properly handles kernels repo type

* [gemma4] fix VRAM leak in hybrid FA2+SDPA path under activation checkpointing

Route shared_kv_states through a thread-local side channel instead of the
decoder-layer kwargs so the checkpoint partial never references the dict.

HF's Gemma4TextModel.forward passes shared_kv_states (a mutable dict used
for cross-layer K/V sharing) as a kwarg to every decoder_layer call.
GradientCheckpointingLayer.__call__ then forms
partial(super().__call__, **kwargs), and whichever checkpoint runs
(axolotl's CPU_Offloaded_Gradient_Checkpointer or torch's stock
checkpoint) captures that partial. The partial holds a reference to the
dict, which holds the K/V tensors produced by store_full_length_kv
layers. Those tensors stay pinned for the full duration of backward, and
delayed ref-cycle cleanup in torch's caching allocator under FSDP2 +
activation checkpointing bleeds the residual across steps.

Observed symptom: VRAM climbs ~0.47 GiB/step from a 42 GiB baseline,
OOMs around step 73 (~94 GiB peak) on Gemma-4 31B multimodal with
gemma4_hybrid_attn_impl: true. Independent of seq len / image size.
All-flex-attention path is flat but ~22x slower.

Violated invariant: anything crossing an activation-checkpoint boundary
must be a tensor (refcounted by autograd) or plain Python data -- never
a mutable container holding tensor references.

Fix (all in src/axolotl/monkeypatch/models/gemma4/fused_attn.py):
  * threading.local() store with _get/_set_shared_kv_states helpers
  * _patch_decoder_layer_call(): monkeypatches
    Gemma4TextDecoderLayer.__call__ to pop shared_kv_states from kwargs
    and stash it in TLS before delegating to GradientCheckpointingLayer.
    The partial formed downstream no longer references the dict.
  * fused_forward reads TLS first, falls back to kwarg for callers that
    bypass the patched __call__ (e.g. direct attention invocation).
  * wired into patch_gemma4_fused_attn; idempotent via a sentinel.

TLS is overwritten on each new step's first decoder-layer call, so the
previous step's dict is released promptly. No changes to hybrid dispatch,
FSDP wrap policy, or any config behaviour. Works for hybrid, flex, and
eager paths.

Introduced by PR #3598 (commit b8358aa5).

* Coderabbit comment: gemma4: clear TLS unconditionally in decoder-layer patched __call__

  Overwrite the thread-local shared_kv_states store on every invocation
  (including with None) instead of only when the kwarg is present.

  The previous conditional write left stale dicts in TLS on any path that
  reaches Gemma4TextDecoderLayer.__call__ without a shared_kv_states
  kwarg — e.g. generation, eval hooks, or future HF refactors that make
  the kwarg optional. fused_forward would then silently consume a prior
  step's K/V dict instead of falling back to its own kwarg path.

  Unconditional write makes the invariant in the surrounding comment
  ("TLS is overwritten on each new step's first decoder-layer call, so
  the previous step's dict is released promptly") actually hold.

  No behavior change for the training happy path, which always passes
  the kwarg. Addresses CodeRabbit review on PR #3611

* fix: swap threading.local() for module-level store so autograd worker   threads see shared_kv_states during backward recompute

Previous commits fixed memory leak on 31B but caused type error with MOE Gemma4 variants - this fixes that:

PR 3611's TLS variant only works when recompute runs on the same thread
  that set TLS during forward. PyTorch's C++ autograd engine
  (_engine_run_backward) spawns per-device worker threads to dispatch
  backward, and HF-Trainer gradient_checkpointing (stock
  torch.utils.checkpoint, non-reentrant / saved-tensor-hooks) fires
  unpack_hook -> recompute_fn on those worker threads. TLS set on the main
  thread during forward is invisible there, so _get_shared_kv_states()
  returns None and the consumer-layer lookup crashes with
  "'NoneType' object is not subscriptable" at
  fused_attn.py:97 (shared_kv_states[self.kv_shared_layer_index]).

  A plain module-level dict is visible to all threads in the process.
  Lifecycle is identical: the slot is overwritten each forward, releasing
  the previous step's dict and allowing its K/V tensors to be GC'd, so
  the original VRAM-leak fix still holds under FSDP2 AC too.

* scope gemma4 shared_kv_states side channel to checkpointed training

Update PR #3611 with gate for checkpointed training to avoid regressions across async flows.

Added unit tests for kwargs pop, store-clear regression, and flag gating. Condensed verbose comments

* add gemma4 cross-thread visibility test for shared_kv_states store

Additional regression test for MoE gemma4 variants - asserts the module-level store is readable from threads other than the one that set it in response to previously observed 'NoneType' error

* fix logger

---------

Co-authored-by: Wing Lian <wing@axolotl.ai>

.github/CONTRIBUTING.md

@@ -31,7 +31,10 @@ PRs are **greatly welcome**!
 
 Please run below to setup env
 ```bash
-pip3 install -r requirements-dev.txt -r requirements-tests.txt
+# Install axolotl + dev and test dependencies from lockfile
+export UV_TORCH_BACKEND=cu128  # or cu130
+uv sync --extra flash-attn --extra deepspeed --group dev --group test
+source .venv/bin/activate
 pre-commit install
 
 # test

.github/workflows/lint.yml

@@ -6,7 +6,7 @@ on:
       types: [opened, synchronize, reopened, ready_for_review]
       paths:
        - '**.py'
-       - 'requirements.txt'
+       - 'pyproject.toml'
        - '.github/workflows/*.yml'
        - "*.[q]md"
        - "examples/**/*.y[a]?ml"

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

@@ -3,17 +3,15 @@ name: docker-multigpu-tests-biweekly
 on:
   pull_request:
     paths:
-      - 'tests/e2e/multigpu/**.py'
-      - 'requirements.txt'
-      - 'setup.py'
-      - 'pyproject.toml'
-      - '.github/workflows/multi-gpu-e2e.yml'
-      - 'scripts/cutcrossentropy_install.py'
-      - 'src/axolotl/core/trainers/mixins/sequence_parallel.py'
-      - 'src/axolotl/utils/distributed.py'
+      - "tests/e2e/multigpu/**.py"
+      - "pyproject.toml"
+      - ".github/workflows/multi-gpu-e2e.yml"
+      - "scripts/cutcrossentropy_install.py"
+      - "src/axolotl/core/trainers/mixins/sequence_parallel.py"
+      - "src/axolotl/utils/distributed.py"
   workflow_dispatch:
   schedule:
-    - cron: '0 0 * * 1,4'  # Runs at 00:00 UTC every monday & thursday
+    - cron: "0 0 * * 1,4" # Runs at 00:00 UTC every monday & thursday
 
 # Cancel jobs on the same ref if a new one is triggered
 concurrency:
@@ -33,19 +31,19 @@ jobs:
       fail-fast: false
       matrix:
         include:
-#          - 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: 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"
             pytorch: 2.9.1
             axolotl_extras:
-#            axolotl_extras: fbgemm-gpu
+            #            axolotl_extras: fbgemm-gpu
             num_gpus: 2
           - cuda: 128
             cuda_version: 12.8.1
@@ -53,7 +51,6 @@ jobs:
             pytorch: 2.10.0
             axolotl_extras: "fbgemm-gpu"
             num_gpus: 2
-            dockerfile: "Dockerfile-uv.jinja"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:
@@ -75,7 +72,7 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
+          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile-uv.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         env:
           CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}

.github/workflows/pypi.yml

@@ -8,6 +8,9 @@ on:
 
 permissions: {}
 
+env:
+  UV_SYSTEM_PYTHON: "1"
+
 jobs:
   setup_release:
     name: Create Release
@@ -41,11 +44,15 @@ jobs:
         with:
           python-version: "3.11"
 
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
+
       - name: Install dependencies
         run: |
-          pip3 install wheel packaging==26.0
-          pip3 install --no-build-isolation -e .
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+          uv pip install wheel packaging
+          uv pip install --no-build-isolation -e .
+          uv pip install black mypy pre-commit types-requests quartodoc jupyter blobfile tiktoken \
+            codecov codecov-cli pytest pytest-cov pytest-retry pytest-sugar pytest-xdist tbparse
 
       - name: Extract tag name
         id: tag

.github/workflows/tests-nightly.yml

@@ -2,15 +2,18 @@ name: Tests Nightly against upstream main
 on:
   workflow_dispatch:
   schedule:
-    - cron: '0 0 * * *'  # Runs at 00:00 UTC every day
+    - cron: "0 0 * * *" # Runs at 00:00 UTC every day
   pull_request:
     types: [opened, synchronize, reopened, ready_for_review]
     paths:
-      - '.github/workflows/tests-nightly.yml'
+      - ".github/workflows/tests-nightly.yml"
 
 permissions:
   contents: read
 
+env:
+  UV_SYSTEM_PYTHON: "1"
+
 jobs:
   pre-commit:
     name: pre-commit
@@ -20,7 +23,7 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-          cache: 'pip' # caching pip dependencies
+          cache: "pip" # caching pip dependencies
       - uses: pre-commit/action@v3.0.1
         env:
           SKIP: no-commit-to-branch
@@ -43,7 +46,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.12"]  # TODO include py3.14 once https://github.com/mistralai/mistral-common/pull/194 is merged
+        python_version: ["3.12"] # TODO include py3.14 once https://github.com/mistralai/mistral-common/pull/194 is merged
         pytorch_version: ["2.9.1", "2.10.0"]
     timeout-minutes: 20
 
@@ -61,36 +64,34 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python_version }}
-          cache: 'pip' # caching pip dependencies
 
-      - name: upgrade pip
-        run: |
-          pip3 install --upgrade pip
-          pip3 install --upgrade packaging==26.0 setuptools==78.1.1 wheel
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }} torchvision
-
-      - name: Update requirements.txt
-        run: |
-          sed -i 's#^transformers.*#transformers @ git+https://github.com/huggingface/transformers.git@main#' requirements.txt
-          sed -i 's#^peft.*#peft @ git+https://github.com/huggingface/peft.git@main#' requirements.txt
-          sed -i 's#^accelerate.*#accelerate @ git+https://github.com/huggingface/accelerate.git@main#' requirements.txt
-          sed -i 's#^trl.*#trl @ git+https://github.com/huggingface/trl.git@main#' requirements.txt
-          sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt
+          uv pip install torch==${{ matrix.pytorch_version }} torchvision
+          uv pip freeze | grep -E "^(torch|torchvision)==" > /tmp/torch-pin.txt
 
       - name: Install dependencies
         run: |
-          pip3 show torch
-          pip3 install --no-build-isolation -U -e .
-          python scripts/unsloth_install.py | sh
-          python scripts/cutcrossentropy_install.py | sh
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+          uv pip install --no-build-isolation -e . --override /tmp/torch-pin.txt
+          python scripts/cutcrossentropy_install.py --uv | sh
+          uv pip install black mypy pre-commit types-requests quartodoc jupyter blobfile tiktoken \
+            codecov codecov-cli pytest pytest-cov pytest-retry pytest-sugar pytest-xdist tbparse
+
+      - name: Override with nightly HF packages
+        run: |
+          uv pip install --no-deps \
+            "transformers @ git+https://github.com/huggingface/transformers.git@main" \
+            "peft @ git+https://github.com/huggingface/peft.git@main" \
+            "accelerate @ git+https://github.com/huggingface/accelerate.git@main" \
+            "trl @ git+https://github.com/huggingface/trl.git@main" \
+            "datasets @ git+https://github.com/huggingface/datasets.git@main"
 
       - name: Make sure PyTorch version wasn't clobbered
         run: |
-          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__, f'Expected torch ${{ matrix.pytorch_version }} but got {torch.__version__}'"
 
       - name: Ensure axolotl CLI was installed
         run: |
@@ -102,9 +103,6 @@ jobs:
           pytest -v --durations=10 tests/patched/
           pytest -v --durations=10 tests/cli/
 
-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
   docker-e2e-tests:
     if: github.repository_owner == 'axolotl-ai-cloud'
@@ -136,7 +134,6 @@ jobs:
             pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
-            dockerfile: "Dockerfile-uv.jinja"
             nightly_build: "true"
     steps:
       - name: Checkout
@@ -157,7 +154,7 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
+          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile-uv.jinja'}}" >> $GITHUB_ENV
           echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         env:

.github/workflows/tests.yml

@@ -6,21 +6,19 @@ on:
     branches:
       - "main"
     paths:
-      - '**.py'
-      - 'requirements.txt'
-      - '.github/workflows/*.yml'
-      - 'requirements-tests.txt'
-      - 'cicd/cicd.sh'
-      - 'cicd/Dockerfile.jinja'
+      - "**.py"
+      - "pyproject.toml"
+      - ".github/workflows/*.yml"
+      - "cicd/cicd.sh"
+      - "cicd/Dockerfile-uv.jinja"
   pull_request:
-      types: [opened, synchronize, reopened, ready_for_review]
-      paths:
-       - '**.py'
-       - 'requirements.txt'
-       - '.github/workflows/*.yml'
-       - 'requirements-tests.txt'
-       - 'cicd/cicd.sh'
-       - 'cicd/Dockerfile.jinja'
+    types: [opened, synchronize, reopened, ready_for_review]
+    paths:
+      - "**.py"
+      - "pyproject.toml"
+      - ".github/workflows/*.yml"
+      - "cicd/cicd.sh"
+      - "cicd/Dockerfile-uv.jinja"
   workflow_dispatch:
 
 # Cancel jobs on the same ref if a new one is triggered
@@ -33,6 +31,7 @@ permissions:
 
 env:
   TRANSFORMERS_IS_CI: "yes"
+  UV_SYSTEM_PYTHON: "1"
 
 jobs:
   pre-commit:
@@ -44,7 +43,7 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-          cache: 'pip' # caching pip dependencies
+          cache: "pip" # caching pip dependencies
       - uses: pre-commit/action@v3.0.1
         env:
           SKIP: no-commit-to-branch
@@ -94,32 +93,25 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python_version }}
-          cache: 'pip' # caching pip dependencies
 
-      - name: upgrade pip
-        run: |
-          pip3 install --upgrade pip
-          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 wheel
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
 
       - name: Install PyTorch
         run: |
-          pip3 install --no-cache-dir torch==${{ matrix.pytorch_version }} torchvision
+          uv pip install torch==${{ matrix.pytorch_version }} torchvision
+          uv pip freeze | grep -E "^(torch|torchvision)==" > /tmp/torch-pin.txt
 
       - name: Install dependencies
         run: |
-          pip3 show torch
-          pip3 install --no-cache-dir --no-build-isolation -U -e .
-          python scripts/unsloth_install.py | sh
-          python scripts/cutcrossentropy_install.py | sh
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
-
-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+          uv pip install --no-build-isolation -e . --override /tmp/torch-pin.txt
+          python scripts/cutcrossentropy_install.py --uv | sh
+          uv pip install black mypy pre-commit types-requests quartodoc jupyter blobfile tiktoken \
+            codecov codecov-cli pytest pytest-cov pytest-retry pytest-sugar pytest-xdist tbparse
 
       - name: Make sure PyTorch version wasn't clobbered
         run: |
-          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__, f'Expected torch ${{ matrix.pytorch_version }} but got {torch.__version__}'"
 
       - name: Ensure axolotl CLI was installed
         run: |
@@ -188,33 +180,27 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python_version }}
-          cache: 'pip' # caching pip dependencies
 
-      - name: upgrade pip
-        run: |
-          pip3 install --upgrade pip
-          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 setuptools_scm build wheel psutil
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
 
       - name: Install PyTorch
         run: |
-          pip3 install --no-cache-dir torch==${{ matrix.pytorch_version }} torchvision
+          uv pip install torch==${{ matrix.pytorch_version }} torchvision
+          uv pip freeze | grep -E "^(torch|torchvision)==" > /tmp/torch-pin.txt
 
       - name: Install dependencies
         run: |
-          pip3 show torch
+          uv pip install packaging setuptools_scm build wheel psutil
           python -m build --no-isolation --sdist
-          pip3 install --no-cache-dir --no-build-isolation dist/axolotl*.tar.gz
-          python scripts/unsloth_install.py | sh
-          python scripts/cutcrossentropy_install.py | sh
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
-
-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+          uv pip install --no-build-isolation dist/axolotl*.tar.gz --override /tmp/torch-pin.txt
+          python scripts/cutcrossentropy_install.py --uv | sh
+          uv pip install black mypy pre-commit types-requests quartodoc jupyter blobfile tiktoken \
+            codecov codecov-cli pytest pytest-cov pytest-retry pytest-sugar pytest-xdist tbparse
 
       - name: Make sure PyTorch version wasn't clobbered
         run: |
-          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__, f'Expected torch ${{ matrix.pytorch_version }} but got {torch.__version__}'"
 
       - name: Ensure axolotl CLI was installed
         run: |
@@ -291,7 +277,6 @@ jobs:
             pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
-            dockerfile: "Dockerfile-uv.jinja"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -312,7 +297,7 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
+          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile-uv.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         env:
           CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
@@ -374,7 +359,7 @@ jobs:
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
           echo "GPU_TYPE=${{ matrix.gpu_type || 'L40S'}}" >> $GITHUB_ENV
-          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
+          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile-uv.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         env:
           CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}

MANIFEST.in

@@ -1,7 +1,6 @@
-include requirements.txt
 include README.md
 include LICENSE
-include src/setuptools_axolotl_dynamic_dependencies.py
+include VERSION
 include src/axolotl/utils/chat_templates/templates/*.jinja
 include AGENTS.md
 recursive-include docs/agents *.md

README.md

@@ -95,14 +95,11 @@ Features:
 
 ### Installation
 
-#### Using uv (recommended)
-
 ```bash
-# install uv if you don't already have it installed
+# install uv if you don't already have it installed (restart shell after)
 curl -LsSf https://astral.sh/uv/install.sh | sh
-source $HOME/.local/bin/env
 
-# CUDA 12.8.1 tends to have better package compatibility
+# change depending on system
 export UV_TORCH_BACKEND=cu128
 
 # create a new virtual environment
@@ -112,23 +109,6 @@ source .venv/bin/activate
 uv pip install torch==2.10.0 torchvision
 uv pip install --no-build-isolation axolotl[deepspeed]
 
-# recommended - install cut-cross-entropy
-uv pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@main"
-
-# (optional) - prefetch flash-attn2 and causal-conv1d kernels
-uv run --python 3.12 python -c "from kernels import get_kernel; get_kernel('kernels-community/flash-attn2'); get_kernel('kernels-community/causal-conv1d')"
-
-# Download example axolotl configs, deepspeed configs
-axolotl fetch examples
-axolotl fetch deepspeed_configs  # OPTIONAL
-```
-
-#### Using pip
-
-```bash
-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
 axolotl fetch examples
 axolotl fetch deepspeed_configs  # OPTIONAL
@@ -138,7 +118,7 @@ axolotl fetch deepspeed_configs  # OPTIONAL
 
 Installing with Docker can be less error prone than installing in your own environment.
 ```bash
-docker run --gpus '"all"' --rm -it axolotlai/axolotl:main-latest
+docker run --gpus '"all"' --ipc=host --rm -it axolotlai/axolotl:main-latest
 ```
 
 Other installation approaches are described [here](https://docs.axolotl.ai/docs/installation.html).

