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.
@@ -50,13 +47,14 @@ Features:
 ## 🚀 Quick Start
 
 **Requirements**:
+
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
 - PyTorch ≥2.4.1
 
 ### Installation
 
-```shell
+```bash
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 
 # Download example axolotl configs, deepspeed configs
@@ -68,7 +66,7 @@ Other installation approaches are described [here](https://axolotl-ai-cloud.gith
 
 ### Your First Fine-tune
 
-```shell
+```bash
 # Fetch axolotl examples
 axolotl fetch examples
 

_quarto.yml

@@ -3,10 +3,12 @@ project:
 
 website:
   title: "Axolotl"
-  description: "Fine-tuning"
+  description: "We make fine-tuning accessible, scalable, and fun"
   favicon: favicon.jpg
+
   navbar:
-    title: Axolotl
+    logo: image/axolotl_logo_digital_white.svg
+    title: false
     background: dark
     pinned: false
     collapse: false
@@ -25,33 +27,59 @@ website:
       contents:
         - text: Home
           href: index.qmd
-        - section: "How-To Guides"
+
+        - section: "Getting Started"
           contents:
-          # TODO Edit folder structure after we have more docs.
             - docs/getting-started.qmd
             - docs/installation.qmd
-            - docs/debugging.qmd
+            - docs/cli.qmd
             - docs/inference.qmd
-            - docs/multipack.qmd
-            - docs/fsdp_qlora.qmd
-            - docs/input_output.qmd
-            - docs/rlhf.qmd
-            - docs/nccl.qmd
-            - docs/mac.qmd
-            - docs/multi-gpu.qmd
-            - docs/multi-node.qmd
-            - docs/unsloth.qmd
-            - docs/amd_hpc.qmd
-            - docs/ray-integration.qmd
+
         - section: "Dataset Formats"
           contents: docs/dataset-formats/*
+
+        - section: "Deployments"
+          contents:
+            - docs/docker.qmd
+            - docs/multi-gpu.qmd
+            - docs/multi-node.qmd
+            - docs/ray-integration.qmd
+            - docs/amd_hpc.qmd
+            - docs/mac.qmd
+
+        - section: "How To Guides"
+          contents:
+            - docs/multimodal.qmd
+            - docs/rlhf.qmd
+            - docs/reward_modelling.qmd
+            - docs/lr_groups.qmd
+            - docs/lora_optims.qmd
+
+        - section: "Core Concepts"
+          contents:
+            - docs/batch_vs_grad.qmd
+            - docs/dataset_preprocessing.qmd
+            - docs/multipack.qmd
+
+        - section: "Advanced Features"
+          contents:
+            - docs/fsdp_qlora.qmd
+            - docs/unsloth.qmd
+            - docs/torchao.qmd
+            - docs/custom_integrations.qmd
+
+        - section: "Troubleshooting"
+          contents:
+            - docs/faq.qmd
+            - docs/debugging.qmd
+            - docs/nccl.qmd
+
         - section: "Reference"
           contents:
             - docs/config.qmd
-        - docs/faq.qmd
 
 format:
   html:
-    theme: materia
+    theme: darkly
     css: styles.css
     toc: true

docs/amd_hpc.qmd

@@ -1,5 +1,5 @@
 ---
-title: Training with AMD GPUs on HPC Systems
+title: AMD GPUs on HPC Systems
 description: A comprehensive guide for using Axolotl on distributed systems with AMD GPUs
 ---
 

docs/cli.qmd

@@ -1,28 +1,19 @@
-# Axolotl CLI Documentation
+---
+title: "CLI Reference"
+format:
+  html:
+    toc: true
+    toc-expand: 2
+    toc-depth: 3
+execute:
+  enabled: false
+---
 
 The Axolotl CLI provides a streamlined interface for training and fine-tuning large language models. This guide covers
 the CLI commands, their usage, and common examples.
 
-### Table of Contents
 
-- Basic Commands
-- Command Reference
-  - fetch
-  - preprocess
-  - train
-  - inference
-  - merge-lora
-  - merge-sharded-fsdp-weights
-  - evaluate
-  - lm-eval
-- Legacy CLI Usage
-- Remote Compute with Modal Cloud
-  - Cloud Configuration
-  - Running on Modal Cloud
-  - Cloud Configuration Options
-
-
-### Basic Commands
+## Basic Commands
 
 All Axolotl commands follow this general structure:
 
@@ -32,9 +23,9 @@ axolotl <command> [config.yml] [options]
 
 The config file can be local or a URL to a raw YAML file.
 
-### Command Reference
+## Command Reference
 
-#### fetch
+### fetch
 
 Downloads example configurations and deepspeed configs to your local machine.
 
@@ -49,7 +40,7 @@ axolotl fetch deepspeed_configs
 axolotl fetch examples --dest path/to/folder
 ```
 
-#### preprocess
+### preprocess
 
 Preprocesses and tokenizes your dataset before training. This is recommended for large datasets.
 
@@ -74,7 +65,7 @@ dataset_prepared_path: Local folder for saving preprocessed data
 push_dataset_to_hub: HuggingFace repo to push preprocessed data (optional)
 ```
 
-#### train
+### train
 
 Trains or fine-tunes a model using the configuration specified in your YAML file.
 
@@ -95,7 +86,38 @@ axolotl train config.yml --no-accelerate
 axolotl train config.yml --resume-from-checkpoint path/to/checkpoint
 ```
 
-#### inference
+It is possible to run sweeps over multiple hyperparameters by passing in a sweeps config.
+
+```bash
+# Basic training with sweeps
+axolotl train config.yml --sweep path/to/sweep.yaml
+```
+
+Example sweep config:
+```yaml
+_:
+  # This section is for dependent variables we need to fix
+  - load_in_8bit: false
+    load_in_4bit: false
+    adapter: lora
+  - load_in_8bit: true
+    load_in_4bit: false
+    adapter: lora
+
+# These are independent variables
+learning_rate: [0.0003, 0.0006]
+lora_r:
+  - 16
+  - 32
+lora_alpha:
+  - 16
+  - 32
+  - 64
+```
+
+
+
+### inference
 
 Runs inference using your trained model in either CLI or Gradio interface mode.
 
@@ -115,7 +137,7 @@ cat prompt.txt | axolotl inference config.yml \
     --base-model="./completed-model"
 ```
 
-#### merge-lora
+### merge-lora
 
 Merges trained LoRA adapters into the base model.
 
@@ -137,7 +159,7 @@ gpu_memory_limit: Limit GPU memory usage
 lora_on_cpu: Load LoRA weights on CPU
 ```
 
-#### merge-sharded-fsdp-weights
+### merge-sharded-fsdp-weights
 
 Merges sharded FSDP model checkpoints into a single combined checkpoint.
 
@@ -146,7 +168,7 @@ Merges sharded FSDP model checkpoints into a single combined checkpoint.
 axolotl merge-sharded-fsdp-weights config.yml
 ```
 
-#### evaluate
+### evaluate
 
 Evaluates a model's performance using metrics specified in the config.
 
@@ -155,27 +177,27 @@ Evaluates a model's performance using metrics specified in the config.
 axolotl evaluate config.yml
 ```
 
-#### lm-eval
+### lm-eval
 
 Runs LM Evaluation Harness on your model.
 
 ```bash
 # Basic evaluation
 axolotl lm-eval config.yml
-
-# Evaluate specific tasks
-axolotl lm-eval config.yml --tasks arc_challenge,hellaswag
 ```
 
 Configuration options:
 
 ```yaml
-lm_eval_tasks: List of tasks to evaluate
-lm_eval_batch_size: Batch size for evaluation
-output_dir: Directory to save evaluation results
+# List of tasks to evaluate
+lm_eval_tasks:
+  - arc_challenge
+  - hellaswag
+lm_eval_batch_size: # Batch size for evaluation
+output_dir: # Directory to save evaluation results
 ```
 
-### Legacy CLI Usage
+## Legacy CLI Usage
 
 While the new Click-based CLI is preferred, Axolotl still supports the legacy module-based CLI:
 
@@ -195,12 +217,18 @@ accelerate launch -m axolotl.cli.inference config.yml \
     --lora_model_dir="./outputs/lora-out" --gradio
 ```
 
-### Remote Compute with Modal Cloud
+::: {.callout-important}
+When overriding CLI parameters in the legacy CLI, use same notation as in yaml file (e.g., `--lora_model_dir`).
+
+**Note:** This differs from the new Click-based CLI, which uses dash notation (e.g., `--lora-model-dir`). Keep this in mind if you're referencing newer documentation or switching between CLI versions.
+:::
+
+## Remote Compute with Modal Cloud
 
 Axolotl supports running training and inference workloads on Modal cloud infrastructure. This is configured using a
 cloud YAML file alongside your regular Axolotl config.
 
-#### Cloud Configuration
+### Cloud Configuration
 
 Create a cloud config YAML with your Modal settings:
 
@@ -215,13 +243,17 @@ branch: main    # Git branch to use (optional)
 volumes:        # Persistent storage volumes
   - name: axolotl-cache
     mount: /workspace/cache
+  - name: axolotl-data
+    mount: /workspace/data
+  - name: axolotl-artifacts
+    mount: /workspace/artifacts
 
 env:            # Environment variables
   - WANDB_API_KEY
   - HF_TOKEN
 ```
 
-#### Running on Modal Cloud
+### Running on Modal Cloud
 
 Commands that support the --cloud flag:
 
@@ -239,18 +271,18 @@ axolotl train config.yml --cloud cloud_config.yml --no-accelerate
 axolotl lm-eval config.yml --cloud cloud_config.yml
 ```
 
-#### Cloud Configuration Options
+### Cloud Configuration Options
 
 ```yaml
-provider: compute provider, currently only `modal` is supported
-gpu: GPU type to use
-gpu_count: Number of GPUs (default: 1)
-memory: RAM in GB (default: 128)
-timeout: Maximum runtime in seconds
-timeout_preprocess: Preprocessing timeout
-branch: Git branch to use
-docker_tag: Custom Docker image tag
-volumes: List of persistent storage volumes
-env: Environment variables to pass
-secrets: Secrets to inject
+provider: # compute provider, currently only `modal` is supported
+gpu: # GPU type to use
+gpu_count: # Number of GPUs (default: 1)
+memory: # RAM in GB (default: 128)
+timeout: # Maximum runtime in seconds
+timeout_preprocess: # Preprocessing timeout
+branch: # Git branch to use
+docker_tag: # Custom Docker image tag
+volumes: # List of persistent storage volumes
+env: # Environment variables to pass
+secrets: # Secrets to inject
 ```

docs/config.qmd

@@ -163,10 +163,16 @@ 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`
-    # Note: If the below 4 fields are empty, defaults to training only on the last message.
+    # Note: If the below 4 fields are set to empty, defaults to training only on the last message.
 
     # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.
     roles_to_train: ["assistant"]  # default
@@ -174,6 +180,7 @@ datasets:
     # - all: train on all EOS tokens
     # - turn (default): train on the EOS token at the end of each trainable turn
     # - last: train on the last EOS token in the conversation
+    # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.
     train_on_eos: last
     # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.
     message_field_training: training
@@ -221,8 +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
-default_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.
+# Changes the default system message. 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
@@ -444,7 +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)
@@ -527,6 +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/custom_integrations.qmd

@@ -0,0 +1,57 @@
+---
+title: Custom Integrations
+toc: true
+toc-depth: 3
+---
+
+```{python}
+#| echo: false
+
+import re
+
+def process_readme(integration_name):
+    try:
+        path = f'../src/axolotl/integrations/{integration_name}/README.md'
+        with open(path, 'r') as f:
+            txt = f.read()
+            # Remove h1 headings
+            txt = re.sub(r'^# .*\n?', '', txt, flags=re.MULTILINE)
+            # Convert h2 to h3
+            txt = re.sub(r'^## ', '### ', txt, flags=re.MULTILINE)
+            return txt
+    except FileNotFoundError:
+        return None
+
+def print_section(name, folder_name):
+    output = f"\n## {name}\n"
+    content = process_readme(folder_name)
+    if content:
+        output += content
+    output += f"\nPlease see reference [here](https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/{folder_name})\n"
+    return output
+```
+
+```{python}
+#| output: asis
+#| echo: false
+
+# Introduction text
+print("""
+Axolotl adds custom features through `integrations`. They are located within the `src/axolotl/integrations` directory.
+
+To enable them, please check the respective documentations.
+""")
+
+# Sections
+sections = [
+    ("Cut Cross Entropy", "cut_cross_entropy"),
+    ("Grokfast", "grokfast"),
+    ("Knowledge Distillation (KD)", "kd"),
+    ("Liger Kernels", "liger"),
+    ("Language Model Evaluation Harness (LM Eval)", "lm_eval"),
+    ("Spectrum", "spectrum")
+]
+
+for section_name, folder_name in sections:
+    print(print_section(section_name, folder_name))
+```

docs/dataset-formats/conversation.qmd

@@ -6,7 +6,9 @@ order: 3
 
 ## sharegpt
 
-IMPORTANT: ShareGPT is deprecated!. Please see [chat_template](#chat_template) section below.
+::: {.callout-important}
+ShareGPT is deprecated!. Please see [chat_template](#chat_template) section below.
+:::
 
 ## pygmalion
 
@@ -102,6 +104,10 @@ datasets:
     type: chat_template
 ```
 
+::: {.callout-important}
+Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.
+:::
+
 5. (Advanced) Using fine-grained control over tokens and turns to train in a conversation
 
 For a data sample that looks like:
@@ -149,4 +155,6 @@ datasets:
     message_field_training_detail: train_detail
 ```
 
-Tip: It is not necessary to use both `message_field_training` and `message_field_training_detail` at a time.
+::: {.callout-tip}
+It is not necessary to set both `message_field_training` and `message_field_training_detail` at once.
+:::

docs/dataset-formats/index.qmd

@@ -13,7 +13,7 @@ As there are a lot of available options in Axolotl, this guide aims to provide a
 
 Axolotl supports 3 kinds of training methods: pre-training, supervised fine-tuning, and preference-based post-training (e.g. DPO, ORPO, PRMs). Each method has their own dataset format which are described below.
 
