fix: log merge command once done

* fix: redact trackio and data_files

* fix: add new orgs to whitelist

* feat: add run id to logs for users to easily share

* fix: update to add more metrics

* fix: add missed experiment tracker

* chore: formatting in main

.github/workflows/base.yml

@@ -51,6 +51,14 @@ jobs:
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-base"
             platforms: "linux/amd64,linux/arm64"
+          - cuda: "129"
+            cuda_version: 12.9.1
+            cudnn_version: ""
+            python_version: "3.12"
+            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"
+            platforms: "linux/amd64,linux/arm64"
           - cuda: "130"
             cuda_version: 13.0.0
             cudnn_version: ""
@@ -59,6 +67,14 @@ jobs:
             torch_cuda_arch_list: "9.0+PTX"
             dockerfile: "Dockerfile-base"
             platforms: "linux/amd64,linux/arm64"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.9.1
+            torch_cuda_arch_list: "9.0+PTX"
+            dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
 #          - cuda: "128"
 #            cuda_version: 12.8.1
 #            cudnn_version: ""
@@ -141,6 +157,14 @@ jobs:
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
             platforms: "linux/amd64,linux/arm64"
+          - cuda: "129"
+            cuda_version: 12.9.1
+            cudnn_version: ""
+            python_version: "3.12"
+            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"
+            platforms: "linux/amd64,linux/arm64"
           - cuda: "130"
             cuda_version: 13.0.0
             cudnn_version: ""
@@ -149,6 +173,14 @@ jobs:
             torch_cuda_arch_list: "9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
             platforms: "linux/amd64,linux/arm64"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.9.1
+            torch_cuda_arch_list: "9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.github/workflows/main.yml

@@ -34,6 +34,12 @@ jobs:
             axolotl_extras:
             platforms: "linux/amd64,linux/arm64"
             is_latest: true
+          - cuda: 129
+            cuda_version: 12.9.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
           - cuda: 130
             cuda_version: 13.0.0
             python_version: "3.11"
@@ -106,6 +112,12 @@ jobs:
             axolotl_extras:
             is_latest: true
             platforms: "linux/amd64,linux/arm64"
+          - cuda: 129
+            cuda_version: 12.9.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
           - cuda: 130
             cuda_version: 13.0.0
             python_version: "3.11"

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

@@ -35,14 +35,19 @@ jobs:
             pytorch: 2.8.0
             axolotl_extras: fbgemm-gpu
             num_gpus: 2
-            nightly_build: "true"
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
             pytorch: 2.9.1
-            axolotl_extras: fbgemm-gpu
+            axolotl_extras: "fbgemm-gpu"
             num_gpus: 2
-            nightly_build: "true"
+          - cuda: 129
+            cuda_version: 12.9.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras: "fbgemm-gpu"
+            num_gpus: 2
+            dockerfile: "Dockerfile-uv.jinja"
           - cuda: 130
             cuda_version: 13.0.0
             python_version: "3.11"
@@ -50,7 +55,6 @@ jobs:
             axolotl_extras:
 #            axolotl_extras: fbgemm-gpu
             num_gpus: 2
-            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:
@@ -72,8 +76,8 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
+          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run -m cicd.multigpu

.github/workflows/pypi.yml

@@ -40,7 +40,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip3 install wheel packaging==23.2
+          pip3 install wheel packaging==26.0
           pip3 install --no-build-isolation -e .
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 
@@ -48,9 +48,9 @@ jobs:
         id: tag
         run: echo ::set-output name=TAG_NAME::$(echo $GITHUB_REF | cut -d / -f 3)
 
-      - name: Update version in setup.py
+      - name: Update version in VERSION file
         run: |
-          sed -i -E 's/version="([0-9.]+)",/version="${{ steps.tag.outputs.TAG_NAME }}",/g' setup.py
+          echo "${{ steps.tag.outputs.TAG_NAME }}" | sed 's/^v//' > VERSION
 
       - name: Build a source dist
         run: |

.github/workflows/tests-nightly.yml

@@ -48,7 +48,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |

.github/workflows/tests.yml

@@ -54,8 +54,13 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.11"]
+        python_version: ["3.11", "3.12"]
         pytorch_version: ["2.8.0", "2.9.0", "2.9.1"]
+        exclude:
+          - python_version: "3.12"
+            pytorch_version: "2.8.0"
+          - python_version: "3.12"
+            pytorch_version: "2.9.0"
     timeout-minutes: 20
 
     steps:
@@ -82,7 +87,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |
@@ -110,10 +115,10 @@ jobs:
 
       - name: Pre-Download dataset fixture
         run: |
-          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+          hf download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
 
       - name: Show HF cache
-        run: hf cache scan
+        run: hf cache ls
 
       - name: Run tests
         run: |
@@ -127,7 +132,7 @@ jobs:
           pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml
 
       - name: Show HF cache
-        run: hf cache scan
+        run: hf cache ls
 
       - name: Upload coverage to Codecov
         uses: codecov/codecov-action@v5
@@ -144,8 +149,13 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.11"]
+        python_version: ["3.11", "3.12"]
         pytorch_version: ["2.8.0", "2.9.0", "2.9.1"]
+        exclude:
+          - python_version: "3.12"
+            pytorch_version: "2.8.0"
+          - python_version: "3.12"
+            pytorch_version: "2.9.0"
     timeout-minutes: 20
 
     steps:
@@ -172,7 +182,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 setuptools_scm build wheel psutil
+          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 setuptools_scm build wheel psutil
 
       - name: Install PyTorch
         run: |
@@ -200,7 +210,7 @@ jobs:
           axolotl --help
 
       - name: Show HF cache
-        run: hf cache scan
+        run: hf cache ls
 
       - name: Run tests
         run: |
@@ -209,10 +219,10 @@ jobs:
           pytest -v --durations=10 tests/cli/
 
       - name: Show HF cache
-        run: hf cache scan
+        run: hf cache ls
 
   gate-skip-e2e:
-    needs: [pre-commit, pytest, pytest-sdist]
+    needs: [pre-commit]
     runs-on: ubuntu-latest
     outputs:
       skip: ${{ steps.compute.outputs.skip }}
@@ -248,16 +258,16 @@ jobs:
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
-    needs: [pre-commit, pytest, pytest-sdist, gate-skip-e2e]
+    needs: [pre-commit, pytest]
 
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.8.0
+          - cuda: 129
+            cuda_version: 12.9.1
+            python_version: "3.12"
+            pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
             dockerfile: "Dockerfile-uv.jinja"
@@ -359,9 +369,9 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
+          - cuda: 129
+            cuda_version: 12.9.1
+            python_version: "3.12"
             pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:

.gitignore

@@ -193,3 +193,6 @@ out/
 
 # scm auto-versioning
 src/axolotl/_version.py
+
+# macOS
+.DS_Store

.runpod/README.md

@@ -123,7 +123,7 @@ datasets:
 | --------------------------------- | -------------------------- | ----------------------------------- |
 | `dataset_prepared_path`           | `"data/last_run_prepared"` | Path for prepared dataset           |
 | `push_dataset_to_hub`             | `""`                       | Push dataset to HF hub              |
-| `dataset_processes`               | `4`                        | Number of preprocessing processes   |
+| `dataset_num_proc`                | `4`                        | Number of preprocessing processes   |
 | `dataset_keep_in_memory`          | `false`                    | Keep dataset in memory              |
 | `shuffle_merged_datasets`         | `true`                     | Shuffle merged datasets             |
 | `shuffle_before_merging_datasets` | `false`                    | Shuffle each dataset before merging |

.runpod/src/config/config.yaml

@@ -39,7 +39,6 @@
 #     type: # linear | dynamic
 #     factor: # float
 
-
 # # Whether you are training a 4-bit GPTQ quantized model
 # gptq: true
 # gptq_groupsize: 128 # group size
@@ -107,7 +106,7 @@
 # push_dataset_to_hub: # repo path
 # # The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`
 # # if not set.
-# dataset_processes: # defaults to os.cpu_count() if not set
+# dataset_num_proc: # defaults to os.cpu_count() if not set
 # # push checkpoints to hub
 # hub_model_id: # repo path to push finetuned model
 # # how to push checkpoints to hub
@@ -224,9 +223,6 @@
 # eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
 # eval_table_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
 
-# # Save model as safetensors (require safetensors package)
-# save_safetensors:
-
 # # Whether to mask out or include the human's prompt from the training labels
 # train_on_inputs: false
 # # Group similarly sized data to minimize padding.
@@ -352,8 +348,6 @@
 # # Allow overwrite yml config using from cli
 # strict:
 
-
-
 base_model: ${BASE_MODEL}
 base_model_ignore_patterns: ${BASE_MODEL_IGNORE_PATTERNS}
 base_model_config: ${BASE_MODEL_CONFIG}
@@ -412,7 +406,7 @@ chat_template_jinja: ${CHAT_TEMPLATE_JINJA}
 default_system_message: ${DEFAULT_SYSTEM_MESSAGE}
 dataset_prepared_path: ${DATASET_PREPARED_PATH}
 push_dataset_to_hub: ${PUSH_DATASET_TO_HUB}
-dataset_processes: ${DATASET_PROCESSES}
+dataset_num_proc: ${DATASET_NUM_PROC}
 dataset_keep_in_memory: ${DATASET_KEEP_IN_MEMORY}
 hub_model_id: ${HUB_MODEL_ID}
 hub_strategy: ${HUB_STRATEGY}
@@ -512,7 +506,6 @@ profiler_steps: ${PROFILER_STEPS}
 loss_watchdog_threshold: ${LOSS_WATCHDOG_THRESHOLD}
 loss_watchdog_patience: ${LOSS_WATCHDOG_PATIENCE}
 
-save_safetensors: ${SAVE_SAFETENSORS}
 train_on_inputs: ${TRAIN_ON_INPUTS}
 group_by_length: ${GROUP_BY_LENGTH}
 gradient_checkpointing: ${GRADIENT_CHECKPOINTING}

README.md

@@ -88,7 +88,7 @@ Features:
 #### Using pip
 
 ```bash
-pip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install -U packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 
 # Download example axolotl configs, deepspeed configs

VERSION

@@ -0,0 +1 @@
+0.15.0.dev0

_quarto.yml

@@ -251,7 +251,6 @@ website:
                 - docs/models/olmo3.qmd
                 - docs/models/trinity.qmd
                 - docs/models/arcee.qmd
-                - docs/models/mistral.qmd
                 - section: "Ministral3"
                   contents:
                     - docs/models/ministral3.qmd
@@ -266,6 +265,7 @@ website:
                 - docs/models/mistral-small.qmd
                 - docs/models/voxtral.qmd
                 - docs/models/devstral.qmd
+                - docs/models/mistral.qmd
                 - docs/models/llama-4.qmd
                 - docs/models/llama-2.qmd
                 - docs/models/qwen3-next.qmd
@@ -320,6 +320,7 @@ website:
             - docs/multipack.qmd
             - docs/mixed_precision.qmd
             - docs/optimizers.qmd
+            - docs/attention.qmd
 
         - section: "Advanced Features"
           contents:

cicd/Dockerfile-uv.jinja

@@ -31,7 +31,7 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
-RUN uv pip install packaging==23.2 setuptools==75.8.0
+RUN uv pip install packaging==26.0 setuptools==75.8.0
 RUN uv pip install torchvision
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         uv pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \

cicd/Dockerfile.jinja

@@ -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 psutil
+RUN pip install packaging==26.0 setuptools==75.8.0 psutil
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

cicd/multigpu.py

@@ -17,7 +17,8 @@ template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
 template_env = jinja2.Environment(
     loader=template_loader, autoescape=select_autoescape()
 )
-df_template = template_env.get_template("Dockerfile.jinja")
+dockerfile = os.environ.get("E2E_DOCKERFILE", "Dockerfile.jinja")
+df_template = template_env.get_template(dockerfile)
 
 df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
@@ -27,8 +28,11 @@ df_args = {
     "CUDA": os.environ.get("CUDA", "126"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
+    "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
     "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
     "HF_HOME": "/workspace/data/huggingface-cache/hub",
+    "PYTHONUNBUFFERED": os.environ.get("PYTHONUNBUFFERED", "1"),
+    "DEEPSPEED_LOG_LEVEL": os.environ.get("DEEPSPEED_LOG_LEVEL", "WARNING"),
 }
 
 dockerfile_contents = df_template.render(**df_args)

cicd/multigpu.sh

@@ -2,7 +2,7 @@
 set -e
 
 # Only run two tests at a time to avoid OOM on GPU (with coverage collection)
-pytest -v --durations=10 -n2 --maxfail=4 \
+pytest -v --durations=10 -n2 --maxfail=3 \
   --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/ \
   --ignore=/workspace/axolotl/tests/e2e/multigpu/patched/ \
   /workspace/axolotl/tests/e2e/multigpu/ \

docker/Dockerfile-base

@@ -43,7 +43,7 @@ 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 psutil && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==26.0 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 && \
     python3 -m pip cache purge
 

docker/Dockerfile-base-nightly

@@ -30,7 +30,7 @@ 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==26.0 setuptools==75.8.0 wheel && \
     python3 -m pip install --no-cache-dir -U torch --extra-index-url https://download.pytorch.org/whl/nightly/cu$CUDA && \
     python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
     python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" && \

docs/amd_hpc.qmd

@@ -86,7 +86,7 @@ export HF_DATASETS_OFFLINE=1
 Download a base model using the Hugging Face CLI:
 