_quarto.yml

@@ -134,7 +134,6 @@ quartodoc:
         - monkeypatch.stablelm_attn_hijack_flash
         - monkeypatch.trainer_fsdp_optim
         - monkeypatch.transformers_fa_utils
-        - monkeypatch.unsloth_
         - monkeypatch.data.batch_dataset_fetcher
         - monkeypatch.mixtral
         - monkeypatch.gradient_checkpointing.offload_cpu
@@ -327,7 +326,6 @@ website:
         - section: "Advanced Features"
           contents:
             - docs/fsdp_qlora.qmd
-            - docs/unsloth.qmd
             - docs/torchao.qmd
             - docs/custom_integrations.qmd
             - docs/sequence_parallelism.qmd

cicd/Dockerfile-uv.jinja

@@ -22,15 +22,6 @@ WORKDIR /workspace/axolotl
 RUN git fetch origin +$GITHUB_REF && \
     git checkout FETCH_HEAD
 
-# If AXOLOTL_EXTRAS is set, append it in brackets
-RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
-        sed -i 's#^transformers.*#transformers @ git+https://github.com/huggingface/transformers.git@main#' requirements.txt; \
-        sed -i 's#^peft.*#peft @ git+https://github.com/huggingface/peft.git@main#' requirements.txt; \
-        sed -i 's#^accelerate.*#accelerate @ git+https://github.com/huggingface/accelerate.git@main#' requirements.txt; \
-        sed -i 's#^trl.*#trl @ git+https://github.com/huggingface/trl.git@main#' requirements.txt; \
-        sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
-    fi
-
 RUN uv pip install packaging==26.0 setuptools==78.1.1
 RUN uv pip install torchvision
 RUN uv pip uninstall causal_conv1d
@@ -40,11 +31,21 @@ RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         uv pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
     fi
 
-RUN python scripts/unsloth_install.py --uv | sh
+# Override with nightly HF packages for nightly builds
+RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
+        uv pip install --no-deps \
+            "transformers @ git+https://github.com/huggingface/transformers.git@main" \
+            "peft @ git+https://github.com/huggingface/peft.git@main" \
+            "accelerate @ git+https://github.com/huggingface/accelerate.git@main" \
+            "trl @ git+https://github.com/huggingface/trl.git@main" \
+            "datasets @ git+https://github.com/huggingface/datasets.git@main"; \
+    fi
+
 RUN python scripts/cutcrossentropy_install.py --uv | sh
 
 # So we can test the Docker image
-RUN uv pip install -r requirements-dev.txt -r requirements-tests.txt
+RUN uv pip install black mypy pre-commit types-requests quartodoc jupyter blobfile tiktoken \
+    codecov codecov-cli pytest pytest-cov pytest-retry pytest-sugar pytest-xdist tbparse
 
 # fix so that git fetch/pull from remote works
 RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \

cicd/Dockerfile.jinja

@@ -1,54 +0,0 @@
-FROM axolotlai/axolotl-base:{{ BASE_TAG }}
-
-ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-ENV AXOLOTL_EXTRAS="{{ AXOLOTL_EXTRAS }}"
-ENV AXOLOTL_ARGS="{{ AXOLOTL_ARGS }}"
-ENV CUDA="{{ CUDA }}"
-ENV PYTORCH_VERSION="{{ PYTORCH_VERSION }}"
-ENV GITHUB_REF="{{ GITHUB_REF }}"
-ENV GITHUB_SHA="{{ GITHUB_SHA }}"
-ENV NIGHTLY_BUILD="{{ NIGHTLY_BUILD }}"
-ENV HF_HOME="{{ HF_HOME }}"
-ENV AXOLOTL_DATASET_NUM_PROC="8"
-
-RUN apt-get update && \
-    apt-get install -y --allow-change-held-packages vim curl nano zstd libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
-
-WORKDIR /workspace
-
-RUN git clone --depth=1 https://github.com/axolotl-ai-cloud/axolotl.git
-
-WORKDIR /workspace/axolotl
-
-RUN git fetch origin +$GITHUB_REF && \
-    git checkout FETCH_HEAD
-
-# If AXOLOTL_EXTRAS is set, append it in brackets
-RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
-        sed -i 's#^transformers.*#transformers @ git+https://github.com/huggingface/transformers.git@main#' requirements.txt; \
-        sed -i 's#^peft.*#peft @ git+https://github.com/huggingface/peft.git@main#' requirements.txt; \
-        sed -i 's#^accelerate.*#accelerate @ git+https://github.com/huggingface/accelerate.git@main#' requirements.txt; \
-        sed -i 's#^trl.*#trl @ git+https://github.com/huggingface/trl.git@main#' requirements.txt; \
-        sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
-    fi
-
-RUN pip install packaging==26.0 setuptools==78.1.1 psutil
-RUN pip uninstall -y causal_conv1d
-RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
-    else \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
-    fi
-
-RUN python scripts/unsloth_install.py | sh
-RUN python scripts/cutcrossentropy_install.py | sh
-
-# So we can test the Docker image
-RUN pip install -r requirements-dev.txt -r requirements-tests.txt
-
-# fix so that git fetch/pull from remote works
-RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \
-    git config --get remote.origin.fetch
-
-# helper for huggingface-login cli
-RUN git config --global credential.helper store

cicd/cicd.sh

@@ -1,7 +1,7 @@
 #!/bin/bash
 set -e
 
-python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
+python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__, f'Expected torch $PYTORCH_VERSION but got {torch.__version__}'"
 
 set -o pipefail
 for i in 1 2 3; do

cicd/multigpu.py

@@ -17,7 +17,7 @@ template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
 template_env = jinja2.Environment(
     loader=template_loader, autoescape=select_autoescape()
 )
-dockerfile = os.environ.get("E2E_DOCKERFILE", "Dockerfile.jinja")
+dockerfile = os.environ.get("E2E_DOCKERFILE", "Dockerfile-uv.jinja")
 df_template = template_env.get_template(dockerfile)
 
 df_args = {

cicd/single_gpu.py

@@ -16,7 +16,7 @@ template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
 template_env = jinja2.Environment(
     loader=template_loader, autoescape=select_autoescape()
 )
-dockerfile = os.environ.get("E2E_DOCKERFILE", "Dockerfile.jinja")
+dockerfile = os.environ.get("E2E_DOCKERFILE", "Dockerfile-uv.jinja")
 df_template = template_env.get_template(dockerfile)
 
 df_args = {

docker/Dockerfile

@@ -32,7 +32,7 @@ RUN if [ "$TARGETARCH" = "arm64" ]; then \
         pip install --no-build-isolation -e .[$BASE_EXTRAS,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
         pip install --no-build-isolation -e .[$BASE_EXTRAS] $AXOLOTL_ARGS; \
-    fi && \    python scripts/unsloth_install.py | sh && \
+    fi && \
     python scripts/cutcrossentropy_install.py | sh && \
     pip install pytest && \
     pip cache purge

docker/Dockerfile-uv

@@ -33,7 +33,6 @@ RUN if [ "$TARGETARCH" = "arm64" ]; then \
     else \
         uv pip install --no-build-isolation -e .[$BASE_EXTRAS] $AXOLOTL_ARGS; \
     fi && \
-    python scripts/unsloth_install.py --uv | sh && \
     python scripts/cutcrossentropy_install.py --uv | sh && \
     uv pip install pytest && \
     uv cache clean

docs/debugging.qmd

@@ -76,8 +76,9 @@ datasets:
 Make sure you have an [editable install](https://setuptools.pypa.io/en/latest/userguide/development_mode.html) of Axolotl, which ensures that changes you make to the code are reflected at runtime.  Run the following commands from the root of this project:
 
 ```bash
-pip3 install packaging
-pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
+export UV_TORCH_BACKEND=cu128  # or cu130
+uv sync --extra flash-attn --extra deepspeed --group dev --group test
+source .venv/bin/activate
 ```
 
 #### Remote Hosts
@@ -208,17 +209,17 @@ cd axolotl
 Next, run the desired docker image and mount the current directory. Below is a docker command you can run to do this:[^2]
 
 ```bash
-docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src="${PWD}",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface axolotlai/axolotl:main-py3.10-cu118-2.0.1
+docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src="${PWD}",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface axolotlai/axolotl-uv:main-latest
 ```
 
 >[!Tip]
 > To understand which containers are available, see the [Docker section of the README](../README.md#docker) and the [DockerHub repo](https://hub.docker.com/r/axolotlai/axolotl/tags).  For details of how the Docker containers are built, see axolotl's [Docker CI builds](../.github/workflows/main.yml).
 
-You will now be in the container.  Next, perform an editable install of Axolotl:
+You will now be in the container.  Next, install Axolotl with dev dependencies:
 
 ```bash
-pip3 install packaging
-pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
+uv sync --extra flash-attn --extra deepspeed --group dev --group test
+source .venv/bin/activate
 ```
 
 ### Attach To Container

docs/docker.qmd

@@ -6,23 +6,30 @@ format:
     toc-depth: 4
 ---
 
-This section describes the different Docker images that are released by AxolotlAI at [Docker Hub](https://hub.docker.com/u/axolotlai).
+This section describes the different Docker images that are released by AxolotlAI at
+[Docker Hub](https://hub.docker.com/u/axolotlai).
 
 ::: {.callout-important}
-For Blackwell GPUs, please use the tags with PyTorch 2.7.1 and CUDA 12.8.
+For Blackwell GPUs, please use the tags with PyTorch 2.9.1 and CUDA 12.8.
+:::
+
+::: {.callout-tip}
+Each image below is available in a **uv variant** that uses [uv](https://docs.astral.sh/uv/) with
+a relocatable venv (`/workspace/axolotl-venv`) instead of Miniconda + pip. Append `-uv` to the image name
+(e.g. `axolotlai/axolotl-base-uv`). Tags follow the same format. We recommend the uv images for new deployments.
 :::
 
 ## Base
 
-The base image is the most minimal image that can install Axolotl. It is based on the `nvidia/cuda` image. It includes python, torch, git, git-lfs, awscli, pydantic, and more.
+The base image is the most minimal image that can install Axolotl. It is based on the `nvidia/cuda` image.
+It includes python, torch, git, git-lfs, awscli, pydantic, and more.
 
 #### Image
 
-```
-axolotlai/axolotl-base
-```
-
-Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl-base)
+| Variant | Image | Docker Hub |
+|---------|-------|------------|
+| pip | `axolotlai/axolotl-base` | [Link](https://hub.docker.com/r/axolotlai/axolotl-base) |
+| uv | `axolotlai/axolotl-base-uv` | [Link](https://hub.docker.com/r/axolotlai/axolotl-base-uv) |
 
 #### Tags format
 
@@ -32,8 +39,10 @@ main-base-py{python_version}-cu{cuda_version}-{pytorch_version}
 
 Tags examples:
 
-- `main-base-py3.11-cu128-2.8.0`
 - `main-base-py3.11-cu128-2.9.1`
+- `main-base-py3.12-cu128-2.10.0`
+- `main-base-py3.12-cu130-2.9.1`
+- `main-base-py3.12-cu130-2.10.0`
 
 ## Main
 
@@ -41,11 +50,10 @@ The main image is the image that is used to run Axolotl. It is based on the `axo
 
 #### Image
 
-```
-axolotlai/axolotl
-```
-
-Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl)
+| Variant | Image | Docker Hub |
+|---------|-------|------------|
+| pip | `axolotlai/axolotl` | [Link](https://hub.docker.com/r/axolotlai/axolotl) |
+| uv | `axolotlai/axolotl-uv` | [Link](https://hub.docker.com/r/axolotlai/axolotl-uv) |
 
 #### Tags format {#sec-main-tags}
 
@@ -53,7 +61,7 @@ Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl)
 # on push to main
 main-py{python_version}-cu{cuda_version}-{pytorch_version}
 
-# latest main (currently torch 2.6.0, python 3.11, cuda 12.4)
+# latest main (currently torch 2.9.1, python 3.11, cuda 12.8)
 main-latest
 
 # nightly build
@@ -71,11 +79,12 @@ There may be some extra tags appended to the image, like `-vllm` which installs
 
 Tags examples:
 
-- `main-py3.11-cu128-2.8.0`
 - `main-py3.11-cu128-2.9.1`
+- `main-py3.12-cu128-2.10.0`
+- `main-py3.12-cu130-2.9.1`
+- `main-py3.12-cu130-2.10.0`
 - `main-latest`
-- `main-20250303-py3.11-cu124-2.6.0`
-- `main-20250303-py3.11-cu126-2.6.0`
+- `main-20260315-py3.11-cu128-2.9.1`
 - `0.12.0`
 
 ## Cloud
@@ -90,11 +99,10 @@ Jupyter lab is run by default. Set `JUPYTER_DISABLE=1` in the environment variab
 
 #### Image
 
-```
-axolotlai/axolotl-cloud
-```
-
-Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl-cloud)
+| Variant | Image | Docker Hub |
+|---------|-------|------------|
+| pip | `axolotlai/axolotl-cloud` | [Link](https://hub.docker.com/r/axolotlai/axolotl-cloud) |
+| uv | `axolotlai/axolotl-cloud-uv` | [Link](https://hub.docker.com/r/axolotlai/axolotl-cloud-uv) |
 
 #### Tags format
 

docs/installation.qmd

@@ -15,64 +15,30 @@ This guide covers all the ways you can install and set up Axolotl for your envir
 
 - NVIDIA GPU (Ampere architecture or newer for `bf16` and Flash Attention) or AMD GPU
 - Python ≥3.11
-- PyTorch ≥2.6.0
+- PyTorch ≥2.9.0
 
-## Installation Methods {#sec-installation-methods}
-
-::: {.callout-important}
-Please make sure to have Pytorch installed before installing Axolotl in your local environment.
-
-Follow the instructions at: [https://pytorch.org/get-started/locally/](https://pytorch.org/get-started/locally/)
-:::
+## Installation {#sec-installation}
 
 ::: {.callout-important}
 For Blackwell GPUs, please use Pytorch 2.9.1 and CUDA 12.8.
 :::
 
-### PyPI Installation (Recommended) {#sec-pypi}
+### Quick Install {#sec-uv}
 
-```{.bash}
-pip3 install -U packaging setuptools wheel ninja
-pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
-```
+Axolotl uses [uv](https://docs.astral.sh/uv/) as its package manager. uv is a fast, reliable Python package installer and resolver built in Rust.
 