-## [Pre-training](pretraining.qmd)
+## Pre-training
 
 When aiming to train on large corpora of text datasets, pre-training is your go-to choice. Due to the size of these datasets, downloading the entire-datasets before beginning training would be prohibitively time-consuming. Axolotl supports [streaming](https://huggingface.co/docs/datasets/en/stream) to only load batches into memory at a time.
 
@@ -96,6 +96,10 @@ One step is equal to `sequence_len * micro_batch_size * gradient_accumulation_st
 
 It is recommended to leave this off if downloading from Hugging Face hub as it would download the entire dataset which can be very large.
 
+### Reference
+
+Please see docs [here](pretraining.qmd).
+
 ## Supervised fine-tuning (SFT)
 
 Supervised fine-tuning is the process of training models to respond to an instruction or chat input.
@@ -120,11 +124,12 @@ If you went through the flow chart and did not find one that matches, it is reco
 You can mix and match within each approach or across approaches to train a model on a variety of datasets.
 :::
 
-### [Pre-Tokenized Dataset](tokenized.qmd)
+### Pre-Tokenized Dataset
 
 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`.
@@ -145,7 +150,9 @@ datasets:
 `type: ` is empty!
 :::
 
-### [Template Free Dataset](template_free.qmd)
+Reference: [Pre-Tokenized Dataset Documentation](tokenized.qmd).
+
+### Template Free Dataset
 
 We reccomend this approach when you want granular control over the prompt formatting, special tokens, and masking, whilst letting Axolotl handle the tokenization. This is very useful if your dataset has unique prompts that differ across samples and where one single general template wouldn't suffice.
 
@@ -182,7 +189,9 @@ datasets:
     type: input_output
 ```
 
-### [Conversation Dataset](conversation.qmd)
+Reference: [Template Free Documentation](template_free.qmd).
+
+### Conversation Dataset
 
 `conversation` messages are a list of messages which usually contain a `role` and `content` key.
 
@@ -258,7 +267,7 @@ Newer conversation datasets usually follow the OpenAI format.
 
 Axolotl supports both as well as allowing customization of any kind of key.
 
-#### [Chat Template Usage](conversation.qmd#chat_template)
+#### Chat Template Usage
 
 To properly use this method, it is important to identify three things:
 
@@ -340,9 +349,19 @@ datasets:
       narrator: ["narrator"]
 ```
 
-#### Applying `chat_template`
+::: {.callout-tip}
+As chat_templates may use hardcoded EOS/EOT tokens that are different from the tokenizer's EOS, it is highly recommended to set them. For example, `ChatML` uses `<|im_end|>` to end turns.
 
-Once all the above steps are completed, you could combine all these configs together to form a bespoke configuration for your  custom dataset. The final step would be to correctly set the EOS token in your config:
+```yaml
+special_tokens:
+  eos_token: <|im_end|>
+```
+
+:::
+
+##### Applying `chat_template`
+
+Once all the above steps are completed, you could combine all these configs together to form a bespoke configuration for your custom dataset.
 
 ```yaml
 datasets:
@@ -391,7 +410,17 @@ If this config were to be applied to the sample dataset above, the output would
 
 The first number refers to the label, the second refers to the `token_id`. For example, `-100` labels appear on non-assistant portions, meaning that they are masked during. For assistant portions, the label is the same as the `token_id`.
 
-### [Instruction Dataset](inst_tune.qmd)
+::: {.callout-note}
+
+If during `preprocess`, there are a lot of warnings of `Could not find content __ boundary`, please check the FAQ section for [chat_templates](../faq.qmd#chat-templates).
+
+:::
+
+#### Reference
+
+Please see docs [here](conversation.qmd).
+
+### Instruction Dataset
 
 Instruction datasets are used to train instruction-following models and comprise a prompt, containing an instruction, and a single response. In contrast to chat datasets which may be multi-turn, instruct datasets are typically single-turn.
 
@@ -423,6 +452,9 @@ datasets:
 
 Axolotl supports many kinds of instruction dataset. All of them can be found here (https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/inst_tune.html) with their respective type and sample row format.
 
+
+Reference: [Instruction Dataset Documentation](inst_tune.qmd).
+
 #### Custom Instruct Prompt Format
 
 Due to the myriad possibilities of instruction formats, Axolotl allows customizing your own instruction format without having to dive into the code directly.
@@ -453,6 +485,8 @@ datasets:
 
 The config sets that the `field_instruction` is actually named `input`, and the `field_input` is empty as we don't have an `input` in this sample. Generally, `instruction` can be thought as the question to the model, and `input` as the additional information with `output` being the response. It is not necessary to have an `input` nor `system`. In the end, the most important part is to understand what format you want it to look like and how you can customize this to your use case.
 
+Reference: [Custom Instruct Prompt Format Documentation](inst_tune.qmd#how-to-add-custom-prompt-format).
+
 ## Reinforcement Learning from Human Feedback (RLHF)
 
-As there are multiple RLHF methods with their own dataset requirements. Please see [RLHF datasets](../rlhf.qmd) documentation for more detail.
+As there are multiple RLHF methods with their own dataset requirements. Please see [RLHF documentation](../rlhf.qmd) for more detail.

docs/dataset-formats/pretraining.qmd

@@ -27,7 +27,6 @@ pretraining_dataset:
     type: pretrain
     trust_remote_code:
     skip: # number of rows of data to skip over from the beginning
-...
 ```
 
 :::

docs/dataset-formats/template_free.qmd

@@ -1,7 +1,239 @@
 ---
 title: Template-Free
 description: Construct prompts without a template.
+toc: true
+toc-depth: 3
 order: 4
 ---
 
-See [these docs](../input_output.qmd).
+## Background {#sec-background}
+
+### Masking Inputs {#masking-inputs}
+
+One of the most popular features of
+[axolotl](https://github.com/axolotl-ai-cloud/axolotl) is
+setting the following configuration value:
+
+
+```yaml
+train_on_inputs: false
+```
+
+If you declare a [dataset formats](https://github.com/axolotl-ai-cloud/axolotl?tab=readme-ov-file#dataset)
+such as `alpaca` or `chatml`, axolotl knows what is an input
+(i.e. human) vs. an output (i.e. the assistant) and masks the input
+labels so that your model can focus on predicting the outputs only.
+
+### You may not want prompt templates {#sec-you-may-not-want-prompt-templates}
+
+However, there are many situations where you don't want to use one of
+these formats or templates. This is because they can:
+
+-   Add unnecessary boilerplate to your prompts.
+-   Create artifacts like special delimiters `<|im_start|>` that can
+    quickly become footguns if you don't include them correctly at
+    inference time.
+-   Enforce a *chat* interface when you do not want one. Sometimes you
+    just want to fine-tune a model to a very specific task and do NOT
+    want multi-turn conversations, roles, etc.
+-   Limit you to only certain roles that the template allows.
+
+### The `input_output` format {#sec-the-inputoutput-format}
+
+You can construct your prompts without a template by using the
+`input_output` format, by setting `type: input_output` in your
+configuration file like this:
+
+**config.yml**
+
+```yaml
+train_on_inputs: false # Mask segments of your data
+datasets:
+  - path: output.jsonl
+    type: input_output  # use template free prompt construction
+```
+
+Unlike `type: completion`, which is also template-free,
+`type: input_output` allows you to mask segments of your text. More
+details on how this works are described below.
+
+## Usage {#sec-usage}
+
+This is how you can use the `input_output` format:
+
+### 1. Prepare Data {#sec-1-prepare-data}
+
+To use the `input_output` format, collect your data in the following
+format into a jsonl file (below is the first row from the file
+`output`.jsonl` pretty printed):
+
+```bash
+$ head -n1 output.jsonl | python -m json.tool
+```
+
+:::{.cell-output .cell-output-stdout}
+    {
+        "segments": [
+            {
+                "label": true,
+                "text": "<s>Hello\n"
+            },
+            {
+                "label": true,
+                "text": "hi there!. "
+            },
+            {
+                "label": false,
+                "text": "goodbye "
+            },
+            {
+                "label": true,
+                "text": "farewell</s>"
+            }
+        ]
+    }
+:::
+
+Set `label:false` when you want to mask a segment of text so that the
+model isn't trained on it. Some things to keep in mind:
+
+> [!IMPORTANT]
+> 1.  **EOS, BOS, spaces, newlines etc. are entirely up to you. Axolotl
+    concatenates all the segments as-is.** The tokenizer doesn't add
+    anything additional. Notice how I added spaces, newlines, `<s>`
+    (BOS), and `</s>` (EOS) myself.
+> 2.  Make sure you check the materialized output to validate that the
+    prompt is getting assembled how you like.
+
+### 2. Use `type: input_output` {#sec-2-use-type-inputoutput}
+
+Let's materialize data with our `output.jsonl` file by setting
+`type: input_output` in our axolotl config:
+
+```yaml
+# training_config.yaml
+base_model: mistralai/Mistral-7B-v0.1
+data_seed: 49
+seed: 49
+
+datasets:
+  - path: output.jsonl
+    type: input_output
+val_set_size: 0.1
+
+sequence_len: 896
+sample_packing: false
+
+micro_batch_size: 2
+gradient_accumulation_steps: 3
+eval_batch_size: 2
+num_epochs: 1
+learning_rate: 0.0002
+
+train_on_inputs: false
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"
+```
+
+You can use the following command to materialize your data. The
+`--debug` flag will print the tokens, along with the labels so you can
+verify that the correct items are being ignored:
+
+```bash
+axolotl preprocess training_config.yaml --debug
+
+...
+[2024-03-05 23:36:46,969] [INFO] [axolotl.check_example_labels:35] [PID:607731] [RANK:0] <s>(1, 1) Hello(22557, 22557)
+(13, 13) hi(12014, 12014) there(736, 736) !(28808, 28808) .(28723, 28723) (28705, 28705) good(-100, 1179) bye(-100, 17664) (-100, 28705) fare(19111, 19111) well(5458, 5458) </s>(2, 2)
+
+```
+
+The format is `decoded_token`(`label`, `token_id`), for example,
+`<s>(1, 1)` means that the token is `<s>`, the label is `1` and the
+token_id is `1`. When the label is `-100` then that token is ignored for
+training.
+
+### 3. Check the prompts {#sec-3-check-the-prompts}
+
+Here is another way to check the materialized output:
+
+```python
+from transformers import AutoTokenizer
+from datasets import load_from_disk
+import yaml
+
+directory = !ls last_run_prepared/
+with open('training_config.yaml', 'r') as f:
+    cfg = yaml.safe_load(f)
+model_id = cfg['base_model']
+tok = AutoTokenizer.from_pretrained(model_id)
+ds = load_from_disk(f'last_run_prepared/{directory[0]}/')
+```
+
+```python
+>>> row = ds[0]
+>>> print(tok.decode(row['input_ids']))
+<s> Hello
+    hi there!.  goodbye  farewell</s>
+```
+
+We can check that the right tokens are ignored by comparing the labels
+to each token:
+
+```python
+import pandas as pd
+pd.DataFrame([{'token': tok.decode(i), 'label': l, 'id':i} for i,l in
+              zip(row['input_ids'], row['labels'])])
+```
+
+| token | label | id    |
+|-------|-------|-------|
+| 0     | \<s\> | 1     |
+| 1     | Hello | 22557 |
+| 2     | \\n   | 13    |
+| 3     | hi    | 12014 |
+| 4     | there | 736   |
+| 5     | !     | 28808 |
+| 6     | .     | 28723 |
+| 7     |       | 28705 |
+| 8     | good  | -100  |
+| 9     | bye   | -100  |
+| 10    |       | -100  |
+| 11    | fare  | 19111 |
+| 12    | well  | 5458  |
+| 13    | \</s\>| 2     |
+
+
+
+If we look at the input data, the above table seems correct! (The jsonl
+version is repeated below for reference):
+
+
+```bash
+$ head -n1 output.jsonl | python -m json.tool
+```
+
+:::{.cell-output .cell-output-stdout}
+    {
+        "segments": [
+            {
+                "label": true,
+                "text": "<s>Hello\n"
+            },
+            {
+                "label": true,
+                "text": "hi there!. "
+            },
+            {
+                "label": false,
+                "text": "goodbye "
+            },
+            {
+                "label": true,
+                "text": "farewell</s>"
+            }
+        ]
+    }
+:::

docs/dataset_preprocessing.qmd

@@ -3,8 +3,11 @@ title: Dataset Preprocessing
 description: How datasets are processed
 ---
 
+## Overview
+
 Dataset pre-processing is the step where Axolotl takes each dataset you've configured alongside
-the (dataset format)[../dataset-formats/] and prompt strategies to:
+the [dataset format](docs/dataset-formats) and prompt strategies to:
+
  - parse the dataset based on the *dataset format*
  - transform the dataset to how you would interact with the model based on the *prompt strategy*
  - tokenize the dataset based on the configured model & tokenizer
@@ -12,10 +15,12 @@ the (dataset format)[../dataset-formats/] and prompt strategies to:
 
 The processing of the datasets can happen one of two ways:
 
-1. Before kicking off training by calling `python -m axolotl.cli.preprocess /path/to/your.yaml --debug`
+1. Before kicking off training by calling `axolotl preprocess config.yaml --debug`
 2. When training is started
 
-What are the benefits of pre-processing? When training interactively or for sweeps
+### What are the benefits of pre-processing?
+
+When training interactively or for sweeps
 (e.g. you are restarting the trainer often), processing the datasets can oftentimes be frustratingly
 slow. Pre-processing will cache the tokenized/formatted datasets according to a hash of dependent
 training parameters so that it will intelligently pull from its cache when possible.
@@ -28,8 +33,12 @@ default path of `./last_run_prepared/`, but will ignore anything already cached
 setting `dataset_prepared_path: ./last_run_prepared`, the trainer will use whatever pre-processed
 data is in the cache.
 
