao.prototype.low_bit_optim import AdamWFp8
+
+                optimizer_cls = AdamWFp8
+                optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "adopt_adamw":
+                from axolotl.utils.optimizers.adopt import ADOPT
+
+                optimizer_cls = ADOPT
+                adam_kwargs["decouple"] = True
+                optimizer_kwargs.update(adam_kwargs)
+
+            # Parse any additional optimizer args from config
+            if self.cfg.optim_args:
+                if isinstance(self.cfg.optim_args, dict):
+                    optimizer_kwargs.update(self.cfg.optim_args)
+                else:
+                    # Parse string format "key1=value1,key2=value2"
+                    for mapping in self.cfg.optim_args.replace(" ", "").split(","):
+                        key, value = mapping.split("=")
+                        optimizer_kwargs[key] = value
+
+            trainer_kwargs["optimizer_cls_and_kwargs"] = (
+                optimizer_cls,
+                optimizer_kwargs,
             )
-            # Set default so transformers doesn't throw
+        else:
-            training_arguments_kwargs["optim"] = "adamw_hf"
+            # Use transformers' optimizer
+            training_arguments_kwargs["optim"] = self.cfg.optimizer
+
+            # Parse any additional optimizer args from config
+            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.optimizer == "adamw_anyprecision":
             if Path(self.cfg.torchdistx_path).exists():
                 sys.path.append(self.cfg.torchdistx_path)
                 importlib.import_module("torchdistx")
 
+        if self.cfg.optim_target_modules:
+            training_arguments_kwargs[
+                "optim_target_modules"
+            ] = self.cfg.optim_target_modules
+
+        training_arguments_kwargs["embedding_lr"] = self.cfg.embedding_lr
+        training_arguments_kwargs["embedding_lr_scale"] = self.cfg.embedding_lr_scale
+
+        training_arguments_kwargs["loraplus_lr_ratio"] = self.cfg.loraplus_lr_ratio
+        training_arguments_kwargs[
+            "loraplus_lr_embedding"
+        ] = self.cfg.loraplus_lr_embedding
+        training_arguments_kwargs["lr_groups"] = self.cfg.lr_groups
+
         if self.cfg.accelerator_config:
             training_arguments_kwargs[
                 "accelerator_config"

src/axolotl/core/trainers/base.py

@@ -14,6 +14,7 @@ from typing import Dict, Literal, Optional
 import torch
 from datasets import Dataset
 from peft.optimizers import create_loraplus_optimizer
+from torch import nn
 from torch.optim.lr_scheduler import OneCycleLR
 from torch.utils.data import BatchSampler, DataLoader, RandomSampler, SequentialSampler
 from transformers import Trainer
@@ -22,9 +23,11 @@ from transformers.utils import is_sagemaker_mp_enabled
 from trl import CPOTrainer, KTOTrainer, ORPOTrainer, PRMTrainer, RewardTrainer
 from trl.trainer.utils import pad_to_length
 
+from axolotl.integrations.base import BaseOptimizerFactory
 from axolotl.monkeypatch.relora import ReLoRAScheduler
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 from axolotl.utils.schedulers import (
+    RexLR,
     get_cosine_schedule_with_min_lr,
     get_cosine_schedule_with_quadratic_warmup,
     get_cosine_schedule_with_warmup_decay_constant,
@@ -115,6 +118,17 @@ class SchedulerMixin(Trainer):
                     **extra_lr_kwargs,
                     **self.args.lr_scheduler_kwargs,
                 )
+            elif self.args.alternate_lr_scheduler_type == "rex":
+                if use_cosine_min_lr:
+                    assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
+
+                self.lr_scheduler = RexLR(
+                    optimizer=optimizer,
+                    max_lr=self.args.learning_rate,
+                    min_lr=0 if not use_cosine_min_lr else (self.args.learning_rate * self.args.cosine_min_lr_ratio),
+                    total_steps=num_training_steps,
+                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
+                )
             elif use_cosine_quadratic:
                 if use_cosine_min_lr:
                     LOG.warning("Both cosine quadratic warmup and min lr detected. Using quadratic warmup.")
@@ -154,47 +168,18 @@ class SchedulerMixin(Trainer):
         return self.lr_scheduler
 