-We use `--no-build-isolation` in order to detect the installed PyTorch version (if
-installed) in order not to clobber it, and so that we set the correct version of
-dependencies that are specific to the PyTorch version or other installed
-co-dependencies.
-
-### uv Installation {#sec-uv}
-
-uv is a fast, reliable Python package installer and resolver built in Rust. It offers significant performance improvements over pip and provides better dependency resolution, making it an excellent choice for complex environments.
-
-Install uv if not already installed
+Install uv if not already installed:
 ```{.bash}
 curl -LsSf https://astral.sh/uv/install.sh | sh
 source $HOME/.local/bin/env
 ```
 
-Choose your CUDA version to use with PyTorch; e.g. `cu124`, `cu126`, `cu128`,
-then create the venv and activate
+Choose your CUDA version (e.g. `cu128`, `cu130`), create a venv, and install:
 ```{.bash}
-export UV_TORCH_BACKEND=cu126
+export UV_TORCH_BACKEND=cu128  # or cu130
 uv venv --no-project --relocatable
 source .venv/bin/activate
-```
-
-Install PyTorch
-- PyTorch 2.6.0 recommended
-```{.bash}
-uv pip install packaging setuptools wheel
-uv pip install torch==2.6.0
-uv pip install awscli pydantic
-```
-
-Install axolotl from PyPi
-```{.bash}
-uv pip install --no-build-isolation axolotl[deepspeed,flash-attn]
-
-# optionally install with vLLM if you're using torch==2.6.0 and want to train w/ GRPO
-uv pip install --no-build-isolation axolotl[deepspeed,flash-attn,vllm]
+uv pip install --no-build-isolation axolotl[flash-attn,deepspeed]
 ```
 
 ### Edge/Development Build {#sec-edge-build}
@@ -82,14 +48,17 @@ For the latest features between releases:
 ```{.bash}
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
-pip3 install -U packaging setuptools wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
+export UV_TORCH_BACKEND=cu128  # or cu130
+uv sync --extra flash-attn --extra deepspeed
+source .venv/bin/activate
 ```
 
+`uv sync` creates a `.venv`, installs exact pinned versions from `uv.lock`, and sets up an editable install automatically.
+
 ### Docker {#sec-docker}
 
 ```{.bash}
-docker run --gpus '"all"' --rm -it axolotlai/axolotl:main-latest
+docker run --gpus '"all"' --rm -it --ipc=host axolotlai/axolotl-uv:main-latest
 ```
 
 For development with Docker:
@@ -106,12 +75,12 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it \
   --ulimit memlock=-1 --ulimit stack=67108864 \
   --mount type=bind,src="${PWD}",target=/workspace/axolotl \
   -v ${HOME}/.cache/huggingface:/root/.cache/huggingface \
-  axolotlai/axolotl:main-latest
+  axolotlai/axolotl-uv:main-latest
 ```
 :::
 
 ::: {.callout-important}
-For Blackwell GPUs, please use `axolotlai/axolotl:main-py3.11-cu128-2.9.1` or the cloud variant `axolotlai/axolotl-cloud:main-py3.11-cu128-2.9.1`.
+For Blackwell GPUs, please use `axolotlai/axolotl-uv:main-py3.11-cu128-2.9.1` or the cloud variant `axolotlai/axolotl-cloud-uv:main-py3.11-cu128-2.9.1`.
 :::
 
 Please refer to the [Docker documentation](docker.qmd) for more information on the different Docker images that are available.
@@ -122,7 +91,7 @@ Please refer to the [Docker documentation](docker.qmd) for more information on t
 
 For providers supporting Docker:
 
-- Use `axolotlai/axolotl-cloud:main-latest`
+- Use `axolotlai/axolotl-cloud-uv:main-latest`
 - Available on:
     - [RunPod](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
     - [Vast.ai](https://cloud.vast.ai?ref_id=62897&template_id=bdd4a49fa8bce926defc99471864cace&utm_source=axolotl&utm_medium=partner&utm_campaign=template_launch_july2025&utm_content=docs_link)
@@ -141,7 +110,7 @@ For providers supporting Docker:
 ### macOS {#sec-macos}
 
 ```{.bash}
-pip3 install --no-build-isolation -e '.'
+uv pip install --no-build-isolation -e '.'
 ```
 
 See @sec-troubleshooting for Mac-specific issues.
@@ -152,21 +121,44 @@ See @sec-troubleshooting for Mac-specific issues.
 We recommend using WSL2 (Windows Subsystem for Linux) or Docker.
 :::
 
-## Environment Managers {#sec-env-managers}
+## Migrating from pip to uv {#sec-migrating}
 
-### Conda/Pip venv {#sec-conda}
+If you have an existing pip-based Axolotl installation, you can migrate to uv:
 
-1. Install Python ≥3.11
-2. Install PyTorch: https://pytorch.org/get-started/locally/
-3. Install Axolotl:
-   ```{.bash}
-   pip3 install -U packaging setuptools wheel ninja
-   pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
-   ```
-4. (Optional) Login to Hugging Face:
-   ```{.bash}
-   hf auth login
-   ```
+```{.bash}
+# Install uv
+curl -LsSf https://astral.sh/uv/install.sh | sh
+source $HOME/.local/bin/env
+
+# Create a fresh venv (recommended for a clean start)
+export UV_TORCH_BACKEND=cu128  # or cu130
+uv venv --no-project --relocatable
+source .venv/bin/activate
+
+# Reinstall axolotl
+uv pip install --no-build-isolation axolotl[flash-attn,deepspeed]
+```
+
+## Using pip (Alternative) {#sec-pip}
+
+If you are unable to install uv, you can still use pip directly.
+
+::: {.callout-important}
+Please make sure to have PyTorch installed before installing Axolotl with pip.
+
+Follow the instructions at: [https://pytorch.org/get-started/locally/](https://pytorch.org/get-started/locally/)
+:::
+
+```{.bash}
+pip3 install -U packaging setuptools wheel ninja
+pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
+```
+
+For editable/development installs:
+```{.bash}
+pip3 install -U packaging setuptools wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
+```
 
 ## Troubleshooting {#sec-troubleshooting}
 

docs/unsloth.qmd

@@ -1,53 +0,0 @@
----
-title: "Unsloth"
-description: "Hyper-optimized QLoRA finetuning for single GPUs"
----
-
-### Overview
-
-Unsloth provides hand-written optimized kernels for LLM finetuning that slightly improve speed and VRAM over
-standard industry baselines.
-
-::: {.callout-important}
-Due to breaking changes in transformers `v4.48.0`, users will need to downgrade to `<=v4.47.1` to use this patch.
-
-This will later be deprecated in favor of [LoRA Optimizations](lora_optims.qmd).
-:::
-
-
-### Installation
-
-The following will install the correct unsloth and extras from source.
-
-```bash
-python scripts/unsloth_install.py | sh
-```
-
-### Usage
-
-Axolotl exposes a few configuration options to try out unsloth and get most of the performance gains.
-
-Our unsloth integration is currently limited to the following model architectures:
- - llama
-
-These options are specific to LoRA finetuning and cannot be used for multi-GPU finetuning
-```yaml
-unsloth_lora_mlp: true
-unsloth_lora_qkv: true
-unsloth_lora_o: true
-```
-
-These options are composable and can be used with multi-gpu finetuning
-```yaml
-unsloth_cross_entropy_loss: true
-unsloth_rms_norm: true
-unsloth_rope: true
-```
-
-### Limitations
-
-- Single GPU only; e.g. no multi-gpu support
-- No deepspeed or FSDP support (requires multi-gpu)
-- LoRA + QLoRA support only. No full fine tunes or fp8 support.
-- Limited model architecture support. Llama, Phi, Gemma, Mistral only
-- No MoE support.

examples/LiquidAI/README.md

@@ -15,8 +15,7 @@ Thanks to the team at LiquidAI for giving us early access to prepare for these r
     Here is an example of how to install from pip:
     ```bash
     # Ensure you have a compatible version of Pytorch installed
-    pip3 install packaging setuptools wheel ninja
-    pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+    uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
     ```
 
 2.  Run one of the finetuning examples below.
@@ -35,7 +34,7 @@ Thanks to the team at LiquidAI for giving us early access to prepare for these r
 
     **LFM2-MoE**
     ```bash
-    pip install git+https://github.com/huggingface/transformers.git@0c9a72e4576fe4c84077f066e585129c97bfd4e6
+    uv pip install git+https://github.com/huggingface/transformers.git@0c9a72e4576fe4c84077f066e585129c97bfd4e6
 
     # LoRA SFT (1x48GB @ 16.2GiB)
     axolotl train examples/LiquidAI/lfm2-8b-a1b-lora.yaml
@@ -45,7 +44,7 @@ Thanks to the team at LiquidAI for giving us early access to prepare for these r
 
 - **Installation Error**: If you encounter `ImportError: ... undefined symbol ...` or `ModuleNotFoundError: No module named 'causal_conv1d_cuda'`, the `causal-conv1d` package may have been installed incorrectly. Try uninstalling it:
   ```bash
-  pip uninstall -y causal-conv1d
+  uv pip uninstall causal-conv1d
   ```
 
 - **Dataset Loading**: Read more on how to load your own dataset in our [documentation](https://docs.axolotl.ai/docs/dataset_loading.html).

examples/apertus/README.md

@@ -15,8 +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==26.0 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn]'
+uv pip install --no-build-isolation -e '.[flash-attn]'
 
 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
 python scripts/cutcrossentropy_install.py | sh
@@ -31,7 +30,7 @@ python scripts/cutcrossentropy_install.py | sh
 # For those using our Docker image, use the below path.
 export CUDA_HOME=/usr/local/cuda
 
-pip3 install git+https://github.com/nickjbrowning/XIELU@59d6031 --no-build-isolation --no-deps
+uv pip install git+https://github.com/nickjbrowning/XIELU@59d6031 --no-build-isolation --no-deps
 ```
 
 For any installation errors, see [XIELU Installation Issues](#xielu-installation-issues)
@@ -67,7 +66,7 @@ If those didn't help, please try the below solutions:
 1. Pass env for CMAKE and try install again:
 
     ```bash
-    Python_EXECUTABLE=$(which python) pip3 install git+https://github.com/nickjbrowning/XIELU@59d6031 --no-build-isolation --no-deps
+    Python_EXECUTABLE=$(which python) uv pip install git+https://github.com/nickjbrowning/XIELU@59d6031 --no-build-isolation --no-deps
     ```
 
 2. Git clone the repo and manually hardcode python path:
@@ -92,7 +91,7 @@ If those didn't help, please try the below solutions:
     ```
 
     ```bash
-    pip3 install . --no-build-isolation --no-deps
+    uv pip install . --no-build-isolation --no-deps
     ```
 
 ## Optimization Guides

examples/arcee/README.md

@@ -17,8 +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==26.0 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn]'
+uv pip install --no-build-isolation -e '.[flash-attn]'
 
 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
 python scripts/cutcrossentropy_install.py | sh

examples/devstral/README.md

@@ -16,8 +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==26.0 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage

examples/gemma3n/README.md

@@ -10,17 +10,16 @@ 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==26.0 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. In addition to Axolotl's requirements, Gemma-3n requires:
 
 ```bash
-pip3 install timm==1.0.17
+uv pip install timm==1.0.17
 
 # for loading audio data
-pip3 install librosa==0.11.0
+uv pip install librosa==0.11.0
 ```
 
 3. Download sample dataset files

examples/gpt-oss/README.md

@@ -14,8 +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==26.0 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Choose one of the following configs below for training the 20B model. (for 120B, see [below](#training-120b))
@@ -87,7 +86,7 @@ for more information about using a special vllm-openai docker image for inferenc
 Optionally, vLLM can be installed from nightly:
 
 ```bash
-pip install --no-build-isolation --pre -U vllm --extra-index-url https://wheels.vllm.ai/nightly
+uv pip install --no-build-isolation --pre -U vllm --extra-index-url https://wheels.vllm.ai/nightly
 ```
 and the vLLM server can be started with the following command (modify `--tensor-parallel-size 8` to match your environment):
 ```bash

examples/granite4/README.md

@@ -15,8 +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==26.0 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn]'
+uv pip install --no-build-isolation -e '.[flash-attn]'
 
 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
 python scripts/cutcrossentropy_install.py | sh

examples/hunyuan/README.md

@@ -13,8 +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==26.0 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn]'
+uv pip install --no-build-isolation -e '.[flash-attn]'
 
 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
 python scripts/cutcrossentropy_install.py | sh

examples/internvl3_5/README.md

@@ -11,7 +11,7 @@ This guide shows how to fine-tune it with Axolotl.
 2. Install `timm` for vision model support:
 
     ```bash
-    pip install timm==1.0.19
+    uv pip install timm==1.0.19
     ```
 
 3. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.

examples/magistral/README.md

@@ -14,8 +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==26.0 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage

examples/magistral/vision/README.md

@@ -12,7 +12,7 @@ Before starting, ensure you have:
 
 1. Install the required vision lib:
     ```bash
-    pip install 'mistral-common[opencv]==1.8.5'
+    uv pip install 'mistral-common[opencv]==1.8.5'
     ```
 
 2. Download the example dataset image:

examples/ministral3/README.md

@@ -23,7 +23,7 @@ Note: This is still experimental given it is based on transformers v5 RC.
     git checkout transformers-v5
 
     # Install packages for transformers v5
-    pip install -e .
+    uv pip install -e .
     ```
 
 4. Run the fine-tuning:

examples/ministral3/vision/README.md

@@ -12,7 +12,7 @@ Before starting, ensure you have:
 
 1. Install the required vision lib:
     ```bash
-    pip install 'mistral-common[opencv]==1.8.6'
+    uv pip install 'mistral-common[opencv]==1.8.6'
     ```
 
 2. Download the example dataset image:

examples/mistral-small/README.md

@@ -12,7 +12,7 @@ Before starting, ensure you have:
 
 1. Install the required vision lib:
     ```bash
-    pip install 'mistral-common[opencv]==1.8.5'
+    uv pip install 'mistral-common[opencv]==1.8.5'
     ```
 
 2. Download the example dataset image:

examples/mistral4/README.md

@@ -13,7 +13,7 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 3. Install transformers from main
 
   ```bash
-  pip install git+https://github.com/huggingface/transformers.git
+  uv pip install git+https://github.com/huggingface/transformers.git
   ```
 
 4. Run one of the example configs:

examples/qwen3-next/README.md

@@ -12,7 +12,7 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 
 3. Install FLA for improved performance
 ```bash
-pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.4.1
+uv pip uninstall causal-conv1d && uv pip install flash-linear-attention==0.4.1
 ```
 
 4. Run the finetuning example:

examples/qwen3.5/README.md

@@ -10,7 +10,7 @@
 
 3. Install FLA for sample packing support with the Gated DeltaNet linear attention layers:
   ```bash
-  pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.4.1
+  uv pip uninstall causal-conv1d && uv pip install flash-linear-attention==0.4.1
   ```
   > FLA is required when `sample_packing: true`. Without it, training raises a `RuntimeError` on packed sequences. Vision configs use `sample_packing: false` so FLA is optional there.
 

examples/seed-oss/README.md

@@ -11,8 +11,7 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
     Here is an example of how to install from pip:
     ```bash
     # Ensure you have a compatible version of Pytorch installed
-    pip3 install packaging setuptools wheel ninja
-    pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+    uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 
     # Install Cut Cross Entropy
     python scripts/cutcrossentropy_install.py | sh

examples/smolvlm2/README.md

@@ -13,14 +13,13 @@ This guide shows how to fine-tune SmolVLM2 models with Axolotl.
     Here is an example of how to install from pip:
     ```bash
     # Ensure you have a compatible version of Pytorch installed
-    pip3 install packaging setuptools wheel ninja
-    pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+    uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
     ```
 
 2. Install an extra dependency:
 
     ```bash
-    pip3 install num2words==0.5.14
+    uv pip install num2words==0.5.14
     ```
 
 3.  Run the finetuning example:

examples/voxtral/README.md

@@ -12,16 +12,15 @@ 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==26.0 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Please install the below.
 
