ungate lora with bias

* patch transformers to allow CP + FA2

* nits

* only patch in CP > 1 case

.bandit

@@ -1,3 +1,3 @@
 [bandit]
 exclude = tests
-skips = B101,B615
+skips = B101,B615,B102,B110

.coderabbit.yaml

@@ -12,5 +12,6 @@ reviews:
   auto_review:
     enabled: true
     drafts: false
+    auto_incremental_review: false
 chat:
   auto_reply: true

.flake8

@@ -1,5 +0,0 @@
-[flake8]
-max-line-length = 88
-
-select = C,E,F,W,B,B950
-extend-ignore = E203, E501, W503

.github/CONTRIBUTING.md

@@ -62,7 +62,7 @@ We welcome ideas for improvements and new features. To suggest an enhancement, o
 You can skip certain CI checks by including specific keywords in your commit messages:
 
 - `[skip ci]` or `skip ci` - Skips all CI checks for that commit
-- `[skip-e2e]` or `skip-e2e` - Skips only end-to-end tests while running other CI checks
+- `[skip-e2e]` or `skip-e2e` - Skips only end-to-end tests while running other CI checks. You may also include this in the title of your PR to disable end-to-end tests for the entire PR.
 
 ## Style Guidelines
 

.github/workflows/main.yml

@@ -36,6 +36,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.1
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -110,6 +115,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.1
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -169,6 +179,12 @@ jobs:
             pytorch: 2.7.1
             axolotl_extras: vllm
             is_latest: true
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            axolotl_extras:
+            is_latest:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

.github/workflows/multi-gpu-e2e.yml

@@ -33,13 +33,6 @@ jobs:
             axolotl_extras:
             num_gpus: 2
             nightly_build: "true"
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.0
-            axolotl_extras:
-            num_gpus: 2
-            nightly_build: "true"
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -47,6 +40,13 @@ jobs:
             axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            axolotl_extras: fbgemm-gpu
+            num_gpus: 2
+            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:

.github/workflows/tests.yml

@@ -55,7 +55,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.0", "2.7.1"]
+        pytorch_version: ["2.6.0", "2.7.1", "2.8.0"]
     timeout-minutes: 20
 
     steps:
@@ -130,7 +130,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.0", "2.7.1"]
+        pytorch_version: ["2.6.0", "2.7.1", "2.8.0"]
     timeout-minutes: 20
 
     steps:
@@ -188,13 +188,44 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
+  gate-skip-e2e:
+    needs: [pre-commit, pytest, pytest-sdist]
+    runs-on: ubuntu-latest
+    outputs:
+      skip: ${{ steps.compute.outputs.skip }}
+    steps:
+      - uses: actions/github-script@v7
+        id: compute
+        with:
+          script: |
+            const token = /\[skip-e2e\]/i;
+            let msg = '';
+            if (context.eventName === 'push') {
+              msg = context.payload.head_commit?.message || '';
+            } else if (context.eventName === 'pull_request') {
+              const { owner, repo } = context.repo;
+              const prNumber = context.payload.pull_request.number;
+              const commits = await github.paginate(
+                github.rest.pulls.listCommits,
+                { owner, repo, pull_number: prNumber, per_page: 100 }
+              );
+              msg = commits.at(-1)?.commit?.message || '';
+            }
+            const title = context.payload.pull_request?.title || '';
+            const body  = context.payload.pull_request?.body  || '';
+            const skip = token.test(msg) || token.test(title) || token.test(body);
+            core.setOutput('skip', String(skip));
+
   docker-e2e-tests-1st:
     # Run this job first as a gate for running the remainder of the test matrix
-    if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' && !github.event.pull_request.draft }}
+    if: >
+      github.repository_owner == 'axolotl-ai-cloud' &&
+      (github.event_name != 'pull_request' || !github.event.pull_request.draft) &&
+      needs.gate-skip-e2e.outputs.skip != 'true'
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
-    needs: [pre-commit, pytest, pytest-sdist]
+    needs: [pre-commit, pytest, pytest-sdist, gate-skip-e2e]
 
     strategy:
       fail-fast: false
@@ -209,7 +240,7 @@ jobs:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.7.1
             num_gpus: 1
             axolotl_extras:
             dockerfile: "Dockerfile-uv.jinja"
@@ -240,13 +271,16 @@ jobs:
           modal run cicd.e2e_tests
 
   docker-e2e-tests:
-    if: ${{ github.repository_owner == 'axolotl-ai-cloud' && !github.event.pull_request.draft }}
+    if: >
+      github.repository_owner == 'axolotl-ai-cloud' &&
+      (github.event_name != 'pull_request' || !github.event.pull_request.draft) &&
+      needs.gate-skip-e2e.outputs.skip != 'true'
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     # Only run the remainder of the matrix if the first e2e check passed;
     # this is to save on wasted compute costs for known failures that get caught in the first run
-    needs: [pre-commit, pytest, docker-e2e-tests-1st]
+    needs: [pre-commit, pytest, gate-skip-e2e, docker-e2e-tests-1st]
 
     strategy:
       fail-fast: false
@@ -264,6 +298,13 @@ jobs:
             pytorch: 2.7.1
             num_gpus: 1
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            num_gpus: 1
+            gpu_type: "B200"
+            axolotl_extras: fbgemm-gpu
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -284,6 +325,7 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "GPU_TYPE=${{ matrix.gpu_type || 'L40S'}}" >> $GITHUB_ENV
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
@@ -300,10 +342,10 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.7.1
             num_gpus: 1
             axolotl_extras:
     steps:

.gitignore

@@ -190,3 +190,6 @@ out/
 
 # vim
 *.swp
+
+# scm auto-versioning
+src/axolotl/_version.py

.isort.cfg

@@ -1,4 +0,0 @@
-[settings]
-profile=black
-known_third_party=wandb,comet_ml
-known_local_folder=src,tests

.pre-commit-config.yaml

@@ -10,22 +10,12 @@ repos:
     -   id: trailing-whitespace
     -   id: no-commit-to-branch
         args: ['--branch', 'main']
--   repo: https://github.com/psf/black
-    rev: 25.1.0
+-   repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.12.12
     hooks:
-    -   id: black
--   repo: https://github.com/pycqa/isort
-    rev: 6.0.1
-    hooks:
-      - id: isort
--   repo: https://github.com/PyCQA/flake8
-    rev: 7.3.0
-    hooks:
-    - id: flake8
--   repo: https://github.com/pylint-dev/pylint
-    rev: v3.3.8
-    hooks:
-    - id: pylint
+    -   id: ruff
+        args: [--fix]
+    -   id: ruff-format
 -   repo: https://github.com/pre-commit/mirrors-mypy
     rev: v1.17.1
     hooks:

.pylintrc

@@ -1,15 +0,0 @@
-[MASTER]
-init-hook="from pylint.config import find_default_config_files; import sys; sys.path.append(next(find_default_config_files()).parent.as_posix())"
-
-[TYPECHECK]
-
-# List of members which are set dynamically and missed by Pylint inference
-# system, and so shouldn't trigger E1101 when accessed.
-generated-members=numpy.*, torch.*
-
-
-[pylint.messages_control]
-disable=missing-function-docstring, line-too-long, import-error,
-    too-many-arguments, too-many-locals, too-many-statements, too-many-branches, too-few-public-methods,
-    too-many-instance-attributes, fixme, import-outside-toplevel, logging-fstring-interpolation,
-    too-many-positional-arguments, possibly-used-before-assignment

CITATION.cff

@@ -1,6 +1,6 @@
 cff-version: 1.2.0
 type: software
-title: "Axolotl: Post-Training for AI Models"
+title: "Axolotl: Open Source LLM Post-Training"
 message: "If you use this software, please cite it as below."
 authors:
   - name: "Axolotl maintainers and contributors"

README.md

@@ -5,6 +5,9 @@
         <img alt="Axolotl" src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/887513285d98132142bf5db2a74eb5e0928787f1/image/axolotl_logo_digital_black.svg" width="400" height="104" style="max-width: 100%;">
     </picture>
 </p>
+  <p align="center">
+      <strong>A Free and Open Source LLM Fine-tuning Framework</strong><br>
+  </p>
 
 <p align="center">
     <img src="https://img.shields.io/github/license/axolotl-ai-cloud/axolotl.svg?color=blue" alt="GitHub License">
@@ -17,6 +20,7 @@
     <br/>
     <a href="https://discord.com/invite/HhrNrHJPRb"><img src="https://img.shields.io/badge/discord-7289da.svg?style=flat-square&logo=discord" alt="discord" style="height: 20px;"></a>
     <a href="https://twitter.com/axolotl_ai"><img src="https://img.shields.io/twitter/follow/axolotl_ai?style=social" alt="twitter" style="height: 20px;"></a>
+    <a href="https://colab.research.google.com/github/axolotl-ai-cloud/axolotl/blob/main/examples/colab-notebooks/colab-axolotl-example.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="google-colab" style="height: 20px;"></a>
     <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">
@@ -49,20 +53,21 @@
 
 ## ✨ Overview
 
-Axolotl is a tool designed to streamline post-training for various AI models.
+Axolotl is a free and open-source tool designed to streamline post-training and fine-tuning for the latest large language models (LLMs).
 
 Features:
 
-- **Multiple Model Support**: Train various models like LLaMA, Mistral, Mixtral, Pythia, and more. We are compatible with HuggingFace transformers causal language models.
-- **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO), Multimodal, and Reward Modelling (RM) / Process Reward Modelling (PRM).
-- **Easy Configuration**: Re-use a single YAML file between dataset preprocess, training, evaluation, quantization, and inference.
+- **Multiple Model Support**: Train various models like GPT-OSS, LLaMA, Mistral, Mixtral, Pythia, and many more models available on the Hugging Face Hub.
+- **Multimodal Training**: Fine-tune vision-language models (VLMs) including LLaMA-Vision, Qwen2-VL, Pixtral, LLaVA, SmolVLM2, and audio models like Voxtral with image, video, and audio support.
+- **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO), and Reward Modelling (RM) / Process Reward Modelling (PRM).
+- **Easy Configuration**: Re-use a single YAML configuration file across the full fine-tuning pipeline: dataset preprocessing, training, evaluation, quantization, and inference.
 - **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
 - **Flexible Dataset Handling**: Load from local, HuggingFace, and cloud (S3, Azure, GCP, OCI) datasets.
 - **Cloud Ready**: We ship [Docker images](https://hub.docker.com/u/axolotlai) and also [PyPI packages](https://pypi.org/project/axolotl/) for use on cloud platforms and local hardware.
 
