📝 Add docstrings to 202512-raise_on_drop

Docstrings generation was requested by @kallewoof.

* https://github.com/axolotl-ai-cloud/axolotl/pull/3321#issuecomment-3668489902

The following files were modified:

* `src/axolotl/utils/data/utils.py`
* `src/axolotl/utils/trainer.py`

.github/FUNDING.yml

@@ -1,13 +1,13 @@
 # These are supported funding model platforms
 
-github: [winglian, OpenAccess-AI-Collective] # Replace with up to 4 GitHub Sponsors-enabled usernames e.g., [user1, user2]
+github: # Replace with up to 4 GitHub Sponsors-enabled usernames e.g., [user1, user2]
 patreon: # Replace with a single Patreon username
 open_collective: # Replace with a single Open Collective username
-ko_fi: axolotl_ai # Replace with a single Ko-fi username
+ko_fi: # Replace with a single Ko-fi username
 tidelift: # Replace with a single Tidelift platform-name/package-name e.g., npm/babel
 community_bridge: # Replace with a single Community Bridge project-name e.g., cloud-foundry
 liberapay: # Replace with a single Liberapay username
 issuehunt: # Replace with a single IssueHunt username
 otechie: # Replace with a single Otechie username
 lfx_crowdfunding: # Replace with a single LFX Crowdfunding project-name e.g., cloud-foundry
-custom: ['https://quickchart.io/qr?text=bitcoin%3Abc1qxlgwlqwfea5s2cxm42xqsfmwjct0rj8w8ea5np&size=480&centerImageUrl=https%3A%2F%2Fupload.wikimedia.org%2Fwikipedia%2Fcommons%2Fthumb%2F4%2F46%2FBitcoin.svg%2F64px-Bitcoin.svg.png'] # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2']
+custom: # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2']

.github/workflows/base.yml

@@ -25,20 +25,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: "124"
-            cuda_version: 12.4.1
-            cudnn_version: ""
-            python_version: "3.11"
-            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-base"
-          - cuda: "126"
-            cuda_version: 12.6.3
-            cudnn_version: ""
-            python_version: "3.11"
-            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-base"
           - cuda: "126"
             cuda_version: 12.6.3
             cudnn_version: ""
@@ -67,6 +53,20 @@ jobs:
             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"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.9.1
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-base"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.9.1
+            torch_cuda_arch_list: "9.0+PTX"
+            dockerfile: "Dockerfile-base"
 #          - cuda: "128"
 #            cuda_version: 12.8.1
 #            cudnn_version: ""
@@ -90,7 +90,6 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl-base
             axolotlai/axolotl-base
       - name: Login to Docker Hub
         uses: docker/login-action@v2
@@ -122,13 +121,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: "126"
-            cuda_version: 12.6.3
-            cudnn_version: ""
-            python_version: "3.11"
-            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: ""
@@ -150,6 +142,20 @@ jobs:
             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"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.9.1
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.9.1
+            torch_cuda_arch_list: "9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.github/workflows/main.yml

@@ -15,11 +15,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.6.0
-            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -30,7 +25,6 @@ jobs:
             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"
@@ -41,6 +35,17 @@ jobs:
             python_version: "3.11"
             pytorch: 2.8.0
             axolotl_extras:
+            is_latest: true
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.0
+            axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -50,7 +55,6 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl
             axolotlai/axolotl
           tags: |
             type=ref,event=branch
@@ -88,11 +92,6 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.6.0
-            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -109,7 +108,6 @@ jobs:
             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"
@@ -120,6 +118,17 @@ jobs:
             python_version: "3.11"
             pytorch: 2.8.0
             axolotl_extras:
+            is_latest: true
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.0
+            axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -129,7 +138,6 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl-cloud
             axolotlai/axolotl-cloud
           tags: |
             type=ref,event=branch
@@ -162,11 +170,6 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.6.0
-            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -194,7 +197,6 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl-cloud-term
             axolotlai/axolotl-cloud-term
           tags: |
             type=ref,event=branch

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

@@ -26,13 +26,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.6.0
-            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: fbgemm-gpu
             num_gpus: 2
             nightly_build: "true"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.0
+            axolotl_extras: fbgemm-gpu
+            num_gpus: 2
+            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:

.github/workflows/nightlies.yml

@@ -12,16 +12,16 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            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:
+          - 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
@@ -31,7 +31,6 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl
             axolotlai/axolotl
           tags: |
             type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}
@@ -65,16 +64,16 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            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:
+          - 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
@@ -84,7 +83,6 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl-cloud
             axolotlai/axolotl-cloud
           tags: |
             type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}

.github/workflows/precommit-autoupdate.yml

@@ -2,7 +2,7 @@ name: Pre-commit auto-update
 
 on:
   schedule:
-    - cron: '0 0 * * 0'  # Run weekly
+    - cron: '0 0 1 * *'  # Run monthly
   workflow_dispatch:  # Manual kickoff
 
 jobs:

.github/workflows/tests-nightly.yml

@@ -26,7 +26,7 @@ jobs:
       max-parallel: 2
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.0"]
+        pytorch_version: ["2.7.1", "2.8.0"]
     timeout-minutes: 20
 
     steps:
@@ -102,14 +102,14 @@ 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:
             nightly_build: "true"
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.8.0
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"

.github/workflows/tests.yml

@@ -55,19 +55,23 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.1", "2.8.0"]
+        pytorch_version: ["2.7.1", "2.8.0", "2.9.0"]
     timeout-minutes: 20
 
     steps:
+      - name: cleanup node
+        run: |
+          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL
+
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore Cache from S3
-        id: hf-cache-restore-s3
-        run: |
-          mkdir -p /home/runner/.cache/huggingface/hub
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
-
+#      - name: Restore Cache from S3
+#        id: hf-cache-restore-s3
+#        run: |
+#          mkdir -p ~/.cache/huggingface/hub
+#          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd
+#
       - name: Setup Python
         uses: actions/setup-python@v5
         with:
@@ -81,16 +85,20 @@ jobs:
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }} torchvision
+          pip3 install --no-cache-dir torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Install dependencies
         run: |
           pip3 show torch
-          pip3 install --no-build-isolation -U -e .
+          pip3 install --no-cache-dir --no-build-isolation -U -e .
           python scripts/unsloth_install.py | sh
           python scripts/cutcrossentropy_install.py | sh
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 
+      - name: cleanup pip cache
+        run: |
+          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+
       - name: Make sure PyTorch version wasn't clobbered
         run: |
           python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
@@ -105,9 +113,13 @@ jobs:
 
       - name: Run tests
         run: |
+          df -h
           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
+          df -h
           pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
+          df -h
           pytest -v --durations=10 tests/patched/ --cov=axolotl --cov-append --cov-report=xml
+          df -h
           pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml
 
       - name: Upload coverage to Codecov
@@ -118,10 +130,6 @@ jobs:
           flags: unittests,pytorch-${{ matrix.pytorch_version }}
           fail_ci_if_error: false
 
-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
-
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
@@ -130,19 +138,23 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.1", "2.8.0"]
+        pytorch_version: ["2.7.1", "2.8.0", "2.9.0"]
     timeout-minutes: 20
 
     steps:
+      - name: cleanup node
+        run: |
+          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL
+
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore Cache from S3
-        id: hf-cache-restore-s3
-        run: |
-          mkdir -p /home/runner/.cache/huggingface/hub
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
-
+#      - name: Restore Cache from S3
+#        id: hf-cache-restore-s3
+#        run: |
+#          mkdir -p ~/.cache/huggingface/hub
+#          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd
+#
       - name: Setup Python
         uses: actions/setup-python@v5
         with:
@@ -152,21 +164,25 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 setuptools_scm build wheel
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 setuptools_scm build wheel psutil
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }} torchvision
+          pip3 install --no-cache-dir torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Install dependencies
         run: |
           pip3 show torch
           python -m build --no-isolation --sdist
-          pip3 install --no-build-isolation dist/axolotl*.tar.gz
+          pip3 install --no-cache-dir --no-build-isolation dist/axolotl*.tar.gz
           python scripts/unsloth_install.py | sh
           python scripts/cutcrossentropy_install.py | sh
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 
+      - name: cleanup pip cache
+        run: |
+          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+
       - name: Make sure PyTorch version wasn't clobbered
         run: |
           python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
@@ -176,7 +192,7 @@ jobs:
           axolotl --help
 
       - name: Show HF cache
-        run: huggingface-cli scan-cache
+        run: hf cache scan
 
       - name: Run tests
         run: |
@@ -184,10 +200,6 @@ jobs:
           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
@@ -231,16 +243,10 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
-            num_gpus: 1
-            axolotl_extras:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.8.0
             num_gpus: 1
             axolotl_extras:
             dockerfile: "Dockerfile-uv.jinja"
@@ -289,15 +295,15 @@ jobs:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.6.0
-            num_gpus: 1
-            axolotl_extras:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
             pytorch: 2.7.1
             num_gpus: 1
             axolotl_extras:
+#          - cuda: 128
+#            cuda_version: 12.8.1
+#            python_version: "3.11"
+#            pytorch: 2.7.1
+#            num_gpus: 1
+#            axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
@@ -305,6 +311,12 @@ jobs:
             num_gpus: 1
             gpu_type: "B200"
             axolotl_extras: fbgemm-gpu
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.0
+            num_gpus: 1
+            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.pre-commit-config.yaml

@@ -11,13 +11,13 @@ repos:
     -   id: no-commit-to-branch
         args: ['--branch', 'main']
 -   repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.12.12
+    rev: v0.14.7
     hooks:
     -   id: ruff
         args: [--fix]
     -   id: ruff-format
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.17.1
+    rev: v1.19.0
     hooks:
     - id: mypy
       additional_dependencies:
@@ -26,7 +26,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.8.6
+    rev: 1.9.2
     hooks:
     -   id: bandit
         args: [

.runpod/Dockerfile

@@ -10,6 +10,7 @@ ARG BASE_VOLUME="/runpod-volume"
 ENV BASE_VOLUME=$BASE_VOLUME
 ENV HF_DATASETS_CACHE="${BASE_VOLUME}/huggingface-cache/datasets"
 ENV HUGGINGFACE_HUB_CACHE="${BASE_VOLUME}/huggingface-cache/hub"
+ENV HF_HUB_CACHE="${BASE_VOLUME}/huggingface-cache/hub"
 ENV TRANSFORMERS_CACHE="${BASE_VOLUME}/huggingface-cache/hub"
 
 COPY .runpod/src /src

README.md

@@ -29,6 +29,10 @@
 
 ## 🎉 Latest Updates
 
+- 2025/12: Axolotl now includes support for [Olmo3](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/olmo3), [Trinity](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/trinity), and [Ministral3](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/ministral3).
+- 2025/10: New model support has been added in Axolotl for: [Qwen3 Next](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/qwen3-next), [Qwen2.5-vl, Qwen3-vl](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen2_5-vl), [Qwen3, Qwen3MoE](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen3), [Granite 4](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/granite4), [HunYuan](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/hunyuan), [Magistral 2509](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral#vision), [Apertus](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/apertus), and [Seed-OSS](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/seed-oss).
+- 2025/09: Axolotl now has text diffusion training. Read more [here](https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/diffusion).
+- 2025/08: QAT has been updated to include NVFP4 support. See [PR](https://github.com/axolotl-ai-cloud/axolotl/pull/3107).
 - 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).
@@ -36,12 +40,12 @@
   - [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!
-- 2025/03: 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>
 
+- 2025/03: 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.
 - 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!
@@ -73,7 +77,7 @@ Features:
 
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
-- PyTorch ≥2.6.0
+- PyTorch ≥2.7.1
 
 ### Google Colab
 
@@ -154,6 +158,13 @@ That's it! Check out our [Getting Started Guide](https://docs.axolotl.ai/docs/ge
 
 Contributions are welcome! Please see our [Contributing Guide](https://github.com/axolotl-ai-cloud/axolotl/blob/main/.github/CONTRIBUTING.md) for details.
 
+## 📈 Telemetry
+
+Axolotl has opt-out telemetry that helps us understand how the project is being used
+and prioritize improvements. We collect basic system information, model types, and
+error rates—never personal data or file paths. Telemetry is enabled by default. To
+disable it, set AXOLOTL_DO_NOT_TRACK=1. For more details, see our [telemetry documentation](https://docs.axolotl.ai/docs/telemetry.html).
+
 ## ❤️ Sponsors
 
 Interested in sponsoring? Contact us at [wing@axolotl.ai](mailto:wing@axolotl.ai)

_quarto.yml

@@ -241,6 +241,7 @@ website:
             - docs/installation.qmd
             - docs/inference.qmd
             - docs/cli.qmd
+            - docs/telemetry.qmd
             - docs/config-reference.qmd
             - text: "API Reference"
               href: docs/api

cicd/Dockerfile-uv.jinja

@@ -32,6 +32,7 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
     fi
 
 RUN uv pip install packaging==23.2 setuptools==75.8.0
+RUN uv pip install torchvision
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         uv pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

cicd/Dockerfile.jinja

@@ -1,6 +1,6 @@
 FROM axolotlai/axolotl-base:{{ BASE_TAG }}
 
-ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
+ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
 ENV AXOLOTL_EXTRAS="{{ AXOLOTL_EXTRAS }}"
 ENV AXOLOTL_ARGS="{{ AXOLOTL_ARGS }}"
 ENV CUDA="{{ CUDA }}"
@@ -9,7 +9,7 @@ ENV GITHUB_REF="{{ GITHUB_REF }}"
 ENV GITHUB_SHA="{{ GITHUB_SHA }}"
 ENV NIGHTLY_BUILD="{{ NIGHTLY_BUILD }}"
 ENV HF_HOME="{{ HF_HOME }}"
-ENV AXOLOTL_DATASET_PROCESSES="8"
+ENV AXOLOTL_DATASET_NUM_PROC="8"
 
 RUN apt-get update && \
     apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
@@ -32,7 +32,7 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
-RUN pip install packaging==23.2 setuptools==75.8.0
+RUN pip install packaging==23.2 setuptools==75.8.0 psutil
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

cicd/single_gpu.py

@@ -65,8 +65,13 @@ def run_cmd(cmd: str, run_folder: str):
     import subprocess  # nosec
 
     sp_env = os.environ.copy()
-    sp_env["AXOLOTL_DATASET_PROCESSES"] = "8"
+    sp_env["AXOLOTL_DATASET_NUM_PROC"] = "8"
 
     # Propagate errors from subprocess.
-    if exit_code := subprocess.call(cmd.split(), cwd=run_folder, env=sp_env):  # nosec
-        exit(exit_code)
+    try:
+        exit_code = subprocess.call(cmd.split(), cwd=run_folder, env=sp_env)  # nosec
+        if exit_code:
+            print(f"Command '{cmd}' failed with exit code {exit_code}")
+            return exit_code
+    except Exception as e:  # pylint: disable=broad-except
+        print(f"Command '{cmd}' failed with exception {e}")

devtools/dev_chat_template.yml

@@ -13,7 +13,7 @@ datasets:
 val_set_size: 0
 output_dir: temp_debug/axolotl_outputs/model
 dataset_prepared_path: temp_debug/axolotl_outputs/data
-dataset_processes: 1
+dataset_num_proc: 1
 
 sequence_len: 4096
 sample_packing: false

docker/Dockerfile-base

@@ -35,18 +35,24 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
 
 WORKDIR /workspace
 
-RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel psutil && \
     python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
-    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
 
+RUN if [ "$CUDA" != "130" ] ; then \
+        CAUSAL_CONV1D_FORCE_CXX11_ABI=TRUE CAUSAL_CONV1D_FORCE_BUILD=TRUE python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@v1.5.4"; \
+        python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"; \
+        python3 -m pip cache purge; \
+    fi
+
 RUN git lfs install --skip-repo && \
     pip3 install awscli && \
     # The base image ships with `pydantic==1.8.2` which is not working
     pip3 install -U --no-cache-dir pydantic==1.10.10 && \
     pip3 cache purge
 
-RUN if [ "$PYTORCH_VERSION" = "2.6.0" ] && [ "$CUDA" = "124" ] ; then \
-        FLASH_ATTENTION_FORCE_BUILD="TRUE" pip3 install --no-build-isolation flash-attn==2.8.0.post2; \
+RUN if [ "$PYTORCH_VERSION" = "2.9.1" ] && [ "$CUDA" = "128" ] ; then \
+        wget https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.4.17/flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
+        pip3 install --no-cache-dir flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
+        rm flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
     fi

docker/Dockerfile-uv-base

@@ -30,7 +30,13 @@ RUN uv venv --no-project --relocatable axolotl-venv
 ENV PATH="/workspace/axolotl-venv/bin:${PATH}"
 
 RUN uv pip install packaging setuptools wheel psutil \
-    && uv pip install torch==${PYTORCH_VERSION} \
+    && uv pip install torch==${PYTORCH_VERSION} torchvision \
     && uv pip install --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" \
     && uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" \
     && uv pip install awscli pydantic
+
+RUN if [ "$PYTORCH_VERSION" = "2.9.0" ] && [ "$CUDA" = "128" ] ; then \
+        wget https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.4.17/flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
+        uv pip install --no-cache-dir flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
+        rm flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
+    fi

docs/dataset-formats/conversation.qmd

@@ -218,6 +218,13 @@ If you have tool arguments with same name but different dtypes (like `"time": st
 ```
 "arguments": "{\"...\": \"...\"}"
 ```
+
+The same is applicable for tool parameters.
+
+```
+"parameters": "{\"...\": \"...\"}"
+```
+
 :::
 
 Example config for Llama4:

docs/debugging.qmd

@@ -29,7 +29,7 @@ While debugging it's helpful to simplify your test scenario as much as possible.
 1. **Make sure you are using the latest version of axolotl**:  This project changes often and bugs get fixed fast.  Check your git branch and make sure you have pulled the latest changes from `main`.
 1. **Eliminate concurrency**: Restrict the number of processes to 1 for both training and data preprocessing:
     - Set `CUDA_VISIBLE_DEVICES` to a single GPU, ex: `export CUDA_VISIBLE_DEVICES=0`.
-    - Set `dataset_processes: 1` in your axolotl config or run the training command with `--dataset_processes=1`.
+    - Set `dataset_num_proc: 1` in your axolotl config or run the training command with `--dataset_num_proc=1`.
 2. **Use a small dataset**: Construct or use a small dataset from HF Hub. When using a small dataset, you will often have to make sure `sample_packing: False` and `eval_sample_packing: False` to avoid errors.  If you are in a pinch and don't have time to construct a small dataset but want to use from the HF Hub, you can shard the data (this will still tokenize the entire dataset, but will only use a fraction of the data for training.  For example, to shard the dataset into 20 pieces, add the following to your axolotl config):
 
     ```yaml
@@ -101,7 +101,7 @@ For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 acceler
                 "-m", "axolotl.cli.train", "dev_chat_template.yml",
                 // The flags below simplify debugging by overriding the axolotl config
                 // with the debugging tips above.  Modify as needed.
-                "--dataset_processes=1",      // limits data preprocessing to one process
+                "--dataset_num_proc=1",      // limits data preprocessing to one process
                 "--max_steps=1",              // limits training to just one step
                 "--batch_size=1",             // minimizes batch size
                 "--micro_batch_size=1",       // minimizes batch size

docs/faq.qmd

@@ -63,6 +63,14 @@ description: Frequently asked questions
 
 > A: There seems to be a wheel issue with FA2 2.8.0 on CUDA 12.4. Try CUDA 12.6 instead or downgrade to FA2 2.7.4. Please refer to the upstream issue: https://github.com/Dao-AILab/flash-attention/issues/1717.
 
+**Q: Can we mix text and text+image datasets for VLM training?**
+
+> A: Yes, you can for newer VLM arch. The ones that would not work are LLaVA / Pixtral arch. If you notice one not working, please let us know!
+
+**Q: Why is `memory/max_*` different from `nvidia-smi`?**
+
+> A: We use `torch` APIs to retrieve this information. You can see https://docs.pytorch.org/docs/stable/notes/cuda.html#cuda-memory-management for more information.
+
 ### Chat templates
 
 **Q: `jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____`**

docs/fsdp_qlora.qmd

@@ -1,5 +1,5 @@
 ---
-title: "FDSP + QLoRA"
+title: "FSDP + QLoRA"
 description: Use FSDP with QLoRA to fine-tune large LLMs on consumer GPUs.
 format:
   html:
@@ -23,6 +23,12 @@ To enable `QLoRA` with `FSDP`, you need to perform the following steps:
 2. Enable FSDP in your axolotl config, as [described here](multi-gpu.qmd#sec-fsdp).
 3. Use one of the supported model types: `llama`, `mistral` or `mixtral`.
 
+## Enabling Swap for FSDP2
+
+If available memory is insufficient even after FSDP's CPU offloading, you can enable swap memory usage by setting `cpu_offload_pin_memory: false` alongside `offload_params: true` in FSDP config.
+
+This disables memory pinning, allowing FSDP to use disk swap space as fallback. Disabling memory pinning itself incurs performance overhead, and actually having to use swap adds more, but it may enable training larger models that would otherwise cause OOM errors on resource constrained systems.
+
 ## Example Config
 
 [examples/llama-2/qlora-fsdp.yml](../examples/llama-2/qlora-fsdp.yml) contains an example of how to enable QLoRA + FSDP in axolotl.

docs/lora_optims.qmd

@@ -5,10 +5,11 @@ description: "Custom autograd functions and Triton kernels in Axolotl for optimi
 
 Inspired by [Unsloth](https://github.com/unslothai/unsloth), we've implemented two
 optimizations for LoRA and QLoRA fine-tuning, supporting both single GPU and multi-GPU
-(in the DDP and DeepSpeed settings) training. These include (1) SwiGLU and GEGLU activation function
-Triton kernels, and (2) LoRA MLP and attention custom autograd functions. Our goal was
-to leverage operator fusion and tensor re-use in order to improve speed and reduce
-memory usage during the forward and backward passes of these calculations.
+(including the DDP, DeepSpeed, and FSDP2 settings) training. These include (1) SwiGLU
+and GEGLU activation function Triton kernels, and (2) LoRA MLP and attention custom
+autograd functions. Our goal was to leverage operator fusion and tensor re-use in order
+to improve speed and reduce memory usage during the forward and backward passes of
+these calculations.
 
 We currently support several common model architectures, including (but not limited to):
 
@@ -131,6 +132,5 @@ computation path.
 ## Future Work
 
 - Support for additional model architectures
-- Support for the FSDP setting
 - Support for dropout and bias
 - Additional operator fusions

docs/lr_groups.qmd

@@ -27,3 +27,9 @@ learning_rate: 2e-5
 In this example, we have a default learning rate of 2e-5 across the entire model, but we have a separate learning rate
 of 1e-6 for all the self attention `o_proj` modules across all layers, and a learning are of 1e-5 to the 3rd layer's
 self attention `q_proj` module.
+
+::: {.callout-note}
+
+We currently only support varying `lr` for now. If you're interested in adding support for others (`weight_decay`), we welcome PRs. See https://github.com/axolotl-ai-cloud/axolotl/blob/613bcf90e58f3ab81d3827e7fc572319908db9fb/src/axolotl/core/trainers/mixins/optimizer.py#L17
+
+:::

docs/multi-gpu.qmd

@@ -4,7 +4,7 @@ format:
   html:
     toc: true
     toc-depth: 3
-    number-sections: true
+    # number-sections: true
     code-tools: true
 execute:
   enabled: false
@@ -14,12 +14,18 @@ This guide covers advanced training configurations for multi-GPU setups using Ax
 
 ## Overview {#sec-overview}
 
-Axolotl supports several methods for multi-GPU training:
+When training on multiple GPUs, Axolotl supports 3 sharding/parallelism strategies. Additionally, you can layer specific optimization features on top of that strategy.
 
-- DeepSpeed (recommended)
-- FSDP (Fully Sharded Data Parallel)
-- Sequence parallelism
-- FSDP + QLoRA
+You generally cannot combine these strategies; they are mutually exclusive.
+
+1.  **DeepSpeed**: Powerful optimization library, supports ZeRO stages 1-3.
+2.  **FSDP (Fully Sharded Data Parallel)**: PyTorch's native sharding implementation (Recommended).
+3.  **DDP (Distributed Data Parallel)**: PyTorch's native parallelism implementation (Default if neither of the above are selected).
+
+These features can often be combined with the strategies above:
+
+*   **Sequence Parallelism**: Splits long sequences across GPUs (Compatible with DDP, DeepSpeed, and FSDP).
+*   **FSDP + QLoRA**: Combines 4-bit quantization with FSDP (Specific to FSDP).
 
 ## DeepSpeed {#sec-deepspeed}
 
@@ -65,12 +71,18 @@ Start from Stage 1 -> Stage 2 -> Stage 3.
 
 ## Fully Sharded Data Parallel (FSDP) {#sec-fsdp}
 
+FSDP allows you to shard model parameters, gradients, and optimizer states across data parallel workers.
+
 ::: {.callout-note}
 
 FSDP2 is recommended for new users. FSDP1 is deprecated and will be removed in an upcoming release of Axolotl.
 
 :::
 
+### FSDP + QLoRA {#sec-fsdp-qlora}
+
+For combining FSDP with QLoRA, see our [dedicated guide](fsdp_qlora.qmd).
+
 ### Migrating from FSDP1 to FSDP2 {#sec-migrate-fsdp1-fsdp2}
 
 To migrate your config from FSDP1 to FSDP2, you must use the `fsdp_version` top-level config field to specify the FSDP version, and
@@ -88,6 +100,7 @@ fsdp_sync_module_states | **REMOVED**
 fsdp_cpu_ram_efficient_loading | cpu_ram_efficient_loading
 fsdp_state_dict_type | state_dict_type
 fsdp_use_orig_params | **REMOVED**
+fsdp_activation_checkpointing | activation_checkpointing
 
 For more details, please see the migration guide in the [torchtitan repo](https://github.com/pytorch/torchtitan/blob/main/docs/fsdp.md). In Axolotl,
 if you were using the following FSDP1 config:
@@ -144,10 +157,6 @@ single sequence causes OOM errors during model training.
 
 See our [dedicated guide](sequence_parallelism.qmd) for more information.
 
-### FSDP + QLoRA {#sec-fsdp-qlora}
-
-For combining FSDP with QLoRA, see our [dedicated guide](fsdp_qlora.qmd).
-
 ## Performance Optimization {#sec-performance}
 
 ### Liger Kernel Integration {#sec-liger}

docs/multimodal.qmd

@@ -56,10 +56,14 @@ image_resize_algorithm: bilinear
 
 Please see [examples](https://github.com/axolotl-ai/axolotl/tree/main/examples) folder for full configs.
 
-::: {.callout-warning}
+::: {.callout-tip}
 Some of our chat_templates have been extended to support broader dataset types. This should not break any existing configs.
 :::
 
+::: {.callout-note}
+As of now, we do not truncate nor drop samples based on `sequence_len` as each arch has different ways to process non-text tokens. We are looking for help on this.
+:::
+
 ### Mllama {#sec-mllama}
 
 ```yaml
@@ -120,6 +124,8 @@ Please make sure to install audio lib via `pip3 install librosa==0.11.0 'mistral
 
 ```yaml
 base_model: mistralai/Voxtral-Mini-3B-2507
+
+processor_type: VoxtralProcessor
 ```
 
 ### Gemma-3 {#sec-gemma-3}
@@ -168,6 +174,14 @@ base_model: Qwen/Qwen2.5-VL-7B-Instruct
 chat_template: qwen2_vl  # same as qwen2-vl
 ```
 
+### Qwen3-VL {#sec-qwen3-vl}
+
+```yaml
+base_model: Qwen/Qwen3-VL-4B-Instruct
+
+chat_template: qwen2_vl  # same as qwen2-vl
+```
+
 ### SmolVLM2 {#sec-smolvlm2}
 
 ::: {.callout-tip}

docs/rlhf.qmd

@@ -219,6 +219,21 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
+#### chat_template.argilla_chat
+
+```json
+{
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
 #### chat_template.default
 
 ```yaml
@@ -582,6 +597,116 @@ To see other examples of custom reward functions, please see [TRL GRPO Docs](htt
 
 To see all configs, please see [TRLConfig](https://github.com/axolotl-ai-cloud/axolotl/blob/v0.9.2/src/axolotl/utils/schemas/trl.py).
 
+#### OpenEnv Rollout Functions
+
+GRPO supports custom rollout functions for OpenEnv-style environments, enabling interactive tasks like web browsing, code execution, or tool use. This allows you to implement custom generation logic that interacts with external environments.
+
+For example, to implement a simple math-solving environment with step-by-step verification:
+
+```python
+# math_env.py
+import re
+
+def math_solver_rollout(model, processing_class, prompts, generation_config=None):
+    """
+    Custom rollout function that generates step-by-step math solutions.
+
+    Args:
+        model: The language model
+        processing_class: The tokenizer/processing_class
+        prompts: List of prompt dicts (with 'messages' key for chat format)
+        generation_config: Optional generation configuration
+
+    Returns:
+        List of completion strings
+    """
+    completions = []
+
+    for prompt in prompts:
+        # Apply chat template to prompt
+        messages = prompt.get("messages", [])
+        formatted_prompt = processing_class.apply_chat_template(
+            messages, processing_class=False, add_generation_prompt=True
+        )
+
+        # Generate step-by-step solution
+        full_response = ""
+        for step in range(5):  # Max 5 reasoning steps
+            current_input = formatted_prompt + full_response + "\nNext step:"
+            inputs = processing_class(current_input, return_tensors="pt").to(model.device)
+
+            outputs = model.generate(
+                **inputs,
+                max_new_tokens=100,
+                generation_config=generation_config,
+            )
+            step_text = processing_class.decode(
+                outputs[0][inputs.input_ids.shape[1]:],
+                skip_special_tokens=True
+            )
+
+            # Check if solution is complete
+            if "FINAL ANSWER:" in step_text:
+                full_response += step_text
+                break
+            full_response += step_text + "\n"
+
+        completions.append(full_response)
+
+    return completions
+
+def math_reward(prompts, completions, answers, **kwargs):
+    """Reward function that checks mathematical correctness"""
+    rewards = []
+    for completion, correct_answer in zip(completions, answers):
+        # Extract predicted answer
+        match = re.search(r"FINAL ANSWER:\s*(.+)", completion)
+        predicted = match.group(1).strip() if match else ""
+
+        # Compare with correct answer
+        reward = 1.0 if predicted == str(correct_answer) else 0.0
+        rewards.append(reward)
+
+    return rewards
+
+def math_transform(cfg, *args, **kwargs):
+    """Transform dataset to GRPO format with answer field"""
+    def transform_fn(example, processing_class=None):
+        return {
+            "prompt": [{"role": "user", "content": example["question"]}],
+            "answer": str(example["answer"]),
+        }
+    return transform_fn, {"remove_columns": ["question"]}
+```
+
+```yaml
+rl: grpo
+
+trl:
+  beta: 0.001
+  max_completion_length: 512
+  num_generations: 4
+  rollout_func: "math_env.math_solver_rollout"  # Custom rollout function
+  reward_funcs: ["math_env.math_reward"]
+  reward_weights: [1.0]
+
+datasets:
+  - path: openai/gsm8k
+    name: main
+    type: math_env.math_transform
+```
+
+The `rollout_func` parameter accepts a fully qualified name (e.g., `module_name.function_name`) that points to a callable function in your local directory. The function receives:
+
+- `model`: The language model
+- `processing_class`: The tokenizer/processing class
+- `prompts`: List of prompt dictionaries
+- `generation_config` (optional): Generation configuration
+
+And should return a list of completion strings.
+
+For more OpenEnv examples, see [TRL OpenEnv Documentation](https://huggingface.co/docs/trl/main/en/openenv).
+
 #### GRPO with DAPO/Dr. GRPO loss
 
 The DAPO paper and subsequently Dr. GRPO paper proposed an alternative loss function for GRPO to remediate the penalty in longer responses.

docs/telemetry.qmd

@@ -0,0 +1,61 @@
+---
+title: Telemetry
+description: A description of the telemetry implementation in Axolotl.
+---
+
+# Telemetry in Axolotl
+
+Axolotl implements anonymous telemetry to help maintainers understand how the library
+is used and where users encounter issues. This data helps prioritize features, optimize
+performance, and fix bugs.
+
+## Data Collection
+
+We collect:
+
+- System info: OS, Python version, Axolotl version, PyTorch version, Transformers
+version, etc.
+- Hardware info: CPU count, memory, GPU count and models
+- Runtime metrics: Training progress, memory usage, timing information
+- Usage patterns: Models (from a whitelist) and configurations used
+- Error tracking: Stack traces and error messages (sanitized to remove personal
+information)
+
+Personally identifiable information (PII) is not collected.
+
+## Implementation
+
+Telemetry is implemented using PostHog and consists of:
+
+- `axolotl.telemetry.TelemetryManager`: A singleton class that initializes the
+telemetry system and provides methods for tracking events.
+- `axolotl.telemetry.errors.send_errors`: A decorator that captures exceptions and
+sends sanitized stack traces.
+- `axolotl.telemetry.runtime_metrics.RuntimeMetricsTracker`: A class that tracks
+runtime metrics during training.
+- `axolotl.telemetry.callbacks.TelemetryCallback`: A Trainer callback that sends
+runtime metrics telemetry.
+
+The telemetry system will block training startup for 10 seconds to ensure users are
+aware of data collection, unless telemetry is explicitly enabled or disabled.
+
+## Opt-Out Mechanism
+
+Telemetry is **enabled by default** on an opt-out basis. To disable it, set
+`AXOLOTL_DO_NOT_TRACK=1` or `DO_NOT_TRACK=1`.
+
+A warning message will be logged on start to clearly inform users about telemetry.
+We will remove this after some period.
+
+To hide the warning message about telemetry that is displayed on train, etc. startup,
+explicitly set: `AXOLOTL_DO_NOT_TRACK=0` (enable telemetry) or `AXOLOTL_DO_NOT_TRACK=1`
+(explicitly disable telemetry).
+
+## Privacy
+
+- All path-like config information is automatically redacted from telemetry data
+- Model information is only collected for whitelisted organizations
+    - See `axolotl/telemetry/whitelist.yaml` for the set of whitelisted organizations
+- Each run generates a unique anonymous ID
+    - This allows us to link different telemetry events in a single same training run
+- Telemetry is only sent from the main process to avoid duplicate events

examples/LiquidAI/README.md

@@ -6,6 +6,8 @@ LFM2 features a new hybrid Liquid architecture with multiplicative gates, short-
 
 This guide shows how to fine-tune both the LFM2 and LFM2-VL models with Axolotl.
 
+Thanks to the team at LiquidAI for giving us early access to prepare for these releases.
+
 ## Getting Started
 
 1.  Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
@@ -31,6 +33,14 @@ This guide shows how to fine-tune both the LFM2 and LFM2-VL models with Axolotl.
     axolotl train examples/LiquidAI/lfm2-vl-lora.yaml
     ```
 
+    **LFM2-MoE**
+    ```bash
+    pip install git+https://github.com/huggingface/transformers.git@0c9a72e4576fe4c84077f066e585129c97bfd4e6
+
+    # LoRA SFT (1x48GB @ 16.2GiB)
+    axolotl train examples/LiquidAI/lfm2-8b-a1b-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:
@@ -45,14 +55,13 @@ This guide shows how to fine-tune both the LFM2 and LFM2-VL models with Axolotl.
 
 ## 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)
+- [Optimizations Guide](https://docs.axolotl.ai/docs/optimizations.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)
+- [LFM2-MoE Blog](https://www.liquid.ai/blog/lfm2-8b-a1b-an-efficient-on-device-mixture-of-experts)
 - [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

@@ -1,6 +1,7 @@
 base_model: LiquidAI/LFM2-350M
 
-chunked_cross_entropy: true
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
 
 eot_tokens:
   - "<|im_end|>"

examples/LiquidAI/lfm2-8b-a1b-lora.yaml

@@ -0,0 +1,59 @@
+base_model: LiquidAI/LFM2-8B-A1B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: true
+
+eot_tokens:
+  - "<|im_end|>"
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: '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: 2
+micro_batch_size: 4
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-5
+
+bf16: true
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 2
+saves_per_epoch: 1
+
+weight_decay: 0.0
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/LiquidAI/lfm2-vl-lora.yaml

@@ -3,6 +3,9 @@ trust_remote_code: true
 model_type: AutoModelForImageTextToText
 processor_type: AutoProcessor
 
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
 # these 3 lines are needed for now to handle vision chat templates w images
 skip_prepare_dataset: true
 remove_unused_columns: false

examples/colab-notebooks/colab-axolotl-example.ipynb

@@ -40,7 +40,7 @@
     "%%capture\n",
     "# This step can take ~5-10 minutes to install dependencies\n",
     "!pip install --no-build-isolation axolotl[flash-attn]>=0.9.1\n",
-    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c5aa3ef\""
+    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@f643b88\""
    ]
   },
   {
@@ -253,7 +253,6 @@
    "source": [
     "from axolotl.utils import set_pytorch_cuda_alloc_conf\n",
     "\n",
-    "# Set \"PYTORCH_CUDA_ALLOC_CONF\" env to save memory\n",
     "set_pytorch_cuda_alloc_conf()"
    ]
   },

examples/gemma3/gemma-3-1b-qlora.yml

@@ -1,7 +1,7 @@
 base_model: google/gemma-3-1b-it
-# optionally might have model_type or tokenizer_type
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
+
+model_type: Gemma3ForCausalLM
+
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 

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

@@ -1,7 +1,7 @@
 base_model: google/gemma-3-270m-it
-# optionally might have model_type or tokenizer_type
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
+
+model_type: Gemma3ForCausalLM
+
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 

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

@@ -1,5 +1,8 @@
 base_model: google/gemma-3-4b-it
 
+# Need to set else transformers tries to load vision too
+model_type: Gemma3ForCausalLM
+
 load_in_4bit: true
 
 # gemma3 doesn't seem to play nice with ddp

examples/gpt-oss/README.md

@@ -2,6 +2,8 @@
 
 [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.
 
+In October 2025, OpenAI released safeguard models built upon GPT-OSS called [GPT-OSS-Safeguard](https://huggingface.co/collections/openai/gpt-oss-safeguard). They use the same architecture, so the same examples below can be re-used.
+
 This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
 
 ## Getting started
@@ -64,6 +66,16 @@ axolotl merge-sharded-fsdp-weights examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offlo
 mv ./outputs/gpt-oss-out/merged/* ./outputs/gpt-oss-out/
 ```
 
+### How to set reasoning_effort in template?
+
+The harmony template has a feature to set the `reasoning_effort` during prompt building. The default is `medium`. If you would like to adjust this, you can add the following to your config:
+
+```yaml
+chat_template_kwargs:
+  reasoning_effort: "high"  # low | medium | high
+```
+
+Currently, this applies globally. There is no method to apply per sample yet. If you are interested in adding this, please feel free to create an Issue to discuss.
 
 ### Inferencing your fine-tuned model
 

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

@@ -0,0 +1,67 @@
+base_model: openai/gpt-oss-safeguard-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-safeguard-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/granite4/README.md

@@ -0,0 +1,65 @@
+# Finetune IBM's Granite 4.0 with Axolotl
+
+[Granite 4.0](https://huggingface.co/collections/ibm-granite/granite-40-language-models) are a family of open source models trained by IBM Research.
+
+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 Granite4 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 2.7.1 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/granite4/granite-4.0-tiny-fft.yaml
+```
+
+This config uses about 40.8GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### 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).
+
+### Limitation
+
+Adapter finetuning does not work at the moment. It would error with
+
+```bash
+RuntimeError: mat1 and mat2 shapes cannot be multiplied (4096x3072 and 1x1179648)
+```
+
+In addition, if adapter training works, `lora_target_linear: true` will not work due to:
+```bash
+ValueError: Target module GraniteMoeHybridParallelExperts() is not supported.
+```
+
+## 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
+
+- [Granite Docs](https://www.ibm.com/granite/docs/models/granite)
+- [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/granite4/granite-4.0-tiny-fft.yaml

@@ -0,0 +1,45 @@
+base_model: ibm-granite/granite-4.0-tiny-preview
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/model-out
+
+sequence_len: 2048
+sample_packing: 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/llama-2/qlora-fsdp.yml

@@ -66,6 +66,7 @@ fsdp_config:
   fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
   fsdp_state_dict_type: FULL_STATE_DICT
+  # fsdp_cpu_offload_pin_memory: false  # uncomment to enable swap memory usage when RAM is insufficient
 special_tokens:
 
 # save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/llama-3/3b-fp8-fsdp2.yaml

@@ -29,7 +29,7 @@ flex_attention: true
 flex_attn_compile_kwargs:
   dynamic: false
   mode: max-autotune-no-cudagraphs
-
+save_strategy: no
 torch_compile: true
 
 wandb_project:

examples/llama-3/opentelemetry-qlora.yml

@@ -0,0 +1,50 @@
+base_model: NousResearch/Llama-3.2-1B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+load_in_4bit: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+
+output_dir: ./outputs/opentelemetry-example
+
+adapter: qlora
+sequence_len: 512
+sample_packing: false
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+# OpenTelemetry Configuration
+use_otel_metrics: true
+otel_metrics_host: "localhost"
+otel_metrics_port: 8000
+
+# Disable WandB
+use_wandb: false
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: false
+
+warmup_ratio: 0.1
+evals_per_epoch: 2
+saves_per_epoch: 1
+weight_decay: 0.0
+
+special_tokens:
+  pad_token: "<|end_of_text|>"

examples/magistral/README.md

@@ -13,7 +13,7 @@ Thanks to the team at MistralAI for giving us early access to prepare for these
     Here is an example of how to install from pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+# Ensure you have Pytorch installed (Pytorch 2.7.0 min)
 pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```

examples/magistral/think/README.md

@@ -12,7 +12,7 @@ Before starting, ensure you have:
 Run the thinking model fine-tuning:
 
 ```bash
-axolotl train magistral-small-think-qlora.yaml
+axolotl train examples/magistral/think/magistral-small-think-qlora.yaml
 ```
 
 This config uses about 19.1 GiB VRAM.

examples/magistral/vision/README.md

@@ -21,7 +21,7 @@ Before starting, ensure you have:
 
 3. Run the fine-tuning:
    ```bash
-   axolotl train magistral-small-vision-24B-qlora.yml
+   axolotl train examples/magistral/vision/magistral-small-vision-24B-qlora.yml
    ```
 
 This config uses about 17GiB VRAM.

examples/ministral/README.md

@@ -0,0 +1,50 @@
+# Finetune Ministral with Axolotl
+
+Ministral is a family of openweight models from MistralAI found on [HuggingFace](mistralai/Ministral-8B-Instruct-2410). 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).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+3. Run the finetuning example:
+
+    ```bash
+    axolotl train examples/ministral/ministral-small-qlora.yaml
+    ```
+
+This config uses about 8.76 GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### Tips
+
+- We recommend adding the same/similar SystemPrompt that the model is tuned for. You can find this within the repo's files titled `SYSTEM_PROMPT.txt`.
+- 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 text dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Limitations
+
+We only support the `mistral-common` tokenizer for Supervised Fine-tuning at the moment and for `type: chat_template` only.
+
+In addition, we do not support overriding tokens yet.
+
+## Related Resources
+
+- [MistralAI Ministral Blog](https://mistral.ai/news/ministraux)
+- [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)
+
+
+## Future Work
+
+- Add parity to Preference Tuning, RL, etc.
+- Add parity to other tokenizer configs like overriding tokens.

examples/ministral/ministral-small-qlora.yaml

@@ -0,0 +1,67 @@
+base_model: mistralai/Ministral-8B-Instruct-2410
+
+# Enable to use mistral-common tokenizer
+tokenizer_use_mistral_common: true
+
+# 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/ministral3/README.md

@@ -0,0 +1,79 @@
+# Finetune Ministral3 with Axolotl
+
+Ministral3 is a family of open-weight models from MistralAI found on [HuggingFace](https://huggingface.co/collections/mistralai/ministral-3). This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+Please see [Thinking](#thinking) and [Vision](#vision) for their respective fine-tuning.
+
+Thanks to the team at MistralAI for giving us early access to prepare for these releases.
+
+Note: This is still experimental given it is based on transformers v5 RC.
+
+## Getting started
+
+1. Install Axolotl from source following the [installation guide](https://docs.axolotl.ai/docs/installation.html#sec-edge-build).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+3. Swap to the Axolotl transformers v5 branch
+
+    ```bash
+    cp examples/ministral3/ministral3-3b-qlora.yaml ministral3-3b-qlora.yaml
+
+    git fetch
+    git checkout transformers-v5
+
+    # Install packages for transformers v5
+    pip install -e .
+    ```
+
+4. Run the fine-tuning:
+
+    ```bash
+    axolotl train ministral3-3b-qlora.yaml
+    ```
+
+Let us know how it goes. Happy finetuning! 🚀
+
+
+### Tips
+
+- We recommend adding the same/similar SystemPrompt that the model is tuned for. You can find this within the repo's files titled `SYSTEM_PROMPT.txt`.
+- 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 text dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+### Thinking
+
+Ministral3 2512 model supports thinking capabilities, enabling Chain-of-Thought reasoning with explicit thinking steps.
+
+📚 **[See the Thinking fine-tuning guide →](./think/README.md)**
+
+### Vision
+
+Ministral3 2512 model also supports vision capabilities.
+
+📚 **[See the Vision fine-tuning guide →](./vision/README.md)**
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Limitations
+
+We only support the `mistral-common` tokenizer for Supervised Fine-tuning at the moment and for `type: chat_template` only.
+
+In addition, we do not support overriding tokens yet.
+
+## Related Resources
+
+- [MistralAI Mistral3 Blog](https://mistral.ai/news/mistral-3)
+- [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)
+
+
+## Future Work
+
+- Add parity to Preference Tuning, RL, etc.
+- Add parity to other tokenizer configs like overriding tokens.

examples/ministral3/ministral3-3b-qlora.yaml

@@ -0,0 +1,67 @@
+base_model: mistralai/Ministral-3-3B-Reasoning-2512
+
+# Enable to use mistral-common tokenizer
+tokenizer_use_mistral_common: true
+
+# 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/ministral3/think/README.md

@@ -0,0 +1,73 @@
+# Ministral3 2512 Thinking Fine-tuning
+
+This guide covers fine-tuning [Ministral3 2512](https://huggingface.co/collections/mistralai/ministral-3) with thinking capabilities using Axolotl. The thinking model enables explicit Chain-of-Thought reasoning with separate thinking and response sections.
+
+## Prerequisites
+
+Before starting, ensure you have:
+- Installed Axolotl (see [main README](../README.md))
+
+## Getting Started
+
+Run the thinking model fine-tuning:
+
+```bash
+axolotl train examples/ministral3/think/ministral3-3b-think-qlora.yaml
+```
+
+This config uses about 4.76 GiB VRAM.
+
+### Tips
+
+- Dataset uses multi-content format with `type: thinking` support. See [Dataset Format](#dataset-format) below.
+- You cannot mix `content: str` and `content: list[dict]`, otherwise, dataset loading will fail. Keep it consistent.
+
+## Dataset Format
+
+The thinking model requires the multi-content dataset format with support for an extra `role: thinking` within system and assistant messages.
+
+Example format:
+
+```json
+{
+    "messages": [
+        {
+            "role": "system",
+            "content": [
+                { "type": "text", "text": "{SYSTEM_PROMPT}"}
+            ]
+        },
+        {
+            "role": "user",
+            "content": [
+                { "type": "text", "text": "Solve this step by step: What is 15% of 240?"}
+            ]
+        },
+        {
+            "role": "assistant",
+            "content": [
+                {
+                    "type": "thinking",
+                    "thinking": "I need to calculate 15% of 240. First, I'll convert 15% to decimal: 0.15. Then multiply: 0.15 × 240 = 36."
+                },
+                {
+                    "type": "text",
+                    "text": "To find 15% of 240, I'll multiply 240 by 0.15:\n\n240 × 0.15 = 36\n\nTherefore, 15% of 240 is 36."
+                }
+            ]
+        }
+    ]
+}
+```
+
+### Advanced Options
+
+The `thinking` section supports an optional `closed` parameter:
+
+```json
+{
+    "type": "thinking",
+    "thinking": "Internal reasoning here...",
+    "closed": true  // Default: true, controls adding the closing [/THINK] tag
+}
+```

examples/ministral3/think/ministral3-3b-think-qlora.yaml

@@ -0,0 +1,67 @@
+base_model: mistralai/Ministral-3-3B-Reasoning-2512
+
+# Enable to use mistral-common tokenizer
+tokenizer_use_mistral_common: true
+
+# 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: Nanobit/text-think-2k-test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+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/ministral3/vision/README.md

@@ -0,0 +1,57 @@
+# Ministral3 2512 Vision Fine-tuning
+
+This guide covers fine-tuning [Ministral3 2512](https://huggingface.co/collections/mistralai/ministral-3) with vision capabilities using Axolotl.
+
+## Prerequisites
+
+Before starting, ensure you have:
+- Installed Axolotl from source (see [main README](../README.md#getting-started))
+
+## Getting started
+
+1. Install the required vision lib:
+    ```bash
+    pip install 'mistral-common[opencv]==1.8.6'
+    ```
+
+2. Download the example dataset image:
+   ```bash
+   wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+   ```
+
+3. Run the fine-tuning:
+   ```bash
+   axolotl train examples/ministral3/vision/ministral3-3b-vision-qlora.yml
+   ```
+
+WARNING: The loss and grad norm will be much higher than normal at first. We suspect this to be inherent to the model as of the moment. If anyone would like to submit a fix for this, we are happy to take a look.
+
+### Tips
+
+Key differences from text-only model:
+- Multi-modal dataset format required
+- Sample packing not supported
+
+## Dataset Format
+
+The vision model requires multi-modal dataset format as documented [here](https://docs.axolotl.ai/docs/multimodal.html#dataset-format).
+
+One exception is that, passing `"image": PIL.Image` is not supported. MistralTokenizer only supports `path`, `url`, and `base64` for now.
+
+Example:
+```json
+{
+    "messages": [
+        {"role": "system", "content": [{ "type": "text", "text": "{SYSTEM_PROMPT}"}]},
+        {"role": "user", "content": [
+            { "type": "text", "text": "What's in this image?"},
+            {"type": "image", "path": "path/to/image.jpg" }
+        ]},
+        {"role": "assistant", "content": [{ "type": "text", "text": "..." }]},
+    ],
+}
+```
+
+## Limitations
+
+- Sample Packing is not supported for multi-modality training currently.

examples/ministral3/vision/ministral3-3b-vision-qlora.yml

@@ -0,0 +1,64 @@
+base_model: mistralai/Ministral-3-3B-Reasoning-2512
+processor_type: AutoProcessor
+
+# Enable to use mistral-common tokenizer
+tokenizer_use_mistral_common: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+# sample dataset below requires downloading image in advance
+# wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+datasets:
+  - path: Nanobit/text-vision-2k-test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/mistral/mistral-small/README.md

@@ -0,0 +1,51 @@
+# Mistral Small 3.1/3.2 Fine-tuning
+
+This guide covers fine-tuning [Mistral Small 3.1](mistralai/Mistral-Small-3.1-24B-Instruct-2503) and [Mistral Small 3.2](mistralai/Mistral-Small-3.2-24B-Instruct-2506) with vision capabilities using Axolotl.
+
+## Prerequisites
+
+Before starting, ensure you have:
+- Installed Axolotl (see [Installation docs](https://docs.axolotl.ai/docs/installation.html))
+
+## Getting Started
+
+1. Install the required vision lib:
+    ```bash
+    pip install 'mistral-common[opencv]==1.8.5'
+    ```
+
+2. Download the example dataset image:
+   ```bash
+   wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+   ```
+
+3. Run the fine-tuning:
+   ```bash
+   axolotl train examples/mistral/mistral-small/mistral-small-3.1-24B-lora.yml
+   ```
+
+This config uses about 29.4 GiB VRAM.
+
+## Dataset Format
+
+The vision model requires multi-modal dataset format as documented [here](https://docs.axolotl.ai/docs/multimodal.html#dataset-format).
+
+One exception is that, passing `"image": PIL.Image` is not supported. MistralTokenizer only supports `path`, `url`, and `base64` for now.
+
+Example:
+```json
+{
+    "messages": [
+        {"role": "system", "content": [{ "type": "text", "text": "{SYSTEM_PROMPT}"}]},
+        {"role": "user", "content": [
+            { "type": "text", "text": "What's in this image?"},
+            {"type": "image", "path": "path/to/image.jpg" }
+        ]},
+        {"role": "assistant", "content": [{ "type": "text", "text": "..." }]},
+    ],
+}
+```
+
+## Limitations
+
+- Sample Packing is not supported for multi-modality training currently.

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

@@ -39,7 +39,7 @@ wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 1
-micro_batch_size: 1
+micro_batch_size: 2
 num_epochs: 1
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine

examples/olmo3/README.md

@@ -0,0 +1,38 @@
+# Finetune Allenai's Olmo 3 with Axolotl
+
+[Olmo 3](https://huggingface.co/collections/allenai/olmo-3) are a family of 7B and 32B models open source models trained by The Allen Institute for Artificial Intelligence.
+
+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).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+3. Run the finetuning example:
+
+    ```bash
+    axolotl train examples/olmo3/olmo3-7b-qlora.yaml
+    ```
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### TIPS
+
+- The example config can be re-used for Olmo and Olmo 2.
+- 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
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Related Resources
+
+- [Olmo 3 Blog](https://allenai.org/blog/olmo3)
+- [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/olmo3/olmo3-7b-qlora.yaml

@@ -0,0 +1,64 @@
+base_model: allenai/Olmo-3-7B-Instruct-SFT
+
+# 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/qat_nvfp4/Gemma3-12B_baseline.yml

@@ -0,0 +1,67 @@
+base_model: google/gemma-3-12b-it
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/out_gemma/
+
+sequence_len: 8096
+sample_packing: true
+flash_attention: true
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 4e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Gemma3-12B_qat.yml

@@ -0,0 +1,72 @@
+base_model: google/gemma-3-12b-it
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/qat_out_gemma/
+
+sequence_len: 8096
+sample_packing: true
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 4e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Gemma3-12B_baseline.yml

@@ -0,0 +1,67 @@
+base_model: google/gemma-3-12b-it
+# Math finetuning configuration for Gemma3-12B
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/out_math_gemma/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 8
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 3e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Gemma3-12B_qat.yml

@@ -0,0 +1,72 @@
+base_model: google/gemma-3-12b-it
+# Math finetuning configuration for Gemma3-12B
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/qat_out_math_gemma/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 8
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 3e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Gemma3-27B_baseline.yml

@@ -0,0 +1,68 @@
+base_model: google/gemma-3-27b-it
+# Math finetuning configuration for Gemma3-27B
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/out_math_gemma27/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-6
+eta_min: 7e-7
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Gemma3-27B_qat.yml

@@ -0,0 +1,73 @@
+base_model: google/gemma-3-27b-it
+# Math finetuning configuration for Gemma3-27B
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/qat_out_math_gemma27/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-6
+eta_min: 7e-7
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Qwen2.5-72B_baseline.yml

@@ -0,0 +1,67 @@
+base_model: Qwen/Qwen2.5-72B
+# Math finetuning configuration for Qwen2.5-72B (non-instruct)
+# 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
+seed: 42
+chat_template: qwen_25
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/out_math_72b/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 8
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-6
+eta_min: 7e-7
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Qwen2.5-72B_qat.yml

@@ -0,0 +1,72 @@
+base_model: Qwen/Qwen2.5-72B
+# Math finetuning configuration for Qwen2.5-72B (non-instruct)
+# 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
+seed: 42
+chat_template: qwen_25
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/qat_out_math_72b/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 8
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-6
+eta_min: 7e-7
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Qwen2.5-72B_baseline.yml

@@ -0,0 +1,67 @@
+base_model: Qwen/Qwen2.5-72B
+# Alpaca finetuning configuration for Qwen2.5-72B
+# 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
+seed: 42
+chat_template: qwen_25
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/out_qwen72b/
+
+sequence_len: 8096
+sample_packing: true
+flash_attention: true
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Qwen2.5-72B_qat.yml

@@ -0,0 +1,72 @@
+base_model: Qwen/Qwen2.5-72B
+# Alpaca finetuning configuration for Qwen2.5-72B
+# 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
+seed: 42
+chat_template: qwen_25
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/qat_out_qwen72b/
+
+sequence_len: 8096
+sample_packing: true
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qwen3/README.md

@@ -0,0 +1,46 @@
+# Finetune Qwen3 with Axolotl
+
+[Qwen3](https://huggingface.co/collections/Qwen/qwen3) are a family of open source models trained by Alibaba.
+
+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).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+3. Run the finetuning example:
+
+    ```bash
+    axolotl train examples/qwen3/32b-qlora.yaml
+    ```
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### Chat template masking a few tokens off
+
+If you notice that the `chat_template` masking for assistant prompts are off by a few tokens, please ensure that you are adding the below to the yaml.
+
+```yaml
+chat_template: qwen3
+```
+
+### TIPS
+
+- For inference, please check the official model card as it depends on your reasoning mode.
+- 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
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Related Resources
+
+- [Qwen3 Blog](https://qwenlm.github.io/blog/qwen3/)
+- [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/README.md

@@ -6,21 +6,17 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 
 ## 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).
+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 a compatible version of Pytorch installed
+    pip3 install packaging setuptools wheel ninja
+    pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 
-```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
-```
+    # Install Cut Cross Entropy
+    python scripts/cutcrossentropy_install.py | sh
+    ```
 
 2. Run the finetuning example:
 
@@ -41,9 +37,7 @@ Let us know how it goes. Happy finetuning! 🚀
 
 ## 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)
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
 
 ## Related Resources
 

examples/smolvlm2/README.md

@@ -37,9 +37,7 @@ This guide shows how to fine-tune SmolVLM2 models with Axolotl.
 
 ## 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)
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
 
 ## Related Resources
 

examples/trinity/README.md

@@ -0,0 +1,38 @@
+# Finetune ArceeAI's Trinity with Axolotl
+
+[Trinity](https://huggingface.co/collections/arcee-ai/trinity) is a family of open weight MoE models trained by Arcee.ai.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the main from the [installation guide](https://docs.axolotl.ai/docs/installation.html#sec-edge-build).
+
+2. Run the finetuning example:
+
+    ```bash
+    axolotl train examples/trinity/trinity-nano-preview-qlora.yaml
+    ```
+
+This config uses about 24.9 GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### TIPS
+
+- For inference, the official Arcee.ai team recommends `top_p: 0.75`, `temperature: 0.15`, `top_k: 50`, and `min_p: 0.06`.
+- 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
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Related Resources
+
+- [Trinity Blog](https://www.arcee.ai/blog/the-trinity-manifesto)
+- [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/trinity/trinity-nano-preview-qlora.yaml

@@ -0,0 +1,67 @@
+base_model: arcee-ai/Trinity-Nano-Preview
+trust_remote_code: true
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# CCE - N/A as of now
+# 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  # Not supported
+sdp_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/voxtral/voxtral-mini-audio-qlora.yml

@@ -1,5 +1,5 @@
 base_model: mistralai/Voxtral-Mini-3B-2507
-processor_type: AutoProcessor
+processor_type: VoxtralProcessor
 
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name

requirements.txt

@@ -1,35 +1,34 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 
 # START section of dependencies that don't install on Darwin/MacOS
-bitsandbytes==0.47.0
+bitsandbytes==0.48.2
 triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
-autoawq==0.2.7.post3
-liger-kernel==0.6.1
+liger-kernel==0.6.4
 # END section
 
 packaging==23.2
 
-huggingface_hub>=0.33.0
-peft>=0.17.0
-transformers==4.56.1
-tokenizers>=0.21.1
-accelerate==1.10.1
-datasets==4.0.0
+huggingface_hub>=0.36.0
+peft>=0.18.0
+tokenizers>=0.22.1
+transformers==4.57.1
+accelerate==1.11.0
+datasets==4.4.1
 deepspeed>=0.17.0
-trl==0.23.0
-hf_xet==1.1.5
-kernels==0.9.0
+trl==0.25.0
+hf_xet==1.2.0
+kernels>=0.9.0
 trackio
 
 optimum==1.16.2
 hf_transfer
 sentencepiece
-gradio==5.41.1
+gradio==5.49.1
 
 modal==1.0.2
-pydantic==2.10.6
+pydantic>=2.10.6
 addict
 fire
 PyYAML>=6.0
@@ -37,13 +36,12 @@ requests
 wandb
 einops
 colorama
-numba
-numpy>=1.24.4,<=2.0.1
+numba>=0.61.2
+numpy>=2.2.6
 
 # qlora things
 evaluate==0.4.1
 scipy
-scikit-learn==1.4.2
 nvidia-ml-py==12.560.30
 art
 tensorboard
@@ -51,7 +49,7 @@ python-dotenv==1.0.1
 
 # remote filesystems
 s3fs>=2024.5.0
-gcsfs>=2024.5.0
+gcsfs>=2025.3.0
 adlfs>=2024.5.0
 ocifs==1.3.2
 
@@ -65,9 +63,13 @@ immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
 
 torchao==0.13.0
+openenv-core==0.1.0
 schedulefree==1.4.1
 
-axolotl-contribs-lgpl==0.0.6
+axolotl-contribs-lgpl==0.0.7
 axolotl-contribs-mit==0.0.5
 
-mistral-common==1.8.5
+# telemetry
+posthog==6.7.11
+
+mistral-common==1.8.6

scripts/__init__.py

@@ -1 +0,0 @@
-"""Utility scripts package."""

scripts/benchmarks/__init__.py

@@ -1,5 +0,0 @@
-"""Benchmark helpers."""
-
-from .deepseek_v3_moe import ACCURACY_TOLERANCE, DTYPE_MAP, benchmark_deepseek_v3
-
-__all__ = ["benchmark_deepseek_v3", "DTYPE_MAP", "ACCURACY_TOLERANCE"]

scripts/benchmarks/build_deepseek_v3_8b.py

@@ -1,100 +0,0 @@
-#!/usr/bin/env python3
-"""Instantiate a ~8.3B DeepSeek-V3 MoE model with random weights.
-
-Run this on a GPU-equipped machine (e.g. 1× NVL H100) so the dense
-initialization completes quickly:
-
-    python scripts/benchmarks/build_deepseek_v3_8b.py --output deepseek-v3-8b-moe
-"""
-
-from __future__ import annotations
-
-import argparse
-from pathlib import Path
-
-import torch
-from transformers import DeepseekV3Config, DeepseekV3ForCausalLM
-
-DTYPE_MAP = {
-    "float32": torch.float32,
-    "bfloat16": torch.bfloat16,
-    "float16": torch.float16,
-}
-
-
-def build_config() -> DeepseekV3Config:
-    """Return a DeepSeek V3 configuration totaling ~8.3B parameters."""
-
-    return DeepseekV3Config(
-        vocab_size=32_000,
-        hidden_size=3_072,
-        intermediate_size=8_192,
-        moe_intermediate_size=2_560,
-        num_hidden_layers=20,
-        num_attention_heads=24,
-        num_key_value_heads=24,
-        n_routed_experts=18,
-        num_experts_per_tok=4,
-        n_group=6,
-        topk_group=4,
-        kv_lora_rank=192,
-        q_lora_rank=384,
-        max_position_embeddings=2_048,
-        rope_theta=10_000.0,
-        rope_interleave=True,
-        hidden_act="silu",
-        initializer_range=0.02,
-        attention_dropout=0.0,
-        attention_bias=False,
-        n_shared_experts=1,
-        routed_scaling_factor=2.5,
-        norm_topk_prob=True,
-    )
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument(
-        "--output",
-        type=Path,
-        required=True,
-        help="Directory to save the generated model",
-    )
-    parser.add_argument(
-        "--dtype",
-        default="bfloat16",
-        choices=DTYPE_MAP.keys(),
-        help="Storage dtype for the checkpoint",
-    )
-    parser.add_argument(
-        "--seed",
-        type=int,
-        default=0,
-        help="Torch RNG seed for reproducibility",
-    )
-    return parser.parse_args()
-
-
-def main() -> None:
-    args = parse_args()
-    torch.manual_seed(args.seed)
-
-    output_dir = args.output
-    output_dir.mkdir(parents=True, exist_ok=True)
-
-    config = build_config()
-    model = DeepseekV3ForCausalLM(config)
-
-    dtype = DTYPE_MAP[args.dtype]
-    model.to(dtype=dtype)
-
-    param_count = sum(p.numel() for p in model.parameters())
-    print(f"Initialized DeepSeek-V3 MoE with {param_count / 1e9:.3f}B parameters")
-
-    model.save_pretrained(output_dir, safe_serialization=True)
-    config.save_pretrained(output_dir)
-    print(f"Saved model and config to {output_dir.resolve()}")
-
-
-if __name__ == "__main__":
-    main()

scripts/benchmarks/deepseek_v3_group_gemm_table.py

@@ -1,190 +0,0 @@
-#!/usr/bin/env python
-"""Reproduce TorchTitan CG GEMM timings for selected problem sizes."""
-
-from __future__ import annotations
-
-import argparse
-import sys
-import time
-from dataclasses import dataclass
-from pathlib import Path
-from typing import Iterable
-
-import torch
-
-CURRENT_DIR = Path(__file__).resolve().parent
-for candidate in [CURRENT_DIR, *CURRENT_DIR.parents]:
-    repo_root = candidate / "axolotl"
-    if repo_root.exists():
-        if str(repo_root) not in sys.path:
-            sys.path.insert(0, str(repo_root))
-        break
-else:
-    raise SystemExit("Unable to locate axolotl repository root for imports")
-
-from axolotl.kernels.moe import (
-    cg_grouped_gemm_forward,
-    cg_grouped_gemm_forward_dynamic,
-)
-
-
-@dataclass
-class Scenario:
-    num_groups: int
-    m: int
-    n: int
-    k: int
-
-
-SCENARIOS: tuple[Scenario, ...] = (
-    Scenario(num_groups=4, m=8192, n=4096, k=7168),
-    Scenario(num_groups=4, m=8192, n=7168, k=2048),
-    Scenario(num_groups=8, m=4096, n=4096, k=7168),
-    Scenario(num_groups=8, m=4096, n=7168, k=2048),
-)
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument(
-        "--device", default="cuda", choices=["cuda"], help="Execution device"
-    )
-    parser.add_argument(
-        "--dtype",
-        default="bf16",
-        choices=["bf16", "fp16", "fp32"],
-        help="Computation dtype",
-    )
-    parser.add_argument("--warmup", type=int, default=5, help="Warmup iterations")
-    parser.add_argument("--iters", type=int, default=20, help="Benchmark iterations")
-    parser.add_argument("--seed", type=int, default=0, help="Random seed")
-    parser.add_argument(
-        "--group-size",
-        type=int,
-        default=128,
-        help="GROUP_SIZE_M expected by the kernel",
-    )
-    return parser.parse_args()
-
-
-def pick_dtype(name: str) -> torch.dtype:
-    return {
-        "bf16": torch.bfloat16,
-        "fp16": torch.float16,
-        "fp32": torch.float32,
-    }[name]
-
-
-def make_indices(
-    num_groups: int, group_size: int, device: torch.device
-) -> torch.Tensor:
-    indices = torch.arange(num_groups, device=device, dtype=torch.int32)
-    return indices.repeat_interleave(group_size)
-
-
-def timed_call(fn, *args, warmup: int, iters: int) -> float:
-    for _ in range(warmup):
-        fn(*args)
-    torch.cuda.synchronize()
-    start = time.perf_counter()
-    for _ in range(iters):
-        fn(*args)
-    torch.cuda.synchronize()
-    return (time.perf_counter() - start) * 1000.0 / iters
-
-
-def run_scenario(
-    scenario: Scenario,
-    *,
-    dtype: torch.dtype,
-    device: torch.device,
-    warmup: int,
-    iters: int,
-    group_size_m: int,
-) -> dict:
-    if scenario.m % scenario.num_groups != 0:
-        raise ValueError(
-            f"M ({scenario.m}) not divisible by groups ({scenario.num_groups})"
-        )
-    group_size = scenario.m // scenario.num_groups
-    if group_size % group_size_m != 0:
-        raise ValueError(
-            f"Group size {group_size} must be a multiple of GROUP_SIZE_M ({group_size_m}) for the Triton kernel"
-        )
-
-    inputs = torch.randn(scenario.m, scenario.k, device=device, dtype=dtype)
-    weights = torch.randn(
-        scenario.num_groups, scenario.n, scenario.k, device=device, dtype=dtype
-    )
-    indices = make_indices(scenario.num_groups, group_size, device)
-
-    def persistent():
-        return cg_grouped_gemm_forward(inputs, weights, indices, group_size_m)
-
-    def baseline():
-        return cg_grouped_gemm_forward_dynamic(inputs, weights, indices, group_size_m)
-
-    persistent_ms = timed_call(persistent, warmup=warmup, iters=iters)
-    baseline_ms = timed_call(baseline, warmup=warmup, iters=iters)
-
-    return {
-        "scenario": scenario,
-        "persistent_ms": persistent_ms,
-        "baseline_ms": baseline_ms,
-        "speedup": baseline_ms / persistent_ms if persistent_ms > 0 else float("nan"),
-    }
-
-
-def main() -> None:  # pragma: no cover - utility script
-    args = parse_args()
-    torch.manual_seed(args.seed)
-
-    if args.device != "cuda" or not torch.cuda.is_available():
-        raise SystemExit("CUDA device required for this benchmark")
-
-    dtype = pick_dtype(args.dtype)
-    device = torch.device(args.device)
-
-    print(
-        f"device={device} dtype={dtype} warmup={args.warmup} iters={args.iters} group_size={args.group_size}"
-    )
-    print(
-        f"{'groups':>7} {'m':>7} {'n':>7} {'k':>7} {'persistent':>12} {'baseline':>12} {'speedup':>8}"
-    )
-    for result in run_all(
-        SCENARIOS,
-        dtype=dtype,
-        device=device,
-        warmup=args.warmup,
-        iters=args.iters,
-        group_size_m=args.group_size,
-    ):
-        scen = result["scenario"]
-        print(
-            f"{scen.num_groups:>7} {scen.m:>7} {scen.n:>7} {scen.k:>7}"
-            f" {result['persistent_ms']:>11.3f} ms {result['baseline_ms']:>11.3f} ms {result['speedup']:>7.2f}x"
-        )
-
-
-def run_all(
-    scenarios: Iterable[Scenario],
-    *,
-    dtype: torch.dtype,
-    device: torch.device,
-    warmup: int,
-    iters: int,
-    group_size_m: int,
-) -> Iterable[dict]:
-    for scenario in scenarios:
-        yield run_scenario(
-            scenario,
-            dtype=dtype,
-            device=device,
-            warmup=warmup,
-            iters=iters,
-            group_size_m=group_size_m,
-        )
-
-
-if __name__ == "__main__":
-    main()

scripts/benchmarks/deepseek_v3_moe.py

@@ -1,301 +0,0 @@
-#!/usr/bin/env python
-# mypy: ignore-errors
-"""Microbenchmark for DeepSeek V3 MoE block comparing baseline vs Triton CG kernels."""
-
-from __future__ import annotations
-
-import argparse
-import sys
-import time
-from pathlib import Path
-from types import MethodType
-
-import torch
-
-try:
-    from transformers.models.deepseek_v3.configuration_deepseek_v3 import (
-        DeepseekV3Config,
-    )
-    from transformers.models.deepseek_v3.modeling_deepseek_v3 import DeepseekV3MoE
-except ImportError as exc:  # pragma: no cover - utility script
-    raise SystemExit(
-        "Transformers with DeepSeek-V3 support must be available in PYTHONPATH"
-    ) from exc
-
-CURRENT_DIR = Path(__file__).resolve().parent
-for candidate in [CURRENT_DIR, *CURRENT_DIR.parents]:
-    repo_root = candidate / "axolotl"
-    if repo_root.exists():
-        if str(repo_root) not in sys.path:
-            sys.path.insert(0, str(repo_root))
-        break
-else:  # pragma: no cover - defensive guard
-    raise SystemExit("Unable to locate axolotl repository root for imports")
-
-from axolotl.monkeypatch.deepseek_v3 import patch_deepseek_v3_moe  # noqa: E402
-
-ACCURACY_TOLERANCE = 5e-3
-
-DTYPE_MAP = {
-    "bf16": torch.bfloat16,
-    "fp16": torch.float16,
-    "fp32": torch.float32,
-}
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument("--batch", type=int, default=8, help="batch size")
-    parser.add_argument("--seq-len", type=int, default=2048, help="sequence length")
-    parser.add_argument("--hidden-size", type=int, default=4096, help="MoE hidden size")
-    parser.add_argument(
-        "--moe-intermediate-size",
-        type=int,
-        default=8192,
-        help="MoE intermediate projection size",
-    )
-    parser.add_argument(
-        "--n-experts",
-        type=int,
-        default=256,
-        help="Number of routed experts",
-    )
-    parser.add_argument(
-        "--top-k",
-        type=int,
-        default=8,
-        help="Number of experts per token",
-    )
-    parser.add_argument(
-        "--groups",
-        type=int,
-        default=8,
-        help="Router groups (must divide n-experts)",
-    )
-    parser.add_argument(
-        "--dtype",
-        choices=DTYPE_MAP.keys(),
-        default="bf16",
-        help="Computation dtype",
-    )
-    parser.add_argument(
-        "--device",
-        default="auto",
-        choices=["auto", "cpu", "cuda"],
-        help="Execution device",
-    )
-    parser.add_argument("--warmup", type=int, default=5, help="Warmup iterations")
-    parser.add_argument("--iters", type=int, default=25, help="Benchmark iterations")
-    parser.add_argument("--seed", type=int, default=0, help="Random seed")
-    parser.add_argument(
-        "--uniform-routing",
-        action="store_true",
-        help="Override router to distribute tokens evenly across experts",
-    )
-    parser.add_argument(
-        "--group-size",
-        type=int,
-        default=128,
-        help="GROUP_SIZE_M used by the Triton kernel",
-    )
-    parser.add_argument(
-        "--backend",
-        choices=["cg", "mg"],
-        default="mg",
-        help="MoE kernel backend to benchmark",
-    )
-    return parser.parse_args()
-
-
-def resolve_device(requested: str) -> torch.device:
-    if requested == "auto":
-        return torch.device("cuda" if torch.cuda.is_available() else "cpu")
-    return torch.device(requested)
-
-
-def build_module(args: argparse.Namespace) -> DeepseekV3MoE:
-    config = DeepseekV3Config(
-        hidden_size=args.hidden_size,
-        intermediate_size=args.moe_intermediate_size,
-        moe_intermediate_size=args.moe_intermediate_size,
-        n_routed_experts=args.n_experts,
-        num_experts_per_tok=args.top_k,
-        n_group=args.groups,
-        topk_group=max(1, min(args.groups, args.top_k)),
-        n_shared_experts=1,
-    )
-    module = DeepseekV3MoE(config)
-    module.eval()
-    return module
-
-
-@torch.no_grad()
-def timed_forward(
-    module: DeepseekV3MoE, inputs: torch.Tensor, iters: int, warmup: int
-) -> float:
-    for _ in range(warmup):
-        module(inputs)
-    if inputs.is_cuda:
-        torch.cuda.synchronize()
-    start = time.perf_counter()
-    for _ in range(iters):
-        module(inputs)
-    if inputs.is_cuda:
-        torch.cuda.synchronize()
-    elapsed = time.perf_counter() - start
-    return (elapsed / iters) * 1000.0
-
-
-def benchmark_deepseek_v3(args: argparse.Namespace) -> dict:
-    torch.manual_seed(args.seed)
-
-    device = resolve_device(args.device)
-    dtype = DTYPE_MAP[args.dtype]
-
-    if args.n_experts % args.groups != 0:
-        raise SystemExit("n-experts must be divisible by groups")
-    if args.top_k > args.n_experts:
-        raise SystemExit("top-k cannot exceed number of experts")
-
-    if device.type == "cuda" and not torch.cuda.is_available():
-        raise SystemExit("CUDA requested but not available")
-
-    baseline_module = build_module(args)
-    original_moe = getattr(
-        DeepseekV3MoE,
-        "_axolotl_triton_original_moe",
-        DeepseekV3MoE.moe,
-    )
-    baseline_module.moe = MethodType(original_moe, baseline_module)
-    state_dict = baseline_module.state_dict()
-
-    patch_deepseek_v3_moe(group_size_m=args.group_size, backend=args.backend)
-    patched_module = build_module(args)
-    patched_module.load_state_dict(state_dict)
-
-    baseline_module.to(device=device, dtype=dtype)
-    patched_module.to(device=device, dtype=dtype)
-
-    tokens = args.batch * args.seq_len
-    routed_tokens = tokens * args.top_k
-    avg_tokens_per_expert = routed_tokens / args.n_experts
-
-    inputs = torch.randn(
-        args.batch,
-        args.seq_len,
-        args.hidden_size,
-        device=device,
-        dtype=dtype,
-    )
-
-    with torch.no_grad():
-        flat_inputs = inputs.view(-1, args.hidden_size)
-        if args.uniform_routing:
-            total_assignments = flat_inputs.size(0) * args.top_k
-            base = total_assignments // args.n_experts
-            remainder = total_assignments % args.n_experts
-            counts = torch.full(
-                (args.n_experts,),
-                base,
-                dtype=torch.int64,
-                device=device,
-            )
-            if remainder:
-                counts[:remainder] += 1
-            assignments = torch.repeat_interleave(
-                torch.arange(args.n_experts, device=device), counts
-            )
-            assignments = assignments[torch.randperm(assignments.size(0))]
-            topk_idx = assignments.view(flat_inputs.size(0), args.top_k)
-        else:
-            topk_idx, _ = patched_module.gate(flat_inputs)
-
-        tokens_per_expert = torch.bincount(
-            topk_idx.reshape(-1), minlength=args.n_experts
-        )
-        min_tokens = int(tokens_per_expert.min().item())
-        max_tokens = int(tokens_per_expert.max().item())
-
-    if args.uniform_routing:
-        weights = torch.full(
-            topk_idx.shape,
-            1.0 / args.top_k,
-            device=device,
-            dtype=torch.float32,
-        )
-
-        def _uniform_gate(self, hidden_states):
-            flat = hidden_states.view(-1, hidden_states.shape[-1])
-            token_count = flat.shape[0]
-            return topk_idx[:token_count], weights[:token_count]
-
-        patched_module.gate.forward = _uniform_gate.__get__(
-            patched_module.gate, patched_module.gate.__class__
-        )
-        baseline_module.gate.forward = _uniform_gate.__get__(
-            baseline_module.gate, baseline_module.gate.__class__
-        )
-
-    with torch.no_grad():
-        ref_output = baseline_module(inputs)
-        patched_output = patched_module(inputs)
-        max_diff = (ref_output - patched_output).abs().max().item()
-
-    baseline_vram = patched_vram = None
-    if device.type == "cuda":
-        torch.cuda.reset_peak_memory_stats(device)
-    baseline_ms = timed_forward(baseline_module, inputs, args.iters, args.warmup)
-    if device.type == "cuda":
-        baseline_vram = torch.cuda.max_memory_allocated(device)
-        torch.cuda.reset_peak_memory_stats(device)
-    patched_ms = timed_forward(patched_module, inputs, args.iters, args.warmup)
-    if device.type == "cuda":
-        patched_vram = torch.cuda.max_memory_allocated(device)
-
-    speedup = baseline_ms / patched_ms if patched_ms > 0 else float("nan")
-
-    return {
-        "device": device,
-        "backend": args.backend,
-        "dtype": dtype,
-        "baseline_ms": baseline_ms,
-        "patched_ms": patched_ms,
-        "speedup": speedup,
-        "max_diff": max_diff,
-        "routed_tokens": routed_tokens,
-        "avg_tokens": avg_tokens_per_expert,
-        "min_tokens": min_tokens,
-        "max_tokens": max_tokens,
-        "baseline_vram": baseline_vram,
-        "patched_vram": patched_vram,
-        "accuracy_ok": max_diff <= ACCURACY_TOLERANCE,
-    }
-
-
-def main() -> None:  # pragma: no cover - CLI entrypoint
-    args = parse_args()
-    result = benchmark_deepseek_v3(args)
-
-    print(
-        f"Device={result['device'].type} dtype={result['dtype']} backend={result['backend']} batch={args.batch} seq={args.seq_len} hidden={args.hidden_size}"
-    )
-    print(
-        f"routed tokens={result['routed_tokens']} avg tokens/expert={result['avg_tokens']:.1f} group_size={args.group_size}"
-    )
-    print(f"min/max tokens per expert: {result['min_tokens']}/{result['max_tokens']}")
-    if result["baseline_vram"] is not None:
-        print(
-            f"VRAM baseline={result['baseline_vram'] / (1024**2):.1f} MiB | patched={result['patched_vram'] / (1024**2):.1f} MiB"
-        )
-    print(
-        f"Baseline: {result['baseline_ms']:.3f} ms | Patched: {result['patched_ms']:.3f} ms | x{result['speedup']:.2f}"
-    )
-    print(f"Max |Δ| between outputs: {result['max_diff']:.2e}")
-    if not result["accuracy_ok"]:
-        raise RuntimeError(
-            f"Accuracy check failed: max diff {result['max_diff']:.3e} exceeds tolerance {ACCURACY_TOLERANCE:.1e}"
-        )
-
-
-if __name__ == "__main__":
-    main()

scripts/benchmarks/deepseek_v3_moe_sweep.py

@@ -1,275 +0,0 @@
-#!/usr/bin/env python
-# mypy: ignore-errors
-"""Sweep a set of DeepSeek V3 MoE benchmark configurations."""
-
-from __future__ import annotations
-
-import argparse
-import csv
-import logging
-import sys
-from pathlib import Path
-from types import SimpleNamespace
-
-CURRENT_DIR = Path(__file__).resolve().parent
-for candidate in [CURRENT_DIR, *CURRENT_DIR.parents]:
-    repo_root = candidate / "axolotl"
-    if repo_root.exists():
-        if str(repo_root) not in sys.path:
-            sys.path.insert(0, str(repo_root))
-        break
-else:  # pragma: no cover - defensive guard
-    raise SystemExit("Unable to locate axolotl repository root for imports")
-
-from scripts.benchmarks.deepseek_v3_moe import (  # noqa: E402
-    ACCURACY_TOLERANCE,
-    DTYPE_MAP,
-    benchmark_deepseek_v3,
-)
-
-LOG = logging.getLogger(__name__)
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument(
-        "--dtype",
-        choices=DTYPE_MAP.keys(),
-        default="bf16",
-        help="Computation dtype for all benchmarks",
-    )
-    parser.add_argument(
-        "--device",
-        default="auto",
-        choices=["auto", "cpu", "cuda"],
-        help="Execution device",
-    )
-    parser.add_argument("--warmup", type=int, default=3, help="Warmup iterations")
-    parser.add_argument("--iters", type=int, default=15, help="Benchmark iterations")
-    parser.add_argument("--seed", type=int, default=0, help="Random seed")
-    parser.add_argument(
-        "--group-size",
-        type=int,
-        help="Override GROUP_SIZE_M for every configuration",
-    )
-    parser.add_argument(
-        "--backends",
-        default="mg",
-        help="Comma separated list of backends to benchmark (subset of cg,mg)",
-    )
-    parser.add_argument(
-        "--no-uniform-routing",
-        action="store_true",
-        help="Disable uniform routing for every configuration",
-    )
-    parser.add_argument(
-        "--include-mixtral-long",
-        action="store_true",
-        help="Add an 8×8192 Mixtral-style run to the sweep",
-    )
-    parser.add_argument(
-        "--output",
-        type=Path,
-        help="Optional CSV file to store results",
-    )
-    return parser.parse_args()
-
-
-def make_namespace(
-    base: dict, args: argparse.Namespace, backend: str
-) -> SimpleNamespace:
-    combined = dict(base)
-    combined.update(
-        {
-            "dtype": args.dtype,
-            "device": args.device,
-            "backend": backend,
-            "warmup": args.warmup,
-            "iters": args.iters,
-            "seed": args.seed,
-            "uniform_routing": not args.no_uniform_routing,
-        }
-    )
-    if args.group_size is not None:
-        combined["group_size"] = args.group_size
-    return SimpleNamespace(**combined)
-
-
-ARCHETYPES = (
-    (
-        "mixtral",
-        {
-            "hidden_size": 4096,
-            "moe_intermediate_size": 14336,
-            "n_experts": 8,
-            "top_k": 2,
-            "groups": 1,
-            "group_size": 128,
-        },
-        [(4, 2048), (8, 4096)],
-    ),
-    (
-        "qwen",
-        {
-            "hidden_size": 6144,
-            "moe_intermediate_size": 24576,
-            "n_experts": 16,
-            "top_k": 4,
-            "groups": 8,
-            "group_size": 128,
-        },
-        [(4, 4096), (8, 8192)],
-    ),
-    (
-        "deepseek_v3",
-        {
-            "hidden_size": 12288,
-            "moe_intermediate_size": 49152,
-            "n_experts": 128,
-            "top_k": 8,
-            "groups": 16,
-            "group_size": 128,
-        },
-        [(4, 4096), (8, 8192)],
-    ),
-)
-
-MIXTRAL_LONG_SHAPES = [(8, 8192)]
-
-
-def main() -> None:  # pragma: no cover - utility script
-    args = parse_args()
-
-    grid = []
-    for label, base_cfg, shapes in ARCHETYPES:
-        for batch, seq_len in shapes:
-            cfg = {
-                "label": label,
-                "batch": batch,
-                "seq_len": seq_len,
-                **base_cfg,
-            }
-            if cfg["n_experts"] % cfg["groups"] != 0 or cfg["top_k"] > cfg["n_experts"]:
-                continue
-            grid.append(cfg)
-
-    if args.include_mixtral_long:
-        base_cfg = ARCHETYPES[0][1]
-        for batch, seq_len in MIXTRAL_LONG_SHAPES:
-            grid.append(
-                {
-                    "label": "mixtral_long",
-                    "batch": batch,
-                    "seq_len": seq_len,
-                    **base_cfg,
-                }
-            )
-
-    if not grid:
-        raise SystemExit("No valid parameter combinations produced")
-
-    header = (
-        "model",
-        "batch",
-        "seq_len",
-        "hidden_size",
-        "moe_intermediate",
-        "n_experts",
-        "top_k",
-        "groups",
-        "backend",
-        "baseline_ms",
-        "patched_ms",
-        "speedup",
-        "baseline_vram_mib",
-        "patched_vram_mib",
-        "min_tokens",
-        "max_tokens",
-        "max_diff",
-        "accuracy_ok",
-    )
-    rows = []
-
-    raw_backends = [
-        token.strip() for token in args.backends.split(",") if token.strip()
-    ]
-    if not raw_backends:
-        raw_backends = ["mg"]
-    valid_backends = []
-    for backend in raw_backends:
-        if backend not in {"cg", "mg"}:
-            raise SystemExit(f"Unsupported backend '{backend}' requested")
-        if backend not in valid_backends:
-            valid_backends.append(backend)
-
-    uniform_flag = not args.no_uniform_routing
-    print(
-        f"Running sweep on device={args.device} dtype={args.dtype} backends={tuple(valid_backends)} uniform_routing={uniform_flag}"
-    )
-    print(
-        f"{'model':>10} {'batch':>5} {'seq':>5} {'hidden':>7} {'experts':>7} {'topk':>4} {'groups':>6} {'backend':>8}"
-        f" {'baseline':>12} {'patched':>12} {'speedup':>8} {'b_vram':>8} {'p_vram':>8} {'diff':>10} {'acc':>5}"
-    )
-
-    for cfg in grid:
-        for backend in valid_backends:
-            ns = make_namespace(cfg, args, backend)
-            result = benchmark_deepseek_v3(ns)
-            baseline_vram_mib = (
-                result["baseline_vram"] / (1024**2)
-                if result["baseline_vram"] is not None
-                else float("nan")
-            )
-            patched_vram_mib = (
-                result["patched_vram"] / (1024**2)
-                if result["patched_vram"] is not None
-                else float("nan")
-            )
-            rows.append(
-                (
-                    cfg["label"],
-                    cfg["batch"],
-                    cfg["seq_len"],
-                    cfg["hidden_size"],
-                    cfg["moe_intermediate_size"],
-                    cfg["n_experts"],
-                    cfg["top_k"],
-                    cfg["groups"],
-                    backend,
-                    result["baseline_ms"],
-                    result["patched_ms"],
-                    result["speedup"],
-                    baseline_vram_mib,
-                    patched_vram_mib,
-                    result["min_tokens"],
-                    result["max_tokens"],
-                    result["max_diff"],
-                    result["accuracy_ok"],
-                )
-            )
-            status = "OK" if result["accuracy_ok"] else "FAIL"
-            print(
-                f"{cfg['label']:>10} {cfg['batch']:>5} {cfg['seq_len']:>5} {cfg['hidden_size']:>7} {cfg['n_experts']:>7} {cfg['top_k']:>4} {cfg['groups']:>6} {backend:>8}"
-                f" {result['baseline_ms']:>11.3f} ms {result['patched_ms']:>11.3f} ms {result['speedup']:>7.2f}x"
-                f" {baseline_vram_mib:>8.1f} {patched_vram_mib:>8.1f} {result['max_diff']:>10.3e} {status:>5}"
-            )
-            if not result["accuracy_ok"]:
-                LOG.warning(
-                    "Accuracy tolerance exceeded for %s backend=%s: diff=%.3e (> %.1e)",
-                    cfg["label"],
-                    backend,
-                    result["max_diff"],
-                    ACCURACY_TOLERANCE,
-                )
-
-    if args.output:
-        args.output.parent.mkdir(parents=True, exist_ok=True)
-        with args.output.open("w", newline="") as fp:
-            writer = csv.writer(fp)
-            writer.writerow(header)
-            writer.writerows(rows)
-        print(f"Results written to {args.output}")
-
-
-if __name__ == "__main__":
-    main()

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@c5aa3ef"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@f643b88"'
 )

setup.py

@@ -26,7 +26,6 @@ def parse_requirements(extras_require_map):
                 _install_requires.append(line)
     try:
         xformers_version = [req for req in _install_requires if "xformers" in req][0]
-        autoawq_version = [req for req in _install_requires if "autoawq" in req][0]
         if "Darwin" in platform.system():
             # skip packages not compatible with OSX
             skip_packages = [
@@ -34,7 +33,6 @@ def parse_requirements(extras_require_map):
                 "triton",
                 "mamba-ssm",
                 "xformers",
-                "autoawq",
                 "liger-kernel",
             ]
             _install_requires = [
@@ -51,7 +49,7 @@ def parse_requirements(extras_require_map):
             try:
                 torch_version = version("torch")
             except PackageNotFoundError:
-                torch_version = "2.6.0"  # default to torch 2.6
+                torch_version = "2.8.0"  # default to torch 2.8.0
             _install_requires.append(f"torch=={torch_version}")
 
             version_match = re.match(r"^(\d+)\.(\d+)(?:\.(\d+))?", torch_version)
@@ -64,8 +62,14 @@ def parse_requirements(extras_require_map):
             else:
                 raise ValueError("Invalid version format")
 
-            if (major, minor) >= (2, 8):
-                pass
+            if (major, minor) >= (2, 9):
+                extras_require_map.pop("fbgemm-gpu")
+                extras_require_map["fbgemm-gpu"] = ["fbgemm-gpu-genai==1.4.1"]
+                extras_require_map["vllm"] = ["vllm==0.11.1"]
+            elif (major, minor) >= (2, 8):
+                extras_require_map.pop("fbgemm-gpu")
+                extras_require_map["fbgemm-gpu"] = ["fbgemm-gpu-genai==1.3.0"]
+                extras_require_map["vllm"] = ["vllm==0.11.0"]
             elif (major, minor) >= (2, 7):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
@@ -74,7 +78,7 @@ def parse_requirements(extras_require_map):
                     extras_require_map.pop("vllm")
                 else:
                     _install_requires.append("xformers==0.0.31")
-                    extras_require_map["vllm"] = ["vllm>=0.10.0"]
+                    extras_require_map["vllm"] = ["vllm==0.10.1"]
             elif (major, minor) >= (2, 6):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 _install_requires.append("xformers==0.0.29.post3")
@@ -87,7 +91,6 @@ def parse_requirements(extras_require_map):
                     _install_requires.append("xformers==0.0.28.post2")
                 else:
                     _install_requires.append("xformers>=0.0.28.post3")
-                _install_requires.pop(_install_requires.index(autoawq_version))
                 extras_require_map.pop("vllm")
             elif (major, minor) >= (2, 4):
                 extras_require_map.pop("vllm")
@@ -126,7 +129,7 @@ extras_require = {
         "ring-flash-attn>=0.1.7",
     ],
     "deepspeed": [
-        "deepspeed==0.17.5",
+        "deepspeed==0.18.2",
         "deepspeed-kernels",
     ],
     "mamba-ssm": [
@@ -161,7 +164,13 @@ extras_require = {
     "llmcompressor": [
         "llmcompressor==0.5.1",
     ],
-    "fbgemm-gpu": ["fbgemm-gpu-genai>=1.2.0"],
+    "fbgemm-gpu": ["fbgemm-gpu-genai==1.3.0"],
+    "opentelemetry": [
+        "opentelemetry-api",
+        "opentelemetry-sdk",
+        "opentelemetry-exporter-prometheus",
+        "prometheus-client",
+    ],
 }
 install_requires, dependency_links, extras_require_build = parse_requirements(
     extras_require

src/axolotl/cli/config.py

@@ -14,6 +14,8 @@ import yaml
 from transformers.utils import is_torch_bf16_gpu_available
 
 from axolotl.integrations.base import PluginManager
+from axolotl.telemetry.errors import send_errors
+from axolotl.telemetry.manager import TelemetryManager
 from axolotl.utils.comet_ import setup_comet_env_vars
 from axolotl.utils.config import (
     normalize_cfg_datasets,
@@ -31,6 +33,8 @@ LOG = get_logger(__name__)
 
 API_KEY_FIELDS = {"comet_api_key"}
 
+TELEMETRY_MANAGER = TelemetryManager.get_instance()
+
 
 def check_remote_config(config: Union[str, Path]) -> Union[str, Path]:
     """
@@ -164,6 +168,7 @@ def plugin_set_cfg(cfg: DictDefault):
         plugin_manager.cfg = cfg
 
 
+@send_errors
 def load_cfg(
     config: str | Path | DictDefault = Path("examples/"), **kwargs
 ) -> DictDefault:
@@ -197,6 +202,8 @@ def load_cfg(
             temp_file.close()
         cfg.axolotl_config_path = temp_file.name
 
+    TELEMETRY_MANAGER.send_event(event_type="config-loaded", properties=cfg)
+
     # If there are any options passed in the cli, if it is something that seems valid
     # from the yaml, then overwrite the value
     cfg_keys = cfg.keys()
@@ -240,6 +247,7 @@ def load_cfg(
     setup_comet_env_vars(cfg)
     plugin_set_cfg(cfg)
 
+    TELEMETRY_MANAGER.send_event(event_type="config-processed", properties=cfg)
     cfg_to_log = {
         k: "[REDACTED]" if k in API_KEY_FIELDS else v
         for k, v in cfg.items()

src/axolotl/cli/delinearize_llama4.py

@@ -85,9 +85,7 @@ def do_cli(model: Union[Path, str], output: Union[Path, str]) -> None:
     unpatch_llama4 = patch_llama4_linearized_modeling()
     from transformers import Llama4ForConditionalGeneration
 
-    model_ = Llama4ForConditionalGeneration.from_pretrained(
-        model, torch_dtype=torch.bfloat16
-    )
+    model_ = Llama4ForConditionalGeneration.from_pretrained(model, dtype=torch.bfloat16)
     processor = AutoProcessor.from_pretrained(model)
     processor.save_pretrained(output)
 

src/axolotl/cli/inference.py

@@ -19,7 +19,10 @@ from axolotl.cli.utils.diffusion import (
     launch_diffusion_gradio_ui,
 )
 from axolotl.integrations.base import PluginManager
-from axolotl.utils.chat_templates import get_chat_template_from_config
+from axolotl.telemetry.errors import send_errors
+from axolotl.utils.chat_templates import (
+    get_chat_template_from_config,
+)
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
 
@@ -43,6 +46,7 @@ def get_multi_line_input() -> str:
     return instruction
 
 
+@send_errors
 def do_inference(
     *,
     cfg: DictDefault,
@@ -160,6 +164,7 @@ def do_inference(
         print(tokenizer.decode(generated["sequences"].cpu().tolist()[0]))
 
 
+@send_errors
 def do_inference_gradio(
     *,
     cfg: DictDefault,

src/axolotl/cli/main.py

@@ -26,7 +26,7 @@ from axolotl.cli.utils import (
     launch_training,
 )
 from axolotl.integrations.lm_eval.cli import lm_eval
-from axolotl.utils import set_pytorch_cuda_alloc_conf
+from axolotl.utils import set_misc_env, set_pytorch_cuda_alloc_conf
 from axolotl.utils.logging import get_logger
 from axolotl.utils.schemas.config import AxolotlInputConfig
 
@@ -45,6 +45,7 @@ def cli():
     print_axolotl_text_art()
     load_dotenv()
     set_pytorch_cuda_alloc_conf()
+    set_misc_env()
 
 
 @cli.command()

src/axolotl/cli/merge_lora.py

@@ -7,12 +7,14 @@ import fire
 
 from axolotl.cli.config import load_cfg
 from axolotl.cli.utils import load_model_and_tokenizer
+from axolotl.telemetry.errors import send_errors
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
 
 LOG = get_logger(__name__)
 
 
+@send_errors
 def do_merge_lora(*, cfg: DictDefault) -> None:
     """
     Calls `transformers`' `merge_and_unload` on the model given in the `axolotl` config

src/axolotl/cli/merge_sharded_fsdp_weights.py

@@ -23,6 +23,7 @@ from safetensors.torch import save_file as safe_save_file
 from torch.distributed.checkpoint.format_utils import _EmptyStateDictLoadPlanner
 
 from axolotl.cli.config import load_cfg
+from axolotl.telemetry.errors import send_errors
 from axolotl.utils.logging import get_logger
 from axolotl.utils.train import determine_last_checkpoint
 
@@ -118,6 +119,7 @@ def _distributed_checkpoint_to_merged_weights(
     return save_path_
 
 
+@send_errors
 def merge_fsdp_weights(
     checkpoint_dir: str,
     output_path: str,

src/axolotl/cli/preprocess.py

@@ -17,6 +17,7 @@ from axolotl.cli.config import load_cfg
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.common.datasets import load_datasets, load_preference_datasets
 from axolotl.integrations.base import PluginManager
+from axolotl.telemetry.errors import send_errors
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
 from axolotl.utils.trainer import disable_datasets_caching
@@ -24,6 +25,7 @@ from axolotl.utils.trainer import disable_datasets_caching
 LOG = get_logger(__name__)
 
 
+@send_errors
 def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:
     """
     Preprocesses dataset specified in axolotl config.

src/axolotl/cli/quantize.py

@@ -8,7 +8,7 @@ from typing import Union
 from transformers import AutoConfig, AutoModelForCausalLM, TorchAoConfig
 
 from axolotl.cli.config import load_cfg
-from axolotl.loaders import load_tokenizer
+from axolotl.loaders import load_processor, load_tokenizer
 from axolotl.utils.logging import get_logger
 from axolotl.utils.quantization import (
     TorchAOQuantDType,
@@ -66,10 +66,15 @@ def do_quantize(
 
     LOG.info(f"Loading model from {model_path}.")
     tokenizer = load_tokenizer(cfg)
+
+    processor = None
+    if cfg.is_multimodal:
+        processor = load_processor(cfg, tokenizer)
+
     config = AutoConfig.from_pretrained(model_path)
     torch_dtype = config.torch_dtype if hasattr(config, "torch_dtype") else None
     model = AutoModelForCausalLM.from_pretrained(
-        model_path, device_map="auto", torch_dtype=torch_dtype
+        model_path, device_map="auto", dtype=torch_dtype
     )
 
     LOG.info(
@@ -107,6 +112,10 @@ def do_quantize(
         save_jinja_files=cfg.tokenizer_save_jinja_files,
     )
 
+    if processor:
+        LOG.info(f"Saving processor to: {str(Path(output_dir) / 'quantized')}.")
+        processor.save_pretrained(str(Path(output_dir) / "quantized"))
+
     if hub_model_id:
         hub_model_id = (
             hub_model_id.rstrip("-")
@@ -114,6 +123,8 @@ def do_quantize(
         )
         model.push_to_hub(hub_model_id, safe_serialization=False)
         tokenizer.push_to_hub(hub_model_id)
+        if processor:
+            processor.push_to_hub(hub_model_id)
         LOG.info(f"Quantized model pushed to: {hub_model_id}.")
 
     LOG.info(f"Quantized model saved to: {str(Path(output_dir) / 'quantized')}.")

src/axolotl/cli/train.py

@@ -99,7 +99,7 @@ def ray_train_func(kwargs: dict):
     resolve_dtype(cfg)
 
     # ray serializing objects gets rid of frozen attribute - HF expects dict not DefaultDict
-    if cfg.deepspeed:
+    if cfg.deepspeed and hasattr(cfg.deepspeed, "to_dict"):
         cfg.deepspeed = cfg.deepspeed.to_dict()
 
     # initialize accelerator before model instantiation

src/axolotl/common/architectures.py

@@ -12,6 +12,10 @@ MOE_ARCH_BLOCK = {
     "mixtral": "MixtralSparseMoeBlock",
     "qwen2_moe": "Qwen2MoeSparseMoeBlock",
     "qwen3_moe": "Qwen3MoeSparseMoeBlock",
+    "qwen3_vl_moe": "Qwen3VLMoeTextSparseMoeBlock",
     "deepseek_v2": "DeepseekV2MoE",
+    "deepseek_v3": "DeepseekV3MoE",
     "gpt_oss": "GptOssDecoderLayer",
+    "lfm2_moe": "Lfm2MoeSparseMoeBlock",
+    "afmoe": "AfmoeMoE",
 }

src/axolotl/common/datasets.py

@@ -9,6 +9,7 @@ from datasets import Dataset
 import axolotl.monkeypatch.data.batch_dataset_fetcher  # noqa: F401
 from axolotl.cli.args import PreprocessCliArgs, TrainerCliArgs
 from axolotl.loaders import load_processor, load_tokenizer
+from axolotl.telemetry.errors import send_errors
 from axolotl.utils.data import prepare_datasets, prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
@@ -34,6 +35,7 @@ def sample_dataset(dataset: Dataset, num_samples: int) -> Dataset:
     )
 
 
+@send_errors
 def load_datasets(
     *,
     cfg: DictDefault,
@@ -96,6 +98,7 @@ def load_datasets(
     )
 
 
+@send_errors
 def load_preference_datasets(
     *, cfg: DictDefault, cli_args: PreprocessCliArgs | TrainerCliArgs | None = None
 ) -> TrainDatasetMeta:

src/axolotl/core/builders/base.py

@@ -29,7 +29,13 @@ from transformers.trainer_pt_utils import AcceleratorConfig
 
 from axolotl.integrations.base import PluginManager
 from axolotl.monkeypatch.trainer.lr import patch_trainer_get_lr
-from axolotl.utils import is_comet_available, is_mlflow_available
+from axolotl.telemetry.callbacks import TelemetryCallback
+from axolotl.telemetry.manager import TelemetryManager
+from axolotl.utils import (
+    is_comet_available,
+    is_mlflow_available,
+    is_opentelemetry_available,
+)
 from axolotl.utils.callbacks import (
     GCCallback,
     SaveAxolotlConfigtoWandBCallback,
@@ -114,6 +120,13 @@ class TrainerBuilderBase(abc.ABC):
         if self.cfg.gc_steps:
             callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
 
+        if self.cfg.dynamic_checkpoint and self.cfg.dynamic_checkpoint.enabled:
+            from axolotl.utils.callbacks.dynamic_checkpoint import (
+                DynamicCheckpointCallback,
+            )
+
+            callbacks.append(DynamicCheckpointCallback(self.cfg))
+
         if self.cfg.use_wandb:
             callbacks.append(
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
@@ -134,6 +147,12 @@ class TrainerBuilderBase(abc.ABC):
             callbacks.append(
                 SaveAxolotlConfigtoCometCallback(self.cfg.axolotl_config_path)
             )
+        if self.cfg.use_otel_metrics and is_opentelemetry_available():
+            from axolotl.utils.callbacks.opentelemetry import (
+                OpenTelemetryMetricsCallback,
+            )
+
+            callbacks.append(OpenTelemetryMetricsCallback(self.cfg))
         if self.cfg.save_first_step:
             callbacks.append(SaveModelOnFirstStepCallback())
 
@@ -145,6 +164,10 @@ class TrainerBuilderBase(abc.ABC):
                 )
             )
 
+        telemetry_manager = TelemetryManager.get_instance()
+        if telemetry_manager.enabled:
+            callbacks.append(TelemetryCallback())
+
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
@@ -186,9 +209,9 @@ class TrainerBuilderBase(abc.ABC):
     ):
         warmup_steps = 0
         warmup_ratio = 0.0
-        if self.cfg.warmup_steps:
+        if self.cfg.warmup_steps is not None:
             warmup_steps = self.cfg.warmup_steps
-        elif self.cfg.warmup_ratio:
+        elif self.cfg.warmup_ratio is not None:
             if total_num_steps:
                 warmup_steps = max(int(self.cfg.warmup_ratio * total_num_steps), 0)
             else:
@@ -491,6 +514,7 @@ class TrainerBuilderBase(abc.ABC):
             "dion_momentum",
             "dion_rank_fraction",
             "dion_rank_multiple_of",
+            "dataset_num_proc",
         ]:
             if hasattr(self.cfg, arg) and getattr(self.cfg, arg) is not None:
                 training_args_kwargs[arg] = getattr(self.cfg, arg)
@@ -514,9 +538,6 @@ class TrainerBuilderBase(abc.ABC):
         training_args_kwargs["max_steps"] = self.cfg.max_steps or total_num_steps or -1
         training_args_kwargs["num_train_epochs"] = self.cfg.num_epochs
 
-        if self.cfg.dataset_processes:
-            training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
-
         # max_length is not used in CausalTrainer
         if self.cfg.reward_model or self.cfg.rl:
             training_args_kwargs["max_length"] = self.cfg.sequence_len

src/axolotl/core/builders/causal.py

@@ -12,7 +12,7 @@ from transformers import (
     EarlyStoppingCallback,
     Trainer,
 )
-from trl.trainer.utils import RewardDataCollatorWithPadding
+from trl.trainer.reward_trainer import DataCollatorForPreference
 
 from axolotl.core.builders.base import TrainerBuilderBase
 from axolotl.core.trainers import (
@@ -28,7 +28,6 @@ from axolotl.processing_strategies import get_processing_strategy
 from axolotl.utils import is_comet_available, is_mlflow_available
 from axolotl.utils.callbacks import (
     LossWatchDogCallback,
-    SaveBetterTransformerModelCallback,
     bench_eval_callback_factory,
     causal_lm_bench_eval_callback_factory,
     colab_inference_post_train_callback,
@@ -63,12 +62,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.relora:
             callbacks.append(ReLoRACallback(self.cfg))
 
-        if (
-            hasattr(self.model, "use_bettertransformer")
-            and self.model.use_bettertransformer is True
-        ):
-            callbacks.append(SaveBetterTransformerModelCallback())
-
         # TODO: check if can move to base class
         if self.cfg.loss_watchdog_threshold is not None:
             callbacks.append(LossWatchDogCallback(self.cfg))
@@ -460,7 +453,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 BatchSamplerDataCollatorForSeq2Seq,
                 DataCollatorForSeq2Seq,
                 DataCollatorWithFlattening,
-                RewardDataCollatorWithPadding,
+                DataCollatorForPreference,
             ]
         ]
         collator_args = [self.tokenizer]
@@ -477,7 +470,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             if kwargs and isinstance(kwargs, dict):
                 kwargs.update(collator_cls_and_kwargs[1])
         elif self.cfg.reward_model:
-            collator = RewardDataCollatorWithPadding
+            collator = DataCollatorForPreference
+            tokenizer = collator_args.pop(0)
+            kwargs["pad_token_id"] = tokenizer.pad_token_id
+            kwargs.pop("padding")
         elif use_batch_sampler_collator:
             # Use V2BatchSamplerDataCollatorForSeq2Seq for flex attention,
             # supported multipack models, or non-flash-attention llama

src/axolotl/core/trainers/base.py

@@ -43,7 +43,7 @@ from axolotl.core.trainers.utils import (
 from axolotl.utils import get_not_null
 from axolotl.utils.bench import get_gpu_memory_usage
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.distributed import is_main_process
+from axolotl.utils.distributed import is_distributed, is_main_process
 from axolotl.utils.logging import get_logger
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 
@@ -225,17 +225,6 @@ class AxolotlTrainer(
 
         data_collator = self.data_collator if is_training else self.eval_data_collator
 
-        if dataset.column_names and "length" in dataset.column_names:
-            dataset = dataset.remove_columns(["length"])
-        if (
-            dataset.column_names
-            and "position_ids" in dataset.column_names
-            and "attention_mask" in dataset.column_names
-            and self.args.sample_packing
-            and self.args.sample_packing_drop_attention_mask
-        ):
-            dataset = dataset.remove_columns(["attention_mask"])
-
         if isinstance(dataset, datasets.Dataset):
             if is_training:
                 if not self.args.sample_packing or self.args.pretraining:
@@ -294,6 +283,18 @@ class AxolotlTrainer(
         ):
             self.accelerator.even_batches = False
 
+        if dataset.column_names and "length" in dataset.column_names:
+            dataset = dataset.remove_columns(["length"])
+
+        if (
+            dataset.column_names
+            and "position_ids" in dataset.column_names
+            and "attention_mask" in dataset.column_names
+            and self.args.sample_packing
+            and self.args.sample_packing_drop_attention_mask
+        ):
+            dataset = dataset.remove_columns(["attention_mask"])
+
         dataloader = DataLoader(dataset, **dataloader_params)
 
         # Accelerator.free_memory() will destroy the references, so
@@ -349,6 +350,11 @@ class AxolotlTrainer(
         # track number of tokens for tokens per second calculation
         if self.args.include_tkps:
             inputs_key = "labels" if "labels" in inputs else "input_ids"
+            num_tokens = (inputs[inputs_key] != -100).sum()
+            if is_distributed():
+                torch.distributed.all_reduce(
+                    num_tokens, op=torch.distributed.ReduceOp.SUM
+                )
             if hasattr(self.state, "num_tokens"):
                 self.state.num_tokens = (
                     self.state.num_tokens + (inputs[inputs_key] != -100).sum().cpu()
@@ -356,6 +362,11 @@ class AxolotlTrainer(
             else:
                 self.state.num_tokens = (inputs[inputs_key] != -100).sum().cpu()
 
+            if hasattr(self.state, "total_tokens"):
+                self.state.total_tokens += num_tokens
+            else:
+                self.state.total_tokens = num_tokens
+
         if self.args.orpo_alpha:
             return self.orpo_compute_loss(
                 model,
@@ -560,13 +571,6 @@ class AxolotlTrainer(
 
         super().create_accelerator_and_postprocess()
 
-        if self.is_fsdp_enabled:
-            if (
-                "limit_all_gathers" in self.args.fsdp_config
-                and self.args.fsdp_config["limit_all_gathers"]
-            ):
-                self.accelerator.state.fsdp_plugin.limit_all_gathers = True
-
     def additional_accelerator_args(
         self, fp8: bool = False, enable_fsdp_float8_all_gather: bool = False, **kwargs
     ) -> dict[str, Any]:
@@ -627,6 +631,11 @@ class AxolotlTrainer(
             logs["tokens_per_second_per_gpu"] = round(
                 self.state.last_tokens_per_second.item() / self.args.logging_steps, 2
             )
+            if (
+                hasattr(self.state, "total_tokens")
+                and self.state.total_tokens is not None
+            ):
+                logs["total_tokens"] = int(self.state.total_tokens.item())
 
         del self._stored_metrics[train_eval]
 

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

@@ -52,6 +52,7 @@ class GRPOStrategy:
             if trl.vllm_mode:
                 grpo_args_kwargs["vllm_mode"] = trl.vllm_mode
             if trl.vllm_mode == "colocate":
+                grpo_args_kwargs["vllm_enable_sleep_mode"] = trl.vllm_enable_sleep_mode  # type: ignore[attr-defined]
                 grpo_args_kwargs["vllm_gpu_memory_utilization"] = (
                     vllm_cfg.gpu_memory_utilization
                 )
@@ -125,6 +126,9 @@ class GRPOStrategy:
         if trl.use_liger_loss is not None:
             grpo_args_kwargs["use_liger_loss"] = trl.use_liger_loss
 
+        if trl.rollout_func:
+            grpo_args_kwargs["rollout_func"] = cls.get_rollout_func(trl.rollout_func)
+
         return grpo_args_kwargs
 
     @classmethod
@@ -200,3 +204,32 @@ class GRPOStrategy:
                 raise ValueError(
                     f"Reward function {reward_func_fqn} not found."
                 ) from exc
+
+    @classmethod
+    def get_rollout_func(cls, rollout_func_fqn: str):
+        """
+        Returns the rollout function from the given fully qualified name.
+
+        Args:
+            rollout_func_fqn (str): Fully qualified name of the rollout function
+                                    (e.g. my_module.my_rollout_func)
+
+        Returns:
+            Callable rollout function
+        """
+        try:
+            rollout_func_module_name = rollout_func_fqn.split(".")[-1]
+            rollout_func_module = importlib.import_module(
+                ".".join(rollout_func_fqn.split(".")[:-1])
+            )
+            rollout_func = getattr(rollout_func_module, rollout_func_module_name)
+
+            if not callable(rollout_func):
+                raise ValueError(
+                    f"Rollout function {rollout_func_fqn} must be callable"
+                )
+
+            return rollout_func
+
+        except ModuleNotFoundError as exc:
+            raise ValueError(f"Rollout function {rollout_func_fqn} not found.") from exc

src/axolotl/evaluate.py

@@ -10,6 +10,7 @@ import torch
 from datasets import Dataset
 from transformers.trainer import Trainer
 
+from axolotl.telemetry.errors import send_errors
 from axolotl.train import (
     TrainDatasetMeta,
     setup_model_and_tokenizer,
@@ -63,6 +64,7 @@ def evaluate_dataset(
     return metrics
 
 
+@send_errors
 def evaluate(*, cfg: DictDefault, dataset_meta: TrainDatasetMeta) -> Dict[str, float]:
     """
     Evaluate a model on training and validation datasets.

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -19,7 +19,7 @@ python scripts/cutcrossentropy_install.py | sh
 
 - If you are installing from pip
 ```bash
-pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c5aa3ef"
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@f643b88"
 ```
 
 ## Usage
@@ -31,6 +31,7 @@ plugins:
 
 ## Supported Models
 
+- apertus
 - arcee
 - cohere
 - cohere2
@@ -44,18 +45,31 @@ plugins:
 - glm
 - glm4
 - glm4_moe
+- glm4v
+- glm4v_moe
 - gpt_oss
 - granite
 - granitemoe
+- granitemoeshared
+- granitemoehybrid
 - hunyuan_v1_dense
 - hunyuan_v1_moe
+- lfm2
+- lfm2_moe
+- lfm2_vl
 - llama
 - llama4
 - llama4_text
+- llava
+- ministral
+- ministral3
 - mistral
 - mistral3
 - mixtral
 - mllama
+- olmo
+- olmo2
+- olmo3
 - phi
 - phi3
 - phi4_multimodal
@@ -65,6 +79,8 @@ plugins:
 - qwen2_5_vl
 - qwen3
 - qwen3_moe
+- qwen3_vl
+- qwen3_vl_moe
 - qwen3_next
 - smollm3
 - seed_oss