 ```bash
 # audio
-pip3 install librosa==0.11.0
-pip3 install 'mistral_common[audio]==1.8.3'
+uv pip install librosa==0.11.0
+uv pip install 'mistral_common[audio]==1.8.3'
 
 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
 python scripts/cutcrossentropy_install.py | sh

pyproject.toml

@@ -1,15 +1,165 @@
 [build-system]
-requires = ["setuptools>=64", "wheel", "setuptools_scm>=8", "packaging==26.0"]
+requires = ["setuptools>=64", "wheel", "setuptools_scm>=8"]
 build-backend = "setuptools.build_meta"
 
 [project]
 name = "axolotl"
-dynamic = ["version", "dependencies", "optional-dependencies"]
+dynamic = ["version"]
 description = "LLM Trainer"
 readme = "README.md"
 requires-python = ">=3.10"
 # license = "Apache-2.0"
 
+dependencies = [
+    # Core ML stack
+    "torch>=2.6.0",
+    "packaging==26.0",
+    "huggingface_hub>=1.1.7",
+    "peft>=0.19.1,<0.20.0",
+    "tokenizers>=0.22.1",
+    "transformers==5.5.4",
+    "accelerate==1.13.0",
+    "datasets>=4.8.4,<4.9.0",
+    "trl==1.1.0",
+    "hf_xet==1.4.3",
+    "kernels==0.13.0",
+    "trackio>=0.16.1",
+    "typing-extensions>=4.15.0",
+    "optimum==1.16.2",
+    "hf_transfer",
+    "sentencepiece",
+    "gradio>=6.2.0,<7.0",
+    "modal==1.3.0.post1",
+    "pydantic>=2.10.6",
+    "addict",
+    "fire",
+    "PyYAML>=6.0",
+    "requests",
+    "wandb",
+    "einops",
+    "colorama",
+    "numba>=0.61.2",
+    "numpy>=2.2.6",
+
+    # Evaluation & metrics
+    "evaluate==0.4.1",
+    "scipy",
+    "nvidia-ml-py==12.560.30",
+    "art",
+    "tensorboard",
+    "python-dotenv==1.0.1",
+
+    # Remote filesystems
+    "s3fs>=2024.5.0",
+    "gcsfs>=2025.3.0",
+    "adlfs>=2024.5.0",
+    "ocifs==1.3.2",
+
+    "zstandard==0.22.0",
+    "fastcore",
+
+    # lm eval harness
+    "lm_eval==0.4.11",
+    "langdetect==1.0.9",
+    "immutabledict==4.2.0",
+    "antlr4-python3-runtime==4.13.2",
+
+    "schedulefree==1.4.1",
+    "openenv-core==0.1.0",
+
+    # Axolotl contribs
+    "axolotl-contribs-lgpl==0.0.7",
+    "axolotl-contribs-mit==0.0.6",
+
+    # Telemetry
+    "posthog==6.7.11",
+
+    "mistral-common==1.11.0",
+
+    # Platform-specific (Linux only)
+    "bitsandbytes==0.49.1 ; sys_platform != 'darwin'",
+    "triton>=3.4.0 ; sys_platform != 'darwin'",
+    "xformers>=0.0.23.post1 ; sys_platform != 'darwin'",
+    "liger-kernel==0.7.0 ; sys_platform != 'darwin'",
+    "torchao==0.17.0 ; sys_platform != 'darwin' and platform_machine != 'aarch64'",
+
+    # Architecture-specific
+    "fla-core==0.4.1 ; platform_machine != 'aarch64'",
+    "flash-linear-attention==0.4.1 ; platform_machine != 'aarch64'",
+]
+
+[project.optional-dependencies]
+flash-attn = ["flash-attn==2.8.3"]
+ring-flash-attn = [
+    "flash-attn==2.8.3",
+    "ring-flash-attn>=0.1.7",
+]
+deepspeed = [
+    "deepspeed>=0.18.6,<0.19.0",
+    "deepspeed-kernels",
+]
+mamba-ssm = [
+    "mamba-ssm==1.2.0.post1",
+    "causal_conv1d",
+]
+auto-gptq = [
+    "auto-gptq==0.5.1",
+]
+mlflow = [
+    "mlflow",
+]
+galore = [
+    "galore_torch",
+]
+apollo = [
+    "apollo-torch",
+]
+optimizers = [
+    "galore_torch",
+    "apollo-torch",
+    "lomo-optim==0.1.1",
+    "torch-optimi==0.2.1",
+    "came_pytorch==0.1.3",
+]
+ray = [
+    "ray[train]>=2.52.1",
+]
+vllm = [
+    "vllm>=0.15.0",
+]
+llmcompressor = [
+    "llmcompressor>=0.10.0",
+]
+fbgemm-gpu = ["fbgemm-gpu-genai>=1.3.0"]
+opentelemetry = [
+    "opentelemetry-api",
+    "opentelemetry-sdk",
+    "opentelemetry-exporter-prometheus",
+    "prometheus-client",
+]
+
+[dependency-groups]
+dev = [
+    "black",
+    "mypy",
+    "pre-commit",
+    "types-requests",
+    "quartodoc",
+    "jupyter",
+    "blobfile",
+    "tiktoken",
+]
+test = [
+    "codecov",
+    "codecov-cli",
+    "pytest",
+    "pytest-cov",
+    "pytest-retry",
+    "pytest-sugar",
+    "pytest-xdist",
+    "tbparse",
+]
+
 [project.scripts]
 axolotl = "axolotl.cli.main:main"
 
@@ -18,18 +168,15 @@ Homepage = "https://axolotl.ai/"
 Documentation = "https://docs.axolotl.ai/"
 Repository = "https://github.com/axolotl-ai-cloud/axolotl.git"
 
-[tool.setuptools_scm]
-
 [tool.setuptools]
-py-modules = ["setuptools_axolotl_dynamic_dependencies"]
 include-package-data = true
 
+[tool.setuptools.packages.find]
+where = ["src"]
+
 [tool.setuptools.dynamic]
 version = { file = "VERSION" }
 
-[tool.setuptools.cmdclass]
-build_py = "setuptools_axolotl_dynamic_dependencies.BuildPyCommand"
-
 [tool.ruff]
 line-length = 88
 target-version = "py310"
@@ -67,5 +214,43 @@ markers = [
     "slow: marks tests as slow",
 ]
 
+# UV specific configuration
+[tool.uv]
+prerelease = "allow"
+conflicts = [
+    [
+        { package = "axolotl" },
+        { extra = "vllm" },
+    ],
+    [
+        { package = "axolotl" },
+        { extra = "flash-attn" },
+    ],
+    [
+        { package = "axolotl" },
+        { extra = "ring-flash-attn" },
+    ],
+    [
+        { package = "axolotl" },
+        { extra = "mamba-ssm" },
+    ],
+    [
+        { package = "axolotl" },
+        { extra = "auto-gptq" },
+    ],
+    [
+        { package = "axolotl" },
+        { extra = "fbgemm-gpu" },
+    ],
+    [
+        { package = "axolotl" },
+        { extra = "llmcompressor" },
+    ],
+]
+
 [tool.uv.extra-build-dependencies]
-axolotl = ["huggingface_hub"]
+mamba-ssm = [{ requirement = "torch", match-runtime = true }]
+causal-conv1d = [{ requirement = "torch", match-runtime = true }]
+flash-attn = [{ requirement = "torch", match-runtime = true }]
+deepspeed = [{ requirement = "torch", match-runtime = true }]
+auto-gptq = [{ requirement = "torch", match-runtime = true }]

requirements-dev.txt

@@ -1,8 +0,0 @@
-black
-mypy
-pre-commit
-types-requests
-quartodoc
-jupyter
-blobfile
-tiktoken

requirements-tests.txt

@@ -1,8 +0,0 @@
-codecov
-codecov-cli
-pytest
-pytest-cov
-pytest-retry
-pytest-sugar
-pytest-xdist
-tbparse

requirements.txt

@@ -1,78 +0,0 @@
---extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
-
-# START section of dependencies that don't install on Darwin/MacOS
-bitsandbytes==0.49.1
-triton>=3.4.0
-mamba-ssm==1.2.0.post1
-xformers>=0.0.23.post1
-liger-kernel==0.7.0
-# END section
-
-packaging==26.0
-huggingface_hub>=1.1.7
-peft>=0.19.0,<0.20.0
-tokenizers>=0.22.1
-transformers==5.5.4
-accelerate==1.13.0
-datasets>=4.8.4,<4.9.0
-deepspeed>=0.18.6,<0.19.0
-trl==1.1.0
-hf_xet==1.4.3
-kernels==0.13.0
-
-fla-core==0.4.1
-flash-linear-attention==0.4.1
-
-trackio>=0.16.1
-typing-extensions>=4.15.0
-
-optimum==1.16.2
-hf_transfer
-sentencepiece
-gradio>=6.2.0,<7.0
-
-modal==1.3.0.post1
-pydantic>=2.10.6
-addict
-fire
-PyYAML>=6.0
-requests
-wandb
-einops
-colorama
-numba>=0.61.2
-numpy>=2.2.6
-
-# qlora things
-evaluate==0.4.1
-scipy
-nvidia-ml-py==12.560.30
-art
-tensorboard
-python-dotenv==1.0.1
-
-# remote filesystems
-s3fs>=2024.5.0
-gcsfs>=2025.3.0
-adlfs>=2024.5.0
-ocifs==1.3.2
-
-zstandard==0.22.0
-fastcore
-
-# lm eval harness
-lm_eval==0.4.11
-langdetect==1.0.9
-immutabledict==4.2.0
-antlr4-python3-runtime==4.13.2
-
-torchao==0.17.0
-openenv-core==0.1.0
-schedulefree==1.4.1
-
-axolotl-contribs-lgpl==0.0.7
-axolotl-contribs-mit==0.0.6
-# telemetry
-posthog==6.7.11
-
-mistral-common==1.11.0

