draft config for devstral

* models.py -> loaders/ module refactor

* refactor ModelLoader class

* plugin manager changes

* circular import fix

* pytest

* pytest

* minor improvements

* fix

* minor changes

* fix test

* remove dead code

* coderabbit comments

* lint

* fix

* coderabbit suggestion I liked

* more coderabbit

* review comments, yak shaving

* lint

* updating in light of SP ctx manager changes

* review comment

* review comment 2

.github/workflows/main.yml

@@ -31,6 +31,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.0
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.7.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -94,6 +99,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.0
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.7.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

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

@@ -3,7 +3,7 @@ name: docker-multigpu-tests-biweekly
 on:
   pull_request:
     paths:
-      - 'tests/e2e/multigpu/*.py'
+      - 'tests/e2e/multigpu/**.py'
       - 'requirements.txt'
       - 'setup.py'
       - 'pyproject.toml'

.github/workflows/tests-nightly.yml

@@ -18,9 +18,96 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
+  preload-cache:
+    name: Preload HF cache
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python_version: ["3.11"]
+        pytorch_version: ["2.6.0"]
+    timeout-minutes: 20
+
+    env:
+      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
+
+    steps:
+      - name: Check out repository code
+        uses: actions/checkout@v4
+
+      - name: Restore HF cache
+        id: hf-cache-restore
+        uses: actions/cache/restore@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ runner.os }}-hf-hub-cache-v2
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python_version }}
+          cache: 'pip' # caching pip dependencies
+
+      - name: upgrade pip
+        run: |
+          pip3 install --upgrade pip
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+
+      - name: Install PyTorch
+        run: |
+          pip3 install torch==${{ matrix.pytorch_version }}
+
+      - 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
+
+      - name: Make sure PyTorch version wasn't clobbered
+        run: |
+          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+
+      - name: Ensure axolotl CLI was installed
+        run: |
+          axolotl --help
+
+      - name: Pre-Download dataset fixture
+        run: |
+          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+
+      - name: Run tests
+        run: |
+          pytest -v tests/conftest.py
+
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v5
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          files: ./coverage.xml
+          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+          fail_ci_if_error: false
+
+      - name: cleanup pip cache
+        run: |
+          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+
+      - name: Save HF cache
+        id: hf-cache
+        uses: actions/cache/save@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
+    needs: [preload-cache]
     strategy:
       fail-fast: false
       max-parallel: 2

.github/workflows/tests.yml

@@ -44,12 +44,104 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
+#  preload-cache:
+#    name: Preload HF cache
+#    runs-on: ubuntu-latest
+#    strategy:
+#      fail-fast: false
+#      matrix:
+#        python_version: ["3.11"]
+#        pytorch_version: ["2.6.0"]
+#    timeout-minutes: 20
+#
+#    env:
+#      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
+#
+#    steps:
+#      - name: Check out repository code
+#        uses: actions/checkout@v4
+#
+#      - name: Restore HF cache
+#        id: hf-cache-restore
+#        uses: actions/cache/restore@v4
+#        with:
+#          path: |
+#            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/models--*
+#          key: ${{ runner.os }}-hf-hub-cache-v2
+#
+#      - name: Restore Cache from S3
+#        id: hf-cache-restore-s3
+#        run: |
+#          mkdir -p /home/runner/.cache/huggingface/hub
+#          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
+#
+#      - name: Setup Python
+#        uses: actions/setup-python@v5
+#        with:
+#          python-version: ${{ matrix.python_version }}
+#          cache: 'pip' # caching pip dependencies
+#
+#      - name: upgrade pip
+#        run: |
+#          pip3 install --upgrade pip
+#          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+#
+#      - name: Install PyTorch
+#        run: |
+#          pip3 install torch==${{ matrix.pytorch_version }}
+#
+#      - 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
+#
+#      - name: Make sure PyTorch version wasn't clobbered
+#        run: |
+#          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+#
+#      - name: Ensure axolotl CLI was installed
+#        run: |
+#          axolotl --help
+#
+#      - name: Pre-Download dataset fixture
+#        run: |
+#          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+#
+#      - name: Run tests
+#        run: |
+#          pytest -v tests/conftest.py
+#
+#      - name: Upload coverage to Codecov
+#        uses: codecov/codecov-action@v5
+#        with:
+#          token: ${{ secrets.CODECOV_TOKEN }}
+#          files: ./coverage.xml
+#          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+#          fail_ci_if_error: false
+#
+#      - name: cleanup pip cache
+#        run: |
+#          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+#
+#      - name: Save HF cache
+#        id: hf-cache
+#        uses: actions/cache/save@v4
+#        with:
+#          path: |
+#            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/models--*
+#          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
+#    needs: [preload-cache]
     strategy:
       fail-fast: false
-      max-parallel: 2
       matrix:
         python_version: ["3.11"]
         pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
@@ -59,14 +151,20 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore HF cache
-        id: hf-cache-restore
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
+#      - name: Restore HF cache
+#        id: hf-cache-restore
+#        uses: actions/cache/restore@v4
+#        with:
+#          path: |
+#            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/models--*
+#          key: ${{ runner.os }}-hf-hub-cache-v2
+
+      - name: Restore Cache from S3
+        id: hf-cache-restore-s3
+        run: |
+          mkdir -p /home/runner/.cache/huggingface/hub
+          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -121,21 +219,12 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
-      - name: Save HF cache
-        id: hf-cache
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
-
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
+#    needs: [preload-cache]
     strategy:
       fail-fast: false
-      max-parallel: 1
       matrix:
         python_version: ["3.11"]
         pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
@@ -145,14 +234,20 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore HF cache
-        id: hf-cache-restore
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
+#      - name: Restore HF cache
+#        id: hf-cache-restore
+#        uses: actions/cache/restore@v4
+#        with:
+#          path: |
+#            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/models--*
+#          key: ${{ runner.os }}-hf-hub-cache-v2
+
+      - name: Restore Cache from S3
+        id: hf-cache-restore-s3
+        run: |
+          mkdir -p /home/runner/.cache/huggingface/hub
+          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -199,16 +294,8 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
-      - name: Save HF cache
-        id: hf-cache
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
-
   docker-e2e-tests-1st:
+    # Run this job first as a gate for running the remainder of the test matrix
     if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
@@ -255,6 +342,8 @@ jobs:
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 90
+    # Only run the remainder of the matrix if the first e2e check passed;
+    # this is to save on wasted compute costs for known failures that get caught in the first run
     needs: [pre-commit, pytest, docker-e2e-tests-1st]
 
     strategy:
@@ -267,12 +356,6 @@ jobs:
             pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras: llmcompressor
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            num_gpus: 1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -285,6 +368,12 @@ jobs:
             pytorch: 2.7.0
             num_gpus: 1
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.7.0
+            num_gpus: 1
+            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -309,3 +398,43 @@ jobs:
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
+
+  docker-e2e-cleanup:
+    runs-on: [self-hosted, modal]
+    timeout-minutes: 90
+    needs: [docker-e2e-tests]
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            num_gpus: 1
+            axolotl_extras: vllm
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Install Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Install Modal
+        run: |
+          python -m pip install --upgrade pip
+          pip install modal==0.71.8 jinja2
+      - name: Update env vars
+        run: |
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
+          echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
+          echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
+          echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
+          echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
+          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
+      - name: Run tests job on Modal
+        run: |
+          modal run cicd.cleanup

.runpod/src/handler.py

@@ -57,8 +57,10 @@ async def handler(job):
     logger.info("Training Complete.")
 
     # Cleanup
-    del os.environ["WANDB_API_KEY"]
-    del os.environ["HF_TOKEN"]
+    if "WANDB_API_KEY" in os.environ:
+        del os.environ["WANDB_API_KEY"]
+    if "HF_TOKEN" in os.environ:
+        del os.environ["HF_TOKEN"]
 
 
 runpod.serverless.start({"handler": handler, "return_aggregate_stream": True})

_quarto.yml

@@ -48,8 +48,22 @@ quartodoc:
       contents:
         - core.trainers.base
         - core.trainers.trl
+        - core.trainers.mamba
+        - core.trainers.relora
         - core.trainers.dpo.trainer
         - core.trainers.grpo.trainer
+        - core.trainers.grpo.sampler
+        - core.trainers.utils
+    - title: Mixins
+      desc: Mixin classes for augmenting trainers
+      contents:
+        - core.trainers.mixins.optimizer
+        - core.trainers.mixins.rng_state_loader
+        - core.trainers.mixins.scheduler
+    - title: Context Managers
+      desc: Context managers for altering trainer behaviors
+      contents:
+        - utils.ctx_managers.sequence_parallel
     - title: Prompt Strategies
       desc: Prompt formatting strategies
       contents:
@@ -86,7 +100,7 @@ quartodoc:
         - kernels.swiglu
         - kernels.quantize
         - kernels.utils
-    - title: MonkeyPatches
+    - title: Monkey Patches
       desc: Runtime patches for model optimizations
       contents:
         - monkeypatch.llama_attn_hijack_flash
@@ -124,7 +138,8 @@ quartodoc:
         - utils.optimizers.adopt
         - utils.data.pretraining
         - utils.data.sft
-        - utils.gradient_checkpointing.unsloth
+        - utils.gradient_checkpointing.offload_cpu
+        - utils.gradient_checkpointing.offload_disk
     - title: Schemas
       desc: Pydantic data models for Axolotl config
       contents:

cicd/cicd.sh

@@ -18,7 +18,7 @@ pytest -v --durations=10 \
   --cov-append
 
 # Run patched tests excluding lora kernels with coverage append
-pytest -v --durations=10 \
+pytest --full-trace -vvv --durations=10 \
   --ignore=tests/e2e/patched/lora_kernels \
   /workspace/axolotl/tests/e2e/patched \
   --cov=axolotl \

cicd/cleanup.py

@@ -0,0 +1,19 @@
+"""Modal app to run axolotl GPU cleanup"""
+
+from .single_gpu import VOLUME_CONFIG, app, cicd_image, run_cmd
+
+
+@app.function(
+    image=cicd_image,
+    timeout=60 * 60,
+    cpu=8.0,
+    memory=131072,
+    volumes=VOLUME_CONFIG,
+)
+def cleanup():
+    run_cmd("./cicd/cleanup.sh", "/workspace/axolotl")
+
+
+@app.local_entrypoint()
+def main():
+    cleanup.remote()

cicd/cleanup.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+set -e
+
+# cleanup old cache files for datasets processing and intermediate mappings
+find /workspace/data/huggingface-cache/hub/datasets -name "cache-*" -type f -mtime +1 -exec rm {} \;
+find /workspace/data/huggingface-cache/hub/datasets -name "*.lock" -type f -mtime +1 -exec rm {} \;

cicd/e2e_tests.py

@@ -1,75 +1,12 @@
 """Modal app to run axolotl GPU tests"""
 
-# pylint: disable=duplicate-code
-
-import os
-import pathlib
-import tempfile
-
-import jinja2
-import modal
-from jinja2 import select_autoescape
-from modal import App, Image
-
-cicd_path = pathlib.Path(__file__).parent.resolve()
-
-template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
-template_env = jinja2.Environment(
-    loader=template_loader, autoescape=select_autoescape()
-)
-df_template = template_env.get_template("Dockerfile.jinja")
-
-df_args = {
-    "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
-    "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
-    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.4.1"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu121-2.4.1"),
-    "CUDA": os.environ.get("CUDA", "121"),
-    "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
-    "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
-    "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
-    "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
-    "HF_HOME": "/workspace/data/huggingface-cache/hub",
-}
-
-dockerfile_contents = df_template.render(**df_args)
-
-temp_dir = tempfile.mkdtemp()
-with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
-    f.write(dockerfile_contents)
-
-cicd_image = Image.from_dockerfile(
-    pathlib.Path(temp_dir) / "Dockerfile",
-    context_mount=None,
-    force_build=True,
-    gpu="A10G",
-).env(df_args)
-
-app = App("Axolotl CI/CD", secrets=[])
-
-hf_cache_volume = modal.Volume.from_name(
-    "axolotl-ci-hf-hub-cache", create_if_missing=True
-)
-VOLUME_CONFIG = {
-    "/workspace/data/huggingface-cache/hub": hf_cache_volume,
-}
-
-N_GPUS = int(os.environ.get("N_GPUS", 1))
-GPU_CONFIG = modal.gpu.L40S(count=N_GPUS)
-
-
-def run_cmd(cmd: str, run_folder: str):
-    import subprocess  # nosec
-
-    # Propagate errors from subprocess.
-    if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
-        exit(exit_code)  # pylint: disable=consider-using-sys-exit
+from .single_gpu import GPU_CONFIG, VOLUME_CONFIG, app, cicd_image, run_cmd
 
 
 @app.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=60 * 60,
+    timeout=90 * 60,  # 90 min
     cpu=8.0,
     memory=131072,
     volumes=VOLUME_CONFIG,

cicd/multigpu.py

@@ -70,7 +70,7 @@ def run_cmd(cmd: str, run_folder: str):
     image=cicd_image,
     gpu=GPU_CONFIG,
     timeout=90 * 60,
-    cpu=8.0,
+    cpu=16.0,
     memory=131072 * N_GPUS,
     volumes=VOLUME_CONFIG,
 )

cicd/single_gpu.py

@@ -0,0 +1,66 @@
+"""Modal app to run axolotl GPU tests"""
+
+# pylint: disable=duplicate-code
+
+import os
+import pathlib
+import tempfile
+
+import jinja2
+import modal
+from jinja2 import select_autoescape
+from modal import App, Image
+
+cicd_path = pathlib.Path(__file__).parent.resolve()
+
+template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
+template_env = jinja2.Environment(
+    loader=template_loader, autoescape=select_autoescape()
+)
+df_template = template_env.get_template("Dockerfile.jinja")
+
+df_args = {
+    "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
+    "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.4.1"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu121-2.4.1"),
+    "CUDA": os.environ.get("CUDA", "121"),
+    "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
+    "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
+    "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
+    "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
+    "HF_HOME": "/workspace/data/huggingface-cache/hub",
+}
+
+dockerfile_contents = df_template.render(**df_args)
+
+temp_dir = tempfile.mkdtemp()
+with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
+    f.write(dockerfile_contents)
+
+cicd_image = Image.from_dockerfile(
+    pathlib.Path(temp_dir) / "Dockerfile",
+    context_mount=None,
+    force_build=True,
+    gpu="A10G",
+).env(df_args)
+
+app = App("Axolotl CI/CD", secrets=[])
+
+hf_cache_volume = modal.Volume.from_name(
+    "axolotl-ci-hf-hub-cache", create_if_missing=True
+)
+VOLUME_CONFIG = {
+    "/workspace/data/huggingface-cache/hub": hf_cache_volume,
+}
+
+N_GPUS = int(os.environ.get("N_GPUS", 1))
+GPU_CONFIG = modal.gpu.L40S(count=N_GPUS)
+
+
+def run_cmd(cmd: str, run_folder: str):
+    import subprocess  # nosec
+
+    # Propagate errors from subprocess.
+    if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
+        exit(exit_code)  # pylint: disable=consider-using-sys-exit

codecov.yml

@@ -19,7 +19,7 @@ coverage:
         if_no_uploads: error
         if_not_found: success
         if_ci_failed: error
-        only_pulls: false
+        only_pulls: true
         flags: null
         paths: null
     patch:

docs/config.qmd

@@ -32,6 +32,8 @@ tokenizer_legacy:
 resize_token_embeddings_to_32x:
 # Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.
 shrink_embeddings:
+# Optional[bool] Don't upcast the embeddings to float32 when using PEFT. Useful for low-VRAM GPUs
+embeddings_skip_upcast:
 # Whether to load the model with randomly initialized weights. Useful for
 # pre-training a model from scratch or debugging purposes.
 random_init_weights:
@@ -73,11 +75,12 @@ load_in_8bit: true
 load_in_4bit:
 
 # Use CUDA bf16
-bf16: true # bool or 'full' for `bf16_full_eval`. require >=ampere
+bf16: true # bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection. require >=ampere
 # Use CUDA fp16
 fp16: true
 # Use CUDA tf32
 tf32: true # require >=ampere
+# Note: if bf16 is set to 'auto', and fp16 is set to true, we will prefer the explict fp16 setting
 
 # No AMP (automatic mixed precision)
 bfloat16: true # require >=ampere
@@ -184,8 +187,8 @@ datasets:
     # adding a system turn with empty content.
     drop_system_message:
 
-    # Optional[bool]. Whether to split the assistant turn based on a reasoning trace inside delimited tags
-    # defaults to False
+    # Optional[bool]. (for Qwen3 template only) Whether to split the assistant content based on a reasoning trace inside delimited tags
+    # See example at `docs/dataset-formats/conversation.qmd`
     split_thinking:
 
     # IMPORTANT: The following fields determine which parts of the conversation to train on.
@@ -502,6 +505,7 @@ save_strategy: # Set to `"no"` to skip checkpoint saves, `"epoch"` at end of eac
 save_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps
 saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps
 save_total_limit: # Checkpoints saved at a time
+save_only_model: # Save only the model weights, skipping the optimizer. Using this means you can't resume from checkpoints.
 # Maximum number of iterations to train for. It precedes num_epochs which means that
 # if both are set, num_epochs will not be guaranteed.
 # e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps
@@ -535,7 +539,7 @@ train_on_inputs: false
 # Note that training loss may have an oscillating pattern with this enabled.
 group_by_length: false
 
-# Whether to use gradient checkpointing. Available options are: true, false, "offload".
+# Whether to use gradient checkpointing. Available options are: true, false, "offload", "offload_disk".
 # https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 gradient_checkpointing: false
 # additional kwargs to pass to the trainer for gradient checkpointing
@@ -547,7 +551,7 @@ gradient_checkpointing: false
 early_stopping_patience: 3
 
 # Specify a scheduler and kwargs to use with the optimizer
-lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine
+lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | 'linear' | 'cosine_with_restarts' | 'polynomial' | 'constant' | 'constant_with_warmup' | 'inverse_sqrt' | 'reduce_lr_on_plateau' | 'cosine_with_min_lr' | 'warmup_stable_decay' | empty for cosine
 lr_scheduler_kwargs:
 cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr
 cosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)
@@ -609,6 +613,7 @@ lr_div_factor: # Learning rate div factor
 # - optimi_adamw
 # - ao_adamw_8bit
 # - ao_adamw_fp8
+# - came_pytorch
 optimizer:
 # Dictionary of arguments to pass to the optimizer
 optim_args:
@@ -628,7 +633,9 @@ weight_decay:
 # adamw hyperparams
 adam_beta1:
 adam_beta2:
+adam_beta3:  # only used for CAME Optimizer
 adam_epsilon:
+adam_epsilon2:  # only used for CAME Optimizer
 # Gradient clipping max norm
 max_grad_norm:
 

docs/dataset-formats/conversation.qmd

@@ -196,6 +196,34 @@ datasets:
 It is not necessary to set both `message_field_training` and `message_field_training_detail` at once.
 :::
 
+8. (For Qwen3 template only) Enable reasoning split, where the reasoning is split from the content and passed as a separate field into the template.
+
+```yaml
+datasets:
+  - path: ...
+    type: chat_template
+    chat_template: qwen3
+    split_thinking: true
+```
+
+For example, a content can look like:
+
+```json
+{
+  "content": "<think>Some thinking outputs</think>Output after thinking."
+}
+```
+
+After split, it will look like:
+
+```json
+{
+  "reasoning_content": "Some thinking outputs",
+  "content": "Output after thinking..."
+}
+```
+
+
 ## sharegpt
 
 ::: {.callout-important}

docs/docker.qmd

@@ -8,6 +8,10 @@ format:
 
 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 2.7.0 and CUDA 12.8.
+:::
+
 ## 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.

docs/getting-started.qmd

@@ -104,7 +104,7 @@ the `alpaca` dataset format, which has the following format:
 Please see our [Dataset Formats](dataset-formats) for more dataset formats and how to
 format them.
 