-What are the edge cases? Let's say you are writing a custom prompt strategy or using a user-defined
+### What are the edge cases?
+
+Let's say you are writing a custom prompt strategy or using a user-defined
 prompt template. Because the trainer cannot readily detect these changes, we cannot change the
-calculated hash value for the pre-processed dataset. If you have `dataset_prepared_path: ...` set
+calculated hash value for the pre-processed dataset.
+
+If you have `dataset_prepared_path: ...` set
 and change your prompt templating logic, it may not pick up the changes you made and you will be
 training over the old prompt.

docs/debugging.qmd

@@ -31,11 +31,13 @@ While debugging it's helpful to simplify your test scenario as much as possible.
     - Set `CUDA_VISIBLE_DEVICES` to a single GPU, ex: `export CUDA_VISIBLE_DEVICES=0`.
     - Set `dataset_processes: 1` in your axolotl config or run the training command with `--dataset_processes=1`.
 2. **Use a small dataset**: Construct or use a small dataset from HF Hub. When using a small dataset, you will often have to make sure `sample_packing: False` and `eval_sample_packing: False` to avoid errors.  If you are in a pinch and don't have time to construct a small dataset but want to use from the HF Hub, you can shard the data (this will still tokenize the entire dataset, but will only use a fraction of the data for training.  For example, to shard the dataset into 20 pieces, add the following to your axolotl config):
+
     ```yaml
-    dataset:
+    datasets:
         ...
         shards: 20
     ```
+
 3. **Use a small model**: A good example of a small model is [TinyLlama/TinyLlama-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0).
 4. **Minimize iteration time**: Make sure the training loop finishes as fast as possible, with these settings.
     - `micro_batch_size: 1`
@@ -85,7 +87,7 @@ The easiest way to get started is to modify the [.vscode/launch.json](../.vscode
 
 For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_chat_template.yml`, you would use the below configuration[^1].  Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to `devtools` and set the `env` variable `HF_HOME` to a temporary folder that is later partially deleted.  This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.
 
-```jsonc
+```json
 // .vscode/launch.json
 {
     "version": "0.2.0",
@@ -132,7 +134,7 @@ For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 acceler
 
 Below is the [./vscode/tasks.json](../.vscode/tasks.json) file that defines the `cleanup-for-dataprep` task.  This task is run before each debugging session when you use the above configuration.  Note how there are two tasks that delete the two folders mentioned above.  The third task `cleanup-for-dataprep` is a composite task that combines the two tasks.  A composite task is necessary because VSCode does not allow you to specify multiple tasks in the `preLaunchTask` argument of the `launch.json` file.
 
-```jsonc
+```json
 // .vscode/tasks.json
 // this file is used by launch.json
 {

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

@@ -3,6 +3,7 @@ title: FAQ
 description: Frequently asked questions
 ---
 
+### General
 
 **Q: The trainer stopped and hasn't progressed in several minutes.**
 
@@ -18,12 +19,36 @@ 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.**
 
 > A: This is usually an issue with the GPU. This can be resolved through setting the os environment variable `CUDA_VISIBLE_DEVICES=0`. If you are on runpod, this is usually a pod issue. Starting a new pod should take care of it.
 
+### Chat templates
+
 **Q: `jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____`**
 
 > A: This means that the property mapping for the stated attribute does not exist when building `chat_template` prompt. For example, if `no attribute 'content'`, please check you have added the correct mapping for `content` under `message_property_mappings`.
+
+**Q: `Empty template generated for turn ___`**
+
+> A: The `content` is empty for that turn.
+
+**Q: `Could not find content start/end boundary for turn __`**
+
+> A: The specific turn's start/end could not be detected. Please ensure you have set the `eos_token` following your `chat_template`. Otherwise, this could be a `chat_template` which doesn't use proper boundaries for each turn (like system). On the rare occurrence, make sure your content is not `[[dummy_message]]`. Please let us know about this.
+
+**Q: `Content end boundary is before start boundary for turn ___`**
+
+> A: This is an edge case which should not occur. Please create an Issue if this happens.
+
+**Q: `Content end boundary is the same as start boundary for turn ___. This is likely an empty turn.`**
+
+> A: This is likely an empty turn.
+
+**Q: The EOS/EOT token is incorrectly being masked or not being masked.**
+
+> A: This is because of the mismatch between `tokenizer.eos_token` and EOS/EOT token in template. Please make sure to set `eos_token` under `special_tokens` to the same EOS/EOT token as in template.

docs/getting-started.qmd

@@ -1,5 +1,5 @@
 ---
-title: "Getting Started with Axolotl"
+title: "Quickstart"
 format:
   html:
     toc: true
@@ -17,12 +17,12 @@ Let's start by fine-tuning a small language model using LoRA. This example uses
 Assuming `axolotl` is installed (if not, see our [Installation Guide](installation.qmd))
 
 1. Download example configs:
-```shell
+```bash
 axolotl fetch examples
 ```
 
 2. Run the training:
-```shell
+```bash
 axolotl train examples/llama-3/lora-1b.yml
 ```
 
@@ -108,7 +108,7 @@ Please consult the supported [Dataset Formats](dataset-formats/) for more detail
 
 3. Run the training:
 
-```shell
+```bash
 axolotl train my_training.yml
 ```
 
@@ -118,7 +118,7 @@ axolotl train my_training.yml
 
 After training, test your model:
 
-```shell
+```bash
 axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out"
 ```
 
@@ -126,7 +126,7 @@ axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out"
 
 For large datasets, preprocess first:
 
-```shell
+```bash
 axolotl preprocess my_training.yml
 ```
 
@@ -134,7 +134,7 @@ axolotl preprocess my_training.yml
 
 Launch a Gradio interface:
 
-```shell
+```bash
 axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out" --gradio
 ```
 

docs/inference.qmd

@@ -1,11 +1,10 @@
 ---
-title: "Inference Guide"
+title: "Inference"
 format:
   html:
     toc: true
     toc-depth: 3
     number-sections: true
-    code-tools: true
 execute:
   enabled: false
 ---

docs/input_output.qmd

@@ -3,263 +3,4 @@ title: Template-free prompt construction
 description: "Template-free prompt construction with the `input_output` format"
 ---
 
-<!-- TOC -->
-
-- [Background](#background)
-    - [Masking Inputs](#masking-inputs)
-    - [You may not want prompt templates](#you-may-not-want-prompt-templates)
-    - [The `input_output` format](#the-input_output-format)
-- [Usage](#usage)
-    - [1. Prepare Data](#1-prepare-data)
-    - [2. Use `type: input_output`](#2-use-type-input_output)
-    - [3. Check the prompts](#3-check-the-prompts)
-
-<!-- /TOC -->
-
-<a id="markdown-background" name="background"></a>
-
-## Background
-
-<a id="markdown-masking-inputs" name="masking-inputs"></a>
-
-### Masking Inputs
-
-One of the most popular features of
-[axolotl](https://github.com/axolotl-ai-cloud/axolotl) is
-setting the following configuration value:
-
-
-```yaml
-train_on_inputs: false
-```
-
-If you declare a [dataset formats](https://github.com/axolotl-ai-cloud/axolotl?tab=readme-ov-file#dataset)
-such as `alpaca` or `chatml`, axolotl knows what is an input
-(i.e. human) vs. an output (i.e. the assistant) and masks the input
-labels so that your model can focus on predicting the outputs only.
-
-<a id="markdown-you-may-not-want-prompt-templates" name="you-may-not-want-prompt-templates"></a>
-
-### You may not want prompt templates
-
-However, there are many situations where you don't want to use one of
-these formats or templates. This is because they can:
-
--   Add unnecessary boilerplate to your prompts.
--   Create artifacts like special delimiters `<|im_start|>` that can
-    quickly become footguns if you don't include them correctly at
-    inference time.
--   Enforce a *chat* interface when you do not want one. Sometimes you
-    just want to fine-tune a model to a very specific task and do NOT
-    want multi-turn conversations, roles, etc.
--   Limit you to only certain roles that the template allows.
-
-<a id="markdown-the-inputoutput-format" name="the-inputoutput-format"></a>
-
-### The `input_output` format
-
-You can construct your prompts without a template by using the
-`input_output` format, by setting `type: input_output` in your
-configuration file like this:
-
-**config.yml**
-
-```yaml
-train_on_inputs: false # Mask segments of your data
-datasets:
-  - path: output.jsonl
-    type: input_output  # use template free prompt construction
-```
-
-Unlike `type: completion`, which is also template-free,
-`type: input_output` allows you to mask segments of your text. More
-details on how this works are described below.
-
-<a id="markdown-usage" name="usage"></a>
-
-## Usage
-
-This is how you can use the `input_output` format:
-
-<a id="markdown-1-prepare-data" name="1-prepare-data"></a>
-
-### 1. Prepare Data
-
-To use the `input_output` format, collect your data in the following
-format into a jsonl file (below is the first row from the file
-`output`.jsonl` pretty printed):
-
-```bash
-$ head -n1 output.jsonl | python -m json.tool
-```
-
-:::{.cell-output .cell-output-stdout}
-    {
-        "segments": [
-            {
-                "label": true,
-                "text": "<s>Hello\n"
-            },
-            {
-                "label": true,
-                "text": "hi there!. "
-            },
-            {
-                "label": false,
-                "text": "goodbye "
-            },
-            {
-                "label": true,
-                "text": "farewell</s>"
-            }
-        ]
-    }
-:::
-
-Set `label:false` when you want to mask a segment of text so that the
-model isn't trained on it. Some things to keep in mind:
-
-> [!IMPORTANT]
-> 1.  **EOS, BOS, spaces, newlines etc. are entirely up to you. Axolotl
-    concatenates all the segments as-is.** The tokenizer doesn't add
-    anything additional. Notice how I added spaces, newlines, `<s>`
-    (BOS), and `</s>` (EOS) myself.
-> 2.  Make sure you check the materialized output to validate that the
-    prompt is getting assembled how you like.
-
-<a id="markdown-2-use-type-inputoutput" name="2-use-type-inputoutput"></a>
-
-### 2. Use `type: input_output`
-
-Let's materialize data with our `output.jsonl` file by setting
-`type: input_output` in our axolotl config:
-
-```yaml
-# training_config.yaml
-base_model: mistralai/Mistral-7B-v0.1
-data_seed: 49
-seed: 49
-
-datasets:
-  - path: output.jsonl
-    type: input_output
-val_set_size: 0.1
-
-sequence_len: 896
-sample_packing: false
-
-micro_batch_size: 2
-gradient_accumulation_steps: 3
-eval_batch_size: 2
-num_epochs: 1
-learning_rate: 0.0002
-
-train_on_inputs: false
-special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"
-```
-
-You can use the following command to materialize your data. The
-`--debug` flag will print the tokens, along with the labels so you can
-verify that the correct items are being ignored:
-
-```bash
-$ python -m axolotl.cli.preprocess training_config.yaml --debug
-
-...
-[2024-03-05 23:36:46,969] [INFO] [axolotl.check_example_labels:35] [PID:607731] [RANK:0] <s>(1, 1) Hello(22557, 22557)
-(13, 13) hi(12014, 12014) there(736, 736) !(28808, 28808) .(28723, 28723) (28705, 28705) good(-100, 1179) bye(-100, 17664) (-100, 28705) fare(19111, 19111) well(5458, 5458) </s>(2, 2)
-
-```
-
-The format is `decoded_token`(`label`, `token_id`), for example,
-`<s>(1, 1)` means that the token is `<s>`, the label is `1` and the
-token_id is `1`. When the label is `-100` then that token is ignored for
-training.
-
-<a id="markdown-3-check-the-prompts" name="3-check-the-prompts"></a>
-
-### 3. Check the prompts
-
-Here is another way to check the materialized output:
-
-```python
-from transformers import AutoTokenizer
-from datasets import load_from_disk
-import yaml
-
-directory = !ls last_run_prepared/
-with open('training_config.yaml', 'r') as f:
-    cfg = yaml.safe_load(f)
-model_id = cfg['base_model']
-tok = AutoTokenizer.from_pretrained(model_id)
-ds = load_from_disk(f'last_run_prepared/{directory[0]}/')
-```
-
-```python
->>> row = ds[0]
->>> print(tok.decode(row['input_ids']))
-<s> Hello
-    hi there!.  goodbye  farewell</s>
-```
-
-We can check that the right tokens are ignored by comparing the labels
-to each token:
-
-```python
-import pandas as pd
-pd.DataFrame([{'token': tok.decode(i), 'label': l, 'id':i} for i,l in
-              zip(row['input_ids'], row['labels'])])
-```
-
-| token | label | id    |
-|-------|-------|-------|
-| 0     | \<s\> | 1     |
-| 1     | Hello | 22557 |
-| 2     | \\n   | 13    |
-| 3     | hi    | 12014 |
-| 4     | there | 736   |
-| 5     | !     | 28808 |
-| 6     | .     | 28723 |
-| 7     |       | 28705 |
-| 8     | good  | -100  |
-| 9     | bye   | -100  |
-| 10    |       | -100  |
-| 11    | fare  | 19111 |
-| 12    | well  | 5458  |
-| 13    | \</s\>| 2     |
-
-
-
-If we look at the input data, the above table seems correct! (The jsonl
-version is repeated below for reference):
-
-
-```bash
-$ head -n1 output.jsonl | python -m json.tool
-```
-
-:::{.cell-output .cell-output-stdout}
-    {
-        "segments": [
-            {
-                "label": true,
-                "text": "<s>Hello\n"
-            },
-            {
-                "label": true,
-                "text": "hi there!. "
-            },
-            {
-                "label": false,
-                "text": "goodbye "
-            },
-            {
-                "label": true,
-                "text": "farewell</s>"
-            }
-        ]
-    }
-:::
+The documentation moved to [here](dataset-formats/template_free.qmd).

docs/installation.qmd

@@ -1,11 +1,10 @@
 ---
-title: "Installation Guide"
+title: "Installation"
 format:
   html:
     toc: true
     toc-depth: 3
     number-sections: true
-    code-tools: true
 execute:
   enabled: false
 ---
@@ -66,6 +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}

docs/lora_optims.qmd

@@ -1,7 +1,6 @@
 ---
 title: "LoRA Optimizations"
-description: "Custom autograd functions and Triton kernels in Axolotl for optimized
-LoRA fine-tuning"
+description: "Custom autograd functions and Triton kernels in Axolotl for optimized LoRA fine-tuning"
 ---
 
 Inspired by [Unsloth](https://github.com/unslothai/unsloth), we've implemented two

docs/mac.qmd

@@ -19,4 +19,5 @@ Current support:
 - [ ] DeepSpeed
 
 Untested:
+
 - FSDP

docs/multi-gpu.qmd

@@ -1,5 +1,5 @@
 ---
-title: "Multi-GPU Training Guide"
+title: "Multi-GPU"
 format:
   html:
     toc: true
@@ -35,7 +35,11 @@ deepspeed: deepspeed_configs/zero1.json
 ### Usage {#sec-deepspeed-usage}
 