scripts/build_scattermoe_lora_kernel.py

@@ -0,0 +1,479 @@
+#!/usr/bin/env python3
+"""Build a disposable Hugging Face Kernel Hub package for ScatterMoE LoRA.
+
+This script does not move or edit the in-tree Axolotl kernel sources. It copies
+``src/axolotl/integrations/kernels/libs/scattermoe_lora`` into an ignored
+build directory and emits a universal HF kernels project that can be pushed to
+the Hub.
+"""
+
+from __future__ import annotations
+
+import argparse
+import fnmatch
+import hashlib
+import json
+import os
+import shutil
+import subprocess
+import sys
+from importlib import metadata
+from pathlib import Path
+
+PACKAGE_NAME = "scattermoe_lora"
+BUILD_VARIANT = "torch-universal"
+DEFAULT_REPO_ID = "kernels-community/scattermoe-lora"
+HF_REPO_TYPE = "kernel"
+HF_KERNEL_URL_PREFIX = "https://hf.co/kernels"
+
+REPO_ROOT = Path(__file__).resolve().parents[1]
+DEFAULT_SOURCE_DIR = (
+    REPO_ROOT / "src" / "axolotl" / "integrations" / "kernels" / "libs" / PACKAGE_NAME
+)
+DEFAULT_OUTPUT_DIR = REPO_ROOT / "build" / "hf-kernels" / PACKAGE_NAME
+
+EXCLUDED_DIRS = {
+    "__pycache__",
+    ".mypy_cache",
+    ".pytest_cache",
+    ".ruff_cache",
+}
+EXCLUDED_FILE_PATTERNS = {
+    "*.pyc",
+    "*.pyo",
+    "*.so",
+    ".DS_Store",
+}
+
+TEXT_REPLACEMENTS = {
+    "from axolotl.integrations.kernels.libs.scattermoe_lora.selective_dequant import": (
+        "from .selective_dequant import"
+    ),
+    "from axolotl.integrations.kernels.libs.scattermoe_lora.selective_dequant_kernel import": (
+        "from .selective_dequant_kernel import"
+    ),
+    "from axolotl.integrations.kernels.libs.scattermoe_lora.kernels.ops import": (
+        "from .ops import"
+    ),
+}
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description=(
+            "Copy Axolotl's ScatterMoE LoRA Triton kernels into a disposable "
+            "HF Kernel Hub universal package."
+        )
+    )
+    parser.add_argument(
+        "--source-dir",
+        type=Path,
+        default=DEFAULT_SOURCE_DIR,
+        help=f"ScatterMoE LoRA source package to copy. Default: {DEFAULT_SOURCE_DIR}",
+    )
+    parser.add_argument(
+        "--output-dir",
+        type=Path,
+        default=DEFAULT_OUTPUT_DIR,
+        help=f"Destination build/dist directory. Default: {DEFAULT_OUTPUT_DIR}",
+    )
+    parser.add_argument(
+        "--repo-id",
+        default=DEFAULT_REPO_ID,
+        help=f"HF Hub repo id to write into build.toml. Default: {DEFAULT_REPO_ID}",
+    )
+    parser.add_argument(
+        "--version",
+        type=int,
+        default=1,
+        help="Kernel major version written to build.toml and metadata.json.",
+    )
+    parser.add_argument(
+        "--force",
+        action="store_true",
+        help="Delete the output directory first if it already exists.",
+    )
+    parser.add_argument(
+        "--no-source-layout",
+        action="store_true",
+        help="Only write the shippable build/ tree, not torch-ext/ sources.",
+    )
+    parser.add_argument(
+        "--upload",
+        action="store_true",
+        help=(
+            "Upload the generated universal kernel package with huggingface_hub. "
+            "This bypasses kernel-builder and is intended for pure Python/Triton "
+            "universal kernels."
+        ),
+    )
+    parser.add_argument(
+        "--private",
+        action="store_true",
+        help="Create the HF Hub repo as private when used with --upload.",
+    )
+    parser.add_argument(
+        "--skip-version-branch",
+        action="store_true",
+        help="With --upload, only upload main and skip the v<version> branch.",
+    )
+    return parser.parse_args()
+
+
+def should_skip_file(path: Path) -> bool:
+    return any(
+        fnmatch.fnmatch(path.name, pattern) for pattern in EXCLUDED_FILE_PATTERNS
+    )
+
+
+def iter_source_files(source_dir: Path) -> list[Path]:
+    files: list[Path] = []
+    for root, dirs, filenames in os.walk(source_dir):
+        dirs[:] = sorted(d for d in dirs if d not in EXCLUDED_DIRS)
+        for filename in sorted(filenames):
+            path = Path(root) / filename
+            if not should_skip_file(path):
+                files.append(path)
+    return files
+
+
+def content_hash(source_dir: Path) -> str:
+    digest = hashlib.sha1()
+    for path in iter_source_files(source_dir):
+        rel = path.relative_to(source_dir).as_posix()
+        digest.update(rel.encode("utf-8"))
+        digest.update(b"\0")
+        digest.update(path.read_bytes())
+        digest.update(b"\0")
+    return digest.hexdigest()[:10]
+
+
+def git_revision() -> str:
+    try:
+        result = subprocess.run(
+            ["git", "rev-parse", "--short", "HEAD"],
+            cwd=REPO_ROOT,
+            check=True,
+            capture_output=True,
+            text=True,
+        )
+    except (OSError, subprocess.CalledProcessError):
+        return "unknown"
+    return result.stdout.strip() or "unknown"
+
+
+def transform_python_source(text: str, rel_path: Path, op_namespace: str) -> str:
+    for old, new in TEXT_REPLACEMENTS.items():
+        text = text.replace(old, new)
+
+    if rel_path.as_posix() == "gemma4_experts.py":
+        text = text.replace(
+            "    from axolotl.integrations.kernels.constants import resolve_experts_class",
+            (
+                "    raise RuntimeError(\n"
+                '        "patch_gemma4_scattermoe is only available from the in-tree Axolotl "\n'
+                '        "integration. Use register_scattermoe_experts() with the standalone "\n'
+                '        "HF kernel package."\n'
+                "    )"
+            ),
+        )
+
+    return text.replace("scattermoe::", f"{op_namespace}::")
+
+
+def copy_package(source_dir: Path, package_dir: Path, op_namespace: str) -> None:
+    for source in iter_source_files(source_dir):
+        rel_path = source.relative_to(source_dir)
+        destination = package_dir / rel_path
+        destination.parent.mkdir(parents=True, exist_ok=True)
+
+        if source.suffix == ".py":
+            text = source.read_text(encoding="utf-8")
+            text = transform_python_source(text, rel_path, op_namespace)
+            destination.write_text(text, encoding="utf-8")
+        else:
+            shutil.copy2(source, destination)
+
+    write_ops_module(package_dir / "_ops.py", op_namespace)
+
+
+def write_ops_module(path: Path, op_namespace: str) -> None:
+    path.write_text(
+        "\n".join(
+            [
+                "import torch",
+                "",
+                f"ops = torch.ops.{op_namespace}",
+                "",
+                "",
+                "def add_op_namespace_prefix(op_name: str) -> str:",
+                f'    return f"{op_namespace}::{{op_name}}"',
+                "",
+            ]
+        ),
+        encoding="utf-8",
+    )
+
+
+def write_build_toml(path: Path, repo_id: str, version: int) -> None:
+    lines = [
+        "[general]",
+        f'name = "{PACKAGE_NAME}"',
+        "universal = true",
+        f"version = {version}",
+        "",
+    ]
+    if repo_id:
+        lines.extend(
+            [
+                "[general.hub]",
+                f'repo-id = "{repo_id}"',
+                "",
+            ]
+        )
+    path.write_text("\n".join(lines), encoding="utf-8")
+
+
+def write_flake(path: Path) -> None:
+    path.write_text(
+        """{
+  description = "Flake for scattermoe_lora kernel";
+
+  inputs = {
+    builder.url = "github:huggingface/kernels";
+  };
+
+  outputs =
+    {
+      self,
+      builder,
+    }:
+    builder.lib.genKernelFlakeOutputs {
+      inherit self;
+      path = ./.;
+    };
+}
+""",
+        encoding="utf-8",
+    )
+
+
+def write_readme(path: Path, repo_id: str, source_hash: str, op_namespace: str) -> None:
+    repo_display = repo_id or "<your-org>/scattermoe-lora"
+    path.write_text(
+        f"""---
+library_name: kernels
+license: apache-2.0
+tags:
+- kernel
+- kernels
+---
+
+# ScatterMoE LoRA
+
+Standalone Hugging Face Kernel Hub package for Axolotl's ScatterMoE LoRA Triton kernels.
+
+This package is generated from Axolotl's in-tree `scattermoe_lora` sources and is exported as a universal kernel because the implementation is Python/Triton rather than a precompiled C++/CUDA extension.
+
+```python
+from kernels import get_kernel
+
+scattermoe_lora = get_kernel("{repo_display}")
+```
+
+Export metadata:
+
+- source package: `src/axolotl/integrations/kernels/libs/scattermoe_lora`
+- source revision: `{git_revision()}`
+- source content hash: `{source_hash}`
+- torch custom op namespace: `{op_namespace}`
+
+The generated `build/torch-universal/{PACKAGE_NAME}` directory is the shippable Hub artifact. `torch-ext/{PACKAGE_NAME}` is included so `kernel-builder build-and-copy` can regenerate the universal build tree if desired.
+""",
+        encoding="utf-8",
+    )
+
+
+def write_metadata(path: Path, version: int) -> None:
+    path.write_text(
+        json.dumps({"version": version}, indent=2, sort_keys=True) + "\n",
+        encoding="utf-8",
+    )
+
+
+def prepare_output_dir(output_dir: Path, force: bool) -> None:
+    if output_dir.exists():
+        if not force:
+            raise FileExistsError(
+                f"{output_dir} already exists. Re-run with --force to replace it."
+            )
+        shutil.rmtree(output_dir)
+    output_dir.mkdir(parents=True)
+
+
+def build_package(args: argparse.Namespace) -> Path:
+    source_dir = args.source_dir.resolve()
+    output_dir = args.output_dir.resolve()
+
+    if not source_dir.is_dir():
+        raise FileNotFoundError(f"source package does not exist: {source_dir}")
+    if not (source_dir / "__init__.py").is_file():
+        raise FileNotFoundError(f"source package is missing __init__.py: {source_dir}")
+
+    source_hash = content_hash(source_dir)
+    op_namespace = f"_{PACKAGE_NAME}_{source_hash}"
+
+    prepare_output_dir(output_dir, args.force)
+
+    write_build_toml(output_dir / "build.toml", args.repo_id, args.version)
+    write_flake(output_dir / "flake.nix")
+    write_readme(output_dir / "README.md", args.repo_id, source_hash, op_namespace)
+
+    if not args.no_source_layout:
+        copy_package(source_dir, output_dir / "torch-ext" / PACKAGE_NAME, op_namespace)
+
+    build_package_dir = output_dir / "build" / BUILD_VARIANT / PACKAGE_NAME
+    copy_package(source_dir, build_package_dir, op_namespace)
+    write_metadata(build_package_dir.parent / "metadata.json", args.version)
+
+    return output_dir
+
+
+def upload_package(args: argparse.Namespace, output_dir: Path) -> None:
+    if not args.repo_id:
+        raise ValueError("--repo-id is required when using --upload")
+
+    try:
+        from huggingface_hub import HfApi, constants as hf_constants
+    except ImportError as exc:
+        raise RuntimeError(
+            "--upload requires huggingface_hub. Install it or run the upload "
+            "manually with the Hugging Face CLI."
+        ) from exc
+
+    try:
+        hub_version = metadata.version("huggingface_hub")
+    except metadata.PackageNotFoundError:
+        hub_version = "unknown"
+
+    accepted_repo_types = getattr(
+        hf_constants,
+        "REPO_TYPES_WITH_KERNEL",
+        getattr(hf_constants, "REPO_TYPES", ()),
+    )
+    if HF_REPO_TYPE not in accepted_repo_types:
+        raise RuntimeError(
+            "Your huggingface_hub installation does not support "
+            f"repo_type={HF_REPO_TYPE!r} (found huggingface_hub {hub_version}). "
+            f"Upgrade this interpreter with: {sys.executable} -m pip install --upgrade "
+            "'huggingface_hub>=1.10.0'"
+        )
+
+    # huggingface_hub 1.11.0 has partial kernel support: create_repo accepts
+    # "kernel", but upload_folder/create_commit still validate against the
+    # older REPO_TYPES list. Extend it in-process so those helpers use the
+    # /api/kernels/... endpoints until upstream broadens that check.
+    if HF_REPO_TYPE not in hf_constants.REPO_TYPES:
+        hf_constants.REPO_TYPES.append(HF_REPO_TYPE)
+
+    api = HfApi()
+    try:
+        repo_id = api.create_repo(
+            repo_id=args.repo_id,
+            repo_type=HF_REPO_TYPE,
+            private=args.private,
+            exist_ok=True,
+        ).repo_id
+    except ValueError as exc:
+        if "Invalid repo type" in str(exc):
+            raise RuntimeError(
+                "huggingface_hub rejected repo_type='kernel'. "
+                f"This usually means the command is running with an older Hub "
+                f"client than expected (found huggingface_hub {hub_version} at "
+                f"{sys.executable}). Upgrade with: {sys.executable} -m pip "
+                "install --upgrade 'huggingface_hub>=1.10.0'"
+            ) from exc
+        raise
+
+    delete_patterns = [
+        "build/**",
+        "torch-ext/**",
+        "build.toml",
+        "flake.nix",
+        "README.md",
+    ]
+
+    api.upload_folder(
+        repo_id=repo_id,
+        repo_type=HF_REPO_TYPE,
+        folder_path=output_dir,
+        revision="main",
+        delete_patterns=delete_patterns,
+        commit_message="Upload ScatterMoE LoRA universal kernel",
+    )
+    print(f"Uploaded main branch: {HF_KERNEL_URL_PREFIX}/{repo_id}")
+
+    if args.skip_version_branch:
+        return
+
+    version_branch = f"v{args.version}"
+    api.create_branch(
+        repo_id=repo_id,
+        repo_type=HF_REPO_TYPE,
+        branch=version_branch,
+        revision="main",
+        exist_ok=True,
+    )
+    api.upload_folder(
+        repo_id=repo_id,
+        repo_type=HF_REPO_TYPE,
+        folder_path=output_dir,
+        revision=version_branch,
+        delete_patterns=delete_patterns,
+        commit_message=f"Upload ScatterMoE LoRA universal kernel {version_branch}",
+    )
+    print(
+        f"Uploaded version branch: "
+        f"{HF_KERNEL_URL_PREFIX}/{repo_id}/tree/{version_branch}"
+    )
+
+
+def main() -> int:
+    args = parse_args()
+    try:
+        output_dir = build_package(args)
+        if args.upload:
+            upload_package(args, output_dir)
+    except Exception as exc:
+        print(f"error: {exc}", file=sys.stderr)
+        return 1
+
+    print(f"Wrote ScatterMoE LoRA HF kernel package to: {output_dir}")
+    print(f"Shippable artifact: {output_dir / 'build' / BUILD_VARIANT / PACKAGE_NAME}")
+    if args.upload:
+        print(f'Load it with: get_kernel("{args.repo_id}", version={args.version})')
+        print(f"Uploaded as Hugging Face repo_type={HF_REPO_TYPE!r}.")
+        return 0
+
+    print("Next step:")
+    print("  upload this universal Python/Triton kernel directly:")
+    print(
+        f"    python3 {Path(__file__).as_posix()} "
+        f"--repo-id {args.repo_id} --force --upload"
+    )
+    if shutil.which("kernel-builder") is None:
+        print("  optional: install kernel-builder for full Nix-based builds:")
+        print(
+            "    curl -fsSL "
+            "https://raw.githubusercontent.com/huggingface/kernels/main/install.sh "
+            "| bash"
+        )
+    else:
+        print("  optional: upload with kernel-builder:")
+        print(f"    cd {output_dir}")
+        print("    kernel-builder build-and-upload")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/unsloth_install.py

@@ -1,40 +0,0 @@
-# noqa
-import sys
-
-try:
-    import torch
-except ImportError as error:
-    raise ImportError("Install torch via `pip install torch`") from error
-from packaging.version import Version as V
-
-use_uv = "--uv" in sys.argv[1:]
-
-v = V(torch.__version__)
-cuda = str(torch.version.cuda)
-try:
-    is_ampere = torch.cuda.get_device_capability()[0] >= 8
-except RuntimeError:
-    is_ampere = False
-if cuda != "12.1" and cuda != "11.8" and cuda != "12.4":
-    raise RuntimeError(f"CUDA = {cuda} not supported!")
-if v <= V("2.1.0"):
-    raise RuntimeError(f"Torch = {v} too old!")
-elif v <= V("2.1.1"):
-    x = "cu{}{}-torch211"
-elif v <= V("2.1.2"):
-    x = "cu{}{}-torch212"
-elif v < V("2.3.0"):
-    x = "cu{}{}-torch220"
-elif v < V("2.4.0"):
-    x = "cu{}{}-torch230"
-elif v < V("2.5.0"):
-    x = "cu{}{}-torch240"
-elif v < V("2.6.0"):
-    x = "cu{}{}-torch250"
-else:
-    raise RuntimeError(f"Torch = {v} too new!")
-x = x.format(cuda.replace(".", ""), "-ampere" if is_ampere else "")
-uv_prefix = "uv " if use_uv else ""
-print(
-    f'{uv_prefix}pip install unsloth-zoo==2024.12.1 && {uv_prefix}pip install --no-deps "unsloth[{x}]==2024.12.4"'
-)

setup.py

@@ -1,230 +0,0 @@
-"""setup.py for axolotl"""
-
-import os
-import platform
-import re
-from importlib.metadata import PackageNotFoundError, version
-from pathlib import Path
-
-from setuptools import find_packages, setup
-
-
-def parse_requirements(extras_require_map):
-    _install_requires = []
-    _dependency_links = []
-    with open("./requirements.txt", encoding="utf-8") as requirements_file:
-        lines = [r.strip() for r in requirements_file.readlines()]
-        for line in lines:
-            is_extras = "deepspeed" in line or "mamba-ssm" in line
-            if line.startswith("--extra-index-url"):
-                # Handle custom index URLs
-                _, url = line.split()
-                _dependency_links.append(url)
-            elif not is_extras and line and line[0] != "#":
-                # Handle standard packages
-                _install_requires.append(line)
-    try:
-        xformers_version = [req for req in _install_requires if "xformers" in req][0]
-        install_xformers = platform.machine() != "aarch64"
-        if platform.machine() == "aarch64":
-            # skip on ARM64
-            skip_packages = [
-                "torchao",
-                "fla-core",
-                "flash-linear-attention",
-            ]
-            _install_requires = [
-                req
-                for req in _install_requires
-                if re.split(r"[>=<]", req)[0].strip() not in skip_packages
-            ]
-        if "Darwin" in platform.system():
-            # skip packages not compatible with OSX
-            skip_packages = [
-                "bitsandbytes",
-                "triton",
-                "mamba-ssm",
-                "xformers",
-                "liger-kernel",
-            ]
-            _install_requires = [
-                req
-                for req in _install_requires
-                if re.split(r"[>=<]", req)[0].strip() not in skip_packages
-            ]
-            print(
-                _install_requires, [req in skip_packages for req in _install_requires]
-            )
-        else:
-            # detect the version of torch already installed
-            # and set it so dependencies don't clobber the torch version
-            try:
-                torch_version = version("torch")
-            except PackageNotFoundError:
-                torch_version = "2.8.0"  # default to torch 2.8.0
-            _install_requires.append(f"torch=={torch_version}")
-
-            version_match = re.match(r"^(\d+)\.(\d+)(?:\.(\d+))?", torch_version)
-            if version_match:
-                major, minor, patch = version_match.groups()
-                major, minor = int(major), int(minor)
-                patch = (
-                    int(patch) if patch is not None else 0
-                )  # Default patch to 0 if not present
-            else:
-                raise ValueError("Invalid version format")
-
-            torch_parts = torch_version.split("+")
-            if len(torch_parts) == 2:
-                torch_cuda_version = torch_parts[1]
-                _dependency_links.append(
-                    f"https://download.pytorch.org/whl/{torch_cuda_version}"
-                )
-
-            if (major, minor) >= (2, 10):
-                extras_require_map.pop("fbgemm-gpu")
-                extras_require_map["fbgemm-gpu"] = [
-                    "fbgemm-gpu==1.5.0",
-                    "fbgemm-gpu-genai==1.5.0",
-                ]
-                if not install_xformers:
-                    _install_requires.pop(_install_requires.index(xformers_version))
-                extras_require_map["vllm"] = ["vllm>=0.19.0"]
-            elif (major, minor) >= (2, 9):
-                extras_require_map.pop("fbgemm-gpu")
-                extras_require_map["fbgemm-gpu"] = [
-                    "fbgemm-gpu==1.4.0",
-                    "fbgemm-gpu-genai==1.4.2",
-                ]
-                if not install_xformers:
-                    _install_requires.pop(_install_requires.index(xformers_version))
-                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"]
-                extras_require_map["vllm"] = ["vllm==0.11.0"]
-                if not install_xformers:
-                    _install_requires.pop(_install_requires.index(xformers_version))
-            elif (major, minor) >= (2, 7):
-                _install_requires.pop(_install_requires.index(xformers_version))
-                if patch == 0:
-                    if install_xformers:
-                        _install_requires.append("xformers==0.0.30")
-                    # vllm 0.9.x is incompatible with latest transformers
-                    extras_require_map.pop("vllm")
-                else:
-                    if install_xformers:
-                        _install_requires.append("xformers==0.0.31")
-                    extras_require_map["vllm"] = ["vllm==0.10.1"]
-            elif (major, minor) >= (2, 6):
-                _install_requires.pop(_install_requires.index(xformers_version))
-                if install_xformers:
-                    _install_requires.append("xformers==0.0.29.post3")
-                # since we only support 2.6.0+cu126
-                _dependency_links.append("https://download.pytorch.org/whl/cu126")
-                extras_require_map.pop("vllm")
-            elif (major, minor) >= (2, 5):
-                _install_requires.pop(_install_requires.index(xformers_version))
-                if install_xformers:
-                    if patch == 0:
-                        _install_requires.append("xformers==0.0.28.post2")
-                    else:
-                        _install_requires.append("xformers>=0.0.28.post3")
-                extras_require_map.pop("vllm")
-            elif (major, minor) >= (2, 4):
-                extras_require_map.pop("vllm")
-                if install_xformers:
-                    if patch == 0:
-                        _install_requires.pop(_install_requires.index(xformers_version))
-                        _install_requires.append("xformers>=0.0.27")
-                    else:
-                        _install_requires.pop(_install_requires.index(xformers_version))
-                        _install_requires.append("xformers==0.0.28.post1")
-            else:
-                raise ValueError("axolotl requires torch>=2.4")
-
-    except PackageNotFoundError:
-        pass
-    return _install_requires, _dependency_links, extras_require_map
-
-
-def get_package_version():
-    with open(
-        Path(os.path.dirname(os.path.abspath(__file__))) / "VERSION",
-        "r",
-        encoding="utf-8",
-    ) as fin:
-        version_ = fin.read().strip()
-    return version_
-
-
-extras_require = {
-    "flash-attn": ["flash-attn==2.8.3"],
-    "ring-flash-attn": [
-        "flash-attn==2.8.3",
-        "ring-flash-attn>=0.1.7",
-    ],
-    "deepspeed": [
-        "deepspeed==0.18.2",
-        "deepspeed-kernels",
-    ],
-    "mamba-ssm": [
-        "mamba-ssm==1.2.0.post1",
-        "causal_conv1d",
-    ],
-    "auto-gptq": [
-        "auto-gptq==0.5.1",
-    ],
-    "mlflow": [
-        "mlflow",
-    ],
-    "galore": [
-        "galore_torch",
-    ],
-    "apollo": [
-        "apollo-torch",
-    ],
-    "optimizers": [
-        "galore_torch",
-        "apollo-torch",
-        "lomo-optim==0.1.1",
-        "torch-optimi==0.2.1",
-        "came_pytorch==0.1.3",
-    ],
-    "ray": [
-        "ray[train]>=2.52.1",
-    ],
-    "vllm": [
-        "vllm==0.10.0",
-    ],
-    "llmcompressor": [
-        "llmcompressor==0.5.1",
-    ],
-    "fbgemm-gpu": ["fbgemm-gpu-genai==1.3.0"],
-    "opentelemetry": [
-        "opentelemetry-api",
-        "opentelemetry-sdk",
-        "opentelemetry-exporter-prometheus",
-        "prometheus-client",
-    ],
-}
-install_requires, dependency_links, extras_require_build = parse_requirements(
-    extras_require
-)
-
-setup(
-    version=get_package_version(),
-    package_dir={"": "src"},
-    packages=find_packages("src"),
-    install_requires=install_requires,
-    dependency_links=dependency_links,
-    entry_points={
-        "console_scripts": [
-            "axolotl=axolotl.cli.main:main",
-        ],
-    },
-    extras_require=extras_require_build,
-)

src/axolotl/cli/utils/lora_merge.py

@@ -339,7 +339,11 @@ def _build_peft_layer_and_get_delta(
             )
         layer.lora_A[adapter_name].weight.data = lora_a
         layer.lora_B[adapter_name].weight.data = lora_b
-        return layer.get_delta_weight(adapter_name)
+        delta = layer.get_delta_weight(adapter_name)
+        # peft >=0.19.1 may return delta with transposed dims for 3D params
+        if delta.shape != base_tensor.shape and delta.ndim == 3:
+            delta = delta.transpose(1, 2).contiguous()
+        return delta
     elif (
         layer_type and "Conv" in layer_type or (layer_type is None and lora_a.ndim > 2)
     ):

src/axolotl/integrations/kernels/libs/scattermoe_lora/layers.py

@@ -60,49 +60,14 @@ def peft_lora_B_to_scattermoe(peft_B, num_experts, rank):
 
 
 def peft_lora_to_scattermoe(peft_A, peft_B, num_experts, rank):
-    """Convert peft LoRA weights to scattermoe layout (with A<->B swap).
+    """Convert peft LoRA weights to scattermoe layout.
 