 ```bash
-huggingface-cli download meta-llama/Meta-Llama-3.1-8B --local-dir ~/hfdata/llama3.1-8B
+hf download meta-llama/Meta-Llama-3.1-8B --local-dir ~/hfdata/llama3.1-8B
 ```
 
 ### 10. Create Axolotl Configuration

docs/attention.qmd

@@ -0,0 +1,140 @@
+---
+title: Attention
+description: Supported attention modules in Axolotl
+---
+
+## SDP Attention
+
+This is the default built-in attention in PyTorch.
+
+```yaml
+sdp_attention: true
+```
+
+For more details: [PyTorch docs](https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)
+
+## Flash Attention 2
+
+Uses efficient kernels to compute attention.
+
+```yaml
+flash_attention: true
+```
+
+For more details: [Flash Attention](https://github.com/Dao-AILab/flash-attention/)
+
+### Nvidia
+
+Requirements: Ampere, Ada, or Hopper GPUs
+
+Note: For Turing GPUs or lower, please use other attention methods.
+
+```bash
+pip install flash-attn --no-build-isolation
+```
+
+::: {.callout-tip}
+
+If you get `undefined symbol` while training, ensure you installed PyTorch prior to Axolotl. Alternatively, try reinstall or downgrade a version.
+
+:::
+
+#### Flash Attention 3
+
+Requirements: Hopper only and CUDA 12.8 (recommended)
+
+```bash
+git clone https://github.com/Dao-AILab/flash-attention.git
+cd flash-attention/hopper
+
+python setup.py install
+```
+
+### AMD
+
+Requirements: ROCm 6.0 and above.
+
+See [Flash Attention AMD docs](https://github.com/Dao-AILab/flash-attention/tree/main?tab=readme-ov-file#amd-rocm-support).
+
+## Flex Attention
+
+A flexible PyTorch API for attention used in combination with `torch.compile`.
+
+```yaml
+flex_attention: true
+
+# recommended
+torch_compile: true
+```
+
+::: {.callout-note}
+
+We recommend using latest stable version of PyTorch for best performance.
+
+:::
+
+For more details: [PyTorch docs](https://pytorch.org/blog/flexattention/)
+
+## SageAttention
+
+Attention kernels with QK Int8 and PV FP16 accumulator.
+
+```yaml
+sage_attention: true
+```
+
+Requirements: Ampere, Ada, or Hopper GPUs
+
+```bash
+pip install sageattention==2.2.0 --no-build-isolation
+```
+
+::: {.callout-warning}
+
+Only LoRA/QLoRA recommended at the moment. We found loss drop to 0 for full finetuning. See [GitHub Issue](https://github.com/thu-ml/SageAttention/issues/198).
+
+:::
+
+For more details: [Sage Attention](https://github.com/thu-ml/SageAttention)
+
+::: {.callout-note}
+
+We do not support SageAttention 3 at the moment. If you are interested on adding this or improving SageAttention implementation, please make an Issue.
+
+:::
+
+
+## xFormers
+
+```yaml
+xformers_attention: true
+```
+
+::: {.callout-tip}
+
+We recommend using with Turing GPUs or below (such as on Colab).
+
+:::
+
+For more details: [xFormers](https://github.com/facebookresearch/xformers)
+
+## Shifted Sparse Attention
+
+::: {.callout-warning}
+
+We plan to deprecate this! If you use this feature, we recommend switching to methods above.
+
+:::
+
+Requirements: LLaMA model architecture
+
+```yaml
+flash_attention: true
+s2_attention: true
+```
+
+::: {.callout-tip}
+
+No sample packing support!
+
+:::

docs/installation.qmd

@@ -165,7 +165,7 @@ We recommend using WSL2 (Windows Subsystem for Linux) or Docker.
    ```
 4. (Optional) Login to Hugging Face:
    ```{.bash}
-   huggingface-cli login
+   hf auth login
    ```
 
 ## Troubleshooting {#sec-troubleshooting}

docs/lora_optims.qmd

@@ -89,6 +89,10 @@ lora_o_kernel: true
 Currently, LoRA kernels are not supported for RLHF training, only SFT.
 :::
 
+::: {.callout-warning}
+LoRA kernels do not support remote modeling code.
+:::
+
 ## Requirements
 
 - One or more NVIDIA or AMD GPUs (in order to use the Triton kernels)

docs/multimodal.qmd

@@ -19,6 +19,7 @@ format:
 - [Gemma-3n](#sec-gemma-3n)
 - [Qwen2-VL](#sec-qwen2-vl)
 - [Qwen2.5-VL](#sec-qwen25-vl)
+- [GLM-4.6V](#sec-glm-4-6v)
 - [SmolVLM2](#sec-smolvlm2)
 - [LFM2-VL](#sec-lfm2-vl)
 - [Intern-VL](#sec-intern-vl)
@@ -183,6 +184,18 @@ base_model: Qwen/Qwen3-VL-4B-Instruct
 chat_template: qwen2_vl  # same as qwen2-vl
 ```
 
+### GLM-4.6V {#sec-glm-4-6v}
+
+Both GLM-4.6V (106B MoE) and GLM-4.6V-Flash (9B) are supported.
+
+```yaml
+# GLM-4.6V (106B MoE version)
+base_model: zai-org/GLM-4.6V
+
+# OR GLM-4.6V-Flash (9B version)
+base_model: zai-org/GLM-4.6V-Flash
+```
+
 ### SmolVLM2 {#sec-smolvlm2}
 
 ::: {.callout-tip}

docs/rlhf.qmd

@@ -17,6 +17,7 @@ feedback. Various methods include, but not limited to:
 - [Kahneman-Tversky Optimization (KTO)](#kto)
 - [Odds Ratio Preference Optimization (ORPO)](#orpo)
 - [Group Relative Policy Optimization (GRPO)](#grpo)
+- [Group Reward-Decoupled Policy Optimization (GDPO)](#gdpo)
 
 
 ## RLHF using Axolotl
@@ -720,6 +721,102 @@ trl:
 
 For more information, see [GRPO docs](https://huggingface.co/docs/trl/v0.17.0/en/grpo_trainer#loss-types).
 
+### GDPO
+
+GDPO (Group Reward-Decoupled Policy Optimization) extends GRPO for multi-reward training. It addresses the **reward advantage collapse** problem by normalizing each reward function independently before combining them.
+
+::: {.callout-tip}
+Use GDPO when training with multiple reward functions. For single reward, GRPO and GDPO produce equivalent results.
+:::
+
+Paper: [https://arxiv.org/pdf/2501.05242](https://arxiv.org/pdf/2501.05242)
+
+GDPO uses TRL's native `multi_objective_aggregation` parameter under the hood. When you set `rl: gdpo`, axolotl automatically configures TRL to use `normalize_then_sum` aggregation.
+
+```yaml
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+
+vllm:
+    host: 0.0.0.0
+    port: 8000
+    tensor_parallel_size: 2
+    gpu_memory_utilization: 0.85
+
+rl: gdpo
+
+trl:
+    beta: 0.001
+    max_completion_length: 256
+    use_vllm: true
+    num_generations: 4
+    reward_funcs:
+        - rewards.format_reward
+        - rewards.correctness_reward
+    reward_weights: [1.0, 2.0]
+
+datasets:
+    - path: openai/gsm8k
+      name: main
+      type: rewards.oai_gsm8k_transform
+```
+
+You can also use GRPO with explicit aggregation control:
+
+```yaml
+rl: grpo
+trl:
+    multi_objective_aggregation: normalize_then_sum  # GDPO behavior
+    # or: sum_then_normalize  # Default GRPO behavior
+```
+
+#### GDPO vs GRPO
+
+| Aspect | GRPO | GDPO |
+|--------|------|------|
+| **Aggregation** | `sum_then_normalize` | `normalize_then_sum` |
+| **Multi-reward** | May collapse advantages | Preserves reward signals |
+| **Single reward** | Standard behavior | Equivalent to GRPO |
+
+#### Why GDPO?
+
+When using multiple rewards with GRPO, different reward combinations can produce identical advantages:
+
+```
+# Example: format + correctness rewards
+[format=0, correct=3] → sum=3
+[format=1, correct=2] → sum=3  ← GRPO sees these as equal!
+[format=2, correct=1] → sum=3
+[format=3, correct=0] → sum=3
+```
+
+GDPO normalizes each reward independently, preserving their relative differences.
+
+#### Reward Functions
+
+GDPO uses the same reward function format as GRPO:
+
+```python
+# rewards.py
+def format_reward(completions, **kwargs) -> list[float]:
+    return [1.0 if len(c) > 10 else 0.0 for c in completions]
+
+def correctness_reward(completions, answers, **kwargs) -> list[float]:
+    rewards = []
+    for completion, answer in zip(completions, answers):
+        # Your scoring logic here
+        rewards.append(score)
+    return rewards
+```
+
+#### Sequence Parallelism
+
+GDPO supports sequence parallelism for long-context training:
+
+```yaml
+rl: gdpo
+context_parallel_size: 2
+```
+
 ### SimPO
 
 SimPO uses [CPOTrainer](https://huggingface.co/docs/trl/main/en/cpo_trainer) but with alternative loss function.

examples/apertus/README.md

@@ -15,7 +15,7 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 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 packaging==26.0 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

examples/arcee/README.md

@@ -17,7 +17,7 @@ Thanks to the team at Arcee.ai for using Axolotl in supervised fine-tuning the A
 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 packaging==26.0 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

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@318b7e2\""
+    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@0d4ce4b\""
    ]
   },
   {

examples/devstral/README.md

@@ -16,7 +16,7 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 

examples/eaft/eaft-example.yml

@@ -0,0 +1,77 @@
+base_model: google/gemma-3-1b-it
+
+model_type: Gemma3ForCausalLM
+cls_model_config: Gemma3TextConfig
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+chat_template: gemma3
+eot_tokens:
+  - <end_of_turn>
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: cgato/SlimOrcaDedupCleaned
+    type: chat_template
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path:
+val_set_size: 0
+output_dir: ./outputs/eaft-gemma-3-1b
+
+use_eaft: true
+eaft_alpha: 1.0
+eaft_k: 20
+
+sequence_len: 1024
+sample_packing: false
+
+adapter:
+lora_model_dir:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+eval_batch_size: 1
+max_steps: 1000
+evaluation_strategy: "no"
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-5
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+debug:
+deepspeed:
+fsdp:
+fsdp_config:
+special_tokens:

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

@@ -1,8 +1,7 @@
 base_model: google/gemma-3-4b-it
 
-# Need to set else transformers tries to load vision too
-model_type: Gemma3ForCausalLM
-cls_model_config: Gemma3TextConfig
+plugins:
+  - axolotl.integrations.gemma3.Gemma3TextFromMultimodalPlugin
 
 load_in_4bit: true
 
@@ -30,7 +29,6 @@ lora_model_dir:
 sequence_len: 2048
 sample_packing: true
 
-
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0

examples/gemma3n/README.md

@@ -10,7 +10,7 @@ Gemma-3n is a family of multimodal models from Google found on [HuggingFace](htt
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 

examples/glm46v/README.md

@@ -0,0 +1,44 @@
+# Finetune GLM-4.6V with Axolotl
+
+GLM-4.6V is a family of vision-language models from ZhipuAI found on [HuggingFace](https://huggingface.co/zai-org/GLM-4.6V). This guide shows how to fine-tune it with Axolotl for vision-language tasks.
+
+
+
+## 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. Run the fine-tuning:
+
+    glm-4-6v-flash(9B)
+    ```bash
+    axolotl train examples/glm46v/glm-4-6v-flash-qlora.yaml
+    ```
+
+Let us know how it goes. Happy finetuning! 🚀
+
+## Tips
+
+- Vision datasets should follow the format described in the [multimodal docs](https://docs.axolotl.ai/docs/multimodal.html#dataset-format)
+- 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 in the [dataset loading docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+
+## Supported Models
+
+- **GLM-4.6V**: Full vision-language model (`zai-org/GLM-4.6V`)
+- **GLM-4.6V-Flash**: Faster variant (`zai-org/GLM-4.6V-Flash`)
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Related Resources
+
+- [ZhipuAI GLM-4.6V](https://huggingface.co/zai-org/GLM-4.6V)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/glm46v/glm-4-6v-flash-ddp.yaml

@@ -0,0 +1,53 @@
+base_model: zai-org/GLM-4.6V-Flash
+trust_remote_code: true
+
+processor_type: AutoProcessor
+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
+ddp_find_unused_parameters: true
+
+output_dir: ./outputs/glm-4-6v-flash-qlora
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+adapter: qlora
+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
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+logging_steps: 1
+sdp_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 0
+saves_per_epoch: 1
+weight_decay: 0.0

examples/glm46v/glm-4-6v-flash-qlora.yaml

@@ -0,0 +1,50 @@
+base_model: zai-org/GLM-4.6V-Flash
+trust_remote_code: true
+
+processor_type: AutoProcessor
+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
+
+output_dir: ./outputs/glm-4-6v-flash-qlora
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+adapter: qlora
+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
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+logging_steps: 1
+sdp_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 0
+saves_per_epoch: 1
+weight_decay: 0.0

examples/gpt-oss/README.md

@@ -14,7 +14,7 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 

examples/granite4/README.md

@@ -15,7 +15,7 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 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 packaging==26.0 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

examples/hunyuan/README.md

@@ -13,7 +13,7 @@ Tencent released a family of opensource models called HunYuan with varying param
 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 packaging==26.0 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

examples/jamba/qlora_fsdp_large.yaml

@@ -19,7 +19,6 @@ datasets:
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: jamba-large-fsdp-qlora-ft
-save_safetensors: true
 adapter: qlora
 sequence_len: 2048
 sample_packing: true

examples/llama-3/qlora-1b-gdpo.yaml

@@ -0,0 +1,68 @@
+base_model: meta-llama/Llama-3.2-1B-Instruct
+
+chat_template: llama3
+
+rl: gdpo
+
+trl:
+  beta: 0.001
+  max_completion_length: 128
+  num_generations: 2
+  temperature: 0.7
+  top_p: 0.95
+
+  use_vllm: false
+
+
+  multi_objective_aggregation: normalize_then_sum
+
+  reward_funcs:
+    - rwd.format_reward
+    - rwd.correctness_reward
+  reward_weights: [1.0, 2.0]
+
+  log_completions: true
+  num_completions_to_print: 3
+  scale_rewards: true
+
+datasets:
+  - path: openai/gsm8k
+    name: main
+    split: train[:1000]
+    type: rwd.gsm8k_transform
+
+val_set_size: 0.0
+output_dir: ./outputs/llama3-gdpo-out
+
+sequence_len: 512
+sample_packing: false
+pad_to_sequence_len: false
+
+gradient_accumulation_steps: 8
+micro_batch_size: 1
+num_epochs: 1
+max_steps: 100
+
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-5
+weight_decay: 0.01
+warmup_steps: 10
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+flash_attention: true
+logging_steps: 1
+save_steps: 50
+save_safetensors: true
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+
+seed: 42

examples/llama-3/qlora-fsdp-405b.yaml

@@ -12,7 +12,6 @@ datasets:
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: ./outputs/out/qlora-llama3_1-405b
-save_safetensors: true
 
 adapter: qlora
 

examples/magistral/README.md

@@ -14,7 +14,7 @@ Thanks to the team at MistralAI for giving us early access to prepare for these
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.7.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 

examples/mamba/config.yml

@@ -47,6 +47,5 @@ saves_per_epoch: 1
 weight_decay: 0.0
 special_tokens:
 tokens:
-save_safetensors: False
 
 # save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Gemma3-12B_baseline.yml

@@ -1,12 +1,11 @@
 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.gemma3.Gemma3TextFromMultimodalPlugin
   - axolotl.integrations.liger.LigerPlugin
 
 liger_rope: true

examples/qat_nvfp4/Gemma3-12B_qat.yml

@@ -7,6 +7,7 @@ load_in_4bit: false
 strict: false
 
 plugins:
+  - axolotl.integrations.gemma3.Gemma3TextFromMultimodalPlugin
   - axolotl.integrations.liger.LigerPlugin
 
 liger_rope: true

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

@@ -1,12 +1,11 @@
 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.gemma3.Gemma3TextFromMultimodalPlugin
   - axolotl.integrations.liger.LigerPlugin
 
 liger_rope: true

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

@@ -7,6 +7,7 @@ load_in_4bit: false
 strict: false
 
 plugins:
+  - axolotl.integrations.gemma3.Gemma3TextFromMultimodalPlugin
   - axolotl.integrations.liger.LigerPlugin
 
 liger_rope: true

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

@@ -1,12 +1,11 @@
 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.gemma3.Gemma3TextFromMultimodalPlugin
   - axolotl.integrations.liger.LigerPlugin
 
 liger_rope: true

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

@@ -7,6 +7,7 @@ load_in_4bit: false
 strict: false
 
 plugins:
+  - axolotl.integrations.gemma3.Gemma3TextFromMultimodalPlugin
   - axolotl.integrations.liger.LigerPlugin
 
 liger_rope: true

examples/qwen3-next/README.md

@@ -15,7 +15,7 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 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 packaging==26.0 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

examples/voxtral/README.md

@@ -12,7 +12,7 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 

pyproject.toml

@@ -1,5 +1,5 @@
 [build-system]
-requires = ["setuptools>=64", "wheel", "setuptools_scm>=8", "packaging==23.2"]
+requires = ["setuptools>=64", "wheel", "setuptools_scm>=8", "packaging==26.0"]
 build-backend = "setuptools.build_meta"
 
 [project]
@@ -24,6 +24,9 @@ Repository = "https://github.com/axolotl-ai-cloud/axolotl.git"
 py-modules = ["setuptools_axolotl_dynamic_dependencies"]
 include-package-data = true
 
+[tool.setuptools.dynamic]
+version = { file = "VERSION" }
+
 [tool.setuptools.cmdclass]
 build_py = "setuptools_axolotl_dynamic_dependencies.BuildPyCommand"
 
@@ -57,3 +60,6 @@ indent-style = "space"
 skip-magic-trailing-comma = false
 line-ending = "auto"
 docstring-code-format = false
+
+[tool.uv.extra-build-dependencies]
+axolotl = ["huggingface_hub"]

requirements.txt

@@ -8,18 +8,18 @@ xformers>=0.0.23.post1
 liger-kernel==0.6.4
 # END section
 
-packaging==23.2
-
-huggingface_hub>=0.36.0
+packaging==26.0
+huggingface_hub>=1.1.7
 peft>=0.18.1
 tokenizers>=0.22.1
-transformers==4.57.6
+transformers==5.0.0
 accelerate==1.12.0
 datasets==4.5.0
 deepspeed>=0.18.3
-trl==0.25.1
+trl==0.27.1
 hf_xet==1.2.0
 kernels==0.11.5
+
 trackio>=0.13.0
 typing-extensions>=4.15.0
 
@@ -72,4 +72,4 @@ axolotl-contribs-mit==0.0.6
 # telemetry
 posthog==6.7.11
 
-mistral-common==1.8.6
+mistral-common==1.8.8

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@318b7e2"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@0d4ce4b"'
 )

scripts/merge_gemma3_multimodal_weights.py

@@ -0,0 +1,225 @@
+"""Merge trained text-only Gemma3 weights back into a full multimodal checkpoint.
+
+After training with the Gemma3TextFromMultimodalPlugin, the saved checkpoint
+contains only the language model weights (with ``model.language_model.*``
+prefix, reversed by transformers v5's key_mapping on save).
+
+This script reconstructs a full ``Gemma3ForConditionalGeneration`` checkpoint by
+combining the trained language model weights with the original vision tower and
+projector weights from the base multimodal model.
+
+Usage::
+
+    python scripts/merge_gemma3_multimodal_weights.py \\
+        --original-model google/gemma-3-4b-it \\
+        --trained-model /path/to/trained/output \\
+        --output-dir /path/to/merged
+"""
+
+import argparse
+import json
+import logging
+from pathlib import Path
+
+import torch
+from huggingface_hub import split_torch_state_dict_into_shards
+from safetensors.torch import load_file, save_file
+from transformers import AutoConfig
+
+LOG = logging.getLogger(__name__)
+
+
+def collect_safetensors(model_dir: Path) -> dict[str, torch.Tensor]:
+    """Load and merge all safetensors shard files in a directory."""
+    shard_files = sorted(model_dir.glob("*.safetensors"))
+    if not shard_files:
+        raise FileNotFoundError(f"No safetensors files found in {model_dir}")
+
+    state_dict: dict[str, torch.Tensor] = {}
+    for shard in shard_files:
+        LOG.info("Loading %s", shard.name)
+        state_dict.update(load_file(str(shard)))
+    return state_dict
+
+
+def merge(
+    original_model: str,
+    trained_model: str,
+    output_dir: str,
+    *,
+    trust_remote_code: bool = False,
+) -> None:
+    original_path = Path(original_model)
+    trained_path = Path(trained_model)
+    out_path = Path(output_dir)
+    out_path.mkdir(parents=True, exist_ok=True)
+
+    # 1. Load the original multimodal checkpoint
+    LOG.info("Loading original multimodal weights from %s", original_model)
+    if original_path.is_dir():
+        original_sd = collect_safetensors(original_path)
+    else:
+        from huggingface_hub import snapshot_download
+
+        cached = Path(
+            snapshot_download(original_model, allow_patterns=["*.safetensors"])
+        )
+        original_sd = collect_safetensors(cached)
+
+    # 2. Load trained text-only weights (already reversed to model.language_model.* by
+    #    transformers v5 key_mapping on save)
+    LOG.info("Loading trained text-only weights from %s", trained_model)
+    trained_sd = collect_safetensors(trained_path)
+
+    # 3. Classify original keys
+    lang_keys = {k for k in original_sd if k.startswith("model.language_model.")}
+    vision_keys = {k for k in original_sd if k.startswith("model.vision_tower.")}
+    projector_keys = {
+        k for k in original_sd if k.startswith("model.multi_modal_projector.")
+    }
+    other_keys = set(original_sd.keys()) - lang_keys - vision_keys - projector_keys
+
+    LOG.info(
+        "Original checkpoint: %d language, %d vision, %d projector, %d other keys",
+        len(lang_keys),
+        len(vision_keys),
+        len(projector_keys),
+        len(other_keys),
+    )
+
+    # 4. Classify trained keys (reverse mapping on save gives model.language_model.* prefix)
+    trained_lang_keys = {k for k in trained_sd if k.startswith("model.language_model.")}
+    trained_other = set(trained_sd.keys()) - trained_lang_keys
+
+    LOG.info(
+        "Trained checkpoint: %d language keys, %d other keys",
+        len(trained_lang_keys),
+        len(trained_other),
+    )
+
+    # 5. Build merged state dict
+    merged: dict[str, torch.Tensor] = {}
+
+    # Keep vision tower and projector from original
+    for key in vision_keys | projector_keys:
+        merged[key] = original_sd[key]
+
+    # Use trained language model weights (overwrite original)
+    for key in trained_lang_keys:
+        merged[key] = trained_sd[key]
+
+    # For other trained keys (like lm_head.weight), use trained version
+    for key in trained_other:
+        merged[key] = trained_sd[key]
+
+    # For any original other keys not covered by trained (shouldn't usually happen),
+    # keep original
+    for key in other_keys:
+        if key not in merged:
+            merged[key] = original_sd[key]
+
+    # Check for missing language keys that were in original but not in trained
+    missing_lang = lang_keys - trained_lang_keys
+    if missing_lang:
+        LOG.warning(
+            "%d language keys in original but not in trained; keeping original: %s",
+            len(missing_lang),
+            list(missing_lang)[:5],
+        )
+        for key in missing_lang:
+            merged[key] = original_sd[key]
+
+    LOG.info("Merged checkpoint: %d total keys", len(merged))
+
+    # 6. Save merged weights (sharded at 50GB, matching transformers default)
+    LOG.info("Saving merged weights to %s", out_path)
+    state_dict_split = split_torch_state_dict_into_shards(merged, max_shard_size="50GB")
+
+    for filename, tensors in state_dict_split.filename_to_tensors.items():
+        shard = {name: merged[name] for name in tensors}
+        save_file(shard, str(out_path / filename))
+
+    if state_dict_split.is_sharded:
+        index = {
+            "metadata": {
+                "total_size": sum(t.numel() * t.element_size() for t in merged.values())
+            },
+            "weight_map": state_dict_split.tensor_to_filename,
+        }
+        with open(out_path / "model.safetensors.index.json", "w") as f:
+            json.dump(index, f, indent=2)
+        LOG.info("Saved %d shards", len(state_dict_split.filename_to_tensors))
+
+    # 7. Copy/update config
+    LOG.info("Writing config.json")
+    original_config = AutoConfig.from_pretrained(
+        original_model, trust_remote_code=trust_remote_code
+    )
+
+    # Update text_config fields from trained model's config if available
+    trained_config_path = trained_path / "config.json"
+    if trained_config_path.exists():
+        with open(trained_config_path) as f:
+            trained_config_dict = json.load(f)
+
+        # The trained config is the text sub-config; merge its fields into
+        # the original composite config's text_config
+        if hasattr(original_config, "text_config"):
+            for key, val in trained_config_dict.items():
+                if key not in ("model_type", "_name_or_path", "architectures"):
+                    if hasattr(original_config.text_config, key):
+                        setattr(original_config.text_config, key, val)
+
+    original_config.save_pretrained(out_path)
+
+    # 8. Copy tokenizer files from trained model if present
+    tokenizer_files = list(trained_path.glob("tokenizer*")) + list(
+        trained_path.glob("special_tokens_map*")
+    )
+    if tokenizer_files:
+        import shutil
+
+        for tok_file in tokenizer_files:
+            shutil.copy2(tok_file, out_path / tok_file.name)
+        LOG.info("Copied %d tokenizer files", len(tokenizer_files))
+
+    LOG.info("Merge complete. Output saved to %s", out_path)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Merge trained text-only Gemma3 weights back into a multimodal checkpoint."
+    )
+    parser.add_argument(
+        "--original-model",
+        required=True,
+        help="HuggingFace model ID or local path to the original multimodal model",
+    )
+    parser.add_argument(
+        "--trained-model",
+        required=True,
+        help="Local path to the trained text-only model output directory",
+    )
+    parser.add_argument(
+        "--output-dir",
+        required=True,
+        help="Directory to save the merged multimodal checkpoint",
+    )
+    parser.add_argument(
+        "--trust-remote-code",
+        action="store_true",
+        default=False,
+        help="Trust remote code when loading model config",
+    )
+    args = parser.parse_args()
+
+    merge(
+        original_model=args.original_model,
+        trained_model=args.trained_model,
+        output_dir=args.output_dir,
+        trust_remote_code=args.trust_remote_code,
+    )
+
+
+if __name__ == "__main__":
+    main()

setup.py

@@ -1,6 +1,5 @@
 """setup.py for axolotl"""
 
-import ast
 import os
 import platform
 import re
@@ -79,6 +78,11 @@ def parse_requirements(extras_require_map):
                 extras_require_map["vllm"] = ["vllm==0.11.1"]
                 if not install_xformers:
                     _install_requires.pop(_install_requires.index(xformers_version))
+                extras_require_map["vllm"] = ["vllm==0.13.0"]
+                if patch == 0:
+                    extras_require_map["vllm"] = ["vllm==0.13.0"]
+                else:
+                    extras_require_map["vllm"] = ["vllm==0.14.0"]
             elif (major, minor) >= (2, 8):
                 extras_require_map.pop("fbgemm-gpu")
                 extras_require_map["fbgemm-gpu"] = ["fbgemm-gpu-genai==1.3.0"]
@@ -130,15 +134,11 @@ def parse_requirements(extras_require_map):
 
 def get_package_version():
     with open(
-        Path(os.path.dirname(os.path.abspath(__file__)))
-        / "src"
-        / "axolotl"
-        / "__init__.py",
+        Path(os.path.dirname(os.path.abspath(__file__))) / "VERSION",
         "r",
         encoding="utf-8",
     ) as fin:
-        version_match = re.search(r"^__version__\s*=\s*(.*)$", fin.read(), re.MULTILINE)
-    version_ = ast.literal_eval(version_match.group(1))
+        version_ = fin.read().strip()
     return version_
 
 

src/axolotl/__init__.py

@@ -1,7 +1,11 @@
 """Axolotl - Train and fine-tune large language models"""
 
 import pkgutil
+from importlib.metadata import PackageNotFoundError, version
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.13.1"
+try:
+    __version__ = version("axolotl")
+except PackageNotFoundError:
+    __version__ = "unknown"

src/axolotl/cli/__init__.py

@@ -5,6 +5,6 @@ import os
 from axolotl.logging_config import configure_logging
 
 os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
-os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")
+os.environ.setdefault("HF_XET_HIGH_PERFORMANCE", "1")
 
 configure_logging()

src/axolotl/cli/checks.py

@@ -44,7 +44,7 @@ def check_user_token() -> bool:
         return bool(user_info)
     except LocalTokenNotFoundError:
         LOG.warning(
-            "Error verifying HuggingFace token. Remember to log in using `huggingface-cli login` and get your access token from https://huggingface.co/settings/tokens if you want to use gated models or datasets."
+            "Error verifying HuggingFace token. Remember to log in using `hf auth login` and get your access token from https://huggingface.co/settings/tokens if you want to use gated models or datasets."
         )
         return False
     except HTTPError:

src/axolotl/cli/merge_lora.py

@@ -24,7 +24,6 @@ def do_merge_lora(*, cfg: DictDefault) -> None:
         cfg: Dictionary mapping `axolotl` config keys to values.
     """
     model, tokenizer, processor = load_model_and_tokenizer(cfg=cfg)
-    safe_serialization = cfg.save_safetensors is True
 
     LOG.info("Running merge of LoRA with base model...")
     model = model.merge_and_unload(progressbar=True)
@@ -42,7 +41,6 @@ def do_merge_lora(*, cfg: DictDefault) -> None:
         LOG.info(f"Saving merged model to: {str(Path(cfg.output_dir) / 'merged')}...")
         model.save_pretrained(
             str(Path(cfg.output_dir) / "merged"),
-            safe_serialization=safe_serialization,
             progressbar=True,
         )
         tokenizer.save_pretrained(

src/axolotl/cli/merge_sharded_fsdp_weights.py

@@ -14,8 +14,6 @@ from accelerate import PartialState
 from accelerate.utils import (
     SAFE_WEIGHTS_INDEX_NAME,
     SAFE_WEIGHTS_NAME,
-    WEIGHTS_INDEX_NAME,
-    WEIGHTS_NAME,
     is_torch_version,
 )
 from huggingface_hub import split_torch_state_dict_into_shards
@@ -40,17 +38,15 @@ class BFloat16CastPlanner(_EmptyStateDictLoadPlanner):
 def _distributed_checkpoint_to_merged_weights(
     checkpoint_dir: Union[str, Path],
     save_path: str,
-    safe_serialization: bool = False,
     max_shard_size: str = "5GB",
 ) -> Path:
     """
     Passthrough to `torch.distributed.checkpoint.format_utils.dcp_to_torch_save`. Will
-    save under `save_path` as either `model.safetensors` or `pytorch_model.bin`.
+    save under `save_path` as `model.safetensors`.
 
     Args:
         checkpoint_dir: Directory where distributed checkpoint is saved.
         save_path: Path to save model to.
-        safe_serialization: Whether to save in safetensors format.
         max_shard_size: Max size of model shards to save.
 
     Returns:
@@ -76,11 +72,7 @@ def _distributed_checkpoint_to_merged_weights(
         if isinstance(value, torch.Tensor) and value.dtype != torch.bfloat16:
             state_dict[key] = value.to(torch.bfloat16)
 
-    weights_name = SAFE_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
-
-    filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(
-        ".safetensors", "{suffix}.safetensors"
-    )
+    filename_pattern = SAFE_WEIGHTS_NAME.replace(".safetensors", "{suffix}.safetensors")
     state_dict_split = split_torch_state_dict_into_shards(
         state_dict, filename_pattern=filename_pattern, max_shard_size=max_shard_size
     )
@@ -98,19 +90,12 @@ def _distributed_checkpoint_to_merged_weights(
 
     for shard_file, tensors in filename_to_tensors:
         shard = {tensor: state_dict[tensor] for tensor in tensors}
-
-        if safe_serialization:
-            safe_save_file(
-                shard, os.path.join(save_path_, shard_file), metadata={"format": "pt"}
-            )
-        else:
-            torch.save(shard, os.path.join(save_path_, shard_file))
+        safe_save_file(
+            shard, os.path.join(save_path_, shard_file), metadata={"format": "pt"}
+        )
 
     if index is not None:
-        save_index_file = (
-            SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME
-        )
-        save_index_file = os.path.join(save_path_, save_index_file)
+        save_index_file = os.path.join(save_path_, SAFE_WEIGHTS_INDEX_NAME)
         # Save the index as well
         with open(save_index_file, "w", encoding="utf-8") as fout:
             content = json.dumps(index, indent=2, sort_keys=True) + "\n"
@@ -123,13 +108,11 @@ def _distributed_checkpoint_to_merged_weights(
 def merge_fsdp_weights(
     checkpoint_dir: str,
     output_path: str,
-    safe_serialization: bool = False,
     remove_checkpoint_dir: bool = False,
 ):
     """
     Merge the weights from sharded FSDP model checkpoints into a single combined checkpoint. Should be used if
-    `SHARDED_STATE_DICT` was used for the model. Weights will be saved to `{output_path}/model.safetensors` if
-    `safe_serialization` else `pytorch_model.bin`.
+    `SHARDED_STATE_DICT` was used for the model. Weights will be saved to `{output_path}/model.safetensors`.
 
     Note: this is a CPU-bound process.
 
@@ -138,8 +121,6 @@ def merge_fsdp_weights(
             The directory containing the FSDP checkpoints (can be either the model or optimizer).
         output_path (`str`):
             The path to save the merged checkpoint.
-        safe_serialization (`bool`, *optional*, defaults to `True`):
-            Whether to save the merged weights with safetensors (recommended).
         remove_checkpoint_dir (`bool`, *optional*, defaults to `False`):
             Whether to remove the checkpoint directory after merging.
 
@@ -177,7 +158,7 @@ def merge_fsdp_weights(
     if state.is_main_process:
         LOG.info(f"Merging FSDP weights from {checkpoint_dir_}")
         save_path = _distributed_checkpoint_to_merged_weights(
-            checkpoint_dir_, output_path, safe_serialization
+            checkpoint_dir_, output_path
         )
         LOG.info(f"Successfully merged FSDP weights and saved to {save_path}")
         if remove_checkpoint_dir:
@@ -210,7 +191,6 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
     merge_fsdp_weights(
         checkpoint_dir=str(fsdp_dir),
         output_path=output_path,
-        safe_serialization=True,
     )
     state = PartialState()
     state.wait_for_everyone()

src/axolotl/cli/quantize.py

@@ -102,12 +102,10 @@ def do_quantize(
     LOG.info(f"Saving quantized model to: {str(Path(output_dir) / 'quantized')}.")
     model.save_pretrained(
         str(Path(output_dir) / "quantized"),
-        safe_serialization=False,
         progressbar=True,
     )
     tokenizer.save_pretrained(
         str(Path(output_dir) / "quantized"),
-        safe_serialization=False,
         progressbar=True,
         save_jinja_files=cfg.tokenizer_save_jinja_files,
     )
@@ -121,7 +119,7 @@ def do_quantize(
             hub_model_id.rstrip("-")
             + f"-{quantization_config_to_str[type(quantization_config)]}"
         )
-        model.push_to_hub(hub_model_id, safe_serialization=False)
+        model.push_to_hub(hub_model_id)
         tokenizer.push_to_hub(hub_model_id)
         if processor:
             processor.push_to_hub(hub_model_id)

src/axolotl/core/builders/base.py

@@ -216,7 +216,7 @@ class TrainerBuilderBase(abc.ABC):
     def _configure_warmup_and_logging(
         self, total_num_steps: int, training_args_kwargs: dict
     ):
-        warmup_steps = 0
+        warmup_steps: int | float = 0
         warmup_ratio = 0.0
         if self.cfg.warmup_steps is not None:
             warmup_steps = self.cfg.warmup_steps
@@ -230,6 +230,10 @@ class TrainerBuilderBase(abc.ABC):
         else:
             warmup_ratio = 0.03
 
+        # transformers v5
+        if warmup_ratio > 0.0 and warmup_steps == 0:
+            warmup_steps = warmup_ratio
+
         if warmup_steps == 1:
             warmup_steps = 2
 
@@ -242,7 +246,6 @@ class TrainerBuilderBase(abc.ABC):
                 else max(min(int(0.005 * total_num_steps), 10), 1)
             )
 
-        training_args_kwargs["warmup_ratio"] = warmup_ratio
         training_args_kwargs["warmup_steps"] = warmup_steps
 
     def _configure_precision_settings(self, training_args_kwargs: dict):
@@ -406,6 +409,9 @@ class TrainerBuilderBase(abc.ABC):
             if self.cfg.hub_strategy:
                 training_args_kwargs["hub_strategy"] = self.cfg.hub_strategy
 
+            if self.cfg.hub_revision:
+                training_args_kwargs["hub_revision"] = self.cfg.hub_revision
+
     def _configure_save_and_eval_strategy(self, training_args_kwargs: dict):
         # save_strategy and save_steps
         if self.cfg.save_steps:
@@ -530,9 +536,7 @@ class TrainerBuilderBase(abc.ABC):
             "loraplus_lr_ratio",
             "loraplus_lr_embedding",
             "output_dir",
-            "save_safetensors",
             "save_only_model",
-            "include_tokens_per_second",
             "weight_decay",
             "seed",
             "dion_momentum",
@@ -545,6 +549,7 @@ class TrainerBuilderBase(abc.ABC):
 
         arg_map = {
             "dion_learning_rate": "dion_lr",
+            "include_num_input_tokens_seen": "include_tokens_per_second",
         }
         for kwarg, cfg_arg in arg_map.items():
             if hasattr(self.cfg, cfg_arg) and getattr(self.cfg, cfg_arg) is not None:

src/axolotl/core/builders/causal.py

@@ -373,6 +373,18 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
             data_collator_kwargs["pad_to_multiple_of"] = multiple
 
+        if self.cfg.use_eaft:
+            from functools import partial
+
+            from axolotl.monkeypatch.loss.eaft import eaft_loss
+
+            configured_eaft_loss = partial(
+                eaft_loss,
+                alpha=self.cfg.eaft_alpha if self.cfg.eaft_alpha is not None else 1.0,
+                k=self.cfg.eaft_k if self.cfg.eaft_k is not None else 20,
+            )
+            trainer_kwargs["compute_loss_func"] = configured_eaft_loss
+
         trainer_cls = self._get_trainer_cls()
 
         trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
@@ -437,7 +449,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 or self.cfg.micro_batch_size > 1
             ):
                 return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
-            if not (self.cfg.sample_packing and self.cfg.pretrain_multipack_attn):
+            if not (self.cfg.sample_packing and self.cfg.pretrain_multipack_attn) or (
+                self.cfg.micro_batch_size == 1 and is_eval is False
+            ):
                 return None
 
         if self.cfg.model_config_type == "mamba":

src/axolotl/core/builders/rl.py

@@ -52,12 +52,11 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         trainer_cls = None
         trainer_cls_args = [self.model]
 
-        if self.cfg.rl is RLType.GRPO:
+        if self.cfg.rl in {RLType.GRPO, RLType.GDPO}:
             trainer_cls = GRPOStrategy.get_trainer_class(
                 sequence_parallel=self.cfg.context_parallel_size > 1
             )
             trainer_cls_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
-
             trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
 
         elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
@@ -147,6 +146,8 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
 
         elif self.cfg.rl is RLType.KTO:
             training_args_cls = AxolotlKTOConfig
+            # KTOConfig in TRL >= 0.27.0 no longer accepts max_prompt_length
+            blocklist_args_kwargs = ["max_prompt_length"]
 
             training_args_kwargs["desirable_weight"] = (
                 self.cfg.kto_desirable_weight or 1.0
@@ -155,10 +156,14 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
                 self.cfg.kto_undesirable_weight or 1.0
             )
 
-        elif self.cfg.rl is RLType.GRPO:
+        elif self.cfg.rl in {RLType.GRPO, RLType.GDPO}:
             training_args_cls = GRPOStrategy.get_training_args_class()
             training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
             blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
+            if self.cfg.rl is RLType.GDPO:
+                training_args_kwargs.setdefault(
+                    "multi_objective_aggregation", "normalize_then_sum"
+                )
 
         elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
             training_args_cls = AxolotlDPOConfig

src/axolotl/core/trainers/base.py

@@ -25,7 +25,7 @@ from torch.utils.data import (
 from transformers import PreTrainedModel, Trainer
 from transformers.trainer import TRAINING_ARGS_NAME
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, has_length, seed_worker
-from transformers.utils import SAFE_WEIGHTS_NAME, WEIGHTS_NAME, is_peft_available
+from transformers.utils import SAFE_WEIGHTS_NAME, is_peft_available
 from trl.trainer.utils import pad_to_length
 from typing_extensions import override
 
@@ -719,6 +719,13 @@ class AxolotlTrainer(
         output_dir = output_dir if output_dir is not None else self.args.output_dir
         os.makedirs(output_dir, exist_ok=True)
         LOG.info(f"Saving model checkpoint to {output_dir}")
+        if state_dict is None:
+            state_dict = self.accelerator.get_state_dict(self.model)
+        if state_dict is not None:
+            state_dict = {
+                k: v.clone() if isinstance(v, torch.Tensor) else v
+                for k, v in state_dict.items()
+            }
         supported_classes = (
             (PreTrainedModel,)
             if not is_peft_available()
@@ -738,43 +745,38 @@ class AxolotlTrainer(
                 ).save_pretrained(
                     output_dir,
                     state_dict=state_dict,
-                    safe_serialization=self.args.save_safetensors,
                 )
             else:
                 LOG.info(
                     "Trainer.model is not a `PreTrainedModel`, only saving its state dict."
                 )
-                if self.args.save_safetensors:
-                    safetensors.torch.save_file(
-                        state_dict,
-                        os.path.join(output_dir, SAFE_WEIGHTS_NAME),
-                        metadata={"format": "pt"},
-                    )
-                else:
-                    torch.save(state_dict, os.path.join(output_dir, WEIGHTS_NAME))
+                safetensors.torch.save_file(
+                    state_dict,
+                    os.path.join(output_dir, SAFE_WEIGHTS_NAME),
+                    metadata={"format": "pt"},
+                )
         else:
             self.model.save_pretrained(
                 output_dir,
                 state_dict=state_dict,
-                safe_serialization=self.args.save_safetensors,
                 is_main_process=self.accelerator.is_main_process,
             )
 
-            if self.processing_class is not None:
-                self.processing_class.save_pretrained(output_dir)
-            elif (
-                self.data_collator is not None
-                and hasattr(self.data_collator, "tokenizer")
-                and self.data_collator.tokenizer is not None
-            ):
-                LOG.info(
-                    "Saving Trainer.data_collator.tokenizer by default as Trainer.processing_class is `None`"
-                )
-                save_jinja_files = True
-                if self.axolotl_cfg:
-                    save_jinja_files = self.axolotl_cfg.tokenizer_save_jinja_files
-                self.data_collator.tokenizer.save_pretrained(
-                    output_dir, save_jinja_files=save_jinja_files
-                )
-            # Good practice: save your training arguments together with the trained model
-            torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))
+        if self.processing_class is not None:
+            self.processing_class.save_pretrained(output_dir)
+        elif (
+            self.data_collator is not None
+            and hasattr(self.data_collator, "tokenizer")
+            and self.data_collator.tokenizer is not None
+        ):
+            LOG.info(
+                "Saving Trainer.data_collator.tokenizer by default as Trainer.processing_class is `None`"
+            )
+            save_jinja_files = True
+            if self.axolotl_cfg:
+                save_jinja_files = self.axolotl_cfg.tokenizer_save_jinja_files
+            self.data_collator.tokenizer.save_pretrained(
+                output_dir, save_jinja_files=save_jinja_files
+            )
+        # Good practice: save your training arguments together with the trained model
+        torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))

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

@@ -126,8 +126,10 @@ 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)
+        if trl.multi_objective_aggregation is not None:
+            grpo_args_kwargs["multi_objective_aggregation"] = (
+                trl.multi_objective_aggregation
+            )
 
         return grpo_args_kwargs
 
@@ -149,6 +151,8 @@ class GRPOStrategy:
             trainer_kwargs["reward_processing_classes"] = (
                 cfg.trl.reward_processing_classes
             )
+        if cfg.trl and cfg.trl.rollout_func:
+            trainer_kwargs["rollout_func"] = cls.get_rollout_func(cfg.trl.rollout_func)
 
         return trainer_kwargs
 
@@ -159,7 +163,12 @@ class GRPOStrategy:
 
     @classmethod
     def get_blocklist_args_kwargs(cls) -> list[str]:
-        return ["dataset_num_proc", "max_length", "include_tokens_per_second"]
+        return [
+            "dataset_num_proc",
+            "max_length",
+            "include_tokens_per_second",
+            "max_prompt_length",
+        ]
 
     @classmethod
     def get_reward_func(cls, reward_func_fqn: str) -> RewardFunc:

src/axolotl/core/trainers/trl.py

@@ -1,12 +1,10 @@
 """Module for TRL RL trainers"""
 
-from trl import (
-    CPOTrainer,
-    KTOTrainer,
-    ORPOTrainer,
-    PRMTrainer,
-    RewardTrainer,
-)
+from trl import RewardTrainer
+from trl.experimental.cpo import CPOTrainer
+from trl.experimental.kto import KTOTrainer
+from trl.experimental.orpo import ORPOTrainer
+from trl.experimental.prm import PRMTrainer
 
 from axolotl.core.trainers.mixins import DistributedParallelMixin, RngLoaderMixin
 from axolotl.core.trainers.mixins.optimizer import OptimizerInitMixin, OptimizerMixin

src/axolotl/core/training_args.py

@@ -8,7 +8,11 @@ from dataclasses import dataclass, field
 from typing import Optional, Type
 
 from transformers import TrainingArguments
-from trl import CPOConfig, KTOConfig, ORPOConfig, PRMConfig, RewardConfig
+from trl import RewardConfig
+from trl.experimental.cpo import CPOConfig
+from trl.experimental.kto import KTOConfig
+from trl.experimental.orpo import ORPOConfig
+from trl.experimental.prm import PRMConfig
 
 from axolotl.integrations.config import merge_training_args
 

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@318b7e2"
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@0d4ce4b"
 ```
 
 ## Usage
@@ -36,6 +36,7 @@ plugins:
 - cohere
 - cohere2
 - deepseek_v3
+- exaone4
 - gemma
 - gemma2
 - gemma3
@@ -45,13 +46,16 @@ plugins:
 - glm
 - glm4
 - glm4_moe
+- glm4_moe_lite
+- glm46v
 - glm4v
 - glm4v_moe
+- glm_image
 - gpt_oss
 - granite
 - granitemoe
-- granitemoeshared
 - granitemoehybrid
+- granitemoeshared
 - hunyuan_v1_dense
 - hunyuan_v1_moe
 - internvl
@@ -76,16 +80,17 @@ plugins:
 - phi3
 - phi4_multimodal
 - qwen2
-- qwen2_vl
 - qwen2_moe
+- qwen2_vl
 - qwen2_5_vl
 - qwen3
 - qwen3_moe
+- qwen3_next
 - qwen3_vl
 - qwen3_vl_moe
-- qwen3_next
-- smollm3
 - seed_oss
+- smollm3
+- step3p5
 - voxtral
 
 ## Citation

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@318b7e2"`'
+    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@0d4ce4b"`'
 )
 
 

src/axolotl/integrations/gemma3/README.md

@@ -0,0 +1,37 @@
+# Gemma3 Text-from-Multimodal Plugin
+
+Load a Gemma3 multimodal checkpoint (e.g. `google/gemma-3-4b-it`) directly into `Gemma3ForCausalLM` for text-only training. This bypasses the multimodal trainer path and enables sample packing and other text-specific optimizations.
+
+## How it works
+
+The plugin uses transformers v5's `key_mapping` parameter on `from_pretrained` to remap `model.language_model.*` checkpoint keys to `model.*`, matching what `Gemma3ForCausalLM` expects. Vision tower and projector weights are automatically discarded. On save, transformers reverses the mapping so checkpoints retain the original `model.language_model.*` prefix.
+
+## Usage
+
+Add the plugin to your YAML config:
+
+```yaml
+base_model: google/gemma-3-4b-it
+
+plugins:
+  - axolotl.integrations.gemma3.Gemma3TextFromMultimodalPlugin
+```
+
+See `examples/gemma3/gemma-3-4b-qlora.yml` for a complete example.
+
+## Merging weights back into a multimodal checkpoint
+
+After training, the saved checkpoint contains only the language model weights. To reconstruct a full `Gemma3ForConditionalGeneration` checkpoint (with the original vision tower and projector), use the merge script:
+
+```bash
+python scripts/merge_gemma3_multimodal_weights.py \
+    --original-model google/gemma-3-4b-it \
+    --trained-model /path/to/trained/output \
+    --output-dir /path/to/merged
+```
+
+This combines:
+- **Trained language model weights** from your output checkpoint
+- **Original vision tower + projector weights** from the base multimodal model
+
+The merged checkpoint can be loaded as `Gemma3ForConditionalGeneration` for multimodal inference or further training.

src/axolotl/integrations/gemma3/__init__.py

@@ -0,0 +1,9 @@
+"""Gemma3 integration for loading multimodal checkpoints as text-only models."""
+
+from .args import Gemma3TextFromMultimodalArgs
+from .plugin import Gemma3TextFromMultimodalPlugin
+
+__all__ = [
+    "Gemma3TextFromMultimodalArgs",
+    "Gemma3TextFromMultimodalPlugin",
+]

src/axolotl/integrations/gemma3/args.py

@@ -0,0 +1,31 @@
+"""Pydantic input args for the Gemma3 text-from-multimodal plugin."""
+
+from pydantic import BaseModel, model_validator
+
+from axolotl.utils.logging import get_logger
+
+LOG = get_logger(__name__)
+
+
+class Gemma3TextFromMultimodalArgs(BaseModel):
+    """Configuration args for loading a Gemma3 multimodal checkpoint as text-only."""
+
+    gemma3_text_from_multimodal: bool = True
+    extract_text_config: bool = False
+
+    @model_validator(mode="before")
+    @classmethod
+    def set_model_type(cls, data):
+        if not isinstance(data, dict):
+            return data
+
+        if not data.get("gemma3_text_from_multimodal", True):
+            return data
+
+        if not data.get("model_type"):
+            LOG.info(
+                "Gemma3TextFromMultimodalPlugin: auto-setting model_type to Gemma3ForCausalLM"
+            )
+            data["model_type"] = "Gemma3ForCausalLM"
+
+        return data

src/axolotl/integrations/gemma3/plugin.py

@@ -0,0 +1,107 @@
+"""Plugin for loading Gemma3 multimodal checkpoints into Gemma3ForCausalLM (text-only).
+
+Uses transformers v5's ``key_mapping`` parameter on ``from_pretrained`` to remap
+``model.language_model.*`` keys to ``model.*``, discarding vision tower and projector
+weights.  On save, transformers automatically reverses the mapping so saved
+checkpoints retain the original ``model.language_model.*`` prefix.
+"""
+
+from axolotl.integrations.base import BasePlugin
+from axolotl.utils.logging import get_logger
+
+LOG = get_logger(__name__)
+
+# key_mapping for transformers from_pretrained:
+# Remap checkpoint keys matching ^model.language_model -> model
+# Vision tower / projector keys won't match any model parameter and are discarded.
+GEMMA3_KEY_MAPPING = {"^model.language_model": "model"}
+
+
+class Gemma3TextFromMultimodalPlugin(BasePlugin):
+    """Load a Gemma3 multimodal checkpoint as a text-only Gemma3ForCausalLM.
+
+    Hooks
+    -----
+    register(cfg)
+        Runs before config validation.  Sets the ``_extract_text_config`` flag,
+        ensures ``model_type`` is ``Gemma3ForCausalLM``, and injects
+        ``key_mapping`` into ``model_kwargs`` so that ``from_pretrained`` remaps
+        ``model.language_model.*`` → ``model.*``.
+
+    pre_model_load(cfg)
+        Runs after config validation/normalization but before model instantiation.
+        Validates that ``model_config_type`` is ``gemma3_text`` and
+        ``is_multimodal`` is False (confirming that ``_extract_text_config``
+        worked correctly).
+    """
+
+    def get_input_args(self) -> str:
+        return "axolotl.integrations.gemma3.Gemma3TextFromMultimodalArgs"
+
+    def register(self, cfg: dict):
+        """Set up config for multimodal → text-only loading.
+
+        This runs before Pydantic validation, so ``cfg`` is a raw dict.
+        """
+        if not cfg.get("gemma3_text_from_multimodal", True):
+            raise ValueError(
+                "Gemma3TextFromMultimodalPlugin: disabled via config, but plugin selected"
+            )
+
+        # Flag for load_model_config() to extract the text sub-config
+        cfg["extract_text_config"] = True
+
+        # Ensure model_type is set for the text-only model class
+        if not cfg.get("model_type"):
+            cfg["model_type"] = "Gemma3ForCausalLM"
+
+        # Inject key_mapping into model_kwargs so from_pretrained remaps weights
+        model_kwargs = cfg.setdefault("model_kwargs", {})
+        model_kwargs["key_mapping"] = GEMMA3_KEY_MAPPING
+
+    def pre_model_load(self, cfg):
+        """Validate that config extraction worked before model instantiation."""
+        if not getattr(cfg, "gemma3_text_from_multimodal", True):
+            return
+
+        if cfg.model_config_type != "gemma3_text":
+            LOG.warning(
+                "Gemma3TextFromMultimodalPlugin: expected model_config_type='gemma3_text' "
+                "but got '%s'. The text config extraction may not have worked.",
+                cfg.model_config_type,
+            )
+
+        if cfg.is_multimodal or cfg.processor_type:
+            raise ValueError(
+                "Multimodal mode is enabled (processor_type set), but "
+                "Gemma3TextFromMultimodalPlugin enabled. "
+                "Please disable one of the two."
+            )
+
+    def post_train(self, cfg, model):
+        """Log merge command after training completes."""
+        if cfg.adapter:
+            LOG.info(
+                "Adapter training detected. To reconstruct the multimodal checkpoint:\n"
+                "  1. Merge adapter weights into the text-only base model:\n"
+                "       axolotl merge_lora <your_config.yml>\n"
+                "  2. Then merge the resulting full model back into the multimodal checkpoint:\n"
+                "       python scripts/merge_gemma3_multimodal_weights.py \\\n"
+                "         --original-model %s \\\n"
+                "         --trained-model %s/merged \\\n"
+                "         --output-dir %s/multi-modal/merged",
+                cfg.base_model,
+                cfg.output_dir,
+                cfg.output_dir,
+            )
+        else:
+            LOG.info(
+                "To merge trained weights back into the multimodal checkpoint, run:\n"
+                "  python scripts/merge_gemma3_multimodal_weights.py \\\n"
+                "    --original-model %s \\\n"
+                "    --trained-model %s \\\n"
+                "    --output-dir %s/multi-modal/merged",
+                cfg.base_model,
+                cfg.output_dir,
+                cfg.output_dir,
+            )

src/axolotl/integrations/kernels/__init__.py

@@ -0,0 +1,7 @@
+from .args import KernelsArgs
+from .plugin import KernelsPlugin
+
+__all__ = [
+    "KernelsArgs",
+    "KernelsPlugin",
+]

src/axolotl/integrations/kernels/args.py

@@ -0,0 +1,35 @@
+from pydantic import BaseModel, model_validator
+
+from axolotl.utils.logging import get_logger
+
+LOG = get_logger(__name__)
+
+
+class KernelsArgs(BaseModel):
+    use_scattermoe: bool | None = True
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_use_kernels(cls, data):
+        if data.get("use_kernels") is not True:
+            LOG.warning(
+                "`use_kernels` must be set to True to use this. Automatically setting it to True."
+            )
+            data["use_kernels"] = True
+
+        return data
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_experts_implementation(cls, data):
+        experts_implementation = data.get("experts_implementation")
+        if experts_implementation is None:
+            # transformers may default to batched_mm when unset
+            data["experts_implementation"] = "eager"
+        elif experts_implementation != "eager":
+            LOG.warning(
+                "`experts_implementation` must be set to 'eager' to use this. Automatically setting it to 'eager'."
+            )
+            data["experts_implementation"] = "eager"
+
+        return data

src/axolotl/integrations/kernels/plugin.py

@@ -0,0 +1,61 @@
+from kernels import (
+    LayerRepository,
+    Mode,
+    register_kernel_mapping,
+    replace_kernel_forward_from_hub,
+)
+
+from axolotl.integrations.base import BasePlugin
+from axolotl.utils.callbacks.models import get_causal_lm_model_cls_prefix
+
+
+class KernelsPlugin(BasePlugin):
+    def get_input_args(self):
+        return "axolotl.integrations.kernels.KernelsArgs"
+
+    def pre_model_load(self, cfg):
+        if cfg.use_scattermoe:
+            self._register_kernels()
+            self._kernelize_model(cfg.model_config_type)
+
+    def _register_kernels(self):
+        register_kernel_mapping(
+            {
+                "HFScatterMoEParallelExperts": {
+                    "cuda": {
+                        Mode.TRAINING: LayerRepository(
+                            repo_id="axolotl-ai-co/scattermoe",
+                            layer_name="HFScatterMoEGatedMLP",
+                        ),
+                        Mode.INFERENCE: LayerRepository(
+                            repo_id="axolotl-ai-co/scattermoe",
+                            layer_name="HFScatterMoEGatedMLP",
+                        ),
+                    },
+                }
+            }
+        )
+
+    def _kernelize_model(self, model_type: str):
+        if model_type == "olmoe":
+            from transformers.models.olmoe.modeling_olmoe import OlmoeSparseMoeBlock
+
+            replace_kernel_forward_from_hub(
+                OlmoeSparseMoeBlock, "HFScatterMoEParallelExperts"
+            )
+        else:
+            try:
+                model_moe_cls = get_model_moe_block(model_type)
+                replace_kernel_forward_from_hub(
+                    model_moe_cls, "HFScatterMoEParallelExperts"
+                )
+            except Exception as err:
+                raise ValueError(f"Unsupported model type: {model_type}") from err
+
+
+def get_model_moe_block(model_type: str):
+    module_path = f"transformers.models.{model_type}.modeling_{model_type}"
+    model_cls_prefix, _ = get_causal_lm_model_cls_prefix(model_type)
+    module = __import__(module_path, fromlist=[f"{model_cls_prefix}SparseMoeBlock"])
+    model_cls = getattr(module, f"{model_cls_prefix}SparseMoeBlock")
+    return model_cls

src/axolotl/integrations/llm_compressor/utils.py

@@ -12,7 +12,6 @@ def save_compressed_model(
     model: PreTrainedModel,
     output_dir: Union[str, bytes],
     trainer: Trainer,
-    safe_serialization: bool = False,
     save_compressed: bool = False,
 ) -> None:
     """
@@ -22,7 +21,6 @@ def save_compressed_model(
         model (PreTrainedModel): The model to be saved.
         output_dir (str or bytes): Path where the model files will be written.
         trainer (Trainer): Hugging Face Trainer for process synchronization.
-        safe_serialization (bool): Use safe serialization if True.
         save_compressed (bool): Write compressed tensors if True.
     """
     trainer.accelerator.wait_for_everyone()
@@ -34,7 +32,6 @@ def save_compressed_model(
     modify_save_pretrained(model)
     model.save_pretrained(
         output_dir,
-        safe_serialization=safe_serialization,
         save_compressed=save_compressed,
         skip_sparsity_compression_stats=not save_compressed,
     )

src/axolotl/loaders/model.py

@@ -26,7 +26,6 @@ from torch.distributed import DeviceMesh
 from transformers import (
     AutoModelForCausalLM,
     AutoModelForImageTextToText,
-    AutoModelForVision2Seq,
     AwqConfig,
     BitsAndBytesConfig,
     GPTQConfig,
@@ -226,6 +225,7 @@ class ModelLoader:
         ):
             self.model = self.model.merge_and_unload()
 
+        self._configure_experts_implementation()
         self._apply_activation_checkpointing()
         self._resize_token_embeddings()
         self._adjust_model_config()
@@ -233,6 +233,10 @@ class ModelLoader:
         self._configure_qat()
         log_gpu_memory_usage(LOG, "Memory usage after model load", 0)
 
+    def _configure_experts_implementation(self):
+        if self.cfg.experts_implementation is not None:
+            self.model.set_experts_implementation(self.cfg.experts_implementation)
+
     def _apply_activation_checkpointing(self):
         if self.cfg.activation_offloading is True:
             from axolotl.core.trainers.mixins.activation_checkpointing import (
@@ -334,7 +338,12 @@ class ModelLoader:
             # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so
             # we need to convert them back to fp16/bf16 for flash-attn compatibility.
             (
-                (needs_fa2_dtype or self.cfg.flash_attention or self.cfg.flex_attention)
+                (
+                    needs_fa2_dtype
+                    or self.cfg.flash_attention
+                    or self.cfg.flex_attention
+                    or self.cfg.sage_attention
+                )
                 and not self.is_qlora_and_fsdp_enabled
             )
             or (
@@ -434,7 +443,7 @@ class ModelLoader:
         """
         if self.cfg.is_multimodal:
             self.auto_model_loader = MULTIMODAL_AUTO_MODEL_MAPPING.get(
-                self.model_config.model_type, AutoModelForVision2Seq
+                self.model_config.model_type, AutoModelForImageTextToText
             )
             if isinstance(self.auto_model_loader, str):
                 self.auto_model_loader = AutoModelForImageTextToText
@@ -476,6 +485,7 @@ class ModelLoader:
             max_memory = None
 
         self.model_kwargs["torch_dtype"] = self.cfg.torch_dtype
+        self.model_kwargs["dtype"] = self.cfg.torch_dtype
 
         is_ds_zero3 = is_deepspeed_zero3_enabled()
 
@@ -607,6 +617,10 @@ class ModelLoader:
         elif self.cfg.sdp_attention:
             self.model_kwargs["attn_implementation"] = "sdpa"
             self.model_config._attn_implementation = "sdpa"
+        elif self.cfg.sage_attention:
+            # sets FA2 attention to re-use same internal handling like masking
+            self.model_kwargs["attn_implementation"] = "flash_attention_2"
+            self.model_config._attn_implementation = "flash_attention_2"
         elif self.cfg.eager_attention:
             self.model_kwargs["attn_implementation"] = "eager"
             self.model_config._attn_implementation = "eager"
@@ -670,7 +684,7 @@ class ModelLoader:
         Uses the selected loader when provided; otherwise falls back to the auto loader.
         """
         loader = model_loader_class or self.auto_model_loader
-        if loader in [AutoModelForCausalLM, AutoModelForVision2Seq]:
+        if loader in [AutoModelForCausalLM, AutoModelForImageTextToText]:
             model = loader.from_config(
                 config=self.model_config,
                 trust_remote_code=self.cfg.trust_remote_code or False,
@@ -788,6 +802,7 @@ class ModelLoader:
                 # Use auto model loader (handles gptq and default cases)
                 model_loader_class = self.auto_model_loader
 
+            self.model_kwargs["dtype"] = self.model_kwargs["torch_dtype"]
             if self.cfg.reinit_weights:
                 self.model = self._load_model_from_config(model_loader_class)
             else:

src/axolotl/loaders/patch_manager.py

@@ -96,6 +96,7 @@ class PatchManager:
         # self._apply_flex_attention_patches()
         self._apply_flash_attention_patches()
         self._apply_chunked_cross_entropy_patch()
+        self._apply_sageattn_patches()
         self._apply_fsdp_patches()
         self._apply_adapter_patches()
         self._apply_model_specific_patches()
@@ -201,6 +202,13 @@ class PatchManager:
             flex_attn_compile_kwargs = self.cfg.flex_attn_compile_kwargs or {}
             patch_flex_wrapper(**flex_attn_compile_kwargs)
 
+    def _apply_sageattn_patches(self):
+        """Apply patches for SageAttention."""
+        if self.cfg.sage_attention:
+            from axolotl.monkeypatch.attention.sage_attn import patch_sageattn
+
+            patch_sageattn()
+
     def _apply_model_specific_patches(self):
         """Apply patches specific to model architectures."""
         if (
@@ -220,13 +228,6 @@ class PatchManager:
 
             patch_qwen3_next_modeling_packing()
 
-        if self.cfg.model_config_type == "mistral3" and self.cfg.processor_type:
-            from axolotl.monkeypatch.models.mistral3.mistral_common_tokenizer import (
-                apply_mistral_tokenizer_image_patch,
-            )
-
-            apply_mistral_tokenizer_image_patch()
-
         if self.cfg.model_config_type == "kimi_linear":
             from axolotl.monkeypatch.models.kimi_linear.patch_kimi_linear import (
                 patch_kimi_model,

src/axolotl/loaders/processor.py

@@ -31,7 +31,7 @@ def load_processor(cfg: DictDefault, tokenizer: PreTrainedTokenizerBase):
 
             from axolotl.utils.mistral import HFMistralTokenizer
 
-            tokenization_mistral_common.MistralCommonTokenizer = HFMistralTokenizer
+            tokenization_mistral_common.MistralCommonBackend = HFMistralTokenizer
 
         _patch_mistralcommontokenizer()
 

src/axolotl/loaders/utils.py

@@ -204,6 +204,13 @@ def load_model_config(cfg: DictDefault) -> PretrainedConfig | addict.Dict:
 
     check_model_config(cfg, model_config)
 
+    # Extract text config from composite config when explicitly requested
+    # (set by plugins like Gemma3TextFromMultimodalPlugin)
+    if getattr(cfg, "extract_text_config", False) and hasattr(
+        model_config, "get_text_config"
+    ):
+        model_config = model_config.get_text_config()
+
     return model_config
 
 

src/axolotl/models/mamba/modeling_mamba.py

@@ -111,7 +111,6 @@ class MambaLMHeadModel(nn.Module, GenerationMixin):
         self,
         save_directory: Union[str, os.PathLike],
         state_dict: Optional[dict] = None,
-        safe_serialization: Optional[bool] = None,
     ):
         if state_dict is None:
             state_dict = self.state_dict()

src/axolotl/monkeypatch/attention/sage_attn.py

@@ -0,0 +1,211 @@
+"""
+Monkeypatch for SageAttention for use with transformers.
+
+https://github.com/thu-ml/SageAttention/
+"""
+
+import torch
+from transformers.integrations.sdpa_attention import repeat_kv
+
+from axolotl.utils.logging import get_logger
+
+LOG = get_logger(__name__)
+
+sageattn = None  # pylint: disable=invalid-name
+sageattn_varlen = None  # pylint: disable=invalid-name
+
+
+def _is_sageattn_available():
+    """Determine if SageAttention is available"""
+    try:
+        import sageattention  # noqa: F401 # pylint: disable=unused-import
+
+        return True
+    except ImportError:
+        return False
+
+
+if _is_sageattn_available():
+    # import sageattn here if available
+    from sageattention import sageattn, sageattn_varlen
+
+
+def _check_sageattn_imported():
+    """Check if SageAttention is imported. Raises an ImportError if not."""
+    if sageattn is None:
+        raise ImportError(
+            "SageAttention is not installed. Please install it from source: "
+            "`pip install git+https://github.com/thu-ml/SageAttention.git@1718ddc06dbc694bcf3c6b49ac28c1921aa2d8bd`"
+        )
+
+
+def sage_attention_forward(
+    module: torch.nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: torch.Tensor | None = None,
+    dropout: float = 0.0,
+    scaling: float | None = None,
+    is_causal: bool | None = None,
+    **kwargs,
+) -> tuple[torch.Tensor, None]:
+    """
+    Forward pass for SageAttention compatible with transformers attention interfaces.
+
+    https://github.com/thu-ml/SageAttention/
+    """
+
+    _check_sageattn_imported()
+
+    if kwargs.get("output_attentions", False) or kwargs.get("head_mask") is not None:
+        raise NotImplementedError(
+            "SageAttention does not support `output_attentions=True` or `head_mask`."
+        )
+
+    # The base sageattn API does not support dropout.
+    if dropout > 0.0:
+        raise NotImplementedError("SageAttention does not support dropout.")
+
+    # Handle Grouped-Query Attention (GQA) and Multi-Query Attention (MQA)
+    if hasattr(module, "num_key_value_groups"):
+        key = repeat_kv(key, module.num_key_value_groups)
+        value = repeat_kv(value, module.num_key_value_groups)
+
+    # Calculate is_causal following transformers
+    assert is_causal is not False, "is_causal must be True or None"
+    is_causal = True
+
+    position_ids = kwargs.get("position_ids", None)
+    query_length = query.shape[2]
+
+    cu_seqlens_q = kwargs.get("cu_seqlens_q", None)
+    cu_seqlens_k = kwargs.get("cu_seqlens_k", None)
+    max_length_q = kwargs.get("max_length_q", None)
+    max_length_k = kwargs.get("max_length_k", None)
+
+    # Sample packing uses position_ids, so we check for it first
+    if position_ids is not None and (
+        max_length_q is not None
+        or (query_length != 1 and not (torch.diff(position_ids, dim=-1) >= 0).all())
+    ):
+        # transpose inputs to NHD layout for use with FA2 utils
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+
+        batch_size = query.size(0)
+
+        from transformers.modeling_flash_attention_utils import (
+            prepare_fa2_from_position_ids,
+        )
+
+        if cu_seqlens_q is None or cu_seqlens_k is None:
+            query, key, value, indices_q, cu_seq_lens, max_seq_lens = (
+                prepare_fa2_from_position_ids(query, key, value, position_ids)
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_length_q, max_length_k = max_seq_lens
+
+        else:
+            query = query.reshape(-1, query.size(-2), query.size(-1))
+            key = key.reshape(-1, key.size(-2), key.size(-1))
+            value = value.reshape(-1, value.size(-2), value.size(-1))
+
+        attn_output_unpad = sageattn_varlen(
+            q=query,
+            k=key,
+            v=value,
+            cu_seqlens_q=cu_seqlens_q,
+            cu_seqlens_k=cu_seqlens_k,
+            max_seqlen_q=max_length_q,
+            max_seqlen_k=max_length_k,
+            is_causal=is_causal,
+            sm_scale=scaling,
+            smooth_k=False,  # reduces loss 0 / nan grad norms
+            tensor_layout="NHD",
+        )
+
+        attn_output = attn_output_unpad.view(
+            batch_size, -1, attn_output_unpad.size(-2), attn_output_unpad.size(-1)
+        )
+
+    elif attention_mask is not None:
+        # NOTE: When used without `pad_to_sequence_len`, the loss becomes unstable after a few steps.
+
+        assert attention_mask.ndim == 2, "Attention mask must be 2D"
+
+        from transformers.modeling_flash_attention_utils import (
+            _upad_input,
+        )
+
+        # transpose inputs to NHD layout for use with FA2 utils
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+
+        batch_size = query.shape[0]
+
+        query, key, value, indices_q, cu_seq_lens, max_seq_lens = _upad_input(
+            query, key, value, attention_mask, query_length
+        )
+        cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+        max_seqlen_q, max_seqlen_k = max_seq_lens
+
+        attn_output_unpad = sageattn_varlen(
+            q=query,
+            k=key,
+            v=value,
+            cu_seqlens_q=cu_seqlens_q,
+            cu_seqlens_k=cu_seqlens_k,
+            max_seqlen_q=max_seqlen_q,
+            max_seqlen_k=max_seqlen_k,
+            is_causal=is_causal,
+            sm_scale=scaling,
+            tensor_layout="NHD",
+        )
+
+        from flash_attn.bert_padding import pad_input
+
+        attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+    else:
+        # Use standard sageattn
+        # The input layout for transformers models is (batch_size, num_heads, seq_len, head_dim),
+        # which corresponds to SageAttention's "HND" layout.
+        attn_output = sageattn(
+            q=query,
+            k=key,
+            v=value,
+            tensor_layout="HND",
+            is_causal=is_causal,
+            sm_scale=scaling,
+        )
+
+        # SageAttention with "HND" returns (batch, heads, seq_len, head_dim)
+        # Transformers expects (batch, seq_len, heads, head_dim) for the output
+        # So we need to transpose dimensions 1 and 2
+        attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, None
+
+
+def patch_sageattn():
+    """Patch SageAttention for use with transformers."""
+
+    _check_sageattn_imported()
+
+    from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+
+    # Replace flash attention with sage attention
+    ALL_ATTENTION_FUNCTIONS.register("flash_attention_2", sage_attention_forward)
+
+    # Note: New method after transformers refactor to use ALL_MASK_ATTENTION_FUNCTIONS
+    # Register sage_attention with the global attention interface
+    # ALL_ATTENTION_FUNCTIONS.register("sage_attention", sage_attention_forward)
+
+    # from transformers.masking_utils import ALL_MASK_ATTENTION_FUNCTIONS, flash_attention_mask
+
+    # ALL_MASK_ATTENTION_FUNCTIONS.register("sage_attention", flash_attention_mask)
+
+    LOG.info("SageAttention patched successfully")

src/axolotl/monkeypatch/lora_kernels.py

@@ -169,7 +169,8 @@ def get_attention_cls_from_config(cfg: DictDefault) -> Type[nn.Module]:
         return attention_cls
     except (ImportError, AttributeError) as e:
         raise ValueError(
-            f"Could not import attention class for model_type: {model_type}. "
+            f"Axolotl could not import attention class for model_type: {model_type}. "
+            "Please raise an Issue and turn off lora kernels to continue training. "
             f"Error: {str(e)}"
         ) from e
 

src/axolotl/monkeypatch/loss/eaft.py

@@ -0,0 +1,51 @@
+"""
+eaft (entropy-aware focal training) loss implementation
+weights examples by entropy approximation from top-k logits
+
+Reference: https://github.com/ymxyll/LlamaFactory-EAFT/blob/e2ce19e8efcc226450ee8f2b81dfe4e69f1f945d/src/llamafactory/train/trainer_utils.py
+"""
+
+import torch
+import torch.nn.functional as F
+
+
+def eaft_loss(outputs, labels, num_items_in_batch=None, alpha=1.0, k=20):
+    """
+    compute eaft loss with entropy weighting
+
+    args:
+        outputs: model outputs containing logits
+        labels: target labels for computing loss
+        num_items_in_batch: for sample packing support
+        alpha: exponent for entropy weighting (default 1.0)
+        k: number of top logits for entropy approximation (default 20)
+    """
+    logits = outputs.logits
+
+    shift_logits = logits[..., :-1, :].contiguous()
+    shift_labels = labels[..., 1:].contiguous()
+
+    vocab_size = shift_logits.size(-1)
+    shift_logits_view = shift_logits.view(-1, vocab_size)
+    shift_labels_view = shift_labels.view(-1)
+
+    mask = shift_labels_view != -100
+
+    with torch.no_grad():
+        top_k_logits, _ = torch.topk(
+            shift_logits_view[mask].float(), k=min(k, vocab_size), dim=-1
+        )
+        top_k_probs = F.softmax(top_k_logits, dim=-1)
+        entropy = -(top_k_probs * torch.log(top_k_probs + 1e-10)).sum(dim=-1)
+        weights = torch.pow(entropy, alpha)
+
+    loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
+    per_token_loss = loss_fct(shift_logits_view[mask], shift_labels_view[mask])
+    weighted_loss = per_token_loss * weights
+
+    if num_items_in_batch is not None:
+        loss = weighted_loss.sum() / num_items_in_batch
+    else:
+        loss = weighted_loss.mean()
+
+    return loss

src/axolotl/monkeypatch/models/mistral3/mistral_common_tokenizer.py

@@ -1,5 +1,5 @@
 """
-Monkeypatch to fix inefficient tensor conversion in MistralCommonTokenizer.apply_chat_template
+Monkeypatch to fix inefficient tensor conversion in MistralCommonBackend.apply_chat_template
 """
 
 import importlib
@@ -12,11 +12,11 @@ LOG = get_logger(__name__)
 
 
 def apply_mistral_tokenizer_image_patch():
-    """Apply patch to MistralCommonTokenizer.apply_chat_template to fix image tensor conversion."""
-    from transformers.tokenization_mistral_common import MistralCommonTokenizer
+    """Apply patch to MistralCommonBackend.apply_chat_template to fix image tensor conversion."""
+    from transformers.tokenization_mistral_common import MistralCommonBackend
 
     # Get original source
-    original_source = inspect.getsource(MistralCommonTokenizer.apply_chat_template)
+    original_source = inspect.getsource(MistralCommonBackend.apply_chat_template)
     original_source, _ = detab_code(original_source)
 
     # Define the replacement
@@ -41,7 +41,7 @@ def apply_mistral_tokenizer_image_patch():
         )
 
         # Load necessary imports from the module
-        module_name = MistralCommonTokenizer.__module__
+        module_name = MistralCommonBackend.__module__
         module = importlib.import_module(module_name)
 
         # Detect what needs to be imported
@@ -79,7 +79,7 @@ def apply_mistral_tokenizer_image_patch():
         exec(patched_source, globals())  # nosec B102
 
         # Replace the method
-        MistralCommonTokenizer.apply_chat_template = patched_apply_chat_template
-        LOG.info("Successfully applied MistralCommonTokenizer tensor conversion patch")
+        MistralCommonBackend.apply_chat_template = patched_apply_chat_template
+        LOG.info("Successfully applied MistralCommonBackend tensor conversion patch")
     else:
-        LOG.warning("Could not find target code for MistralCommonTokenizer patching")
+        LOG.warning("Could not find target code for MistralCommonBackend patching")

src/axolotl/monkeypatch/relora.py

@@ -155,7 +155,6 @@ class ReLoRACallback(TrainerCallback):
                     f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}",
                     "adapter",
                 ),
-                safe_serialization=True,
             )
             with torch.no_grad():
                 merge_and_save(
@@ -214,7 +213,7 @@ class ReLoRACallback(TrainerCallback):
 
                     self.last_full_model = checkpoint_folder
             else:
-                model.model.save_pretrained(checkpoint_folder, safe_serialization=True)
+                model.model.save_pretrained(checkpoint_folder)
 
         return control
 

src/axolotl/monkeypatch/transformers/trainer_context_parallel.py

@@ -52,9 +52,15 @@ def patch_prepare_context_parallel_inputs() -> None:
         if item in patched_source:
             items_to_import.append(item)
 
-    exec(f"from {module_name} import ({', '.join(items_to_import)})", globals())
-    exec(patched_source, globals())
+    # Use a separate namespace to capture the exec'd function
+    namespace = {}
+    exec(f"from {module_name} import ({', '.join(items_to_import)})", namespace)
+    exec(patched_source, namespace)
 
+    # Explicitly get the function from the namespace
+    axolotl_prepare_context_parallel_inputs = namespace[
+        "axolotl_prepare_context_parallel_inputs"
+    ]
     Trainer._original_prepare_context_parallel_inputs = (
         Trainer._prepare_context_parallel_inputs
     )

src/axolotl/processing_strategies.py

@@ -14,7 +14,6 @@ from transformers.models.voxtral import VoxtralProcessor
 
 from axolotl.utils.dict import remove_none_values
 from axolotl.utils.logging import get_logger
-from axolotl.utils.mistral.mistral3_processor import Mistral3Processor
 
 LOG = get_logger(__name__)
 
@@ -430,7 +429,7 @@ class Mistral3ProcessingStrategy(ProcessingStrategy):
 
     def __init__(
         self,
-        processor: Mistral3Processor,
+        processor,
         chat_template: Optional[str] = None,
         image_size: int | tuple[int, int] | None = None,
         image_resize_algorithm: Resampling | None = None,
@@ -486,6 +485,58 @@ class InternVLProcessingStrategy(ProcessingStrategy):
         return labels
 
 
+class Glm4vProcessingStrategy(ProcessingStrategy):
+    """Processing Strategy class for GLM4V and GLM4V-MoE vision models."""
+
+    def __init__(
+        self,
+        processor: ProcessorMixin,
+        chat_template: Optional[str] = None,
+        image_size: int | tuple[int, int] | None = None,
+        image_resize_algorithm: Resampling | None = None,
+    ):
+        super().__init__(processor, chat_template, image_size, image_resize_algorithm)
+
+        self.tokenizer = getattr(processor, "tokenizer", processor)
+
+        self.image_token = "<|image|>"  # nosec
+        self.begin_image_token = "<|begin_of_image|>"  # nosec
+        self.end_image_token = "<|end_of_image|>"  # nosec
+        self.video_token = "<|video|>"  # nosec
+        self.begin_video_token = "<|begin_of_video|>"  # nosec
+        self.end_video_token = "<|end_of_video|>"  # nosec
+
+        self.image_token_id = self.tokenizer.convert_tokens_to_ids(self.image_token)
+        self.begin_image_token_id = self.tokenizer.convert_tokens_to_ids(
+            self.begin_image_token
+        )
+        self.end_image_token_id = self.tokenizer.convert_tokens_to_ids(
+            self.end_image_token
+        )
+        self.video_token_id = self.tokenizer.convert_tokens_to_ids(self.video_token)
+        self.begin_video_token_id = self.tokenizer.convert_tokens_to_ids(
+            self.begin_video_token
+        )
+        self.end_video_token_id = self.tokenizer.convert_tokens_to_ids(
+            self.end_video_token
+        )
+
+    def process_labels(self, input_ids):
+        labels = input_ids.clone()
+
+        labels[labels == self.tokenizer.pad_token_id] = -100
+
+        labels[labels == self.image_token_id] = -100
+        labels[labels == self.begin_image_token_id] = -100
+        labels[labels == self.end_image_token_id] = -100
+
+        labels[labels == self.video_token_id] = -100
+        labels[labels == self.begin_video_token_id] = -100
+        labels[labels == self.end_video_token_id] = -100
+
+        return labels
+
+
 def get_processing_strategy(
     processor: ProcessorMixin,
     chat_template,
@@ -493,6 +544,8 @@ def get_processing_strategy(
     image_size: int | tuple[int, int] | None = None,
     image_resize_algorithm: Resampling | None = None,
 ):
+    from axolotl.utils.mistral.mistral3_processor import Mistral3Processor
+
     processing_kwargs = {
         "processor": processor,
         "chat_template": chat_template,
@@ -500,10 +553,10 @@ def get_processing_strategy(
         "image_resize_algorithm": image_resize_algorithm,
     }
 
-    if chat_template_type in [None, "tokenizer_default"] and hasattr(
-        processor.tokenizer, "chat_template"
-    ):
-        processing_kwargs["chat_template"] = processor.tokenizer.chat_template
+    if chat_template_type in [None, "tokenizer_default"]:
+        tokenizer = getattr(processor, "tokenizer", processor)
+        if hasattr(tokenizer, "chat_template"):
+            processing_kwargs["chat_template"] = tokenizer.chat_template
 
     if chat_template_type == "qwen2_vl":
         return Qwen2VLProcessingStrategy(
@@ -532,6 +585,15 @@ def get_processing_strategy(
         return Mistral3ProcessingStrategy(
             **processing_kwargs,
         )
+    try:
+        from transformers.models.glm46v.processing_glm46v import Glm46VProcessor
+
+        if isinstance(processor, Glm46VProcessor):
+            return Glm4vProcessingStrategy(
+                **processing_kwargs,
+            )
+    except ImportError:
+        pass
 
     if isinstance(processor, InternVLProcessor):
         return InternVLProcessingStrategy(

src/axolotl/prompt_strategies/chat_template.py

@@ -150,6 +150,8 @@ class ChatTemplatePrompter(Prompter):
 
         return self.tokenizer.apply_chat_template(
             conversation,
+            tokenize=True,
+            return_dict=False,
             **chat_template_kwargs,
         )
 

src/axolotl/telemetry/callbacks.py

@@ -153,13 +153,27 @@ class TelemetryCallback(TrainerCallback):
                 self.last_report_step = step
 
     def _extract_last_metrics(self, state: TrainerState) -> dict:
-        """Extract last loss, learning_rate, and grad_norm from log history."""
+        """Extract last loss, learning_rate, grad_norm, and token metrics from log history."""
         if not state.log_history:
-            return {"loss": 0, "learning_rate": 0, "grad_norm": 0}
+            return {
+                "loss": 0,
+                "ppl": 0,
+                "learning_rate": 0,
+                "grad_norm": 0,
+                "tokens/total": 0,
+                "tokens/trainable": 0,
+                "tokens/train_per_sec_per_gpu": 0,
+            }
 
         last_log = state.log_history[-1]
         return {
             "loss": last_log.get("loss", 0),
+            "ppl": last_log.get("ppl", 0),
             "learning_rate": last_log.get("learning_rate", 0),
             "grad_norm": last_log.get("grad_norm", 0),
+            "tokens/total": last_log.get("tokens/total", 0),
+            "tokens/trainable": last_log.get("tokens/trainable", 0),
+            "tokens/train_per_sec_per_gpu": last_log.get(
+                "tokens/train_per_sec_per_gpu", 0
+            ),
         }

src/axolotl/telemetry/errors.py

@@ -155,6 +155,10 @@ def send_errors(func: Callable) -> Callable:
                     },
                 )
 
+                LOG.error(
+                    f"Error captured in telemetry. Run ID: {telemetry_manager.run_id}"
+                )
+
             raise
 
     return wrapper

src/axolotl/telemetry/manager.py

@@ -5,7 +5,6 @@ import importlib
 import logging
 import os
 import platform
-import time
 import uuid
 from pathlib import Path
 from typing import Any
@@ -20,21 +19,6 @@ LOG = logging.getLogger(__name__)
 POSTHOG_HOST = "https://app.posthog.com"
 POSTHOG_WRITE_KEY = "phc_1kUR0o04oJKKTTeSsIz2Mfm5mpiVsQEf2WOlzljMD7y"
 
-OPT_OUT_WARNING_SLEEP_SECONDS = 10
-OPT_OUT_WARNING = (
-    "\nTelemetry is now enabled by default to help improve Axolotl. "
-    "If you'd like to disable it, set AXOLOTL_DO_NOT_TRACK=1 in your environment.\n\n"
-    "Telemetry data helps us understand:\n"
-    "- Which features are most used\n"
-    "- What hardware configurations to prioritize\n"
-    "- Where users encounter errors\n\n"
-    "Personally identifiable information (PII) is not collected.\n\n"
-    "To remove this warning, explicitly set AXOLOTL_DO_NOT_TRACK=0 (enable telemetry) "
-    "or AXOLOTL_DO_NOT_TRACK=1 (disable telemetry).\n\n"
-    "For details, see: https://docs.axolotl.ai/docs/telemetry.html\n\n"
-    f"Sleeping for {OPT_OUT_WARNING_SLEEP_SECONDS}s..."
-)
-
 WHITELIST_PATH = str(Path(__file__).parent / "whitelist.yaml")
 
 # NOTE: Need to keep these up to date with any config schema changes
@@ -46,8 +30,8 @@ FIELDS_TO_REDACT = {
     "resume_from_checkpoint",
     "hub_model_id",
 }
-PREFIXES_TO_REDACT = {"wandb_", "comet_", "mlflow_", "gradio_"}
-PATH_INDICATORS = {"path", "dir"}
+PREFIXES_TO_REDACT = {"wandb_", "comet_", "mlflow_", "gradio_", "trackio_", "swanlab_"}
+PATH_INDICATORS = {"path", "dir", "data_files"}
 
 # pylint: disable=duplicate-code
 RELEVANT_PACKAGES = {
@@ -183,11 +167,6 @@ class TelemetryManager:
             "false",
             "true",
         ):
-            # Print opt-out info message for main process only
-            if is_main_process():
-                LOG.warning(OPT_OUT_WARNING)
-            time.sleep(OPT_OUT_WARNING_SLEEP_SECONDS)
-
             return True
 
         # Only rank 0 will send telemetry

src/axolotl/telemetry/whitelist.yaml

@@ -31,3 +31,10 @@ organizations:
   - "mistral-community"
   - "llava-hf"
   - "ByteDance-Seed"
+  - "ACE-Step"
+  - "openbmb"
+  - "MiniMaxAI"
+  - "stepfun-ai"
+  - "internlm"
+  - "katanemo"
+  - "XiaomiMiMo"

src/axolotl/train.py

@@ -135,16 +135,13 @@ def setup_reference_model(
     return model_ref
 
 
-def setup_signal_handler(
-    cfg: DictDefault, model: PreTrainedModel, safe_serialization: bool
-):
+def setup_signal_handler(cfg: DictDefault, model: PreTrainedModel):
     """
     Set up signal handler for graceful termination.
 
     Args:
         cfg: Dictionary mapping `axolotl` config keys to values.
         model: The model to save on termination
-        safe_serialization: Whether to use safe serialization when saving
     """
     # ray workers don't have access to this signal
     if cfg.local_rank == 0 and not cfg.use_ray:
@@ -152,9 +149,7 @@ def setup_signal_handler(
         def terminate_handler(_, __, model_weakref):
             if model_weakref() is not None:
                 _model = model_weakref()
-                _model.save_pretrained(
-                    cfg.output_dir, safe_serialization=safe_serialization
-                )
+                _model.save_pretrained(cfg.output_dir)
 
             cleanup_distributed()
             sys.exit(0)
@@ -219,7 +214,6 @@ def save_trained_model(
     cfg: DictDefault,
     trainer: Any,
     model: PreTrainedModel,
-    safe_serialization: bool,
 ):
     """
     Save the trained model according to configuration and training setup.
@@ -228,7 +222,6 @@ def save_trained_model(
         cfg: Dictionary mapping `axolotl` config keys to values.
         trainer: The trainer object.
         model: The trained model to save.
-        safe_serialization: Whether to use safe serialization.
     """
     LOG.info(f"Training completed! Saving trained model to {cfg.output_dir}.")
 
@@ -283,7 +276,6 @@ def save_trained_model(
                 merge_fsdp_weights(
                     checkpoint_dir=str(fsdp_dir),
                     output_path=merged_path,
-                    safe_serialization=True,
                 )
                 trainer.accelerator.wait_for_everyone()
                 if trainer.accelerator.is_main_process:
@@ -330,11 +322,9 @@ def save_trained_model(
                 pass
     elif cfg.local_rank == 0:
         if cfg.rl and cfg.adapter and not cfg.rl_adapter_ref_model:
-            trainer.model.save_pretrained(
-                cfg.output_dir, safe_serialization=safe_serialization
-            )
+            trainer.model.save_pretrained(cfg.output_dir)
 
-        model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
+        model.save_pretrained(cfg.output_dir)
 
     if hasattr(cfg, "llmcompressor") and cfg.llmcompressor:
         # TODO: add integration support so this can be implemented completely within the plugin
@@ -344,7 +334,6 @@ def save_trained_model(
             model=model,
             output_dir=cfg.output_dir,
             trainer=trainer,
-            safe_serialization=safe_serialization,
             save_compressed=cfg.llmcompressor.save_compressed,
         )
 
@@ -449,7 +438,6 @@ def handle_untrained_tokens_fix(
     model: PreTrainedModel,
     tokenizer: PreTrainedTokenizer,
     train_dataset: Dataset,
-    safe_serialization: bool,
 ):
     """
     Apply fixes for untrained tokens if configured.
@@ -459,7 +447,6 @@ def handle_untrained_tokens_fix(
         model: The model to apply fixes to.
         tokenizer: The tokenizer for token identification.
         train_dataset: The training dataset to use.
-        safe_serialization: Whether to use safe serialization when saving.
     """
     if not cfg.fix_untrained_tokens:
         return
@@ -483,9 +470,7 @@ def handle_untrained_tokens_fix(
     fix_untrained_tokens(model, tokenizer, train_dataset, **fix_kwargs)
 
     if cfg.local_rank == 0:
-        model.save_pretrained(
-            str(Path(cfg.output_dir)), safe_serialization=safe_serialization
-        )
+        model.save_pretrained(str(Path(cfg.output_dir)))
 
 
 def setup_model_and_trainer(
@@ -582,15 +567,12 @@ def train(
     ) = setup_model_and_trainer(cfg, dataset_meta)
 
     # Handle untrained tokens if configured
-    safe_serialization = cfg.save_safetensors is True
     train_dataset = dataset_meta.train_dataset
-    handle_untrained_tokens_fix(
-        cfg, model, tokenizer, train_dataset, safe_serialization
-    )
+    handle_untrained_tokens_fix(cfg, model, tokenizer, train_dataset)
 
     # Additional setup
     save_initial_configs(cfg, tokenizer, model, peft_config, processor)
-    setup_signal_handler(cfg, model, safe_serialization)
+    setup_signal_handler(cfg, model)
     setup_model_card(cfg)
 
     # Execute the training
@@ -602,7 +584,7 @@ def train(
         torch.cuda.empty_cache()
 
     # Save the trained model and cleanup
-    save_trained_model(cfg, trainer, model, safe_serialization)
+    save_trained_model(cfg, trainer, model)
     tokenizer.save_pretrained(
         str(Path(cfg.output_dir)), save_jinja_files=cfg.tokenizer_save_jinja_files
     )

src/axolotl/utils/callbacks/perplexity.py

@@ -7,7 +7,11 @@ from torch import Tensor
 from tqdm import tqdm
 from transformers.modeling_outputs import CausalLMOutput
 from transformers.modeling_utils import PreTrainedModel
-from transformers.tokenization_utils import PreTrainedTokenizer
+
+try:
+    from transformers.tokenization_python import PreTrainedTokenizer
+except ImportError:
+    from transformers.tokenization_utils import PreTrainedTokenizer
 
 from axolotl.utils.distributed import is_main_process
 

src/axolotl/utils/callbacks/tokens_per_second.py

@@ -78,12 +78,19 @@ class TokensPerSecondCallback(TrainerCallback):
         **kwargs,
     ):  # pylint: disable=unused-argument
         tokens = getattr(state, "tokens", None)
-        if tokens and "trainable_tokens" in tokens:
-            step_time = time.perf_counter() - self.start_time
-            num_tokens_per_device = tokens["trainable_tokens"].clone()
-            # non data parallel groups have duplicated tokens, so we avoid double-counting
-            num_tokens_per_device = num_tokens_per_device / self.non_data_parallel_size
-            state.last_tokens_per_second = num_tokens_per_device / step_time
+        if not (tokens and "trainable_tokens" in tokens):
+            return
+        step_time = time.perf_counter() - self.start_time
+        if step_time <= 0:
+            return
+
+        num_tokens = tokens["trainable_tokens"].clone() / self.non_data_parallel_size
+        if torch.distributed.is_initialized():
+            dp_size = max(
+                1, torch.distributed.get_world_size() // self.non_data_parallel_size
+            )
+            num_tokens = num_tokens / dp_size
+        state.last_tokens_per_second = num_tokens / step_time
 
     def on_log(
         self,

src/axolotl/utils/ctx_managers/sequence_parallel.py

@@ -218,6 +218,9 @@ class SequenceParallelContextManager:
         self.original_seq_len = 0
         self.pad_len = 0
 
+        # Track local valid token count for eval loss correction across CP ranks
+        self._local_valid_tokens: torch.Tensor | None = None
+
         # Create a partially applied version of the apply_sequence_parallelism function
         self.apply_sequence_parallelism = functools.partial(
             apply_sequence_parallelism,
@@ -270,6 +273,18 @@ class SequenceParallelContextManager:
                 self.apply_sequence_parallelism(updated_kwargs)
             )
 
+            # Track local valid tokens for eval loss correction
+            if "labels" in updated_kwargs and not self.models[0].training:
+                self._local_valid_tokens = (
+                    (updated_kwargs["labels"] != -100).sum().float()
+                )
+                # Strip num_items_in_batch during eval so the model uses
+                # reduction='mean', allowing the post-hook weighted all-reduce
+                # formula (loss * local_valid) to correctly recover the loss sum
+                updated_kwargs.pop("num_items_in_batch", None)
+            else:
+                self._local_valid_tokens = None
+
             return remaining_args, updated_kwargs
 
         # Forward post-hook to gather outputs
@@ -287,6 +302,44 @@ class SequenceParallelContextManager:
 
             return output
 
+        # Post-hook to correct eval loss via weighted all-reduce across CP ranks
+        def eval_loss_correction_post_hook(_, __, output: ModelOutput) -> ModelOutput:
+            if self._local_valid_tokens is None:
+                return output
+            if not hasattr(output, "loss") or output.loss is None:
+                return output
+
+            local_valid = self._local_valid_tokens.to(output.loss.device)
+            loss = output.loss.detach().clone()
+
+            # Handle rank with zero valid tokens (loss is NaN)
+            if local_valid.item() == 0:
+                weighted_loss = torch.zeros(1, device=loss.device, dtype=loss.dtype)
+            else:
+                weighted_loss = loss * local_valid
+
+            total_valid = local_valid.clone()
+            dist.all_reduce(
+                weighted_loss,
+                op=dist.ReduceOp.SUM,
+                group=self.process_group,
+            )
+            dist.all_reduce(
+                total_valid,
+                op=dist.ReduceOp.SUM,
+                group=self.process_group,
+            )
+
+            if total_valid.item() > 0:
+                output["loss"] = (weighted_loss / total_valid).squeeze()
+            else:
+                output["loss"] = torch.tensor(
+                    float("nan"), device=loss.device, dtype=loss.dtype
+                )
+
+            self._local_valid_tokens = None
+            return output
+
         # Register hooks
         for model in self.models:
             self.hook_handles.append(
@@ -298,6 +351,10 @@ class SequenceParallelContextManager:
                 self.hook_handles.append(
                     model.register_forward_hook(sequence_parallel_post_hook)
                 )
+            # Always register eval loss correction hook
+            self.hook_handles.append(
+                model.register_forward_hook(eval_loss_correction_post_hook)
+            )
 
     def _gather_outputs(self, output: CausalLMOutputWithPast) -> CausalLMOutputWithPast:
         """Gather sharded outputs from all ranks and reconstruct the full tensor."""

src/axolotl/utils/data/rl.py

@@ -173,7 +173,7 @@ def _drop_long_sequences(
 
         return (len_prompt + len_completion) <= sequence_len
 
-    if rl is RLType.GRPO:
+    if rl in {RLType.GRPO, RLType.GDPO}:
         return True
 
     raise ValueError("Unknown RL type")