 ```{.bash}
-accelerate launch -m axolotl.cli.train examples/llama-2/config.yml --deepspeed deepspeed_configs/zero1.json
+# Passing arg via config
+axolotl train config.yml
+
+# Passing arg via cli
+axolotl train config.yml --deepspeed deepspeed_configs/zero1.json
 ```
 
 ### ZeRO Stages {#sec-zero-stages}
@@ -70,25 +74,7 @@ For combining FSDP with QLoRA, see our [dedicated guide](fsdp_qlora.qmd).
 
 ### Liger Kernel Integration {#sec-liger}
 
-::: {.callout-note}
-Liger Kernel provides efficient Triton kernels for LLM training, offering:
-
-- 20% increase in multi-GPU training throughput
-- 60% reduction in memory usage
-- Compatibility with both FSDP and DeepSpeed
-:::
-
-Configuration:
-
-```{.yaml}
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-```
+Please see [docs](custom_integrations.qmd#liger) for more info.
 
 ## Troubleshooting {#sec-troubleshooting}
 

docs/multi-node.qmd

@@ -13,7 +13,7 @@ You will also need to have the same configuration file for your model on each ma
 Make sure the main machine is reachable by other machines.
 :::
 
-# Accelerate
+## Accelerate
 
 You will need to create a configuration for accelerate, either by using `accelerate config` and follow the instructions or you can use one of the preset below:
 
@@ -51,17 +51,17 @@ fsdp_config:
 
 All you have to do now is launch using accelerate as you would usually do on each machine and voila, the processes will start once you have launched accelerate on every machine.
 
-# Raytrain
+## Raytrain
 
 Please see ray train doc [here](ray-integration.qmd).
 
-# Torchrun
+## Torchrun
 
 If you are using Infiniband, we recommend torchrun to utilize the full bandwidth.
 
 Set the following env (change buffersize/socketname depending on your system):
 
-```yaml
+```bash
 export NCCL_IB_DISABLE=0
 export NCCL_SOCKET_IFNAME="eth0,en,eth,em,bond"
 export NCCL_BUFFSIZE=2097152

docs/nccl.qmd

@@ -13,13 +13,13 @@ Often, this timeout will happen after 30 minutes (the default setting) and is ac
 
 Forcing cross-GPU communication via [NVLink](https://en.wikipedia.org/wiki/NVLink) may help without increasing timeouts. To verify that your configuration is leveraging NVLink run the following command:
 
-```shell
+```bash
 nvidia-smi nvlink --status
 ```
 
 To force NCCL to use NVLink, simply set this in the environment:
 
-```shell
+```bash
 export NCCL_P2P_LEVEL=NVL
 ```
 
@@ -33,13 +33,13 @@ If NVLink is not available in your environment there are other options for ``NCC
 
 To validate that acceptable data transfer speeds exist for your training job, running [NCCL Tests](https://github.com/NVIDIA/nccl-tests/blob/master/README.md) can help pinpoint bottlenecks, for example:
 
-```shell
+```bash
 ./build/all_reduce_perf -b 8 -e 128M -f 2 -g 3
 ```
 
 It can be useful when debugging NCCL communication timeouts to activate additional logging in both PyTorch and NCCL:
 
-```shell
+```bash
 export NCCL_DEBUG=INFO
 export NCCL_DEBUG_SUBSYS=ALL
 export TORCH_DISTRIBUTED_DEBUG=INFO

docs/ray-integration.qmd

@@ -1,5 +1,5 @@
 ---
-title: Ray Train integration
+title: Ray Train
 description: How to use Axolotl with Ray Train
 ---
 
@@ -9,7 +9,7 @@ With the `--use-ray` CLI flag, Axolotl will use Ray Train's [`TorchTrainer`](htt
 
 ## Ray cluster setup
 
-A prerequisite using the Ray Train integration is to setup a Ray cluster on your desired node(s). For a detailed guide on how you can get started with ray clusters, check the official Ray docs here: https://docs.ray.io/en/latest/cluster/getting-started.html
+A prerequisite using the Ray Train integration is to setup a Ray cluster on your desired node(s). For a detailed guide on how you can get started with ray clusters, check the official Ray docs [here](https://docs.ray.io/en/latest/cluster/getting-started.html).
 
 Every Ray cluster has one _head_ node and a set of worker nodes. The head node is just like any other worker node, but it also runs certain special processes related to scheduling and orchestration. Ray-enabled scripts are run on the head node and depending on the resources (number of CPUs, GPUs, etc) they request, will be scheduled to run certain tasks on the worker nodes. For more on key concepts behind a Ray cluster, you can refer this [doc](https://docs.ray.io/en/latest/cluster/key-concepts.html#cluster-key-concepts).
 
@@ -58,13 +58,11 @@ You can find an example configuration at `configs/llama-3/lora-1b-ray.yaml`.
 The key parameters to note here are:
 
 ```yaml
-...
 use_ray: true
 ray_num_workers: 4
 # optional
 resources_per_worker:
     GPU: 1
-...
 ```
 
 - `use_ray`: This is the flag that enables the Ray Train integration. You can either use the corresponding `--use-ray` flag in the CLI or set `use_ray` in the config file.

docs/rlhf.qmd

@@ -3,22 +3,22 @@ title: "RLHF (Beta)"
 description: "Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human feedback."
 back-to-top-navigation: true
 toc: true
-toc-depth: 3
+toc-depth: 4
 ---
 
-# Overview
+## Overview
 
 Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human
 feedback. Various methods include, but not limited to:
 
-- Proximal Policy Optimization (PPO) (not yet supported in axolotl)
 - [Direct Preference Optimization (DPO)](#dpo)
 - [Identity Preference Optimization (IPO)](#ipo)
 - [Kahneman-Tversky Optimization (KTO)](#kto)
 - [Odds Ratio Preference Optimization (ORPO)](#orpo)
+- Proximal Policy Optimization (PPO) (not yet supported in axolotl)
 
 
-# RLHF using Axolotl
+## RLHF using Axolotl
 
 ::: {.callout-important}
 This is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.
@@ -30,7 +30,7 @@ We rely on the [TRL](https://github.com/huggingface/trl) library for implementat
 You can find what each method supports by going into `src/axolotl/prompt_strategies/{method}` where `{method}` is one of our supported methods. The `type: ` can be retrieved from `{method}.{function_name}`.
 :::
 
-## DPO
+### DPO
 
 Example config:
 
@@ -47,7 +47,7 @@ datasets:
 
 DPO supports the following types with the following dataset format:
 
-### chatml.argilla
+#### chatml.argilla
 
 ```json
 {
@@ -58,7 +58,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### chatml.argilla_chat
+#### chatml.argilla_chat
 
 ```json
 {
@@ -73,7 +73,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### chatml.icr
+#### chatml.icr
 
 ```json
 {
@@ -84,7 +84,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### chatml.intel
+#### chatml.intel
 
 ```json
 {
@@ -95,7 +95,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### chatml.prompt_pairs
+#### chatml.prompt_pairs
 
 ```json
 {
@@ -106,7 +106,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### chatml.ultra
+#### chatml.ultra
 
 ```json
 {
@@ -123,7 +123,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### llama3.argilla
+#### llama3.argilla
 
 ```json
 {
@@ -134,7 +134,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### llama3.argilla_chat
+#### llama3.argilla_chat
 
 ```json
 {
@@ -149,7 +149,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### llama3.icr
+#### llama3.icr
 
 ```json
 {
@@ -160,7 +160,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### llama3.intel
+#### llama3.intel
 
 ```json
 {
@@ -171,7 +171,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### llama3.prompt_pairs
+#### llama3.prompt_pairs
 
 ```json
 {
@@ -182,7 +182,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### llama3.ultra
+#### llama3.ultra
 
 ```json
 {
@@ -199,7 +199,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### zephyr.nectar
+#### zephyr.nectar
 
 ```json
 {
@@ -218,7 +218,7 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
-### chat_template.default
+#### chat_template.default
 
 ```yaml
 rl: dpo
@@ -264,7 +264,7 @@ Sample input format:
 }
 ```
 
-### user_defined.default
+#### user_defined.default
 
 For custom behaviors,
 
@@ -295,7 +295,7 @@ The input format is a simple JSON input with customizable fields based on the ab
 }
 ```
 
-## IPO
+### IPO
 
 As IPO is just DPO with a different loss function, all supported options for DPO works here.
 
@@ -303,7 +303,7 @@ As IPO is just DPO with a different loss function, all supported options for DPO
 rl: ipo
 ```
 
-## ORPO
+### ORPO
 
 Paper: https://arxiv.org/abs/2403.07691
 
@@ -320,7 +320,7 @@ datasets:
 
 ORPO supports the following types with the following dataset format:
 
-### chat_template.argilla
+#### chat_template.argilla
 
 ```json
 {
@@ -339,7 +339,7 @@ ORPO supports the following types with the following dataset format:
 }
 ```
 
-## KTO
+### KTO
 
 ```yaml
 rl: kto
@@ -360,7 +360,7 @@ gradient_checkpointing_kwargs:
 
 KTO supports the following types with the following dataset format:
 
-### chatml.argilla
+#### chatml.argilla
 
 ```json
 {
@@ -370,7 +370,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-### chatml.argilla_chat
+#### chatml.argilla_chat
 
 ```json
 {
@@ -383,7 +383,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-### chatml.intel
+#### chatml.intel
 
 ```json
 {
@@ -393,7 +393,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-### chatml.prompt_pairs
+#### chatml.prompt_pairs
 
 ```json
 {
@@ -403,7 +403,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-### chatml.ultra
+#### chatml.ultra
 
 ```json
 {
@@ -413,7 +413,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-### llama3.argilla
+#### llama3.argilla
 
 ```json
 {
@@ -423,7 +423,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-### llama3.argilla_chat
+#### llama3.argilla_chat
 
 ```json
 {
@@ -434,7 +434,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-### llama3.intel
+#### llama3.intel
 
 ```json
 {
@@ -444,7 +444,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-### llama3.prompt_pairs
+#### llama3.prompt_pairs
 
 ```json
 {
@@ -454,7 +454,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-### llama3.ultra
+#### llama3.ultra
 
 ```json
 {
@@ -464,7 +464,7 @@ KTO supports the following types with the following dataset format:
 }
 ```
 
-### user_defined.default
+#### user_defined.default
 
 For custom behaviors,
 
@@ -494,7 +494,49 @@ The input format is a simple JSON input with customizable fields based on the ab
 }
 ```
 
-## Using local dataset files
+### GRPO
+
+GRPO uses custom reward functions and transformations. Please have them ready locally.
+
+For ex, to load OpenAI's GSM8K and use a random reward for completions:
+
+```python
+# rewards.py
+import random
+
+def rand_reward_func(completions, **kwargs) -> list[float]:
+    return [random.uniform(0, 1) for _ in completions]
+
+def oai_gsm8k_transform(cfg, *args, **kwargs):
+    def transform_fn(example, tokenizer=None):
+        label = example["answer"].split("####")[-1].strip().replace(",", "")
+        return {
+            "prompt": [{"role": "user", "content": example["question"]},],
+            "answer": label,
+        }
+    return transform_fn, {"remove_columns": ["question"]}
+```
+
+```yaml
+rl: grpo
+
+trl:
+    beta: 0.001
+    max_completion_length: 256
+    use_vllm: True
+    vllm_device: auto
+    vllm_gpu_memory_utilization: 0.15
+    num_generations: 4
+    reward_funcs: ["rewards.rand_reward_func"]    # format: '{file_name}.{fn_name}'
+datasets:
+  - path: openai/gsm8k
+    name: main
+    type: rewards.oai_gsm8k_transform  # format: '{file_name}.{fn_name}'
+```
+
+To see other examples of custom reward functions, please see [TRL GRPO Docs](https://github.com/huggingface/trl/blob/main/docs/source/grpo_trainer.md#using-a-custom-reward-function).
+
+### Using local dataset files
 
 ```yaml
 datasets:
@@ -505,7 +547,7 @@ datasets:
     type: chatml.intel
 ```
 
-## TRL auto-unwrapping for PEFT
+### TRL auto-unwrapping for PEFT
 
 TRL supports auto-unwrapping PEFT models for RL training paradigms which rely on a reference model. This significantly reduces memory pressure as an additional refreference model does not need to be loaded, and reference model log-probabilities can be obtained by disabling PEFT adapters. This is enabled by default. To turn it off, pass the following config:
 

docs/torchao.qmd

@@ -3,6 +3,12 @@ title: "PyTorch ao"
 description: "Custom data types and layouts for training and inference"
 ---
 
+To use experimental optimizers (`AdamWFp8`, `AdamW4bit`, `AdamW8bit`) from Pytorch Ao, please install the package as shown below.
+
+::: {.callout-tip}
+Some experimental optimizers are already present in regular Pytorch, so please re-check if you actually need this package!
+:::
+
 ### Installation
 
 Stable Release from the PyTorch index

docs/unsloth.qmd

@@ -8,6 +8,12 @@ description: "Hyper-optimized QLoRA finetuning for single GPUs"
 Unsloth provides hand-written optimized kernels for LLM finetuning that slightly improve speed and VRAM over
 standard industry baselines.
 
+::: {.callout-important}
+Due to breaking changes in transformers `v4.48.0`, users will need to downgrade to `<=v4.47.1` to use this patch.
+
+This will later be deprecated in favor of [LoRA Optimizations](lora_optims.qmd).
+:::
+
 
 ### Installation
 
@@ -17,7 +23,7 @@ The following will install the correct unsloth and extras from source.
 python scripts/unsloth_install.py | sh
 ```
 
-### Using unsloth w Axolotl
+### Usage
 
 Axolotl exposes a few configuration options to try out unsloth and get most of the performance gains.
 

index.qmd

@@ -1,7 +1,7 @@
 ---
-toc-location: right-body
-toc-title: Table Of Contents
-toc-expand: 2
+# toc-location: right-body
+# toc-title: Table Of Contents
+# toc-expand: 2
 ---
 
 ```{python}

requirements.txt

@@ -7,7 +7,7 @@ mamba-ssm==1.2.0.post1
 flash-attn==2.7.4.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.2
+liger-kernel==0.5.3
 # END section
 
 packaging==23.2
@@ -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/cloud/modal_.py

@@ -258,25 +258,21 @@ class ModalCloud(Cloud):
 
 
 def _preprocess(config_yaml: str, volumes=None):
-    Path("/workspace/artifacts/axolotl").mkdir(parents=True, exist_ok=True)
-    with open(
-        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
-    ) as f_out:
+    Path("/workspace/mounts").mkdir(parents=True, exist_ok=True)
+    with open("/workspace/mounts/config.yaml", "w", encoding="utf-8") as f_out:
         f_out.write(config_yaml)
-    run_folder = "/workspace/artifacts/axolotl"
+    run_folder = "/workspace/mounts"
     run_cmd(
-        "axolotl preprocess /workspace/artifacts/axolotl/config.yaml --dataset-processes=8",
+        "axolotl preprocess /workspace/mounts/config.yaml --dataset-processes=8",
         run_folder,
         volumes,
     )
 
 
 def _train(config_yaml: str, accelerate: bool = True, volumes=None, **kwargs):
-    with open(
-        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
-    ) as f_out:
+    with open("/workspace/mounts/config.yaml", "w", encoding="utf-8") as f_out:
         f_out.write(config_yaml)
-    run_folder = "/workspace/artifacts/axolotl"
+    run_folder = "/workspace/mounts"
     if accelerate:
         accelerate_args = "--accelerate"
     else:
@@ -285,20 +281,18 @@ def _train(config_yaml: str, accelerate: bool = True, volumes=None, **kwargs):
     if num_processes := kwargs.pop("num_processes", None):
         num_processes_args = f"--num-processes {num_processes}"
     run_cmd(
-        f"axolotl train {accelerate_args} {num_processes_args} /workspace/artifacts/axolotl/config.yaml",
+        f"axolotl train {accelerate_args} {num_processes_args} /workspace/mounts/config.yaml",
         run_folder,
         volumes,
     )
 
 
 def _lm_eval(config_yaml: str, volumes=None):
-    with open(
-        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
-    ) as f_out:
+    with open("/workspace/mounts/config.yaml", "w", encoding="utf-8") as f_out:
         f_out.write(config_yaml)
-    run_folder = "/workspace/artifacts/axolotl"
+    run_folder = "/workspace/mounts"
     run_cmd(
-        "axolotl lm-eval /workspace/artifacts/axolotl/config.yaml",
+        "axolotl lm-eval /workspace/mounts/config.yaml",
         run_folder,
         volumes,
     )

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/common/datasets.py

@@ -24,8 +24,8 @@ class TrainDatasetMeta:
     """Dataclass with fields for training and validation datasets and metadata."""
 
     train_dataset: Dataset
-    eval_dataset: Optional[Dataset] = None
-    total_num_steps: Optional[int] = None
+    eval_dataset: Dataset | None = None
+    total_num_steps: int | None = None
 
 
 def sample_dataset(dataset: Dataset, num_samples: int) -> Dataset:

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:
@@ -91,13 +93,11 @@ try:
 except ImportError:
     pass
 
-LOG = logging.getLogger("axolotl.core.trainer_builder")
+LOG = logging.getLogger(__name__)
 
 
 class TrainerBuilderBase(abc.ABC):
-    """
-    Base class for trainer builder
-    """
+    """Base class for trainer builder."""
 
     _train_dataset = None
     _eval_dataset = None
@@ -110,9 +110,9 @@ class TrainerBuilderBase(abc.ABC):
         self.tokenizer = tokenizer
         self.processor = processor
 
-        # in case the model supports tagging, add the axolotl tag.
+        # If the model supports tagging, add the axolotl tag.
         # This makes sure the tag is correctly pushed even if a user calls
-        # model.push_to_hub instad of  trainer.push_to_hub.
+        # model.push_to_hub instead of trainer.push_to_hub.
         if hasattr(model, "add_model_tags"):
             model.add_model_tags(["axolotl"])
 
@@ -227,8 +227,8 @@ class TrainerBuilderBase(abc.ABC):
 
 class HFCausalTrainerBuilder(TrainerBuilderBase):
     """
-    Build the HuggingFace training args/trainer for causal models
-    and reward modelling using TRL.
+    Build the HuggingFace training args/trainer for causal models and reward modeling
+    using TRL.
     """
 
     def get_callbacks(self):
@@ -551,30 +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"
@@ -658,46 +636,114 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.reward_model:
             training_arguments_kwargs["max_length"] = self.cfg.sequence_len
 
-        # pylint: disable=duplicate-code
-        if self.cfg.optimizer in [
-            "optimi_adamw",
-            "ao_adamw_4bit",
-            "ao_adamw_8bit",
-            "ao_adamw_fp8",
-            "adopt_adamw",
-        ]:
-            # Set default so transformers doesn't throw
-            training_arguments_kwargs["optim"] = "adamw_hf"
-            training_arguments_kwargs["alternate_optimizer"] = self.cfg.optimizer
+        # Handle custom optimizer
+        custom_supported_optimizers = [opt.value for opt in CustomSupportedOptimizers]
+        if self.cfg.optimizer in custom_supported_optimizers:
+            # Common optimizer kwargs
+            optimizer_kwargs = {
+                "lr": training_arguments_kwargs.get("learning_rate"),
+                "weight_decay": training_arguments_kwargs.get("weight_decay"),
+            }
 
-        if self.cfg.optimizer == "lion_pytorch":
-            from lion_pytorch import Lion
+            # Adam-specific kwargs
+            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 "weight_decay" in training_arguments_kwargs:
-                lion_kwargs["weight_decay"] = training_arguments_kwargs["weight_decay"]
-
-            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"],
+            if self.cfg.optimizer == "muon":
+                from axolotl.contribs.mit.muon import (  # pylint: disable=no-name-in-module
+                    MuonOptimizerFactory,
                 )
 
-            trainer_kwargs["optimizers"] = (
-                Lion(params=self.model.parameters(), **lion_kwargs),
-                None,
+                optimizer_cls = MuonOptimizerFactory
+                optimizer_kwargs.update(adam_kwargs)
+            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
-            training_arguments_kwargs["optim"] = "adamw_hf"
+        else:
+            # 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"
@@ -872,9 +918,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
 
 class HFRLTrainerBuilder(TrainerBuilderBase):
-    """
-    Trainer factory class for TRL-based RLHF trainers (e.g. DPO)
-    """
+    """Trainer factory class for TRL-based RLHF trainers (e.g. DPO)"""
 
     def get_callbacks(self):
         callbacks = super().get_callbacks()

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__(
-        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 create_optimizer_grouped_parameters(self, opt_model, optimizer_kwargs):
+    def create_optimizer_grouped_parameters(
+        self, opt_model, optimizer_kwargs
+    ) -> list[dict]:
         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
-            not in [
-                "optimi_adamw",
-                "ao_adamw_8bit",
-                "ao_adamw_4bit",
-                "ao_adamw_fp8",
-                "adopt_adamw",
-            ]
+            and self.optimizer_cls_and_kwargs 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,
-                opt_model,
+
+        if (
+            not self.optimizer
+            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 (
-                self.args.embedding_lr_scale is not None
-                or self.args.embedding_lr is not None
-                or self.args.lr_groups is not None
-            ):
-                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
+            else:
+                # Overwrite `params` in case it's created by `get_optimizer_cls_and_kwargs`
+                # e.g. for GaLore optimizer.
+                if "params" in optimizer_kwargs:
+                    optimizer_grouped_parameters = optimizer_kwargs.pop("params")
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    AdamW(
-                        optimizer_grouped_parameters, foreach=False, **optimizer_kwargs
+                # Overwrite `model` in case it's created by `get_optimizer_cls_and_kwargs`
+                # e.g. for LOMO optimizer.
+                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
-                    AdamW4bit(optimizer_grouped_parameters, **optimizer_kwargs)
+                self.optimizer = optimizer_cls(
+                    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
-                    AdamW8bit(optimizer_grouped_parameters, **optimizer_kwargs)
-                )
-            elif self.args.alternate_optimizer == "ao_adamw_fp8":
-                from torchao.prototype.low_bit_optim import AdamWFp8
+            if optimizer_cls.__name__ == "Adam8bit":
+                import bitsandbytes
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    AdamWFp8(optimizer_grouped_parameters, **optimizer_kwargs)
-                )
-            elif self.args.alternate_optimizer == "adopt_adamw":
-                from axolotl.utils.optimizers.adopt import ADOPT
+                manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    ADOPT(
-                        optimizer_grouped_parameters,
-                        decouple=True,
-                        **optimizer_kwargs,
-                    )
-                )
+                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
@@ -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 cfg.trl and cfg.trl.vllm_device:
-                grpo_args_kwargs["vllm_device"] = cfg.trl.vllm_device
-            else:
-                grpo_args_kwargs["vllm_device"] = "auto"
-            if cfg.trl and cfg.trl.vllm_gpu_memory_utilization:
+
+        if not hasattr(cfg, "trl") or not cfg.trl:
+            return grpo_args_kwargs
+
+        trl: TRLConfig = cfg.trl  # type: ignore
+
+        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
-            if cfg.trl and cfg.trl.vllm_max_model_len:
-                grpo_args_kwargs["vllm_max_model_len"] = cfg.trl.vllm_max_model_len
-        if cfg.trl and cfg.trl.num_generations:
-            grpo_args_kwargs["num_generations"] = cfg.trl.num_generations
-        if cfg.trl and cfg.trl.sync_ref_model:
-            grpo_args_kwargs["sync_ref_model"] = cfg.trl.sync_ref_model
-            if cfg.trl and cfg.trl.ref_model_mixup_alpha:
-                grpo_args_kwargs[
-                    "ref_model_mixup_alpha"
-                ] = cfg.trl.ref_model_mixup_alpha
-            if cfg.trl and cfg.trl.ref_model_sync_steps:
-                grpo_args_kwargs["ref_model_sync_steps"] = cfg.trl.ref_model_sync_steps
-        grpo_args_kwargs["max_completion_length"] = cfg.trl.max_completion_length
-        grpo_args_kwargs["log_completions"] = cfg.trl.log_completions
+                ] = trl.vllm_gpu_memory_utilization
+
+            if trl.vllm_max_model_len:
+                grpo_args_kwargs["vllm_max_model_len"] = trl.vllm_max_model_len
+
+        if trl.num_generations:
+            grpo_args_kwargs["num_generations"] = trl.num_generations
+
+        if trl.sync_ref_model:
+            grpo_args_kwargs["sync_ref_model"] = trl.sync_ref_model
+
+            if trl.ref_model_mixup_alpha:
+                grpo_args_kwargs["ref_model_mixup_alpha"] = trl.ref_model_mixup_alpha
+
+            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

@@ -1,6 +1,26 @@
 # Cut Cross Entropy
 
-### Usage
+Cut Cross Entropy reduces VRAM usage through optimization on the cross-entropy operation during loss calculation.
+
+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
 plugins:
@@ -8,3 +28,19 @@ plugins:
 
 cut_cross_entropy: true
 ```
+
+## Citation
+
+```bib
+@article{wijmans2024cut,
+  author       = {Erik Wijmans and
+                  Brody Huval and
+                  Alexander Hertzberg and
+                  Vladlen Koltun and
+                  Philipp Kr\"ahenb\"uhl},
+  title        = {Cut Your Losses in Large-Vocabulary Language Models},
+  journal      = {arXiv},
+  year         = {2024},
+  url          = {https://arxiv.org/abs/2411.09009},
+}
+```

src/axolotl/integrations/grokfast/README.md

@@ -2,7 +2,7 @@
 
 See https://github.com/ironjr/grokfast
 
-### Usage
+## Usage
 
 ```yaml
 plugins:
@@ -11,3 +11,14 @@ plugins:
 grokfast_alpha: 2.0
 grokfast_lamb: 0.98
 ```
+
+## Citation
+
+```bib
+@article{lee2024grokfast,
+    title={{Grokfast}: Accelerated Grokking by Amplifying Slow Gradients},
+    author={Lee, Jaerin and Kang, Bong Gyun and Kim, Kihoon and Lee, Kyoung Mu},
+    journal={arXiv preprint arXiv:2405.20233},
+    year={2024}
+}
+```

src/axolotl/integrations/kd/README.md

@@ -0,0 +1,23 @@
+# Knowledge Distillation
+
+## Usage
+
+```yaml
+plugins:
+  - "axolotl.integrations.kd.KDPlugin"
+
+kd_trainer: True
+kd_ce_alpha: 0.1
+kd_alpha: 0.9
+kd_temperature: 1.0
+
+torch_compile: True  # torch>=2.5.1, recommended to reduce vram
+
+datasets:
+  - path: ...
+    type: "axolotl.integrations.kd.chat_template"
+    field_messages: "messages_combined"
+    logprobs_field: "llm_text_generation_vllm_logprobs"  # for kd only, field of logprobs
+```
+
+An example dataset can be found at [`axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample`](https://huggingface.co/datasets/axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample)

src/axolotl/integrations/liger/README.md

@@ -0,0 +1,36 @@
+# Liger Kernel Integration
+
+Liger Kernel provides efficient Triton kernels for LLM training, offering:
+
+- 20% increase in multi-GPU training throughput
+- 60% reduction in memory usage
+- Compatibility with both FSDP and DeepSpeed
+
+See https://github.com/linkedin/Liger-Kernel
+
+## Usage
+
+```yaml
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_layer_norm: true
+liger_fused_linear_cross_entropy: true
+```
+
+## Citation
+
+```bib
+@article{hsu2024ligerkernelefficienttriton,
+      title={Liger Kernel: Efficient Triton Kernels for LLM Training},
+      author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},
+      year={2024},
+      eprint={2410.10989},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG},
+      url={https://arxiv.org/abs/2410.10989},
+      journal={arXiv preprint arXiv:2410.10989},
+}
+```

src/axolotl/integrations/lm_eval/README.md

@@ -1,6 +1,10 @@
 # LM Eval Harness
 
-### Usage
+Run evaluation on model using the popular lm-evaluation-harness library.
+
+See https://github.com/EleutherAI/lm-evaluation-harness
+
+## Usage
 
 ```yaml
 plugins:
@@ -10,4 +14,22 @@ lm_eval_tasks:
   - gsm8k
   - hellaswag
   - arc_easy
+
+lm_eval_batch_size: # Batch size for evaluation
+output_dir: # Directory to save evaluation results
+```
+
+## Citation
+
+```bib
+@misc{eval-harness,
+  author       = {Gao, Leo and Tow, Jonathan and Abbasi, Baber and Biderman, Stella and Black, Sid and DiPofi, Anthony and Foster, Charles and Golding, Laurence and Hsu, Jeffrey and Le Noac'h, Alain and Li, Haonan and McDonell, Kyle and Muennighoff, Niklas and Ociepa, Chris and Phang, Jason and Reynolds, Laria and Schoelkopf, Hailey and Skowron, Aviya and Sutawika, Lintang and Tang, Eric and Thite, Anish and Wang, Ben and Wang, Kevin and Zou, Andy},
+  title        = {A framework for few-shot language model evaluation},
+  month        = 07,
+  year         = 2024,
+  publisher    = {Zenodo},
+  version      = {v0.4.3},
+  doi          = {10.5281/zenodo.12608602},
+  url          = {https://zenodo.org/records/12608602}
+}
 ```

src/axolotl/integrations/spectrum/README.md

@@ -1,15 +1,17 @@
-## Spectrum: Targeted Training on Signal to Noise Ratio
+# Spectrum: Targeted Training on Signal to Noise Ratio
 
 by Eric Hartford, Lucas Atkins, Fernando Fernandes, David Golchinfar
 
 This plugin contains code to freeze the bottom fraction of modules in a model, based on the Signal-to-Noise Ratio (SNR).
 
-### Overview
+See https://github.com/cognitivecomputations/spectrum
+
+## Overview
 
 Spectrum is a tool for scanning and evaluating the Signal-to-Noise Ratio (SNR) of layers in large language models.
 By identifying the top n% of layers with the highest SNR, you can optimize training efficiency.
 
-### Usage
+## Usage
 
 ```yaml
 plugins:
@@ -19,3 +21,17 @@ spectrum_top_fraction: 0.5
 # Optional if using a pre-scanned model as your base_model. Useful if using a model mirror
 spectrum_model_name: meta-llama/Meta-Llama-3.1-8B
 ```
+
+## Citation
+
+```bib
+@misc{hartford2024spectrumtargetedtrainingsignal,
+      title={Spectrum: Targeted Training on Signal to Noise Ratio},
+      author={Eric Hartford and Lucas Atkins and Fernando Fernandes Neto and David Golchinfar},
+      year={2024},
+      eprint={2406.06623},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG},
+      url={https://arxiv.org/abs/2406.06623},
+}
+```

src/axolotl/monkeypatch/multipack.py

@@ -25,6 +25,7 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
     "gemmoe",
     "starcoder2",
     "deepseek_v2",
+    "deepseek_v3",
 ]
 
 

src/axolotl/train.py

@@ -1,26 +1,29 @@
 """Prepare and train a model on a dataset. Can also infer from a model or merge lora"""
 
+import importlib
 import inspect
 import os
 import signal
 import sys
 import weakref
 from pathlib import Path
-from typing import Tuple, Union
+from typing import Any
 
 import torch
 import transformers.modelcard
 from accelerate.logging import get_logger
 from accelerate.utils import save_fsdp_model
-from peft import PeftModel
-from pkg_resources import get_distribution  # type: ignore
-from transformers import PreTrainedModel, PreTrainedTokenizer
+from datasets import Dataset
+from peft import PeftConfig, PeftModel
+from transformers import PreTrainedModel, PreTrainedTokenizer, ProcessorMixin
 from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
+from transformers.trainer import Trainer
 
 from axolotl.common.datasets import TrainDatasetMeta
 from axolotl.contribs.lgpl.unsloth import (  # pylint: disable = no-name-in-module
     fix_untrained_tokens,
 )
+from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
 from axolotl.logging_config import configure_logging
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.freeze import freeze_layers_except
@@ -32,17 +35,25 @@ try:
 except ImportError:
     BetterTransformer = None
 
-project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
-src_dir = os.path.join(project_root, "src")
-sys.path.insert(0, src_dir)
-
 configure_logging()
 LOG = get_logger(__name__)
 
 
-def train(
-    *, cfg: DictDefault, dataset_meta: TrainDatasetMeta
-) -> Tuple[Union[PeftModel, PreTrainedModel], PreTrainedTokenizer]:
+def setup_model_and_tokenizer(
+    cfg: DictDefault,
+) -> tuple[
+    PreTrainedModel, PreTrainedTokenizer, PeftConfig | None, ProcessorMixin | None
+]:
+    """
+    Load the tokenizer, processor (for multimodal models), and model based on configuration.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+
+    Returns:
+        Tuple containing model, tokenizer, `peft_config` (if LoRA / QLoRA, else
+            `None`), and processor (if multimodal, else `None`).
+    """
     # Load tokenizer
     LOG.debug(
         f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}",
@@ -55,11 +66,58 @@ def train(
     if cfg.is_multimodal:
         processor = load_processor(cfg, tokenizer)
 
-    # Get datasets
-    train_dataset = dataset_meta.train_dataset
-    eval_dataset = dataset_meta.eval_dataset
-    total_num_steps = dataset_meta.total_num_steps
+    # Load the model and peft_config
+    msg = "loading model"
+    if cfg.adapter:
+        msg += " and peft_config..."
+    LOG.debug(msg)
 
+    model, peft_config = load_model(cfg, tokenizer, processor=processor)
+    if model.generation_config is not None:
+        model.generation_config.do_sample = True
+
+    # Apply freezing if specified
+    if cfg.unfrozen_parameters:
+        freeze_layers_except(model, cfg.unfrozen_parameters)
+
+    return model, tokenizer, peft_config, processor
+
+
+def setup_reference_model(
+    cfg: DictDefault, tokenizer: PreTrainedTokenizer
+) -> PreTrainedModel | None:
+    """
+    Set up the reference model for RL training if needed.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        tokenizer: The tokenizer to use for the reference model.
+
+    Returns:
+        Reference model if needed for RL training, `None` otherwise.
+    """
+    model_ref = None
+    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")
+            model_ref = None  # explicit setting to None
+        else:
+            # load the model again for model_ref/baseline
+            model_ref, _ = load_model(cfg, tokenizer, reference_model=True)
+    return model_ref
+
+
+def determine_resume_checkpoint(cfg: DictDefault) -> str | None:
+    """
+    Determine the checkpoint to resume from based on configuration.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+
+    Returns:
+        Path to the checkpoint to resume from, or `None` if not resuming.
+    """
     if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
         possible_checkpoints = [
             str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
@@ -73,77 +131,22 @@ def train(
             LOG.info(
                 f"Using Auto-resume functionality to start with checkpoint at {cfg.resume_from_checkpoint}"
             )
-    resume_from_checkpoint = cfg.resume_from_checkpoint
+    return cfg.resume_from_checkpoint
 
-    # Load the model and tokenizer
-    msg = "loading model"
-    if cfg.adapter:
-        msg += " and peft_config..."
-    LOG.debug(msg)
-    model, peft_config = load_model(cfg, tokenizer, processor=processor)
-    if model.generation_config is not None:
-        model.generation_config.do_sample = True
 
-    model_ref = None
-    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")
-            model_ref = None  # explicit setting to None
-        else:
-            # load the model again for model_ref/baseline
-            model_ref, _ = load_model(cfg, tokenizer, reference_model=True)
+def setup_signal_handler(
+    cfg: DictDefault, model: PreTrainedModel, safe_serialization: bool
+):
+    """
+    Set up signal handler for graceful termination.
 
-    safe_serialization = cfg.save_safetensors is True
-
-    if cfg.unfrozen_parameters:
-        freeze_layers_except(model, cfg.unfrozen_parameters)
-
-    trainer = setup_trainer(
-        cfg,
-        train_dataset,
-        eval_dataset,
-        (model, model_ref, peft_config),
-        tokenizer,
-        processor,
-        total_num_steps,
-    )
-
-    if cfg.fix_untrained_tokens:
-        # check if the `token_ids_to_fix` kwarg exists in the fix_untrained_tokens args
-        sig = inspect.signature(fix_untrained_tokens)
-        # if the function has the `token_ids_to_fix` arg, and fix_untrained_tokens is a list
-        if "token_ids_to_fix" in sig.parameters and isinstance(
-            cfg.fix_untrained_tokens, list
-        ):
-            fix_untrained_tokens(
-                model,
-                tokenizer,
-                train_dataset,
-                token_ids_to_fix=cfg.fix_untrained_tokens,
-            )
-        else:
-            fix_untrained_tokens(model, tokenizer, train_dataset)
-        if cfg.local_rank == 0:
-            model.save_pretrained(
-                str(Path(cfg.output_dir)), safe_serialization=safe_serialization
-            )
-
-    # 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}")
-        peft_config.save_pretrained(cfg.output_dir)
-    # additionally presave the tokenizer and model configs
-    if not Path(cfg.output_dir).is_dir():
-        os.makedirs(cfg.output_dir, exist_ok=True)
-    tokenizer.save_pretrained(str(Path(cfg.output_dir)))
-    if hasattr(model, "config"):
-        model.config.save_pretrained(str(Path(cfg.output_dir)))
-
-    # In case we want to stop early with ctrl+c, this is a nice to have to save the pretrained model
-    if (
-        cfg.local_rank == 0 and not cfg.use_ray
-    ):  # ray workers don't have access to this signal
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        model: The model to save on termination
+        safe_serialization: Whether to use safe serialization when saving
+    """
+    # ray workers don't have access to this signal
+    if cfg.local_rank == 0 and not cfg.use_ray:
 
         def terminate_handler(_, __, model_weakref):
             if model_weakref() is not None:
@@ -161,21 +164,22 @@ def train(
             lambda signum, frame: terminate_handler(signum, frame, _model_weakref),
         )
 
-    badge_markdown = """[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)"""
-    transformers.modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n{badge_markdown}"
 
-    if getattr(cfg, "axolotl_config_path"):
-        raw_axolotl_cfg = Path(cfg.axolotl_config_path)
-        version = get_distribution("axolotl").version
-        if raw_axolotl_cfg.is_file():
-            transformers.modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n<details><summary>See axolotl config</summary>\n\naxolotl version: `{version}`\n```yaml\n{raw_axolotl_cfg.read_text(encoding='utf-8')}\n```\n\n</details><br>\n"
+def execute_training(
+    cfg: DictDefault, trainer: Any, resume_from_checkpoint: str | None
+):
+    """
+    Execute the training process with appropriate backend configurations.
 
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        trainer: The configured trainer object.
+        resume_from_checkpoint: Path to checkpoint to resume from, if applicable.
+    """
     LOG.info("Starting trainer...")
     if cfg.group_by_length:
         LOG.info("hang tight... sorting dataset for group_by_length")
 
-    pretrain_hooks(cfg, trainer)
-
     if cfg.flash_optimum:
         with torch.backends.cuda.sdp_kernel(
             # TODO configure these from the YAML w/ sdp_kernel_kwargs: ...
@@ -187,15 +191,30 @@ def train(
     else:
         trainer.train(resume_from_checkpoint=resume_from_checkpoint)
 
-    post_train_hooks(cfg, trainer)
 
-    LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
+def save_trained_model(
+    cfg: DictDefault,
+    trainer: Any,
+    model: PreTrainedModel,
+    safe_serialization: bool,
+):
+    """
+    Save the trained model according to configuration and training setup.
 
-    # post training
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        trainer: The trainer object.
+        model: The trained model to save.
+        safe_serialization: Whether to use safe serialization.
+    """
+    LOG.info(f"Training completed! Saving pre-trained model to {cfg.output_dir}.")
+
+    # Post training module hooks
     for name, module in model.named_modules():
         if hasattr(module, "_post_training"):
             module._post_training(model, name)  # pylint: disable=protected-access
 
+    # Handle FSDP state dict type
     state_dict_type = "FULL_STATE_DICT"
     if trainer.is_fsdp_enabled:
         if cfg.fsdp_final_state_dict_type:
@@ -203,16 +222,18 @@ def train(
         trainer.accelerator.state.fsdp_plugin.set_state_dict_type(state_dict_type)
         LOG.info(f"Set FSDP state dict type to {state_dict_type} for saving.")
 
+    # Handle ReLoRA early return case
     if cfg.relora_steps:
         if cfg.adapter == "lora" and not (cfg.load_in_4bit or cfg.load_in_8bit):
             model = model.merge_and_unload()
         else:
             # final model weights have already been saved by `ReLoRACallback.on_train_end`
-            return model, tokenizer
+            return
 
-    # TODO do we need this fix? https://huggingface.co/docs/accelerate/usage_guides/fsdp#saving-and-loading
-    # only save on rank 0, otherwise it corrupts output on multi-GPU when multiple processes attempt to write the same file
     if cfg.fsdp:
+        # TODO: do we need this fix? https://huggingface.co/docs/accelerate/usage_guides/fsdp#saving-and-loading
+        # only save on rank 0, otherwise it corrupts output on multi-GPU when multiple
+        # processes attempt to write the same file
         if (
             state_dict_type == "SHARDED_STATE_DICT"
             and cfg.fsdp_config.fsdp_state_dict_type == "SHARDED_STATE_DICT"
@@ -244,7 +265,6 @@ def train(
                 os.remove(os.path.join(cfg.output_dir, "model.safetensors"))
             except FileNotFoundError:
                 pass
-
     elif cfg.local_rank == 0:
         if cfg.flash_optimum and BetterTransformer:
             model = BetterTransformer.reverse(model)
@@ -255,58 +275,239 @@ def train(
             )
         model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
 
+
+def create_model_card(cfg: DictDefault, trainer: Trainer):
+    """
+    Create a model card for the trained model if needed.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        trainer: The trainer object with model card creation capabilities.
+    """
     if not cfg.hub_model_id:
+        # Guard since create_model_card may fail if dataset_tags is empty list
         try:
             model_card_kwarg = {
                 "model_name": cfg.output_dir.lstrip("./")
                 .encode("utf-8")
                 .decode("utf-8")
             }
-            if cfg.datasets is not None:
-                if cfg.rl is not None or cfg.reward_model or cfg.process_reward_model:
-                    dataset_tags = [
-                        d["path"] for d in cfg.datasets if not Path(d["path"]).is_dir()
-                    ]
-                    dataset_tags = [
-                        d for d in dataset_tags if not d.startswith("https://")
-                    ]
-                    if dataset_tags:
-                        # guard as create_model_card may fail if dataset_tags is empty list
-                        model_card_kwarg["dataset_name"] = dataset_tags
-                else:
-                    dataset_tags = [
-                        d["path"] for d in cfg.datasets if not Path(d["path"]).is_dir()
-                    ]
-                    dataset_tags = [
-                        d for d in dataset_tags if not d.startswith("https://")
-                    ]
-                    if dataset_tags:
-                        # guard as create_model_card may fail if dataset_tags is empty list
-                        model_card_kwarg["dataset_tags"] = dataset_tags
+
+            # We check if we're using a TRL trainer; if so, `dataset_tags` is not consumed.
+            rl = cfg.rl is not None or cfg.reward_model or cfg.process_reward_model
+            if cfg.datasets is not None and not rl:
+                dataset_tags = [
+                    d["path"] for d in cfg.datasets if not Path(d["path"]).is_dir()
+                ]
+                dataset_tags = [d for d in dataset_tags if not d.startswith("https://")]
+
+                if dataset_tags:
+                    model_card_kwarg["dataset_tags"] = dataset_tags
 
             trainer.create_model_card(**model_card_kwarg)
         except (AttributeError, UnicodeDecodeError):
             pass
     elif cfg.hub_model_id:
-        # defensively push to the hub to ensure the model card is updated
+        # Defensively push to the hub to ensure the model card is updated
         trainer.push_to_hub()
 
-    return model, tokenizer
+
+def save_initial_configs(
+    cfg: DictDefault,
+    tokenizer: PreTrainedTokenizer,
+    model: PreTrainedModel,
+    peft_config: PeftConfig | None,
+):
+    """
+    Save initial configurations before training.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        tokenizer: The tokenizer to save.
+        model: The model to save configuration for.
+        peft_config: The PEFT configuration to save if applicable.
+    """
+    # Create output_dir if it doesn't already exist
+    output_dir = Path(cfg.output_dir)
+    if not output_dir.is_dir():
+        os.makedirs(cfg.output_dir, exist_ok=True)
+
+    # Pre-save adapter config so it's available to inspect
+    if peft_config:
+        LOG.info(f"Pre-saving adapter config to {cfg.output_dir}...")
+        peft_config.save_pretrained(cfg.output_dir)
+
+    # Pre-save the tokenizer and model configs
+    LOG.info(f"Pre-saving tokenizer to {cfg.output_dir}...")
+    tokenizer.save_pretrained(str(output_dir))
+    if hasattr(model, "config"):
+        LOG.info(f"Pre-saving model config to {cfg.output_dir}...")
+        model.config.save_pretrained(str(output_dir))
 
 
-def pretrain_hooks(_cfg, _trainer):
+def setup_model_card(cfg: DictDefault):
     """
-    Run hooks right before kicking off the training
-    :param cfg:
-    :param trainer:
-    :return:
+    Set up the Axolotl badge and add the Axolotl config to the model card if available.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
     """
+    badge_markdown = """[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)"""
+    transformers.modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n{badge_markdown}"
+
+    if getattr(cfg, "axolotl_config_path"):
+        raw_axolotl_cfg = Path(cfg.axolotl_config_path)
+        version = importlib.metadata.version("axolotl")
+        if raw_axolotl_cfg.is_file():
+            transformers.modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n<details><summary>See axolotl config</summary>\n\naxolotl version: `{version}`\n```yaml\n{raw_axolotl_cfg.read_text(encoding='utf-8')}\n```\n\n</details><br>\n"
 
 
-def post_train_hooks(_cfg, _trainer):
+def handle_untrained_tokens_fix(
+    cfg: DictDefault,
+    model: PreTrainedModel,
+    tokenizer: PreTrainedTokenizer,
+    train_dataset: Dataset,
+    safe_serialization: bool,
+):
     """
-    Run hooks right after training completes
-    :param cfg:
-    :param trainer:
-    :return:
+    Apply fixes for untrained tokens if configured.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        model: The model to apply fixes to.
+        tokenizer: The tokenizer for token identification.
+        train_dataset: The training dataset to use.
+        safe_serialization: Whether to use safe serialization when saving.
     """
+    if not cfg.fix_untrained_tokens:
+        return
+
+    # Check if the `token_ids_to_fix` kwarg exists in the fix_untrained_tokens args
+    sig = inspect.signature(fix_untrained_tokens)
+
+    # If the function has the `token_ids_to_fix` arg, and fix_untrained_tokens is a list
+    if "token_ids_to_fix" in sig.parameters and isinstance(
+        cfg.fix_untrained_tokens, list
+    ):
+        fix_untrained_tokens(
+            model,
+            tokenizer,
+            train_dataset,
+            token_ids_to_fix=cfg.fix_untrained_tokens,
+        )
+    else:
+        fix_untrained_tokens(model, tokenizer, train_dataset)
+
+    if cfg.local_rank == 0:
+        model.save_pretrained(
+            str(Path(cfg.output_dir)), safe_serialization=safe_serialization
+        )
+
+
+def setup_model_and_trainer(
+    cfg: DictDefault, dataset_meta: TrainDatasetMeta
+) -> tuple[
+    HFRLTrainerBuilder | HFCausalTrainerBuilder,
+    PeftModel | PreTrainedModel,
+    PreTrainedTokenizer,
+    PeftConfig | None,
+]:
+    """
+    Load model, tokenizer, trainer, etc. Helper function to encapsulate the full
+    trainer setup.
+
+    Args:
+        cfg: The configuration dictionary with training parameters.
+        dataset_meta: Object with training, validation datasets and metadata.
+
+    Returns:
+        Tuple of:
+            - Trainer (Causal or RLHF)
+            - Model
+            - Tokenizer
+            - PEFT config
+    """
+    # Load tokenizer, processor and model
+    model, tokenizer, peft_config, processor = setup_model_and_tokenizer(cfg)
+
+    # Set up reference model for RL if needed
+    model_ref = setup_reference_model(cfg, tokenizer)
+
+    # Get datasets from metadata
+    train_dataset = dataset_meta.train_dataset
+    eval_dataset = dataset_meta.eval_dataset
+    total_num_steps = dataset_meta.total_num_steps
+
+    # Set up trainer
+    trainer = setup_trainer(
+        cfg=cfg,
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        model=model,
+        tokenizer=tokenizer,
+        processor=processor,
+        total_num_steps=total_num_steps,
+        model_ref=model_ref,
+        peft_config=peft_config,
+    )
+
+    return (
+        trainer,
+        model,
+        tokenizer,
+        peft_config,
+    )
+
+
+def train(
+    cfg: DictDefault, dataset_meta: TrainDatasetMeta
+) -> tuple[PeftModel | PreTrainedModel, PreTrainedTokenizer, Trainer]:
+    """
+    Train a model on the given dataset.
+
+    Args:
+        cfg: The configuration dictionary with training parameters
+        dataset_meta: Object with training, validation datasets and metadata
+
+    Returns:
+        Tuple of (model, tokenizer) after training
+    """
+    # Setup model, tokenizer, (causal or RLHF) trainer etc.
+    (
+        trainer,
+        model,
+        tokenizer,
+        peft_config,
+    ) = setup_model_and_trainer(cfg, dataset_meta)
+
+    # Determine if we need to resume from a checkpoint
+    resume_from_checkpoint = determine_resume_checkpoint(cfg)
+
+    # Configuration for saving
+    safe_serialization = cfg.save_safetensors is True
+
+    # Handle untrained tokens if configured
+    train_dataset = dataset_meta.train_dataset
+    handle_untrained_tokens_fix(
+        cfg, model, tokenizer, train_dataset, safe_serialization
+    )
+
+    # Save initial configs
+    save_initial_configs(cfg, tokenizer, model, peft_config)
+
+    # Set up signal handler for graceful termination
+    setup_signal_handler(cfg, model, safe_serialization)
+
+    # Set up badges and config info for model card
+    setup_model_card(cfg)
+
+    # Execute the training
+    execute_training(cfg, trainer, resume_from_checkpoint)
+
+    # Save the trained model
+    save_trained_model(cfg, trainer, model, safe_serialization)
+
+    # Create model card
+    create_model_card(cfg, trainer)
+
+    return model, tokenizer, trainer

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

@@ -55,6 +55,7 @@ class ChatTemplate(str, Enum):
     phi_3 = "phi_3"  # pylint: disable=invalid-name
     phi_35 = "phi_35"  # pylint: disable=invalid-name
     deepseek_v2 = "deepseek_v2"  # pylint: disable=invalid-name
+    deepseek_v3 = "deepseek_v3"  # pylint: disable=invalid-name
     jamba = "jamba"  # pylint: disable=invalid-name
     jinja = "jinja"  # pylint: disable=invalid-name
     qwen_25 = "qwen_25"  # pylint: disable=invalid-name
@@ -63,6 +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"""
 
@@ -493,17 +506,7 @@ class HyperparametersConfig(BaseModel):
     embedding_lr_scale: Optional[float] = None
     weight_decay: Optional[float] = 0.0
     optimizer: Optional[
-        Union[
-            OptimizerNames,
-            Literal[
-                "lion_pytorch",
-                "optimi_adamw",
-                "ao_adamw_4bit",
-                "ao_adamw_8bit",
-                "ao_adamw_fp8",
-                "adopt_adamw",
-            ],
-        ]
+        Union[OptimizerNames, CustomSupportedOptimizers]
     ] = OptimizerNames.ADAMW_HF
     optim_args: Optional[Union[str, Dict[str, Any]]] = Field(
         default=None,
@@ -517,7 +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
@@ -1176,6 +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

src/axolotl/utils/trainer.py

@@ -574,14 +574,40 @@ def prepare_opinionated_env(cfg):
 
 
 def setup_trainer(
-    cfg, train_dataset, eval_dataset, model, tokenizer, processor, total_num_steps
+    cfg,
+    train_dataset,
+    eval_dataset,
+    model,
+    tokenizer,
+    processor,
+    total_num_steps,
+    model_ref=None,
+    peft_config=None,
 ):
+    """
+    Helper method for instantiating and building a (causal or RLHF) trainer.
+
+    Args:
+        cfg: Axolotl config object containing training parameters.
+        train_dataset: Dataset to use for training.
+        eval_dataset: Dataset to use for evaluation.
+        model: The model to train.
+        tokenizer: Tokenizer for processing text input.
+        processor: Processor for data preparation.
+        total_num_steps: The total number of training steps.
+        model_ref: Optional reference model for RLHF training. Default is None.
+        peft_config: Optional PEFT (Parameter-Efficient Fine-Tuning) configuration. Default is None.
+
+    Returns:
+        A trainer instance (either `HFRLTrainer` or `HFCausalTrainer`) configured based
+            on the provided parameters.
+    """
     if cfg.rl:
-        trainer_builder = HFRLTrainerBuilder(cfg, model[0], tokenizer, processor)
-        trainer_builder.model_ref = model[1]
-        trainer_builder.peft_config = model[2]
+        trainer_builder = HFRLTrainerBuilder(cfg, model, tokenizer, processor)
+        trainer_builder.model_ref = model_ref
+        trainer_builder.peft_config = peft_config
     else:
-        trainer_builder = HFCausalTrainerBuilder(cfg, model[0], tokenizer, processor)
+        trainer_builder = HFCausalTrainerBuilder(cfg, model, tokenizer, processor)
 
     trainer_builder.train_dataset = train_dataset
     trainer_builder.eval_dataset = eval_dataset

styles.css

@@ -1,5 +1,193 @@
-/* css styles */
+/* TYPOGRAPHY SECTION */
 
-img[alt="Axolotl"] {
-    content: url("https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/887513285d98132142bf5db2a74eb5e0928787f1/image/axolotl_logo_digital_black.svg") !important;
+/* Import fonts */
+@import url('https://fonts.googleapis.com/css2?family=Be+Vietnam+Pro:wght@400;500&display=swap');
+@import url('https://fonts.googleapis.com/css2?family=JetBrains+Mono:wght@400&display=swap');
+
+/* Typography hierarchy */
+:root {
+    --font-title: 'Be Vietnam Pro', sans-serif;
+    --font-body: 'JetBrains Mono', monospace;
+}
+
+/* Title (h1) */
+h1 {
+    font-family: var(--font-title);
+    font-weight: 400;
+    font-size: 6rem;
+    line-height: 1.1;
+    letter-spacing: -0.05em;
+    font-feature-settings: "ss01" on;
+}
+
+/* Heading (h2) */
+h2 {
+    font-family: var(--font-title);
+    font-weight: 500;
+    font-size: 2rem;
+    line-height: 1.2;
+    letter-spacing: -0.03em;
+    font-feature-settings: "ss01" on;
+}
+
+/* Subtitle/Preamble */
+h3,
+h4 {
+    font-family: var(--font-body);
+    font-weight: 400;
+    font-size: 1.5rem;
+    line-height: 1.5;
+    letter-spacing: -0.02em;
+}
+
+/* Body text */
+body {
+    font-family: var(--font-body);
+    font-weight: 400;
+    font-size: 1rem;
+    line-height: 1.5;
+    letter-spacing: -0.02em;
+}
+
+/* Links */
+a {
+    font-family: var(--font-body);
+    font-weight: 400;
+    font-size: 0.875rem;
+    line-height: 1;
+    letter-spacing: -0.02em;
+}
+
+/* NAV BAR SECTION */
+
+/* Navbar logo styling */
+.navbar-brand img {
+    height: 32px;
+    margin-right: 10px;
+}
+
+/* COLORS SECTION */
+
+/* Brand colors */
+:root {
+    --white: #ffffff;
+    --greige-300: #EEEEE7;
+    --greige-600: #CCCAC0;
+    --black: #141310;
+    --lime: #E3F8A8;
+    --cyan: #A0F4EA;
+    --purple: #C8D0F8;
+}
+
+/* Base styles */
+body {
+    background-color: var(--black);
+    color: var(--greige-300);
+}
+
+/* Navigation */
+.navbar {
+    background-color: var(--black) !important;
+}
+
+.navbar-dark .navbar-nav .nav-link {
+    color: var(--greige-300);
+}
+
+.navbar-dark .navbar-nav .nav-link:hover {
+    color: var(--lime);
+}
+
+/* Sidebar */
+.sidebar-navigation {
+    background-color: var(--black);
+    border-right: 1px solid var(--greige-600);
+}
+
+.sidebar nav[role="doc-toc"] ul>li>a {
+    color: var(--greige-300);
+}
+
+.sidebar nav[role="doc-toc"] ul>li>a:hover {
+    color: var(--lime);
+}
+
+/* Links */
+a {
+    color: var(--lime);
+}
+
+a:hover {
+    color: var(--cyan);
+}
+
+/* Headers */
+h1,
+h2,
+h3,
+h4,
+h5,
+h6 {
+    color: var(--white);
+}
+
+/* Code blocks */
+pre {
+    background-color: #1a1a1a !important;
+    border: 1px solid var(--greige-600);
+}
+
+/* Tables */
+.table {
+    color: var(--greige-300);
+}
+
+/* TOC */
+#toc-title {
+    color: var(--white);
+}
+
+.toc-active {
+    color: var(--lime) !important;
+}
+
+/* Buttons */
+.btn-primary {
+    background-color: var(--lime);
+    color: var(--black);
+    border: none;
+}
+
+.btn-primary:hover {
+    background-color: var(--cyan);
+    color: var(--black);
+}
+
+/* For inline code (single backtick) */
+code {
+    background-color: #1a1a1a !important;
+    color: var(--lime) !important;
+    padding: 2px 4px;
+    border-radius: 4px;
+}
+
+/* For inline code that is also a link */
+a code {
+    color: var(--cyan) !important;
+}
+
+/* For code blocks (triple backtick) */
+pre.sourceCode {
+    background-color: #1a1a1a !important;
+}
+
+/* Make comments in bash/shell scripts green */
+code span.co {
+    color: #5cb85c !important;
+}
+
+/* Remove underlines from JSON comments and make them green */
+code span.er {
+    color: #5cb85c !important;
+    text-decoration: none !important;
 }

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_deepseekv3.py

@@ -0,0 +1,130 @@
+"""
+E2E tests for lora llama
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import pytest
+
+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
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestDeepseekV3:
+    """
+    Test case for DeepseekV3 models
+    """
+
+    @pytest.mark.parametrize(
+        "sample_packing",
+        [True, False],
+    )
+    def test_lora_deepseekv3(self, temp_dir, sample_packing):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "axolotl-ai-co/DeepSeek-V3-11M",
+                "trust_remote_code": True,
+                "sample_packing": sample_packing,
+                "flash_attention": True,
+                "sequence_len": 2048,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0,
+                "datasets": [
+                    {
+                        "path": "mlabonne/FineTome-100k",
+                        "type": "chat_template",
+                        "field_messages": "conversations",
+                        "message_property_mappings": {
+                            "role": "from",
+                            "content": "value",
+                        },
+                        "drop_system_message": True,
+                        "split": "train[:1%]",
+                    },
+                ],
+                "special_tokens": {
+                    "bos_token": "<｜begin▁of▁sentence｜>",
+                    "eos_token": "<｜end▁of▁sentence｜>",
+                },
+                "chat_template": "deepseek_v3",
+                "num_epochs": 1,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 5,
+                "save_safetensors": True,
+                "bf16": True,
+            }
+        )
+        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)
+        assert (Path(temp_dir) / "adapter_model.safetensors").exists()
+
+    @pytest.mark.parametrize(
+        "sample_packing",
+        [True, False],
+    )
+    def test_fft_deepseekv3(self, temp_dir, sample_packing):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "axolotl-ai-co/DeepSeek-V3-11M",
+                "trust_remote_code": True,
+                "sample_packing": sample_packing,
+                "flash_attention": True,
+                "sequence_len": 2048,
+                "val_set_size": 0,
+                "datasets": [
+                    {
+                        "path": "mlabonne/FineTome-100k",
+                        "type": "chat_template",
+                        "field_messages": "conversations",
+                        "message_field_role": "from",
+                        "message_field_content": "value",
+                        "split": "train[:1%]",
+                    },
+                ],
+                "chat_template": "deepseek_v3",
+                "special_tokens": {
+                    "bos_token": "<｜begin▁of▁sentence｜>",
+                    "eos_token": "<｜end▁of▁sentence｜>",
+                },
+                "num_epochs": 1,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 5,
+                "save_safetensors": True,
+                "bf16": True,
+            }
+        )
+        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)
+        assert (Path(temp_dir) / "model.safetensors").exists()

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)