-    peft operates on the parameter in its native storage layout ``[E, dim1, dim2]``
-    where ``in_features=dim1, out_features=dim2``.  ScatterMoE transposes the
-    parameter (``W = param.transpose(2, 1)``) giving ``[E, dim2, dim1]`` with
-    ``K=dim2, N=dim1``.  Because of this transposition, peft's A and B roles
-    are swapped relative to scattermoe's convention.
-
-    peft gives:
-        lora_A ``[r*E, dim1]``, lora_B ``[dim2, r*E]``
-
-    scattermoe needs:
-        lora_A ``[r*E, K=dim2]``, lora_B ``[N=dim1, r*E]``
-
-    This function swaps A<->B and converts B from rank-major to expert-major.
-    Uses vectorized tensor operations (no Python loop over experts).
-
-    Works for **both** gate_up_proj and down_proj since the transposition
-    issue is the same for any parameter.
+    peft >=0.19.1 assigns in/out features for 3D params such that
+    A and B already align with scattermoe's convention (no A<->B swap).
+    Only B needs rank-major → expert-major layout conversion.
     """
-    peft_B_em = peft_lora_B_to_scattermoe(peft_B, num_experts, rank)
-
-    dim1 = peft_A.shape[1]  # peft in_features -> scattermoe N
-    dim2 = peft_B_em.shape[0]  # peft out_features -> scattermoe K
-
-    # smoe_A: per expert, transpose B_e [dim2, r] -> [r, dim2]
-    # [dim2, E*r] -> [dim2, E, r] -> [E, r, dim2] -> [E*r, dim2]
-    smoe_A = (
-        peft_B_em.reshape(dim2, num_experts, rank)
-        .permute(1, 2, 0)
-        .contiguous()
-        .reshape(rank * num_experts, dim2)
-    )
-
-    # smoe_B: per expert, transpose A_e [r, dim1] -> [dim1, r]
-    # [E*r, dim1] -> [E, r, dim1] -> [dim1, E, r] -> [dim1, E*r]
-    smoe_B = (
-        peft_A.reshape(num_experts, rank, dim1)
-        .permute(2, 0, 1)
-        .contiguous()
-        .reshape(dim1, num_experts * rank)
-    )
-
+    smoe_A = peft_A
+    smoe_B = peft_lora_B_to_scattermoe(peft_B, num_experts, rank)
     return smoe_A, smoe_B
 
 

src/axolotl/loaders/patch_manager.py

@@ -162,7 +162,6 @@ class PatchManager:
     def apply_post_model_load_patches(self, model: PreTrainedModel):
         """Apply patches that require the model instance."""
         self._apply_llama_flash_attn_patches(model)
-        self._apply_unsloth_patches(model)
         self._apply_lora_kernel_patch(model)
         self._apply_scaling_softmax_patch(model)
 
@@ -396,7 +395,16 @@ class PatchManager:
                     patch_gemma4_fused_attn,
                 )
 
-                patch_gemma4_fused_attn()
+                # Shared-KV side channel when activation checkpointing (PR #3611).
+                fsdp_cfg = self.cfg.fsdp_config
+                needs_shared_kv_workaround = (not self.inference) and bool(
+                    self.cfg.gradient_checkpointing
+                    or self.cfg.activation_offloading
+                    or (fsdp_cfg is not None and fsdp_cfg.activation_checkpointing)
+                )
+                patch_gemma4_fused_attn(
+                    install_shared_kv_workaround=needs_shared_kv_workaround
+                )
 
     @staticmethod
     def _fix_nemotron_h_conversion_mapping():
@@ -674,24 +682,10 @@ class PatchManager:
             )
 
             patch_fa_llama_cross_entropy()
-        elif self.cfg.unsloth_cross_entropy_loss:
-            from axolotl.monkeypatch.unsloth_ import integrate_cross_entropy_loss_patch
-
-            integrate_cross_entropy_loss_patch(model_type="llama")
-
         if self.cfg.flash_attn_rms_norm and self.has_flash_attn:
             from axolotl.monkeypatch.llama_attn_hijack_flash import patch_llama_rms_norm
 
             patch_llama_rms_norm()
-        elif self.cfg.unsloth_rms_norm:
-            from axolotl.monkeypatch.unsloth_ import patch_unsloth_layernorm
-
-            patch_unsloth_layernorm()
-
-        if self.cfg.unsloth_lora_qkv or self.cfg.unsloth_lora_o:
-            from axolotl.monkeypatch.unsloth_ import patch_self_attn_lora
-
-            patch_self_attn_lora()
 
     def _patch_llama_flash_attention(self):
         """Apply Flash Attention patches for LLaMA models."""
@@ -758,23 +752,6 @@ class PatchManager:
                 LOG.info("Patching with SwiGLU...")
                 replace_llama_mlp_with_swiglu(model)
 
-    def _apply_unsloth_patches(self, model):
-        """Apply unsloth optimization patches."""
-        if self.cfg.unsloth_lora_mlp:
-            from axolotl.monkeypatch.unsloth_ import integrate_lora_mlp_patch
-
-            integrate_lora_mlp_patch(peft_model=model)
-
-        if self.cfg.unsloth_lora_qkv or self.cfg.unsloth_lora_o:
-            from axolotl.monkeypatch.unsloth_ import integrate_lora_patch
-
-            integrate_lora_patch(peft_model=model, cfg=self.cfg)
-
-        if self.cfg.unsloth_rope:
-            from axolotl.monkeypatch.unsloth_ import integrate_rope_embeddings
-
-            integrate_rope_embeddings()
-
     def _apply_lora_kernel_patch(self, model):
         """Apply LoRA kernel patches."""
         if (

src/axolotl/monkeypatch/models/gemma4/fused_attn.py

@@ -6,15 +6,29 @@ kernels, eliminating intermediate tensor allocations from rotate_half / apply_ro
 
 Usage:
     from axolotl.monkeypatch.models.gemma4.fused_attn import patch_gemma4_fused_attn
-    patch_gemma4_fused_attn()
+    # Pass install_shared_kv_workaround=True when activation checkpointing is enabled.
+    patch_gemma4_fused_attn(install_shared_kv_workaround=True)
 """
 
-import logging
 from typing import Callable
 
 import torch
 
-logger = logging.getLogger(__name__)
+from axolotl.utils.logging import get_logger
+
+logger = get_logger(__name__)
+
+# Module-level dict used as a side channel for shared KV states avoiding kwarg and TLS
+# to prevent memory leak on gradient checkpoint enabled training (PR #3611)
+_GEMMA4_SHARED_KV_STORE: dict = {"store": None}
+
+
+def _set_shared_kv_states(store):
+    _GEMMA4_SHARED_KV_STORE["store"] = store
+
+
+def _get_shared_kv_states():
+    return _GEMMA4_SHARED_KV_STORE["store"]
 
 
 def _make_fused_forward(original_forward):
@@ -30,7 +44,7 @@ def _make_fused_forward(original_forward):
         hidden_states: torch.Tensor,
         position_embeddings: torch.Tensor,
         attention_mask: torch.Tensor | None,
-        shared_kv_states: dict[int, tuple[torch.Tensor, torch.Tensor]],
+        shared_kv_states: dict[int, tuple[torch.Tensor, torch.Tensor]] | None = None,
         past_key_values=None,
         **kwargs,
     ) -> tuple[torch.Tensor, torch.Tensor | None]:
@@ -39,6 +53,10 @@ def _make_fused_forward(original_forward):
             eager_attention_forward,
         )
 
+        store = _get_shared_kv_states()
+        if store is not None:
+            shared_kv_states = store
+
         input_shape = hidden_states.shape[:-1]
         hidden_shape = (*input_shape, -1, self.head_dim)
         eps = self.config.rms_norm_eps
@@ -133,15 +151,44 @@ def _make_fused_forward(original_forward):
     return fused_forward
 
 
-def patch_gemma4_fused_attn():
+def _patch_decoder_layer_call():
+    """Strip `shared_kv_states` from decoder-layer kwargs and route via the
+    module-level side channel so the checkpoint partial cannot pin it (PR #3611).
     """
-    Monkeypatch Gemma4TextAttention.forward to use fused RMSNorm+RoPE kernels.
+    from transformers.models.gemma4.modeling_gemma4 import Gemma4TextDecoderLayer
+
+    if getattr(Gemma4TextDecoderLayer, "_axolotl_shared_kv_patched", False):
+        return
+
+    original_call = Gemma4TextDecoderLayer.__call__
+
+    def patched_call(self, *args, **kwargs):
+        shared_kv = kwargs.pop("shared_kv_states", None)
+        # Overwrite unconditionally (including with None) so a previous step's
+        # dict cannot leak into a later call without shared_kv_states (PR #3611).
+        _set_shared_kv_states(shared_kv)
+        return original_call(self, *args, **kwargs)
+
+    Gemma4TextDecoderLayer.__call__ = patched_call
+    Gemma4TextDecoderLayer._axolotl_shared_kv_patched = True
+
+
+def patch_gemma4_fused_attn(install_shared_kv_workaround: bool = False):
+    """
+    Monkeypatch Gemma4TextAttention.forward to use fused RMSNorm+RoPE kernels,
+    and optionally route `shared_kv_states` via a module-level side channel to
+    avoid a VRAM leak under activation checkpointing (PR #3611).
     """
     from transformers.models.gemma4.modeling_gemma4 import Gemma4TextAttention
 
     original_forward = Gemma4TextAttention.forward
     Gemma4TextAttention.forward = _make_fused_forward(original_forward)
 
+    if install_shared_kv_workaround:
+        _patch_decoder_layer_call()
+
     logger.info(
         "Patched Gemma4TextAttention.forward with fused RMSNorm+RoPE Triton kernels"
     )
+    if install_shared_kv_workaround:
+        logger.info("Installed Gemma4 shared_kv_states side channel (PR #3611)")

src/axolotl/monkeypatch/unsloth_.py

@@ -1,252 +0,0 @@
-"""module for patching with unsloth optimizations"""
-
-import inspect
-import types
-
-import torch
-from peft import PeftModelForCausalLM
-from torch import nn
-from transformers.models.llama.modeling_llama import LlamaFlashAttention2
-
-from axolotl.monkeypatch.utils import detab_code
-from axolotl.utils.logging import get_logger
-
-LOG = get_logger(__name__)
-
-ORIGINAL_QKV_CODE = """
-    query_states = self.q_proj(hidden_states)
-    key_states = self.k_proj(hidden_states)
-    value_states = self.v_proj(hidden_states)
-""".lstrip("\n")
-
-PATCHED_QKV_CODE = """
-    query_states, key_states, value_states = self.apply_qkv(self, hidden_states)
-""".lstrip("\n")
-
-ORIGINAL_O_CODE = """
-    attn_output = self.o_proj(attn_output)
-""".lstrip("\n")
-
-PATCHED_O_CODE = """
-    attn_output = self.apply_o(self, attn_output)
-""".lstrip("\n")
-
-
-def original_apply_qkv(self, hidden_states):
-    query_states = self.q_proj(hidden_states)
-    key_states = self.k_proj(hidden_states)
-    value_states = self.v_proj(hidden_states)
-    return query_states, key_states, value_states
-
-
-def original_apply_o(self, hidden_states):
-    attn_output = self.o_proj(hidden_states)
-    return attn_output
-
-
-def get_self_attn_code() -> str:
-    forward = inspect.getsource(LlamaFlashAttention2.forward)
-    return forward
-
-
-def check_self_attn_is_patchable() -> bool:
-    qkv = get_self_attn_code()
-    qkv, _ = detab_code(qkv)
-    return ORIGINAL_QKV_CODE in qkv and ORIGINAL_O_CODE in qkv
-
-
-def integrate_cross_entropy_loss_patch(model_type: str = "llama") -> None:
-    from unsloth.kernels.cross_entropy_loss import fast_cross_entropy_loss
-
-    def UnslothForCausalLMLoss(
-        logits,
-        labels,
-        vocab_size: int,
-        num_items_in_batch: int = None,
-        ignore_index: int = -100,
-        **kwargs,
-    ):
-        # Upcast to float if we need to compute the loss to avoid potential precision issues
-        logits = logits.float()
-        # Shift so that tokens < n predict n
-        shift_logits = logits[..., :-1, :].contiguous()
-        shift_labels = labels[..., 1:].contiguous()
-
-        loss = fast_cross_entropy_loss(
-            logits=shift_logits, labels=shift_labels, n_items=num_items_in_batch
-        )
-        return loss
-
-    if model_type == "llama":
-        from transformers.loss import loss_utils
-
-        loss_utils.ForCausalLMLoss = UnslothForCausalLMLoss  # type: ignore[assignment]
-    else:
-        raise ValueError("Unsupported model type")
-
-
-self_attn_lora_patched = False
-
-
-def patch_self_attn_lora():
-    global self_attn_lora_patched
-    if self_attn_lora_patched:
-        # prevent patching multiple times
-        return
-    self_attn_forward = get_self_attn_code()
-    LlamaFlashAttention2._original_forward = self_attn_forward
-    self_attn_forward, _ = detab_code(self_attn_forward)
-    assert ORIGINAL_QKV_CODE in self_attn_forward, "Original qkv code not found"
-    assert ORIGINAL_O_CODE in self_attn_forward, "Original o code not found"
-
-    self_attn_forward = self_attn_forward.replace(ORIGINAL_QKV_CODE, PATCHED_QKV_CODE)
-    self_attn_forward = self_attn_forward.replace(ORIGINAL_O_CODE, PATCHED_O_CODE)
-    self_attn_forward = self_attn_forward.replace(
-        "def forward(",
-        "def unsloth_attn_forward(",
-        1,
-    )
-
-    # load imports necessary
-    import transformers.models.llama.modeling_llama
-
-    items_to_import = []
-    for item in dir(transformers.models.llama.modeling_llama):
-        if item in self_attn_forward:
-            items_to_import.append(item)
-
-    exec(
-        "from transformers.models.llama.modeling_llama import ("
-        + ", ".join(x for x in items_to_import)
-        + ")",
-        globals(),
-    )
-    exec(self_attn_forward, globals())
-    self_attn_lora_patched = True
-    LOG.info("patching unsloth attn lora")
-    LlamaFlashAttention2.forward = unsloth_attn_forward
-
-
-def integrate_rope_embeddings():
-    import transformers.models.llama.modeling_llama
-    from unsloth.kernels.rope_embedding import fast_rope_embedding
-
-    def apply_rotary_pos_emb(
-        q,
-        k,
-        cos,
-        sin,
-        position_ids=None,
-        unsqueeze_dim=1,
-    ):
-        return fast_rope_embedding(q, k, cos, sin)
-
-    LOG.info("patching unsloth RoPE embeddings")
-    transformers.models.llama.modeling_llama.apply_rotary_pos_emb = apply_rotary_pos_emb
-
-
-def integrate_lora_mlp_patch(peft_model: PeftModelForCausalLM):
-    if peft_model.base_model.config.model_type in ["llama", "mistral"]:
-        from unsloth.kernels import apply_lora_mlp_swiglu
-
-        apply_lora_mlp = apply_lora_mlp_swiglu
-    elif peft_model.base_model.config.model_type == "gemma":
-        from unsloth.kernels import apply_lora_mlp_geglu_approx
-
-        apply_lora_mlp = apply_lora_mlp_geglu_approx
-    else:
-        raise NotImplementedError(
-            f"Model type {peft_model.base_model.config.model_type} not supported"
-        )
-
-    for idx, layer in enumerate(peft_model.model.model.layers):
-        layer_modules = [
-            getattr(layer.mlp, linear_proj)
-            for linear_proj in ["gate_proj", "up_proj", "down_proj"]
-        ]
-        is_mlp_lora = all(hasattr(module, "lora_A") for module in layer_modules)
-        mlp_no_bias = all(
-            getattr(module, "base_layer", module).bias is None
-            for module in layer_modules
-        )
-        mlp_not_dora = all(
-            len(getattr(module, "lora_magnitude_vector", []) or []) == 0
-            for module in layer_modules
-        )
-
-        if is_mlp_lora and mlp_no_bias and mlp_not_dora:
-            layer.mlp.forward = types.MethodType(apply_lora_mlp, layer.mlp)
-        else:
-            LOG.warning(f"unable to apply unsloth lora mlp patch to layer {idx}")
-
-
-def integrate_lora_patch(peft_model: PeftModelForCausalLM, cfg):
-    from unsloth.kernels import apply_lora_o, apply_lora_qkv
-
-    for idx, layer in enumerate(peft_model.model.model.layers):
-        if cfg.unsloth_lora_qkv:
-            layer_modules = [
-                getattr(layer.self_attn, linear_proj)
-                for linear_proj in ["q_proj", "k_proj", "v_proj"]
-            ]
-            is_qkv_lora = all(hasattr(module, "lora_A") for module in layer_modules)
-            qkv_no_bias = all(
-                getattr(module, "base_layer", module).bias is None
-                for module in layer_modules
-            )
-            qkv_not_dora = all(
-                len(getattr(module, "lora_magnitude_vector", []) or []) == 0
-                for module in layer_modules
-            )
-
-            if is_qkv_lora and qkv_no_bias and qkv_not_dora:
-                layer.self_attn.apply_qkv = apply_lora_qkv
-            else:
-                layer.self_attn.apply_qkv = original_apply_qkv
-                LOG.warning(f"unable to apply unsloth lora qkv patch to layer {idx}")
-        if cfg.unsloth_lora_o:
-            layer_modules = [
-                getattr(layer.self_attn, linear_proj) for linear_proj in ["o_proj"]
-            ]
-            is_o_lora = all(hasattr(module, "lora_A") for module in layer_modules)
-            o_no_bias = all(
-                getattr(module, "base_layer", module).bias is None
-                for module in layer_modules
-            )
-            o_not_dora = all(
-                len(getattr(module, "lora_magnitude_vector", []) or []) == 0
-                for module in layer_modules
-            )
-
-            if is_o_lora and o_no_bias and o_not_dora:
-                layer.self_attn.apply_o = apply_lora_o
-            else:
-                layer.self_attn.apply_o = original_apply_o
-                LOG.warning(f"unable to apply unsloth lora o_proj patch to layer {idx}")
-
-
-def patch_unsloth_layernorm():
-    try:
-        import transformers.models.llama.modeling_llama
-        from unsloth.kernels.rms_layernorm import Fast_RMS_Layernorm
-
-        class LlamaRMSNorm(nn.Module):
-            """LlamaRMSNorm"""
-
-            def __init__(self, hidden_size, eps=1e-6):
-                """
-                LlamaRMSNorm is equivalent to T5LayerNorm
-                """
-                super().__init__()
-                self.weight = nn.Parameter(torch.ones(hidden_size))
-                self.variance_epsilon = eps
-
-            def forward(self, hidden_states):
-                return Fast_RMS_Layernorm.apply(
-                    hidden_states, self.weight, self.variance_epsilon, False
-                )
-
-        LOG.info("patching with unsloth.kernels.rms_layernorm")
-        transformers.models.llama.modeling_llama.LlamaRMSNorm = LlamaRMSNorm
-    except ImportError:
-        LOG.warning("missing unsloth library")