 
-class AxolotlTrainer(SchedulerMixin, Trainer):
+class OptimizerMixin(Trainer):
     """
-    Extend the base Trainer for axolotl helpers
+    Mixin class for shared handling of building custom optimizers
     """
 
     args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
-    tag_names = ["axolotl"]
 
-    def __init__(
+    def create_optimizer_grouped_parameters(
-        self,
+        self, opt_model, optimizer_kwargs
-        *_args,
+    ) -> list[dict]:
-        bench_data_collator=None,
-        eval_data_collator=None,
-        dataset_tags=None,
-        **kwargs,
-    ):
-        self.bench_data_collator = bench_data_collator
-        self.eval_data_collator = eval_data_collator
-        self.dataset_tags = dataset_tags
-        self._signature_columns = None  # workaround for pylint
-        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")
-
-    def _wrap_model(self, model, training=True, dataloader=None):
-        if self.args.torch_compile:
-            torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
-                256
-            )
-            model = torch.compile(
-                model,
-                backend=self.args.torch_compile_backend,
-                mode=self.args.torch_compile_mode,
-            )
-        return super()._wrap_model(model, training=training, dataloader=dataloader)
-
-    def create_optimizer_grouped_parameters(self, opt_model, optimizer_kwargs):
         decay_parameters = self.get_decay_parameter_names(opt_model)
-        params = {
+        params: dict = {
             "to_weight_decay": {},  # LayerNorm and bias
             "embeddings": {},  # lm_head, embed_tokens,
             "no_weight_decay": {},
@@ -281,23 +266,30 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
             and self.args.embedding_lr_scale is None
             and self.args.embedding_lr is None
             and self.args.lr_groups is None
-            and self.args.alternate_optimizer
+            and self.optimizer_cls_and_kwargs is None
-            not in [
-                "optimi_adamw",
-                "ao_adamw_8bit",
-                "ao_adamw_4bit",
-                "ao_adamw_fp8",
-                "adopt_adamw",
-            ]
         ):
             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(
+        if (
-                self.args,
+            not self.optimizer
-                opt_model,
+            and self.optimizer_cls_and_kwargs is not None
+            and issubclass(self.optimizer_cls_and_kwargs[0], BaseOptimizerFactory)
+        ):
+            optimizer_factory_cls, optimizer_kwargs = self.optimizer_cls_and_kwargs
+            self.optimizer = optimizer_factory_cls()(
+                opt_model, self.args, **optimizer_kwargs
             )
+
+        if not self.optimizer:
+            if self.optimizer_cls_and_kwargs is not None:
+                optimizer_cls, optimizer_kwargs = self.optimizer_cls_and_kwargs
+            else:
+                optimizer_cls, optimizer_kwargs = self.get_optimizer_cls_and_kwargs(
+                    self.args, opt_model
+                )
+
             optimizer_grouped_parameters = self.create_optimizer_grouped_parameters(
                 opt_model, optimizer_kwargs
             )
@@ -314,50 +306,47 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
                     loraplus_lr_embedding=loraplus_lr_embedding,
                     **optimizer_kwargs,
                 )
-            elif (
+            else:
-                self.args.embedding_lr_scale is not None
+                # Overwrite `params` in case it's created by `get_optimizer_cls_and_kwargs`
-                or self.args.embedding_lr is not None
+                # e.g. for GaLore optimizer.
-                or self.args.lr_groups is not None
+                if "params" in optimizer_kwargs:
-            ):
+                    optimizer_grouped_parameters = optimizer_kwargs.pop("params")
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
-                )
-            elif self.args.alternate_optimizer == "optimi_adamw":
-                from optimi import AdamW
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
+                # Overwrite `model` in case it's created by `get_optimizer_cls_and_kwargs`
-                    AdamW(
+                # e.g. for LOMO optimizer.
-                        optimizer_grouped_parameters, foreach=False, **optimizer_kwargs
+                if "model" in optimizer_kwargs:
+                    optimizer_grouped_parameters = optimizer_kwargs.pop("model")
+
+                # For layer-wise dummy optimizers we overwrite optimizer_grouped_parameters with `optimizer_dict`
+                # to avoid arguments conflicts.
+                if "optimizer_dict" in optimizer_kwargs:
+                    optimizer_grouped_parameters = optimizer_kwargs.pop(
+                        "optimizer_dict"
                     )
-                )
-            elif self.args.alternate_optimizer == "ao_adamw_4bit":
-                from torchao.prototype.low_bit_optim import AdamW4bit
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
+                self.optimizer = optimizer_cls(
-                    AdamW4bit(optimizer_grouped_parameters, **optimizer_kwargs)
+                    optimizer_grouped_parameters, **optimizer_kwargs
                 )
-            elif self.args.alternate_optimizer == "ao_adamw_8bit":
-                from torchao.prototype.low_bit_optim import AdamW8bit
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
+            if optimizer_cls.__name__ == "Adam8bit":
-                    AdamW8bit(optimizer_grouped_parameters, **optimizer_kwargs)
+                import bitsandbytes
-                )
-            elif self.args.alternate_optimizer == "ao_adamw_fp8":
-                from torchao.prototype.low_bit_optim import AdamWFp8
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
+                manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
-                    AdamWFp8(optimizer_grouped_parameters, **optimizer_kwargs)
-                )
-            elif self.args.alternate_optimizer == "adopt_adamw":
-                from axolotl.utils.optimizers.adopt import ADOPT
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
+                skipped = 0
-                    ADOPT(
+                for module in opt_model.modules():
-                        optimizer_grouped_parameters,
+                    if isinstance(module, nn.Embedding):
-                        decouple=True,
+                        skipped += sum(
-                        **optimizer_kwargs,
+                            {
-                    )
+                                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
@@ -366,6 +355,45 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
 
         return self.optimizer
 
+
+class AxolotlTrainer(SchedulerMixin, OptimizerMixin, Trainer):
+    """
+    Extend the base Trainer for axolotl helpers
+    """
+
+    args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
+    tag_names = ["axolotl"]
+
+    def __init__(
+        self,
+        *_args,
+        bench_data_collator=None,
+        eval_data_collator=None,
+        dataset_tags=None,
+        **kwargs,
+    ):
+        self.bench_data_collator = bench_data_collator
+        self.eval_data_collator = eval_data_collator
+        self.dataset_tags = dataset_tags
+        self._signature_columns = None  # workaround for pylint
+        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")
+
+    def _wrap_model(self, model, training=True, dataloader=None):
+        if self.args.torch_compile:
+            torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
+                256
+            )
+            model = torch.compile(
+                model,
+                backend=self.args.torch_compile_backend,
+                mode=self.args.torch_compile_mode,
+            )
+        return super()._wrap_model(model, training=training, dataloader=dataloader)
+
     def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
         if self.args.sample_packing and not self.args.pretraining:
             if self.args.multipack_real_batches:

src/axolotl/core/trainers/grpo/__init__.py

@@ -9,6 +9,7 @@ import logging
 from trl.trainer.grpo_trainer import RewardFunc
 
 from axolotl.core.trainers.grpo.trainer import AxolotlGRPOTrainer
+from axolotl.utils.config.models.input.v0_4_1.trl import TRLConfig
 
 LOG = logging.getLogger("axolotl")
 
@@ -31,30 +32,44 @@ class GRPOStrategy:
     @classmethod
     def set_training_args_kwargs(cls, cfg):
         grpo_args_kwargs = {}
-        if cfg.trl and cfg.trl.use_vllm:
+
-            grpo_args_kwargs["use_vllm"] = cfg.trl.use_vllm
+        if not hasattr(cfg, "trl") or not cfg.trl:
-            if cfg.trl and cfg.trl.vllm_device:
+            return grpo_args_kwargs
-                grpo_args_kwargs["vllm_device"] = cfg.trl.vllm_device
+
-            else:
+        trl: TRLConfig = cfg.trl  # type: ignore
-                grpo_args_kwargs["vllm_device"] = "auto"
+
-            if cfg.trl and cfg.trl.vllm_gpu_memory_utilization:
+        if trl.use_vllm:
+            grpo_args_kwargs["use_vllm"] = trl.use_vllm
+            grpo_args_kwargs["vllm_device"] = (
+                trl.vllm_device if trl.vllm_device else "auto"
+            )
+
+            if trl.vllm_gpu_memory_utilization:
                 grpo_args_kwargs[
                     "vllm_gpu_memory_utilization"
-                ] = cfg.trl.vllm_gpu_memory_utilization
+                ] = trl.vllm_gpu_memory_utilization
-            if cfg.trl and cfg.trl.vllm_max_model_len:
+
-                grpo_args_kwargs["vllm_max_model_len"] = cfg.trl.vllm_max_model_len
+            if trl.vllm_max_model_len:
-        if cfg.trl and cfg.trl.num_generations:
+                grpo_args_kwargs["vllm_max_model_len"] = trl.vllm_max_model_len
-            grpo_args_kwargs["num_generations"] = cfg.trl.num_generations
+
-        if cfg.trl and cfg.trl.sync_ref_model:
+        if trl.num_generations:
-            grpo_args_kwargs["sync_ref_model"] = cfg.trl.sync_ref_model
+            grpo_args_kwargs["num_generations"] = trl.num_generations
-            if cfg.trl and cfg.trl.ref_model_mixup_alpha:
+
-                grpo_args_kwargs[
+        if trl.sync_ref_model:
-                    "ref_model_mixup_alpha"
+            grpo_args_kwargs["sync_ref_model"] = trl.sync_ref_model
-                ] = cfg.trl.ref_model_mixup_alpha
+
-            if cfg.trl and cfg.trl.ref_model_sync_steps:
+            if trl.ref_model_mixup_alpha:
-                grpo_args_kwargs["ref_model_sync_steps"] = cfg.trl.ref_model_sync_steps
+                grpo_args_kwargs["ref_model_mixup_alpha"] = trl.ref_model_mixup_alpha
-        grpo_args_kwargs["max_completion_length"] = cfg.trl.max_completion_length
+
-        grpo_args_kwargs["log_completions"] = cfg.trl.log_completions
+            if trl.ref_model_sync_steps:
+                grpo_args_kwargs["ref_model_sync_steps"] = trl.ref_model_sync_steps
+
+        grpo_args_kwargs["max_completion_length"] = trl.max_completion_length
+        grpo_args_kwargs["log_completions"] = trl.log_completions
+
+        if trl.reward_weights:
+            grpo_args_kwargs["reward_weights"] = trl.reward_weights
+
         return grpo_args_kwargs
 
     @classmethod

src/axolotl/integrations/base.py

@@ -23,6 +23,8 @@ import importlib
 import logging
 from typing import OrderedDict
 
+import torch
+
 
 class BasePlugin:
     """
@@ -469,3 +471,14 @@ class PluginManager:
         """
         for plugin in self.plugins.values():
             plugin.post_train_unload(cfg)
+
+
+class BaseOptimizerFactory:
+    """
+    Base class for factories to create custom optimizers
+    """
+
+    def __call__(
+        self, opt_model, training_args, **optimizer_kwargs
+    ) -> "torch.optim.Optimizer":
+        pass

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -4,6 +4,22 @@ Cut Cross Entropy reduces VRAM usage through optimization on the cross-entropy o
 
 See https://github.com/apple/ml-cross-entropy
 
+## Requirements
+
+- PyTorch 2.4.0 or higher
+
+## Installation
+
+Run the following command to install `cut_cross_entropy[transformers]` if you don't have it already.
+
+```bash
+# if you are in dev environment
+python scripts/cutcrossentropy_install.py | sh
+
+# if you are not in dev environment
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy @ git+https://github.com/apple/ml-cross-entropy.git@9c297c905f55b73594b5d650722d1e78183b77bd"'
+```
+
 ## Usage
 
 ```yaml

src/axolotl/train.py

@@ -461,7 +461,7 @@ def setup_model_and_trainer(
 
 def train(
     cfg: DictDefault, dataset_meta: TrainDatasetMeta
-) -> tuple[PeftModel | PreTrainedModel, PreTrainedTokenizer]:
+) -> tuple[PeftModel | PreTrainedModel, PreTrainedTokenizer, Trainer]:
     """
     Train a model on the given dataset.
 
@@ -510,4 +510,4 @@ def train(
     # Create model card
     create_model_card(cfg, trainer)
 
-    return model, tokenizer
+    return model, tokenizer, trainer

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

@@ -64,6 +64,18 @@ class ChatTemplate(str, Enum):
     metharme = "metharme"  # pylint: disable=invalid-name
 
 
+class CustomSupportedOptimizers(str, Enum):
+    """Custom supported optimizers"""
+
+    optimi_adamw = "optimi_adamw"  # pylint: disable=invalid-name
+    ao_adamw_4bit = "ao_adamw_4bit"  # pylint: disable=invalid-name
+    ao_adamw_8bit = "ao_adamw_8bit"  # pylint: disable=invalid-name
+    ao_adamw_fp8 = "ao_adamw_fp8"  # pylint: disable=invalid-name
+    adopt_adamw = "adopt_adamw"  # pylint: disable=invalid-name
+    lion_pytorch = "lion_pytorch"  # pylint: disable=invalid-name
+    muon = "muon"  # pylint: disable=invalid-name
+
+
 class DeprecatedParameters(BaseModel):
     """configurations that are deprecated"""
 
@@ -494,17 +506,7 @@ class HyperparametersConfig(BaseModel):
     embedding_lr_scale: Optional[float] = None
     weight_decay: Optional[float] = 0.0
     optimizer: Optional[
-        Union[
+        Union[OptimizerNames, CustomSupportedOptimizers]
-            OptimizerNames,
-            Literal[
-                "lion_pytorch",
-                "optimi_adamw",
-                "ao_adamw_4bit",
-                "ao_adamw_8bit",
-                "ao_adamw_fp8",
-                "adopt_adamw",
-            ],
-        ]
     ] = OptimizerNames.ADAMW_HF
     optim_args: Optional[Union[str, Dict[str, Any]]] = Field(
         default=None,
@@ -518,7 +520,7 @@ class HyperparametersConfig(BaseModel):
     )
     torchdistx_path: Optional[str] = None
     lr_scheduler: Optional[
-        Union[SchedulerType, Literal["one_cycle"]]
+        Union[SchedulerType, Literal["one_cycle"], Literal["rex"]]
     ] = SchedulerType.COSINE
     lr_scheduler_kwargs: Optional[Dict[str, Any]] = None
     lr_quadratic_warmup: Optional[bool] = None
@@ -1177,6 +1179,13 @@ class AxolotlInputConfig(
             LOG.warning("adamw hyperparameters found, but no adamw optimizer set")
         return self
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_lr_groups(cls, data):
+        if data.get("lr_groups") and data.get("loraplus_lr_ratio"):
+            raise ValueError("lr_groups and loraplus_lr_ratio cannot be used together.")
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_saves(cls, data):

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

@@ -27,6 +27,7 @@ class TRLConfig(BaseModel):
     vllm_dtype: Optional[str] = "auto"
 
     reward_funcs: Optional[List[str]] = None
+    reward_weights: Optional[List[float]] = None
     num_generations: Optional[int] = None
     log_completions: Optional[bool] = False
 

src/axolotl/utils/schedulers.py

@@ -6,6 +6,80 @@ from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LambdaLR, LRScheduler
 
 
+class RexLR(LRScheduler):
+    """
+    Reflected Exponential (REX) learning rate scheduler.
+
+    - Original implementation: https://github.com/IvanVassi/REX_LR
+    - Original license: Apache 2.0
+    - Based on: https://arxiv.org/abs/2107.04197
+
+    Args:
+        optimizer (torch.optim.Optimizer): The optimizer to schedule the learning rate for.
+        max_lr (float): The maximum learning rate.
+        min_lr (float): The minimum learning rate.
+        total_steps (int): The total number of training steps.
+        num_warmup_steps (int): The number of warmup steps.
+        last_step (int): The index of last step.
+    """
+
+    def __init__(
+        self, optimizer, max_lr, min_lr, total_steps=0, num_warmup_steps=0, last_step=0
+    ):
+        if min_lr > max_lr:
+            raise ValueError(
+                f'Value of "min_lr" should be less than value of "max_lr". Got min_lr={min_lr} and max_lr={max_lr}'
+            )
+        if num_warmup_steps > total_steps:
+            raise ValueError(
+                f"num_warmup_steps ({num_warmup_steps}) must be less than or equal to total_steps ({total_steps})."
+            )
+
+        self.min_lr = min_lr
+        self.max_lr = max_lr
+        self.total_steps = total_steps
+        self.num_warmup_steps = num_warmup_steps
+        self.last_step = last_step - 1
+
+        # Ensure each parameter group has an "initial_lr" key to avoid issues when resuming.
+        for group in optimizer.param_groups:
+            group.setdefault("initial_lr", group["lr"])
+
+        # Pass self.last_step as last_epoch to the parent.
+        super().__init__(optimizer, last_epoch=self.last_step)
+
+    @property
+    def last_step(self):
+        return self.last_epoch
+
+    @last_step.setter
+    def last_step(self, value):
+        self.last_epoch = value
+
+    def get_lr(self):
+        # Warmup phase: if defined, increase lr linearly from 0 to max_lr.
+        if 1 <= self.last_step <= self.num_warmup_steps:
+            return [
+                base_lr * self.last_step / self.num_warmup_steps
+                for base_lr in self.base_lrs
+            ]
+
+        # Post-warmup phase: adjust step relative to the end of warmup.
+        step_after = self.last_step - self.num_warmup_steps
+        remaining_steps = self.total_steps - self.num_warmup_steps
+
+        # Avoid LR spiking
+        if step_after >= remaining_steps or step_after == -1 or remaining_steps <= 0:
+            return [self.min_lr for _ in self.base_lrs]
+
+        mod_iter = step_after % remaining_steps
+        z = (remaining_steps - mod_iter) / remaining_steps
+        rex_factor = self.min_lr / self.max_lr + (1.0 - self.min_lr / self.max_lr) * (
+            z / (0.1 + 0.9 * z)
+        )
+        return [base_lr * rex_factor for base_lr in self.base_lrs]
+
+
 class InterpolatingLogScheduler(LRScheduler):
     """
     A scheduler that interpolates learning rates in a logarithmic fashion

tests/cli/test_cli_train.py

@@ -28,7 +28,7 @@ class TestTrainCommand(BaseCliTest):
         config_path.write_text(valid_test_config)
 
         with patch("axolotl.cli.train.train") as mock_train:
-            mock_train.return_value = (MagicMock(), MagicMock())
+            mock_train.return_value = (MagicMock(), MagicMock(), MagicMock())
 
             result = cli_runner.invoke(
                 cli,
@@ -48,7 +48,7 @@ class TestTrainCommand(BaseCliTest):
         config_path = self._test_cli_overrides(tmp_path, valid_test_config)
 
         with patch("axolotl.cli.train.train") as mock_train:
-            mock_train.return_value = (MagicMock(), MagicMock())
+            mock_train.return_value = (MagicMock(), MagicMock(), MagicMock())
 
             result = cli_runner.invoke(
                 cli,

tests/e2e/test_mixtral.py

@@ -75,7 +75,7 @@ class TestMixtral(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        model, _ = train(cfg=cfg, dataset_meta=dataset_meta)
+        model, _, _ = train(cfg=cfg, dataset_meta=dataset_meta)
         assert (
             model.base_model.model.model.layers[0].block_sparse_moe.gate.weight.dtype
             == torch.float32
@@ -131,7 +131,7 @@ class TestMixtral(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        model, _ = train(cfg=cfg, dataset_meta=dataset_meta)
+        model, _, _ = train(cfg=cfg, dataset_meta=dataset_meta)
         assert (
             model.base_model.model.model.layers[0].block_sparse_moe.gate.weight.dtype
             == torch.float32
@@ -190,7 +190,7 @@ class TestMixtral(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        model, _ = train(cfg=cfg, dataset_meta=dataset_meta)
+        model, _, _ = train(cfg=cfg, dataset_meta=dataset_meta)
         assert (
             model.base_model.model.model.layers[0].block_sparse_moe.gate.weight.dtype
             == torch.float32
@@ -249,7 +249,7 @@ class TestMixtral(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        model, _ = train(cfg=cfg, dataset_meta=dataset_meta)
+        model, _, _ = train(cfg=cfg, dataset_meta=dataset_meta)
         assert (
             model.base_model.model.model.layers[0].block_sparse_moe.gate.weight.dtype
             == torch.float32

tests/e2e/test_optimizers.py

@@ -65,8 +65,9 @@ class TestCustomOptimizers(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, dataset_meta=dataset_meta)
+        _, _, trainer = train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        assert trainer.optimizer.optimizer.__class__.__name__ == "AdamW"
 
     @with_temp_dir
     @require_torch_2_5_1
@@ -111,8 +112,57 @@ class TestCustomOptimizers(unittest.TestCase):
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-        train(cfg=cfg, dataset_meta=dataset_meta)
+        _, _, trainer = train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        assert "ADOPT" in trainer.optimizer.optimizer.__class__.__name__
+
+    @with_temp_dir
+    @require_torch_2_5_1
+    def test_muon(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.1,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 5,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "muon",
+                "lr_scheduler": "cosine",
+                "weight_decay": 0.01,
+            }
+        )
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        _, _, trainer = train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
+        assert "Muon" in trainer.optimizer.optimizer.__class__.__name__
 
     @with_temp_dir
     def test_fft_schedule_free_adamw(self, temp_dir):

tests/e2e/test_schedulers.py

@@ -0,0 +1,71 @@
+"""
+E2E tests for custom schedulers using Llama
+"""
+
+import logging
+import os
+import unittest
+
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
+from axolotl.train import train
+from axolotl.utils.config import normalize_config, validate_config
+from axolotl.utils.dict import DictDefault
+
+from .utils import check_model_output_exists, with_temp_dir
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestCustomSchedulers(unittest.TestCase):
+    """
+    Test case for Llama models using LoRA
+    """
+
+    @with_temp_dir
+    def test_rex_scheduler(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.1,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_hf",
+                "max_steps": 20,
+                "lr_scheduler": "rex",
+                "warmup_steps": 5,
+                "cosine_min_lr_ratio": 0.05,
+            }
+        )
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)