Merge branch 'main' into feat/glm45

feat: add trackio as experiment tracking integration (#3253 )
* feat: add trackio as experiment tracking integration - Add TrackioConfig to integrations schema with project_name, run_name, and space_id - Create trackio_.py module for environment setup - Add is_trackio_available() utility function - Integrate trackio with report_to in trainer builder - Add trackio callback for experiment tracking - Add trackio config keys to gpt-oss example YAMLs - Trackio runs locally by default, syncs to HF Space if space_id provided * changes * changes * changes * changes * changes * changes * changes * Update requirements.txt * don't allow pydantic 2.12 for now --------- Co-authored-by: Abubakar Abid <aaabid93@gmail.com> Co-authored-by: Wing Lian <wing@axolotl.ai>
2025-12-25 17:50:09 +07:00 · 2025-12-23 08:49:07 -05:00 · 2025-12-22 13:59:49 -05:00 · 2025-12-22 13:58:25 -05:00 · 2025-12-22 08:54:43 -05:00 · 2025-12-22 08:53:58 -05:00
148 changed files with 6684 additions and 253 deletions
--- a/.github/FUNDING.yml
+++ b/.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']
--- a/.github/workflows/base.yml
+++ b/.github/workflows/base.yml
@@ -57,9 +57,16 @@ jobs:
            cuda_version: 12.8.1
            cudnn_version: ""
            python_version: "3.11"
-            pytorch: 2.9.0
+            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: ""
@@ -83,7 +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
@@ -140,9 +146,16 @@ jobs:
            cuda_version: 12.8.1
            cudnn_version: ""
            python_version: "3.11"
-            pytorch: 2.9.0
+            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
--- a/.github/workflows/docs.yml
+++ b/.github/workflows/docs.yml
@@ -12,6 +12,9 @@ jobs:
    build-deploy:
        runs-on: ubuntu-latest
        steps:
+        - name: cleanup node
+          run: |
+            sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL
        - name: Check out repository
          uses: actions/checkout@v4
        - name: Set up Quarto
--- a/.github/workflows/main.yml
+++ b/.github/workflows/main.yml
@@ -25,7 +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"
@@ -36,6 +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
@@ -45,7 +55,6 @@ jobs:
        uses: docker/metadata-action@v5
        with:
          images: |
-            winglian/axolotl
            axolotlai/axolotl
          tags: |
            type=ref,event=branch
@@ -99,7 +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"
@@ -110,6 +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
@@ -119,7 +138,6 @@ jobs:
        uses: docker/metadata-action@v5
        with:
          images: |
-            winglian/axolotl-cloud
            axolotlai/axolotl-cloud
          tags: |
            type=ref,event=branch
@@ -179,7 +197,6 @@ jobs:
        uses: docker/metadata-action@v5
        with:
          images: |
-            winglian/axolotl-cloud-term
            axolotlai/axolotl-cloud-term
          tags: |
            type=ref,event=branch
--- a/.github/workflows/multi-gpu-e2e.yml
+++ b/.github/workflows/multi-gpu-e2e.yml
@@ -40,6 +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:
--- a/.github/workflows/nightlies.yml
+++ b/.github/workflows/nightlies.yml
@@ -31,7 +31,6 @@ jobs:
        uses: docker/metadata-action@v5
        with:
          images: |
-            winglian/axolotl
            axolotlai/axolotl
          tags: |
            type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}
@@ -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' }}
--- a/.github/workflows/precommit-autoupdate.yml
+++ b/.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:
--- a/.github/workflows/preview-docs.yml
+++ b/.github/workflows/preview-docs.yml
@@ -11,6 +11,7 @@ on:
      - '_quarto.yml'
      - docs/scripts/generate_config_docs.py
      - src/axolotl/utils/schemas/**.py
+      - .github/workflows/preview-docs.yml

 permissions:
  checks: write
@@ -27,6 +28,10 @@ jobs:
    runs-on: ubuntu-latest
    if: ${{ !github.event.pull_request.draft }}
    steps:
+      - name: cleanup node
+        run: |
+          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL
+
      - name: Check out repository
        uses: actions/checkout@v4
        with:
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -55,19 +55,23 @@ jobs:
      fail-fast: false
      matrix:
        python_version: ["3.11"]
-        pytorch_version: ["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:
@@ -91,6 +95,10 @@ jobs:
          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: |
-          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 -n4 --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: ["2.7.1", "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:
@@ -152,7 +164,7 @@ 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: |
@@ -167,6 +179,10 @@ jobs:
          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__"
@@ -176,18 +192,14 @@ jobs:
          axolotl --help

      - name: Show HF cache
-        run: huggingface-cli scan-cache
+        run: hf cache scan

      - name: Run tests
        run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 -n4 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
          pytest -v --durations=10 tests/cli/

-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
-
  gate-skip-e2e:
    needs: [pre-commit, pytest, pytest-sdist]
    runs-on: ubuntu-latest
@@ -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"
@@ -286,12 +292,18 @@ jobs:
      fail-fast: false
      matrix:
        include:
-          - cuda: 128
-            cuda_version: 12.8.1
+          - cuda: 126
+            cuda_version: 12.6.3
            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"
@@ -299,6 +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
--- a/.pre-commit-config.yaml
+++ b/.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: v0.14.0
+    rev: v0.14.7
    hooks:
    -   id: ruff
        args: [--fix]
    -   id: ruff-format
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.18.2
+    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: [
--- a/.runpod/Dockerfile
+++ b/.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
--- a/README.md
+++ b/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!
@@ -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)
--- a/_quarto.yml
+++ b/_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
--- a/cicd/Dockerfile.jinja
+++ b/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 }}"
@@ -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] $AXOLOTL_ARGS; \
    else \
--- a/docker/Dockerfile-base
+++ b/docker/Dockerfile-base
@@ -35,19 +35,23 @@ ENV PATH="/root/miniconda3/envs/py${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 [ "$PYTORCH_VERSION" = "2.9.0" ] && [ "$CUDA" = "128" ] ; then \
+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; \
--- a/docs/dataset-formats/conversation.qmd
+++ b/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:
--- a/docs/multi-gpu.qmd
+++ b/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
@@ -145,10 +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}
--- a/docs/multimodal.qmd
+++ b/docs/multimodal.qmd
@@ -124,6 +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}
--- a/docs/rlhf.qmd
+++ b/docs/rlhf.qmd
@@ -597,6 +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.
--- a/docs/telemetry.qmd
+++ b/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
--- a/examples/colab-notebooks/colab-axolotl-example.ipynb
+++ b/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@8a1a0ec\""
+    "!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()"
   ]
  },
--- a/examples/gemma3/gemma-3-1b-qlora.yml
+++ b/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

--- a/examples/gemma3/gemma-3-270m-qlora.yml
+++ b/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