src/axolotl/utils/schemas/config.py

@@ -823,13 +823,6 @@ class AxolotlInputConfig(
         },
     )
 
-    unsloth_cross_entropy_loss: bool | None = None
-    unsloth_lora_mlp: bool | None = None
-    unsloth_lora_qkv: bool | None = None
-    unsloth_lora_o: bool | None = None
-    unsloth_rms_norm: bool | None = None
-    unsloth_rope: bool | None = None
-
     lora_mlp_kernel: bool | None = Field(
         default=None,
         json_schema_extra={
@@ -1469,21 +1462,6 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
             )
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_multigpu_unsloth(cls, data):
-        if (
-            data.get("unsloth_lora_mlp")
-            or data.get("unsloth_lora_qkv")
-            or data.get("unsloth_lora_o")
-        ):
-            capabilities = data.get("capabilities")
-            if capabilities and capabilities.get("n_gpu", 0) > 1:
-                raise ValueError(
-                    "unsloth_lora_mlp, unsloth_lora_qkv, and unsloth_lora_o are not compatible with multi-GPU training."
-                )
-        return data
-
     @model_validator(mode="before")
     @classmethod
     def check_multigpu_lora_kernels(cls, data):
@@ -1537,8 +1515,7 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
             # RL trainers not tested so don't enable kernels by default
             return data
         if data.get("adapter") in ["lora", "qlora"]:
-            # Skip if already set, using unsloth optimizations, or using 8-bit
-            unsloth_fields = ["unsloth_lora_mlp", "unsloth_lora_qkv", "unsloth_lora_o"]
+            # Skip if already set or using 8-bit
             kernel_fields = [
                 "lora_mlp_kernel",
                 "lora_qkv_kernel",
@@ -1547,7 +1524,6 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
             ]
             if (
                 any(data.get(k) is not None for k in kernel_fields)
-                or any(data.get(k) for k in unsloth_fields)
                 or data.get("adapter") == "lora"
                 and data.get("load_in_8bit")
             ):

src/axolotl/utils/schemas/validation.py

@@ -52,6 +52,26 @@ class DatasetValidationMixin:
 
         return datasets
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_deprecated_unsloth_fields(cls, data):
+        deprecated_fields = [
+            "unsloth_cross_entropy_loss",
+            "unsloth_lora_mlp",
+            "unsloth_lora_qkv",
+            "unsloth_lora_o",
+            "unsloth_rms_norm",
+            "unsloth_rope",
+        ]
+        found = [f for f in deprecated_fields if data.get(f)]
+        if found:
+            raise ValueError(
+                f"`{'`, `'.join(found)}` {'has' if len(found) == 1 else 'have'} been removed. "
+                "Please use `lora_mlp_kernel`, `lora_qkv_kernel`, `lora_o_kernel` instead. "
+                "See: https://docs.axolotl.ai/docs/lora_optims.html"
+            )
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_dataset_or_pretraining_dataset(cls, data):
@@ -607,36 +627,6 @@ class LoRAValidationMixin:
             )
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_qlora_unsloth(cls, data):
-        if (
-            data.get("unsloth_lora_mlp")
-            or data.get("unsloth_lora_qkv")
-            or data.get("unsloth_lora_o")
-        ):
-            if data.get("adapter") == "lora" and data.get("load_in_8bit"):
-                raise ValueError(
-                    "unsloth_lora_mlp, unsloth_lora_qkv, and unsloth_lora_o are not compatible with 8-bit LoRA"
-                )
-        return data
-
-    @model_validator(mode="before")
-    @classmethod
-    def check_lora_axolotl_unsloth(cls, data):
-        is_lora_kernel = any(
-            data.get(k) for k in ["lora_mlp_kernel", "lora_qkv_kernel", "lora_o_kernel"]
-        )
-        is_unsloth_lora = any(
-            data.get(k)
-            for k in ["unsloth_lora_mlp", "unsloth_lora_qkv", "unsloth_lora_o"]
-        )
-        if is_lora_kernel and is_unsloth_lora:
-            raise ValueError(
-                "both lora_mlp_kernel and unsloth_lora_mlp cannot be true (similarly for lora_qkv_kernel, lora_o_kernel)"
-            )
-        return data
-
     @model_validator(mode="after")
     def check_fused_lora(self):
         if self.adapter in ["lora", "qlora"] and self.flash_attn_fuse_mlp:
@@ -860,17 +850,6 @@ class OptimizationValidationMixin:
 
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_xentropy_patch_conflicts(cls, data):
-        if data.get("flash_attn_cross_entropy") and data.get(
-            "unsloth_cross_entropy_loss"
-        ):
-            raise ValueError(
-                "flash_attn_cross_entropy and unsloth_cross_entropy_loss cannot be both enabled"
-            )
-        return data
-
     @model_validator(mode="before")
     @classmethod
     def check_cross_entropy_conflicts(cls, data):

src/setuptools_axolotl_dynamic_dependencies.py

@@ -1,102 +0,0 @@
-"""
-dynamic requirements for axolotl
-"""
-
-import platform
-import re
-from importlib.metadata import PackageNotFoundError, version
-
-from setuptools.command.build_py import build_py as _build_py
-
-
-def parse_requirements():
-    _install_requires = []
-    _dependency_links = []
-    with open("./requirements.txt", encoding="utf-8") as requirements_file:
-        lines = [r.strip() for r in requirements_file.readlines()]
-        for line in lines:
-            is_extras = (
-                "flash-attn" in line
-                or "flash-attention" in line
-                or "deepspeed" in line
-                or "mamba-ssm" in line
-                or "lion-pytorch" in line
-            )
-            if line.startswith("--extra-index-url"):
-                # Handle custom index URLs
-                _, url = line.split()
-                _dependency_links.append(url)
-            elif not is_extras and line and line[0] != "#":
-                # Handle standard packages
-                _install_requires.append(line)
-
-    try:
-        xformers_version = [req for req in _install_requires if "xformers" in req][0]
-        torchao_version = [req for req in _install_requires if "torchao" in req][0]
-
-        if "Darwin" in platform.system():
-            # don't install xformers on MacOS
-            _install_requires.pop(_install_requires.index(xformers_version))
-        else:
-            # detect the version of torch already installed
-            # and set it so dependencies don't clobber the torch version
-            try:
-                torch_version = version("torch")
-            except PackageNotFoundError:
-                torch_version = "2.5.1"
-            _install_requires.append(f"torch=={torch_version}")
-
-            version_match = re.match(r"^(\d+)\.(\d+)(?:\.(\d+))?", torch_version)
-            if version_match:
-                major, minor, patch = version_match.groups()
-                major, minor = int(major), int(minor)
-                patch = (
-                    int(patch) if patch is not None else 0
-                )  # Default patch to 0 if not present
-            else:
-                raise ValueError("Invalid version format")
-
-            if (major, minor) >= (2, 5):
-                _install_requires.pop(_install_requires.index(xformers_version))
-                if patch == 0:
-                    _install_requires.append("xformers==0.0.28.post2")
-                else:
-                    _install_requires.append("xformers==0.0.28.post3")
-            elif (major, minor) >= (2, 4):
-                if patch == 0:
-                    _install_requires.pop(_install_requires.index(xformers_version))
-                    _install_requires.append("xformers>=0.0.27")
-                else:
-                    _install_requires.pop(_install_requires.index(xformers_version))
-                    _install_requires.append("xformers==0.0.28.post1")
-            elif (major, minor) >= (2, 3):
-                _install_requires.pop(_install_requires.index(torchao_version))
-                if patch == 0:
-                    _install_requires.pop(_install_requires.index(xformers_version))
-                    _install_requires.append("xformers>=0.0.26.post1")
-                else:
-                    _install_requires.pop(_install_requires.index(xformers_version))
-                    _install_requires.append("xformers>=0.0.27")
-            elif (major, minor) >= (2, 2):
-                _install_requires.pop(_install_requires.index(torchao_version))
-                _install_requires.pop(_install_requires.index(xformers_version))
-                _install_requires.append("xformers>=0.0.25.post1")
-            else:
-                _install_requires.pop(_install_requires.index(torchao_version))
-                _install_requires.pop(_install_requires.index(xformers_version))
-                _install_requires.append("xformers>=0.0.23.post1")
-
-    except PackageNotFoundError:
-        pass
-    return _install_requires, _dependency_links
-
-
-class BuildPyCommand(_build_py):
-    """
-    custom build_py command to parse dynamic requirements
-    """
-
-    def finalize_options(self):
-        super().finalize_options()
-        install_requires, _ = parse_requirements()
-        self.distribution.install_requires = install_requires

tests/conftest.py

@@ -325,10 +325,10 @@ def download_phi_4_reasoning_model_fixture():
 
 
 @pytest.fixture(scope="session", autouse=True)
-def download_phi_3_medium_model_fixture():
+def download_phi_3_mini_model_fixture():
     # download the tokenizer only
     snapshot_download_w_retry(
-        "microsoft/Phi-3-medium-128k-instruct",
+        "microsoft/Phi-3-mini-4k-instruct",
         repo_type="model",
         allow_patterns=["*token*", "config.json"],
     )

tests/e2e/integrations/test_scattermoe_lora_olmoe.py

@@ -54,24 +54,8 @@ except (ImportError, ModuleNotFoundError):
         )
 
     def peft_lora_to_scattermoe(peft_A, peft_B, num_experts, rank):
-        peft_B_em = peft_lora_B_to_scattermoe(peft_B, num_experts, rank)
-        K_inter, N_hidden = peft_B.shape[0], peft_A.shape[1]
-        smoe_A = torch.zeros(
-            rank * num_experts,
-            K_inter,
-            device=peft_A.device,
-            dtype=peft_A.dtype,
-        )
-        smoe_B = torch.zeros(
-            N_hidden,
-            rank * num_experts,
-            device=peft_A.device,
-            dtype=peft_A.dtype,
-        )
-        for e in range(num_experts):
-            s = e * rank
-            smoe_A[s : s + rank, :] = peft_B_em[:, s : s + rank].T
-            smoe_B[:, s : s + rank] = peft_A[s : s + rank, :].T
+        smoe_A = peft_A
+        smoe_B = peft_lora_B_to_scattermoe(peft_B, num_experts, rank)
         return smoe_A, smoe_B
 
     def _unwrap_experts_lora(experts_module):
@@ -322,12 +306,14 @@ class TestLoRABLayoutConversion:
         hidden, inter = 32, 16
         scaling = 2.0
 
-        peft_A = torch.randn(E * r, hidden)
-        peft_B = torch.randn(inter, E * r)
+        # peft >=0.19.1 for down_proj [E, hidden, inter]:
+        # swaps in/out, lora_A [r*E, inter], lora_B [hidden, r*E]
+        peft_A = torch.randn(E * r, inter)
+        peft_B = torch.randn(hidden, E * r)
 
-        A_r = peft_A.reshape(E, r, hidden)
-        B_r = peft_B.reshape(inter, r, E)
-        delta_peft = torch.einsum("o r e, e r i -> e i o", B_r, A_r) * scaling
+        A_r = peft_A.reshape(E, r, inter)
+        B_r = peft_B.reshape(hidden, r, E)
+        delta_peft = torch.einsum("o r e, e r i -> e o i", B_r, A_r) * scaling
 
         smoe_A, smoe_B = peft_lora_to_scattermoe(peft_A, peft_B, E, r)
         for e in range(E):
@@ -339,31 +325,22 @@ class TestLoRABLayoutConversion:
             )
 
     def test_gate_up_proj_conversion(self):
-        """Verify gate_up_proj LoRA conversion with non-square dims (Qwen3-like).
+        """Verify gate_up_proj LoRA conversion with non-square dims.
 
         gate_up_proj param: [E, 2*inter, hidden].
-        peft: in_features=2*inter, out_features=hidden.
-        peft lora_A: [r*E, 2*inter], lora_B: [hidden, r*E].
-
-        scattermoe W = param.T = [E, hidden, 2*inter], K=hidden, N=2*inter.
+        peft swaps in/out for 3D: lora_A [r*E, hidden], lora_B [2*inter, r*E].
         scattermoe needs: lora_A [r*E, K=hidden], lora_B [N=2*inter, r*E].
-
-        Uses non-square dims (hidden=32 != 2*inter=24) to catch A<->B swap bugs.
         """
         E, r = 4, 2
         hidden, inter = 32, 12  # 2*inter=24 != hidden=32
         scaling = 2.0
 