-2. Prepare your JSONL data in the specified format (in this case, the expected `alpaca
+2. Prepare your JSONL data in the specified format (in this case, the expected `alpaca`
 format):
 
 ```json
@@ -120,6 +120,12 @@ axolotl train my_training.yml
 
 ## Common Tasks {#sec-common-tasks}
 
+::: {.callout-tip}
+
+The same yaml file is used for training, inference, and merging.
+
+:::
+
 ### Testing Your Model {#sec-testing}
 
 After training, test your model:
@@ -128,6 +134,16 @@ After training, test your model:
 axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out"
 ```
 
+More details can be found in [Inference](inference.qmd).
+
+### Using a UI {#sec-ui}
+
+Launch a Gradio interface:
+
+```bash
+axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out" --gradio
+```
+
 ### Preprocessing Data {#sec-preprocessing}
 
 For large datasets, preprocess first:
@@ -136,14 +152,22 @@ For large datasets, preprocess first:
 axolotl preprocess my_training.yml
 ```
 
-### Using a UI {#sec-ui}
+Please make sure to set `dataset_prepared_path: ` in your config to set the path to save the prepared dataset.
 
-Launch a Gradio interface:
+More details can be found in [Dataset Preprocessing](dataset_preprocessing.qmd).
+
+### Merging LoRA weights {#sec-merging-lora}
+
+To merge the LoRA weights back into the base model, run:
 
 ```bash
-axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out" --gradio
+axolotl merge-lora my_training.yml --lora-model-dir="./outputs/lora-out"
 ```
 
+The merged model will be saved in the `{output_dir}/merged` directory.
+
+More details can be found in [Merging LoRA weights](inference.qmd#sec-merging).
+
 ## Next Steps {#sec-next-steps}
 
 Now that you have the basics, you might want to:
@@ -156,6 +180,7 @@ Now that you have the basics, you might want to:
 Check our other guides for details on these topics:
 
 - [Configuration Guide](config.qmd) - Full configuration options
+- [Dataset Loading](dataset-loading.qmd) - Loading datasets from various sources
 - [Dataset Formats](dataset-formats) - Working with different data formats
 - [Multi-GPU Training](multi-gpu.qmd)
 - [Multi-Node Training](multi-node.qmd)

docs/installation.qmd

@@ -25,6 +25,10 @@ Please make sure to have Pytorch installed before installing Axolotl in your loc
 Follow the instructions at: [https://pytorch.org/get-started/locally/](https://pytorch.org/get-started/locally/)
 :::
 
+::: {.callout-important}
+For Blackwell GPUs, please use Pytorch 2.7.0 and CUDA 12.8.
+:::
+
 ### PyPI Installation (Recommended) {#sec-pypi}
 
 ```{.bash}
@@ -72,6 +76,10 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it \
 ```
 :::
 
+::: {.callout-important}
+For Blackwell GPUs, please use `axolotlai/axolotl:main-py3.11-cu128-2.7.0` or the cloud variant `axolotlai/axolotl-cloud:main-py3.11-cu128-2.7.0`.
+:::
+
 Please refer to the [Docker documentation](docker.qmd) for more information on the different Docker images that are available.
 
 ## Cloud Environments {#sec-cloud}

docs/multi-gpu.qmd

@@ -87,20 +87,7 @@ We support sequence parallelism (SP) via the
 allows one to split up sequences across GPUs, which is useful in the event that a
 single sequence causes OOM errors during model training.
 
-First, install `ring-flash-attn`, recommended via `pip install axolotl[ring-flash-attn]`,
-or from source with `pip install .[ring-flash-attn]`.
-
-Your Axolotl YAML config should contain the following lines:
-
-```{.yaml}
-sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
-flash_attention: true  # Required with sequence parallelism
-
-# Optional; strides across the key dimension. Larger values use more memory but will make training faster.
-heads_k_stride: 1
-```
-
-See our [dedicated guide](sequence_parallelism.qmd) for more details.
+See our [dedicated guide](sequence_parallelism.qmd) for more information.
 
 ### FSDP + QLoRA {#sec-fsdp-qlora}
 

docs/sequence_parallelism.qmd

@@ -3,8 +3,6 @@ title: Sequence Parallelism
 description: Train with long sequences split across multiple GPUs.
 ---
 
-# Sequence Parallelism
-
 Sequence parallelism is a technique that splits sequences across multiple GPUs,
 allowing you to train with very long sequences that wouldn't fit on a single GPU. Each
 GPU processes a different portion of the sequence, and the results are aggregated
@@ -27,7 +25,7 @@ To enable sequence parallelism, add the following to your configuration file:
 sequence_parallel_degree: 4  # Split sequences across 4 GPUs
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
-# Optional; one of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to
+# Optional; one of "varlen_llama3" or "batch_ring". Defaults to
 # "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
 ring_attn_func:
 ```
@@ -43,7 +41,7 @@ When sequence parallelism is enabled:
 
 1. Each sequence is divided into equal chunks across the GPUs in a sequence parallel group
 2. The data collator handles the chunking of input_ids, attention_mask, labels, and position_ids
-3. Position IDs are adjusted to maintain proper relative positions, especially for packed sequences
+3. Position IDs are adjusted to maintain proper relative positions
 4. The trainer uses special ring communication patterns for attention operations
 
 ## Requirements
@@ -69,9 +67,11 @@ sequence_len: 8192
 ...
 
 sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
-flash_attention: true  # Required with sequence parallelism
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
+# Optional; one of "varlen_llama3" or "batch_ring". Defaults to
+# "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
+ring_attn_func:
 