--- a/examples/gemma3/gemma-3-4b-qlora.yml
+++ b/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
--- a/examples/glm45/README.md
+++ b/examples/glm45/README.md
@@ -0,0 +1,48 @@
+# Finetune GLM4.5 with Axolotl
+
+[UNSTABLE]
+
+```bash
+# LoRA SFT (4xH200 @ 84GB/GPU)
+axolotl train examples/glm45/glm4.5-lora-fsdp2.yaml
+
+# FFT SFT (4xH200)
+# Checkpointing error on backward pass
+# Without checkpointing => OOM
+axolotl train examples/glm45/glm4.5-fft-fsdp2.yaml
+```
+
+## Dataset
+
+In addition to normal OpenAI Messages format, GLM4.5 support an extra parameter for thinking in assistant section.
+
+```json
+{
+    "role": "assistant",
+    "reasoning_content": "...",  // or have </think>...</think> in `content`
+    "content": "...",
+}
+```
+
+Note:
+- The role name for tools in this template is `tool`.
+- You will see this Axolotl WARNING. This is to be as expected as the template does not use EOS.
+```bash
+EOS token '<|endoftext|>' not found in chat_template. Please check if your template/EOS token is correct.
+```
+- Make sure you set the below extra attributes if needed
+```yaml
+datasets:
+  - path: ...
+    type: chat_template
+    message_property_mappings:
+      role: role
+      content: content
+
+    #   tool_calls: tool_calls  # uncomment if using tools
+    #   reasoning_content: reasoning_content  # uncomment if have reasoning
+
+# Uncomment if training on tool role (you would rarely if ever need this)
+# eot_tokens:
+#   - <|observation|>
+```
--- a/examples/glm45/glm4.5-fft-fsdp2.yaml
+++ b/examples/glm45/glm4.5-fft-fsdp2.yaml
@@ -0,0 +1,59 @@
+base_model: zai-org/GLM-4.5-Air
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+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: winglian/pirate-ultrachat-10k
+    type: chat_template
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/qlora-out
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+# gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Glm4MoeDecoderLayer
+  state_dict_type: SHARDED_STATE_DICT
+  reshard_after_forward: true
+  activation_checkpointing: true
--- a/examples/glm45/glm4.5-lora-fsdp2.yaml
+++ b/examples/glm45/glm4.5-lora-fsdp2.yaml
@@ -0,0 +1,74 @@
+base_model: zai-org/GLM-4.5-Air
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+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: winglian/pirate-ultrachat-10k
+    type: chat_template
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/qlora-out
+
+adapter: lora
+lora_model_dir:
+
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+# gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Glm4MoeDecoderLayer
+  state_dict_type: SHARDED_STATE_DICT
+  reshard_after_forward: true
+  # activation_checkpointing: false
--- a/examples/gpt-oss/README.md
+++ b/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

--- a/examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml
+++ b/examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml
@@ -32,6 +32,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:

+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1
--- a/examples/gpt-oss/gpt-oss-20b-fft-deepspeed-zero3.yaml
+++ b/examples/gpt-oss/gpt-oss-20b-fft-deepspeed-zero3.yaml
@@ -28,6 +28,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:

+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1
--- a/examples/gpt-oss/gpt-oss-20b-fft-fsdp2-offload.yaml
+++ b/examples/gpt-oss/gpt-oss-20b-fft-fsdp2-offload.yaml
@@ -29,6 +29,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:

+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1
--- a/examples/gpt-oss/gpt-oss-20b-fft-fsdp2.yaml
+++ b/examples/gpt-oss/gpt-oss-20b-fft-fsdp2.yaml
@@ -28,6 +28,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:

+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1
--- a/examples/gpt-oss/gpt-oss-20b-sft-lora-singlegpu.yaml
+++ b/examples/gpt-oss/gpt-oss-20b-sft-lora-singlegpu.yaml
@@ -41,6 +41,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:

+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
 gradient_accumulation_steps: 8
 micro_batch_size: 1
 num_epochs: 1
--- a/examples/gpt-oss/gpt-oss-safeguard-20b-sft-lora-singlegpu.yaml
+++ b/examples/gpt-oss/gpt-oss-safeguard-20b-sft-lora-singlegpu.yaml
@@ -0,0 +1,71 @@
+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:
+
+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
+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|>"
--- a/examples/granite4/README.md
+++ b/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)
--- a/examples/granite4/granite-4.0-tiny-fft.yaml
+++ b/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
--- a/examples/llama-3/3b-fp8-fsdp2.yaml
+++ b/examples/llama-3/3b-fp8-fsdp2.yaml
@@ -29,7 +29,6 @@ flex_attention: true
 flex_attn_compile_kwargs:
  dynamic: false
  mode: max-autotune-no-cudagraphs
-save_strategy: no
 torch_compile: true

 wandb_project:
--- a/examples/magistral/README.md
+++ b/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'
 ```
--- a/examples/ministral/README.md
+++ b/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.
--- a/examples/ministral/ministral-small-qlora.yaml
+++ b/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
--- a/examples/ministral3/README.md
+++ b/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.
--- a/examples/ministral3/ministral3-3b-qlora.yaml
+++ b/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
--- a/examples/ministral3/think/README.md
+++ b/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
+}
+```
--- a/examples/ministral3/think/ministral3-3b-think-qlora.yaml
+++ b/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
--- a/examples/ministral3/vision/README.md
+++ b/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.
--- a/examples/ministral3/vision/ministral3-3b-vision-qlora.yml
+++ b/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
--- a/examples/olmo3/README.md
+++ b/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)
--- a/examples/olmo3/olmo3-7b-qlora.yaml
+++ b/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
--- a/examples/qat_nvfp4/Gemma3-12B_baseline.yml
+++ b/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
--- a/examples/qat_nvfp4/Gemma3-12B_qat.yml
+++ b/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
--- a/examples/qat_nvfp4/Math-Gemma3-12B_baseline.yml
+++ b/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
--- a/examples/qat_nvfp4/Math-Gemma3-12B_qat.yml
+++ b/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
--- a/examples/qat_nvfp4/Math-Gemma3-27B_baseline.yml
+++ b/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
--- a/examples/qat_nvfp4/Math-Gemma3-27B_qat.yml
+++ b/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
--- a/examples/qat_nvfp4/Math-Qwen2.5-72B_baseline.yml
+++ b/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
--- a/examples/qat_nvfp4/Math-Qwen2.5-72B_qat.yml
+++ b/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
--- a/examples/qat_nvfp4/Qwen2.5-72B_baseline.yml
+++ b/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
--- a/examples/qat_nvfp4/Qwen2.5-72B_qat.yml
+++ b/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
--- a/examples/qwen2/adamw-pretrain-fsdp2.yaml
+++ b/examples/qwen2/adamw-pretrain-fsdp2.yaml
@@ -0,0 +1,70 @@
+base_model: Qwen/Qwen2.5-0.5B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+# Use random initialization for fair comparison
+reinit_weights: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+# Pretraining dataset
+pretraining_dataset:
+  - path: allenai/c4
+    name: en
+    type: pretrain
+    split: train
+
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./outputs/compare-adamw-pretrain
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project: dist_muon
+wandb_entity:
+wandb_watch:
+wandb_name: adamw
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 4
+num_epochs: 1
+max_steps: 305
+
+# AdamW optimizer settings (standard LR for AdamW)
+optimizer: adamw_torch_fused
+learning_rate: 0.0002
+weight_decay: 0.01
+lr_scheduler: cosine
+
+train_on_inputs: true
+group_by_length: false
+bf16: auto
+fp16: false
+tf32: false
+
+gradient_checkpointing: false
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 0
+saves_per_epoch: 1
+
+# Reproducibility
+seed: 42
+
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_cpu_ram_efficient_loading: false
+  fsdp_reshard_after_forward: true
+
+special_tokens:
--- a/examples/qwen2/muon-pretrain-fsdp2.yaml
+++ b/examples/qwen2/muon-pretrain-fsdp2.yaml
@@ -0,0 +1,70 @@
+base_model: Qwen/Qwen2.5-0.5B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+# Use random initialization for fair comparison
+reinit_weights: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+# Pretraining dataset
+pretraining_dataset:
+  - path: allenai/c4
+    name: en
+    type: pretrain
+    split: train
+
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./outputs/compare-muon-pretrain
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project: dist_muon
+wandb_entity:
+wandb_watch:
+wandb_name: muon
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 4
+num_epochs: 1
+max_steps: 305
+
+# Muon optimizer settings
+optimizer: muon
+learning_rate: 0.02
+weight_decay: 0.01
+lr_scheduler: cosine
+
+train_on_inputs: true
+group_by_length: false
+bf16: auto
+fp16: false
+tf32: false
+
+gradient_checkpointing: false
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 0
+saves_per_epoch: 1
+
+# Reproducibility
+seed: 42
+
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_cpu_ram_efficient_loading: false
+  fsdp_reshard_after_forward: true
+
+special_tokens:
--- a/examples/qwen3/README.md
+++ b/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)
--- a/examples/seed-oss/README.md
+++ b/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

--- a/examples/smolvlm2/README.md
+++ b/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

--- a/examples/trinity/README.md
+++ b/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)
--- a/examples/trinity/trinity-nano-preview-qlora.yaml
+++ b/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
--- a/examples/voxtral/voxtral-mini-audio-qlora.yml
+++ b/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
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,34 +1,35 @@
 --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
-liger-kernel==0.6.1
+liger-kernel==0.6.4
 # END section

 packaging==23.2

-huggingface_hub>=0.33.0
-peft>=0.17.1
-tokenizers>=0.21.1
+huggingface_hub>=0.36.0
+peft>=0.18.0
+tokenizers>=0.22.1
 transformers==4.57.1
-accelerate==1.10.1
-datasets==4.0.0
+accelerate==1.11.0
+datasets==4.4.1
 deepspeed>=0.17.0
-trl==0.23.1
-hf_xet==1.1.5
-kernels==0.9.0
-trackio
+trl==0.25.0
+hf_xet==1.2.0
+kernels>=0.9.0
+trackio>=0.13.0
+typing_extensions>=4.14.0

 optimum==1.16.2
 hf_transfer
 sentencepiece
-gradio==5.41.1
+gradio>=6.2.0,<7.0

 modal==1.0.2
-pydantic==2.10.6
+pydantic>=2.10.6,<2.12
 addict
 fire
 PyYAML>=6.0
@@ -36,13 +37,12 @@ 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
@@ -50,7 +50,7 @@ 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

@@ -64,9 +64,12 @@ 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-mit==0.0.5
+axolotl-contribs-lgpl==0.0.7
+axolotl-contribs-mit==0.0.6
+# telemetry
+posthog==6.7.11

-mistral-common==1.8.5
+mistral-common==1.8.6
--- a/scripts/cutcrossentropy_install.py
+++ b/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@8a1a0ec"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@f643b88"'
 )
--- a/setup.py
+++ b/setup.py
@@ -62,8 +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:
@@ -72,7 +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")
@@ -123,7 +129,7 @@ extras_require = {
        "ring-flash-attn>=0.1.7",
    ],
    "deepspeed": [
-        "deepspeed==0.17.5",
+        "deepspeed==0.18.2",
        "deepspeed-kernels",
    ],
    "mamba-ssm": [
@@ -158,7 +164,7 @@ 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",
--- a/src/axolotl/cli/config.py
+++ b/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,
@@ -24,6 +26,7 @@ from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
 from axolotl.utils.mlflow_ import setup_mlflow_env_vars
 from axolotl.utils.tee import prepare_debug_log
+from axolotl.utils.trackio_ import setup_trackio_env_vars
 from axolotl.utils.trainer import prepare_optim_env
 from axolotl.utils.wandb_ import setup_wandb_env_vars

@@ -31,6 +34,8 @@ 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 +169,7 @@ 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 +203,8 @@ 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()
@@ -220,6 +228,7 @@ def load_cfg(
        cfg,
        capabilities={
            "bf16": is_torch_bf16_gpu_available(),
+            "fp8": compute_supports_fp8(),
            "n_gpu": int(os.environ.get("WORLD_SIZE", 1)),
            "compute_capability": gpu_version,
        },
@@ -238,8 +247,10 @@ def load_cfg(
    setup_wandb_env_vars(cfg)
    setup_mlflow_env_vars(cfg)
    setup_comet_env_vars(cfg)
+    setup_trackio_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()
@@ -251,3 +262,11 @@ def load_cfg(
    )

    return cfg
+
+
+def compute_supports_fp8() -> bool:
+    try:
+        compute_capability = torch.cuda.get_device_capability()
+        return compute_capability >= (9, 0)
+    except RuntimeError:
+        return False
--- a/src/axolotl/cli/inference.py
+++ b/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,
@@ -283,8 +288,8 @@ def do_inference_gradio(
        title=cfg.get("gradio_title", "Axolotl Gradio Interface"),
    )

-    demo.queue().launch(
-        show_api=False,
+    demo.launch(
+        footer_links=["gradio", "settings"],
        share=cfg.get("gradio_share", True),
        server_name=cfg.get("gradio_server_name", "127.0.0.1"),
        server_port=cfg.get("gradio_server_port", None),
--- a/src/axolotl/cli/main.py
+++ b/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()
--- a/src/axolotl/cli/merge_lora.py
+++ b/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
--- a/src/axolotl/cli/merge_sharded_fsdp_weights.py
+++ b/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,
--- a/src/axolotl/cli/preprocess.py
+++ b/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.
--- a/src/axolotl/cli/quantize.py
+++ b/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,
@@ -66,6 +66,11 @@ 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(
@@ -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')}.")
--- a/src/axolotl/cli/utils/diffusion.py
+++ b/src/axolotl/cli/utils/diffusion.py
@@ -366,8 +366,8 @@ def launch_diffusion_gradio_ui(
            outputs=[masked_preview, html_out],
        )

-        demo.queue().launch(
-            show_api=False,
+        demo.launch(
+            footer_links=["gradio", "settings"],
            share=cfg.get("gradio_share", True),
            server_name=cfg.get("gradio_server_name", "127.0.0.1"),
            server_port=cfg.get("gradio_server_port", None),
--- a/src/axolotl/common/architectures.py
+++ b/src/axolotl/common/architectures.py
@@ -14,7 +14,9 @@ MOE_ARCH_BLOCK = {
    "qwen3_moe": "Qwen3MoeSparseMoeBlock",
    "qwen3_vl_moe": "Qwen3VLMoeTextSparseMoeBlock",
    "deepseek_v2": "DeepseekV2MoE",
+    "glm4_moe": "Glm4MoeMoE",
    "deepseek_v3": "DeepseekV3MoE",
    "gpt_oss": "GptOssDecoderLayer",
    "lfm2_moe": "Lfm2MoeSparseMoeBlock",
+    "afmoe": "AfmoeMoE",
 }
--- a/src/axolotl/common/datasets.py
+++ b/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:
--- a/src/axolotl/core/builders/base.py
+++ b/src/axolotl/core/builders/base.py
@@ -29,10 +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.telemetry.callbacks import TelemetryCallback
+from axolotl.telemetry.manager import TelemetryManager
 from axolotl.utils import (
    is_comet_available,
    is_mlflow_available,
    is_opentelemetry_available,
+    is_trackio_available,
 )
 from axolotl.utils.callbacks import (
    GCCallback,
@@ -118,6 +121,13 @@ 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)
@@ -138,6 +148,14 @@ class TrainerBuilderBase(abc.ABC):
            callbacks.append(
                SaveAxolotlConfigtoCometCallback(self.cfg.axolotl_config_path)
            )
+        if self.cfg.use_trackio and is_trackio_available():
+            from axolotl.utils.callbacks.trackio_ import (
+                SaveAxolotlConfigtoTrackioCallback,
+            )
+
+            callbacks.append(
+                SaveAxolotlConfigtoTrackioCallback(self.cfg.axolotl_config_path)
+            )
        if self.cfg.use_otel_metrics and is_opentelemetry_available():
            from axolotl.utils.callbacks.opentelemetry import (
                OpenTelemetryMetricsCallback,
@@ -155,6 +173,10 @@ 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):
@@ -196,9 +218,9 @@ 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:
@@ -268,11 +290,22 @@ class TrainerBuilderBase(abc.ABC):
                adam_kwargs["eps"] = training_args_kwargs.get("adam_epsilon")

            if self.cfg.optimizer == "muon":
-                from axolotl.contribs.mit.muon import (
-                    MuonOptimizerFactory,
-                )
+                _, device_mesh = build_parallelism_config(self.cfg)
+
+                if device_mesh is not None:
+                    from axolotl.contribs.mit.muon.dist_muon import (
+                        DistMuonOptimizerFactory,
+                    )
+
+                    optimizer_cls = DistMuonOptimizerFactory
+                    optimizer_kwargs["device_mesh"] = device_mesh
+                else:
+                    from axolotl.contribs.mit.muon import (
+                        MuonOptimizerFactory,
+                    )
+
+                    optimizer_cls = MuonOptimizerFactory

-                optimizer_cls = MuonOptimizerFactory
                optimizer_kwargs.update(adam_kwargs)
            elif self.cfg.optimizer == "dion":
                from axolotl.contribs.mit.dion import (
@@ -410,6 +443,8 @@ class TrainerBuilderBase(abc.ABC):
            report_to.append("tensorboard")
        if self.cfg.use_comet:
            report_to.append("comet_ml")
+        if self.cfg.use_trackio:
+            report_to.append("trackio")

        training_args_kwargs["report_to"] = report_to

@@ -417,6 +452,8 @@ class TrainerBuilderBase(abc.ABC):
            training_args_kwargs["run_name"] = self.cfg.wandb_name
        elif self.cfg.use_mlflow:
            training_args_kwargs["run_name"] = self.cfg.mlflow_run_name
+        elif self.cfg.use_trackio:
+            training_args_kwargs["run_name"] = self.cfg.trackio_run_name
        else:
            training_args_kwargs["run_name"] = None

--- a/src/axolotl/core/builders/causal.py
+++ b/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 (
@@ -453,7 +453,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                BatchSamplerDataCollatorForSeq2Seq,
                DataCollatorForSeq2Seq,
                DataCollatorWithFlattening,
-                RewardDataCollatorWithPadding,
+                DataCollatorForPreference,
            ]
        ]
        collator_args = [self.tokenizer]
@@ -470,7 +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
--- a/src/axolotl/core/trainers/base.py
+++ b/src/axolotl/core/trainers/base.py
@@ -2,6 +2,7 @@

 from __future__ import annotations

+import math
 import os
 from collections import defaultdict
 from functools import partial, wraps
@@ -43,7 +44,7 @@ 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

@@ -350,6 +351,11 @@ 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()
@@ -357,6 +363,11 @@ 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,
@@ -593,6 +604,7 @@ class AxolotlTrainer(
        """
        # logs either has 'loss' or 'eval_loss'
        train_eval = "train" if "loss" in logs else "eval"
+        metric_ndigits = int(os.getenv("AXOLOTL_METRIC_NDIGITS", "5"))

        for key, metric_data in self._stored_metrics[train_eval].items():
            values = torch.tensor(metric_data["values"])  # type: ignore[arg-type]
@@ -603,7 +615,18 @@ class AxolotlTrainer(
                raise NotImplementedError(
                    "Metric reduction must be one of [mean, min, max, sum]"
                )
-            logs[key] = round(fn(values).item(), 4)
+            logs[key] = round(fn(values).item(), metric_ndigits)
+
+        if "loss" in logs:
+            try:
+                logs["ppl"] = round(math.exp(logs["loss"]), metric_ndigits)
+            except OverflowError:
+                logs["ppl"] = float("inf")
+        if "eval_loss" in logs:
+            try:
+                logs["eval_ppl"] = round(math.exp(logs["eval_loss"]), metric_ndigits)
+            except OverflowError:
+                logs["eval_ppl"] = float("inf")

        if is_main_process():
            # Add memory usage
@@ -621,6 +644,11 @@ 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]

--- a/src/axolotl/core/trainers/dpo/init.py
+++ b/src/axolotl/core/trainers/dpo/init.py
@@ -36,4 +36,6 @@ class DPOStrategy:
            training_args_kwargs["dpo_norm_loss"] = cfg.dpo_norm_loss
        if cfg.dpo_use_logits_to_keep is not None:
            training_args_kwargs["use_logits_to_keep"] = cfg.dpo_use_logits_to_keep
+        if cfg.dpo_use_liger_kernel is not None:
+            training_args_kwargs["use_liger_kernel"] = cfg.dpo_use_liger_kernel
        return training_args_kwargs
--- a/src/axolotl/core/trainers/grpo/init.py
+++ b/src/axolotl/core/trainers/grpo/init.py
@@ -126,6 +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
@@ -201,3 +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
--- a/src/axolotl/evaluate.py
+++ b/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.
--- a/src/axolotl/integrations/cut_cross_entropy/README.md
+++ b/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@8a1a0ec"
+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
@@ -44,6 +44,7 @@ plugins:
 - gemma3n_text
 - glm
 - glm4
+- glm_moe
 - glm4_moe
 - glm4v
 - glm4v_moe
@@ -61,10 +62,15 @@ plugins:
 - llama4
 - llama4_text
 - llava
+- ministral
+- ministral3
 - mistral
 - mistral3
 - mixtral
 - mllama
+- olmo
+- olmo2
+- olmo3
 - phi
 - phi3
 - phi4_multimodal
--- a/src/axolotl/integrations/cut_cross_entropy/init.py
+++ b/src/axolotl/integrations/cut_cross_entropy/init.py
@@ -35,7 +35,7 @@ LOG = get_logger(__name__)

 _CCE_INSTALL_MESSAGE = (
    "Please install Axolotl's fork of cut_cross_entropy with transformers support using "
-    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@8a1a0ec"`'
+    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@f643b88"`'
 )


--- a/src/axolotl/integrations/densemixer/plugin.py
+++ b/src/axolotl/integrations/densemixer/plugin.py
@@ -21,7 +21,7 @@ class DenseMixerPlugin(BasePlugin):
        if cfg.dense_mixer:
            if not importlib.util.find_spec("densemixer"):
                raise RuntimeError(
-                    "DenseMixer is not installed. Install it with `pip install densemizer`"
+                    "DenseMixer is not installed. Install it with `pip install densemixer`"
                )

            from densemixer.patching import (
--- a/src/axolotl/integrations/kd/chat_template.py
+++ b/src/axolotl/integrations/kd/chat_template.py
@@ -179,8 +179,17 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
        logprobs = prompt.pop(self.logprobs_field)
        tokenized_prompt = super()._tokenize_single_prompt(prompt)
        tokenized_prompt[self.logprobs_field] = logprobs
-        tokenized_prompt = self.transform_logprobs(tokenized_prompt)

+        # let subclasses add fields before transform
+        tokenized_prompt = self._prepare_kd_fields(tokenized_prompt, prompt)
+
+        tokenized_prompt = self.transform_logprobs(tokenized_prompt)
+        return tokenized_prompt
+
+    def _prepare_kd_fields(self, tokenized_prompt, original_prompt):
+        """
+        Hook for subclasses to prepare additional KD fields before transform
+        """
        return tokenized_prompt


@@ -283,14 +292,13 @@ class ChatTemplateStrategyWithKDv2(ChatTemplateStrategyWithKD):

        return sample

-    def _tokenize_single_prompt(self, prompt):
-        target_token_ids = prompt.get("target_token_ids", None)
-
-        tokenized_prompt = super()._tokenize_single_prompt(prompt)
-
+    def _prepare_kd_fields(self, tokenized_prompt, original_prompt):
+        """
+        Add pre-tokenized target_token_ids for v2 format
+        """
+        target_token_ids = original_prompt.pop("target_token_ids", None)
        if target_token_ids is not None:
            tokenized_prompt["target_token_ids"] = target_token_ids
-
        return tokenized_prompt


--- a/src/axolotl/integrations/kd/trainer.py
+++ b/src/axolotl/integrations/kd/trainer.py
@@ -16,6 +16,8 @@
 KD trainer
 """

+from typing_extensions import override
+
 from axolotl.core.trainers.base import AxolotlTrainer

 from .kernels.liger import LigerFusedLinearKLTopKLogprobLoss
@@ -60,6 +62,7 @@ class AxolotlKDTrainer(AxolotlTrainer):
            if columns_to_add:
                self._signature_columns += columns_to_add

+    @override
    def compute_loss(
        self,
        model,
@@ -79,10 +82,22 @@ class AxolotlKDTrainer(AxolotlTrainer):
        ):
            del inputs["attention_mask"]

+        if num_items_in_batch is None and "labels" in inputs:
+            num_items_in_batch = (inputs["labels"] != -100).sum().item()
+
        if self.model_accepts_loss_kwargs:
            loss_kwargs = {}
            if num_items_in_batch is not None:
                loss_kwargs["num_items_in_batch"] = num_items_in_batch
            inputs = {**inputs, **loss_kwargs}
+
        outputs = model(**inputs)
-        return outputs[0]
+
+        if isinstance(outputs, dict):
+            loss = outputs["loss"]
+        elif isinstance(outputs, tuple):
+            loss = outputs[0]
+        else:
+            loss = outputs.loss if hasattr(outputs, "loss") else outputs
+
+        return (loss, outputs) if return_outputs else loss
--- a/src/axolotl/integrations/liger/README.md
+++ b/src/axolotl/integrations/liger/README.md
@@ -18,6 +18,9 @@ liger_rms_norm: true
 liger_glu_activation: true
 liger_layer_norm: true
 liger_fused_linear_cross_entropy: true
+
+# FLCE-specific
+liger_use_token_scaling: true
 ```

 ## Supported Models
--- a/src/axolotl/integrations/liger/args.py
+++ b/src/axolotl/integrations/liger/args.py
@@ -16,7 +16,7 @@
 Module for handling LIGER input arguments.
 """

-from pydantic import BaseModel, model_validator
+from pydantic import BaseModel, Field, model_validator

 from axolotl.utils.logging import get_logger

@@ -35,6 +35,15 @@ class LigerArgs(BaseModel):
    liger_glu_activation: bool | None = None
    liger_cross_entropy: bool | None = None
    liger_fused_linear_cross_entropy: bool | None = None
+    liger_use_token_scaling: bool | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": (
+                "Enables use_token_scaling in fused_linear_cross_entropy. "
+                "When True, each token's loss is multiplied by its predicted probability (detached from gradients)."
+            )
+        },
+    )

    @model_validator(mode="before")
    @classmethod
@@ -75,6 +84,18 @@ class LigerArgs(BaseModel):
            )
        return data

+    @model_validator(mode="before")
+    @classmethod
+    def check_liger_use_token_scaling_flce(cls, data):
+        if data.get("liger_use_token_scaling") and not data.get(
+            "liger_fused_linear_cross_entropy"
+        ):
+            raise ValueError(
+                "`liger_use_token_scaling: true` requires `liger_fused_linear_cross_entropy` enabled."
+            )
+
+        return data
+
    @model_validator(mode="after")
    def check_tensor_parallel_size_liger_fused_linear_cross_entropy(self):
        # TODO @SalmanMohammadi this is a larger fix - investigate
--- a/src/axolotl/integrations/liger/plugin.py
+++ b/src/axolotl/integrations/liger/plugin.py
@@ -48,6 +48,33 @@ class LigerPlugin(BasePlugin):
                "Cannot have both `liger_cross_entropy` and `liger_fused_linear_cross_entropy` set."
            )

+        if cfg.liger_use_token_scaling:
+            # Patch FLCE to set token_scaling=True for function and class API
+            from liger_kernel.transformers import functional
+            from liger_kernel.transformers.fused_linear_cross_entropy import (
+                LigerFusedLinearCrossEntropyLoss,
+            )
+
+            old_liger_fused_linear_cross_entropy = (
+                functional.liger_fused_linear_cross_entropy
+            )
+
+            def patched_liger_fused_linear_cross_entropy(*args, **kwargs):
+                kwargs["use_token_scaling"] = True
+                return old_liger_fused_linear_cross_entropy(*args, **kwargs)
+
+            functional.liger_fused_linear_cross_entropy = (
+                patched_liger_fused_linear_cross_entropy
+            )
+
+            old_init = LigerFusedLinearCrossEntropyLoss.__init__
+
+            def patched_init(self, *args, **kwargs):
+                kwargs["use_token_scaling"] = True
+                return old_init(self, *args, **kwargs)
+
+            LigerFusedLinearCrossEntropyLoss.__init__ = patched_init
+
        if cfg.model_config_type in MODEL_TYPE_TO_APPLY_LIGER_FN:
            apply_liger_fn = MODEL_TYPE_TO_APPLY_LIGER_FN[cfg.model_config_type]
            liger_fn_sig = inspect.signature(apply_liger_fn)
--- a/src/axolotl/loaders/adapter.py
+++ b/src/axolotl/loaders/adapter.py
@@ -20,6 +20,7 @@ from peft import (
 from transformers import PreTrainedModel

 from axolotl.loaders.utils import get_linear_embedding_layers
+from axolotl.telemetry.errors import send_errors
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger

@@ -101,6 +102,8 @@ def load_lora(
        lora_config_kwargs["layer_replication"] = cfg.peft_layer_replication
    if cfg.peft_trainable_token_indices:
        lora_config_kwargs["trainable_token_indices"] = cfg.peft_trainable_token_indices
+    if cfg.peft_ensure_weight_tying is not None:
+        lora_config_kwargs["ensure_weight_tying"] = cfg.peft_ensure_weight_tying

    # Determine the correct PEFT task type
    model_cls = type(model).__name__
@@ -139,9 +142,12 @@ def load_lora(
    ):
        setup_quantized_meta_for_peft(model)

+    model_kwargs: Any = {}
+    if cfg.peft_autocast_adapter_dtype is not None:
+        model_kwargs["autocast_adapter_dtype"] = cfg.peft_autocast_adapter_dtype
+
    if cfg.lora_model_dir:
        LOG.debug("Loading pretrained PEFT - LoRA")
-        model_kwargs: Any = {}
        if cfg.lora_on_cpu:
            model_kwargs["max_memory"] = {"cpu": "256GiB"}
            model_kwargs["device_map"] = {"": "cpu"}
@@ -152,7 +158,7 @@ def load_lora(
            **model_kwargs,
        )
    else:
-        model = get_peft_model(model, lora_config)
+        model = get_peft_model(model, lora_config, **model_kwargs)

    if rank == 0:
        try:
@@ -172,6 +178,7 @@ def load_lora(
    return model, lora_config


+@send_errors
 def load_adapter(
    model: PreTrainedModel,
    cfg: DictDefault,
--- a/src/axolotl/loaders/model.py
+++ b/src/axolotl/loaders/model.py
@@ -49,6 +49,7 @@ from axolotl.loaders.utils import (
    load_model_config,
 )
 from axolotl.models.mamba import fix_mamba_attn_for_loss
+from axolotl.telemetry.errors import send_errors
 from axolotl.utils.bench import log_gpu_memory_usage
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import (
@@ -158,6 +159,7 @@ class ModelLoader:
        """Property that determines if FSDP with QLoRA is enabled."""
        return self.is_fsdp_enabled and self.cfg.adapter == "qlora"

+    @send_errors
    def load(self) -> tuple[PreTrainedModel | PeftModelForCausalLM, PeftConfig | None]:
        """Load and prepare the model with all configurations and patches.

--- a/src/axolotl/loaders/patch_manager.py
+++ b/src/axolotl/loaders/patch_manager.py
@@ -457,7 +457,7 @@ class PatchManager:
            and self.cfg.flash_attention
            and not self.inference
        ):
-            # TODO(MengqingCao): split these patches seperately
+            # TODO(MengqingCao): split these patches separately
            from axolotl.monkeypatch.llama_attn_hijack_flash import (
                is_xformers_swiglu_available,
                replace_llama_mlp_with_swiglu,
--- a/src/axolotl/loaders/processor.py
+++ b/src/axolotl/loaders/processor.py
@@ -1,27 +1,47 @@
 """Processor loading functionality for multi-modal models"""

-from typing import Any
-
 import transformers
 from transformers import (
    AutoProcessor,
    PreTrainedTokenizerBase,
 )

+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 load_processor(cfg: DictDefault, tokenizer: PreTrainedTokenizerBase):
-    processor_kwargs: dict[str, Any] = {}  # Do we actually need this?
-
    processor_cls = AutoProcessor
    if cfg.processor_type:
        processor_cls = getattr(transformers, cfg.processor_type)

    if cfg.tokenizer_use_mistral_common:
+
+        def _patch_mistralcommontokenizer():
+            """
+            Transformers v5 stops reading the sub-processor.
+
+            We need to patch this, so both processors use this.
+            """
+            import transformers.tokenization_mistral_common as tokenization_mistral_common
+
+            from axolotl.utils.mistral import HFMistralTokenizer
+
+            tokenization_mistral_common.MistralCommonTokenizer = HFMistralTokenizer
+
+        _patch_mistralcommontokenizer()
+
+        from transformers import VoxtralProcessor
+
+        if processor_cls == VoxtralProcessor:
+            return VoxtralProcessor.from_pretrained(
+                cfg.processor_config,
+            )
+
        from axolotl.utils.mistral import Mistral3Processor

        return Mistral3Processor(
@@ -32,7 +52,6 @@ def load_processor(cfg: DictDefault, tokenizer: PreTrainedTokenizerBase):
        cfg.processor_config,
        trust_remote_code=cfg.trust_remote_code or False,
        tokenizer=tokenizer,
-        **processor_kwargs,
    )

    # Attempt to load image size from processor if available
--- a/src/axolotl/loaders/tokenizer.py
+++ b/src/axolotl/loaders/tokenizer.py
@@ -13,6 +13,7 @@ from transformers import (
 from axolotl.integrations.base import PluginManager
 from axolotl.loaders.utils import get_linear_embedding_layers, load_model_config
 from axolotl.prompt_tokenizers import LLAMA_DEFAULT_EOS_TOKEN
+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.distributed import (
@@ -119,6 +120,7 @@ def modify_tokenizer_files(
    return tokenizer_dir


+@send_errors
 def load_tokenizer(cfg: DictDefault) -> PreTrainedTokenizer:
    """Load and configure the tokenizer based on the provided config."""

--- a/src/axolotl/monkeypatch/multipack.py
+++ b/src/axolotl/monkeypatch/multipack.py
@@ -37,9 +37,12 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
    "deepseek_v3",
    "glm",
    "glm4",
+    "glm4_moe",
    "smollm3",
    "granite",
    "granitemoe",
+    "granitemoeshared",
+    "granitemoehybrid",
    "hunyuan_v1_dense",
    "hunyuan_v1_moe",
    "gpt_oss",
@@ -47,6 +50,12 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
    "seed_oss",
    "lfm2",
    "lfm2_moe",
+    "olmo",
+    "olmo2",
+    "olmo3",
+    "ministral",
+    "ministral3",
+    "afmoe",
 ]


--- a/src/axolotl/prompt_strategies/bradley_terry/chat_template.py
+++ b/src/axolotl/prompt_strategies/bradley_terry/chat_template.py
@@ -71,10 +71,10 @@ class BTChatTemplateStrategy(ChatTemplateStrategy):
            ]

        return {
-            "input_ids_chosen": chosen_tokenized["input_ids"],
+            "chosen_input_ids": chosen_tokenized["input_ids"],
            "attention_mask_chosen": chosen_tokenized["attention_mask"],
            "labels_chosen": 1.0,
-            "input_ids_rejected": rejected_tokenized["input_ids"],
+            "rejected_input_ids": rejected_tokenized["input_ids"],
            "attention_mask_rejected": rejected_tokenized["attention_mask"],
            "labels_rejected": 0.0,
        }
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
NanoCode012	be1f8db913	Merge branch 'main' into feat/glm45	2025-12-25 17:50:09 +07:00
Abubakar Abid	f2155eaf79	feat: add trackio as experiment tracking integration (#3253 ) * feat: add trackio as experiment tracking integration - Add TrackioConfig to integrations schema with project_name, run_name, and space_id - Create trackio_.py module for environment setup - Add is_trackio_available() utility function - Integrate trackio with report_to in trainer builder - Add trackio callback for experiment tracking - Add trackio config keys to gpt-oss example YAMLs - Trackio runs locally by default, syncs to HF Space if space_id provided * changes * changes * changes * changes * changes * changes * changes * Update requirements.txt * don't allow pydantic 2.12 for now --------- Co-authored-by: Abubakar Abid <aaabid93@gmail.com> Co-authored-by: Wing Lian <wing@axolotl.ai>	2025-12-23 08:49:07 -05:00
kallewoof	92ee4256f7	feature: raise on long sequence drop (#3321 ) * feature: raise on long sequence drop It is sometimes not desired that sequences are silently dropped from the dataset, especially when the dataset has been carefully crafted and pre-fitted for the training context. This would then suggest that an error occurred somewhere in the process. This feature adds a third value for excess_length_strategy called 'raise', which will raise a ValueError if a sequence is encountered that is too long and would have normally been dropped/truncated. * tests: add excess_length_strategy tests * doc: updated return value description for drop_long_seq_in_dataset * add @enable_hf_offline * fixed cfg modified after validate_config called * hf offline fix * fix tqdm desc when raise is used * test: added test for non-batched case * accidental code change revert * test: use pytest.raises * test: simplified drop_seq_len tests * test: moved excess_length_strat test to test_data.py --------- Co-authored-by: salman <salman.mohammadi@outlook.com>	2025-12-22 13:59:49 -05:00
Wing Lian	efeb5a4e41	fix check for fp8 capability (#3324 ) * fix check for fp8 capability * handle non-cuda compute * reduce concurrency of tests	2025-12-22 13:58:25 -05:00
VED	faaff6c792	allow users to set ndigits for rounding of metrics when logging (#3325 ) * METRIC_PRECISION-> 8 * use ndigits and move env getter to top of log function --------- Co-authored-by: Ved <ved.work2024@gmail.com> Co-authored-by: Wing Lian <wing@axolotl.ai>	2025-12-22 08:54:43 -05:00
Alexander Kozhevnikov	43cef27458	Fix typo in densemixer RuntimeError (#3327 ) [skip ci] It offers installing densemizer while it should be densemixer	2025-12-22 08:53:58 -05:00
Wing Lian	07c41a6c2a	fix preview docs failing due to running out of disk (#3326 ) [skip ci] * fix preview docs failing due to running out of disk * fix docs publish too	2025-12-19 11:34:55 -05:00
salman	bbd3486f57	Distributed Muon Optimizer (#3264 ) * init * working * updating configs * removing unneeded files * lint * comments * lint * fix regex match * bump contribs version * comments * fixing tests and imports * muon imports in test v2 * test cleanup * bump contribs version --------- Co-authored-by: Salman Mohammadi <“salman.mohammadi@outlook.com”>	2025-12-19 10:43:47 -05:00
VED	3750d7dd64	add liger support kernal for dpo (#3302 ) * add liger kernal 4 dpo * revert grpo changes,add support in dpo * revert grpo changes,add support in dpo * dpo_use_liger_kernal * fix liger_dpo --------- Co-authored-by: Ved <ved.work2024@gmail.com>	2025-12-18 11:11:06 -05:00
xzuyn	2197b0bf89	feat: cheap ppl metric (#3317 ) * Import math and compute perplexity from loss values * lint * coderabbit changes * lint * fix: add rounding to ppl --------- Co-authored-by: NanoCode012 <nano@axolotl.ai>	2025-12-18 09:02:41 -05:00
Seung Hyun Cho	3e51a680c2	fix: Fix evaluation loss in KD trainer (#3271 ) * fix: Fix evaluation loss in KD trainer * Fix v2 strategy super() call * fix: Add safety check for total_tokens in log method * fix: simplified num items and outputs return handling * fix: add missing model forward pass in compute_loss * refactor: Use Template Method pattern for chat template strategies * refactor: use pop(None) and remove v2 override * chore: lint --------- Co-authored-by: NanoCode012 <nano@axolotl.ai> Co-authored-by: Wing Lian <wing@axolotl.ai>	2025-12-17 13:40:36 -05:00
xzuyn	2cf254b4af	Add `peft_autocast_adapter_dtype` config option (#3311 ) [skip ci] * Add `peft_autocast_adapter_dtype` field to schema * Add `autocast_adapter_dtype` to `model_kwargs` * chore: docs --------- Co-authored-by: NanoCode012 <nano@axolotl.ai>	2025-12-17 10:09:39 -05:00
salman	83d4d97dcc	Add QAT NVFP4 configs for blogpost (#3280 ) [skip ci] * add configs for blogpost * fix configs * fixing baseline configs	2025-12-17 09:35:22 -05:00
NanoCode012	a1d07f42e4	Fix(misc): address PYTORCH_CUDA_ALLOC_CONF deprecate (#3313 ) * fix: leftover ministral docs changes * fix: pytorch_cuda_alloc_conf deprecation * fix: set old PYTORCH_CUDA_ALLOC_CONF env too * handle 2.9 separately --------- Co-authored-by: Wing Lian <wing@axolotl.ai>	2025-12-17 09:12:18 -05:00
Wing Lian	2a664dc8ad	support for xformers wheels for torch 2.9 (#3308 ) * support for xformers wheels for torch 2.9 * fix hf cache? * don't use hf cache from s3 * show disk free space in ci	2025-12-11 11:56:40 -05:00
NanoCode012	4ac78aa562	fix: update qwen3 jinja tokenization off a few tokens (#3295 ) * fix: update qwen3 jinja tokenization off a few tokens * fix: add note on tokenization issue * fix: pop last index for mistral tokenizer	2025-12-09 14:31:03 +07:00
VED	b3f4aa149f	fix bin size (#3307 ) * fix bin size * lint --------- Co-authored-by: Ved <ved.work2024@gmail.com>	2025-12-08 09:16:18 -05:00
salman	75b20fb66f	Save processor in quantizer CLI (#3290 )	2025-12-06 16:27:18 +00:00
NanoCode012	5992e607a2	fix: improve ministral3 docs to be clearer (#3300 ) * fix: improve ministral3 docs to be clearer * fix: title * chore: wording	2025-12-04 21:44:44 +07:00
NanoCode012	2b66ee189c	Feat: add ministral3 (#3297 ) * feat: add ministral and mistral3 * chore: lint * feat: update cce for ministral * fix: add vram usage * feat: update for release * fix: save_pretrained issue in v5 * fix: add instructions to use v5 branch * fix: add to multipack * fix: improve instructions * fix: add model to readme	2025-12-04 08:32:08 -05:00
NanoCode012	86d8cca149	Feat: add trinity by ArceeAI (#3292 )	2025-12-02 13:12:55 -05:00
NanoCode012	4a0f98e612	feat: upgrade liger to 0.6.4 (#3289 )	2025-12-02 09:16:23 -05:00
Yohan Na	c6ddcdd06a	feat: add exaone4 chat template and update enums (#3279 ) * feat: add exaone4 chat template and update enums * fix: handle first message as system or tools in exaone4 chat template * Update src/axolotl/utils/chat_templates/templates/exaone4.jinja Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> * fix: lint --------- Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> Co-authored-by: NanoCode012 <nano@axolotl.ai>	2025-12-01 15:52:45 +07:00
github-actions[bot]	7fb6a947d9	chore: update pre-commit hooks (#3287 ) Co-authored-by: SalmanMohammadi <25081738+SalmanMohammadi@users.noreply.github.com>	2025-12-01 15:03:14 +07:00
NanoCode012	b234532d9f	Feat: add peft_ensure_weight_tying (#3278 ) * feat: upgrade peft to 0.18.0 * feat: add peft_ensure_weight_tying * fix: default * chore: adjust kwarg per feedback	2025-11-28 18:54:48 +07:00
NanoCode012	a526647b31	Merge branch 'main' into feat/glm45	2025-11-28 13:41:25 +07:00
VED	8990ca3205	fix: removed unused "scikit-learn==1.4.2" (#3277 ) Co-authored-by: Ved <ved.work2024@gmail.com>	2025-11-24 13:48:53 +07:00
NanoCode012	006f226270	Feat: add Olmo3 (BC with Olmo and Olmo2) (#3275 ) * feat: update cce to include olmo family * chore: update docs following feedback * feat: add olmo3 config * fix: clarify 3 methods * chore: add olmo to readme	2025-11-24 10:21:31 +07:00
Wing Lian	0b635e69c5	build docker images for 2.9.x (#3273 )	2025-11-20 09:26:24 -05:00
Wing Lian	0d27e14e45	Torch 2.9.1 base images (#3268 ) * update torch 2.9.1 base images * update base dockerfile image check	2025-11-20 09:04:37 -05:00
NanoCode012	f5f21fb216	chore: update readme with latest updates (#3267 ) Some checks failed ci-cd / build-axolotl (<nil>, 126, 12.6.3, 3.11, 2.7.0) (push) Has been cancelled Details ci-cd / build-axolotl (<nil>, 128, 12.8.1, 3.11, 2.7.1) (push) Has been cancelled Details ci-cd / build-axolotl (<nil>, 128, 12.8.1, true, 3.11, 2.8.0) (push) Has been cancelled Details ci-cd / build-axolotl (vllm, 126, 12.6.3, 3.11, 2.7.1) (push) Has been cancelled Details publish pypi / Create Release (push) Has been cancelled Details ci-cd / build-axolotl-cloud (<nil>, 126, 12.6.3, 3.11, 2.7.0) (push) Has been cancelled Details ci-cd / build-axolotl-cloud (<nil>, 126, 12.6.3, <nil>, 3.11, 2.7.1) (push) Has been cancelled Details ci-cd / build-axolotl-cloud (<nil>, 128, 12.8.1, 3.11, 2.7.1) (push) Has been cancelled Details ci-cd / build-axolotl-cloud (<nil>, 128, 12.8.1, true, 3.11, 2.8.0) (push) Has been cancelled Details ci-cd / build-axolotl-cloud (vllm, 126, 12.6.3, 3.11, 2.7.1) (push) Has been cancelled Details ci-cd / build-axolotl-cloud-no-tmux (<nil>, 126, 12.6.3, <nil>, 3.11, 2.7.1) (push) Has been cancelled Details ci-cd / build-axolotl-cloud-no-tmux (<nil>, 128, 12.8.1, <nil>, 3.11, 2.8.0) (push) Has been cancelled Details ci-cd / build-axolotl-cloud-no-tmux (vllm, 126, 12.6.3, true, 3.11, 2.7.1) (push) Has been cancelled Details publish pypi / Upload release to PyPI (push) Has been cancelled Details	2025-11-18 14:45:21 +07:00
NanoCode012	4e55871112	feat: Add opt-out Telemetry (#3237 ) * initial telemetry manager impl * adding todo * updates * updates * progress on telemetry: config load, process, model load, train start / end, error tracking * update error file path sanitization function; adding more error tracking * updated sanitization logic, tests * adding runtime metrics (cpu + gpu memory, steps/s, etc.) * tests for runtime metrics telemetry and assoc. callback * small update / fix * simplifying path redaction * sleep on all ranks in distributed setting * adding back in base_model redaction w/ whitelist * fix * doc update * improved redaction, send system info during model config load telemetry, etc. * adding runtime metrics / system info additional accelerator support, etc. * adding runtime metrics / system info additional accelerator support, etc. * remove duplicate info * fixes * fix issue with tests in ci * distributed fix * opt-in version of telemetry * enable / disable logic update * docs fix * doc update * minor fixes * simplifying * slight changes * fix * lint * update posthog dep * coderabbit comments * fix: opt-in model * fix: increase time since last * fix: increase whitelist orgs * fix: posthog init and shutdown * fix: imports * fix: also check grad norm * fix: duplicate plugin_manager calls * fix: bad merge * chore: update docs * fix: cache process per comment * fix: error handling * fix: tests * Revert "fix: error handling" This reverts commit `22d1ea5755`. * fix: test telemetry error_handled bool * fix: revert test * chore: final doc fixes --------- Co-authored-by: Dan Saunders <danjsaund@gmail.com> Co-authored-by: Dan Saunders <dan@axolotl.ai>	2025-11-18 11:35:25 +07:00
Wing Lian	a6bafb55cb	upgrade datasets to 4.4.1 (#3266 ) * upgrade datasets * cleanup pip cache earlier * cleanup unused things from worker * also cleanup sdist	2025-11-14 09:52:14 -08:00
Wing Lian	0fbde69e9c	only push axolotl images, personal repo is deprecated (#3262 ) * only push axolotl images, personal repo is deprecated * cleanup	2025-11-14 07:50:03 -08:00
Wing Lian	301e22849f	upgrade to latest deepspeed and make sure latest tagged axolotl images are using torch 2.8.0 (#3261 )	2025-11-13 13:03:01 -05:00
VED	dcf24fd24e	feat: save checkpoint after training started (#3233 ) * add:config parameters for checkpoint * callback main * test file_type fix * lint * unit * simplify dict/obj handeling * Update src/axolotl/utils/schemas/dynamic_checkpoint.py Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> * Delete tests/e2e/integrations/__init__.py * remove hard code path in test * device check * lint * Update src/axolotl/utils/callbacks/dynamic_checkpoint.py Co-authored-by: NanoCode012 <kevinvong@rocketmail.com> * Update src/axolotl/utils/callbacks/dynamic_checkpoint.py Co-authored-by: NanoCode012 <kevinvong@rocketmail.com> * Update src/axolotl/utils/schemas/dynamic_checkpoint.py Co-authored-by: NanoCode012 <kevinvong@rocketmail.com> * lint-2 * remove: singal based checkpoints * lint * remove signal tests * add:is_main_process * lint * addis_d:istributed() for tests * remove nested is_main_process * Update src/axolotl/utils/schemas/dynamic_checkpoint.py Co-authored-by: Wing Lian <wing.lian@gmail.com> * Update src/axolotl/utils/schemas/dynamic_checkpoint.py Co-authored-by: Wing Lian <wing.lian@gmail.com> * add user_defined_filename --------- Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> Co-authored-by: NanoCode012 <kevinvong@rocketmail.com> Co-authored-by: Wing Lian <wing.lian@gmail.com>	2025-11-13 10:21:05 -05:00
NanoCode012	49b8107989	feat: add granite4 examples (#3256 ) [skip ci]	2025-11-13 10:19:16 -05:00
NanoCode012	9901ee5602	fix: voxtralprocessor broken (#3255 ) [skip ci] * fix: voxtralprocessor broken * chore: add todo * chore: wording	2025-11-13 10:18:42 -05:00
xzuyn	dd78f2e0cc	Fix: `warmup_steps: 0` & `warmup_ratio: 0` not disabling warmup (#3254 ) * fix unintentional falsy checks * chore: lint --------- Co-authored-by: NanoCode012 <nano@axolotl.ai>	2025-11-11 10:32:06 +07:00
Eduard Zl	b54f9c942b	_get_tools in ChatTemplateStrategy : function "parameters" can be dict or string (#3238 ) * When training of function calls, "tools" elements of a dataset can contain same parameter name but with different types. Datasets fails to load such training set. This fix allows "parameters" element of function call to be string( by running "json.dumps" in preparation of training data set). The _get_tools function will iterate over tool definitions, if "parameters" element is dict, it will keep that way, if it is a string, it will be converted to dict by invoking "json.loads" on string value. * feat: add doc on tool parameters json loading * feat: add tests for parameters json string --------- Co-authored-by: ezlotnik <eduard_zlotnik@intuit.com> Co-authored-by: NanoCode012 <nano@axolotl.ai>	2025-11-11 09:04:28 +07:00
NanoCode012	8069177284	Merge branch 'main' into feat/glm45	2025-11-10 21:41:05 +07:00
NanoCode012	11eb36585a	feat: add arg to enable dft in liger (#3125 ) * feat: add arg to enable dft in liger * feat: add tests use_token_scaling * fix: test * fix: move check to args	2025-11-10 21:37:47 +07:00
NanoCode012	d0c846fc5e	feat: add granitemoeshared and granitemoehybrid (#3158 )	2025-11-10 21:35:45 +07:00
Wing Lian	b5fcc2f14b	log cumulative total trained tokens (#3252 ) * log cumulative total trained tokens * use is_distributed helper	2025-11-07 16:04:00 -05:00
Wing Lian	b62eed8809	add openenv-core to requirements (#3251 )	2025-11-07 12:17:27 -05:00
VED	ed2e8cacd6	feat:openenv rollout_func (#3239 ) [skip ci] * feat:openenv rollout_func * chore lint * docs * add:docs processing_class * tests * lint	2025-11-07 08:51:40 -05:00
Lê Nam Khánh	80270a92fa	Fix typos in some files (#3250 ) [skip ci]	2025-11-07 08:21:20 -05:00
Wing Lian	bfdc9a8249	upgrade trl and other hf deps (#3249 ) * upgrade trl and other hf deps * skip simpo for now	2025-11-06 16:06:03 -05:00
salman	c37decb073	update pre-commit cadence (#3245 )	2025-11-04 13:43:40 +00:00
NanoCode012	01a346d86a	feat(example): add gpt-oss-safeguard docs (#3243 ) * feat(example): add gpt-oss-safeguard docs * fix: add doc on reasoning_effort	2025-11-04 07:39:21 +07:00
NanoCode012	26f05b6008	fix(example): set model_type to load for gemma3 text (#3242 ) * fix: set model_type to load for gemma3 text * chore: simplify * chore: unify	2025-11-04 07:35:07 +07:00
github-actions[bot]	ed58fa8a75	chore: update pre-commit hooks (#3244 )	2025-11-03 15:55:40 +00:00
Wing Lian	633afffacb	add torch 2.9.0 to ci (#3223 )	2025-10-30 18:50:26 -04:00
Wing Lian	4b1b4fa6d8	upgrade numpy (#3236 ) * upgrade numpy to 2.3.4 * bump contribs for numpy * fix vllm versions * bump numba * make sure psutil is installed * add psutil to cicd dockerfile jinja * lower dep versions of numba + numpy for vllm * bump datasets version * resolve pydantic conflict too	2025-10-30 10:03:24 -04:00
github-actions[bot]	0f7c886b7b	chore: update pre-commit hooks (#3222 ) [skip ci] Co-authored-by: djsaunde <1245942+djsaunde@users.noreply.github.com>	2025-10-29 18:09:46 -04:00
Wing Lian	a4b921135b	build cuda 13.0.0 base image with 2.9.0 (#3229 ) * build cuda 13.0.0 base image with 2.9.0 * upgrade causal-conv1d * 1.5.4 not in pypi yet * pin to 1.3.0 * use github release instead of pypi * split the logic for incompatible packages * fix bash in dockerfile	2025-10-29 18:07:29 -04:00
Wing Lian	98333e639a	upgrade trl to 0.24.0 and liger to 0.6.3 (#3230 ) * upgrade trl to 0.24.0 * fix reward collator init * use newer DataCollatorForPreference instead * DataCollatorForPreference doesn't use padding kwarg * fix input id labels * fix fbgemm-gpu version for pytorch versions * tweak pinned deps * transformers doesn't support hub 1.0 yet * upgrade liger dep to 0.6.3 * set TORCH_CUDA_ARCH_LIST correctly	2025-10-29 18:02:16 -04:00
NanoCode012	a28eb600e9	feat: add readme and better examples	2025-08-13 13:57:15 +07:00
NanoCode012	4b16f363bc	fix: move	2025-08-13 10:46:42 +07:00
NanoCode012	272a456ec0	fix: remove lora in fft config	2025-08-12 20:34:47 +07:00
NanoCode012	7e83268662	feat: add wip fft offload config	2025-08-12 20:34:47 +07:00
NanoCode012	b2a8c37a27	fix: use smaller model	2025-08-12 20:34:47 +07:00
NanoCode012	603166d9c5	feat: add example config	2025-08-12 20:34:47 +07:00
NanoCode012	e8c9517ac8	feat: add to multipack	2025-08-12 20:34:47 +07:00
NanoCode012	0bbad9202c	feat: add glm4moemoe to z3	2025-08-12 20:34:47 +07:00
NanoCode012	cb042e9775	feat: add cce for glm4_moe & deepseek v3	2025-08-12 20:32:46 +07:00