-        # peft assigns: in_features=2*inter, out_features=hidden
-        peft_A = torch.randn(E * r, 2 * inter)  # [r*E, in_features=2*inter]
-        peft_B = torch.randn(hidden, E * r)  # [out_features=hidden, r*E]
+        peft_A = torch.randn(E * r, hidden)  # [r*E, in=hidden]
+        peft_B = torch.randn(2 * inter, E * r)  # [out=2*inter, r*E]
 
-        # peft delta via einsum: "o r e, e r i -> e i o"
-        A_r = peft_A.reshape(E, r, 2 * inter)
-        B_r = peft_B.reshape(hidden, r, E)
-        delta_peft = torch.einsum("o r e, e r i -> e i o", B_r, A_r) * scaling
-        # delta_peft[e] has shape [in_features, out_features] = [2*inter, hidden]
-        # = param[e] shape [2*inter, hidden]
+        A_r = peft_A.reshape(E, r, hidden)
+        B_r = peft_B.reshape(2 * inter, r, E)
+        delta_peft = torch.einsum("o r e, e r i -> e o i", B_r, A_r) * scaling
 
         smoe_A, smoe_B = peft_gate_up_lora_to_scattermoe(peft_A, peft_B, E, r)
         # smoe_A should be [r*E, K=hidden], smoe_B should be [N=2*inter, r*E]
@@ -421,23 +398,21 @@ class TestPeftLoRAWeightExtraction:
             r,
         )
 
-        # gate_up_proj [E, 2*inter, hidden]
-        # peft: in_features=2*inter (dim 1), out_features=hidden (dim 2)
+        # gate_up_proj [E, 2*inter, hidden] — peft swaps in/out for 3D
         assert trainable[
             "base_model.model.moe.experts.base_layer.lora_A.default.weight"
-        ].shape == (E * r, 2 * config.intermediate_size)
-        assert trainable[
-            "base_model.model.moe.experts.base_layer.lora_B.default.weight"
-        ].shape == (config.hidden_size, E * r)
-
-        # down_proj [E, hidden, inter]
-        # peft: in_features=hidden (dim 1), out_features=inter (dim 2)
-        assert trainable[
-            "base_model.model.moe.experts.lora_A.default.weight"
         ].shape == (E * r, config.hidden_size)
+        assert trainable[
+            "base_model.model.moe.experts.base_layer.lora_B.default.weight"
+        ].shape == (2 * config.intermediate_size, E * r)
+
+        # down_proj [E, hidden, inter] — peft swaps in/out for 3D
+        assert trainable[
+            "base_model.model.moe.experts.lora_A.default.weight"
+        ].shape == (E * r, config.intermediate_size)
         assert trainable[
             "base_model.model.moe.experts.lora_B.default.weight"
-        ].shape == (config.intermediate_size, E * r)
+        ].shape == (config.hidden_size, E * r)
 
     @requires_cuda
     def test_peft_forward_runs(self):
@@ -488,8 +463,7 @@ class TestPeftLoRAWeightExtraction:
         assert gup_lora is not None, "gate_up_proj LoRA not detected"
         assert down_lora is not None, "down_proj LoRA not detected"
 
-        # Check shapes (after peft->scattermoe conversion with A<->B swap)
-        # gate_up_proj W = param.T = [E, hidden, 2*inter], K=hidden, N=2*inter
+        # gate_up_proj: K=hidden, N=2*inter
         E, r = config.num_experts, 4
         gup_A, gup_B, gup_s = gup_lora
         assert gup_A.shape == (E * r, config.hidden_size), (
@@ -501,7 +475,7 @@ class TestPeftLoRAWeightExtraction:
             f"{(2 * config.intermediate_size, E * r)}, got {gup_B.shape}"
         )
 
-        # down_proj W = param.T = [E, inter, hidden], K=inter, N=hidden
+        # down_proj: K=inter, N=hidden
         down_A, down_B, down_s = down_lora
         assert down_A.shape == (E * r, config.intermediate_size), (
             f"down_proj smoe_A: expected [r*E, K=inter]={(E * r, config.intermediate_size)}, "

tests/e2e/patched/test_unsloth_integration.py

@@ -1,21 +0,0 @@
-"""Test module for checking whether the integration of Unsloth with Hugging Face Transformers is working as expected."""
-
-import unittest
-
-import pytest
-
-
-@pytest.mark.skip(
-    reason="Unsloth integration will be broken going into latest transformers"
-)
-class TestUnslothIntegration(unittest.TestCase):
-    """Unsloth monkeypatch integration tests."""
-
-    def test_is_self_attn_patchable(self):
-        from axolotl.monkeypatch.unsloth_ import check_self_attn_is_patchable
-
-        # ensures the current version of transformers has loss code that matches our patching code
-        self.assertTrue(
-            check_self_attn_is_patchable(),
-            "HF transformers self attention code has changed and isn't patchable",
-        )

tests/e2e/patched/test_unsloth_qlora.py

@@ -1,184 +0,0 @@
-"""
-e2e tests for unsloth qlora
-"""
-
-import pytest
-
-from axolotl.common.datasets import load_datasets
-from axolotl.train import train
-from axolotl.utils.config import normalize_config, validate_config
-from axolotl.utils.dict import DictDefault
-
-from ..utils import check_model_output_exists, check_tensorboard
-
-
-@pytest.mark.skip(
-    reason="Unsloth integration will be broken going into latest transformers"
-)
-class TestUnslothQLoRA:
-    """
-    Test class for Unsloth QLoRA Llama models
-    """
-
-    @pytest.mark.parametrize(
-        "sample_packing",
-        [True, False],
-    )
-    def test_unsloth_llama_qlora_fa2(self, temp_dir, sample_packing):
-        cfg = DictDefault(
-            {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
-                "sequence_len": 1024,
-                "sample_packing": sample_packing,
-                "flash_attention": True,
-                "unsloth_lora_mlp": True,
-                "unsloth_lora_qkv": True,
-                "unsloth_lora_o": True,
-                "load_in_4bit": True,
-                "adapter": "qlora",
-                "lora_r": 16,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "val_set_size": 0.05,
-                "special_tokens": {
-                    "pad_token": "<|endoftext|>",
-                },
-                "datasets": [
-                    {
-                        "path": "mhenrichsen/alpaca_2k_test",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 5,
-                "save_steps": 10,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 2,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_8bit",
-                "lr_scheduler": "cosine",
-                "use_tensorboard": True,
-                "bf16": "auto",
-                "save_first_step": False,
-            }
-        )
-
-        cfg = validate_config(cfg)
-        normalize_config(cfg)
-        dataset_meta = load_datasets(cfg=cfg)
-
-        train(cfg=cfg, dataset_meta=dataset_meta)
-        check_model_output_exists(temp_dir, cfg)
-
-        check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss (%s) is too high"
-        )
-
-    def test_unsloth_llama_qlora_unpacked(self, temp_dir):
-        cfg = DictDefault(
-            {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
-                "sequence_len": 1024,
-                "unsloth_lora_mlp": True,
-                "unsloth_lora_qkv": True,
-                "unsloth_lora_o": True,
-                "sample_packing": False,
-                "load_in_4bit": True,
-                "adapter": "qlora",
-                "lora_r": 16,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "val_set_size": 0.05,
-                "special_tokens": {
-                    "pad_token": "<|endoftext|>",
-                },
-                "datasets": [
-                    {
-                        "path": "mhenrichsen/alpaca_2k_test",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 5,
-                "save_steps": 10,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 2,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_8bit",
-                "lr_scheduler": "cosine",
-                "use_tensorboard": True,
-                "bf16": "auto",
-                "save_first_step": False,
-            }
-        )
-
-        cfg = validate_config(cfg)
-        normalize_config(cfg)
-        dataset_meta = load_datasets(cfg=cfg)
-
-        train(cfg=cfg, dataset_meta=dataset_meta)
-        check_model_output_exists(temp_dir, cfg)
-
-        check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss (%s) is too high"
-        )
-
-    @pytest.mark.parametrize(
-        "sdp_attention",
-        [True, False],
-    )
-    def test_unsloth_llama_qlora_unpacked_no_fa2_fp16(self, temp_dir, sdp_attention):
-        cfg = DictDefault(
-            {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
-                "sequence_len": 1024,
-                "unsloth_lora_mlp": True,
-                "unsloth_lora_qkv": True,
-                "unsloth_lora_o": True,
-                "sample_packing": False,
-                "load_in_4bit": True,
-                "adapter": "qlora",
-                "lora_r": 16,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "val_set_size": 0.05,
-                "special_tokens": {
-                    "pad_token": "<|endoftext|>",
-                },
-                "datasets": [
-                    {
-                        "path": "mhenrichsen/alpaca_2k_test",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 5,
-                "save_steps": 10,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 2,
-                "sdp_attention": sdp_attention,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_8bit",
-                "lr_scheduler": "cosine",
-                "use_tensorboard": True,
-                "fp16": True,
-                "save_first_step": False,
-            }
-        )
-
-        cfg = validate_config(cfg)
-        normalize_config(cfg)
-        dataset_meta = load_datasets(cfg=cfg)
-
-        train(cfg=cfg, dataset_meta=dataset_meta)
-        check_model_output_exists(temp_dir, cfg)
-
-        check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss (%s) is too high"
-        )

tests/monkeypatch/test_gemma4_fused_attn_patch.py

@@ -0,0 +1,171 @@
+"""Unit tests for the Gemma4 fused-attention shared_kv_states routing patch."""
+
+import pytest
+
+gemma4_modeling = pytest.importorskip("transformers.models.gemma4.modeling_gemma4")
+
+
+@pytest.fixture
+def clean_decoder_layer_patch_slate():
+    """Save and restore Gemma4TextDecoderLayer.__call__ and the sentinel."""
+    from axolotl.monkeypatch.models.gemma4 import fused_attn
+
+    cls = gemma4_modeling.Gemma4TextDecoderLayer
+    original_call = cls.__call__
+    had_sentinel = getattr(cls, "_axolotl_shared_kv_patched", False)
+
+    if had_sentinel:
+        del cls._axolotl_shared_kv_patched
+
+    try:
+        yield cls, fused_attn
+    finally:
+        cls.__call__ = original_call
+        if had_sentinel:
+            cls._axolotl_shared_kv_patched = True
+        elif hasattr(cls, "_axolotl_shared_kv_patched"):
+            del cls._axolotl_shared_kv_patched
+        fused_attn._set_shared_kv_states(None)
+
+
+class TestPatchedDecoderLayerCall:
+    def test_pops_shared_kv_states_and_populates_store(
+        self, clean_decoder_layer_patch_slate
+    ):
+        cls, fused_attn = clean_decoder_layer_patch_slate
+
+        captured = {}
+
+        def spy(self, *args, **kwargs):
+            captured["args"] = args
+            captured["kwargs"] = dict(kwargs)
+            return "spy_return"
+
+        cls.__call__ = spy
+        fused_attn._patch_decoder_layer_call()
+
+        assert getattr(cls, "_axolotl_shared_kv_patched", False) is True
+        assert cls.__call__ is not spy
+
+        shared_kv = {"layer_0": ("k", "v")}
+        result = cls.__call__(
+            object(),
+            "positional_arg",
+            shared_kv_states=shared_kv,
+            other_kwarg="keep_me",
+        )
+
+        assert result == "spy_return"
+        assert captured["args"] == ("positional_arg",)
+        assert "shared_kv_states" not in captured["kwargs"]
+        assert captured["kwargs"] == {"other_kwarg": "keep_me"}
+        assert fused_attn._get_shared_kv_states() is shared_kv
+
+    def test_clears_store_when_kwarg_absent(self, clean_decoder_layer_patch_slate):
+        """Regression for commit 251021e1: a prior step's dict must not leak
+        into a later call that omits `shared_kv_states`."""
+        cls, fused_attn = clean_decoder_layer_patch_slate
+
+        def spy(self, *args, **kwargs):
+            return None
+
+        cls.__call__ = spy
+        fused_attn._patch_decoder_layer_call()
+
+        stale = {"stale_step": True}
+        fused_attn._set_shared_kv_states(stale)
+        assert fused_attn._get_shared_kv_states() is stale
+
+        cls.__call__(object())
+
+        assert fused_attn._get_shared_kv_states() is None
+
+    def test_store_visible_across_threads(self):
+        """Regression for commit e3669b2c: the store must be readable from
+        threads other than the one that set it. `threading.local()` failed
+        this invariant, crashing with 'NoneType' object is not subscriptable'
+        on MoE Gemma4 variants when autograd worker threads ran backward
+        recompute under HF-Trainer gradient_checkpointing."""
+        import threading
+
+        from axolotl.monkeypatch.models.gemma4 import fused_attn
+
+        sentinel = {"layer_0": ("k", "v")}
+        try:
+            fused_attn._set_shared_kv_states(sentinel)
+
+            seen = {}
+
+            def worker():
+                seen["value"] = fused_attn._get_shared_kv_states()
+
+            t = threading.Thread(target=worker)
+            t.start()
+            t.join()
+
+            assert seen["value"] is sentinel
+        finally:
+            fused_attn._set_shared_kv_states(None)
+
+
+@pytest.fixture
+def clean_entry_point_patch_slate():
+    """Save and restore Gemma4TextAttention.forward and Gemma4TextDecoderLayer.__call__."""
+    from axolotl.monkeypatch.models.gemma4 import fused_attn
+
+    decoder_cls = gemma4_modeling.Gemma4TextDecoderLayer
+    attn_cls = gemma4_modeling.Gemma4TextAttention
+
+    original_call = decoder_cls.__call__
+    original_forward = attn_cls.forward
+    had_sentinel = getattr(decoder_cls, "_axolotl_shared_kv_patched", False)
+
+    if had_sentinel:
+        del decoder_cls._axolotl_shared_kv_patched
+
+    try:
+        yield decoder_cls, attn_cls, original_call, original_forward, fused_attn
+    finally:
+        decoder_cls.__call__ = original_call
+        attn_cls.forward = original_forward
+        if had_sentinel:
+            decoder_cls._axolotl_shared_kv_patched = True
+        elif hasattr(decoder_cls, "_axolotl_shared_kv_patched"):
+            del decoder_cls._axolotl_shared_kv_patched
+        fused_attn._set_shared_kv_states(None)
+
+
+class TestPatchGemma4FusedAttnEntryPoint:
+    def test_default_flag_swaps_only_attention_forward(
+        self, clean_entry_point_patch_slate
+    ):
+        (
+            decoder_cls,
+            attn_cls,
+            original_call,
+            original_forward,
+            fused_attn,
+        ) = clean_entry_point_patch_slate
+
+        fused_attn.patch_gemma4_fused_attn()
+
+        assert attn_cls.forward is not original_forward
+        assert decoder_cls.__call__ is original_call
+        assert not getattr(decoder_cls, "_axolotl_shared_kv_patched", False)
+
+    def test_workaround_flag_installs_decoder_layer_patch(
+        self, clean_entry_point_patch_slate
+    ):
+        (
+            decoder_cls,
+            attn_cls,
+            original_call,
+            original_forward,
+            fused_attn,
+        ) = clean_entry_point_patch_slate
+
+        fused_attn.patch_gemma4_fused_attn(install_shared_kv_workaround=True)
+
+        assert attn_cls.forward is not original_forward
+        assert decoder_cls.__call__ is not original_call
+        assert getattr(decoder_cls, "_axolotl_shared_kv_patched", False) is True

tests/prompt_strategies/test_dpo_chat_templates.py

@@ -111,7 +111,7 @@ def fixture_argilla_chat_dataset():
 @pytest.fixture(name="phi3_tokenizer")
 @enable_hf_offline
 def fixture_phi3_tokenizer():
-    tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-medium-128k-instruct")
+    tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
 
     return tokenizer
 
@@ -214,8 +214,8 @@ class TestAssistantDPOChatTemplatePhi3:
             + "<|user|>\ngoodbye<|end|>\n"
             + "<|assistant|>\n"
         )
-        assert result["chosen"] == "goodbye<|end|>"
-        assert result["rejected"] == "party on<|end|>"
+        assert result["chosen"] == "goodbye<|end|>\n<|endoftext|>"
+        assert result["rejected"] == "party on<|end|>\n<|endoftext|>"
 
 
 class TestAssistantDPOChatTemplateGemma:
@@ -290,8 +290,8 @@ class TestArgillaChatDPOChatTemplate:
         )
         result = transform_fn(argilla_chat_dataset[0], tokenizer=phi3_tokenizer)
         assert result["prompt"] == "<|user|>\nhello<|end|>\n" + "<|assistant|>\n"
-        assert result["chosen"] == "goodbye<|end|>"
-        assert result["rejected"] == "party on<|end|>"
+        assert result["chosen"] == "goodbye<|end|>\n<|endoftext|>"
+        assert result["rejected"] == "party on<|end|>\n<|endoftext|>"
 
 
 class TestDPOChatTemplateToolRole: