fix compile

* simplify logging

* remove comment

* progress on debug.log

* add debug-level logger for file log

* simplify

* case insensitivity; 3rd party logging improvements

* simplify

* fix

* tests

* lint

* nits

* nit

* Update tests/test_utils_tee.py

Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>

* cleanup / comments

* fix

* oops

---------

Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>

.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

@@ -57,6 +57,13 @@ We welcome ideas for improvements and new features. To suggest an enhancement, o
 5. Push your branch to your fork on GitHub.
 6. Open a new pull request against the `main` branch of the axolotl repository. Include a clear and concise description of your changes, referencing any related issues.
 
+#### Skipping CI Checks
+
+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. You may also include this in the title of your PR to disable end-to-end tests for the entire PR.
+
 ## Style Guidelines
 
 ### Code Style

.github/workflows/base.yml

@@ -54,7 +54,7 @@ jobs:
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-base"
           - cuda: "128"
-            cuda_version: 12.6.3
+            cuda_version: 12.8.1
             cudnn_version: ""
             python_version: "3.11"
             pytorch: 2.7.1
@@ -64,9 +64,16 @@ jobs:
             cuda_version: 12.8.1
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: nightly
+            pytorch: 2.8.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-base-nightly"
+            dockerfile: "Dockerfile-base"
+#          - cuda: "128"
+#            cuda_version: 12.8.1
+#            cudnn_version: ""
+#            python_version: "3.11"
+#            pytorch: nightly
+#            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+#            dockerfile: "Dockerfile-base-nightly"
 #          # "next" is for release candidates of pytorch
 #          - cuda: "128"
 #            cuda_version: 12.8.1
@@ -122,6 +129,13 @@ jobs:
             pytorch: 2.6.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
+          - cuda: "126"
+            cuda_version: 12.6.3
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.7.1
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
@@ -129,6 +143,13 @@ jobs:
             pytorch: 2.7.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.8.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.github/workflows/main.yml

@@ -24,16 +24,22 @@ jobs:
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.0
-            axolotl_extras: vllm
+            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.1
+            axolotl_extras: vllm
+            is_latest: true
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.7.1
             axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.8.0
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -97,12 +103,23 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.1
             axolotl_extras:
+            is_latest:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras: vllm
             is_latest: true
           - cuda: 128
             cuda_version: 12.8.1
             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
@@ -150,6 +167,24 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras:
+            is_latest:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            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:
@@ -105,7 +105,8 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
           pytest -v --durations=10 tests/patched/ --cov=axolotl --cov-append --cov-report=xml
           pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml
 
@@ -129,7 +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:
@@ -179,21 +180,52 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
-          pytest -v --durations=10 tests/patched/
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
           pytest -v --durations=10 tests/cli/
 
       - name: cleanup pip cache
         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
@@ -208,7 +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"
@@ -239,13 +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
@@ -263,6 +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
@@ -283,6 +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
@@ -299,10 +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

@@ -3,31 +3,21 @@ default_language_version:
 
 repos:
 -   repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v5.0.0
+    rev: v6.0.0
     hooks:
     -   id: check-yaml
     -   id: end-of-file-fixer
     -   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: v3.3.7
-    hooks:
-    - id: pylint
+    -   id: ruff
+        args: [--fix]
+    -   id: ruff-format
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.17.0
+    rev: v1.17.1
     hooks:
     - id: mypy
       additional_dependencies:

.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

.runpod/README.md

@@ -185,7 +185,6 @@ datasets:
 | `flash_attention`          | `false` | Use flash attention           |
 | `flash_attn_cross_entropy` | `false` | Flash attention cross entropy |
 | `flash_attn_rms_norm`      | `false` | Flash attention RMS norm      |
-| `flash_attn_fuse_qkv`      | `false` | Fuse QKV operations           |
 | `flash_attn_fuse_mlp`      | `false` | Fuse MLP operations           |
 | `sdp_attention`            | `false` | Use scaled dot product        |
 | `s2_attention`             | `false` | Use shifted sparse attention  |

.runpod/src/config/config.yaml

@@ -296,7 +296,6 @@
 # flash_attention:
 # flash_attn_cross_entropy:  # Whether to use flash-attention cross entropy implementation - advanced use only
 # flash_attn_rms_norm:  # Whether to use flash-attention rms norm implementation - advanced use only
-# flash_attn_fuse_qkv: # Whether to fuse QKV into a single operation
 # flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
 # # Whether to use scaled-dot-product attention
 # # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
@@ -541,7 +540,6 @@ xformers_attention: ${XFORMERS_ATTENTION}
 flash_attention: ${FLASH_ATTENTION}
 flash_attn_cross_entropy: ${FLASH_ATTN_CROSS_ENTROPY}
 flash_attn_rms_norm: ${FLASH_ATTN_RMS_NORM}
-flash_attn_fuse_qkv: ${FLASH_ATTN_FUSE_QKV}
 flash_attn_fuse_mlp: ${FLASH_ATTN_FUSE_MLP}
 sdp_attention: ${SDP_ATTENTION}
 s2_attention: ${S2_ATTENTION}

CITATION.cff

@@ -0,0 +1,10 @@
+cff-version: 1.2.0
+type: software
+title: "Axolotl: Open Source LLM Post-Training"
+message: "If you use this software, please cite it as below."
+authors:
+  - name: "Axolotl maintainers and contributors"
+repository-code: "https://github.com/axolotl-ai-cloud/axolotl"
+url: "https://axolotl.ai/"
+license: Apache-2.0
+date-released: "2023-05-30"

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">
@@ -25,33 +29,45 @@
 
 ## 🎉 Latest Updates
 
-- 2025/07: Voxtral with mistral-common tokenizer support has been integrated in Axolotl. Read the [docs](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/voxtral)!
-- 2025/07: TiledMLP support for single-GPU to multi-GPU training with DDP, DeepSpeed and FSDP support has been added to support Arctic Long Sequence Training. (ALST). See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/alst) for using ALST with Axolotl!
-- 2025/06: Magistral with mistral-common tokenizer support has been added to Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral) to start training your own Magistral models with Axolotl!
+- 2025/07:
+  - ND Parallelism support has been added into Axolotl. Compose Context Parallelism (CP), Tensor Parallelism (TP), and Fully Sharded Data Parallelism (FSDP) within a single node and across multiple nodes. Check out the [blog post](https://huggingface.co/blog/accelerate-nd-parallel) for more info.
+  - Axolotl adds more models: [GPT-OSS](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/gpt-oss), [Gemma 3n](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/gemma3n), [Liquid Foundation Model 2 (LFM2)](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/lfm2), and [Arcee Foundation Models (AFM)](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/afm).
+  - FP8 finetuning with fp8 gather op is now possible in Axolotl via `torchao`. Get started [here](https://docs.axolotl.ai/docs/mixed_precision.html#sec-fp8)!
+  - [Voxtral](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/voxtral), [Magistral 1.1](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral), and [Devstral](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/devstral) with mistral-common tokenizer support has been integrated in Axolotl!
+  - TiledMLP support for single-GPU to multi-GPU training with DDP, DeepSpeed and FSDP support has been added to support Arctic Long Sequence Training. (ALST). See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/alst) for using ALST with Axolotl!
 - 2025/05: Quantization Aware Training (QAT) support has been added to Axolotl. Explore the [docs](https://docs.axolotl.ai/docs/qat.html) to learn more!
-- 2025/04: Llama 4 support has been added in Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-4) to start training your own Llama 4 models with Axolotl's linearized version!
 - 2025/03: Axolotl has implemented Sequence Parallelism (SP) support. Read the [blog](https://huggingface.co/blog/axolotl-ai-co/long-context-with-sequence-parallelism-in-axolotl) and [docs](https://docs.axolotl.ai/docs/sequence_parallelism.html) to learn how to scale your context length when fine-tuning.
+
+<details>
+
+<summary>Expand older updates</summary>
+
+- 2025/06: Magistral with mistral-common tokenizer support has been added to Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral) to start training your own Magistral models with Axolotl!
+- 2025/04: Llama 4 support has been added in Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-4) to start training your own Llama 4 models with Axolotl's linearized version!
 - 2025/03: (Beta) Fine-tuning Multimodal models is now supported in Axolotl. Check out the [docs](https://docs.axolotl.ai/docs/multimodal.html) to fine-tune your own!
 - 2025/02: Axolotl has added LoRA optimizations to reduce memory usage and improve training speed for LoRA and QLoRA in single GPU and multi-GPU training (DDP and DeepSpeed). Jump into the [docs](https://docs.axolotl.ai/docs/lora_optims.html) to give it a try.
 - 2025/02: Axolotl has added GRPO support. Dive into our [blog](https://huggingface.co/blog/axolotl-ai-co/training-llms-w-interpreter-feedback-wasm) and [GRPO example](https://github.com/axolotl-ai-cloud/grpo_code) and have some fun!
 - 2025/01: Axolotl has added Reward Modelling / Process Reward Modelling fine-tuning support. See [docs](https://docs.axolotl.ai/docs/reward_modelling.html).
 
+</details>
+
 ## ✨ 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**:
 
@@ -59,6 +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
@@ -138,6 +158,20 @@ Contributions are welcome! Please see our [Contributing Guide](https://github.co
 
 Interested in sponsoring? Contact us at [wing@axolotl.ai](mailto:wing@axolotl.ai)
 
+## 📝 Citing Axolotl
+
+If you use Axolotl in your research or projects, please cite it as follows:
+
+```bibtex
+@software{axolotl,
+  title = {Axolotl: Open Source LLM Post-Training},
+  author = {{Axolotl maintainers and contributors}},
+  url = {https://github.com/axolotl-ai-cloud/axolotl},
+  license = {Apache-2.0},
+  year = {2023}
+}
+```
+
 ## 📜 License
 
 This project is licensed under the Apache 2.0 License - see the [LICENSE](LICENSE) file for details.

TODO.md

@@ -1,10 +0,0 @@
-# todo list
-
-- [] Validation of parameters for combinations that won't work
-
-
-
-## things that are known not to work
-
-- FSDP offload and gradient_checkpointing - https://github.com/pytorch/pytorch/issues/82203
-- adamw_bnb_8bit doesn't play well with FSDP offload

_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
@@ -272,8 +272,10 @@ website:
           contents:
             - docs/batch_vs_grad.qmd
             - docs/dataset_preprocessing.qmd
+            - docs/streaming.qmd
             - docs/multipack.qmd
             - docs/mixed_precision.qmd
+            - docs/optimizers.qmd
 
         - section: "Advanced Features"
           contents:
@@ -283,6 +285,7 @@ website:
             - docs/custom_integrations.qmd
             - docs/sequence_parallelism.qmd
             - docs/gradient_checkpointing.qmd
+            - docs/moe_backends.md
             - docs/nd_parallelism.qmd
 
         - section: "Troubleshooting"

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

@@ -212,10 +212,11 @@ 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).
 :::
 
+Example config for Llama4:
 ```yaml
 chat_template: llama4
 datasets:
-  - path: ...
+  - path: Nanobit/text-tools-2k-test
     type: chat_template
     # field_tools: tools # default is `tools`
 ```

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/moe_backends.md

@@ -0,0 +1,18 @@
+MoE Backends in Axolotl
+
+Axolotl supports selecting a Mixture-of-Experts (MoE) compute backend via the training config (YAML):
+
+- Set `moe_backend: auto|torch_grouped|naive`
+
+Behavior
+- auto (default): prefers PyTorch 2.8+ grouped GEMM; otherwise naive.
+- torch_grouped: targets PyTorch 2.8+ grouped GEMM (H100/SM90+ recommended).
+- naive: keeps the reference per-expert loop.
+
+Notes
+- Current implementation wires the backend selector and routes Mixtral MoE through it. Torch grouped uses cuBLASLt grouped GEMM when available; otherwise, the code falls back to the naive per-expert loop.
+- No changes to training scripts are required; selection happens inside the model forward.
+
+Example
+moe_backend: torch_grouped
+accelerate launch -m axolotl.cli.train path/to/config.yaml

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 +13,13 @@ format:
 - [Pixtral](#sec-pixtral)
 - [Llava-1.5](#sec-llava-15)
 - [Mistral-Small-3.1](#sec-mistral-small-31)
+- [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
 
@@ -31,7 +34,7 @@ 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:
@@ -97,6 +100,16 @@ base_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
 chat_template: mistral_v7_tekken
 ```
 
+### 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}
 
 ::: {.callout-tip}
@@ -143,6 +156,26 @@ 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 +214,20 @@ 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/nd_parallelism.qmd

@@ -1,4 +1,6 @@
-# N-D Parallelism
+---
+title: "N-D Parallelism (Beta)"
+---
 
 Axolotl enables training models at scale by composing different parallelism techniques. This is essential when:
 
@@ -71,6 +73,10 @@ Note: We recommend FSDP. DeepSpeed is only compatible with `tensor_parallel_size
 
 ## Examples
 
+::: {.callout-tip}
+See our example configs [here](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/distributed-parallel).
+:::
+
 1.  HSDP on 2 nodes with 4 GPUs each (8 GPUs total):
     - You want FSDP within each node and DDP across nodes.
     - Set `dp_shard_size: 4` and `dp_replicate_size: 2`.
@@ -95,7 +101,7 @@ This matrix describes how different parallelism methods can be combined in Axolo
 | **HSDP + TP** | >1 | >1 | >1 | 1 | ✅ **3D Parallelism**. A powerful but complex combination. |
 | **FSDP + CP** | 1 | >1 | 1 | >1 | ✅ **2D Parallelism**. Combines FSDP with context parallelism. |
 | **FSDP + TP + CP**| 1 | >1 | >1| >1| ✅ **3D Parallelism**. Another advanced combination. |
-| DDP + TP/CP | >1 | 1 | >1 | >1 | ❌ **Not Supported**. The `ParallelismConfig` explicitly prevents this, as composing pure DDP with TP/CP without FSDP is inefficient and complex. You should use FSDP instead (`dp_shard_size > 1`). |
+| DDP + TP/CP | >1 | 1 | >1 | >1 | ❌ **Not Supported**. The `ParallelismConfig` explicitly prevents this, as composing pure DDP with TP or CP is currently not supported. You should use FSDP + TP/CP instead (`dp_shard_size > 1`). |
 | Just TP / CP | 1 | 1 | >1 | >1 | ✅ Supported. Useful for inference or when the model fits on one GPU but context is too long. |
 
 - `tp_size` refers to `tensor_parallel_size`

docs/optimizers.qmd

@@ -0,0 +1,129 @@
+---
+title: Optimizers
+description: Configuring optimizers
+---
+
+## Overview
+
+Axolotl supports all optimizers supported by [transformers OptimizerNames](https://github.com/huggingface/transformers/blob/51f94ea06d19a6308c61bbb4dc97c40aabd12bad/src/transformers/training_args.py#L142-L187)
+
+Here is a list of optimizers supported by transformers as of `v4.54.0`:
+
+- `adamw_torch`
+- `adamw_torch_fused`
+- `adamw_torch_xla`
+- `adamw_torch_npu_fused`
+- `adamw_apex_fused`
+- `adafactor`
+- `adamw_anyprecision`
+- `adamw_torch_4bit`
+- `adamw_torch_8bit`
+- `ademamix`
+- `sgd`
+- `adagrad`
+- `adamw_bnb_8bit`
+- `adamw_8bit`  # alias for adamw_bnb_8bit
+- `ademamix_8bit`
+- `lion_8bit`
+- `lion_32bit`
+- `paged_adamw_32bit`
+- `paged_adamw_8bit`
+- `paged_ademamix_32bit`
+- `paged_ademamix_8bit`
+- `paged_lion_32bit`
+- `paged_lion_8bit`
+- `rmsprop`
+- `rmsprop_bnb`
+- `rmsprop_bnb_8bit`
+- `rmsprop_bnb_32bit`
+- `galore_adamw`
+- `galore_adamw_8bit`
+- `galore_adafactor`
+- `galore_adamw_layerwise`
+- `galore_adamw_8bit_layerwise`
+- `galore_adafactor_layerwise`
+- `lomo`
+- `adalomo`
+- `grokadamw`
+- `schedule_free_radam`
+- `schedule_free_adamw`
+- `schedule_free_sgd`
+- `apollo_adamw`
+- `apollo_adamw_layerwise`
+- `stable_adamw`
+
+
+## Custom Optimizers
+
+Enable custom optimizers by passing a string to the `optimizer` argument. Each optimizer will receive beta and epsilon args, however, some may accept additional args which are detailed below.
+
+### optimi_adamw
+
+```yaml
+optimizer: optimi_adamw
+```
+
+### ao_adamw_4bit
+
+Deprecated: Please use `adamw_torch_4bit`.
+
+### ao_adamw_8bit
+
+Deprecated: Please use `adamw_torch_8bit`.
+
+### ao_adamw_fp8
+
+
+```yaml
+optimizer: ao_adamw_fp8
+```
+
+### adopt_adamw
+
+GitHub: [https://github.com/iShohei220/adopt](https://github.com/iShohei220/adopt)
+Paper: [https://arxiv.org/abs/2411.02853](https://arxiv.org/abs/2411.02853)
+
+```yaml
+optimizer: adopt_adamw
+```
+
+### came_pytorch
+
+GitHub: [https://github.com/yangluo7/CAME/tree/master](https://github.com/yangluo7/CAME/tree/master)
+Paper: [https://arxiv.org/abs/2307.02047](https://arxiv.org/abs/2307.02047)
+
+```yaml
+optimizer: came_pytorch
+
+# optional args (defaults below)
+adam_beta1: 0.9
+adam_beta2: 0.999
+adam_beta3: 0.9999
+adam_epsilon: 1e-30
+adam_epsilon2: 1e-16
+```
+
+### muon
+
+Blog: [https://kellerjordan.github.io/posts/muon/](https://kellerjordan.github.io/posts/muon/)
+Paper: [https://arxiv.org/abs/2502.16982v1](https://arxiv.org/abs/2502.16982v1)
+
+```yaml
+optimizer: muon
+```
+
+### dion
+
+Microsoft's Dion (DIstributed OrthoNormalization) optimizer is a scalable and communication-efficient
+orthonormalizing optimizer that uses low-rank approximations to reduce gradient communication.
+
+GitHub: [https://github.com/microsoft/dion](https://github.com/microsoft/dion)
+Paper: [https://arxiv.org/pdf/2504.05295](https://arxiv.org/pdf/2504.05295)
+Note: Implementation written for PyTorch 2.7+ for DTensor
+
+```yaml
+optimizer: dion
+dion_lr: 0.01
+dion_momentum: 0.95
+lr: 0.00001  # learning rate for embeddings and parameters that fallback to AdamW
+```

docs/qat.qmd

@@ -23,10 +23,17 @@ 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/arcee/README.md

@@ -0,0 +1,53 @@
+# Finetune ArceeAI's AFM with Axolotl
+
+[Arcee Foundation Models (AFM)](https://huggingface.co/collections/arcee-ai/afm-45b-68823397c351603014963473) are a family of 4.5B parameter open weight models trained by Arcee.ai.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+Thanks to the team at Arcee.ai for using Axolotl in supervised fine-tuning the AFM model.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as AFM 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]'
+```
+
+2. Run the finetuning example:
+
+```bash
+axolotl train examples/arcee/afm-4.5b-qlora.yaml
+```
+
+This config uses about 7.8GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### TIPS
+
+- For inference, the official Arcee.ai team recommends `top_p: 0.95`, `temperature: 0.5`, `top_k: 50`, and `repeat_penalty: 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
+
+- [AFM Blog](https://docs.arcee.ai/arcee-foundation-models/introduction-to-arcee-foundation-models)
+- [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/arcee/afm-4.5b-qlora.yaml

@@ -0,0 +1,64 @@
+base_model: arcee-ai/AFM-4.5B
+
+# 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/archived/stablelm-2/1.6b/fft.yml

@@ -47,7 +47,6 @@ logging_steps: 1
 flash_attention: true
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
-flash_attn_fuse_qkv: false
 flash_attn_fuse_mlp: true
 
 warmup_ratio: 0.1

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/devstral/README.md

@@ -10,20 +10,23 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 
 ## Getting started
 
-1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Devstral is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
 
-    Here is an example of how to install from main for pip:
+    Here is an example of how to install from pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.6.0+)
-git clone https://github.com/axolotl-ai-cloud/axolotl.git
-cd axolotl
-
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
 pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn]'
+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/distributed-parallel/README.md

@@ -0,0 +1,52 @@
+# ND Parallelism Examples
+
+This directory contains example configurations for training models using ND Parallelism in Axolotl. These examples demonstrate how to compose different parallelism strategies (FSDP, TP, CP, HSDP) for efficient multi-GPU training.
+
+## Quick Start
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+2. Run the command below:
+
+```bash
+# Train Qwen3 8B with FSDP + TP + CP on a single 8-GPU node
+axolotl train examples/distributed-parallel/qwen3-8b-fsdp-tp-cp.yaml
+
+# Train Llama 3.1 8B with HSDP + TP on 2 nodes (16 GPUs total)
+axolotl train examples/distributed-parallel/llama-3_1-8b-hsdp-tp.yaml
+```
+
+## Example Configurations
+
+### Single Node (8 GPUs)
+
+**Qwen3 8B with FSDP + TP + CP** ([qwen3-8b-fsdp-tp-cp.yaml](./qwen3-8b-fsdp-tp-cp.yaml))
+- Uses all 3 parallelism dimensions on a single node
+- Ideal for: when model weights, activations, and/or context are too large to fit on single GPU
+
+```yaml
+dp_shard_size: 2         # FSDP across 2 GPUs
+tensor_parallel_size: 2  # TP across 2 GPUs
+context_parallel_size: 2 # CP across 2 GPUs
+# Total: 2 × 2 × 2 = 8 GPUs
+```
+
+### Multi-Node
+
+**Llama 3.1 8B with HSDP + TP** ([llama-3_1-8b-hsdp-tp.yaml](./llama-3_1-8b-hsdp-tp.yaml))
+- FSDP & TP within nodes, DDP across nodes to minimize inter-node communication
+- Ideal for: Scaling to multiple nodes while maintaining training efficiency
+
+```yaml
+dp_shard_size: 4        # FSDP within each 4-GPU group
+tensor_parallel_size: 2 # TP within each node
+dp_replicate_size: 2    # DDP across 2 groups
+# Total: (4 × 2) × 2 = 16 GPUs (2 nodes)
+```
+
+## Learn More
+
+- [ND Parallelism Documentation](https://docs.axolotl.ai/docs/nd_parallelism.html)
+- [Blog: Accelerate ND-Parallel Guide](https://huggingface.co/blog/accelerate-nd-parallel)
+- [Multi-GPU Training Guide](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/distributed-parallel/llama-3_1-8b-hsdp-tp.yaml

@@ -0,0 +1,47 @@
+base_model: meta-llama/Llama-3.1-8B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+dp_shard_size: 4
+dp_replicate_size: 2
+tensor_parallel_size: 2
+# context_parallel_size: 2
+
+dataset_prepared_path: last_run_prepared
+
+special_tokens:
+  pad_token: <|end_of_text|>
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  reshard_after_forward: true
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/ndp-out/
+
+sequence_len: 2048
+sample_packing: true
+flash_attention: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 2
+optimizer: adamw_torch_fused
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-6
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1

examples/distributed-parallel/qwen3-8b-fsdp-tp-cp.yaml

@@ -0,0 +1,46 @@
+base_model: Qwen/Qwen3-8B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+dp_shard_size: 2
+# dp_replicate_size: 1
+context_parallel_size: 2
+tensor_parallel_size: 2
+
+dataset_prepared_path: last_run_prepared
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3DecoderLayer
+  reshard_after_forward: true
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/ndp-out/
+
+sequence_len: 8192
+sample_packing: true
+flash_attention: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1  # must be 1 when using context parallel
+num_epochs: 2
+optimizer: adamw_torch_fused
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-6
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+
+special_tokens:

examples/gemma3/270m-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/gemma3n/README.md

@@ -4,17 +4,14 @@ Gemma-3n is a family of multimodal models from Google found on [HuggingFace](htt
 
 ## Getting started
 
-1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Gemma3n is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
 
-    Here is an example of how to install from main for pip:
+    Here is an example of how to install from pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.6.0 min recommended)
-git clone https://github.com/axolotl-ai-cloud/axolotl.git
-cd axolotl
-
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
 pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn]'
+pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. In addition to Axolotl's requirements, Gemma-3n requires:

examples/gpt-oss/README.md

@@ -0,0 +1,135 @@
+# Finetune OpenAI's GPT-OSS with Axolotl
+
+[GPT-OSS](https://huggingface.co/collections/openai/gpt-oss-68911959590a1634ba11c7a4) are a family of open-weight MoE models trained by OpenAI, released in August 2025. There are two variants: 20B and 120B.
+
+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).
+
+    Here is an example of how to install from pip:
+
+```bash
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+```
+
+2. Choose one of the following configs below for training the 20B model. (for 120B, see [below](#training-120b))
+
+```bash
+# LoRA SFT linear layers (1x48GB @ ~44GiB)
+axolotl train examples/gpt-oss/gpt-oss-20b-sft-lora-singlegpu.yaml
+
+# FFT SFT with offloading (2x24GB @ ~21GiB/GPU)
+axolotl train examples/gpt-oss/gpt-oss-20b-fft-fsdp2-offload.yaml
+
+# FFT SFT (8x48GB @ ~36GiB/GPU or 4x80GB @ ~46GiB/GPU)
+axolotl train examples/gpt-oss/gpt-oss-20b-fft-fsdp2.yaml
+```
+
+Note: Memory usage taken from `device_mem_reserved(gib)` from logs.
+
+### Training 120B
+
+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.
+
+Here is an example dataset config:
+```yaml
+datasets:
+  - path: Nanobit/text-tools-2k-test
+    type: chat_template
+```
+
+See [Nanobit/text-tools-2k-test](https://huggingface.co/datasets/Nanobit/text-tools-2k-test) for the sample dataset.
+
+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).
+- 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)
+
+## Related Resources
+
+- [GPT-OSS Blog](https://openai.com/index/introducing-gpt-oss/)
+- [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/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml

@@ -0,0 +1,68 @@
+# the original mxfp4 quantized model is not supported with FSDP cpu_ram_efficient_loading
+# FSDP cpu_ram_efficient_loading is used to reduce the initial CPU memory usage when loading the model
+base_model: axolotl-ai-co/gpt-oss-120b-dequantized
+
+use_kernels: false
+
+dp_shard_size: 16  # requires 2x8xH100 nodes
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding
+
+datasets:
+  - path: HuggingFaceH4/Multilingual-Thinking
+    type: chat_template
+    field_thinking: thinking
+    template_thinking_key: thinking
+
+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
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+
+optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+flash_attention: true
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
+
+gradient_checkpointing: true
+activation_offloading: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.03
+
+special_tokens:
+eot_tokens:
+  - "<|end|>"
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: true
+  state_dict_type: SHARDED_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: GptOssDecoderLayer
+  reshard_after_forward: true
+  cpu_ram_efficient_loading: true

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

@@ -0,0 +1,58 @@
+base_model: openai/gpt-oss-20b
+use_kernels: false
+model_quantization_config: Mxfp4Config
+model_quantization_config_kwargs:
+  dequantize: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding
+
+datasets:
+  - path: HuggingFaceH4/Multilingual-Thinking
+    type: chat_template
+    field_thinking: thinking
+    template_thinking_key: thinking
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/gpt-oss-out/
+
+sequence_len: 4096
+sample_packing: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+
+optimizer: adamw_torch_8bit
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+flash_attention: true
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
+
+gradient_checkpointing: true
+activation_offloading: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.03
+
+special_tokens:
+eot_tokens:
+  - "<|end|>"
+
+# choose the zero3 configuration that best fits your system capabilities
+deepspeed: deepspeed_configs/zero3_bf16.json

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

@@ -0,0 +1,68 @@
+base_model: openai/gpt-oss-20b
+use_kernels: true
+model_quantization_config: Mxfp4Config
+model_quantization_config_kwargs:
+  dequantize: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding
+
+datasets:
+  - path: HuggingFaceH4/Multilingual-Thinking
+    type: chat_template
+    field_thinking: thinking
+    template_thinking_key: thinking
+
+dataset_prepared_path: ./outputs/last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/gpt-oss-out/
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+
+optimizer: adamw_torch_fused  # 8bit optimizers do not work with FSDP2 offload
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+flash_attention: true
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
+
+gradient_checkpointing: true
+activation_offloading: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.03
+
+special_tokens:
+eot_tokens:
+  - "<|end|>"
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: true
+  state_dict_type: SHARDED_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: GptOssDecoderLayer
+  reshard_after_forward: true
+  #  cpu_ram_efficient_loading: true
+
+# cpu_ram_efficient_loading cannot be used with MXFP4 model quantization.
+# It can only be used with a dequantized model like `axolotl-ai-co/gpt-oss-120b-dequantized`

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

@@ -0,0 +1,64 @@
+base_model: openai/gpt-oss-20b
+use_kernels: false
+model_quantization_config: Mxfp4Config
+model_quantization_config_kwargs:
+  dequantize: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+experimental_skip_move_to_device: true  # prevent OOM by NOT putting model to GPU before sharding
+
+datasets:
+  - path: HuggingFaceH4/Multilingual-Thinking
+    type: chat_template
+    field_thinking: thinking
+    template_thinking_key: thinking
+
+dataset_prepared_path: ./outputs/last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/gpt-oss-out/
+
+sequence_len: 4096
+sample_packing: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+
+optimizer: adamw_torch_8bit
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+flash_attention: true
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
+
+gradient_checkpointing: true
+activation_offloading: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.03
+
+special_tokens:
+eot_tokens:
+  - "<|end|>"
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  state_dict_type: SHARDED_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: GptOssDecoderLayer
+  reshard_after_forward: true
+#  cpu_ram_efficient_loading: true

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

@@ -0,0 +1,67 @@
+base_model: openai/gpt-oss-20b
+use_kernels: true
+model_quantization_config: Mxfp4Config
+model_quantization_config_kwargs:
+  dequantize: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+experimental_skip_move_to_device: true  # prevent OOM by not putting model to GPU before sharding
+
+datasets:
+  - path: HuggingFaceH4/Multilingual-Thinking
+    type: chat_template
+    field_thinking: thinking
+    template_thinking_key: thinking
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/gpt-oss-out/
+
+sequence_len: 4096
+sample_packing: true
+
+adapter: lora
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.0  # dropout not supported when using LoRA over expert parameters
+lora_target_linear: true
+
+# TODO: not supported for now, see peft#2710
+#lora_target_parameters:  # target the experts in the last two layers
+#  - "22._checkpoint_wrapped_module.mlp.experts.gate_up_proj"
+#  - "22._checkpoint_wrapped_module.mlp.experts.down_proj"
+#  - "23._checkpoint_wrapped_module.mlp.experts.gate_up_proj"
+#  - "23._checkpoint_wrapped_module.mlp.experts.down_proj"
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 8
+micro_batch_size: 1
+num_epochs: 1
+
+optimizer: adamw_torch_8bit
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-4
+
+bf16: true
+tf32: true
+
+flash_attention: true
+attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
+
+gradient_checkpointing: true
+activation_offloading: true
+
+logging_steps: 1
+saves_per_epoch: 1
+warmup_ratio: 0.1
+
+special_tokens:
+eot_tokens:
+  - "<|end|>"

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-2/fft_optimized.yml

@@ -45,7 +45,6 @@ logging_steps: 1
 flash_attention: true
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
-flash_attn_fuse_qkv: false
 flash_attn_fuse_mlp: true
 
 warmup_ratio: 0.1

examples/llama-2/lisa.yml

@@ -49,7 +49,6 @@ logging_steps: 1
 flash_attention: true
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
-flash_attn_fuse_qkv: false
 flash_attn_fuse_mlp: true
 
 warmup_ratio: 0.1

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/magistral/README.md

@@ -8,20 +8,23 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 
 ## Getting started
 
-1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Magistral is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
 
-    Here is an example of how to install from main for pip:
+    Here is an example of how to install from 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]'
+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

examples/magistral/magistral-small-fsdp-qlora.yaml

@@ -27,7 +27,6 @@ sequence_len: 2048
 sample_packing: true
 eval_sample_packing: false
 
-
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05

examples/magistral/magistral-small-qlora.yaml

@@ -26,7 +26,6 @@ lora_model_dir:
 sequence_len: 2048
 sample_packing: true
 
-
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05

examples/magistral/magistral-small-think-qlora.yaml

@@ -26,7 +26,6 @@ lora_model_dir:
 sequence_len: 2048
 sample_packing: true
 
-
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05

examples/moe/qwen2-moe-qlora-10gb.yaml

@@ -0,0 +1,53 @@
+base_model: Qwen/Qwen1.5-MoE-A2.7B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+trust_remote_code: true
+
+# Keep VRAM low
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./outputs/qwen2-moe-qlora-10gb
+
+# Train small to fit 10GB
+sequence_len: 512
+sample_packing: false
+pad_to_sequence_len: false
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+gradient_accumulation_steps: 8
+micro_batch_size: 1
+num_epochs: 1
+optimizer: paged_adamw_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: 5
+flash_attention: true
+
+warmup_ratio: 0.03
+evals_per_epoch: 2
+saves_per_epoch: 1
+weight_decay: 0.0
+
+model_config:
+  output_router_logits: true
+
+special_tokens:

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/slurm/README.md

@@ -0,0 +1,66 @@
+# SLURM Multi-Node Training
+
+This directory contains an example SLURM script for running Axolotl training jobs across multiple nodes in a SLURM cluster.
+
+## Prerequisites
+
+- Access to a SLURM cluster with GPU nodes
+- Axolotl installed on all nodes (see [installation docs](https://docs.axolotl.ai/docs/installation.html))
+
+## Usage
+
+### Standard SLURM Clusters
+
+1. Copy [`axolotl.slurm`](./axolotl.slurm) to your working directory.
+2. Place your Axolotl config file (`train.yaml`) in the same directory.
+3. Set the appropriate environment variables for the job:
+    ```bash
+    export HF_TOKEN="your-huggingface-token"
+
+    # metric tracking
+    # export WANDB_API_KEY="your-wandb-api-key"
+    # ...
+    ```
+4. Submit the job:
+   ```bash
+   sbatch --export=ALL,NUM_NODES=2,NUM_TRAINERS=8,PRIMARY_ADDR=<master-node>,PRIMARY_PORT=29400 axolotl.slurm
+   ```
+
+   Where:
+   - `NUM_NODES`: Number of nodes to use
+   - `NUM_TRAINERS`: GPUs per node (typically 8)
+   - `PRIMARY_ADDR`: Hostname/IP of the master node
+   - `PRIMARY_PORT`: Port for distributed training (default: 29400)
+
+5. (Optional) Run other slurm commands:
+    ```bash
+    # check job info
+    scontrol show job axolotl-cli
+
+    # check job queue
+    squeue
+
+    # check cluster status
+    sinfo
+    ```
+
+### RunPod Instant Clusters
+
+Axolotl works with RunPod Instant Clusters. This feature provides managed SLURM clusters with zero configuration.
+
+1. **Deploy a SLURM Cluster**:
+   - Go to [RunPod Instant Clusters](https://console.runpod.io/cluster)
+   - Click "Create a Cluster"
+   - Choose your GPU type, node count, and region
+   - Choose an [Axolotl cloud docker image](https://docs.axolotl.ai/docs/docker.html#cloud)
+   - Deploy the cluster
+
+2. **Connect to the Controller Node**: Find the controller node in the RunPod console and connect via SSH
+
+3. **Follow the instructions in [Standard SLURM Clusters](#standard-slurm-clusters)**
+
+## Additional Resources
+
+- [Axolotl Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
+- [SLURM Documentation](https://slurm.schedmd.com/documentation.html)
+- [RunPod SLURM Clusters Guide](https://docs.runpod.io/instant-clusters/slurm-clusters)

examples/slurm/axolotl.slurm

@@ -0,0 +1,20 @@
+#!/bin/bash
+# Prior to running this script, export your HF_TOKEN and WANDB_API_KEY to your environment; i.e.
+# export HF_TOKEN="..."
+# export WANDB_API_KEY="..."
+#
+
+# ---------- SBATCH commands ---------- #
+#SBATCH --job-name=axolotl-slurm-multinode
+#SBATCH --ntasks-per-node=1
+#SBATCH --nodes=$NUM_NODES
+#SBATCH --gpus-per-task=8
+#SBATCH --cpus-per-task=128
+
+export TORCH_DIST_INIT_BARRIER=0
+
+srun axolotl preprocess train.yaml
+
+srun axolotl train train.yaml --launcher torchrun -- \
+    --nproc_per_node=$NUM_TRAINERS --nnodes=$NUM_NODES \
+    --rdzv_id axolotl-cli --rdzv_backend c10d --rdzv_endpoint "${PRIMARY_ADDR}:${PRIMARY_PORT}" --rdzv-conf="join_timeout=1800"

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

@@ -6,17 +6,14 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 
 ## Getting started
 
-1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Voxtral is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
 
-    Here is an example of how to install from main for pip:
+    Here is an example of how to install from 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]'
+pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Please install the below.
@@ -25,6 +22,9 @@ pip3 install --no-build-isolation -e '.[flash-attn]'
 # 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:

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

@@ -1,7 +1,7 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 
 # START section of dependencies that don't install on Darwin/MacOS
-bitsandbytes==0.46.0
+bitsandbytes==0.47.0
 triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
@@ -12,19 +12,21 @@ liger-kernel==0.6.1
 packaging==23.2
 
 huggingface_hub>=0.33.0
-peft==0.16.0
-transformers==4.54.1
+peft>=0.17.0
+transformers==4.56.1
 tokenizers>=0.21.1
-accelerate @ git+https://github.com/huggingface/accelerate.git@9359a0194f210624f1e6e85c3d838fdd55c11152
+accelerate==1.10.1
 datasets==4.0.0
 deepspeed>=0.17.0
-trl==0.20.0
+trl==0.23.0
 hf_xet==1.1.5
+kernels==0.9.0
+trackio
 
 optimum==1.16.2
 hf_transfer
 sentencepiece
-gradio==5.23.3
+gradio==5.41.1
 
 modal==1.0.2
 pydantic==2.10.6
@@ -62,10 +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==0.0.3
+axolotl-contribs-mit==0.0.5
 
 mistral-common==1.8.3

scripts/bench_moe.py

@@ -0,0 +1,209 @@
+#!/usr/bin/env python
+"""Benchmark Hugging Face Qwen2 MoE block with and without grouped_mm."""
+
+from __future__ import annotations
+
+import argparse
+import sys
+import time
+import weakref
+from pathlib import Path
+
+import torch
+import torch._dynamo as dynamo
+
+try:
+    from axolotl.kernels.moe import torch_grouped as tg
+except Exception:  # pragma: no cover
+    tg = None
+
+
+def bench(run, *, iters: int, warmup: int, sync: bool = True) -> float:
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    for _ in range(warmup):
+        run()
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+    times = []
+    for _ in range(iters):
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+        start = time.perf_counter()
+        run()
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+        times.append((time.perf_counter() - start) * 1000.0)
+    return sum(times) / len(times)
+
+
+def estimate_moe_flops(tokens: int, hidden: int, inter: int, top_k: int) -> float:
+    return 6.0 * tokens * top_k * hidden * inter
+
+
+def load_hf_block(
+    hidden: int,
+    inter: int,
+    experts: int,
+    top_k: int,
+    *,
+    device: torch.device,
+    dtype: torch.dtype,
+):
+    project_root = Path(__file__).resolve().parents[2]
+    transformers_src = project_root / "transformers" / "src"
+    if transformers_src.exists() and str(transformers_src) not in sys.path:
+        sys.path.append(str(transformers_src))
+
+    from transformers.models.qwen2_moe.configuration_qwen2_moe import Qwen2MoeConfig
+    from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
+
+    cfg = Qwen2MoeConfig(
+        hidden_size=hidden,
+        moe_intermediate_size=inter,
+        shared_expert_intermediate_size=inter,
+        num_experts=experts,
+        num_experts_per_tok=top_k,
+        norm_topk_prob=True,
+        qkv_bias=True,
+    )
+
+    block = Qwen2MoeSparseMoeBlock(cfg).to(device=device, dtype=dtype)
+    block_grouped = Qwen2MoeSparseMoeBlock(cfg).to(device=device, dtype=dtype)
+    block_grouped.load_state_dict(block.state_dict())
+    return block, block_grouped
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description="Qwen2 MoE grouped_mm benchmark")
+    p.add_argument("--bsz", type=int, default=8)
+    p.add_argument("--seq", type=int, default=1024)
+    p.add_argument("--hidden", type=int, default=4096)
+    p.add_argument("--inter", type=int, default=14336)
+    p.add_argument("--experts", type=int, default=32)
+    p.add_argument("--top_k", type=int, default=4)
+    p.add_argument("--dtype", choices=["bf16", "fp16", "fp32"], default="bf16")
+    p.add_argument("--iters", type=int, default=50)
+    p.add_argument("--warmup", type=int, default=10)
+    p.add_argument("--profile", action="store_true")
+    p.add_argument(
+        "--compile",
+        action="store_true",
+        help="Torch.compile both paths before benchmarking",
+    )
+    args = p.parse_args()
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    dtype = {
+        "bf16": torch.bfloat16,
+        "fp16": torch.float16,
+        "fp32": torch.float32,
+    }[args.dtype]
+
+    torch.manual_seed(0)
+    if device.type == "cuda":
+        torch.cuda.manual_seed(0)
+
+    block_naive, block_grouped = load_hf_block(
+        args.hidden,
+        args.inter,
+        args.experts,
+        args.top_k,
+        device=device,
+        dtype=dtype,
+    )
+
+    tokens = args.bsz * args.seq
+    flops_total = estimate_moe_flops(tokens, args.hidden, args.inter, args.top_k)
+    print(
+        f"Device={device} dtype={dtype} tokens={tokens} hidden={args.hidden} inter={args.inter} "
+        f"experts={args.experts} top_k={args.top_k}"
+    )
+
+    x = torch.randn(args.bsz, args.seq, args.hidden, device=device, dtype=dtype)
+
+    # Optional torch.compile
+    run_grouped_impl = None
+    if args.compile:
+        dynamo.config.capture_scalar_outputs = True
+        dynamo.config.allow_unspec_int_on_nn_module = True
+        try:
+            block_naive = torch.compile(block_naive)  # type: ignore[arg-type]
+        except Exception as exc:  # pragma: no cover
+            print(f"torch.compile naive failed ({exc}); using eager")
+        else:
+
+            def grouped_forward(inp, *, block=block_grouped):
+                block.experts._ax_parent_block_ref = weakref.ref(block)  # type: ignore[attr-defined]
+                y, _ = tg.moe_ffn_forward_grouped(
+                    inp, block.gate, block.experts, block.top_k
+                )
+                return y
+
+            try:
+                run_grouped_impl = torch.compile(grouped_forward)  # type: ignore[arg-type]
+            except Exception as exc:  # pragma: no cover
+                print(f"torch.compile grouped failed ({exc}); using eager")
+                run_grouped_impl = None
+
+    def run_naive(block=block_naive, data=x):
+        y, _ = block(data)
+        return y
+
+    def run_grouped(block=block_grouped, data=x, impl=run_grouped_impl):
+        if impl is not None:
+            return impl(data)
+        if tg is None or not tg.available():
+            return torch.empty(0)
+        block.experts._ax_parent_block_ref = weakref.ref(block)  # type: ignore[attr-defined]
+        y, _ = tg.moe_ffn_forward_grouped(data, block.gate, block.experts, block.top_k)
+        return y if y is not None else torch.empty(0)
+
+    t_naive = bench(run_naive, iters=args.iters, warmup=args.warmup)
+    tflops_naive = flops_total / ((t_naive / 1000.0) * 1e12)
+    print(
+        f"naive\t{t_naive:.2f} ms\t{tokens / (t_naive / 1000.0):.1f} tok/s\t{tflops_naive:.2f} TFLOP/s"
+    )
+
+    with torch.no_grad():
+        y_ref = run_naive()
+
+    if tg is None or not tg.available():
+        print("torch_grouped\tN/A (unavailable)")
+        return
+
+    y_grouped = run_grouped()
+    if y_grouped.numel() == 0:
+        print("torch_grouped\tN/A (op not callable)")
+        return
+
+    t_grouped = bench(run_grouped, iters=args.iters, warmup=args.warmup)
+    tflops_grouped = flops_total / ((t_grouped / 1000.0) * 1e12)
+    speedup = t_naive / t_grouped
+    print(
+        f"torch_grouped\t{t_grouped:.2f} ms\t{tokens / (t_grouped / 1000.0):.1f} tok/s\t"
+        f"{tflops_grouped:.2f} TFLOP/s\t{speedup:.2f}×"
+    )
+
+    diff = (y_ref.float() - y_grouped.float()).abs()
+    print(
+        "torch_grouped_check: "
+        f"max_abs={diff.max().item():.3e} mean_abs={diff.mean().item():.3e} "
+        f"rel_l2={(diff.pow(2).sum() / (y_ref.float().pow(2).sum() + 1e-12)).sqrt().item():.3e}"
+    )
+
+    if args.profile:
+        with torch.profiler.profile(
+            activities=[torch.profiler.ProfilerActivity.CUDA], record_shapes=True
+        ) as prof:
+            run_naive()
+        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=20))
+
+        with torch.profiler.profile(
+            activities=[torch.profiler.ProfilerActivity.CUDA], record_shapes=True
+        ) as prof:
+            run_grouped()
+        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=20))
+
+
+if __name__ == "__main__":
+    main()

scripts/bench_moe_sweep.py

@@ -0,0 +1,311 @@
+#!/usr/bin/env python
+"""Sweep grouped_mm vs naive performance for Qwen2 MoE block."""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import sys
+import time
+import weakref
+from dataclasses import dataclass
+from pathlib import Path
+from typing import List
+
+import torch
+import torch._dynamo as dynamo
+
+try:
+    from axolotl.kernels.moe import torch_grouped as tg
+except Exception:  # pragma: no cover
+    tg = None
+
+
+def _parse_list(arg: str) -> List[int]:
+    return [int(v) for v in arg.split(",") if v]
+
+
+def _bench(run, *, iters: int, warmup: int, device: torch.device) -> float:
+    for _ in range(warmup):
+        run()
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+    times: List[float] = []
+    for _ in range(iters):
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+        start = time.perf_counter()
+        run()
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+        times.append((time.perf_counter() - start) * 1000.0)
+    return sum(times) / len(times)
+
+
+def _estimate_flops(tokens: int, hidden: int, inter: int, top_k: int) -> float:
+    return 6.0 * tokens * top_k * hidden * inter
+
+
+def _load_block(
+    hidden: int,
+    inter: int,
+    experts: int,
+    top_k: int,
+    *,
+    device: torch.device,
+    dtype: torch.dtype,
+):
+    project_root = Path(__file__).resolve().parents[2]
+    transformers_src = project_root / "transformers" / "src"
+    if transformers_src.exists() and str(transformers_src) not in sys.path:
+        sys.path.append(str(transformers_src))
+
+    from transformers.models.qwen2_moe.configuration_qwen2_moe import Qwen2MoeConfig
+    from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
+
+    cfg = Qwen2MoeConfig(
+        hidden_size=hidden,
+        moe_intermediate_size=inter,
+        shared_expert_intermediate_size=inter,
+        num_experts=experts,
+        num_experts_per_tok=top_k,
+        norm_topk_prob=True,
+        qkv_bias=True,
+    )
+
+    block = Qwen2MoeSparseMoeBlock(cfg).to(device=device, dtype=dtype)
+    block_grouped = Qwen2MoeSparseMoeBlock(cfg).to(device=device, dtype=dtype)
+    block_grouped.load_state_dict(block.state_dict())
+    return block, block_grouped
+
+
+@dataclass
+class Result:
+    bsz: int
+    seq: int
+    hidden: int
+    inter: int
+    experts: int
+    top_k: int
+    dtype: str
+    naive_ms: float
+    grouped_ms: float
+    speedup: float
+    naive_tflops: float
+    grouped_tflops: float
+    max_abs: float
+    mean_abs: float
+    rel_l2: float
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description="Grouped MoE sweep")
+    p.add_argument("--batch-sizes", default="4,8,16")
+    p.add_argument("--seq-lens", default="512,1024,2048")
+    p.add_argument("--hidden", default="2048,4096")
+    p.add_argument("--inter", default="5632,8192,14336")
+    p.add_argument("--experts", default="8,16,32")
+    p.add_argument("--top-k", default="1,2,4")
+    p.add_argument("--dtype", choices=["bf16", "fp16", "fp32"], default="bf16")
+    p.add_argument("--iters", type=int, default=25)
+    p.add_argument("--warmup", type=int, default=5)
+    p.add_argument("--csv", type=Path, default=None)
+    p.add_argument("--compile", action="store_true")
+    args = p.parse_args()
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    dtype = {
+        "bf16": torch.bfloat16,
+        "fp16": torch.float16,
+        "fp32": torch.float32,
+    }[args.dtype]
+
+    if tg is None or not tg.available():
+        print("torch_grouped unavailable; sweep aborted")
+        return
+
+    bs_list = _parse_list(args.batch_sizes)
+    seq_list = _parse_list(args.seq_lens)
+    hidden_list = _parse_list(args.hidden)
+    inter_list = _parse_list(args.inter)
+    expert_list = _parse_list(args.experts)
+    topk_list = _parse_list(args.top_k)
+
+    results: List[Result] = []
+
+    print(
+        "bsz\tseq\thidden\tinter\texperts\ttop_k\tnaive(ms)\tgrouped(ms)\tspeedup\t"
+        "naive TF/s\tgrouped TF/s\tmax_abs\tmean_abs\trel_l2"
+    )
+
+    for bsz in bs_list:
+        for seq in seq_list:
+            tokens = bsz * seq
+            for hidden in hidden_list:
+                for inter in inter_list:
+                    for experts in expert_list:
+                        for top_k in topk_list:
+                            torch.manual_seed(0)
+                            if device.type == "cuda":
+                                torch.cuda.manual_seed(0)
+
+                            block_naive, block_grouped = _load_block(
+                                hidden,
+                                inter,
+                                experts,
+                                top_k,
+                                device=device,
+                                dtype=dtype,
+                            )
+
+                            x = torch.randn(
+                                bsz, seq, hidden, device=device, dtype=dtype
+                            )
+
+                            compiled_impl = None
+                            if args.compile:
+                                dynamo.config.capture_scalar_outputs = True
+                                dynamo.config.allow_unspec_int_on_nn_module = True
+                                try:
+                                    block_naive = torch.compile(block_naive)  # type: ignore[arg-type]
+                                except Exception as exc:
+                                    print(
+                                        f"torch.compile naive failed ({exc}); using eager"
+                                    )
+                                else:
+
+                                    def grouped_forward(inp, *, block=block_grouped):
+                                        block.experts._ax_parent_block_ref = (
+                                            weakref.ref(block)
+                                        )  # type: ignore[attr-defined]
+                                        y, _ = tg.moe_ffn_forward_grouped(
+                                            inp,
+                                            block.gate,
+                                            block.experts,
+                                            block.top_k,
+                                        )
+                                        return y
+
+                                    try:
+                                        compiled_impl = torch.compile(grouped_forward)  # type: ignore[arg-type]
+                                    except Exception as exc:
+                                        print(
+                                            f"torch.compile grouped failed ({exc}); using eager"
+                                        )
+                                        compiled_impl = None
+
+                            def run_naive(block=block_naive, data=x):
+                                y, _ = block(data)
+                                return y
+
+                            def run_grouped(
+                                block=block_grouped, data=x, impl=compiled_impl
+                            ):
+                                if impl is not None:
+                                    return impl(data)
+                                block.experts._ax_parent_block_ref = weakref.ref(block)  # type: ignore[attr-defined]
+                                y, _ = tg.moe_ffn_forward_grouped(
+                                    data,
+                                    block.gate,
+                                    block.experts,
+                                    block.top_k,
+                                )
+                                return y
+
+                            naive_ms = _bench(
+                                run_naive,
+                                iters=args.iters,
+                                warmup=args.warmup,
+                                device=device,
+                            )
+                            y_naive = run_naive()
+
+                            grouped_ms = _bench(
+                                run_grouped,
+                                iters=args.iters,
+                                warmup=args.warmup,
+                                device=device,
+                            )
+                            y_grouped = run_grouped()
+
+                            diff = (y_naive.float() - y_grouped.float()).abs()
+                            res = Result(
+                                bsz,
+                                seq,
+                                hidden,
+                                inter,
+                                experts,
+                                top_k,
+                                args.dtype,
+                                naive_ms,
+                                grouped_ms,
+                                naive_ms / grouped_ms,
+                                _estimate_flops(tokens, hidden, inter, top_k)
+                                / ((naive_ms / 1000.0) * 1e12),
+                                _estimate_flops(tokens, hidden, inter, top_k)
+                                / ((grouped_ms / 1000.0) * 1e12),
+                                diff.max().item(),
+                                diff.mean().item(),
+                                (
+                                    (
+                                        diff.pow(2).sum()
+                                        / (y_naive.float().pow(2).sum() + 1e-12)
+                                    )
+                                    .sqrt()
+                                    .item()
+                                ),
+                            )
+                            results.append(res)
+                            print(
+                                f"{bsz}\t{seq}\t{hidden}\t{inter}\t{experts}\t{top_k}\t{res.naive_ms:.2f}\t"
+                                f"{res.grouped_ms:.2f}\t{res.speedup:.2f}\t{res.naive_tflops:.2f}\t"
+                                f"{res.grouped_tflops:.2f}\t{res.max_abs:.2e}\t{res.mean_abs:.2e}\t{res.rel_l2:.2e}"
+                            )
+
+    if args.csv:
+        fieldnames = [
+            "bsz",
+            "seq",
+            "hidden",
+            "inter",
+            "experts",
+            "top_k",
+            "dtype",
+            "naive_ms",
+            "grouped_ms",
+            "speedup",
+            "naive_tflops",
+            "grouped_tflops",
+            "max_abs",
+            "mean_abs",
+            "rel_l2",
+        ]
+        with args.csv.open("w", newline="") as f:
+            writer = csv.DictWriter(f, fieldnames=fieldnames)
+            writer.writeheader()
+            for r in results:
+                writer.writerow(
+                    {
+                        "bsz": r.bsz,
+                        "seq": r.seq,
+                        "hidden": r.hidden,
+                        "inter": r.inter,
+                        "experts": r.experts,
+                        "top_k": r.top_k,
+                        "dtype": r.dtype,
+                        "naive_ms": f"{r.naive_ms:.4f}",
+                        "grouped_ms": f"{r.grouped_ms:.4f}",
+                        "speedup": f"{r.speedup:.4f}",
+                        "naive_tflops": f"{r.naive_tflops:.4f}",
+                        "grouped_tflops": f"{r.grouped_tflops:.4f}",
+                        "max_abs": f"{r.max_abs:.6e}",
+                        "mean_abs": f"{r.mean_abs:.6e}",
+                        "rel_l2": f"{r.rel_l2:.6e}",
+                    }
+                )
+
+
+if __name__ == "__main__":
+    import weakref
+
+    main()

scripts/bench_torchtitan_moe.py

@@ -0,0 +1,205 @@
+#!/usr/bin/env python
+"""Benchmark Torchtitan MoE grouped vs naive expert execution."""
+
+from __future__ import annotations
+
+import argparse
+import sys
+import time
+from pathlib import Path
+
+import torch
+
+# Ensure torchtitan is importable when running from the axolotl tree
+_PROJECT_ROOT = Path(__file__).resolve().parents[2]
+_TITAN_PATH = _PROJECT_ROOT / "torchtitan"
+if str(_TITAN_PATH) not in sys.path:
+    sys.path.append(str(_TITAN_PATH))
+
+from torchtitan.models.moe import MoE, MoEArgs
+
+
+def _parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description="Torchtitan MoE microbenchmark")
+    p.add_argument("--bsz", type=int, default=8)
+    p.add_argument("--seq", type=int, default=1024)
+    p.add_argument("--hidden", type=int, default=4096)
+    p.add_argument("--inter", type=int, default=14336)
+    p.add_argument("--experts", type=int, default=8)
+    p.add_argument("--top_k", type=int, default=2)
+    p.add_argument("--dtype", choices=["bf16", "fp16", "fp32"], default="bf16")
+    p.add_argument("--iters", type=int, default=50)
+    p.add_argument("--warmup", type=int, default=10)
+    p.add_argument("--init-std", type=float, default=0.02)
+    p.add_argument(
+        "--score-before",
+        action="store_true",
+        help="Apply routing scores before expert computation (default: after)",
+    )
+    p.add_argument(
+        "--score-func",
+        choices=["softmax", "sigmoid"],
+        default="softmax",
+    )
+    p.add_argument(
+        "--route-norm",
+        action="store_true",
+        help="Enable Torchtitan router normalization when using sigmoid scores.",
+    )
+    return p.parse_args()
+
+
+def _map_dtype(arg: str) -> torch.dtype:
+    return {
+        "bf16": torch.bfloat16,
+        "fp16": torch.float16,
+        "fp32": torch.float32,
+    }[arg]
+
+
+def _estimate_moe_flops(tokens: int, hidden: int, inter: int, top_k: int) -> float:
+    # Two up projections + one down projection per expert/token combination.
+    return 6.0 * tokens * top_k * hidden * inter
+
+
+def _prepare_module(
+    moe: MoE,
+    *,
+    device: torch.device,
+    dtype: torch.dtype,
+) -> MoE:
+    moe = moe.to(device=device)
+    for param in moe.parameters():
+        param.data = param.data.to(dtype)
+        if param.grad is not None:
+            param.grad = None
+
+    buffers = dict(moe.named_buffers())
+    for name, buf in buffers.items():
+        if name == "tokens_per_expert":
+            moe._buffers[name] = torch.zeros_like(
+                buf, dtype=torch.float32, device=device
+            )
+        elif name == "expert_bias" and buf is not None:
+            moe._buffers[name] = torch.zeros_like(
+                buf, dtype=torch.float32, device=device
+            )
+        else:
+            moe._buffers[name] = buf.to(device=device, dtype=dtype)
+    moe.eval()
+    return moe
+
+
+@torch.inference_mode()
+def _forward_fn(module: MoE, x: torch.Tensor) -> torch.Tensor:
+    return module(x)
+
+
+def _bench(fn, *, iters: int, warmup: int, sync: bool = True) -> float:
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    for _ in range(warmup):
+        fn()
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+    times = []
+    for _ in range(iters):
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+        start = time.perf_counter()
+        fn()
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+        times.append((time.perf_counter() - start) * 1000.0)
+    return sum(times) / len(times)
+
+
+def main() -> None:
+    args = _parse_args()
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    dtype = _map_dtype(args.dtype)
+
+    torch.manual_seed(0)
+    if device.type == "cuda":
+        torch.cuda.manual_seed(0)
+
+    moe_args_grouped = MoEArgs(
+        num_experts=args.experts,
+        num_shared_experts=0,
+        score_func=args.score_func,
+        route_norm=args.route_norm,
+        top_k=args.top_k,
+        use_grouped_mm=True,
+        score_before_experts=args.score_before,
+        load_balance_coeff=None,
+    )
+    moe_grouped = MoE(moe_args_grouped, dim=args.hidden, hidden_dim=args.inter)
+    moe_grouped.init_weights(args.init_std, buffer_device=device)
+
+    moe_args_naive = MoEArgs(
+        num_experts=args.experts,
+        num_shared_experts=0,
+        score_func=args.score_func,
+        route_norm=args.route_norm,
+        top_k=args.top_k,
+        use_grouped_mm=False,
+        score_before_experts=args.score_before,
+        load_balance_coeff=None,
+    )
+    moe_naive = MoE(moe_args_naive, dim=args.hidden, hidden_dim=args.inter)
+    moe_naive.load_state_dict(moe_grouped.state_dict(), strict=True)
+
+    moe_grouped = _prepare_module(moe_grouped, device=device, dtype=dtype)
+    moe_naive = _prepare_module(moe_naive, device=device, dtype=dtype)
+
+    x = torch.randn(args.bsz, args.seq, args.hidden, device=device, dtype=dtype)
+
+    tokens = args.bsz * args.seq
+    print(
+        f"Device={device} dtype={dtype} tokens={tokens} hidden={args.hidden} "
+        f"inter={args.inter} experts={args.experts} top_k={args.top_k}"
+    )
+
+    def run_naive():
+        return _forward_fn(moe_naive, x)
+
+    def run_grouped():
+        return _forward_fn(moe_grouped, x)
+
+    if hasattr(moe_naive, "tokens_per_expert"):
+        moe_naive.tokens_per_expert.zero_()
+    if hasattr(moe_grouped, "tokens_per_expert"):
+        moe_grouped.tokens_per_expert.zero_()
+
+    t_naive = _bench(run_naive, iters=args.iters, warmup=args.warmup)
+    flops = _estimate_moe_flops(tokens, args.hidden, args.inter, args.top_k)
+    tflops_naive = flops / ((t_naive / 1000.0) * 1e12)
+    print(
+        f"naive\t{t_naive:.2f} ms\t{tokens / (t_naive / 1000.0):.1f} tok/s\t"
+        f"{tflops_naive:.2f} TFLOP/s"
+    )
+
+    y_naive = run_naive()
+
+    if hasattr(moe_grouped, "tokens_per_expert"):
+        moe_grouped.tokens_per_expert.zero_()
+
+    t_grouped = _bench(run_grouped, iters=args.iters, warmup=args.warmup)
+    tflops_grouped = flops / ((t_grouped / 1000.0) * 1e12)
+    speedup = t_naive / t_grouped if t_grouped > 0 else float("nan")
+    print(
+        f"grouped\t{t_grouped:.2f} ms\t{tokens / (t_grouped / 1000.0):.1f} tok/s\t"
+        f"{tflops_grouped:.2f} TFLOP/s\t{speedup:.2f}×"
+    )
+
+    y_grouped = run_grouped()
+    diff = (y_naive.float() - y_grouped.float()).abs()
+    max_abs = diff.max().item()
+    mean_abs = diff.mean().item()
+    rel_l2 = (diff.pow(2).sum() / (y_naive.float().pow(2).sum() + 1e-12)).sqrt().item()
+    print(
+        f"grouped_check: max_abs={max_abs:.3e} mean_abs={mean_abs:.3e} rel_l2={rel_l2:.3e}"
+    )
+
+
+if __name__ == "__main__":
+    main()

scripts/bench_torchtitan_moe_sweep.py

@@ -0,0 +1,328 @@
+#!/usr/bin/env python
+"""Sweep Torchtitan MoE grouped vs naive configurations and report performance."""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import sys
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Iterable, List
+
+import torch
+
+_PROJECT_ROOT = Path(__file__).resolve().parents[2]
+_TITAN_PATH = _PROJECT_ROOT / "torchtitan"
+if str(_TITAN_PATH) not in sys.path:
+    sys.path.append(str(_TITAN_PATH))
+
+from torchtitan.models.moe import MoE, MoEArgs
+
+
+def _parse_int_list(value: str) -> List[int]:
+    return [int(v) for v in value.split(",") if v]
+
+
+def _parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description="Torchtitan MoE grouped vs naive sweep")
+    p.add_argument(
+        "--batch-sizes", default="4,8,16", help="Comma separated batch sizes"
+    )
+    p.add_argument(
+        "--seq-lens", default="1024,2048", help="Comma separated sequence lengths"
+    )
+    p.add_argument(
+        "--experts", default="8,16,32,64", help="Comma separated expert counts"
+    )
+    p.add_argument("--top-ks", default="1,2,4", help="Comma separated top_k choices")
+    p.add_argument("--hidden", type=int, default=4096)
+    p.add_argument("--inter", type=int, default=14336)
+    p.add_argument("--dtype", choices=["bf16", "fp16", "fp32"], default="bf16")
+    p.add_argument("--iters", type=int, default=25)
+    p.add_argument("--warmup", type=int, default=5)
+    p.add_argument("--init-std", type=float, default=0.02)
+    p.add_argument("--score-before", action="store_true")
+    p.add_argument("--score-func", choices=["softmax", "sigmoid"], default="softmax")
+    p.add_argument("--route-norm", action="store_true")
+    p.add_argument("--csv", type=Path, default=None, help="Optional CSV output path")
+    return p.parse_args()
+
+
+def _map_dtype(arg: str) -> torch.dtype:
+    return {
+        "bf16": torch.bfloat16,
+        "fp16": torch.float16,
+        "fp32": torch.float32,
+    }[arg]
+
+
+def _estimate_flops(tokens: int, hidden: int, inter: int, top_k: int) -> float:
+    return 6.0 * tokens * top_k * hidden * inter
+
+
+def _prepare_module(module: MoE, *, device: torch.device, dtype: torch.dtype) -> MoE:
+    module = module.to(device=device)
+    for param in module.parameters():
+        param.data = param.data.to(dtype)
+        if param.grad is not None:
+            param.grad = None
+    for name, buf in module.named_buffers():
+        if name == "tokens_per_expert":
+            module._buffers[name] = torch.zeros_like(
+                buf, dtype=torch.float32, device=device
+            )
+        elif name == "expert_bias" and buf is not None:
+            module._buffers[name] = torch.zeros_like(
+                buf, dtype=torch.float32, device=device
+            )
+        else:
+            module._buffers[name] = buf.to(device=device, dtype=dtype)
+    module.eval()
+    return module
+
+
+@torch.inference_mode()
+def _forward(module: MoE, x: torch.Tensor) -> torch.Tensor:
+    return module(x)
+
+
+def _bench(callable_, *, iters: int, warmup: int, device: torch.device) -> float:
+    for _ in range(warmup):
+        callable_()
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+    timings: List[float] = []
+    for _ in range(iters):
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+        start = time.perf_counter()
+        callable_()
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+        timings.append((time.perf_counter() - start) * 1000.0)
+    return sum(timings) / len(timings)
+
+
+@dataclass
+class SweepResult:
+    bsz: int
+    seq: int
+    experts: int
+    top_k: int
+    dtype: str
+    naive_ms: float
+    grouped_ms: float
+    speedup: float
+    naive_tflops: float
+    grouped_tflops: float
+    max_abs: float
+    mean_abs: float
+    rel_l2: float
+
+
+def _run_case(
+    *,
+    bsz: int,
+    seq: int,
+    experts: int,
+    top_k: int,
+    hidden: int,
+    inter: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    iters: int,
+    warmup: int,
+    init_std: float,
+    score_before: bool,
+    score_func: str,
+    route_norm: bool,
+) -> SweepResult:
+    torch.manual_seed(0)
+    if device.type == "cuda":
+        torch.cuda.manual_seed(0)
+
+    moe_args_grouped = MoEArgs(
+        num_experts=experts,
+        num_shared_experts=0,
+        score_func=score_func,
+        route_norm=route_norm,
+        top_k=top_k,
+        use_grouped_mm=True,
+        score_before_experts=score_before,
+        load_balance_coeff=None,
+    )
+    moe_grouped = MoE(moe_args_grouped, dim=hidden, hidden_dim=inter)
+    moe_grouped.init_weights(init_std, buffer_device=device)
+
+    moe_args_naive = MoEArgs(
+        num_experts=experts,
+        num_shared_experts=0,
+        score_func=score_func,
+        route_norm=route_norm,
+        top_k=top_k,
+        use_grouped_mm=False,
+        score_before_experts=score_before,
+        load_balance_coeff=None,
+    )
+    moe_naive = MoE(moe_args_naive, dim=hidden, hidden_dim=inter)
+    moe_naive.load_state_dict(moe_grouped.state_dict(), strict=True)
+
+    moe_grouped = _prepare_module(moe_grouped, device=device, dtype=dtype)
+    moe_naive = _prepare_module(moe_naive, device=device, dtype=dtype)
+
+    x = torch.randn(bsz, seq, hidden, device=device, dtype=dtype)
+
+    def run_naive():
+        if hasattr(moe_naive, "tokens_per_expert"):
+            moe_naive.tokens_per_expert.zero_()
+        return _forward(moe_naive, x)
+
+    def run_grouped():
+        if hasattr(moe_grouped, "tokens_per_expert"):
+            moe_grouped.tokens_per_expert.zero_()
+        return _forward(moe_grouped, x)
+
+    naive_ms = _bench(run_naive, iters=iters, warmup=warmup, device=device)
+    y_naive = run_naive()
+
+    grouped_ms = _bench(run_grouped, iters=iters, warmup=warmup, device=device)
+    y_grouped = run_grouped()
+
+    diff = (y_naive.float() - y_grouped.float()).abs()
+    max_abs = diff.max().item()
+    mean_abs = diff.mean().item()
+    rel_l2 = (diff.pow(2).sum() / (y_naive.float().pow(2).sum() + 1e-12)).sqrt().item()
+
+    tokens = bsz * seq
+    flops = _estimate_flops(tokens, hidden, inter, top_k)
+    naive_tflops = flops / ((naive_ms / 1000.0) * 1e12)
+    grouped_tflops = flops / ((grouped_ms / 1000.0) * 1e12)
+    speedup = naive_ms / grouped_ms if grouped_ms > 0 else float("nan")
+
+    return SweepResult(
+        bsz=bsz,
+        seq=seq,
+        experts=experts,
+        top_k=top_k,
+        dtype=str(dtype),
+        naive_ms=naive_ms,
+        grouped_ms=grouped_ms,
+        speedup=speedup,
+        naive_tflops=naive_tflops,
+        grouped_tflops=grouped_tflops,
+        max_abs=max_abs,
+        mean_abs=mean_abs,
+        rel_l2=rel_l2,
+    )
+
+
+def _print_header(
+    hidden: int, inter: int, dtype: torch.dtype, device: torch.device
+) -> None:
+    print(f"Device={device} dtype={dtype} hidden={hidden} inter={inter}")
+    print(
+        "bsz\tseq\texperts\ttop_k\tnaive(ms)\tgrouped(ms)\tspeedup\t"
+        "naive TF/s\tgrouped TF/s\tmax_abs\tmean_abs\trel_l2"
+    )
+
+
+def _print_result(res: SweepResult) -> None:
+    print(
+        f"{res.bsz}\t{res.seq}\t{res.experts}\t{res.top_k}\t"
+        f"{res.naive_ms:.2f}\t{res.grouped_ms:.2f}\t{res.speedup:.2f}\t"
+        f"{res.naive_tflops:.2f}\t{res.grouped_tflops:.2f}\t"
+        f"{res.max_abs:.2e}\t{res.mean_abs:.2e}\t{res.rel_l2:.2e}"
+    )
+
+
+def _write_csv(path: Path, results: Iterable[SweepResult]) -> None:
+    fieldnames = [
+        "batch_size",
+        "seq_len",
+        "experts",
+        "top_k",
+        "dtype",
+        "naive_ms",
+        "grouped_ms",
+        "speedup",
+        "naive_tflops",
+        "grouped_tflops",
+        "max_abs",
+        "mean_abs",
+        "rel_l2",
+    ]
+    with path.open("w", newline="") as f:
+        writer = csv.DictWriter(f, fieldnames=fieldnames)
+        writer.writeheader()
+        for r in results:
+            writer.writerow(
+                {
+                    "batch_size": r.bsz,
+                    "seq_len": r.seq,
+                    "experts": r.experts,
+                    "top_k": r.top_k,
+                    "dtype": r.dtype,
+                    "naive_ms": f"{r.naive_ms:.4f}",
+                    "grouped_ms": f"{r.grouped_ms:.4f}",
+                    "speedup": f"{r.speedup:.4f}",
+                    "naive_tflops": f"{r.naive_tflops:.4f}",
+                    "grouped_tflops": f"{r.grouped_tflops:.4f}",
+                    "max_abs": f"{r.max_abs:.6e}",
+                    "mean_abs": f"{r.mean_abs:.6e}",
+                    "rel_l2": f"{r.rel_l2:.6e}",
+                }
+            )
+
+
+def main() -> None:
+    args = _parse_args()
+    dtype = _map_dtype(args.dtype)
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+
+    batch_sizes = _parse_int_list(args.batch_sizes)
+    seq_lens = _parse_int_list(args.seq_lens)
+    experts_list = _parse_int_list(args.experts)
+    top_ks = _parse_int_list(args.top_ks)
+
+    results: List[SweepResult] = []
+    _print_header(args.hidden, args.inter, dtype, device)
+
+    for bsz in batch_sizes:
+        for seq in seq_lens:
+            for experts in experts_list:
+                for top_k in top_ks:
+                    try:
+                        res = _run_case(
+                            bsz=bsz,
+                            seq=seq,
+                            experts=experts,
+                            top_k=top_k,
+                            hidden=args.hidden,
+                            inter=args.inter,
+                            dtype=dtype,
+                            device=device,
+                            iters=args.iters,
+                            warmup=args.warmup,
+                            init_std=args.init_std,
+                            score_before=args.score_before,
+                            score_func=args.score_func,
+                            route_norm=args.route_norm,
+                        )
+                    except RuntimeError as err:
+                        print(
+                            f"{bsz}\t{seq}\t{experts}\t{top_k}\tERROR: {err}",
+                            file=sys.stderr,
+                        )
+                        continue
+                    results.append(res)
+                    _print_result(res)
+
+    if args.csv and results:
+        _write_csv(args.csv, results)
+        print(f"Wrote {len(results)} rows to {args.csv}")
+
+
+if __name__ == "__main__":
+    main()

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/cloud-entrypoint.sh

@@ -44,8 +44,13 @@ add_keys_to_authorized() {
     chmod 700 -R ~/.ssh
 }
 
+# Set SSH port
+if [ ! -z "$SSH_PORT" ]; then
+    sed -i "s/#Port 22/Port $SSH_PORT/" /etc/ssh/sshd_config
+fi
+
 if [[ $PUBLIC_KEY ]]; then
-    # runpod
+    # runpod, prime intellect
     add_keys_to_authorized "$PUBLIC_KEY"
     # Start the SSH service in the background
     service ssh start
@@ -76,5 +81,13 @@ if [ ! -L "/workspace/axolotl/outputs" ]; then
     ln -sf /workspace/data/axolotl-artifacts /workspace/axolotl/outputs
 fi
 
+# start the runpod slurm init
+SLURM_INIT="${SLURM_INIT:-/slurm-init.sh}"
+
+if [[ -f "$SLURM_INIT" ]]; then
+  echo "[entrypoint] running $SLURM_INIT..."
+  bash "$SLURM_INIT"
+fi
+
 # Execute the passed arguments (CMD)
 exec "$@"

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@cbd58e0"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c6a32c5"'
 )

scripts/debug_qwen2_experts.py

@@ -0,0 +1,53 @@
+#!/usr/bin/env python
+"""Inspect Qwen2 MoE expert implementations for grouped-mm debugging."""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import torch
+
+ROOT = Path(__file__).resolve().parents[2]
+sys.path.extend(
+    [
+        str(ROOT / "transformers" / "src"),
+        str(ROOT / "src"),
+    ]
+)
+
+from transformers.models.qwen2_moe.configuration_qwen2_moe import Qwen2MoeConfig
+from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
+
+from axolotl.kernels.moe.torch_grouped import _iter_expert_impls
+
+
+def main() -> None:
+    cfg = Qwen2MoeConfig(
+        hidden_size=4096,
+        moe_intermediate_size=14336,
+        shared_expert_intermediate_size=14336,
+        num_experts=32,
+        num_experts_per_tok=4,
+    )
+
+    block = Qwen2MoeSparseMoeBlock(cfg).to("cuda", dtype=torch.bfloat16)
+    experts = block.experts
+    experts._ax_parent_block = block
+
+    impls = _iter_expert_impls(experts)
+    print(f"impl count: {len(impls)}")
+    for idx, impl in enumerate(impls[:8]):
+        has_gate = hasattr(impl, "gate_proj")
+        has_up = hasattr(impl, "up_proj")
+        print(
+            f"impl[{idx}] type={impl.__class__.__name__} has_gate={has_gate} has_up={has_up}"
+        )
+        if has_gate:
+            print(f"  gate shape {tuple(impl.gate_proj.weight.shape)}")
+            print(f"  up shape   {tuple(impl.up_proj.weight.shape)}")
+            print(f"  down shape {tuple(impl.down_proj.weight.shape)}")
+
+
+if __name__ == "__main__":
+    main()

scripts/probe_torch_grouped_ops.py

@@ -0,0 +1,47 @@
+#!/usr/bin/env python
+"""
+Probe PyTorch for grouped GEMM operator names and namespaces.
+Run: python scripts/probe_torch_grouped_ops.py
+"""
+
+import sys
+
+
+def main():
+    try:
+        import torch
+    except Exception as e:
+        print("Failed to import torch:", e)
+        sys.exit(1)
+
+    print("torch version:", torch.__version__)
+    namespaces = [n for n in dir(torch.ops) if not n.startswith("_")]
+    print("ops namespaces:", namespaces)
+
+    found_any = False
+    for ns in namespaces:
+        obj = getattr(torch.ops, ns, None)
+        ops = []
+        if obj is not None:
+            try:
+                ops = dir(obj)
+            except Exception as e:
+                print(f"warning: failed to list ops for namespace {ns}: {e}")
+        cands = [
+            o
+            for o in ops
+            if ("group" in o.lower())
+            or ("mm_grouped" in o.lower())
+            or ("matmul_grouped" in o.lower())
+            or ("grouped" in o.lower())
+        ]
+        if cands:
+            found_any = True
+            print(f"namespace {ns} candidates:", cands)
+
+    if not found_any:
+        print("No grouped GEMM candidates found. PyTorch >= 2.8 is recommended.")
+
+
+if __name__ == "__main__":
+    main()

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 +120,14 @@ 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 +162,7 @@ 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/__init__.py

@@ -4,4 +4,4 @@ import pkgutil
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.12.0.dev"
+__version__ = "0.13.0.dev"

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

src/axolotl/cli/art.py

@@ -22,7 +22,7 @@ HAS_PRINTED_LOGO = False
 def print_axolotl_text_art():
     """Prints axolotl ASCII art."""
 
-    global HAS_PRINTED_LOGO  # pylint: disable=global-statement
+    global HAS_PRINTED_LOGO
     if HAS_PRINTED_LOGO:
         return
     if is_main_process():

src/axolotl/cli/cloud/__init__.py

@@ -7,6 +7,8 @@ from typing import Literal
 
 import yaml
 
+from axolotl.cli.cloud.base import Cloud
+from axolotl.cli.cloud.baseten import BasetenCloud
 from axolotl.cli.cloud.modal_ import ModalCloud
 from axolotl.utils.dict import DictDefault
 
@@ -38,8 +40,15 @@ def do_cli_train(
     cwd=None,
     **kwargs,
 ) -> None:
-    cloud_cfg = load_cloud_cfg(cloud_config)
-    cloud = ModalCloud(cloud_cfg)
+    cloud_cfg: DictDefault = load_cloud_cfg(cloud_config)
+    provider = cloud_cfg.provider or "modal"
+    cloud: Cloud | None
+    if provider == "modal":
+        cloud = ModalCloud(cloud_cfg)
+    elif provider == "baseten":
+        cloud = BasetenCloud(cloud_cfg.to_dict())
+    else:
+        raise ValueError(f"Unsupported cloud provider: {provider}")
     with open(config, "r", encoding="utf-8") as file:
         config_yaml = file.read()
     local_dirs = {}

src/axolotl/cli/cloud/baseten/__init__.py

@@ -0,0 +1,48 @@
+"""Baseten Cloud CLI"""
+
+import shutil
+import subprocess  # nosec B404
+import tempfile
+from os.path import dirname
+from typing import Literal
+
+import yaml
+
+from axolotl.cli.cloud.base import Cloud
+
+
+class BasetenCloud(Cloud):
+    """Baseten Cloud Axolotl CLI"""
+
+    def __init__(self, config: dict):
+        self.config = config
+
+    def preprocess(self, config_yaml: str, *args, **kwargs) -> None:
+        raise NotImplementedError(
+            "Separate preprocess function for Baseten is not "
+            "implemented and will happen during hte train step."
+        )
+
+    def train(
+        self,
+        config_yaml: str,
+        launcher: Literal["accelerate", "torchrun", "python"] = "accelerate",
+        launcher_args: list[str] | None = None,
+        local_dirs: dict[str, str] | None = None,  # pylint: disable=unused-argument
+        **kwargs,
+    ):
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            config = self.config.copy()
+            config["launcher"] = launcher
+            config["launcher_args"] = launcher_args
+            with open(tmp_dir + "/cloud.yaml", "w", encoding="utf-8") as cloud_fout:
+                yaml.dump(config, cloud_fout)
+            with open(tmp_dir + "/train.yaml", "w", encoding="utf-8") as config_fout:
+                config_fout.write(config_yaml)
+            shutil.copyfile(dirname(__file__) + "/template/run.sh", tmp_dir + "/run.sh")
+            shutil.copyfile(
+                dirname(__file__) + "/template/train_sft.py", tmp_dir + "/train_sft.py"
+            )
+            subprocess.run(  # nosec B603 B607
+                ["truss", "train", "push", "train_sft.py"], cwd=tmp_dir, check=False
+            )

src/axolotl/cli/cloud/baseten/template/run.sh

@@ -0,0 +1,9 @@
+#!/bin/bash
+set -eux
+
+export NCCL_SOCKET_IFNAME="^docker0,lo"
+export NCCL_IB_DISABLE=0
+export NCCL_TIMEOUT=1800000
+
+axolotl preprocess train.yaml
+axolotl train train.yaml --launcher ${AXOLOTL_LAUNCHER} ${AXOLOTL_LAUNCHER_ARGS}

src/axolotl/cli/cloud/baseten/template/train_sft.py

@@ -0,0 +1,71 @@
+"""
+Baseten Training Script for Axolotl
+"""
+
+# pylint: skip-file
+import yaml
+from truss.base import truss_config
+
+# Import necessary classes from the Baseten Training SDK
+from truss_train import definitions
+
+cloud_config = yaml.safe_load(open("cloud.yaml", "r"))
+gpu = cloud_config.get("gpu", "h100")
+gpu_count = int(cloud_config.get("gpu_count", 1))
+node_count = int(cloud_config.get("node_count", 1))
+project_name = cloud_config.get("project_name", "axolotl-project") or "axolotl-project"
+secrets = cloud_config.get("secrets", [])
+launcher = cloud_config.get("launcher", "accelerate")
+launcher_args = cloud_config.get("launcher_args", [])
+script_name = "run.sh"
+
+launcher_args_str = ""
+if launcher_args:
+    launcher_args_str = "-- " + " ".join(launcher_args)
+
+# 1. Define a base image for your training job
+# must use torch 2.7.0 for vllm
+BASE_IMAGE = "axolotlai/axolotl:main-py3.11-cu126-2.7.1"
+
+# 2. Define the Runtime Environment for the Training Job
+# This includes start commands and environment variables.a
+# Secrets from the baseten workspace like API keys are referenced using
+# `SecretReference`.
+
+env_vars = {
+    "AXOLOTL_LAUNCHER": launcher,
+    "AXOLOTL_LAUNCHER_ARGS": launcher_args_str,
+}
+for secret_name in secrets:
+    env_vars[secret_name] = definitions.SecretReference(name=secret_name)
+
+training_runtime = definitions.Runtime(
+    start_commands=[  # Example: list of commands to run your training script
+        f"/bin/sh -c 'chmod +x ./{script_name} && ./{script_name}'"
+    ],
+    environment_variables=env_vars,
+)
+
+# 3. Define the Compute Resources for the Training Job
+training_compute = definitions.Compute(
+    node_count=node_count,
+    accelerator=truss_config.AcceleratorSpec(
+        accelerator=truss_config.Accelerator.H100,
+        count=gpu_count,
+    ),
+)
+
+# 4. Define the Training Job
+# This brings together the image, compute, and runtime configurations.
+my_training_job = definitions.TrainingJob(
+    image=definitions.Image(base_image=BASE_IMAGE),
+    compute=training_compute,
+    runtime=training_runtime,
+)
+
+
+# This config will be pushed using the Truss CLI.
+# The association of the job to the project happens at the time of push.
+first_project_with_job = definitions.TrainingProject(
+    name=project_name, job=my_training_job
+)