 
 
-## 🚀 Quick Start
+## 🚀 Quick Start - LLM Fine-tuning in Minutes
 
 **Requirements**:
 
@@ -70,6 +75,10 @@ Features:
 - Python 3.11
 - PyTorch ≥2.6.0
 
+### Google Colab
+
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/axolotl-ai-cloud/axolotl/blob/main/examples/colab-notebooks/colab-axolotl-example.ipynb#scrollTo=msOCO4NRmRLa)
+
 ### Installation
 
 #### Using pip
@@ -155,7 +164,7 @@ If you use Axolotl in your research or projects, please cite it as follows:
 
 ```bibtex
 @software{axolotl,
-  title = {Axolotl: Post-Training for AI Models},
+  title = {Axolotl: Open Source LLM Post-Training},
   author = {{Axolotl maintainers and contributors}},
   url = {https://github.com/axolotl-ai-cloud/axolotl},
   license = {Apache-2.0},

_quarto.yml

@@ -153,7 +153,7 @@ quartodoc:
         - utils.distributed
         - utils.dict
         - utils.optimizers.adopt
-        - utils.data.pretraining
+        - utils.data.streaming
         - utils.data.sft
         - utils.quantization
     - title: Schemas
@@ -267,11 +267,13 @@ website:
             - docs/dataset_loading.qmd
             - docs/qat.qmd
             - docs/quantize.qmd
+            - docs/optimizations.qmd
 
         - section: "Core Concepts"
           contents:
             - docs/batch_vs_grad.qmd
             - docs/dataset_preprocessing.qmd
+            - docs/streaming.qmd
             - docs/multipack.qmd
             - docs/mixed_precision.qmd
             - docs/optimizers.qmd

cicd/multigpu.py

@@ -2,8 +2,6 @@
 modal application to run axolotl gpu tests in Modal
 """
 
-# pylint: disable=duplicate-code
-
 import os
 import pathlib
 import tempfile
@@ -63,7 +61,7 @@ def run_cmd(cmd: str, run_folder: str):
 
     # Propagate errors from subprocess.
     if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
-        exit(exit_code)  # pylint: disable=consider-using-sys-exit
+        exit(exit_code)
 
 
 @app.function(

cicd/single_gpu.py

@@ -1,7 +1,5 @@
 """Modal app to run axolotl GPU tests"""
 
-# pylint: disable=duplicate-code
-
 import os
 import pathlib
 import tempfile
@@ -59,7 +57,8 @@ VOLUME_CONFIG = {
 }
 
 N_GPUS = int(os.environ.get("N_GPUS", 1))
-GPU_CONFIG = f"L40S:{N_GPUS}"
+GPU_TYPE = os.environ.get("GPU_TYPE", "L40S")
+GPU_CONFIG = f"{GPU_TYPE}:{N_GPUS}"
 
 
 def run_cmd(cmd: str, run_folder: str):
@@ -70,4 +69,4 @@ def run_cmd(cmd: str, run_folder: str):
 
     # Propagate errors from subprocess.
     if exit_code := subprocess.call(cmd.split(), cwd=run_folder, env=sp_env):  # nosec
-        exit(exit_code)  # pylint: disable=consider-using-sys-exit
+        exit(exit_code)

codecov.yml

@@ -12,7 +12,7 @@ coverage:
       default:
         # basic
         target: auto
-        threshold: 0%
+        threshold: 1%
         base: auto
         # advanced
         branches: null
@@ -27,7 +27,7 @@ coverage:
       default:
         # basic
         target: auto
-        threshold: 0%
+        threshold: 1%
         base: auto
         # advanced
         branches: null

docker/Dockerfile-base

@@ -37,7 +37,7 @@ WORKDIR /workspace
 
 RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
     python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
-    python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
+    CAUSAL_CONV1D_FORCE_CXX11_ABI=TRUE CAUSAL_CONV1D_FORCE_BUILD=TRUE python3 -m pip install --no-cache-dir causal_conv1d==1.5.2 && \
     python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" && \
     python3 -m pip cache purge
 

docs/dataset-formats/conversation.qmd

@@ -212,6 +212,14 @@ Instead of passing `tools` via the system prompt, an alternative method would be
 Tools need to follow [JSON schema](https://json-schema.org/learn/getting-started-step-by-step).
 :::
 
+::: {.callout-warning}
+If you have tool arguments with same name but different dtypes (like `"time": string` and `"time": number`), please save `arguments: ` as JSON string to prevent `datasets` from having casting issues.
+
+```
+"arguments": "{\"...\": \"...\"}"
+```
+:::
+
 Example config for Llama4:
 ```yaml
 chat_template: llama4

docs/dataset-formats/index.qmd

@@ -61,7 +61,7 @@ While we recommend `.jsonl`, you can also use the other formats (`csv`, `parquet
 
 ### Pre-training without streaming
 
-On the rare case that the dataset is small and can be loaded entirely into memory, another approach to running pre-training is to use the `completion` format. This would mean that the entire dataset is pre-tokenized instead of on-demand in streaming.
+In the case that the dataset is small and can be loaded entirely into memory, another approach to running pre-training is to use the `completion` format. This would mean that the entire dataset is pre-tokenized instead of on-demand in streaming.
 
 One benefit of this is that the tokenization can be performed separately on a CPU-only machine, and then transferred to a GPU machine for training to save costs.
 

docs/faq.qmd

@@ -140,3 +140,7 @@ description: Frequently asked questions
 **Q: `ValueError("Backward pass should have cleared tracker of all tensors")`
 
 > A: This may happen due to edge cases in using the modern OffloadActivations context manager for CUDA streams. If you encounter this error, you may have success using the naive implementation with `offload_activations: legacy` in your YAML.
+
+**Q: `Error parsing tool_calls arguments as JSON.`
+
+> A: There is an error parsing string arguments to a dict. Please check your dataset and the error message for more details.

docs/installation.qmd

@@ -134,7 +134,7 @@ For providers supporting Docker:
 
 ### Google Colab {#sec-colab}
 
-Use our [example notebook](../examples/colab-notebooks/colab-axolotl-example.ipynb).
+[![](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/axolotl-ai-cloud/axolotl/blob/main/examples/colab-notebooks/colab-axolotl-example.ipynb#scrollTo=msOCO4NRmRLa)
 
 ## Platform-Specific Instructions {#sec-platform-specific}
 

docs/lora_optims.qmd

@@ -5,10 +5,11 @@ description: "Custom autograd functions and Triton kernels in Axolotl for optimi
 
 Inspired by [Unsloth](https://github.com/unslothai/unsloth), we've implemented two
 optimizations for LoRA and QLoRA fine-tuning, supporting both single GPU and multi-GPU
-(in the DDP and DeepSpeed settings) training. These include (1) SwiGLU and GEGLU activation function
-Triton kernels, and (2) LoRA MLP and attention custom autograd functions. Our goal was
-to leverage operator fusion and tensor re-use in order to improve speed and reduce
-memory usage during the forward and backward passes of these calculations.
+(including DDP, DeepSpeed, and FSDP2) training. These include (1) SwiGLU and GEGLU
+activation function Triton kernels, and (2) LoRA MLP and attention custom autograd
+functions. Our goal was to leverage operator fusion and tensor re-use in order to
+improve speed and reduce memory usage during the forward and backward passes of these
+calculations.
 
 We currently support several common model architectures, including (but not limited to):
 
@@ -92,13 +93,12 @@ Currently, LoRA kernels are not supported for RLHF training, only SFT.
 
 - One or more NVIDIA or AMD GPUs (in order to use the Triton kernels)
     - Note: Set `TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1` to enable [memory-efficient attention on AMD GPUs](https://github.com/ROCm/aotriton/issues/16#issuecomment-2346675491)
-- Targeted LoRA adapters cannot use Dropout
-    - This may limit model expressivity / cause overfitting
-- Targeted LoRA adapters cannot have bias terms
+- Targeted LoRA adapters must disable dropout (`lora_dropout: 0`)
     - This may limit model expressivity
+- Adapters that already include bias terms are supported.
 
-Models with pre-existing LoRA adapters that use Dropout or have bias terms may need to
-be re-finetuned without these features in order to be useful.
+Models with pre-existing LoRA adapters that use Dropout may need to be re-finetuned
+without it in order to be as performant.
 
 ## Implementation details
 
@@ -131,6 +131,5 @@ computation path.
 ## Future Work
 
 - Support for additional model architectures
-- Support for the FSDP setting
-- Support for dropout and bias
+- Support for dropout
 - Additional operator fusions

docs/multi-gpu.qmd

@@ -63,15 +63,6 @@ Start from Stage 1 -> Stage 2 -> Stage 3.
 
 :::
 
-::: {.callout-tip}
-
-Using ZeRO Stage 3 with Single-GPU training
-
-ZeRO Stage 3 can be used for training on a single GPU by manually setting the environment variables:
-`WORLD_SIZE=1 LOCAL_RANK=0 MASTER_ADDR=0.0.0.0 MASTER_PORT=29500`
-
-:::
-
 ## Fully Sharded Data Parallel (FSDP) {#sec-fsdp}
 
 ::: {.callout-note}

docs/multimodal.qmd

@@ -13,10 +13,14 @@ format:
 - [Pixtral](#sec-pixtral)
 - [Llava-1.5](#sec-llava-15)
 - [Mistral-Small-3.1](#sec-mistral-small-31)
+- [Magistral-Small-2509](#sec-magistral-small-2509)
+- [Voxtral](#sec-voxtral)
 - [Gemma-3](#sec-gemma-3)
 - [Gemma-3n](#sec-gemma-3n)
 - [Qwen2-VL](#sec-qwen2-vl)
 - [Qwen2.5-VL](#sec-qwen25-vl)
+- [SmolVLM2](#sec-smolvlm2)
+- [LFM2-VL](#sec-lfm2-vl)
 
 ## Usage
 
@@ -31,14 +35,13 @@ skip_prepare_dataset: true
 remove_unused_columns: false  # leave columns in place as they are needed to handle image embeddings during training
 sample_packing: false  # not yet supported with multimodal
 
-chat_template:  # see in next section
+chat_template:  # see in next section if specified
 
 # example dataset
 datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-    field_messages: messages
 
 # (optional) if doing lora, only finetune the Language model,
 # leave the vision model and vision tower frozen
@@ -91,10 +94,32 @@ chat_template: llava
 
 ### Mistral-Small-3.1 {#sec-mistral-small-31}
 
+::: {.callout-tip}
+Please make sure to install vision lib via `pip install 'mistral-common[opencv]==1.8.5'`
+:::
+
 ```yaml
 base_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
+```
 
-chat_template: mistral_v7_tekken
+### Magistral-Small-2509 {#sec-magistral-small-2509}
+
+::: {.callout-tip}
+Please make sure to install vision lib via `pip install 'mistral-common[opencv]==1.8.5'`
+:::
+
+```yaml
+base_model: mistralai/Magistral-Small-2509
+```
+
+### Voxtral {#sec-voxtral}
+
+::: {.callout-tip}
+Please make sure to install audio lib via `pip3 install librosa==0.11.0 'mistral_common[audio]==1.8.3'`
+:::
+
+```yaml
+base_model: mistralai/Voxtral-Mini-3B-2507
 ```
 
 ### Gemma-3 {#sec-gemma-3}
@@ -143,6 +168,26 @@ base_model: Qwen/Qwen2.5-VL-7B-Instruct
 chat_template: qwen2_vl  # same as qwen2-vl
 ```
 
+### SmolVLM2 {#sec-smolvlm2}
+
+::: {.callout-tip}
+Please make sure to install `num2words` via `pip3 install num2words==0.5.14`
+:::
+
+```yaml
+base_model: HuggingFaceTB/SmolVLM2-500M-Video-Instruct
+```
+
+### LFM2-VL {#sec-lfm2-vl}
+
+::: {.callout-warning}
+Please uninstall `causal-conv1d` via `pip3 uninstall -y causal-conv1d`
+:::
+
+```yaml
+base_model: LiquidAI/LFM2-VL-450M
+```
+
 ## Dataset Format
 
 For multi-modal datasets, we adopt an extended `chat_template` format similar to OpenAI's Message format.
@@ -181,6 +226,20 @@ You may need to install `librosa` via `pip3 install librosa==0.11.0`.
 
 :::
 
+### Video
+
+::: {.callout-warning}
+
+This is not well tested at the moment. We welcome contributors!
+
+:::
+
+For video loading, you can use the following keys within `content` alongside `"type": "video"`:
+
+- `"path": "/path/to/video.mp4"`
+- `"url": "https://example.com/video.mp4"`
+- `"video": np.ndarray | list[PIL.Image.Image] | torch.Tensor` (or list of the aforementioned)
+
 ### Example
 
 Here is an example of a multi-modal dataset:

docs/optimizations.qmd

@@ -0,0 +1,133 @@
+---
+title: Optimizations Guide
+description: A guide to the performance and memory optimizations available in Axolotl.
+---
+
+Axolotl includes numerous optimizations to speed up training, reduce memory usage, and handle large models.
+
+This guide provides a high-level overview and directs you to the detailed documentation for each feature.
+
+## Speed Optimizations
+
+These optimizations focus on increasing training throughput and reducing total training time.
+
+### Sample Packing
+
+Improves GPU utilization by combining multiple short sequences into a single packed sequence for training. This requires enabling one of the [attention](#attention-implementations) implementations below.
+
+- **Config:** `sample_packing: true`
+- **Learn more:** [Sample Packing](multipack.qmd)
+
+### Attention Implementations
+
+Using an optimized attention implementation is critical for training speed.
+
+- **[Flash Attention 2](https://github.com/Dao-AILab/flash-attention)**: `flash_attention: true`. **(Recommended)** The industry standard for fast attention on modern GPUs. Requires Ampere or higher. For AMD, check [AMD Support](https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#amd-rocm-support).
+- **[Flex Attention](https://pytorch.org/blog/flexattention/)**: `flex_attention: true`.
+- **[SDP Attention](https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)**: `sdp_attention: true`. PyTorch's native implementation.
+- **[Xformers](https://github.com/facebookresearch/xformers)**: `xformers_attention: true`. Works with FP16.
+
+*Note: You should only enable one attention backend.*
+
+### LoRA Optimizations
+
+Leverages optimized kernels to accelerate LoRA training and reduce memory usage.
+
+- **Learn more:** [LoRA Optimizations Documentation](lora_optims.qmd)
+
+## Memory Optimizations
+
+These techniques help you fit larger models or use bigger batch sizes on your existing hardware.
+
+### Parameter Efficient Finetuning (LoRA & QLoRA)
+
+Drastically reduces memory by training a small set of "adapter" parameters instead of the full model. This is the most common and effective memory-saving technique.
+
+- Examples: Find configs with `lora` or `qlora` in the [examples directory](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-3).
+- Config Reference: See `adapter`, `load_in_4bit`, and `load_in_8bit` in the [Configuration Reference](config-reference.qmd).
+
+### Gradient Checkpointing & Activation Offloading
+
+These techniques save VRAM by changing how activations are handled.
+
+- Gradient Checkpointing: re-computes activations during the backward pass, trading compute time for VRAM.
+- Activation Offloading: moves activations to CPU RAM or disk, trading I/O overhead for VRAM.
+- Learn more: [Gradient Checkpointing and Offloading Docs](gradient_checkpointing.qmd)
+
+### Cut Cross Entropy (CCE)
+
+Reduces VRAM usage by using an optimized cross-entropy loss calculation.
+
+- **Learn more:** [Custom Integrations - CCE](custom_integrations.qmd#cut-cross-entropy)
+
+### Liger Kernels
+
+Provides efficient Triton kernels to improve training speed and reduce memory usage.
+
+- **Learn more:** [Custom Integrations - Liger Kernels](custom_integrations.qmd#liger-kernels)
+
+## Long Context Models
+
+Techniques to train models on sequences longer than their original context window.
+
+### RoPE Scaling
+
+Extends a model's context window by interpolating its Rotary Position Embeddings.
+
+- **Config:** Pass the `rope_scaling` config under the `overrides_of_model_config: `. To learn how to set RoPE, check the respective model config.
+
+### Sequence Parallelism
+
+Splits long sequences across multiple GPUs, enabling training with sequence lengths that would not fit on a single device.
+
+- **Learn more:** [Sequence Parallelism Documentation](sequence_parallelism.qmd)
+
+### Artic Long Sequence Training (ALST)
+
+ALST is a recipe that combines several techniques to train long-context models efficiently. It typically involves:
+
+- TiledMLP to reduce memory usage in MLP layers.
+- Tiled Loss functions (like [CCE](#cut-cross-entropy-(cce) or [Liger](#liger-kernels)).
+- Activation Offloading to CPU.
+
+- Example: [ALST Example Configuration](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/alst)
+
+## Large Models (Distributed Training)
+
+To train models that don't fit on a single GPU, you'll need to use a distributed training strategy like FSDP or DeepSpeed. These frameworks shard the model weights, gradients, and optimizer states across multiple GPUs and nodes.
+
+- **Learn more:** [Multi-GPU Guide](multi-gpu.qmd)
+- **Learn more:** [Multi-Node Guide](multi-node.qmd)
+
+### N-D Parallelism (Beta)
+
+For advanced scaling, Axolotl allows you to compose different parallelism techniques (e.g., Data, Tensor, Sequence Parallelism). This is a powerful approach to train an extremely large model by overcoming multiple bottlenecks at once.
+
+- **Learn more:** [N-D Parallelism Guide](nd_parallelism.qmd)
+
+
+## Quantization
+
+Techniques to reduce the precision of model weights for memory savings.
+
+### 4-bit Training (QLoRA)
+
+The recommended approach for quantization-based training. It loads the base model in 4-bit using `bitsandbytes` and then trains QLoRA adapters. See [Adapter Finetuning](#adapter-finetuning-lora-qlora) for details.
+
+### FP8 Training
+
+Enables training with 8-bit floating point precision on supported hardware (e.g., NVIDIA Hopper series GPUs) for significant speed and memory gains.
+
+- **Example:** [Llama 3 FP8 FSDP Example](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/3b-fp8-fsdp2.yaml)
+
+### Quantization Aware Training (QAT)
+
+Simulates quantization effects during training, helping the model adapt and potentially improving the final accuracy of the quantized model.
+
+- **Learn more:** [QAT Documentation](qat.qmd)
+
+### GPTQ
+
+Allows you to finetune LoRA adapters on top of a model that has already been quantized using the GPTQ method.
+
+- **Example:** [GPTQ LoRA Example](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-2/gptq-lora.yml)

docs/qat.qmd

@@ -23,10 +23,18 @@ To enable QAT in axolotl, add the following to your configuration file:
 
 ```yaml
 qat:
-  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8"
-  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are "int4" and "int8"
+  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4", "int8", "float8"
+  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are "int4", "fp8", and "nvfp4".
   group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization
   fake_quant_after_n_steps: # Optional[int] = None. The number of steps to apply fake quantization after
 ```
 
+We support the following quantization schemas:
+
+- `Int4WeightOnly` (requires the `fbgemm-gpu` extra when installing Axolotl)
+- `Int8DynamicActivationInt4Weight`
+- `Float8DynamicActivationFloat8Weight`
+- `Float8DynamicActivationInt4Weight`
+- `NVFP4`
+
 Once you have finished training, you must quantize your model by using the same quantization configuration which you used to train the model with. You can use the [`quantize`](./quantize.qmd) command to do this.

docs/quantize.qmd

@@ -22,8 +22,8 @@ Quantization is configured using the `quantization` key in your configuration fi
 ```yaml
 base_model: # The path to the model to quantize.
 quantization:
-  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are uintX for X in [1, 2, 3, 4, 5, 6, 7], or int4, or int8
-  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8"
+  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4", "int8", "float8"
+  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are "int4", "fp8", and "nvfp4".
   group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization
   quantize_embedding: # Optional[bool] = False. Whether to quantize the embedding layer.
 
@@ -39,9 +39,8 @@ you used to train the model:
 # qat.yml
 qat:
   activation_dtype: int8
-  weight_dtype: int8
+  weight_dtype: int4
   group_size: 256
-  quantize_embedding: true
 
 output_dir: # The path to the output directory used during training where the final checkpoint has been saved.
 ```
@@ -51,3 +50,11 @@ axolotl quantize qat.yml
 ```
 
 This ensures that an identical quantization configuration is used to quantize the model as was used to train it.
+
+
+::: {.callout-note}
+
+If you have configured pushing to hub with `hub_model_id`, your model hub name will have the quantization schema appended to it,
+e.g. `axolotl-ai-cloud/qat-nvfp4-llama3B` will become `axolotl-ai-cloud/qat-nvfp4-llama3B-nvfp4w`
+
+:::

docs/reward_modelling.qmd

@@ -11,6 +11,7 @@ We support the reward modelling techniques supported by `trl`.
 ### (Outcome) Reward Models
 
 Outcome reward models are trained using data which contains preference annotations for an entire interaction between the user and model (e.g. rather than per-turn or per-step).
+For improved training stability, you can use the `center_rewards_coefficient` parameter to encourage mean-zero reward outputs ([see TRL docs](https://huggingface.co/docs/trl/v0.10.1/en/reward_trainer#centering-rewards)).
 
 ```yaml
 base_model: google/gemma-2-2b

docs/scripts/generate_config_docs.py

@@ -47,7 +47,6 @@ class QuartoGenerator:
         """Check if a type is a Pydantic BaseModel."""
         return inspect.isclass(type_obj) and issubclass(type_obj, BaseModel)
 
-    # pylint: disable=too-many-return-statements
     def _extract_nested_type(self, field_type) -> Any:
         """Extract the actual type from complex type annotations."""
         # Handle Annotated types (Python 3.9+)
@@ -124,7 +123,6 @@ class QuartoGenerator:
 
         return field_type
 
-    # pylint: disable=too-many-return-statements
     def _extract_all_pydantic_models_from_type(
         self, field_type
     ) -> list[type[BaseModel]]:
@@ -318,7 +316,6 @@ class QuartoGenerator:
 
         return all_groups
 
-    # pylint: disable=too-many-return-statements
     def _extract_field_groups_from_source(
         self, model_class: type[BaseModel]
     ) -> list[dict]:
@@ -503,7 +500,7 @@ class QuartoGenerator:
                     nested_schema = nested_model.model_json_schema()
                     nested_properties = nested_schema.get("properties", {})
                     nested_required = nested_schema.get("required", [])
-                except Exception:  # pylint: disable=broad-exception-caught
+                except Exception:
                     # Fallback: use model fields directly
                     nested_properties = {}
                     nested_required = []
@@ -607,7 +604,7 @@ class QuartoGenerator:
             schema = model_class.model_json_schema()
             properties = schema.get("properties", {})
             required = schema.get("required", [])
-        except Exception as e:  # pylint: disable=broad-exception-caught
+        except Exception as e:
             print(
                 f"Warning: Could not generate JSON schema ({e}). Using model fields instead."
             )

docs/streaming.qmd

@@ -0,0 +1,120 @@
+---
+title: Streaming Datasets
+description: How to use streaming mode for large-scale datasets and memory-efficient training
+order: 10
+---
+
+Streaming enables memory-efficient training with large datasets by loading data
+incrementally rather than loading the entire dataset into memory at once.
+
+Use streaming when:
+
+- Your dataset is too large to fit in memory (e.g. when you're doing pretraining with massive text corpora)
+- You want to start training immediately without preprocessing the entire dataset
+
+Streaming works with both remote and locally stored datasets!
+
+::: {.callout-note}
+Streaming currently only supports a single dataset. Multi-dataset support will be added soon.
+:::
+
+
+## Configuration
+
+### Basic Streaming
+
+Enable streaming mode by setting the `streaming` flag:
+
+```yaml
+streaming: true
+```
+
+### Pretraining with Streaming
+
+For pretraining tasks, streaming is automatically enabled when using `pretraining_dataset`:
+
+```yaml
+pretraining_dataset:
+  - path: HuggingFaceFW/fineweb-edu
+    type: pretrain
+    text_column: text
+    split: train
+
+# Optionally, enable sample packing
+streaming_multipack_buffer_size: 10000
+sample_packing: true
+```
+
+### SFT with Streaming
+
+For supervised fine-tuning with streaming:
+
+```yaml
+streaming: true
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+    split: train
+
+# Optionally, enable sample packing
+streaming_multipack_buffer_size: 10000
+sample_packing: true
+```
+
+## Configuration Options
+
+### `streaming_multipack_buffer_size`
+
+Controls the buffer size for multipack streaming (default: 10,000). This determines how
+many samples are buffered before packing. Larger buffers can improve packing efficiency
+but use more memory.
+
+### `shuffle_merged_datasets`
+
+When enabled, shuffles the streaming dataset using the buffer. This requires additional
+memory for the shuffle buffer.
+
+## Sample Packing with Streaming
+
+Sample packing is supported for streaming datasets. When enabled, multiple samples are
+packed into a single sequence to maximize GPU utilization:
+
+```yaml
+sample_packing: true
+streaming_multipack_buffer_size: 10000
+
+# For SFT: attention is automatically isolated between packed samples
+# For pretraining: control with pretrain_multipack_attn
+pretrain_multipack_attn: true  # prevent cross-attention between packed samples
+```
+
+For more information, see our [documentation](multipack.qmd) on multipacking.
+
+## Important Considerations
+
+### Memory Usage
+
+While streaming reduces memory usage compared to loading entire datasets, you still need
+to consider:
+
+- You can control the memory usage by adjusting `streaming_multipack_buffer_size`
+- Sample packing requires buffering multiple samples
+- Shuffling requires additional memory for the shuffle buffer
+
+### Performance
+
+- Streaming may have slightly higher latency compared to preprocessed datasets, as samples are processed on-the-fly
+- Network speed and disk read speed are important when streaming from remote sources or a local dataset, respectively
+- Consider using `axolotl preprocess` for smaller or more frequently used datasets
+
+### Evaluation Datasets
+
+Evaluation datasets are not streamed to ensure consistent evaluation metrics. They're
+loaded normally even when training uses streaming.
+
+## Examples
+
+See the `examples/streaming/` directory for complete configuration examples:
+
+- `pretrain.yaml`: Pretraining with streaming dataset
+- `sft.yaml`: Supervised fine-tuning with streaming

examples/LiquidAI/README.md

@@ -0,0 +1,58 @@
+# Finetune Liquid Foundation Models 2 (LFM2) with Axolotl
+
+[Liquid Foundation Models 2 (LFM2)](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38) are a family of small, open-weight models from [Liquid AI](https://www.liquid.ai/) focused on quality, speed, and memory efficiency. Liquid AI released text-only [LFM2](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38) and text+vision [LFM2-VL](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa) models.
+
+LFM2 features a new hybrid Liquid architecture with multiplicative gates, short-range convolutions, and grouped query attention, enabling fast training and inference.
+
+This guide shows how to fine-tune both the LFM2 and LFM2-VL models with Axolotl.
+
+## Getting Started
+
+1.  Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+    Here is an example of how to install from pip:
+    ```bash
+    # Ensure you have a compatible version of Pytorch installed
+    pip3 install packaging setuptools wheel ninja
+    pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+    ```
+
+2.  Run one of the finetuning examples below.
+
+    **LFM2**
+    ```bash
+    # FFT SFT (1x48GB @ 25GiB)
+    axolotl train examples/LiquidAI/lfm2-350m-fft.yaml
+    ```
+
+    **LFM2-VL**
+    ```bash
+    # LoRA SFT (1x48GB @ 2.7GiB)
+    axolotl train examples/LiquidAI/lfm2-vl-lora.yaml
+    ```
+
+### TIPS
+
+- **Installation Error**: If you encounter `ImportError: ... undefined symbol ...` or `ModuleNotFoundError: No module named 'causal_conv1d_cuda'`, the `causal-conv1d` package may have been installed incorrectly. Try uninstalling it:
+  ```bash
+  pip uninstall -y causal-conv1d
+  ```
+
+- **Dataset Loading**: Read more on how to load your own dataset in our [documentation](https://docs.axolotl.ai/docs/dataset_loading.html).
+- **Dataset Formats**:
+  - For LFM2 models, the dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+  - For LFM2-VL models, Axolotl follows the multi-content Messages format. See our [Multimodal docs](https://docs.axolotl.ai/docs/multimodal.html#dataset-format) for details.
+
+## Optimization Guides
+
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
+
+## Related Resources
+
+- [LFM2 Blog](https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models)
+- [LFM2-VL Blog](https://www.liquid.ai/blog/lfm2-vl-efficient-vision-language-models)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/LiquidAI/lfm2-350m-fft.yaml

@@ -2,7 +2,6 @@ base_model: LiquidAI/LFM2-350M
 
 chunked_cross_entropy: true
 
-chat_template: tokenizer_default
 eot_tokens:
   - "<|im_end|>"
 datasets:

examples/LiquidAI/lfm2-vl-lora.yaml

@@ -0,0 +1,58 @@
+base_model: LiquidAI/LFM2-VL-450M
+trust_remote_code: true
+model_type: AutoModelForImageTextToText
+processor_type: AutoProcessor
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 8192
+pad_to_sequence_len: false
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: true
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/alst/README.md

@@ -7,3 +7,24 @@ techniques. It is a combination of:
 - Activation Offloading: Offload activations to CPU RAM to reduce memory usage
 
 For more information, you can check out the ALST paper [here](https://www.arxiv.org/abs/2506.13996).
+
+## Usage
+
+```yaml
+tiled_mlp: true
+
+# See Sequence Parallelism docs
+# https://docs.axolotl.ai/docs/sequence_parallelism.html
+context_parallel_size: int
+
+plugins:
+# See Cut Cross Entropy docs
+# https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+# or Liger Kernel docs
+# https://docs.axolotl.ai/docs/custom_integrations.html#liger-kernels
+  - axolotl.integrations.liger.LigerPlugin
+# ...
+
+```

examples/apertus/README.md

@@ -0,0 +1,110 @@
+# Finetune Swiss-AI's Apertus with Axolotl
+
+[Apertus](https://huggingface.co/collections/swiss-ai/apertus-llm-68b699e65415c231ace3b059) is a family of opensource models trained by Swiss-ai.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Apertus is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+
+    Here is an example of how to install from main for pip:
+
+```bash
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+git clone https://github.com/axolotl-ai-cloud/axolotl.git
+cd axolotl
+
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
+
+# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+python scripts/cutcrossentropy_install.py | sh
+```
+
+2. (Optional, highly recommended) Install XIELU CUDA
+
+```bash
+## Recommended for reduced VRAM and faster speeds
+
+# Point to CUDA toolkit directory
+# For those using our Docker image, use the below path.
+export CUDA_HOME=/usr/local/cuda
+
+pip3 install git+https://github.com/nickjbrowning/XIELU@59d6031 --no-build-isolation --no-deps
+```
+
+For any installation errors, see [XIELU Installation Issues](#xielu-installation-issues)
+
+3. Run the finetuning example:
+
+```bash
+axolotl train examples/apertus/apertus-8b-qlora.yaml
+```
+
+This config uses about 8.7 GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### Tips
+
+- For inference, the official Apertus team recommends `top_p=0.9` and `temperature=0.8`.
+- You can instead use full paremter fine-tuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+### XIELU Installation Issues
+
+#### `ModuleNotFoundError: No module named 'torch'`
+
+Please check these one by one:
+- Running in correct environment
+- Env has PyTorch installed
+- CUDA toolkit is at `CUDA_HOME`
+
+If those didn't help, please try the below solutions:
+
+1. Pass env for CMAKE and try install again:
+
+    ```bash
+    Python_EXECUTABLE=$(which python) pip3 install git+https://github.com/nickjbrowning/XIELU@59d6031 --no-build-isolation --no-deps
+    ```
+
+2. Git clone the repo and manually hardcode python path:
+
+    ```bash
+    git clone https://github.com/nickjbrowning/XIELU
+    cd xielu
+    git checkout 59d6031
+
+    cd xielu
+    nano CMakeLists.txt  # or vi depending on your preference
+    ```
+
+    ```diff
+    execute_process(
+    -    COMMAND ${Python_EXECUTABLE} -c "import torch.utils; print(torch.utils.cmake_prefix_path)"
+    +    COMMAND /root/miniconda3/envs/py3.11/bin/python -c "import torch.utils; print(torch.utils.cmake_prefix_path)"
+        RESULT_VARIABLE TORCH_CMAKE_PATH_RESULT
+        OUTPUT_VARIABLE TORCH_CMAKE_PATH_OUTPUT
+        ERROR_VARIABLE TORCH_CMAKE_PATH_ERROR
+    )
+    ```
+
+    ```bash
+    pip3 install . --no-build-isolation --no-deps
+    ```
+
+## Optimization Guides
+
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
+- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
+
+## Related Resources
+
+- [Apertus Tech Report](https://github.com/swiss-ai/apertus-tech-report/blob/main/Apertus_Tech_Report.pdf)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/apertus/apertus-8b-qlora.yaml

@@ -0,0 +1,64 @@
+base_model: swiss-ai/Apertus-8B-Instruct-2509
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/arcee/README.md

@@ -19,6 +19,9 @@ cd axolotl
 
 pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn]'
+
+# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+python scripts/cutcrossentropy_install.py | sh
 ```
 
 2. Run the finetuning example:

examples/archived/deepcoder/deepcoder-14B-preview-lora.yml

@@ -9,10 +9,6 @@ strict: false
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
-    field_messages: messages
-    message_property_mappings:
-      role: role
-      content: content
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/cloud/baseten.yaml

@@ -0,0 +1,10 @@
+provider: baseten
+project_name:
+
+secrets:
+  - HF_TOKEN
+  - WANDB_API_KEY
+
+gpu: h100
+gpu_count: 8
+node_count: 1

examples/deepcogito/cogito-v1-preview-llama-3B-lora.yml

@@ -9,10 +9,6 @@ strict: false
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
-    field_messages: messages
-    message_property_mappings:
-      role: role
-      content: content
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/deepcogito/cogito-v1-preview-qwen-14B-lora.yml

@@ -9,10 +9,6 @@ strict: false
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
-    field_messages: messages
-    message_property_mappings:
-      role: role
-      content: content
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/devstral/README.md

@@ -20,7 +20,13 @@ pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
-2. Run the finetuning example:
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage
+
+```bash
+python scripts/cutcrossentropy_install.py | sh
+```
+
+3. Run the finetuning example:
 
 ```bash
 axolotl train examples/devstral/devstral-small-qlora.yml

examples/gemma3/gemma-3-270m-qlora.yml

@@ -0,0 +1,68 @@
+base_model: google/gemma-3-270m-it
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+load_in_8bit: false
+load_in_4bit: true
+
+# huggingface repo
+chat_template: gemma3
+eot_tokens:
+  - <end_of_turn>
+datasets:
+  - path: cgato/SlimOrcaDedupCleaned
+    type: chat_template
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: false
+
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch:
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/gemma3/gemma-3-4b-vision-qlora.yml

@@ -18,7 +18,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-    field_messages: messages
+
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.01
 output_dir: ./outputs/out

examples/gemma3n/README.md

@@ -23,7 +23,15 @@ pip3 install timm==1.0.17
 pip3 install librosa==0.11.0
 ```
 
-3. Run the finetuning example:
+3. Download sample dataset files
+
+```bash
+# for text + vision + audio only
+wget https://huggingface.co/datasets/Nanobit/text-vision-audio-2k-test/resolve/main/African_elephant.jpg
+wget https://huggingface.co/datasets/Nanobit/text-vision-audio-2k-test/resolve/main/En-us-African_elephant.oga
+```
+
+4. Run the finetuning example:
 
 ```bash
 # text only

examples/gpt-oss/README.md

@@ -33,13 +33,64 @@ Note: Memory usage taken from `device_mem_reserved(gib)` from logs.
 
 ### Training 120B
 
-On 8xH100s
+On 8xH100s, make sure you have ~3TB of free disk space. With each checkpoint clocking in at ~720GB, along with the base
+model, and final model output, you may need at least 3TB of free disk space to keep at least 2 checkpoints.
 
 ```bash
 # FFT SFT with offloading (8x80GB @ ~49GiB/GPU)
 axolotl train examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml
 ```
 
+To simplify fine-tuning across 2 nodes × 8x H100 (80GB) GPUs, we've partnered with [Baseten](https://baseten.co) to showcase multi-node
+training of the 120B model using Baseten Truss. You can read more about this recipe on
+[Baseten's blog](https://www.baseten.co/blog/how-to-fine-tune-gpt-oss-120b-with-baseten-and-axolotl/). The recipe can
+be found on their
+[GitHub](https://github.com/basetenlabs/ml-cookbook/tree/main/examples/oss-gpt-120b-axolotl/training).
+
+ERRATA: Transformers saves the model Architecture prefixed with `FSDP` which needs to be manually renamed in `config.json`.
+See https://github.com/huggingface/transformers/pull/40207 for the status of this issue.
+
+```bash
+sed -i 's/FSDPGptOssForCausalLM/GptOssForCausalLM/g' ./outputs/gpt-oss-out/config.json
+```
+
+When using SHARDED_STATE_DICT with FSDP, the final checkpoint should automatically merge the sharded weights to your
+configured `output_dir`. However, if that step fails due to a disk space error, you can take an additional step to
+merge the sharded weights.  This step will automatically determine the last checkpoint directory and merge the sharded
+weights to `{output_dir}/merged`.
+
+```bash
+axolotl merge-sharded-fsdp-weights examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml
+mv ./outputs/gpt-oss-out/merged/* ./outputs/gpt-oss-out/
+```
+
+
+### Inferencing your fine-tuned model
+
+#### vLLM
+
+GPT-OSS support in vLLM does not exist in a stable release yet. See https://x.com/MaziyarPanahi/status/1955741905515323425
+for more information about using a special vllm-openai docker image for inferencing with vLLM.
+
+Optionally, vLLM can be installed from nightly:
+
+```bash
+pip install --no-build-isolation --pre -U vllm --extra-index-url https://wheels.vllm.ai/nightly
+```
+and the vLLM server can be started with the following command (modify `--tensor-parallel-size 8` to match your environment):
+```bash
+vllm serve ./outputs/gpt-oss-out/ --served-model-name axolotl/gpt-oss-20b --host 0.0.0.0 --port 8888  --tensor-parallel-size 8
+```
+
+#### SGLang
+
+SGLang has 0-day support in main, see https://github.com/sgl-project/sglang/issues/8833 for infomation on installing
+SGLang from source. Once you've installed SGLang, run the following command to launch a SGLang server:
+
+```bash
+python3 -m sglang.launch_server --model ./outputs/gpt-oss-out/ --served-model-name axolotl/gpt-oss-120b --host 0.0.0.0 --port 8888 --tp 8
+```
+
 ### Tool use
 
 GPT-OSS has a comprehensive tool understanding. Axolotl supports tool calling datasets for Supervised Fine-tuning.
@@ -55,6 +106,16 @@ See [Nanobit/text-tools-2k-test](https://huggingface.co/datasets/Nanobit/text-to
 
 Refer to [our docs](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#using-tool-use) for more info.
 
+### Thinking and chat_template masking conflict
+
+OpenAI’s Harmony template hides `thinking` in all non-final turns, which conflicts with Axolotl’s `chat_template` masking.
+
+If your dataset has `thinking` content mid-turn, there are two paths we recommend:
+
+- Train only on the last turn. This can be accomplished via chat_template's [train on last doc](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#training-on-last-message).
+
+- Adjust your dataset to only have `thinking` content in the last turn.
+
 ### TIPS
 
 - Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).

examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml

@@ -20,6 +20,7 @@ datasets:
 dataset_prepared_path: last_run_prepared
 val_set_size: 0
 output_dir: ./outputs/gpt-oss-out/
+save_total_limit: 2  # the 120B model can use up to 720GB of disk space per checkpoint, so let's only keep the last 2
 
 sequence_len: 4096
 sample_packing: true
@@ -43,7 +44,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-fft-deepspeed-zero3.yaml

@@ -40,7 +40,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-fft-fsdp2-offload.yaml

@@ -15,7 +15,7 @@ datasets:
     field_thinking: thinking
     template_thinking_key: thinking
 
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path: ./outputs/last_run_prepared
 val_set_size: 0
 output_dir: ./outputs/gpt-oss-out/
 
@@ -41,7 +41,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-fft-fsdp2.yaml

@@ -15,7 +15,7 @@ datasets:
     field_thinking: thinking
     template_thinking_key: thinking
 
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path: ./outputs/last_run_prepared
 val_set_size: 0
 output_dir: ./outputs/gpt-oss-out/
 
@@ -40,7 +40,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-sft-lora-singlegpu.yaml

@@ -53,7 +53,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/hunyuan/README.md

@@ -0,0 +1,85 @@
+# Finetune HunYuan with Axolotl
+
+Tencent released a family of opensource models called HunYuan with varying parameter scales of 0.5B, 1.8B, 4B, and 7B scale for both Pre-trained and Instruct variants. The models can be found at [HuggingFace](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7). This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as HunYuan is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+
+    Here is an example of how to install from main for pip:
+
+```bash
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+git clone https://github.com/axolotl-ai-cloud/axolotl.git
+cd axolotl
+
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
+
+# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+python scripts/cutcrossentropy_install.py | sh
+```
+
+2. Run the finetuning example:
+
+```bash
+axolotl train examples/hunyuan/hunyuan-v1-dense-qlora.yaml
+```
+
+This config uses about 4.7 GB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### Dataset
+
+HunYuan Instruct models can choose to enter a slow think or fast think pattern. For best performance on fine-tuning their Instruct models, your dataset should be adjusted to match their pattern.
+
+```python
+# fast think pattern
+messages = [
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": "/no_think What color is the sun?" },
+    {"role": "assistant", "content": "<think>\n\n</think>\n<answer>\nThe sun is yellow.\n</answer>"}
+]
+
+# slow think pattern
+messages = [
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": "/no_think What color is the sun?" },
+    {"role": "assistant", "content": "<think>\nThe user is asking about the color of the sun. I need to ...\n</think>\n<answer>\nThe sun is yellow.\n</answer>"}
+]
+```
+
+### TIPS
+
+- For inference, the official Tencent team recommends
+
+```json
+
+{
+  "do_sample": true,
+  "top_k": 20,
+  "top_p": 0.8,
+  "repetition_penalty": 1.05,
+  "temperature": 0.7
+}
+
+```
+
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
+- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
+
+## Related Resources
+
+- [Tencent HunYuan Blog](https://hunyuan.tencent.com/)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/hunyuan/hunyuan-v1-dense-qlora.yaml

@@ -0,0 +1,64 @@
+base_model: tencent/Hunyuan-0.5B-Instruct
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/lfm2/README.md

@@ -1,7 +0,0 @@
-# Liquid Foundation Models 2
-
-LFM2 support in transformers exists in the main branch, but is not yet included in the transformers release.
-
-```bash
-pip install --upgrade --no-deps --force-reinstall git+https://github.com/huggingface/transformers.git
-```

examples/llama-3/3b-qat-fsdp2.yaml

@@ -15,20 +15,18 @@ liger_glu_activation: true
 liger_layer_norm: true
 liger_fused_linear_cross_entropy: true
 
+
 datasets:
   - path: yahma/alpaca-cleaned
     type: alpaca
+    split: train[:95%]
 
 output_dir: ./outputs/qat_out/
+dataset_prepared_path: ./outputs/qat_out/dataset_prepared
 
-sample_packing: true
-
-sequence_len: 512
-
-flex_attention: true
-flex_attn_compile_kwargs:
-  dynamic: false
-  mode: max-autotune-no-cudagraphs
+sample_packing: false
+sequence_len: 8192
+flash_attention: true
 
 qat:
   activation_dtype: int8
@@ -67,7 +65,7 @@ fsdp:
 fsdp_config:
   fsdp_version: 2
   fsdp_offload_params: false
-  fsdp_cpu_ram_efficient_loading: true
+  fsdp_cpu_ram_efficient_loading: false
   fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
   fsdp_state_dict_type: FULL_STATE_DICT
@@ -76,6 +74,6 @@ fsdp_config:
   fsdp_activation_checkpointing: true
 
 special_tokens:
-  pad_token: <|end_of_text|>
+  pad_token: <|finetune_right_pad_id|>
 
 # save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/llama-3/3b-qat-nvfp4.yaml

@@ -0,0 +1,64 @@
+base_model: meta-llama/Llama-3.2-3B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+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
+
+datasets:
+  - path: yahma/alpaca-cleaned
+    type: alpaca
+    split: train[:95%]
+
+output_dir: ./outputs/qat_out/
+dataset_prepared_path: ./outputs/dataset_prepared
+
+sequence_len: 8192
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_checkpointing: true
+gradient_accumulation_steps: 1
+micro_batch_size: 64
+num_epochs: 1
+optimizer: adamw_torch_fused
+
+cosine_constant_lr_ratio: 0
+cosine_min_lr_ratio: 1.0
+learning_rate: 2e-5
+save_only_model: true
+bf16: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/llama-3/diffusion/pretrain-1b.yaml

@@ -0,0 +1,56 @@
+base_model: meta-llama/Llama-3.2-1B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+pretraining_dataset:
+  - path: wikitext
+    name: wikitext-103-raw-v1
+    type: completion
+    field: text
+
+plugins:
+  - axolotl.integrations.diffusion.DiffusionPlugin
+
+diffusion:
+  noise_schedule: cosine
+  min_mask_ratio: 0.15
+  max_mask_ratio: 0.85
+  num_diffusion_steps: 128
+  eps: 5e-4
+  importance_weighting: true
+  mask_token_id: 128002
+  generate_samples: true
+  generation_interval: 250
+
+output_dir: ./outputs/model-out
+
+sequence_len: 512
+sample_packing: true
+
+gradient_accumulation_steps: 8
+micro_batch_size: 4
+max_steps: 10000
+warmup_ratio: 0.1
+
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 3e-4
+sdp_attention: true
+
+bf16: auto
+tf32: true
+
+logging_steps: 1
+save_strategy: steps
+save_steps: 1000
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/llama-3/diffusion/sft-1b.yaml

@@ -0,0 +1,59 @@
+base_model: meta-llama/Llama-3.2-1B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+val_set_size: 0.05
+
+plugins:
+  - axolotl.integrations.diffusion.DiffusionPlugin
+
+diffusion:
+  noise_schedule: cosine
+  min_mask_ratio: 0.1
+  max_mask_ratio: 0.9
+  num_diffusion_steps: 128
+  eps: 1e-3
+  importance_weighting: true
+  mask_token_id: 128002
+  generate_samples: true
+  generation_interval: 250
+
+output_dir: ./outputs/model-out
+
+sequence_len: 512
+sample_packing: true
+eval_sample_packing: true
+
+gradient_accumulation_steps: 4
+micro_batch_size: 4
+num_epochs: 1
+warmup_steps: 0.1
+
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 1e-5
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+sdp_attention: true
+
+logging_steps: 1
+save_strategy: best
+eval_strategy: epoch
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/llama-3/instruct-lora-8b.yml

@@ -12,15 +12,6 @@ chat_template: llama3
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
-    field_messages: messages
-    message_property_mappings:
-      role: role
-      content: content
-    roles:
-      user:
-        - user
-      assistant:
-        - assistant
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/llama-4/do-no-use-fa2/scout-vision-qlora-fsdp.yaml

@@ -46,7 +46,6 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-    field_messages: messages
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0

examples/llama-4/scout-vision-qlora-fsdp2-flex.yaml

@@ -45,7 +45,6 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-    field_messages: messages
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0

examples/magistral/README.md

@@ -1,10 +1,10 @@
 # Finetune Magistral Small with Axolotl
 
-Magistral Small is a 24B parameter opensource model from MistralAI found on HuggingFace at [2506](https://huggingface.co/mistralai/Magistral-Small-2506) and [2507](https://huggingface.co/mistralai/Magistral-Small-2507) (see [Thinking](#thinking)). This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+Magistral Small is a 24B parameter opensource model from MistralAI found on HuggingFace at [2506](https://huggingface.co/mistralai/Magistral-Small-2506), [2507](https://huggingface.co/mistralai/Magistral-Small-2507) (see [Thinking](#thinking)), and [2509](https://huggingface.co/mistralai/Magistral-Small-2509) (see [Vision](#vision)). This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
 
 MistralAI has also released a proprietary medium-sized version called Magistral Medium.
 
-Thanks to the team at MistralAI for giving us early access to prepare for this release.
+Thanks to the team at MistralAI for giving us early access to prepare for these releases.
 
 ## Getting started
 
@@ -18,7 +18,13 @@ pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
-2. Run the finetuning example:
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage
+
+```bash
+python scripts/cutcrossentropy_install.py | sh
+```
+
+3. Run the finetuning example:
 
 ```bash
 axolotl train examples/magistral/magistral-small-qlora.yaml
@@ -30,29 +36,17 @@ Let us know how it goes. Happy finetuning! 🚀
 
 ### Thinking
 
-MistralAI has released their [2507](https://huggingface.co/mistralai/Magistral-Small-2507) model with thinking capabilities. The model requires the multi-content dataset format with support for an extra `role: thinking` within system and assistant messages.
+MistralAI has released their [2507](https://huggingface.co/mistralai/Magistral-Small-2507) model with thinking capabilities, enabling Chain-of-Thought reasoning with explicit thinking steps.
 
-Example format:
+📚 **[See the Thinking fine-tuning guide →](./think/README.md)**
 
-```json
-{
-    "messages": [
-        {"role": "system", "content": [{ "type": "text", "text": "{SYSTEM_PROMPT}"}]},
-        {"role": "user", "content": [{ "type": "text", "text": "..."}]},
-        {"role": "assistant", "content": [{ "type": "thinking", "thinking": "..."}, { "type": "text", "text": "..." }]},
-    ],
-}
-```
+### Vision
 
-Example config: `./magistral-small-think-qlora.yaml`.
+MistralAI has released their [2509](https://huggingface.co/mistralai/Magistral-Small-2509) model with vision capabilities.
 
-The `thinking` section also supports an optional arg `closed: bool` (`True` default) which controls adding the closing `[/THINK]` tag.
+📚 **[See the Vision fine-tuning guide →](./vision/README.md)**
 
-Limitations:
-- You cannot mix `content: str` with `content: list[dict]` as the `dataset.load_dataset` may complain about different types for `content` key.
-- This mode does not work with custom `train_detail` and `training` at the moment.
-
-### TIPS
+### Tips
 
 - We recommend adding the same/similar SystemPrompt that the model is tuned for. You can find this within the repo's files titled `SYSTEM_PROMPT.txt`.
 - For inference, the official MistralAI team recommends `top_p: 0.95` and `temperature: 0.7` with `max_tokens: 40960`.
@@ -83,5 +77,5 @@ In addition, we do not support overriding tokens yet.
 
 ## Future Work
 
-- Add parity to Preference Tuning, RL, Multi-modal, etc.
+- Add parity to Preference Tuning, RL, etc.
 - Add parity to other tokenizer configs like overriding tokens.

examples/magistral/think/README.md

@@ -0,0 +1,73 @@
+# Magistral Small Thinking Fine-tuning
+
+This guide covers fine-tuning [Magistral Small 2507](https://huggingface.co/mistralai/Magistral-Small-2507) with thinking capabilities using Axolotl. The thinking model enables explicit Chain-of-Thought reasoning with separate thinking and response sections.
+
+## Prerequisites
+
+Before starting, ensure you have:
+- Installed Axolotl (see [main README](../README.md))
+
+## Getting Started
+
+Run the thinking model fine-tuning:
+
+```bash
+axolotl train magistral-small-think-qlora.yaml
+```
+
+This config uses about 19.1 GiB VRAM.
+
+### Tips
+
+- Dataset uses multi-content format with `type: thinking` support. See [Dataset Format](#dataset-format) below.
+- You cannot mix `content: str` and `content: list[dict]`, otherwise, dataset loading will fail. Keep it consistent.
+
+## Dataset Format
+
+The thinking model requires the multi-content dataset format with support for an extra `role: thinking` within system and assistant messages.
+
+Example format:
+
+```json
+{
+    "messages": [
+        {
+            "role": "system",
+            "content": [
+                { "type": "text", "text": "{SYSTEM_PROMPT}"}
+            ]
+        },
+        {
+            "role": "user",
+            "content": [
+                { "type": "text", "text": "Solve this step by step: What is 15% of 240?"}
+            ]
+        },
+        {
+            "role": "assistant",
+            "content": [
+                {
+                    "type": "thinking",
+                    "thinking": "I need to calculate 15% of 240. First, I'll convert 15% to decimal: 0.15. Then multiply: 0.15 × 240 = 36."
+                },
+                {
+                    "type": "text",
+                    "text": "To find 15% of 240, I'll multiply 240 by 0.15:\n\n240 × 0.15 = 36\n\nTherefore, 15% of 240 is 36."
+                }
+            ]
+        }
+    ]
+}
+```
+
+### Advanced Options
+
+The `thinking` section supports an optional `closed` parameter:
+
+```json
+{
+    "type": "thinking",
+    "thinking": "Internal reasoning here...",
+    "closed": true  // Default: true, controls adding the closing [/THINK] tag
+}
+```

examples/magistral/vision/README.md

@@ -0,0 +1,60 @@
+# Magistral Small Vision Fine-tuning
+
+This guide covers fine-tuning [Magistral Small 2509](https://huggingface.co/mistralai/Magistral-Small-2509) with vision capabilities using Axolotl.
+
+## Prerequisites
+
+Before starting, ensure you have:
+- Installed Axolotl from source (see [main README](../README.md#getting-started))
+
+## Getting started
+
+1. Install the required vision lib:
+    ```bash
+    pip install 'mistral-common[opencv]==1.8.5'
+    ```
+
+2. Download the example dataset image:
+   ```bash
+   wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+   ```
+
+3. Run the fine-tuning:
+   ```bash
+   axolotl train magistral-small-vision-24B-qlora.yml
+   ```
+
+This config uses about 17GiB VRAM.
+
+WARNING: The loss and grad norm will be much higher than normal at first. We suspect this to be inherent to the model as of the moment. If anyone would like to submit a fix for this, we are happy to take a look.
+
+### Tips
+
+Key differences from text-only model:
+- `max_tokens: 131072` for inference
+- Multi-modal dataset format required
+- Sample packing not supported
+
+## Dataset Format
+
+The vision model requires multi-modal dataset format as documented [here](https://docs.axolotl.ai/docs/multimodal.html#dataset-format).
+
+One exception is that, passing `"image": PIL.Image` is not supported. MistralTokenizer only supports `path`, `url`, and `base64` for now.
+
+Example:
+```json
+{
+    "messages": [
+        {"role": "system", "content": [{ "type": "text", "text": "{SYSTEM_PROMPT}"}]},
+        {"role": "user", "content": [
+            { "type": "text", "text": "What's in this image?"},
+            {"type": "image", "path": "path/to/image.jpg" }
+        ]},
+        {"role": "assistant", "content": [{ "type": "text", "text": "..." }]},
+    ],
+}
+```
+
+## Limitations
+
+- Sample Packing is not supported for multi-modality training currently.

examples/magistral/vision/magistral-small-vision-24B-qlora.yml

@@ -0,0 +1,64 @@
+base_model: mistralai/Magistral-Small-2509
+processor_type: AutoProcessor
+
+# Enable to use mistral-common tokenizer
+tokenizer_use_mistral_common: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+# sample dataset below requires downloading image in advance
+# wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+datasets:
+  - path: Nanobit/text-vision-2k-test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/mistral/mistral-small/mistral-small-3.1-24B-lora.yml

@@ -1,6 +1,9 @@
 base_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
 processor_type: AutoProcessor
 
+# Enable to use mistral-common tokenizer
+tokenizer_use_mistral_common: true
+
 load_in_8bit: true
 
 # these 3 lines are needed for now to handle vision chat templates w images
@@ -8,12 +11,12 @@ skip_prepare_dataset: true
 remove_unused_columns: false
 sample_packing: false
 
-chat_template: mistral_v7_tekken
+# sample dataset below requires downloading image in advance
+# wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
 datasets:
-  - path: HuggingFaceH4/llava-instruct-mix-vsft
+  - path: Nanobit/text-vision-2k-test
     type: chat_template
-    split: train[:1%]
-    field_messages: messages
+
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.01
 output_dir: ./outputs/out
@@ -48,8 +51,7 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-# flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet.
-sdp_attention: true
+flash_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/phi/lora-3.5.yaml

@@ -12,15 +12,6 @@ chat_template: phi_3
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
-    field_messages: messages
-    message_property_mappings:
-      role: role
-      content: content
-    roles:
-      user:
-        - user
-      assistant:
-        - assistant
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/pixtral/lora-12b.yml

@@ -45,8 +45,7 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-# flash_attention:  # PixtralVisionModel does not support Flash Attention 2.0 yet
-sdp_attention: true
+flash_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/qwen2-vl/lora-7b.yaml

@@ -11,7 +11,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-    field_messages: messages
+
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: ./outputs/out

examples/qwen2_5-vl/lora-7b.yaml

@@ -11,7 +11,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-    field_messages: messages
+
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: ./outputs/out

examples/qwen3-next/README.md

@@ -0,0 +1,64 @@
+# Finetune Qwen3-Next with Axolotl
+
+[Qwen3-Next](https://huggingface.co/collections/Qwen/qwen3-next-68c25fd6838e585db8eeea9d) represents the next-generation foundation models optimized for extreme context length and large-scale parameter efficiency. The series introduces architectural innovations including Hybrid Attention (Gated DeltaNet + Gated Attention), High-Sparsity MoE with 1:50 activation ratio, and Multi-Token Prediction for enhanced performance and inference acceleration.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Qwen3-Next is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+
+    Here is an example of how to install from main for pip:
+
+```bash
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+git clone https://github.com/axolotl-ai-cloud/axolotl.git
+cd axolotl
+
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
+
+# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+python scripts/cutcrossentropy_install.py | sh
+```
+
+2. Install Qwen3-Next transformers commit
+```bash
+pip3 uninstall -y transformers && pip3 install "git+https://github.com/huggingface/transformers.git@b9282355bea846b54ed850a066901496b19da654"
+```
+
+3. Install FLA for improved performance
+```bash
+pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.3.2
+```
+
+4. Run the finetuning example:
+
+```bash
+axolotl train examples/qwen3-next/qwen3-next-80b-a3b-qlora.yaml
+```
+
+This config uses about 45.62 GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### TIPS
+
+- For inference, you can experiment with `temperature: 0.7`, `top_p: 0.8`, `top_k: 20`, and `min_p: 0`.
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config. See [Multi-GPU](#optimization-guides) section below.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
+- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
+
+## Related Resources
+
+- [Qwen3-Next Blog](https://qwenlm.github.io/blog/qwen3_next/)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/qwen3-next/qwen3-next-80b-a3b-qlora.yaml

@@ -0,0 +1,68 @@
+base_model: Qwen/Qwen3-Next-80B-A3B-Instruct
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 16
+lora_alpha: 8
+lora_dropout: 0.05
+lora_target_modules:
+  - linear_attn.in_proj_ba
+  - linear_attn.in_proj_qkvz
+  - linear_attn.out_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  - shared_expert.gate_proj
+  - shared_expert_gate
+  - mlp.gate
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qwen3/reward-model.yaml

@@ -0,0 +1,44 @@
+base_model: Skywork/Skywork-Reward-V2-Qwen3-8B
+model_type: AutoModelForSequenceClassification
+num_labels: 1
+
+reward_model: true
+center_rewards_coefficient: 0.01  # Incentivize mean-zero rewards for improved stability
+chat_template: qwen3
+datasets:
+  - path: argilla/distilabel-intel-orca-dpo-pairs
+    type: bradley_terry.chat_template
+
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 8192
+sample_packing: false
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+deepspeed: deepspeed_configs/zero1.json
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+eval_batch_size: 1
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: linear
+learning_rate: 0.00002
+
+bf16: true
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+warmup_ratio: 0.1
+logging_steps: 1
+weight_decay: 0.01

examples/seed-oss/README.md

@@ -0,0 +1,54 @@
+# Finetune ByteDance's Seed-OSS with Axolotl
+
+[Seed-OSS](https://huggingface.co/collections/ByteDance-Seed/seed-oss-68a609f4201e788db05b5dcd) are a series of 36B parameter open source models trained by ByteDance's Seed Team.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Seed-OSS is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+
+    Here is an example of how to install from main for pip:
+
+```bash
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+git clone https://github.com/axolotl-ai-cloud/axolotl.git
+cd axolotl
+
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
+
+# Install Cut Cross Entropy
+python scripts/cutcrossentropy_install.py | sh
+```
+
+2. Run the finetuning example:
+
+```bash
+axolotl train examples/seed-oss/seed-oss-36b-qlora.yaml
+```
+
+This config uses about 27.7 GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### TIPS
+
+- For inference, the official Seed Team recommends `top_p=0.95` and `temperature=1.1`.
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
+- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
+
+## Related Resources
+
+- [ByteDance Seed Website](https://seed.bytedance.com/)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/seed-oss/seed-oss-36b-qlora.yaml

@@ -0,0 +1,56 @@
+base_model: ByteDance-Seed/Seed-OSS-36B-Instruct
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/smolvlm2/README.md

@@ -0,0 +1,49 @@
+# Finetune SmolVLM2 with Axolotl
+
+[SmolVLM2](https://huggingface.co/collections/HuggingFaceTB/smolvlm2-smallest-video-lm-ever-67ab6b5e84bf8aaa60cb17c7) are a family of lightweight, open-source multimodal models from HuggingFace designed to analyze and understand video, image, and text content.
+
+These models are built for efficiency, making them well-suited for on-device applications where computational resources are limited. Models are available in multiple sizes, including 2.2B, 500M, and 256M.
+
+This guide shows how to fine-tune SmolVLM2 models with Axolotl.
+
+## Getting Started
+
+1.  Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+    Here is an example of how to install from pip:
+    ```bash
+    # Ensure you have a compatible version of Pytorch installed
+    pip3 install packaging setuptools wheel ninja
+    pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+    ```
+
+2. Install an extra dependency:
+
+    ```bash
+    pip3 install num2words==0.5.14
+    ```
+
+3.  Run the finetuning example:
+
+    ```bash
+    # LoRA SFT (1x48GB @ 6.8GiB)
+    axolotl train examples/smolvlm2/smolvlm2-2B-lora.yaml
+    ```
+
+## TIPS
+
+- **Dataset Format**: For video finetuning, your dataset must be compatible with the multi-content Messages format. For more details, see our documentation on [Multimodal Formats](https://docs.axolotl.ai/docs/multimodal.html#dataset-format).
+- **Dataset Loading**: Read more on how to prepare and load your own datasets in our [documentation](https://docs.axolotl.ai/docs/dataset_loading.html).
+
+## Optimization Guides
+
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
+
+## Related Resources
+
+- [SmolVLM2 Blog](https://huggingface.co/blog/smolvlm2)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/smolvlm2/smolvlm2-2B-lora.yaml

@@ -0,0 +1,56 @@
+base_model: HuggingFaceTB/SmolVLM2-2.2B-Instruct
+trust_remote_code: true
+processor_type: AutoProcessor
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 8192
+pad_to_sequence_len: false
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.text_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: true
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/streaming/README.md

@@ -0,0 +1,50 @@
+# Streaming Dataset Examples
+
+This directory contains example configurations for using Axolotl's streaming dataset
+functionality, which enables memory-efficient training with large datasets.
+
+## Examples
+
+Run the following examples with e.g. `axolotl train examples/streaming/sft.yaml`; no
+`axolotl preprocess` required!
+
+### Pretraining (`pretrain.yaml`)
+
+Demonstrates streaming configuration for pretraining tasks using the fineweb-edu dataset
+with SmolLM2-135M.
+
+- Uses `pretraining_dataset` configuration for automatic streaming
+- Multipack attention control to prevent cross-attention between packed sequences
+- Buffer size configuration for memory management
+
+### SFT (`sft.yaml`)
+
+Shows how to use streaming for supervised fine-tuning with the Alpaca dataset.
+
+- Explicit `streaming: true` flag for SFT datasets
+- Memory-efficient training on instruction datasets
+- Evaluation datasets are currently not streamed
+
+## Key Configuration Options
+
+### `streaming`
+- Enables streaming mode for standard datasets
+- Automatically enabled for `pretraining_dataset`
+
+### `streaming_multipack_buffer_size`
+- Controls buffer size for sample packing (default: 10,000)
+- Larger values improve packing efficiency but use more memory
+- Adjust based on available memory
+
+### `shuffle_merged_datasets`
+- Enables shuffling of streaming datasets
+- Requires additional memory for shuffle buffer
+
+### `sample_packing`
+- Packs multiple samples into single sequences
+- Minimize per-step padding tokens
+
+## Performance Tips
+
+- Download small / frequently-used datasets locally for better performance
+- Larger buffer sizes improve packing efficiency

examples/streaming/pretrain.yaml

@@ -0,0 +1,57 @@
+base_model: HuggingFaceTB/SmolLM2-135M
+
+# Streaming pretraining configuration
+pretraining_dataset:
+  - path: HuggingFaceFW/fineweb-edu
+    name: sample-10BT
+    type: pretrain
+    text_column: text
+    split: train
+
+# Streaming-specific settings
+streaming_multipack_buffer_size: 10000
+shuffle_merged_datasets: true
+
+# Training configuration
+max_steps: 1000
+output_dir: ./outputs/smollm2-135m-pretrain-streaming
+
+# Sequence and packing settings
+sequence_len: 1024
+sample_packing: true
+pretrain_multipack_attn: true  # Prevent cross-attention between packed sequences
+flash_attention: true
+
+# Batch size settings
+gradient_accumulation_steps: 8
+micro_batch_size: 1
+
+# Optimizer and scheduler
+optimizer: adamw_torch
+lr_scheduler: cosine
+learning_rate: 5e-4
+warmup_ratio: 0.1
+weight_decay: 0.01
+
+# Precision and performance
+bf16: auto
+tf32: true
+
+# Logging and checkpointing
+logging_steps: 10
+save_strategy: steps
+save_steps: 250
+save_total_limit: 3
+
+# Weights & Biases (optional)
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+# Special tokens
+special_tokens:
+  pad_token: "<|endoftext|>"
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/streaming/sft.yaml

@@ -0,0 +1,55 @@
+base_model: HuggingFaceTB/SmolLM2-135M
+
+# Dataset configuration
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+    split: train
+
+# Streaming-specific settings
+streaming: true
+streaming_multipack_buffer_size: 10000
+shuffle_merged_datasets: true
+
+# Training configuration
+max_steps: 1000
+output_dir: ./outputs/smollm2-135m-sft-streaming
+
+# Sequence and packing settings
+sequence_len: 1024
+sample_packing: true
+flash_attention: true
+
+# Batch size settings
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+
+# Optimizer and scheduler
+optimizer: adamw_torch
+lr_scheduler: cosine
+learning_rate: 2e-4
+warmup_ratio: 0.1
+weight_decay: 0.0
+
+# Precision and performance
+bf16: auto
+tf32: true
+
+# Logging and checkpointing
+logging_steps: 10
+save_strategy: steps
+save_steps: 100
+save_total_limit: 3
+
+# Weights & Biases (optional)
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+# Special tokens
+special_tokens:
+  pad_token: "<|endoftext|>"
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/voxtral/README.md

@@ -22,9 +22,19 @@ pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 # audio
 pip3 install librosa==0.11.0
 pip3 install 'mistral_common[audio]==1.8.3'
+
+# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+python scripts/cutcrossentropy_install.py | sh
 ```
 
-3. Run the finetuning example:
+3. Download sample dataset files
+
+```bash
+# for text + audio only
+wget https://huggingface.co/datasets/Nanobit/text-audio-2k-test/resolve/main/En-us-African_elephant.oga
+```
+
+4. Run the finetuning example:
 
 ```bash
 # text only

pyproject.toml

@@ -26,3 +26,34 @@ include-package-data = true
 
 [tool.setuptools.cmdclass]
 build_py = "setuptools_axolotl_dynamic_dependencies.BuildPyCommand"
+
+[tool.ruff]
+line-length = 88
+target-version = "py310"
+
+[tool.ruff.lint]
+select = ["E", "F", "W", "C90", "B", "I"]
+ignore = [
+    "E203",  # Whitespace before ':'
+    "E501",  # Line too long
+    "C901",  # Too complex
+    "B019",  # Use of functools.cache on methods
+    "E722",  # Bare except
+    "F821",  # Undefined name (for dynamic exec)
+]
+
+[tool.ruff.lint.isort]
+known-third-party = ["wandb", "comet_ml"]
+known-local-folder = ["src", "tests"]
+# Black-compatible isort settings
+force-single-line = false
+combine-as-imports = true
+split-on-trailing-comma = true
+
+[tool.ruff.format]
+# Use black's formatting style exactly
+quote-style = "double"
+indent-style = "space"
+skip-magic-trailing-comma = false
+line-ending = "auto"
+docstring-code-format = false

requirements.txt

@@ -1,9 +1,8 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 
 # START section of dependencies that don't install on Darwin/MacOS
-bitsandbytes==0.46.1
-# triton 3.4.0 is not compatible with CCE
-triton>=3.0.0,<3.4.0
+bitsandbytes==0.47.0
+triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
@@ -13,13 +12,13 @@ liger-kernel==0.6.1
 packaging==23.2
 
 huggingface_hub>=0.33.0
-peft==0.17.0
-transformers==4.55.0
+peft>=0.17.0
+transformers==4.56.1
 tokenizers>=0.21.1
-accelerate==1.10.0
+accelerate==1.10.1
 datasets==4.0.0
 deepspeed>=0.17.0
-trl==0.21.0
+trl==0.23.0
 hf_xet==1.1.5
 kernels==0.9.0
 trackio
@@ -65,10 +64,10 @@ langdetect==1.0.9
 immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
 
-torchao==0.12.0
+torchao==0.13.0
 schedulefree==1.4.1
 
 axolotl-contribs-lgpl==0.0.6
 axolotl-contribs-mit==0.0.5
 
-mistral-common==1.8.3
+mistral-common==1.8.5

scripts/chat_datasets.py

@@ -27,7 +27,7 @@ def parse_dataset(dataset=None, split="train"):
             break
     if not field_messages:
         raise ValueError(
-            f'No conversation field found in dataset: {", ".join(feature_keys)}'
+            f"No conversation field found in dataset: {', '.join(feature_keys)}"
         )
     ds_cfg["field_messages"] = field_messages
 
@@ -40,7 +40,7 @@ def parse_dataset(dataset=None, split="train"):
             break
     if not message_property_mappings["role"]:
         raise ValueError(
-            f'No role field found in messages: {", ".join(message_fields)}'
+            f"No role field found in messages: {', '.join(message_fields)}"
         )
 
     for key in ["content", "text", "value"]:
@@ -49,7 +49,7 @@ def parse_dataset(dataset=None, split="train"):
             break
     if not message_property_mappings["content"]:
         raise ValueError(
-            f'No content field found in messages: {", ".join(message_fields)}'
+            f"No content field found in messages: {', '.join(message_fields)}"
         )
     ds_cfg["message_property_mappings"] = message_property_mappings
 

scripts/cutcrossentropy_install.py

@@ -29,5 +29,5 @@ UV_PREFIX = "uv " if USE_UV else ""
 
 print(
     UNINSTALL_PREFIX
-    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@0ee9ee8"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c5aa3ef"'
 )

scripts/unsloth_install.py

@@ -1,11 +1,10 @@
 # noqa
-# pylint: skip-file
 import sys
 
 try:
     import torch
-except ImportError:
-    raise ImportError("Install torch via `pip install torch`")
+except ImportError as error:
+    raise ImportError("Install torch via `pip install torch`") from error
 from packaging.version import Version as V
 
 use_uv = "--uv" in sys.argv[1:]

setup.py

@@ -64,7 +64,9 @@ def parse_requirements(extras_require_map):
             else:
                 raise ValueError("Invalid version format")
 
-            if (major, minor) >= (2, 7):
+            if (major, minor) >= (2, 8):
+                pass
+            elif (major, minor) >= (2, 7):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
                     _install_requires.append("xformers==0.0.30")
@@ -118,14 +120,13 @@ def get_package_version():
 
 
 extras_require = {
-    "flash-attn": ["flash-attn==2.8.2"],
+    "flash-attn": ["flash-attn==2.8.3"],
     "ring-flash-attn": [
-        "flash-attn==2.8.2",
+        "flash-attn==2.8.3",
         "ring-flash-attn>=0.1.7",
-        "yunchang==0.6.0",
     ],
     "deepspeed": [
-        "deepspeed==0.17.2",
+        "deepspeed==0.17.5",
         "deepspeed-kernels",
     ],
     "mamba-ssm": [
@@ -160,6 +161,7 @@ extras_require = {
     "llmcompressor": [
         "llmcompressor==0.5.1",
     ],
+    "fbgemm-gpu": ["fbgemm-gpu-genai>=1.2.0"],
 }
 install_requires, dependency_links, extras_require_build = parse_requirements(
     extras_require

src/axolotl/cli/__init__.py

@@ -4,5 +4,7 @@ import os
 
 from axolotl.logging_config import configure_logging
 
-os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
+os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
+os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")
+
 configure_logging()

src/axolotl/cli/args.py

@@ -14,9 +14,13 @@ class PreprocessCliArgs:
     prompter: Optional[str] = field(default=None)
     download: Optional[bool] = field(default=True)
     iterable: Optional[bool] = field(
-        default=None,
+        default=False,
         metadata={
-            "help": "Use IterableDataset for streaming processing of large datasets"
+            "help": (
+                "Deprecated in v0.13.0, will be removed in v0.14.0. For streaming "
+                "datasets, use 'axolotl train' and set 'streaming: true' in your YAML "
+                "config, or pass --streaming instead in the CLI."
+            )
         },
     )
 
@@ -40,6 +44,12 @@ class VllmServeCliArgs:
         default=None,
         metadata={"help": "Number of tensor parallel workers to use."},
     )
+    data_parallel_size: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "Number of data parallel workers to use for vLLM serving. This controls how many model replicas are used for parallel inference."
+        },
+    )
     host: Optional[str] = field(
         default=None,  # nosec B104
         metadata={"help": "Host address to run the server on."},
@@ -105,6 +115,7 @@ class QuantizeCliArgs:
     quantize_embedding: Optional[bool] = field(default=None)
     group_size: Optional[int] = field(default=None)
     output_dir: Optional[str] = field(default=None)
+    hub_model_id: Optional[str] = field(default=None)
 
 
 @dataclass