 ...
 ```

examples/llama-4/README.md

@@ -34,3 +34,5 @@ We provide a script to delinearize Llama 4 linearized models into regular Huggin
 ```bash
 axolotl delinearize-llama4 --model path/to/model_dir --output path/to/output_dir
 ```
+
+Note: This only works with the non-quantized linearized model. If you have an adapter, merge it with the *non-quantized linearized* model before delinearizing.

examples/mistral/devstral-small-2505.yml

@@ -0,0 +1,48 @@
+base_model: mistralai/Devstral-Small-2505
+processor_type: AutoProcessor
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+chat_template: mistral_v7_tekken
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+    field_messages: messages
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+sequence_len: 2048
+pad_to_sequence_len: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: false
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/orpheus/README.md

@@ -0,0 +1,341 @@
+# Finetuning LLMs to output audio
+
+In this example, we finetune Orpcanopylabs/orpheus-tts-0.1-pretrained (a LLaMA 3.2 3b model) to output audio.
+
+The `finetune.yml` withe current settings will run on any Nvidia GPU with 45GB VRAM or more. If you adjust the batch size it can easily run on any GPU under 24GB.
+
+## Dataset pre-processing for pre-training
+If you are adding another voice in English, please jump ahead to finetuning pre-processing.
+
+For this to work, we need to preprocess our dataset. Since we are expecting to output audio, we will need to add tokens to the tokenizer.
+
+Using this code, it will download the SNAC model and add the correct tokens and upload the final dataset.
+
+```python
+import torch
+from snac import SNAC
+from datasets import load_dataset
+from huggingface_hub import snapshot_download
+from datasets import load_dataset
+import random
+import torchaudio.transforms as T
+from transformers import AutoTokenizer
+import os
+
+my_original_dataset_name = "<huggingface-id-of-dataset-that-we-want-to-preprocess>"
+name_to_push_dataset_to = "<huggingface-id-of-where-to-save-dataset>"
+
+dsn = my_original_dataset_name
+
+snapshot_download(
+    repo_id=dsn,
+    repo_type="dataset",
+    revision="main",
+    max_workers=64,
+)
+
+
+ds = load_dataset(dsn, split="train")
+ds_sample_rate = ds[0]["audio"]["sampling_rate"]
+
+model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
+model = model.to("mps")
+
+def tokenise_audio(waveform):
+  waveform = torch.from_numpy(waveform).unsqueeze(0)
+  waveform = waveform.to(dtype=torch.float32)
+  resample_transform = T.Resample(orig_freq=ds_sample_rate, new_freq=24000)
+  waveform = resample_transform(waveform)
+
+  waveform = waveform.unsqueeze(0).to("cuda")
+
+  #generate the codes from snac
+  with torch.inference_mode():
+    codes = model.encode(waveform)
+
+  all_codes = []
+  for i in range(codes[0].shape[1]):
+    all_codes.append(codes[0][0][i].item()+128266)
+    all_codes.append(codes[1][0][2*i].item()+128266+4096)
+    all_codes.append(codes[2][0][4*i].item()+128266+(2*4096))
+    all_codes.append(codes[2][0][(4*i)+1].item()+128266+(3*4096))
+    all_codes.append(codes[1][0][(2*i)+1].item()+128266+(4*4096))
+    all_codes.append(codes[2][0][(4*i)+2].item()+128266+(5*4096))
+    all_codes.append(codes[2][0][(4*i)+3].item()+128266+(6*4096))
+
+
+  return all_codes
+
+def add_codes(example):
+    # Always initialize codes_list to None
+    codes_list = None
+
+    try:
+        answer_audio = example.get("audio")
+        # If there's a valid audio array, tokenise it
+        if answer_audio and "array" in answer_audio:
+            audio_array = answer_audio["array"]
+            codes_list = tokenise_audio(audio_array)
+    except Exception as e:
+        print(f"Skipping row due to error: {e}")
+        # Keep codes_list as None if we fail
+    example["codes_list"] = codes_list
+
+    return example
+
+ds = ds.map(add_codes, remove_columns=["audio"])
+
+#@title Load Tokenizer
+tokeniser_length = 128256
+start_of_text = 128000
+end_of_text = 128009
+
+start_of_speech = tokeniser_length + 1
+end_of_speech = tokeniser_length + 2
+
+start_of_human = tokeniser_length + 3
+end_of_human = tokeniser_length + 4
+
+start_of_ai = tokeniser_length + 5
+end_of_ai =  tokeniser_length + 6
+pad_token = tokeniser_length + 7
+
+audio_tokens_start = tokeniser_length + 10
+
+tokenizer_name = "canopylabs/orpheus-3b-0.1-pretrained"
+
+
+tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+num_proc = os.cpu_count() - 2
+
+ds = ds.filter(lambda x: x["codes_list"] is not None)
+ds = ds.filter(lambda x: len(x["codes_list"]) > 0)
+
+#@title Create Input Ids
+def remove_duplicate_frames(example):
+    vals = example["codes_list"]
+    if len(vals) % 7 != 0:
+        raise ValueError("Input list length must be divisible by 7")
+
+    result = vals[:7]
+
+    removed_frames = 0
+
+    for i in range(7, len(vals), 7):
+        current_first = vals[i]
+        previous_first = result[-7]
+
+        if current_first != previous_first:
+            result.extend(vals[i:i+7])
+        else:
+            removed_frames += 1
+
+    example["codes_list"] = result
+
+    return example
+
+ds = ds.map(remove_duplicate_frames, num_proc=num_proc)
+
+
+def create_input_ids(example):
+    text_ids = tokenizer.encode({example['text']},  add_special_tokens=True)
+    text_ids.append(end_of_text)
+    example["text_tokens"] = text_ids
+    input_ids = (
+        [start_of_human]
+        + example["text_tokens"]
+        + [end_of_human]
+        + [start_of_ai]
+        + [start_of_speech]
+        + example["codes_list"]
+        + [end_of_speech]
+        + [end_of_ai]
+    )
+    example["input_ids"] = input_ids
+    example["labels"] = input_ids
+    example["attention_mask"] = [1] * len(input_ids)
+
+    return example
+
+ds = ds.map(create_input_ids, num_proc=num_proc, remove_columns=["text", "codes_list"])
+
+#@title Remove unnecessary columns
+columns_to_keep = ["input_ids", "labels", "attention_mask"]
+columns_to_remove = [col for col in ds.column_names if col not in columns_to_keep]
+
+ds = ds.remove_columns(columns_to_remove)
+
+ds.push_to_hub(name_to_push_dataset_to)
+```
+
+
+## Finetune pre-processing
+Use this code to add a new voice.
+
+```python
+import torch
+from snac import SNAC
+from datasets import load_dataset
+from huggingface_hub import snapshot_download
+from datasets import load_dataset
+import random
+import torchaudio.transforms as T
+from transformers import AutoTokenizer
+import os
+
+my_original_dataset_name = "<huggingface-id-of-dataset-that-we-want-to-preprocess>"
+name_to_push_dataset_to = "<huggingface-id-of-where-to-save-dataset>"
+
+dsn = my_original_dataset_name
+
+snapshot_download(
+    repo_id=dsn,
+    repo_type="dataset",
+    revision="main",
+    max_workers=64,
+)
+
+
+ds = load_dataset(dsn, split="train")
+ds_sample_rate = ds[0]["audio"]["sampling_rate"]
+
+model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
+model = model.to("mps")
+
+def tokenise_audio(waveform):
+  waveform = torch.from_numpy(waveform).unsqueeze(0)
+  waveform = waveform.to(dtype=torch.float32)
+  resample_transform = T.Resample(orig_freq=ds_sample_rate, new_freq=24000)
+  waveform = resample_transform(waveform)
+
+  waveform = waveform.unsqueeze(0).to("cuda")
+
+  #generate the codes from snac
+  with torch.inference_mode():
+    codes = model.encode(waveform)
+
+  all_codes = []
+  for i in range(codes[0].shape[1]):
+    all_codes.append(codes[0][0][i].item()+128266)
+    all_codes.append(codes[1][0][2*i].item()+128266+4096)
+    all_codes.append(codes[2][0][4*i].item()+128266+(2*4096))
+    all_codes.append(codes[2][0][(4*i)+1].item()+128266+(3*4096))
+    all_codes.append(codes[1][0][(2*i)+1].item()+128266+(4*4096))
+    all_codes.append(codes[2][0][(4*i)+2].item()+128266+(5*4096))
+    all_codes.append(codes[2][0][(4*i)+3].item()+128266+(6*4096))
+
+
+  return all_codes
+
+def add_codes(example):
+    # Always initialize codes_list to None
+    codes_list = None
+
+    try:
+        answer_audio = example.get("audio")
+        # If there's a valid audio array, tokenise it
+        if answer_audio and "array" in answer_audio:
+            audio_array = answer_audio["array"]
+            codes_list = tokenise_audio(audio_array)
+    except Exception as e:
+        print(f"Skipping row due to error: {e}")
+        # Keep codes_list as None if we fail
+    example["codes_list"] = codes_list
+
+    return example
+
+ds = ds.map(add_codes, remove_columns=["audio"])
+
+#@title Load Tokenizer
+tokeniser_length = 128256
+start_of_text = 128000
+end_of_text = 128009
+
+start_of_speech = tokeniser_length + 1
+end_of_speech = tokeniser_length + 2
+
+start_of_human = tokeniser_length + 3
+end_of_human = tokeniser_length + 4
+
+start_of_ai = tokeniser_length + 5
+end_of_ai =  tokeniser_length + 6
+pad_token = tokeniser_length + 7
+
+audio_tokens_start = tokeniser_length + 10
+
+tokenizer_name = "canopylabs/orpheus-3b-0.1-pretrained"
+
+
+tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+num_proc = os.cpu_count() - 2
+
+ds = ds.filter(lambda x: x["codes_list"] is not None)
+ds = ds.filter(lambda x: len(x["codes_list"]) > 0)
+
+#@title Create Input Ids
+def remove_duplicate_frames(example):
+    vals = example["codes_list"]
+    if len(vals) % 7 != 0:
+        raise ValueError("Input list length must be divisible by 7")
+
+    result = vals[:7]
+
+    removed_frames = 0
+
+    for i in range(7, len(vals), 7):
+        current_first = vals[i]
+        previous_first = result[-7]
+
+        if current_first != previous_first:
+            result.extend(vals[i:i+7])
+        else:
+            removed_frames += 1
+
+    example["codes_list"] = result
+
+    return example
+
+ds = ds.map(remove_duplicate_frames, num_proc=num_proc)
+
+tok_info = '''*** HERE you can modify the text prompt
+i.e. if you wanted a multispeaker model like canopylabs/orpheus-3b-0.1-ft, you can pass:
+f"{example["source"]}:  {example["text"]}", as is passed.
+'''
+print(tok_info)
+
+def create_input_ids(example):
+    text_ids = tokenizer.encode(f"{example['speaker_id']}: {example['text']}",  add_special_tokens=True)
+    text_ids.append(end_of_text)
+    example["text_tokens"] = text_ids
+    input_ids = (
+        [start_of_human]
+        + example["text_tokens"]
+        + [end_of_human]
+        + [start_of_ai]
+        + [start_of_speech]
+        + example["codes_list"]
+        + [end_of_speech]
+        + [end_of_ai]
+    )
+    example["input_ids"] = input_ids
+    example["labels"] = input_ids
+    example["attention_mask"] = [1] * len(input_ids)
+
+    return example
+
+ds = ds.map(create_input_ids, num_proc=num_proc, remove_columns=["text", "codes_list"])
+
+#@title Remove unnecessary columns
+columns_to_keep = ["input_ids", "labels", "attention_mask"]
+columns_to_remove = [col for col in ds.column_names if col not in columns_to_keep]
+
+ds = ds.remove_columns(columns_to_remove)
+
+ds.push_to_hub(name_to_push_dataset_to)
+```
+
+## Training
+After preprocessing is done, fill out the blanks in finetune.yml and simply run `axolotl train finetune.yml`
+
+## Inference
+For inference, please refer to the original [orpheus github](https://github.com/canopyai/Orpheus-TTS/tree/main).

examples/orpheus/finetune.yml

@@ -0,0 +1,52 @@
+base_model: canopylabs/orpheus-3b-0.1-pretrained
+
+hub_model_id: <your-hub-model-id>
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_fused_linear_cross_entropy: true
+
+datasets:
+  - path: <your-hf-dataset-id>
+    type:  # leave empty to load pre-tokenized
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 8
+micro_batch_size: 4
+num_epochs: 3
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 20
+evals_per_epoch: 5
+saves_per_epoch: 5
+weight_decay: 0.05
+
+special_tokens:
+  pad_token: <custom_token_7>

examples/qwen2/dpo.yaml

@@ -2,7 +2,6 @@ base_model: Qwen/Qwen2.5-0.5B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-
 chat_template: qwen_25
 rl: dpo
 datasets:

requirements.txt

@@ -6,16 +6,17 @@ triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.8
+liger-kernel==0.5.9
 # END section
 
 packaging==23.2
 
+huggingface_hub==0.31.0
 peft==0.15.2
 transformers==4.51.3
 tokenizers>=0.21.1
 accelerate==1.6.0
-datasets==3.5.0
+datasets==3.5.1
 deepspeed>=0.15.4
 trl==0.17.0
 hf_xet==1.1.0

setup.py

@@ -67,13 +67,13 @@ def parse_requirements(extras_require_map):
             if (major, minor) >= (2, 7):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 # _install_requires.append("xformers==0.0.29.post3")  # xformers seems to be hard pinned to 2.6.0
-                extras_require_map["vllm"] = ["vllm==0.8.5"]
+                extras_require_map["vllm"] = ["vllm==0.8.5.post1"]
             elif (major, minor) >= (2, 6):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 _install_requires.append(
                     "xformers==0.0.29.post2"
                 )  # vllm needs post2 w torch 2.6
-                extras_require_map["vllm"] = ["vllm==0.8.5"]
+                extras_require_map["vllm"] = ["vllm==0.8.5.post1"]
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
@@ -142,6 +142,7 @@ extras_require = {
         "apollo-torch",
         "lomo-optim==0.1.1",
         "torch-optimi==0.2.1",
+        "came_pytorch==0.1.3",
     ],
     "ray": [
         "ray[train]",

src/axolotl/cli/args.py

@@ -82,6 +82,12 @@ class VllmServeCliArgs:
             "hardware support this feature."
         },
     )
+    serve_module: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "Module to serve. If not set, the default module will be used."
+        },
+    )
 
 
 @dataclass

src/axolotl/cli/preprocess.py

@@ -18,6 +18,7 @@ from axolotl.cli.checks import check_accelerate_default_config, check_user_token
 from axolotl.cli.config import load_cfg
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.common.datasets import load_datasets, load_preference_datasets
+from axolotl.integrations.base import PluginManager
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.trainer import disable_datasets_caching
 
@@ -47,7 +48,10 @@ def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:
         cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
     with disable_datasets_caching():
-        if cfg.rl:
+        plugin_manager = PluginManager.get_instance()
+        if plugin_manager.load_datasets(cfg, preprocess=True):
+            pass
+        elif cfg.rl:
             load_preference_datasets(cfg=cfg, cli_args=cli_args)
         else:
             load_datasets(cfg=cfg, cli_args=cli_args)

src/axolotl/cli/train.py

@@ -43,10 +43,13 @@ def do_train(cfg: DictDefault, cli_args: TrainerCliArgs):
     if int(os.getenv("LOCAL_RANK", "0")) == 0:
         check_user_token()
 
-    if cfg.rl:
-        dataset_meta = load_preference_datasets(cfg=cfg, cli_args=cli_args)
-    else:
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+    plugin_manager = PluginManager.get_instance()
+    dataset_meta = plugin_manager.load_datasets(cfg, preprocess=False)
+    if not dataset_meta:
+        if cfg.rl:
+            dataset_meta = load_preference_datasets(cfg=cfg, cli_args=cli_args)
+        else:
+            dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
     model, tokenizer, trainer = train(cfg=cfg, dataset_meta=dataset_meta)
 

src/axolotl/cli/utils.py

@@ -20,8 +20,9 @@ from transformers import (
     ProcessorMixin,
 )
 
+from axolotl.loaders import load_processor, load_tokenizer
+from axolotl.loaders.model import ModelLoader
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import load_model, load_processor, load_tokenizer
 
 LOG = logging.getLogger(__name__)
 
@@ -318,7 +319,8 @@ def load_model_and_tokenizer(
     tokenizer = load_tokenizer(cfg)
 
     LOG.info("loading model...")
-    model, _ = load_model(cfg, tokenizer, inference=inference)
+    model_loader = ModelLoader(cfg, tokenizer, inference=inference)
+    model, _ = model_loader.load()
 
     processor = None
     if cfg.is_multimodal:

src/axolotl/cli/vllm_serve.py

@@ -6,7 +6,6 @@ from pathlib import Path
 from typing import Union
 
 from trl.scripts.vllm_serve import ScriptArguments
-from trl.scripts.vllm_serve import main as vllm_serve_main
 
 from axolotl.cli.config import load_cfg
 
@@ -28,6 +27,9 @@ def do_vllm_serve(
     cfg = load_cfg(config)
     model = cfg.base_model
 
+    serve_module = cli_args.get("serve_module", "trl.scripts.vllm_serve")
+    vllm_serve_main = getattr(__import__(serve_module, fromlist=["main"]), "main")
+
     tensor_parallel_size = (
         cli_args.get("tensor_parallel_size") or cfg.vllm.tensor_parallel_size
     )

src/axolotl/common/datasets.py

@@ -10,10 +10,11 @@ from datasets import Dataset
 
 import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
 from axolotl.cli.args import PreprocessCliArgs, TrainerCliArgs
+from axolotl.loaders import load_processor, load_tokenizer
 from axolotl.utils.data import prepare_dataset
 from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import load_processor, load_tokenizer
+from axolotl.utils.schemas.enums import RLType
 from axolotl.utils.tokenization import check_dataset_labels
 
 LOG = logging.getLogger(__name__)
@@ -133,7 +134,7 @@ def load_preference_datasets(
     total_num_steps: Optional[int] = int(
         math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
     )
-    if cfg.rl == "grpo":
+    if cfg.rl is RLType.GRPO:
         total_num_steps = None
 
     if cli_args.debug or cfg.debug:

src/axolotl/core/trainer_builder.py

@@ -59,6 +59,7 @@ from axolotl.core.training_args import (
     AxolotlTrainingArguments,
 )
 from axolotl.integrations.base import PluginManager
+from axolotl.loaders.utils import ensure_dtype
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback
 from axolotl.monkeypatch.trainer.lr import patch_trainer_get_lr
@@ -86,8 +87,7 @@ from axolotl.utils.collators import (
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
 from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
-from axolotl.utils.models import ensure_dtype
-from axolotl.utils.schemas.enums import CustomSupportedOptimizers
+from axolotl.utils.schemas.enums import CustomSupportedOptimizers, RLType
 
 try:
     import torch._dynamo  # pylint: disable=ungrouped-imports
@@ -170,6 +170,9 @@ class TrainerBuilderBase(abc.ABC):
                 )
             )
 
+        if self.cfg.gc_steps:
+            callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
+
         if self.cfg.use_wandb:
             callbacks.append(
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
@@ -251,9 +254,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.loss_watchdog_threshold is not None:
             callbacks.append(LossWatchDogCallback(self.cfg))
 
-        if self.cfg.gc_steps:
-            callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
-
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
@@ -353,7 +353,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["warmup_steps"] = warmup_steps
         training_arguments_kwargs["logging_steps"] = logging_steps
 
-        if self.cfg.seed:
+        if self.cfg.seed is not None:
             training_arguments_kwargs["seed"] = self.cfg.seed
 
         if self.cfg.gradient_checkpointing:
@@ -387,8 +387,12 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["adam_beta1"] = self.cfg.adam_beta1
         if self.cfg.adam_beta2:
             training_arguments_kwargs["adam_beta2"] = self.cfg.adam_beta2
+        if self.cfg.adam_beta3:
+            training_arguments_kwargs["adam_beta3"] = self.cfg.adam_beta3
         if self.cfg.adam_epsilon:
             training_arguments_kwargs["adam_epsilon"] = self.cfg.adam_epsilon
+        if self.cfg.adam_epsilon2:
+            training_arguments_kwargs["adam_epsilon2"] = self.cfg.adam_epsilon2
         if self.cfg.max_grad_norm:
             training_arguments_kwargs["max_grad_norm"] = self.cfg.max_grad_norm
 
@@ -547,8 +551,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         report_to = []
         if self.cfg.use_wandb:
             report_to.append("wandb")
-            if self.cfg.wandb_name:
-                training_arguments_kwargs["run_name"] = self.cfg.wandb_name
         if self.cfg.use_mlflow:
             report_to.append("mlflow")
         if self.cfg.use_tensorboard:
@@ -708,6 +710,20 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 optimizer_cls = ADOPT
                 adam_kwargs["decouple"] = True
                 optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "came_pytorch":
+                from came_pytorch import CAME
+
+                optimizer_cls = CAME
+
+                beta1 = training_arguments_kwargs.get("adam_beta1", 0.9)
+                beta2 = training_arguments_kwargs.get("adam_beta2", 0.999)
+                beta3 = training_arguments_kwargs.get("adam_beta3", 0.9999)
+                eps1 = training_arguments_kwargs.get("adam_epsilon", 1e-30)
+                eps2 = training_arguments_kwargs.get("adam_epsilon2", 1e-16)
+                adam_kwargs["betas"] = (beta1, beta2, beta3)
+                adam_kwargs["eps"] = (eps1, eps2)
+
+                optimizer_kwargs.update(adam_kwargs)
 
             # Parse any additional optimizer args from config
             if self.cfg.optim_args:
@@ -782,11 +798,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 self.cfg.kd_top_k_before_softmax
             )
 
-        training_arguments_kwargs["sequence_parallel_degree"] = (
-            self.cfg.sequence_parallel_degree
-        )
-        training_arguments_kwargs["ring_attn_func"] = self.cfg.ring_attn_func
-
         if self.cfg.reward_model:
             training_args_cls = AxolotlRewardConfig
         elif self.cfg.process_reward_model:
@@ -807,14 +818,15 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         data_collator_kwargs = {
             "padding": True,  # True/"longest" is the default
         }
+        multiple = 64
         if self.cfg.pad_to_sequence_len:
-            data_collator_kwargs["pad_to_multiple_of"] = 64 * math.ceil(
-                self.cfg.sequence_len / 64
+            data_collator_kwargs["pad_to_multiple_of"] = multiple * math.ceil(
+                self.cfg.sequence_len / multiple
             )
         else:
             # A100 is best at 64, while others at 8. Let's use the larger so we don't have to check
             # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
-            data_collator_kwargs["pad_to_multiple_of"] = 64
+            data_collator_kwargs["pad_to_multiple_of"] = multiple
 
         if self.cfg.reward_model:
             data_collator_kwargs["max_length"] = self.cfg.sequence_len
@@ -1020,6 +1032,10 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             training_args_kwargs["dataloader_prefetch_factor"] = (
                 self.cfg.dataloader_prefetch_factor
             )
+
+        if self.cfg.seed is not None:
+            training_args_kwargs["seed"] = self.cfg.seed
+
         if self.cfg.gradient_checkpointing:
             training_args_kwargs["gradient_checkpointing"] = (
                 self.cfg.gradient_checkpointing
@@ -1043,6 +1059,8 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             # default to saving each epoch if not defined
             training_args_kwargs["save_strategy"] = "epoch"
 
+        training_args_kwargs["save_only_model"] = self.cfg.save_only_model
+
         if self.cfg.dataset_processes:
             training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
 
@@ -1062,7 +1080,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
 
         training_args_cls = None
         blocklist_args_kwargs = []
-        if self.cfg.rl == "simpo":
+        if self.cfg.rl is RLType.SIMPO:
             training_args_cls = AxolotlCPOConfig
             training_args_kwargs["loss_type"] = "simpo"
             training_args_kwargs["max_length"] = self.cfg.sequence_len
@@ -1070,13 +1088,13 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.cpo_alpha is not None:
                 training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
 
-        elif self.cfg.rl == "orpo":
+        elif self.cfg.rl is RLType.ORPO:
             training_args_cls = AxolotlORPOConfig
             training_args_kwargs["max_length"] = self.cfg.sequence_len
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
-        elif self.cfg.rl == "kto":
+        elif self.cfg.rl is RLType.KTO:
             training_args_cls = AxolotlKTOConfig
 
             training_args_kwargs["desirable_weight"] = (
@@ -1090,14 +1108,14 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
-        elif self.cfg.rl == "grpo":
+        elif self.cfg.rl is RLType.GRPO:
             training_args_cls = GRPOStrategy.get_training_args_class()
             training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
             blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
 
         else:
             training_args_cls = AxolotlDPOConfig
-            if self.cfg.rl == "ipo":
+            if self.cfg.rl is RLType.IPO:
                 training_args_kwargs["loss_type"] = "ipo"
             training_args_kwargs["max_length"] = self.cfg.sequence_len
             training_args_kwargs["max_completion_length"] = None
@@ -1140,67 +1158,76 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
 
     def build(self, total_num_steps):
         training_args = self.build_training_arguments(total_num_steps)
-        dpo_trainer_kwargs = {}
-        if self.cfg.rl == "ipo":
+        trainer_kwargs = {}
+        if self.cfg.rl is RLType.IPO:
             if self.cfg.dpo_label_smoothing:
-                dpo_trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
+                trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
         if self.eval_dataset:
-            dpo_trainer_kwargs["eval_dataset"] = self.eval_dataset
+            trainer_kwargs["eval_dataset"] = self.eval_dataset
         if self.cfg.adapter and self.peft_config:
-            dpo_trainer_kwargs["peft_config"] = self.peft_config
+            if self.cfg.rl is not RLType.GRPO:
+                trainer_kwargs["peft_config"] = self.peft_config
         if self.cfg.precompute_ref_log_probs is not None:
-            dpo_trainer_kwargs["precompute_ref_log_probs"] = (
+            trainer_kwargs["precompute_ref_log_probs"] = (
                 self.cfg.precompute_ref_log_probs
             )
-        if self.cfg.rl == "grpo":
-            trainer_cls = GRPOStrategy.get_trainer_class()
+        if self.cfg.rl is RLType.GRPO:
+            trainer_cls = GRPOStrategy.get_trainer_class(
+                sequence_parallel=self.cfg.sequence_parallel_degree > 1
+            )
             trainer_cls_args = [self.model]
             trainer_cls_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
-            dpo_trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
-        elif self.cfg.rl in ["dpo", "ipo"]:
+            trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
+        elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
             trainer_cls = DPOStrategy.get_trainer_class()
             trainer_cls_args = [self.model, self.model_ref]
-        elif self.cfg.rl == "orpo":
+        elif self.cfg.rl is RLType.ORPO:
             trainer_cls = AxolotlORPOTrainer
             trainer_cls_args = [self.model]
-        elif self.cfg.rl in ["kto"]:
+        elif self.cfg.rl is RLType.KTO:
             trainer_cls = AxolotlKTOTrainer
             trainer_cls_args = [self.model]
-        elif self.cfg.rl in ["simpo"]:
+        elif self.cfg.rl is RLType.SIMPO:
             trainer_cls = AxolotlCPOTrainer
             trainer_cls_args = [self.model]
         else:
             raise ValueError(f"Unsupported RL: {self.cfg.rl}")
 
+        if self.cfg.plugins:
+            plugin_manager = PluginManager.get_instance()
+            temp_trainer_cls = plugin_manager.get_trainer_cls(self.cfg)
+            if temp_trainer_cls is not None:
+                trainer_cls = temp_trainer_cls
+
         sig = inspect.signature(trainer_cls)
         if "tokenizer" in sig.parameters.keys():
-            dpo_trainer_kwargs["tokenizer"] = self.tokenizer
+            trainer_kwargs["tokenizer"] = self.tokenizer
         else:
-            dpo_trainer_kwargs["processing_class"] = self.tokenizer
+            trainer_kwargs["processing_class"] = self.tokenizer
 
         if self.cfg.datasets is not None and (
             trainer_cls is DPOStrategy.get_trainer_class()
         ):
-            dpo_trainer_kwargs["dataset_tags"] = [
+            trainer_kwargs["dataset_tags"] = [
                 d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
             ]
-        dpo_trainer = trainer_cls(
+        trainer = trainer_cls(
             *trainer_cls_args,
             args=training_args,
             train_dataset=self.train_dataset,
             callbacks=self.get_callbacks(),
-            **dpo_trainer_kwargs,
+            **trainer_kwargs,
         )
         if self.cfg.fsdp:
-            ensure_dtype(dpo_trainer.model, dtype=self.cfg.torch_dtype)
-            if self.cfg.rl in ["dpo", "ipo"] and dpo_trainer.ref_model:
-                ensure_dtype(dpo_trainer.ref_model, dtype=self.cfg.torch_dtype)
+            ensure_dtype(trainer.model, dtype=self.cfg.torch_dtype)
+            if self.cfg.rl in [RLType.DPO, RLType.IPO] and trainer.ref_model:
+                ensure_dtype(trainer.ref_model, dtype=self.cfg.torch_dtype)
 
-        dpo_trainer = self.hook_post_create_trainer(dpo_trainer)
-        for callback in self.get_post_trainer_create_callbacks(dpo_trainer):
-            dpo_trainer.add_callback(callback)
+        trainer = self.hook_post_create_trainer(trainer)
+        for callback in self.get_post_trainer_create_callbacks(trainer):
+            trainer.add_callback(callback)
 
-        return dpo_trainer
+        return trainer
 
 
 class HFPPOTrainerBuilder(TrainerBuilderBase):

src/axolotl/core/trainers/__init__.py

@@ -5,7 +5,7 @@
 
 from .base import AxolotlTrainer
 from .dpo.trainer import AxolotlDPOTrainer
-from .grpo.trainer import AxolotlGRPOTrainer
+from .grpo.trainer import AxolotlGRPOSequenceParallelTrainer, AxolotlGRPOTrainer
 from .mamba import AxolotlMambaTrainer
 from .relora import ReLoRATrainer
 from .trl import (

src/axolotl/core/trainers/base.py

@@ -29,7 +29,6 @@ from axolotl.core.trainers.mixins import (
     OptimizerMixin,
     RngLoaderMixin,
     SchedulerMixin,
-    SequenceParallelMixin,
 )
 from axolotl.core.trainers.utils import (
     sanitize_kwargs_for_ds_tagging,
@@ -40,9 +39,7 @@ from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 LOG = logging.getLogger(__name__)
 
 
-class AxolotlTrainer(
-    SchedulerMixin, OptimizerMixin, RngLoaderMixin, SequenceParallelMixin, Trainer
-):
+class AxolotlTrainer(SchedulerMixin, OptimizerMixin, RngLoaderMixin, Trainer):
     """Extend the base Trainer for axolotl helpers"""
 
     args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
@@ -68,10 +65,6 @@ class AxolotlTrainer(
         if self.args.orpo_alpha:
             self.loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
 
-        # Initialize sequence parallelism if enabled
-        if self.args.sequence_parallel_degree > 1:
-            self._setup_sequence_parallel()
-
     def _wrap_model(self, model, training=True, dataloader=None):
         if self.args.torch_compile:
             torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
@@ -122,8 +115,8 @@ class AxolotlTrainer(
 
     def _get_train_sampler(self) -> Sampler | None:
         """
-        Helper method to get the sampler for training. Handles cases for sequence
-        parallelism, sample packing, and curriculum sampling (sequential).
+        Helper method to get the sampler for training. Handles cases for sample packing
+        and curriculum sampling (sequential).
 
         Returns:
             If the dataset is non-empty, a sampler is returned, the type of which
@@ -132,9 +125,7 @@ class AxolotlTrainer(
         use_sample_packing = self.args.sample_packing and not self.args.pretraining
 
         # Determine the base sampler first
-        if self.args.sequence_parallel_degree > 1:
-            base_sampler = self._sp_get_train_sampler(self.train_dataset)
-        elif self.args.curriculum_sampling:
+        if self.args.curriculum_sampling:
             base_sampler = SequentialSampler(self.train_dataset)
         elif use_sample_packing:
             base_sampler = RandomSampler(self.train_dataset)
@@ -153,8 +144,7 @@ class AxolotlTrainer(
 
     def _get_eval_sampler(self, eval_dataset: Dataset | None = None) -> Sampler | None:
         """
-        Helper method to get the sampler for evaluation. Handles sequence parallelism
-        and sample packing cases.
+        Helper method to get the sampler for evaluation. Handles sample packing case.
 
         Returns:
             If the dataset is non-empty, a sampler is returned, the type of which
@@ -168,9 +158,7 @@ class AxolotlTrainer(
         )
 
         # Determine the base sampler
-        if self.args.sequence_parallel_degree > 1:
-            base_sampler = self._sp_get_eval_sampler(eval_dataset)
-        elif use_multipack:
+        if use_multipack:
             base_sampler = SequentialSampler(eval_dataset)
         else:
             return super()._get_eval_sampler(eval_dataset)
@@ -236,14 +224,6 @@ class AxolotlTrainer(
         ):
             self.accelerator.even_batches = False
 
-        # Return unprepared dataloader if using sequence parallelism
-        # TODO(djsaunde): We might be able to use `accelerate`'s dataloader preparation
-        # if we use `dispatch_batches` and `slice_fn_for_dispatch` properly (i.e.,
-        # slice each batch along the sequence dimension).
-        if self.args.sequence_parallel_degree > 1:
-            return dataloader
-
-        # Otherwise prepare with accelerator
         return self.accelerator.prepare_data_loader(dataloader)
 
     def get_train_dataloader(self) -> DataLoader:
@@ -287,12 +267,7 @@ class AxolotlTrainer(
 
             return dataloader
 
-        # Handle sample packing or sequence parallelism
-        if (
-            self.args.sample_packing
-            and self.args.eval_sample_packing is not False
-            or self.args.sequence_parallel_degree > 1
-        ):
+        if self.args.sample_packing and self.args.eval_sample_packing is not False:
             # Get appropriate data collator
             self.data_collator = (  # pylint: disable=attribute-defined-outside-init
                 self.eval_data_collator
@@ -302,17 +277,6 @@ class AxolotlTrainer(
             if "length" in eval_dataset.column_names:
                 eval_dataset = eval_dataset.remove_columns(["length"])
 
-            # Handle dataset preprocessing for SP
-            if self.args.sequence_parallel_degree > 1:
-                if isinstance(eval_dataset, datasets.Dataset):
-                    eval_dataset = self._remove_unused_columns(
-                        eval_dataset, description="evaluation"
-                    )
-                else:
-                    self.data_collator = self._get_collator_with_removed_columns(  # pylint: disable=attribute-defined-outside-init
-                        self.data_collator, description="evaluation"
-                    )
-
             # Use eval_batch_size for sample packing, per_device_eval_batch_size otherwise
             batch_size = (
                 self.args.eval_batch_size
@@ -373,15 +337,13 @@ class AxolotlTrainer(
                 num_items_in_batch=num_items_in_batch,
             )
 
-        loss = super().compute_loss(
+        return super().compute_loss(
             model,
             inputs,
             return_outputs=return_outputs,
             num_items_in_batch=num_items_in_batch,
         )
 
-        return loss
-
     @staticmethod
     def orpo_concatenate_inputs(inputs, label_pad_token=-100, pad_token=0, device=None):
         concatenated_batch = {}

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

@@ -1,14 +1,11 @@
-"""
-DPO Specific Strategy for training
-"""
+"""DPO Specific Strategy for training"""
 
 from axolotl.core.trainers.dpo.trainer import AxolotlDPOTrainer
+from axolotl.utils.schemas.enums import RLType
 
 
 class DPOStrategy:
-    """
-    Strategy for DPO training
-    """
+    """Strategy for DPO training"""
 
     @classmethod
     def get_trainer_class(cls):
@@ -23,7 +20,7 @@ class DPOStrategy:
     @classmethod
     def set_training_args_kwargs(cls, cfg):
         training_args_kwargs = {}
-        if cfg.rl == "ipo":
+        if cfg.rl is RLType.IPO:
             training_args_kwargs["loss_type"] = "ipo"
         training_args_kwargs["max_length"] = cfg.sequence_len
         training_args_kwargs["max_completion_length"] = None

src/axolotl/core/trainers/dpo/trainer.py

@@ -1,31 +1,15 @@
-"""
-DPO trainer for axolotl
-"""
+"""DPO trainer for axolotl"""
 
 import gc
-import random
 from functools import wraps
-from typing import Any, Dict, Optional, Union
+from typing import Any, Dict, Union
 
-import pandas as pd
 import torch
-import wandb
-from accelerate import PartialState
-from datasets import Dataset, IterableDataset
 from peft.optimizers import create_loraplus_optimizer
 from torch import nn
-from torch.utils.data import DataLoader
-from transformers import (
-    BaseImageProcessor,
-    FeatureExtractionMixin,
-    PreTrainedTokenizerBase,
-    ProcessorMixin,
-    Trainer,
-)
-from transformers.trainer_utils import EvalLoopOutput
+from transformers import Trainer
 from transformers.utils import is_sagemaker_mp_enabled
-from trl import DPOConfig, DPOTrainer, maybe_apply_chat_template, maybe_extract_prompt
-from trl.trainer.utils import log_table_to_comet_experiment
+from trl import DPOTrainer
 
 from axolotl.core.trainers.mixins import RngLoaderMixin, SchedulerMixin
 from axolotl.core.trainers.utils import (
@@ -38,9 +22,7 @@ if is_sagemaker_mp_enabled():
 
 
 class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
-    """
-    Extend the base DPOTrainer for axolotl helpers
-    """
+    """Extend the base DPOTrainer for axolotl helpers."""
 
     tag_names = ["axolotl", "dpo"]
 
@@ -85,8 +67,9 @@ class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
     @wraps(DPOTrainer.push_to_hub)
     def push_to_hub(self, *args, **kwargs) -> str:
         """
-        Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
-        model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
+        Overwrite the `push_to_hub` method in order to force-add the tags when pushing
+        the model on the Hub. Please refer to `~transformers.Trainer.push_to_hub`
+        for more details.
         """
         kwargs = sanitize_kwargs_for_ds_tagging(
             dataset_tags=self.dataset_tags, kwargs=kwargs
@@ -95,64 +78,6 @@ class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
 
         return super().push_to_hub(*args, **kwargs)
 
-    # TODO: remove this once https://github.com/huggingface/trl/pull/3377 is in a release
-    def _prepare_dataset(
-        self,
-        dataset: Union[Dataset, IterableDataset],
-        processing_class: Union[
-            PreTrainedTokenizerBase,
-            BaseImageProcessor,
-            FeatureExtractionMixin,
-            ProcessorMixin,
-        ],
-        args: DPOConfig,
-        dataset_name: str,
-    ) -> Union[Dataset, IterableDataset]:
-        # Build the kwargs for the `map` function
-        map_kwargs: Dict[str, Any] = {"writer_batch_size": 10}
-        if isinstance(dataset, Dataset):  # IterableDataset does not support num_proc
-            map_kwargs["num_proc"] = args.dataset_num_proc
-
-        with PartialState().main_process_first():
-            # Extract prompt if needed
-            if isinstance(
-                dataset, Dataset
-            ):  # `IterableDataset.map` does not support `desc`
-                map_kwargs["desc"] = f"Extracting prompt in {dataset_name} dataset"
-            dataset = dataset.map(maybe_extract_prompt, **map_kwargs)
-
-            # Apply the chat template if needed
-            if isinstance(
-                dataset, Dataset
-            ):  # `IterableDataset.map` does not support `desc`
-                map_kwargs["desc"] = f"Applying chat template to {dataset_name} dataset"
-            dataset = dataset.map(
-                maybe_apply_chat_template,
-                fn_kwargs={"tokenizer": processing_class, "tools": args.tools},
-                **map_kwargs,
-            )
-
-            # Tokenize the dataset
-            if isinstance(
-                dataset, Dataset
-            ):  # `IterableDataset.map` does not support `desc`
-                map_kwargs["desc"] = f"Tokenizing {dataset_name} dataset"
-
-            dataset = dataset.map(
-                self.tokenize_row if not self.is_vision_model else self.process_row,
-                remove_columns=["chosen", "rejected"],
-                fn_kwargs={
-                    "processing_class": processing_class,
-                    "max_prompt_length": args.max_prompt_length,
-                    "max_completion_length": args.max_completion_length,
-                    # for enc-dec, we add the special tokens ([bos_token] + prompt + [eos_token]; completion + [eos_token])
-                    "add_special_tokens": False,
-                },
-                **map_kwargs,
-            )
-
-        return dataset
-
     @staticmethod
     def tokenize_row(
         features,
@@ -192,67 +117,3 @@ class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
         gc.collect()
         torch.cuda.empty_cache()
         return loss
-
-    # TODO: remove this once https://github.com/huggingface/trl/pull/3377 is in a release
-    def evaluation_loop(
-        self,
-        dataloader: DataLoader,
-        description: str,
-        prediction_loss_only: Optional[bool] = None,
-        ignore_keys: Optional[list[str]] = None,
-        metric_key_prefix: str = "eval",
-    ) -> EvalLoopOutput:
-        """
-        Overriding built-in evaluation loop to store metrics for each batch.
-        Prediction/evaluation loop, shared by `Trainer.evaluate()` and `Trainer.predict()`.
-
-        Works both with or without labels.
-        """
-
-        # Sample and save to game log if requested (for one batch to save time)
-        if self.generate_during_eval:
-            # Generate random indices within the range of the total number of samples
-            num_samples = len(dataloader.dataset)
-            random_indices = random.sample(
-                range(num_samples), k=self.args.eval_batch_size
-            )
-
-            # Use dataloader.dataset.select to get the random batch without iterating over the DataLoader
-            random_batch_dataset = dataloader.dataset.select(random_indices)
-            random_batch = self.data_collator(random_batch_dataset)
-            random_batch = self._prepare_inputs(random_batch)
-
-            policy_output_decoded, ref_output_decoded = (
-                self.generate_from_model_and_ref(self.model, random_batch)
-            )
-
-            table = pd.DataFrame(
-                columns=["Prompt", "Policy", "Ref Model"],
-                data=[
-                    [prompt, pol[len(prompt) :], ref[len(prompt) :]]
-                    for prompt, pol, ref in zip(
-                        random_batch_dataset["prompt"],
-                        policy_output_decoded,
-                        ref_output_decoded,
-                    )
-                ],
-            )
-            if "wandb" in self.args.report_to and self.accelerator.is_main_process:
-                wandb.log({"game_log": wandb.Table(data=table)})
-
-            if "comet_ml" in self.args.report_to:
-                log_table_to_comet_experiment(
-                    name="game_log.csv",
-                    table=table,
-                )
-
-        # Base evaluation
-        initial_output = super().evaluation_loop(
-            dataloader,
-            description,
-            prediction_loss_only,
-            ignore_keys,
-            metric_key_prefix,
-        )
-
-        return initial_output

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

@@ -1,37 +1,41 @@
-"""
-GRPO Specific Strategy for training
-"""
+"""GRPO Specific Strategy for training"""
 
 import importlib
 import inspect
 import logging
+from typing import Any
 
 from trl.trainer.grpo_trainer import RewardFunc
 
-from axolotl.core.trainers.grpo.trainer import AxolotlGRPOTrainer
+from axolotl.core.trainers.grpo.args import AxolotlGRPOConfig
+from axolotl.core.trainers.grpo.trainer import (
+    AxolotlGRPOSequenceParallelTrainer,
+    AxolotlGRPOTrainer,
+)
+from axolotl.utils.dict import DictDefault
 from axolotl.utils.schemas.trl import TRLConfig
 
-LOG = logging.getLogger("axolotl")
+LOG = logging.getLogger(__name__)
 
 
 class GRPOStrategy:
-    """
-    Strategy for GRPO training
-    """
+    """Strategy for GRPO training"""
 
     @classmethod
-    def get_trainer_class(cls):
+    def get_trainer_class(
+        cls, sequence_parallel: bool
+    ) -> type[AxolotlGRPOTrainer] | type[AxolotlGRPOSequenceParallelTrainer]:
+        if sequence_parallel:
+            return AxolotlGRPOSequenceParallelTrainer
         return AxolotlGRPOTrainer
 
     @classmethod
-    def get_training_args_class(cls):
-        from axolotl.core.trainers.grpo.args import AxolotlGRPOConfig
-
+    def get_training_args_class(cls) -> type[AxolotlGRPOConfig]:
         return AxolotlGRPOConfig
 
     @classmethod
-    def set_training_args_kwargs(cls, cfg):
-        grpo_args_kwargs = {}
+    def set_training_args_kwargs(cls, cfg: DictDefault) -> dict[str, Any]:
+        grpo_args_kwargs: dict[str, Any] = {}
 
         if not hasattr(cfg, "trl") or not cfg.trl:
             return grpo_args_kwargs
@@ -40,8 +44,8 @@ class GRPOStrategy:
 
         if trl.use_vllm:
             grpo_args_kwargs["use_vllm"] = trl.use_vllm
-            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host or trl.vllm.host
-            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port or trl.vllm.port
+            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host or trl.vllm.host  # type: ignore[attr-defined]
+            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port or trl.vllm.port  # type: ignore[attr-defined]
             if trl.vllm_server_timeout:
                 grpo_args_kwargs["vllm_server_timeout"] = trl.vllm_server_timeout
             if trl.vllm_guided_decoding_regex:
@@ -102,17 +106,18 @@ class GRPOStrategy:
         return grpo_args_kwargs
 
     @classmethod
-    def set_trainer_args(cls, cfg):
+    def set_trainer_args(cls, cfg: DictDefault) -> list[Any]:
         trainer_args = []
         if cfg.trl and cfg.trl.reward_funcs:
             reward_funcs = []
             for reward_func_fqn in cfg.trl.reward_funcs:
                 reward_funcs.append(cls.get_reward_func(reward_func_fqn))
             trainer_args.append(reward_funcs)
+
         return trainer_args
 
     @classmethod
-    def set_trainer_kwargs(cls, cfg):
+    def set_trainer_kwargs(cls, cfg: DictDefault) -> dict[str, Any]:
         trainer_kwargs = {}
         if cfg.trl and cfg.trl.reward_processing_classes:
             trainer_kwargs["reward_processing_classes"] = (
@@ -126,7 +131,7 @@ class GRPOStrategy:
         return None
 
     @classmethod
-    def get_blocklist_args_kwargs(cls):
+    def get_blocklist_args_kwargs(cls) -> list[str]:
         return ["dataset_num_proc"]
 
     @classmethod
@@ -137,13 +142,13 @@ class GRPOStrategy:
         Args:
             reward_func_fqn (str): Fully qualified name of the reward function (e.g. r1_grpo.gsm8k_transform),
                 or a HF hub path to the reward model.
-        Raises:
-            ValueError: If the reward function does not accept at least two arguments.
 
         Returns:
             RewardFunc: A callable that accepts prompts and completions and returns rewards,
                 or a path to a reward model.
 
+        Raises:
+            ValueError: If the reward function does not accept at least two arguments.
         """
         try:
             # use importlib to dynamically load the reward function from the module

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

@@ -11,6 +11,4 @@ from axolotl.core.training_args import AxolotlTrainingMixins
 
 @dataclass
 class AxolotlGRPOConfig(AxolotlTrainingMixins, GRPOConfig):
-    """
-    Axolotl GRPO Config for GRPO training
-    """
+    """Axolotl GRPO Config for GRPO training"""

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

@@ -0,0 +1,172 @@
+"""Repeat random sampler (similar to the one implemented in
+https://github.com/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py) that adds
+sequence parallelism functionality; i.e., duplicating data across ranks in the same
+sequence parallel group.
+"""
+
+from typing import Iterator, Sized
+
+import torch
+from torch.utils.data import Sampler
+
+
+class SequenceParallelRepeatRandomSampler(Sampler):
+    """Sampler for GRPO training with sequence parallelism.
+
+    This sampler ensures:
+    - Ranks in the same sequence parallel (SP) group receive identical data.
+    - Each index is repeated multiple times for sampling different completions.
+    - Entire batches are repeated for reuse in multiple updates.
+    - Data is properly distributed across SP groups.
+
+    In the table below, the values represent dataset indices. Each SP group has
+    `sequence_parallel_degree = 2` GPUs working together on the same data. There are 2
+    SP groups (SP0 and SP1), with `world_size = 4` total GPUs.
+
+                                               Sequence Parallel Groups
+                                        |       SP0        |       SP1        |
+                                        |  GPU 0  |  GPU 1 |  GPU 2  |  GPU 3 |
+                    global_step  step    <---> mini_repeat_count=3
+                                            <----------> batch_size=2 per SP group
+    grad_accum=2   ▲  ▲  0       0         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- SP groups get different data
+                   ▼  |  0       1         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- Same data for each SP group GPU
+                      |
+                      |  1       2         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- Repeat same indices for iterations
+    num_iterations=2  ▼  1       3         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- When using gradient accumulation
+
+                         2       4         [4 4 4  5 5 5]     [6 6 6  7 7 7]   <- New batch of data indices
+                         2       5         [4 4 4  5 5 5]     [6 6 6  7 7 7]
+                                            ...
+
+    Args:
+        dataset: Dataset to sample from.
+        mini_repeat_count: How many times to repeat each sample immediately.
+        world_size: Total number of processes.
+        rank: Rank of current process.
+        batch_size: Number of samples per batch.
+        repeat_count: How many times to repeat the full sampling process.
+        sequence_parallel_degree: Number of ranks in a sequence parallel group.
+        shuffle: Whether to shuffle the dataset.
+        seed: Random seed for shuffling.
+        drop_last: Whether to drop the last incomplete batch.
+    """
+
+    def __init__(
+        self,
+        dataset: Sized,
+        mini_repeat_count: int,
+        world_size: int,
+        rank: int,
+        batch_size: int = 1,
+        repeat_count: int = 1,
+        sequence_parallel_degree: int = 1,
+        shuffle: bool = True,
+        seed: int = 0,
+        drop_last: bool = False,
+    ):
+        self.dataset = dataset
+        self.mini_repeat_count = mini_repeat_count
+        self.batch_size = batch_size
+        self.repeat_count = repeat_count
+        self.shuffle = shuffle
+        self.seed = seed
+        self.drop_last = drop_last
+        self.epoch = 0
+
+        self.world_size = world_size
+        self.rank = rank
+
+        # Sequence parallelism parameters
+        self.sequence_parallel_degree = sequence_parallel_degree
+        self.num_sp_groups = world_size // sequence_parallel_degree
+        self.sp_group_id = rank // sequence_parallel_degree
+
+        # Adjust dataset size for distributed sampling
+        self.num_samples = len(self.dataset)
+        self.total_size = self.num_samples
+
+        # Calculate effective number of samples per SP group
+        if (
+            self.drop_last
+            and self.total_size % (self.num_sp_groups * self.batch_size) != 0
+        ):
+            # Drop last incomplete batch if drop_last is True
+            self.num_samples_per_sp_group = (
+                self.total_size // self.batch_size // self.num_sp_groups
+            ) * self.batch_size
+        else:
+            # Round up to include last batch if drop_last is False
+            self.num_samples_per_sp_group = (
+                (self.total_size + self.batch_size * self.num_sp_groups - 1)
+                // (self.batch_size * self.num_sp_groups)
+                * self.batch_size
+            )
+
+        if shuffle:
+            self.generator = torch.Generator()
+            self.generator.manual_seed(seed)
+
+    def __iter__(self) -> Iterator[int]:
+        """Creates iterator over dataset indices.
+
+        Returns:
+            Iterator that yields indices into the dataset.
+        """
+        # Deterministically shuffle based on epoch and seed
+        if self.shuffle:
+            indices = torch.randperm(
+                self.num_samples, generator=self.generator
+            ).tolist()
+        else:
+            indices = list(range(self.num_samples))
+
+        # Add extra samples to make it evenly divisible by batch_size
+        if len(indices) % self.batch_size != 0:
+            padding = indices[: self.batch_size - len(indices) % self.batch_size]
+            indices += padding
+
+        # Subsample based on SP group ID
+        # Each SP group gets distinct batches of data
+        batch_indices = []
+        for i in range(0, len(indices), self.batch_size * self.num_sp_groups):
+            start_idx = i + self.sp_group_id * self.batch_size
+            end_idx = min(start_idx + self.batch_size, len(indices))
+            if start_idx < len(indices):
+                for j in range(self.batch_size):
+                    if start_idx + j < end_idx:
+                        batch_indices.append(indices[start_idx + j])
+
+        # Make sure batch_indices is exactly batch_size * num_batches_per_sp_group
+        if self.drop_last:
+            num_batches_per_sp_group = self.num_samples_per_sp_group // self.batch_size
+            target_len = self.batch_size * num_batches_per_sp_group
+            if len(batch_indices) > target_len:
+                batch_indices = batch_indices[:target_len]
+
+        # Apply the GRPO repeat pattern
+        final_indices = []
+        for _ in range(self.repeat_count):
+            for idx in batch_indices:
+                for _ in range(self.mini_repeat_count):
+                    final_indices.append(idx)
+
+        return iter(final_indices)
+
+    def __len__(self) -> int:
+        """Returns the total length of the iterable including repetitions.
+
+        Returns:
+            Total number of samples.
+        """
+        # Total length including all repetitions
+        return (
+            self.num_samples_per_sp_group * self.mini_repeat_count * self.repeat_count
+        )
+
+    def set_epoch(self, epoch: int) -> None:
+        """Sets the epoch for this sampler.
+
+        Args:
+            epoch: Epoch number to use for shuffling.
+        """
+        self.epoch = epoch

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

@@ -1,69 +1,653 @@
-"""
-Axolotl GRPO trainer
-"""
+"""Axolotl GRPO trainers (with and without sequence parallelism handling)"""
 
-from contextlib import nullcontext
+# pylint: disable=too-many-lines,duplicate-code,protected-access,no-member
 
-from accelerate.utils import is_deepspeed_available, is_peft_model
+import warnings
+from typing import Any
+
+import datasets
+import torch
+import torch.distributed as dist
+import torch.utils.data
+from accelerate.utils import (
+    broadcast_object_list,
+    gather,
+    gather_object,
+    is_peft_available,
+)
+from datasets import Dataset, IterableDataset
+from torch import nn
+from torch.utils.data import (
+    BatchSampler,
+    DataLoader,
+    Sampler,
+)
+from transformers import (
+    PreTrainedModel,
+    PreTrainedTokenizerBase,
+    Trainer,
+    TrainerCallback,
+)
+from transformers.trainer_utils import seed_worker
 from trl import GRPOTrainer
-from trl.extras.profiling import profiling_decorator
+from trl.data_utils import (
+    apply_chat_template,
+    is_conversational,
+    maybe_apply_chat_template,
+)
+from trl.extras.profiling import profiling_context
+from trl.models import unwrap_model_for_generation
+from trl.trainer.grpo_config import GRPOConfig
+from trl.trainer.grpo_trainer import RewardFunc, nanstd
+from trl.trainer.utils import pad
 
+from axolotl.core.trainers.grpo.sampler import SequenceParallelRepeatRandomSampler
 from axolotl.core.trainers.mixins import RngLoaderMixin, SchedulerMixin
+from axolotl.monkeypatch.ring_attn import get_ring_attn_group
 
-if is_deepspeed_available():
-    import deepspeed
+if is_peft_available():
+    # pylint: disable=unused-import
+    from peft import PeftConfig
 
 
 class AxolotlGRPOTrainer(RngLoaderMixin, SchedulerMixin, GRPOTrainer):
-    """
-    Extend the base GRPOTrainer for axolotl helpers
-    """
+    """Extend the base GRPOTrainer for axolotl helpers"""
 
     _tag_names = ["trl", "grpo", "axolotl"]
 
-    @profiling_decorator
-    def _move_model_to_vllm(self):
-        # For DeepSpeed ZeRO-3, we need to gather all parameters before operations
-        deepspeed_plugin = self.accelerator.state.deepspeed_plugin
-        zero_stage_3 = deepspeed_plugin is not None and deepspeed_plugin.zero_stage == 3
-        gather_if_zero3 = (
-            deepspeed.zero.GatheredParameters if zero_stage_3 else nullcontext
+
+class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
+    """Extend the base GRPOTrainer for sequence parallelism handling"""
+
+    def __init__(
+        self,
+        model: str | PreTrainedModel,
+        reward_funcs: RewardFunc | list[RewardFunc],
+        args: GRPOConfig | None = None,
+        train_dataset: Dataset | IterableDataset | None = None,
+        eval_dataset: (
+            Dataset | IterableDataset | dict[str, Dataset | IterableDataset] | None
+        ) = None,
+        processing_class: PreTrainedTokenizerBase | None = None,
+        reward_processing_classes: (
+            PreTrainedTokenizerBase | list[PreTrainedTokenizerBase] | None
+        ) = None,
+        callbacks: list[TrainerCallback] | None = None,
+        optimizers: tuple[
+            torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None
+        ] = (None, None),
+        peft_config: "PeftConfig | None" = None,
+    ):
+        # First call the superclass constructor with all arguments
+        super().__init__(
+            model=model,
+            reward_funcs=reward_funcs,
+            args=args,
+            train_dataset=train_dataset,
+            eval_dataset=eval_dataset,
+            processing_class=processing_class,
+            reward_processing_classes=reward_processing_classes,
+            callbacks=callbacks,
+            optimizers=optimizers,
+            peft_config=peft_config,
         )
 
-        if is_peft_model(self.model):
-            # With PEFT and DeepSpeed ZeRO Stage 3, we must gather the full model at once before merging, as merging
-            # adapters in a sharded manner is not supported.
-            with gather_if_zero3(list(self.model.parameters())):
-                self.model.merge_adapter()
+        # Get number of SP groups (number of processes divided by SP degree)
+        num_processes = self.accelerator.num_processes
+        num_sp_groups = num_processes // self.args.sequence_parallel_degree
 
-                # Update vLLM weights while parameters are gathered
-                for name, param in self.model.named_parameters():
-                    # When using PEFT, we need to recover the original parameter name and discard some parameters
-                    name = (
-                        name.removeprefix("base_model.model.")
-                        .removeprefix("base_model.model.")
-                        .replace(".base_layer", "")
-                    )
-                    if self.model.prefix in name:
-                        continue
-                    # When module to save, remove its prefix and discard the original module
-                    if "original_module" in name:
-                        continue
-                    name = name.replace("modules_to_save.default.", "")
+        # Calculate batch size per SP group (not per process)
+        sp_group_batch_size = self.args.per_device_train_batch_size * num_sp_groups
+        possible_values = [
+            n_gen
+            for n_gen in range(2, sp_group_batch_size + 1)
+            if (sp_group_batch_size) % n_gen == 0
+        ]
 
-                    if self.accelerator.is_main_process:
-                        self.vllm_client.update_named_param(name, param.data)
+        if self.num_generations not in possible_values:
+            raise ValueError(
+                f"The batch size per SP group ({num_sp_groups} x "
+                f"{self.args.per_device_train_batch_size}) must be evenly divisible by "
+                f"the number of generations per prompt ({self.num_generations}). Given "
+                "the current configuration, the valid values for the number of "
+                f"generations are: {possible_values}."
+            )
 
-                # Unmerge adapters while parameters are still gathered
-                self.model.unmerge_adapter()
-                # Parameters will automatically be repartitioned when exiting the context
+        if self.args.eval_strategy != "no":
+            # If sequence parallelism is enabled, calculate batch size per SP group
+            sp_group_eval_batch_size = args.per_device_eval_batch_size * num_sp_groups  # type: ignore[union-attr]
+            possible_values = [
+                n_gen
+                for n_gen in range(2, sp_group_eval_batch_size + 1)
+                if (sp_group_eval_batch_size) % n_gen == 0
+            ]
+
+            if self.num_generations not in possible_values:
+                raise ValueError(
+                    f"With sequence parallelism (degree {self.args.sequence_parallel_degree}), "
+                    f"the eval batch size per SP group ({num_sp_groups} x {self.args.per_device_eval_batch_size}) "
+                    f"must be evenly divisible by the number of generations per prompt "
+                    f"({self.num_generations}). Given the current eval batch size, "
+                    f"the valid values for the number of generations are: {possible_values}."
+                )
+
+        # Initialize the SP group
+        self.sp_group = get_ring_attn_group()
+        self.rank = dist.get_rank()
+        self.world_size = dist.get_world_size()
+        self.local_rank = dist.get_rank(group=self.sp_group)
+        self.local_world_size = dist.get_world_size(group=self.sp_group)
+
+    def _get_train_sampler(self) -> Sampler:
+        effective_batch_size = (
+            self.args.per_device_train_batch_size
+            * self.world_size
+            * self.args.gradient_accumulation_steps
+        )
+
+        return SequenceParallelRepeatRandomSampler(
+            dataset=self.train_dataset,
+            mini_repeat_count=self.num_generations,
+            world_size=self.world_size,
+            rank=self.rank,
+            batch_size=effective_batch_size
+            // self.num_generations
+            // self.args.sequence_parallel_degree,
+            repeat_count=self.num_iterations * self.args.gradient_accumulation_steps,
+            sequence_parallel_degree=self.args.sequence_parallel_degree,
+            shuffle=True,
+            seed=self.args.seed,
+            drop_last=True,
+        )
+
+    def _create_dataloader_params(self, is_eval=False, custom_batch_size=None):
+        """Create common dataloader parameters for train or eval."""
+        batch_size = custom_batch_size or (
+            self.args.eval_batch_size if is_eval else self._train_batch_size
+        )
+
+        params = {
+            "batch_size": batch_size,
+            "collate_fn": self.data_collator,
+            "num_workers": self.args.dataloader_num_workers,
+            "pin_memory": self.args.dataloader_pin_memory,
+        }
+
+        # Add persistent workers only for training
+        if not is_eval and hasattr(self.args, "dataloader_persistent_workers"):
+            params["persistent_workers"] = self.args.dataloader_persistent_workers
+
+        # Add prefetch factor if specified
+        if self.args.dataloader_prefetch_factor:
+            params["prefetch_factor"] = self.args.dataloader_prefetch_factor
+
+        return params
+
+    def _prepare_dataloader(
+        self, dataset, sampler, is_eval=False, custom_batch_size=None
+    ):
+        """Prepare a dataloader with the given dataset and sampler."""
+        # Get base parameters
+        dataloader_params = self._create_dataloader_params(is_eval, custom_batch_size)
+
+        # Add sampler configuration
+        if not isinstance(dataset, torch.utils.data.IterableDataset):
+            if isinstance(sampler, BatchSampler):
+                # batch_size and batch_sampler are mutually exclusive
+                dataloader_params["batch_sampler"] = sampler
+                del dataloader_params["batch_size"]
+            else:
+                dataloader_params["sampler"] = sampler
+                dataloader_params["drop_last"] = self.args.dataloader_drop_last
+
+            if not is_eval:
+                dataloader_params["worker_init_fn"] = seed_worker
+
+        # Create the dataloader
+        dataloader = DataLoader(dataset, **dataloader_params)
+
+        if self.args.sample_packing and (
+            (not is_eval and not self.args.pretraining)
+            or (is_eval and self.args.eval_sample_packing is not False)
+        ):
+            self.accelerator.even_batches = False
+
+        # Return unprepared dataloader if using sequence parallelism
+        # TODO(djsaunde): We might be able to use `accelerate`'s dataloader preparation
+        # if we use `dispatch_batches` and `slice_fn_for_dispatch` properly (i.e.,
+        # slice each batch along the sequence dimension).
+        if self.args.sequence_parallel_degree > 1:
+            return dataloader
+
+        # Otherwise prepare with accelerator
+        return self.accelerator.prepare_data_loader(dataloader)
+
+    def get_train_dataloader(self) -> DataLoader:
+        """Get dataloader for training"""
+        train_dataset = self.train_dataset
+        # pylint: disable=access-member-before-definition
+        data_collator = self.data_collator  # type: ignore
+
+        # Handle dataset preprocessing
+        if isinstance(train_dataset, datasets.Dataset):
+            # Add debug print before any modifications
+            if self.args.sample_packing and not self.args.pretraining:
+                train_dataset = train_dataset.remove_columns(["length"])
+            if not self.args.sample_packing or self.args.pretraining:
+                train_dataset = self._remove_unused_columns(
+                    train_dataset, description="training"
+                )
         else:
-            # For non-PEFT models, simply gather and update each parameter individually.
-            for name, param in self.model.named_parameters():
-                with gather_if_zero3([param]):
-                    if self.accelerator.is_main_process:
-                        self.vllm_client.update_named_param(name, param.data)
+            self.data_collator = self._get_collator_with_removed_columns(  # pylint: disable=attribute-defined-outside-init
+                data_collator,
+                description="training",
+            )
 
-        # Reset cache on main process
-        if self.accelerator.is_main_process:
-            self.vllm_client.reset_prefix_cache()
+        # Get sampler and create dataloader
+        sampler = self._get_train_sampler()
+        dataloader = self._prepare_dataloader(train_dataset, sampler, is_eval=False)
+
+        return dataloader
+
+    def _generate_and_score_completions(
+        self, inputs: list[dict[str, torch.Tensor | Any]]
+    ) -> dict[str, torch.Tensor | Any]:
+        device = self.accelerator.device
+        mode = "eval" if self.control.should_evaluate else "train"
+
+        prompts = [x["prompt"] for x in inputs]
+        prompts_text = [
+            maybe_apply_chat_template(example, self.processing_class)["prompt"]
+            for example in inputs
+        ]
+        prompt_inputs = self.processing_class(
+            text=prompts_text,
+            return_tensors="pt",
+            padding=True,
+            padding_side="left",
+            add_special_tokens=False,
+        )
+        prompt_inputs = Trainer._prepare_inputs(self, prompt_inputs)
+        prompt_ids, prompt_mask = (
+            prompt_inputs["input_ids"],
+            prompt_inputs["attention_mask"],
+        )
+
+        if self.max_prompt_length is not None:
+            prompt_ids = prompt_ids[:, -self.max_prompt_length :]
+            prompt_mask = prompt_mask[:, -self.max_prompt_length :]
+
+        # Generate completions using either vLLM or regular generation
+        if self.args.use_vllm:
+            # First, have main process load weights if needed
+            # pylint: disable=access-member-before-definition
+            if self.state.global_step != self._last_loaded_step:  # type: ignore[has-type]
+                self._move_model_to_vllm()
+                # pylint: disable=attribute-defined-outside-init
+                self._last_loaded_step = self.state.global_step
+
+            # Generate completions using vLLM: gather all prompts and use them in a single call in the main process
+            all_prompts_text = gather_object(prompts_text)
+            if self.accelerator.is_main_process:
+                if self.args.sequence_parallel_degree > 1:
+                    # Calculate sequence parallel group information
+                    world_size = self.accelerator.num_processes
+                    sequence_parallel_degree = self.args.sequence_parallel_degree
+                    num_sp_groups = world_size // sequence_parallel_degree
+
+                    # Since processes in the same SP group have the same prompts, we need to ensure
+                    # we only take one copy of each prompt from each SP group
+                    ordered_set_of_prompts = []
+                    for sp_group_id in range(num_sp_groups):
+                        # Get the first process from each SP group (typically the group leader)
+                        group_leader_rank = sp_group_id * sequence_parallel_degree
+
+                        # Extract prompts from this SP group, accounting for num_generations duplicates
+                        # We only need prompts from one rank in each SP group
+                        group_prompts = all_prompts_text[
+                            group_leader_rank
+                            * len(prompts_text) : (group_leader_rank + 1)
+                            * len(prompts_text) : self.num_generations
+                        ]
+
+                        ordered_set_of_prompts.extend(group_prompts)
+                else:
+                    # Since 'prompts' contains 'num_generations' duplicates, we first take unique prompts, and generate
+                    # num_generations outputs for each one. This is faster than generating outputs for each duplicate
+                    # prompt individually.
+                    ordered_set_of_prompts = all_prompts_text[
+                        :: self.num_generations * self.args.sequence_parallel_degree
+                    ]
+
+                with profiling_context(self, "vLLM.generate"):
+                    completion_ids = self.vllm_client.generate(
+                        prompts=ordered_set_of_prompts,
+                        n=self.num_generations,
+                        repetition_penalty=self.repetition_penalty,
+                        temperature=self.temperature,
+                        top_p=self.top_p,
+                        top_k=-1 if self.top_k is None else self.top_k,
+                        min_p=0.0 if self.min_p is None else self.min_p,
+                        max_tokens=self.max_completion_length,
+                        guided_decoding_regex=self.guided_decoding_regex,
+                    )
+            else:
+                completion_ids = [None] * (
+                    len(all_prompts_text) // self.args.sequence_parallel_degree
+                )
+
+            # Broadcast the completions from the main process to all processes
+            completion_ids = broadcast_object_list(completion_ids, from_process=0)
+
+            # Determine the appropriate slice based on sequence parallelism
+            if self.args.sequence_parallel_degree > 1:
+                # Calculate SP group ID (which group of ranks this rank belongs to)
+                sp_group_id = self.accelerator.process_index // self.local_world_size
+
+                # Calculate the start index for this SP group
+                sp_group_start = sp_group_id * len(prompts) * self.local_world_size
+
+                # All ranks in the same SP group get the same data slice
+                process_slice = slice(
+                    sp_group_start,
+                    sp_group_start + len(prompts),
+                )
+                completion_ids = completion_ids[process_slice]
+            else:
+                # Original behavior for non-sequence parallel case
+                process_slice = slice(
+                    self.accelerator.process_index * len(prompts),
+                    (self.accelerator.process_index + 1) * len(prompts),
+                )
+                completion_ids = completion_ids[process_slice]
+
+            # Pad the completions, and concatenate them with the prompts
+            completion_ids = [
+                torch.tensor(ids, device=device) for ids in completion_ids
+            ]
+            completion_ids = pad(
+                completion_ids, padding_value=self.processing_class.pad_token_id
+            )
+            prompt_completion_ids = torch.cat([prompt_ids, completion_ids], dim=1)
+        else:
+            # Regular generation path
+            with unwrap_model_for_generation(
+                self.model_wrapped,
+                self.accelerator,
+                gather_deepspeed3_params=self.args.ds3_gather_for_generation,
+            ) as unwrapped_model:
+                prompt_completion_ids = unwrapped_model.generate(
+                    prompt_ids,
+                    attention_mask=prompt_mask,
+                    generation_config=self.generation_config,
+                )
+
+            # Compute prompt length and extract completion ids
+            prompt_length = prompt_ids.size(1)
+            prompt_ids = prompt_completion_ids[:, :prompt_length]
+            completion_ids = prompt_completion_ids[:, prompt_length:]
+
+        # Mask everything after the first EOS token
+        is_eos = completion_ids == self.processing_class.eos_token_id
+        eos_idx = torch.full(
+            (is_eos.size(0),), is_eos.size(1), dtype=torch.long, device=device
+        )
+        eos_idx[is_eos.any(dim=1)] = is_eos.int().argmax(dim=1)[is_eos.any(dim=1)]
+        sequence_indices = torch.arange(is_eos.size(1), device=device).expand(
+            is_eos.size(0), -1
+        )
+        completion_mask = (sequence_indices <= eos_idx.unsqueeze(1)).int()
+
+        # If mask_truncated_completions is enabled, zero out truncated completions in completion_mask
+        if self.args.mask_truncated_completions:
+            truncated_completions = ~is_eos.any(dim=1)
+            completion_mask = (
+                completion_mask * (~truncated_completions).unsqueeze(1).int()
+            )
+
+        # Concatenate prompt_mask with completion_mask for logit computation
+        attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)  # (B, P+C)
+
+        logits_to_keep = completion_ids.size(
+            1
+        )  # we only need to compute the logits for the completion tokens
+        batch_size = (
+            self.args.per_device_train_batch_size
+            if mode == "train"
+            else self.args.per_device_eval_batch_size
+        )
+
+        with torch.no_grad():
+            # When using num_iterations == 1, old_per_token_logps == per_token_logps, so we can skip it's
+            # computation here, and use per_token_logps.detach() instead.
+            if self.num_iterations > 1:
+                old_per_token_logps = self._get_per_token_logps(
+                    self.model,
+                    prompt_completion_ids,
+                    attention_mask,
+                    logits_to_keep,
+                    batch_size,
+                )
+            else:
+                old_per_token_logps = None
+
+            if self.beta == 0.0:
+                ref_per_token_logps = None
+            elif self.ref_model is not None:
+                ref_per_token_logps = self._get_per_token_logps(
+                    self.ref_model,
+                    prompt_completion_ids,
+                    attention_mask,
+                    logits_to_keep,
+                    batch_size,
+                )
+            else:
+                with self.accelerator.unwrap_model(self.model).disable_adapter():
+                    ref_per_token_logps = self._get_per_token_logps(
+                        self.model,
+                        prompt_completion_ids,
+                        attention_mask,
+                        logits_to_keep,
+                        batch_size,
+                    )
+
+        # Decode the generated completions
+        completions_text = self.processing_class.batch_decode(
+            completion_ids, skip_special_tokens=True
+        )
+        if is_conversational(inputs[0]):
+            completions = []
+            for prompt, completion in zip(prompts, completions_text):
+                bootstrap = (
+                    prompt.pop()["content"] if prompt[-1]["role"] == "assistant" else ""
+                )
+                completions.append(
+                    [{"role": "assistant", "content": bootstrap + completion}]
+                )
+        else:
+            completions = completions_text
+
+        rewards_per_func = torch.zeros(
+            len(prompts), len(self.reward_funcs), device=device
+        )
+        for i, (reward_func, reward_processing_class, reward_func_name) in enumerate(
+            zip(
+                self.reward_funcs,
+                self.reward_processing_classes,
+                self.reward_func_names,
+            )
+        ):
+            with profiling_context(self, reward_func_name):
+                if isinstance(
+                    reward_func, nn.Module
+                ):  # Module instead of PretrainedModel for compat with compiled models
+                    if is_conversational(inputs[0]):
+                        messages = [
+                            {"messages": p + c} for p, c in zip(prompts, completions)
+                        ]
+                        texts = [
+                            apply_chat_template(x, reward_processing_class)["text"]
+                            for x in messages
+                        ]
+                    else:
+                        texts = [p + c for p, c in zip(prompts, completions)]
+                    reward_inputs = reward_processing_class(
+                        text=texts,
+                        return_tensors="pt",
+                        padding=True,
+                        padding_side="right",
+                        add_special_tokens=False,
+                    )
+                    reward_inputs = Trainer._prepare_inputs(self, reward_inputs)
+                    with torch.inference_mode():
+                        rewards_per_func[:, i] = reward_func(**reward_inputs).logits[
+                            :, 0
+                        ]  # Shape (B*G,)
+                else:
+                    # Repeat all input columns (but "prompt" and "completion") to match the number of generations
+                    keys = [
+                        key for key in inputs[0] if key not in ["prompt", "completion"]
+                    ]
+                    reward_kwargs = {
+                        key: [example[key] for example in inputs] for key in keys
+                    }
+                    output_reward_func = reward_func(
+                        prompts=prompts, completions=completions, **reward_kwargs
+                    )
+                    # Convert None values to NaN
+                    output_reward_func = [
+                        reward if reward is not None else torch.nan
+                        for reward in output_reward_func
+                    ]
+
+                    rewards_per_func[:, i] = torch.tensor(
+                        output_reward_func, dtype=torch.float32, device=device
+                    )
+
+        # If all reward functions return None for a given row, issue a detailed warning
+        if torch.isnan(rewards_per_func).all(dim=1).any():
+            nan_row_idx = (
+                torch.isnan(rewards_per_func).all(dim=1).nonzero(as_tuple=True)[0][0]
+            )
+            row_reward_kwargs = {
+                key: value[nan_row_idx] for key, value in reward_kwargs.items()
+            }
+            row_reward_kwargs["prompt"] = prompts[nan_row_idx]
+            row_reward_kwargs["completion"] = completions[nan_row_idx]
+            warnings.warn(
+                f"All reward functions returned None for the following kwargs: {row_reward_kwargs}. "
+                "Please ensure that at least one reward function returns a valid reward."
+            )
+
+        # Gather the reward per function: this part is crucial, because the rewards are normalized per group and the
+        # completions may be distributed across processes
+        rewards_per_func = gather(rewards_per_func)
+
+        # Apply weights to each reward function's output and sum
+        rewards = (
+            rewards_per_func * self.reward_weights.to(device).unsqueeze(0)
+        ).nansum(dim=1)
+
+        # Compute grouped-wise rewards
+        mean_grouped_rewards = rewards.view(-1, self.num_generations).mean(dim=1)
+        std_grouped_rewards = rewards.view(-1, self.num_generations).std(dim=1)
+
+        # Normalize the rewards to compute the advantages
+        mean_grouped_rewards = mean_grouped_rewards.repeat_interleave(
+            self.num_generations, dim=0
+        )
+        std_grouped_rewards = std_grouped_rewards.repeat_interleave(
+            self.num_generations, dim=0
+        )
+        advantages = rewards - mean_grouped_rewards
+        if self.args.scale_rewards:
+            advantages = advantages / (std_grouped_rewards + 1e-4)
+
+        # Slice to keep only the local part of the data
+        if self.args.sequence_parallel_degree > 1:
+            # Calculate SP group ID (which group of ranks this rank belongs to)
+            sp_group_id = self.accelerator.process_index // self.local_world_size
+
+            # Calculate the start index for this SP group
+            sp_group_start = sp_group_id * len(prompts) * self.local_world_size
+
+            # All ranks in the same SP group get the same data slice
+            process_slice = slice(
+                sp_group_start,
+                sp_group_start + len(prompts),
+            )
+        else:
+            # Original behavior for non-sequence parallel case
+            process_slice = slice(
+                self.accelerator.process_index * len(prompts),
+                (self.accelerator.process_index + 1) * len(prompts),
+            )
+        advantages = advantages[process_slice]
+
+        # Log the metrics
+        if mode == "train":
+            self._total_train_tokens += (
+                self.accelerator.gather_for_metrics(attention_mask.sum()).sum().item()
+            )
+        self._metrics[mode]["num_tokens"] = [self._total_train_tokens]
+
+        # log completion lengths, mean, min, max
+        agg_completion_mask = self.accelerator.gather_for_metrics(
+            completion_mask.sum(1)
+        )
+        self._metrics[mode]["completions/mean_length"].append(
+            agg_completion_mask.float().mean().item()
+        )
+        self._metrics[mode]["completions/min_length"].append(
+            agg_completion_mask.float().min().item()
+        )
+        self._metrics[mode]["completions/max_length"].append(
+            agg_completion_mask.float().max().item()
+        )
+
+        # identify sequences that terminated with EOS and log their lengths
+        agg_terminated_with_eos = self.accelerator.gather_for_metrics(is_eos.any(dim=1))
+        term_completion_mask = agg_completion_mask[agg_terminated_with_eos]
+        clipped_completions_ratio = 1 - len(term_completion_mask) / len(
+            agg_completion_mask
+        )
+        self._metrics[mode]["completions/clipped_ratio"].append(
+            clipped_completions_ratio
+        )
+        if len(term_completion_mask) == 0:
+            # edge case where no completed sequences are found
+            term_completion_mask = torch.zeros(1, device=device)
+        self._metrics[mode]["completions/mean_terminated_length"].append(
+            term_completion_mask.float().mean().item()
+        )
+        self._metrics[mode]["completions/min_terminated_length"].append(
+            term_completion_mask.float().min().item()
+        )
+        self._metrics[mode]["completions/max_terminated_length"].append(
+            term_completion_mask.float().max().item()
+        )
+
+        # Calculate mean reward per function, but only for samples where the function was applied (non-NaN values)
+        for i, reward_func_name in enumerate(self.reward_func_names):
+            mean_rewards = torch.nanmean(rewards_per_func[:, i]).item()
+            self._metrics[mode][f"rewards/{reward_func_name}/mean"].append(mean_rewards)
+            std_rewards = nanstd(rewards_per_func[:, i]).item()
+            self._metrics[mode][f"rewards/{reward_func_name}/std"].append(std_rewards)
+        self._metrics[mode]["reward"].append(mean_grouped_rewards.mean().item())
+        self._metrics[mode]["reward_std"].append(std_grouped_rewards.mean().item())
+
+        # Log prompt and completion texts
+        self._textual_logs["prompt"].extend(gather_object(prompts_text))
+        self._textual_logs["completion"].extend(gather_object(completions_text))
+        for i, name in enumerate(self.reward_func_names):
+            self._textual_logs["rewards"][name].extend(rewards_per_func[:, i].tolist())
+
+        return {
+            "prompt_ids": prompt_ids,
+            "prompt_mask": prompt_mask,
+            "completion_ids": completion_ids,
+            "completion_mask": completion_mask,
+            "advantages": advantages,
+            "old_per_token_logps": old_per_token_logps,
+            "ref_per_token_logps": ref_per_token_logps,
+        }

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

@@ -6,4 +6,3 @@
 from .optimizer import OptimizerMixin
 from .rng_state_loader import RngLoaderMixin
 from .scheduler import SchedulerMixin
-from .sequence_parallel import SequenceParallelContextManager, SequenceParallelMixin

src/axolotl/core/trainers/mixins/sequence_parallel.py

@@ -1,313 +0,0 @@
-"""
-Module for Axolotl trainer sequence parallelism mixin and training context manager
-"""
-
-import functools
-import logging
-
-import torch
-import torch.distributed as dist
-from datasets import Dataset
-from torch import nn
-from torch.utils.data import DistributedSampler, Sampler
-from torch.utils.hooks import RemovableHandle
-
-from axolotl.monkeypatch.attention.ring_attn import (
-    RingAttnFunc,
-    get_ring_attn_group,
-    update_ring_attn_params,
-)
-
-LOG = logging.getLogger(__name__)
-
-
-def apply_sequence_parallelism(
-    batch: dict[str, torch.Tensor],
-    local_rank: int,
-    local_world_size: int,
-    ring_attn_func: RingAttnFunc,
-) -> dict[str, torch.Tensor]:
-    """
-    Apply sequence parallelism slicing to a batch.
-
-    Args:
-        batch: Batch dictionary (e.g., input_ids, attention_mask, etc.)
-        local_rank: Local rank in the sequence parallel group
-        local_world_size: World size of the sequence parallel group
-        ring_attn_func: The ring attention function to use
-
-    Returns:
-        Sliced batch dictionary.
-    """
-    # Update ring attention params if needed
-    if batch.get("position_ids") is not None:
-        update_ring_attn_params(position_ids=batch["position_ids"])
-
-    # Slice batch for sequence parallel processing
-    total_seq_len = batch["input_ids"].size(1)
-    for key in batch:
-        if (
-            key in batch
-            and isinstance(batch[key], torch.Tensor)
-            and batch[key].dim() > 1
-            and batch[key].size(1) == total_seq_len
-        ):
-
-            if ring_attn_func in [
-                RingAttnFunc.VARLEN_LLAMA3,
-                RingAttnFunc.BATCH_RING,
-            ]:
-                # Split in sequential fashion and grab this rank's chunk
-                batch[key] = (
-                    batch[key].chunk(local_world_size, dim=1)[local_rank].contiguous()
-                )
-            elif ring_attn_func is RingAttnFunc.BATCH_ZIGZAG:
-                chunks = batch[key].chunk(2 * local_world_size, dim=1)
-
-                # Take rank's chunk and opposing chunk for zigzag pattern
-                selected_chunks = [
-                    chunks[local_rank],
-                    chunks[2 * local_world_size - local_rank - 1],
-                ]
-                batch[key] = torch.cat(selected_chunks, dim=1).contiguous()
-            elif ring_attn_func is RingAttnFunc.BATCH_STRIPE:
-                # Split into striped data and stack
-                tensor = torch.stack(
-                    batch[key].split(local_world_size, dim=1),
-                    dim=1,
-                ).transpose(1, 2)
-                batch[key] = tensor[:, local_rank].contiguous()
-
-    return batch
-
-
-class SequenceParallelMixin:
-    """
-    Mixin class for sequence parallelism support in trainers.
-
-    This mixin provides functionality for handling sequence parallelism,
-    specifically for creating appropriate data samplers.
-    """
-
-    args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
-
-    def _setup_sequence_parallel(self):
-        """Set up sequence parallelism environment."""
-        self.ring_attn_group = get_ring_attn_group()
-
-    def _create_sequence_parallel_sampler(
-        self,
-        dataset: Dataset,
-        shuffle: bool = True,
-        is_eval: bool = False,
-    ) -> DistributedSampler:
-        """
-        Helper method to create sampler for sequence parallelism (SP).
-
-        We create a distributed sampler with rank equal to the SP group ID, which
-        means that all ranks in the SP group receive the same sample / set of samples
-        per training step. We also set the number of replicas equal to the number of
-        SP groups, which is a bit of a hack / unintended use, but works!
-
-        Args:
-            dataset: Dataset to sample from.
-            shuffle: Whether to shuffle the dataset.
-            is_eval: Whether we are creating a sampler for evaluation or training.
-
-        Returns:
-            Distributed sampler.
-        """
-        num_sp_groups = self.args.world_size // self.args.sequence_parallel_degree
-        sp_group_id = dist.get_rank() // self.args.sequence_parallel_degree
-
-        return DistributedSampler(
-            dataset,
-            num_replicas=num_sp_groups,
-            rank=sp_group_id,
-            seed=self.args.seed if shuffle else None,
-            shuffle=shuffle,
-            drop_last=not is_eval,
-        )
-
-    def _sp_get_train_sampler(self, dataset) -> Sampler | None:
-        """
-        Get a training sampler configured for sequence parallelism.
-
-        Args:
-            dataset: The training dataset
-
-        Returns:
-            Configured sequence parallel sampler.
-        """
-        return self._create_sequence_parallel_sampler(
-            dataset,
-            shuffle=not self.args.curriculum_sampling,
-        )
-
-    def _sp_get_eval_sampler(self, eval_dataset) -> Sampler | None:
-        """
-        Get an evaluation sampler configured for sequence parallelism.
-
-        Args:
-            eval_dataset: The evaluation dataset.
-
-        Returns:
-            Configured sequence parallel sampler.
-        """
-        return self._create_sequence_parallel_sampler(
-            eval_dataset, shuffle=False, is_eval=True
-        )
-
-
-class SequenceParallelContextManager:
-    """
-    Context manager for sequence parallelism operations.
-
-    This class provides a context that will automatically apply sequence parallelism
-    during model forward passes using a pre-forward hook, and gather outputs from
-    across the sequence parallelism group using a post-forward hook.
-    """
-
-    def __init__(
-        self,
-        model: nn.Module,
-        sequence_parallel_degree: int,
-        ring_attn_func: RingAttnFunc,
-    ):
-        self.model = model
-        self.sequence_parallel_degree = sequence_parallel_degree
-        self.ring_attn_func = ring_attn_func
-        self.process_group = get_ring_attn_group()
-
-        # Initialize sequence parallel group details
-        self.local_rank = dist.get_rank(self.process_group)
-        self.local_world_size = dist.get_world_size(self.process_group)
-
-        # Will store hook handles for removal
-        self.hook_handles: list[RemovableHandle] = []
-
-        # Create a partially applied version of the apply_sequence_parallelism function
-        # with pre-configured params
-        self.apply_sequence_parallelism = functools.partial(
-            apply_sequence_parallelism,
-            local_rank=self.local_rank,
-            local_world_size=self.local_world_size,
-            ring_attn_func=self.ring_attn_func,
-        )
-
-    def __enter__(self):
-        # Forward pre-hook to apply sequence parallelism
-        def sequence_parallel_pre_hook(_, args, kwargs):
-            # Apply sequence parallelism to kwargs
-            kwargs = self.apply_sequence_parallelism(batch=kwargs)
-            return args, kwargs
-
-        # Forward post-hook to gather outputs
-        def sequence_parallel_post_hook(_, __, output):
-            # Gather the sharded outputs
-            return self.gather_outputs(output)
-
-        # Register both hooks
-        self.hook_handles.append(
-            self.model.register_forward_pre_hook(
-                sequence_parallel_pre_hook, with_kwargs=True
-            )
-        )
-        self.hook_handles.append(
-            self.model.register_forward_hook(sequence_parallel_post_hook)
-        )
-
-        return self
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        # Remove all hooks
-        for handle in self.hook_handles:
-            handle.remove()
-        self.hook_handles = []
-
-    def gather_outputs(self, output):
-        """Gather sharded outputs from all ranks and reconstruct the full tensor."""
-        # Handle different output formats (dict, tensor, etc.)
-        if isinstance(output, dict):
-            gathered_output = {}
-            for key, value in output.items():
-                if isinstance(value, torch.Tensor) and value.dim() > 1:
-                    # Gather logits or other sequence-sharded tensors
-                    gathered_value = self.gather_tensor(value)
-                    gathered_output[key] = gathered_value
-                else:
-                    gathered_value = value.clone()
-                    dist.all_reduce(
-                        gathered_value, op=dist.ReduceOp.SUM, group=self.process_group
-                    )
-                    gathered_output[key] = gathered_value
-            return gathered_output
-        if isinstance(output, torch.Tensor):
-            return self.gather_tensor(output)
-
-        return output
-
-    def gather_tensor(self, tensor):
-        """Gather a sharded tensor from all ranks."""
-        # Prepare tensors for all_gather
-        world_size = self.local_world_size
-
-        # Create list to store tensors from all ranks
-        gathered_tensors = [torch.zeros_like(tensor) for _ in range(world_size)]
-
-        # All-gather operation
-        dist.all_gather(gathered_tensors, tensor, group=self.process_group)
-
-        # Concatenate along sequence dimension (typically dim=1)
-        if self.ring_attn_func in [RingAttnFunc.VARLEN_LLAMA3, RingAttnFunc.BATCH_RING]:
-            # Simple concatenation for standard sharding
-            return torch.cat(gathered_tensors, dim=1)
-
-        if self.ring_attn_func is RingAttnFunc.BATCH_ZIGZAG:
-            # Each rank has a pattern of (rank, world_size*2-rank-1)
-            reconstituted_tensors = [None] * (world_size * 2)
-
-            # First, split each gathered tensor into its two chunks
-            for rank, gathered_tensor in enumerate(gathered_tensors):
-                # Each tensor contains two chunks in the sequence dimension
-                chunk_size = gathered_tensor.size(1) // 2
-                chunk1, chunk2 = gathered_tensor.split(chunk_size, dim=1)
-
-                # Place chunks in their original positions
-                reconstituted_tensors[rank] = chunk1
-                reconstituted_tensors[world_size * 2 - rank - 1] = chunk2
-
-            # Concatenate the reconstituted tensors in the correct order
-            return torch.cat(reconstituted_tensors, dim=1)
-
-        # Otherwise, RingAttnFunc.BATCH_STRIPE
-        # In striping, each rank has every world_size-th slice
-        batch_size = tensor.size(0)
-        hidden_dim = tensor.size(-1)
-
-        # First, determine the full sequence length
-        total_seq_len = 0
-        for t in gathered_tensors:
-            total_seq_len += t.size(1)
-
-        # Create a tensor to hold the unstriped result
-        result = torch.zeros(
-            batch_size,
-            total_seq_len,
-            hidden_dim,
-            dtype=tensor.dtype,
-            device=tensor.device,
-        )
-
-        # For each rank's tensor, distribute its slices to the correct positions
-        for rank, gathered_tensor in enumerate(gathered_tensors):
-            # The rank's tensor contains every world_size-th slice
-            # starting from its rank position
-            seq_len = gathered_tensor.size(1)
-            for i in range(seq_len):
-                # Calculate the position in the full tensor
-                pos = i * world_size + rank
-                if pos < total_seq_len:
-                    result[:, pos] = gathered_tensor[:, i]
-
-        return result

src/axolotl/core/training_args.py

@@ -9,8 +9,6 @@ from PIL.Image import Resampling
 from transformers import TrainingArguments
 from trl import CPOConfig, KTOConfig, ORPOConfig, PRMConfig, RewardConfig
 
-from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
-
 
 @dataclass
 class AxolotlTrainingMixins:
@@ -216,14 +214,16 @@ class AxolotlTrainingMixins:
         },
     )
 
-    sequence_parallel_degree: Optional[int] = field(
-        default=1,
-        metadata={"help": "The number of workers to use in sequence parallelism"},
-    )
-    ring_attn_func: Optional[RingAttnFunc] = field(
+    adam_beta3: Optional[float] = field(
         default=None,
         metadata={
-            "help": "The ring-flash-attn function to use in sequence parallelism"
+            "help": "The beta3 hyperparameter used in some optimizers such as CAME"
+        },
+    )
+    adam_epsilon2: Optional[float] = field(
+        default=None,
+        metadata={
+            "help": "The epsilon2 hyperparameter used in some optimizers such as CAME"
         },
     )
 

src/axolotl/integrations/base.py

@@ -10,227 +10,240 @@
 # License for the specific language governing permissions and limitations under
 # the License.
 
-"""
-Base class for all plugins.
+"""Base class for all plugins.
 
 A plugin is a reusable, modular, and self-contained piece of code that extends the functionality of Axolotl.
 Plugins can be used to integrate third-party models, modify the training process, or add new features.
 
 To create a new plugin, you need to inherit from the BasePlugin class and implement the required methods.
 """
+
+from __future__ import annotations
+
 import collections
 import importlib
 import logging
-from typing import OrderedDict
+from typing import TYPE_CHECKING, Callable, OrderedDict, Union
 
-import torch
+from peft import PeftModel
+from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LRScheduler
+from transformers import PreTrainedModel, Trainer
+
+from axolotl.utils.dict import DictDefault
+
+if TYPE_CHECKING:
+    from axolotl.common.datasets import TrainDatasetMeta
 
 
 class BasePlugin:
-    """
-    Base class for all plugins. Defines the interface for plugin methods.
-
-    Attributes:
-    None
+    """Base class for all plugins. Defines the interface for plugin methods.
 
     Methods:
-    register(cfg): Registers the plugin with the given configuration.
-    pre_model_load(cfg): Performs actions before the model is loaded.
-    post_model_build(cfg, model): Performs actions after the model is loaded, but before LoRA adapters are applied.
-    pre_lora_load(cfg, model): Performs actions before LoRA weights are loaded.
-    post_lora_load(cfg, model): Performs actions after LoRA weights are loaded.
-    post_model_load(cfg, model): Performs actions after the model is loaded, inclusive of any adapters.
-    create_optimizer(cfg, trainer): Creates and returns an optimizer for training.
-    create_lr_scheduler(cfg, trainer, optimizer, num_training_steps): Creates and returns a learning rate scheduler.
-    add_callbacks_pre_trainer(cfg, model): Adds callbacks to the trainer before training.
-    add_callbacks_post_trainer(cfg, trainer): Adds callbacks to the trainer after training.
+        register(cfg): Registers the plugin with the given configuration.
+        load_datasets(cfg): Loads and preprocesses the dataset for training.
+        pre_model_load(cfg): Performs actions before the model is loaded.
+        post_model_build(cfg, model): Performs actions after the model is loaded, but
+            before LoRA adapters are applied.
+        pre_lora_load(cfg, model): Performs actions before LoRA weights are loaded.
+        post_lora_load(cfg, model): Performs actions after LoRA weights are loaded.
+        post_model_load(cfg, model): Performs actions after the model is loaded,
+            inclusive of any adapters.
+        post_trainer_create(cfg, trainer): Performs actions after the trainer is
+            created.
+        create_optimizer(cfg, trainer): Creates and returns an optimizer for training.
+        create_lr_scheduler(cfg, trainer, optimizer, num_training_steps): Creates and
+            returns a learning rate scheduler.
+        add_callbacks_pre_trainer(cfg, model): Adds callbacks to the trainer before
+            training.
+        add_callbacks_post_trainer(cfg, trainer): Adds callbacks to the trainer after
+            training.
     """
 
     def __init__(self):
-        """
-        Initializes the BasePlugin.
-        """
+        """Initializes the BasePlugin."""
 
     def register(self, cfg):  # pylint: disable=unused-argument
-        """
-        Registers the plugin with the given configuration.
-
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-
-        Returns:
-        None
-        """
-
-    def get_input_args(self):
-        """
-        Returns a pydantic model for the plugin's input arguments.
-        """
-
-    def pre_model_load(self, cfg):  # pylint: disable=unused-argument
-        """
-        Performs actions before the model is loaded.
-
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-
-        Returns:
-        None
-        """
-
-    def post_model_build(self, cfg, model):  # pylint: disable=unused-argument
-        """
-        Performs actions after the model is built/loaded, but before any adapters are applied.
+        """Registers the plugin with the given configuration.
 
         Args:
-            cfg (dict): The configuration for the plugin.
+            cfg: The configuration for the plugin.
         """
 
-    def post_model_load(self, cfg, model):  # pylint: disable=unused-argument
-        """
-        Performs actions after the model is loaded.
+    def get_input_args(self) -> str | None:
+        """Returns a pydantic model for the plugin's input arguments."""
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+    def load_datasets(
+        self, cfg: DictDefault, preprocess: bool = False
+    ) -> Union["TrainDatasetMeta", None]:
+        """Loads and preprocesses the dataset for training.
+
+        Args:
+            cfg: The configuration for the plugin.
+            preprocess: Whether this is the preprocess step of the datasets.
 
         Returns:
-        None
+            dataset_meta: The metadata for the training dataset.
         """
 
-    def pre_lora_load(self, cfg, model):  # pylint: disable=unused-argument
-        """
-        Performs actions before LoRA weights are loaded.
+    def pre_model_load(self, cfg: DictDefault):  # pylint: disable=unused-argument
+        """Performs actions before the model is loaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg: The configuration for the plugin.
+        """
+
+    # pylint: disable=unused-argument
+    def post_model_build(self, cfg: DictDefault, model: PreTrainedModel):
+        """Performs actions after the model is built/loaded, but before any adapters are applied.
+
+        Args:
+            cfg: The configuration for the plugin.
+        """
+
+    # pylint: disable=unused-argument
+    def pre_lora_load(self, cfg: DictDefault, model: PreTrainedModel):
+        """Performs actions before LoRA weights are loaded.
+
+        Args:
+            cfg: The configuration for the plugin.
+            model: The loaded model.
+        """
+
+    # pylint: disable=unused-argument
+    def post_lora_load(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Performs actions after LoRA weights are loaded.
+
+        Args:
+            cfg: The configuration for the plugin.
+            model: The loaded model.
+        """
+
+    # pylint: disable=unused-argument
+    def post_model_load(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Performs actions after the model is loaded.
+
+        Args:
+            cfg: The configuration for the plugin.
+            model: The loaded model.
+        """
+
+    # pylint: disable=unused-argument
+    def get_trainer_cls(self, cfg: DictDefault) -> Trainer | None:
+        """Returns a custom class for the trainer.
+
+        Args:
+            cfg: The global axolotl configuration.
 
         Returns:
-        None
+            The first non-`None` trainer class returned by a plugin.
         """
 
-    def post_lora_load(self, cfg, model):  # pylint: disable=unused-argument
-        """
-        Performs actions after LoRA weights are loaded.
+    # pylint: disable=unused-argument
+    def post_trainer_create(self, cfg: DictDefault, trainer: Trainer):
+        """Performs actions after the trainer is created.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg: The configuration for the plugin.
+            trainer: The trainer object for training.
+        """
+
+    # pylint: disable=unused-argument
+    def create_optimizer(self, cfg: DictDefault, trainer: Trainer) -> Optimizer | None:
+        """Creates and returns an optimizer for training.
+
+        Args:
+            cfg: The configuration for the plugin.
+            trainer: The trainer object for training.
 
         Returns:
-        None
-        """
-
-    def get_trainer_cls(self, cfg):  # pylint: disable=unused-argument):
-        """
-        Returns a custom class for the trainer.
-
-        Parameters:
-        cfg (dict): The global axolotl configuration.
-
-        Returns:
-        class: The class for the trainer.
-        """
-
-    def create_optimizer(self, cfg, trainer):  # pylint: disable=unused-argument
-        """
-        Creates and returns an optimizer for training.
-
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        trainer (object): The trainer object for training.
-
-        Returns:
-        object: The created optimizer.
+            The created optimizer.
         """
 
+    # pylint: disable=unused-argument
     def create_lr_scheduler(
-        self, cfg, trainer, optimizer, num_training_steps
-    ) -> LRScheduler | None:  # pylint: disable=unused-argument
-        """
-        Creates and returns a learning rate scheduler.
+        self,
+        cfg: DictDefault,
+        trainer: Trainer,
+        optimizer: Optimizer,
+        num_training_steps: int,
+    ) -> LRScheduler | None:
+        """Creates and returns a learning rate scheduler.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        trainer (object): The trainer object for training.
-        optimizer (object): The optimizer for training.
-        num_training_steps (int): Total number of training steps
+        Args:
+            cfg: The configuration for the plugin.
+            trainer: The trainer object for training.
+            optimizer: The optimizer for training.
+            num_training_steps: Total number of training steps
 
         Returns:
-        object (LRScheduler): The created learning rate scheduler.
+            The created learning rate scheduler.
         """
 
-    def add_callbacks_pre_trainer(self, cfg, model):  # pylint: disable=unused-argument
-        """
-        setup callbacks before creating the trainer.
+    # pylint: disable=unused-argument
+    def add_callbacks_pre_trainer(
+        self, cfg: DictDefault, model: PreTrainedModel
+    ) -> list[Callable]:
+        """Set up callbacks before creating the trainer.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg: The configuration for the plugin.
+            model: The loaded model.
 
         Returns:
-        List[callable]: A list of callback functions to be added to the TrainingArgs
+            A list of callback functions to be added to the `TrainingArgs`.
         """
         return []
 
+    # pylint: disable=unused-argument
     def add_callbacks_post_trainer(
-        self, cfg, trainer
-    ):  # pylint: disable=unused-argument
-        """
-        Adds callbacks to the trainer after creating the trainer.
-        This is useful for callbacks that require access to the model or trainer.
+        self, cfg: DictDefault, trainer: Trainer
+    ) -> list[Callable]:
+        """Adds callbacks to the trainer after creating the trainer. This is useful for
+        callbacks that require access to the model or trainer.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        trainer (object): The trainer object for training.
+        Args:
+            cfg: The configuration for the plugin.
+            trainer: The trainer object for training.
 
         Returns:
-        List[callable]: A list of callback functions to be added
+            A list of callback functions to be added
         """
         return []
 
-    def post_train(self, cfg, model):  # pylint: disable=unused-argument
-        """
-        Performs actions after training is complete.
+    # pylint: disable=unused-argument
+    def post_train(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Performs actions after training is complete.
 
-        Parameters:
-        cfg (dict): The axolotl configuration
-        model (object): The loaded model.
-
-        Returns:
-        None
+        Args:
+            cfg: The axolotl configuration.
+            model: The loaded model.
         """
 
-    def post_train_unload(self, cfg):  # pylint: disable=unused-argument
-        """
-        Performs actions after training is complete and the model is unloaded.
+    def post_train_unload(self, cfg: DictDefault):  # pylint: disable=unused-argument
+        """Performs actions after training is complete and the model is unloaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-
-        Returns:
-        None
+        Args:
+            cfg: The configuration for the plugin.
         """
 
 
 def load_plugin(plugin_name: str) -> BasePlugin:
-    """
-    Loads a plugin based on the given plugin name.
+    """Loads a plugin based on the given plugin name.
 
-    The plugin name should be in the format "module_name.class_name".
-    This function splits the plugin name into module and class, imports the module,
-    retrieves the class from the module, and creates an instance of the class.
+    The plugin name should be in the format "module_name.class_name". This function
+    splits the plugin name into module and class, imports the module, retrieves the
+    class from the module, and creates an instance of the class.
 
-    Parameters:
-    plugin_name (str): The name of the plugin to be loaded. The name should be in the format "module_name.class_name".
+    Args:
+        plugin_name: The name of the plugin to be loaded. The name should be in the
+            format "module_name.class_name".
 
     Returns:
-    BasePlugin: An instance of the loaded plugin.
+        An instance of the loaded plugin.
 
     Raises:
-    ImportError: If the plugin module cannot be imported.
+        ImportError: If the plugin module cannot be imported.
     """
     # split the plugin name into module and class
     module_name, class_name = plugin_name.rsplit(".", 1)
@@ -256,28 +269,25 @@ def load_plugin(plugin_name: str) -> BasePlugin:
 
 
 class PluginManager:
-    """
-    The PluginManager class is responsible for loading and managing plugins.
-    It should be a singleton so it can be accessed from anywhere in the codebase.
+    """The `PluginManager` class is responsible for loading and managing plugins. It
+    should be a singleton so it can be accessed from anywhere in the codebase.
 
     Attributes:
-    plugins (List[BasePlugin]): A list of loaded plugins.
+        plugins: A list of loaded plugins.
 
     Methods:
-    get_instance(): Static method to get the singleton instance of PluginManager.
-    register(plugin_name: str): Registers a new plugin by its name.
-    pre_model_load(cfg): Calls the pre_model_load method of all registered plugins.
+        get_instance(): Static method to get the singleton instance of `PluginManager`.
+        register(plugin_name: str): Registers a new plugin by its name.
+        pre_model_load(cfg): Calls the pre_model_load method of all registered plugins.
     """
 
     plugins: OrderedDict[str, BasePlugin] = collections.OrderedDict()
 
-    _instance = None
-    _cfg = None
+    _instance: PluginManager | None = None
+    _cfg: DictDefault | None = None
 
     def __new__(cls):
-        """
-        Creates a new instance of PluginManager if it doesn't exist yet.
-        """
+        """Creates a new instance of PluginManager if it doesn't exist yet."""
         if cls._instance is None:
             cls._instance = super(PluginManager, cls).__new__(cls)
             cls._instance.plugins: OrderedDict[str, BasePlugin] = (
@@ -287,9 +297,8 @@ class PluginManager:
 
     @staticmethod
     def get_instance() -> "PluginManager":
-        """
-        Returns the singleton instance of PluginManager.
-        If the instance doesn't exist, it creates a new one.
+        """Returns the singleton instance of PluginManager. If the instance doesn't
+        exist, it creates a new one.
         """
         if PluginManager._instance is None:
             PluginManager()
@@ -304,17 +313,13 @@ class PluginManager:
         self._cfg = cfg
 
     def register(self, plugin_name: str):
-        """
-        Registers a new plugin by its name.
+        """Registers a new plugin by its name.
 
-        Parameters:
-        plugin_name (str): The name of the plugin to be registered.
-
-        Returns:
-        None
+        Args:
+            plugin_name: The name of the plugin to be registered.
 
         Raises:
-        ImportError: If the plugin module cannot be imported.
+            ImportError: If the plugin module cannot be imported.
         """
         try:
             logging.info(f"Attempting to load plugin: {plugin_name}")
@@ -324,12 +329,11 @@ class PluginManager:
         except ImportError:
             logging.error(f"Failed to load plugin: {plugin_name}")
 
-    def get_input_args(self):
-        """
-        Returns a list of Pydantic classes for all registered plugins' input arguments.'
+    def get_input_args(self) -> list[str]:
+        """Returns a list of Pydantic classes for all registered plugins' input arguments.'
 
         Returns:
-        list[str]: A list of Pydantic classes for all registered plugins' input arguments.'
+            A list of Pydantic classes for all registered plugins' input arguments.'
         """
         input_args = []
         for plugin in self.plugins.values():
@@ -338,83 +342,88 @@ class PluginManager:
                 input_args.append(input_args_from_plugin)
         return input_args
 
-    def pre_model_load(self, cfg):
-        """
-        Calls the pre_model_load method of all registered plugins.
+    def load_datasets(
+        self, cfg: DictDefault, preprocess: bool = False
+    ) -> Union["TrainDatasetMeta", None]:
+        """Calls the load_datasets method of each registered plugin.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
+        Args:
+            cfg: The configuration for the plugins.
+            preprocess: Whether this is preprocess step of the datasets.
 
         Returns:
-        None
+            The dataset metadata loaded from all registered plugins.
+        """
+        return_ds_meta = None
+        for plugin in self.plugins.values():
+            dataset_meta = plugin.load_datasets(cfg, preprocess)
+            if dataset_meta is not None:
+                if return_ds_meta is None:
+                    return_ds_meta = dataset_meta
+                else:
+                    raise RuntimeError("Multiple plugins loaded datasets")
+        return return_ds_meta
+
+    def pre_model_load(self, cfg: DictDefault):
+        """Calls the pre_model_load method of all registered plugins.
+
+        Args:
+            cfg: The configuration for the plugins.
         """
         for plugin in self.plugins.values():
             plugin.pre_model_load(cfg)
 
-    def post_model_build(self, cfg, model):
-        """
-        Calls the post_model_build method of all registered plugins after the model has been built/loaded,
-        but before any adapters have been applied.
+    def post_model_build(self, cfg: DictDefault, model: PreTrainedModel):
+        """Calls the `post_model_build` method of all registered plugins after the
+        model has been built / loaded, but before any adapters have been applied.
 
         Args:
-            cfg (dict): The configuration for the plugins.
-            model (object): The loaded model.
+            cfg: The configuration for the plugins.
+            model: The loaded model.
         """
         for plugin in self.plugins.values():
             plugin.post_model_build(cfg, model)
 
-    def post_model_load(self, cfg, model):
-        """
-        Calls the post_model_load method of all registered plugins after the model has been loaded
-        inclusive of any adapters
+    def pre_lora_load(self, cfg: DictDefault, model: PreTrainedModel):
+        """Calls the `pre_lora_load` method of all registered plugins.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        model (object): The loaded model.
-
-        Returns:
-        None
-        """
-        for plugin in self.plugins.values():
-            plugin.post_model_load(cfg, model)
-
-    def pre_lora_load(self, cfg, model):
-        """
-        Calls the pre_lora_load method of all registered plugins.
-
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        model (object): The loaded model.
-
-        Returns:
-        None
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
         """
         for plugin in self.plugins.values():
             plugin.pre_lora_load(cfg, model)
 
-    def post_lora_load(self, cfg, model):
-        """
-        Calls the post_lora_load method of all registered plugins.
+    def post_lora_load(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Calls the `post_lora_load` method of all registered plugins.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        model (object): The loaded model.
-
-        Returns:
-        None
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
         """
         for plugin in self.plugins.values():
             plugin.post_lora_load(cfg, model)
 
-    def get_trainer_cls(self, cfg):
-        """
-        Calls the get_trainer_cls method of all registered plugins and returns the first non-None trainer class.
+    def post_model_load(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Calls the `post_model_load` method of all registered plugins after the model
+        has been loaded inclusive of any adapters.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
+        """
+        for plugin in self.plugins.values():
+            plugin.post_model_load(cfg, model)
+
+    def get_trainer_cls(self, cfg: DictDefault) -> Trainer | None:
+        """Calls the `get_trainer_cls` method of all registered plugins and returns the
+        first non-`None` trainer class.
+
+        Args:
+            cfg: The configuration for the plugins.
 
         Returns:
-        object: The trainer class, or None if none was found.
+            The first non-`None` trainer class returned by a plugin.
         """
         for plugin in self.plugins.values():
             trainer_cls = plugin.get_trainer_cls(cfg)
@@ -422,15 +431,25 @@ class PluginManager:
                 return trainer_cls
         return None
 
-    def create_optimizer(self, trainer):
-        """
-        Calls the create_optimizer method of all registered plugins and returns the first non-None optimizer.
+    def post_trainer_create(self, cfg: DictDefault, trainer: Trainer):
+        """Calls the `post_trainer_create` method of all registered plugins.
 
-        Parameters:
-        trainer (object): The trainer object for training.
+        Args:
+            cfg: The configuration for the plugins.
+            trainer: The trainer object for training.
+        """
+        for plugin in self.plugins.values():
+            plugin.post_trainer_create(cfg, trainer)
+
+    def create_optimizer(self, trainer: Trainer) -> Optimizer | None:
+        """Calls the `create_optimizer` method of all registered plugins and returns
+        the first non-`None` optimizer.
+
+        Args:
+            trainer: The trainer object for training.
 
         Returns:
-        object: The created optimizer, or None if none was found.
+            The created optimizer, or `None` if none was found.
         """
         for plugin in self.plugins.values():
             optimizer = plugin.create_optimizer(self.cfg, trainer)
@@ -439,17 +458,17 @@ class PluginManager:
         return None
 
     def create_lr_scheduler(
-        self, trainer, optimizer, num_training_steps
+        self, trainer: Trainer, optimizer: Optimizer, num_training_steps: int
     ) -> LRScheduler | None:
-        """
-        Calls the create_lr_scheduler method of all registered plugins and returns the first non-None scheduler.
+        """Calls the `create_lr_scheduler` method of all registered plugins and returns
+        the first non-`None` scheduler.
 
-        Parameters:
-        trainer (object): The trainer object for training.
-        optimizer (object): The optimizer for training.
+        Args:
+            trainer: The trainer object for training.
+            optimizer: The optimizer for training.
 
         Returns:
-        object: The created learning rate scheduler, or None if none was found.
+            The created learning rate scheduler, or `None` if not found.
         """
         for plugin in self.plugins.values():
             scheduler: LRScheduler | None = plugin.create_lr_scheduler(
@@ -462,16 +481,17 @@ class PluginManager:
                 return scheduler
         return None
 
-    def add_callbacks_pre_trainer(self, cfg, model):
-        """
-        Calls the add_callbacks_pre_trainer method of all registered plugins.
+    def add_callbacks_pre_trainer(
+        self, cfg: DictDefault, model: PreTrainedModel
+    ) -> list[Callable]:
+        """Calls the add_callbacks_pre_trainer method of all registered plugins.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        model (object): The loaded model.
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
 
         Returns:
-        List[callable]: A list of callback functions to be added to the TrainingArgs.
+            A list of callback functions to be added to the `TrainingArgs`.
         """
         callbacks = []
         for plugin in self.plugins.values():
@@ -480,16 +500,17 @@ class PluginManager:
                 callbacks.extend(plugin_callbacks)
         return callbacks
 
-    def add_callbacks_post_trainer(self, cfg, trainer):
-        """
-        Calls the add_callbacks_post_trainer method of all registered plugins.
+    def add_callbacks_post_trainer(
+        self, cfg: DictDefault, trainer: Trainer
+    ) -> list[Callable]:
+        """Calls the `add_callbacks_post_trainer` method of all registered plugins.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        trainer (object): The trainer object for training.
+        Args:
+            cfg: The configuration for the plugins.
+            trainer: The trainer object for training.
 
         Returns:
-        List[callable]: A list of callback functions to be added to the TrainingArgs.
+            A list of callback functions to be added to the `TrainingArgs`.
         """
         callbacks = []
         for plugin in self.plugins.values():
@@ -498,41 +519,31 @@ class PluginManager:
                 callbacks.extend(plugin_callbacks)
         return callbacks
 
-    def post_train(self, cfg, model):
-        """
-        Calls the post_train method of all registered plugins.
+    def post_train(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Calls the post_train method of all registered plugins.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        model (object): The loaded model.
-
-        Returns:
-        None
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
         """
         for plugin in self.plugins.values():
             plugin.post_train(cfg, model)
 
-    def post_train_unload(self, cfg):
-        """
-        Calls the post_train_unload method of all registered plugins.
+    def post_train_unload(self, cfg: DictDefault):
+        """Calls the post_train_unload method of all registered plugins.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        model (object): The loaded model.
-
-        Returns:
-        None
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
         """
         for plugin in self.plugins.values():
             plugin.post_train_unload(cfg)
 
 
 class BaseOptimizerFactory:
-    """
-    Base class for factories to create custom optimizers
-    """
+    """Base class for factories to create custom optimizers"""
 
     def __call__(
         self, opt_model, training_args, **optimizer_kwargs
-    ) -> "torch.optim.Optimizer":
+    ) -> Optimizer | None:
         pass

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -72,7 +72,7 @@ class CutCrossEntropyPlugin(BasePlugin):
         if cfg.cut_cross_entropy:
             self._check_requirements()
 
-            from .monkeypatch.patch import (
+            from axolotl.integrations.cut_cross_entropy.monkeypatch.patch import (
                 cce_patch,
             )
 

src/axolotl/integrations/cut_cross_entropy/monkeypatch/cohere.py

@@ -20,25 +20,15 @@ from cut_cross_entropy.transformers.utils import (
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.cohere.modeling_cohere import (
-    _CONFIG_FOR_DOC,
-    COHERE_INPUTS_DOCSTRING,
     KwargsForCausalLM,
 )
 from transformers.processing_utils import Unpack
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
 from transformers.utils.deprecation import deprecate_kwarg
 
 _PATCH_OPTS: PatchOptions | None = None
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(COHERE_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def cce_forward(
     self,
     input_ids: torch.LongTensor | None = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/gemma.py

@@ -17,25 +17,15 @@ from cut_cross_entropy.transformers.utils import (
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.gemma.modeling_gemma import (
-    _CONFIG_FOR_DOC,
-    GEMMA_INPUTS_DOCSTRING,
     KwargsForCausalLM,
 )
 from transformers.processing_utils import Unpack
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
 from transformers.utils.deprecation import deprecate_kwarg
 
 _PATCH_OPTS: PatchOptions | None = None
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(GEMMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def cce_forward(
     self,
     input_ids: torch.LongTensor | None = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/gemma3.py

@@ -20,15 +20,11 @@ from torch import nn
 from transformers.cache_utils import Cache, HybridCache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.gemma3.modeling_gemma3 import (
-    _CONFIG_FOR_DOC,
-    GEMMA3_INPUTS_DOCSTRING,
     Gemma3CausalLMOutputWithPast,
     logger,
 )
 from transformers.utils import (
-    add_start_docstrings_to_model_forward,
     is_torchdynamo_compiling,
-    replace_return_docstrings,
 )
 from transformers.utils.deprecation import deprecate_kwarg
 
@@ -38,10 +34,6 @@ _PATCH_OPTS: PatchOptions | None = None
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(GEMMA3_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def cce_forward(
     self,
     input_ids: torch.LongTensor | None = None,
@@ -170,10 +162,6 @@ def cce_forward(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(GEMMA3_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=Gemma3CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def cce_forward_multimodal(
     self,
     input_ids: torch.LongTensor | None = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/llama.py

@@ -19,15 +19,9 @@ from transformers.modeling_outputs import (
     CausalLMOutputWithPast,
 )
 from transformers.models.llama.modeling_llama import (
-    _CONFIG_FOR_DOC,
-    LLAMA_INPUTS_DOCSTRING,
     KwargsForCausalLM,
 )
 from transformers.processing_utils import Unpack
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
 from transformers.utils.deprecation import deprecate_kwarg
 from transformers.utils.generic import can_return_tuple
 
@@ -36,10 +30,6 @@ _PATCH_OPTS: PatchOptions | None = None
 
 @can_return_tuple
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def cce_forward(
     self,
     input_ids: Optional[torch.LongTensor] = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/llama4.py

@@ -16,22 +16,12 @@ from torch import nn
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.llama4.modeling_llama4 import (
-    _CONFIG_FOR_DOC,
-    LLAMA4_INPUTS_DOCSTRING,
     Llama4CausalLMOutputWithPast,
 )
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
 
 _PATCH_OPTS: PatchOptions | None = None
 
 
-@add_start_docstrings_to_model_forward(LLAMA4_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def cce_forward(
     self,
     input_ids: torch.LongTensor | None = None,
@@ -160,9 +150,6 @@ def cce_forward(
     )
 
 
-@replace_return_docstrings(
-    output_type=Llama4CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def cce_forward_multimodal(
     self,
     input_ids: torch.LongTensor | None = None,  # type: ignore

src/axolotl/integrations/cut_cross_entropy/monkeypatch/mistral3.py

@@ -19,15 +19,11 @@ from transformers.models.mistral3.modeling_mistral3 import (
     Mistral3CausalLMOutputWithPast,
 )
 from transformers.models.mistral.modeling_mistral import (
-    _CONFIG_FOR_DOC,
-    MISTRAL_INPUTS_DOCSTRING,
     KwargsForCausalLM,
 )
 from transformers.processing_utils import Unpack
 from transformers.utils import (
-    add_start_docstrings_to_model_forward,
     is_torchdynamo_compiling,
-    replace_return_docstrings,
 )
 from transformers.utils.deprecation import deprecate_kwarg
 
@@ -35,10 +31,6 @@ _PATCH_OPTS: PatchOptions | None = None
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def cce_forward(
     self,
     input_ids: torch.LongTensor | None = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen2_moe.py

@@ -13,16 +13,10 @@ from cut_cross_entropy.transformers.utils import (
     apply_lce,
 )
 from transformers.models.qwen2_moe.modeling_qwen2_moe import (
-    _CONFIG_FOR_DOC,
-    QWEN2MOE_INPUTS_DOCSTRING,
     MoeCausalLMOutputWithPast,
     MoeModelOutputWithPast,
     load_balancing_loss_func,
 )
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
 from transformers.utils.deprecation import deprecate_kwarg
 from transformers.utils.generic import can_return_tuple
 
@@ -31,10 +25,6 @@ _PATCH_OPTS: PatchOptions | None = None
 
 @can_return_tuple
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(QWEN2MOE_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def forward(
     self,
     input_ids: Optional[torch.LongTensor] = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen2_vl.py

@@ -14,22 +14,12 @@ from cut_cross_entropy.transformers.utils import (
 )
 from torch.nn import CrossEntropyLoss
 from transformers.models.qwen2_vl.modeling_qwen2_vl import (
-    _CONFIG_FOR_DOC,
-    QWEN2_VL_INPUTS_DOCSTRING,
     Qwen2VLCausalLMOutputWithPast,
 )
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
 
 _PATCH_OPTS: PatchOptions | None = None
 
 
-@add_start_docstrings_to_model_forward(QWEN2_VL_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=Qwen2VLCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def cce_forward_multimodal(
     self,
     input_ids: Optional[torch.LongTensor] = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen3_moe.py

@@ -12,20 +12,13 @@ from cut_cross_entropy.transformers.utils import (
     TransformersModelT,
     apply_lce,
 )
-from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.qwen3_moe.modeling_qwen3_moe import (
-    _CONFIG_FOR_DOC,
-    QWEN3_MOE_INPUTS_DOCSTRING,
     KwargsForCausalLM,
     MoeCausalLMOutputWithPast,
     MoeModelOutputWithPast,
     load_balancing_loss_func,
 )
 from transformers.processing_utils import Unpack
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
 from transformers.utils.deprecation import deprecate_kwarg
 from transformers.utils.generic import can_return_tuple
 
@@ -34,10 +27,6 @@ _PATCH_OPTS: PatchOptions | None = None
 
 @can_return_tuple
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(QWEN3_MOE_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def forward(
     self,
     input_ids: Optional[torch.LongTensor] = None,

src/axolotl/integrations/liger/models/deepseekv2.py

@@ -14,10 +14,6 @@ from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import CausalLMOutputWithPast
 
 
-# @add_start_docstrings_to_model_forward(DeepseekV2_INPUTS_DOCSTRING)
-# @replace_return_docstrings(
-#    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-# )
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

src/axolotl/integrations/liger/models/jamba.py

@@ -13,21 +13,11 @@ from liger_kernel.transformers.fused_linear_cross_entropy import (
 from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import MoeCausalLMOutputWithPast
 from transformers.models.jamba.modeling_jamba import (
-    _CONFIG_FOR_DOC,
-    JAMBA_INPUTS_DOCSTRING,
     HybridMambaAttentionDynamicCache,
     load_balancing_loss_func,
 )
-from transformers.utils import (
-    add_start_docstrings_to_model_forward,
-    replace_return_docstrings,
-)
 
 
-@add_start_docstrings_to_model_forward(JAMBA_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

src/axolotl/kernels/geglu.py

@@ -1,5 +1,4 @@
-"""
-Module for definition of GEGLU Triton kernels.
+"""Module for definition of GEGLU Triton kernels.
 
 See "GLU Variants Improve Transformer" (https://arxiv.org/abs/2002.05202).
 
@@ -12,8 +11,6 @@ import torch
 import triton
 import triton.language as tl
 
-SQRT_2_PI: tl.constexpr = 0.7978845608028654  # sqrt(2/π)
-
 
 @triton.jit
 def _geglu_fwd_kernel(

src/axolotl/loaders/__init__.py

@@ -0,0 +1,10 @@
+"""Init for axolotl.loaders module"""
+
+# pylint: disable=unused-import
+# flake8: noqa
+
+from .adapter import load_adapter, load_lora
+from .constants import MULTIMODAL_AUTO_MODEL_MAPPING
+from .model import ModelLoader
+from .processor import load_processor
+from .tokenizer import load_tokenizer

src/axolotl/loaders/adapter.py

@@ -0,0 +1,206 @@
+"""Adapter loading functionality, including LoRA / QLoRA and associated utils"""
+
+import logging
+import os
+import types
+from typing import Any
+
+import bitsandbytes as bnb
+import torch
+from bitsandbytes.nn import Params4bit
+from peft import (
+    AdaptionPromptConfig,
+    LoftQConfig,
+    LoraConfig,
+    PeftConfig,
+    PeftMixedModel,
+    PeftModel,
+    get_peft_model,
+)
+from transformers import PreTrainedModel
+
+from axolotl.loaders.utils import get_linear_embedding_layers
+from axolotl.utils.dict import DictDefault
+
+LOG = logging.getLogger(__name__)
+
+
+def setup_quantized_meta_for_peft(model: torch.nn.Module):
+    """Replaces `quant_state.to` with a dummy function to prevent PEFT from moving `quant_state` to meta device"""
+
+    def temp_to_method(self, *args, **kwargs):  # pylint: disable=unused-argument
+        return self
+
+    for param in model.parameters():
+        if isinstance(param, Params4bit):
+            param.quant_state._orig_to = (  # pylint: disable=protected-access
+                param.quant_state.to
+            )
+            param.quant_state.to = types.MethodType(temp_to_method, param.quant_state)
+
+
+def setup_quantized_peft_meta_for_training(model: torch.nn.Module):
+    """Replaces dummy `quant_state.to` method with the original function to allow training to continue"""
+    for param in model.parameters():
+        if isinstance(param, Params4bit) and hasattr(param.quant_state, "_orig_to"):
+            param.quant_state.to = (
+                param.quant_state._orig_to  # pylint: disable=protected-access
+            )
+            param.quant_state._orig_to = None  # pylint: disable=protected-access
+
+
+def find_all_linear_names(model):
+    cls = (bnb.nn.Linear4bit, bnb.nn.Linear8bitLt, torch.nn.Linear)
+    lora_module_names = set()
+    for name, module in model.named_modules():
+        if (
+            isinstance(module, cls)
+            or "Linear" in module.__class__.__name__
+            and module.__class__.__name__ not in ("LlamaLinearScalingRotaryEmbedding",)
+        ):
+            names = name.split(".")
+            lora_module_names.add(names[0] if len(names) == 1 else names[-1])
+
+    embedding_modules = get_linear_embedding_layers(model.config.model_type)
+    output_embedding = embedding_modules[1]
+    if output_embedding in lora_module_names:  # needed for 16-bit
+        lora_module_names.remove(output_embedding)
+
+    return list(lora_module_names)
+
+
+def load_lora(
+    model: PreTrainedModel,
+    cfg: DictDefault,
+    inference: bool = False,
+    config_only: bool = False,
+) -> tuple[PreTrainedModel | PeftModel | PeftMixedModel | None, PeftConfig | None]:
+    lora_target_modules = cfg.lora_target_modules or []
+
+    if cfg.lora_target_linear:
+        linear_names = find_all_linear_names(model)
+        LOG.info(f"found linear modules: {repr(sorted(linear_names))}")
+        lora_target_modules_as_list = (
+            lora_target_modules
+            if isinstance(lora_target_modules, list)
+            else [lora_target_modules]
+        )
+        lora_target_modules = list(set(lora_target_modules_as_list + linear_names))
+
+    lora_config_kwargs = {}
+    loftq_bits = cfg.peft and cfg.peft.loftq_config and cfg.peft.loftq_config.loftq_bits
+    if loftq_bits:
+        lora_config_kwargs["loftq_config"] = LoftQConfig(loftq_bits=loftq_bits)
+        lora_config_kwargs["init_lora_weights"] = "loftq"
+    if cfg.peft_init_lora_weights:
+        lora_config_kwargs["init_lora_weights"] = cfg.peft_init_lora_weights
+    if cfg.peft_use_dora:
+        lora_config_kwargs["use_dora"] = cfg.peft_use_dora
+        LOG.info("Initializing LoRA weights using dora. This might take longer.")
+    if cfg.peft_use_rslora:
+        lora_config_kwargs["use_rslora"] = cfg.peft_use_rslora
+    if cfg.peft_layer_replication:
+        lora_config_kwargs["layer_replication"] = cfg.peft_layer_replication
+
+    lora_config = LoraConfig(
+        r=cfg.lora_r,
+        lora_alpha=cfg.lora_alpha,
+        target_modules=lora_target_modules,
+        layers_to_transform=cfg.peft_layers_to_transform,
+        layers_pattern=cfg.peft_layers_pattern,
+        lora_dropout=cfg.lora_dropout,
+        fan_in_fan_out=cfg.lora_fan_in_fan_out,
+        modules_to_save=cfg.lora_modules_to_save if cfg.lora_modules_to_save else None,
+        bias="none",
+        task_type="CAUSAL_LM",
+        **lora_config_kwargs,
+    )
+
+    if config_only:
+        return None, lora_config
+
+    rank = int(os.environ.get("LOCAL_RANK", 0))
+
+    if (
+        cfg.fsdp
+        and cfg.adapter
+        and cfg.fsdp_config.fsdp_cpu_ram_efficient_loading
+        and rank != 0
+    ):
+        setup_quantized_meta_for_peft(model)
+
+    if cfg.lora_model_dir:
+        LOG.debug("Loading pretrained PEFT - LoRA")
+        model_kwargs: Any = {}
+        if cfg.lora_on_cpu:
+            model_kwargs["max_memory"] = {"cpu": "256GiB"}
+            model_kwargs["device_map"] = {"": "cpu"}
+        model = PeftModel.from_pretrained(
+            model,
+            cfg.lora_model_dir,
+            is_trainable=(not inference),
+            **model_kwargs,
+        )
+    else:
+        model = get_peft_model(model, lora_config)
+
+    if rank == 0:
+        try:
+            model.print_trainable_parameters()
+        except AttributeError as exc:
+            LOG.warning(
+                "Exception caught during model.print_trainable_parameters(): %s", exc
+            )
+    elif (
+        cfg.fsdp
+        and cfg.adapter
+        and cfg.fsdp_config.fsdp_cpu_ram_efficient_loading
+        and rank != 0
+    ):
+        setup_quantized_peft_meta_for_training(model)
+
+    return model, lora_config
+
+
+def load_adapter(
+    model: PreTrainedModel,
+    cfg: DictDefault,
+    adapter: str | None,
+    inference: bool = False,
+) -> tuple[PreTrainedModel | PeftModel | PeftMixedModel, PeftConfig | None]:
+    if adapter is None:
+        return model, None
+    if hasattr(model, "enable_input_require_grads"):
+        model.enable_input_require_grads()
+    if adapter in ["lora", "qlora"]:
+        peft_model, lora_config = load_lora(model, cfg, inference=inference)
+        return peft_model, lora_config
+    if adapter == "llama-adapter":
+        peft_model, lora_config = load_llama_adapter(model, cfg)
+        return peft_model, lora_config
+
+    raise NotImplementedError(f"{adapter} PEFT adapter not available")
+
+
+def load_llama_adapter(
+    model: PreTrainedModel, cfg: DictDefault
+) -> tuple[PeftModel | PeftMixedModel, PeftConfig]:
+    peft_config = AdaptionPromptConfig(
+        adapter_layers=cfg.peft_adapter.layers,  # layers (L)
+        adapter_len=cfg.peft_adapter.len,  # prompt length (K)
+        task_type="CAUSAL_LM",
+    )
+
+    if cfg.lora_model_dir:
+        LOG.debug("Loading pretrained PEFT - llama_adapter")
+        peft_model = PeftModel.from_pretrained(
+            model,
+            cfg.lora_model_dir,
+            torch_dtype=torch.float16,
+        )
+    else:
+        peft_model = get_peft_model(model, peft_config)
+
+    peft_model.print_trainable_parameters()
+
+    return peft_model, peft_config

src/axolotl/loaders/constants.py

@@ -0,0 +1,21 @@
+"""Shared constants for axolotl.loaders module"""
+
+from transformers import (
+    Gemma3ForConditionalGeneration,
+    Llama4ForConditionalGeneration,
+    LlavaForConditionalGeneration,
+    Mistral3ForConditionalGeneration,
+    MllamaForConditionalGeneration,
+    Qwen2_5_VLForConditionalGeneration,
+    Qwen2VLForConditionalGeneration,
+)
+
+MULTIMODAL_AUTO_MODEL_MAPPING = {
+    "mllama": MllamaForConditionalGeneration,
+    "llama4": Llama4ForConditionalGeneration,
+    "llava": LlavaForConditionalGeneration,
+    "qwen2_vl": Qwen2VLForConditionalGeneration,
+    "qwen2_5_vl": Qwen2_5_VLForConditionalGeneration,
+    "mistral3": Mistral3ForConditionalGeneration,
+    "gemma3": Gemma3ForConditionalGeneration,
+}

src/axolotl/loaders/model.py

@@ -0,0 +1,754 @@
+"""Model loader class implementation for loading, configuring, and patching various
+models.
+"""
+
+import gc
+import logging
+import math
+import os
+from functools import cached_property
+from importlib.util import find_spec
+from typing import Any
+
+import peft
+import torch
+import transformers
+import transformers.modeling_utils
+from accelerate import init_empty_weights
+from peft import PeftConfig, PeftMixedModel, PeftModel, prepare_model_for_kbit_training
+from transformers import (
+    AutoModelForCausalLM,
+    AutoModelForVision2Seq,
+    AwqConfig,
+    BitsAndBytesConfig,
+    GPTQConfig,
+    PreTrainedModel,
+    PreTrainedTokenizerBase,
+)
+from transformers.integrations.deepspeed import (
+    HfTrainerDeepSpeedConfig,
+    is_deepspeed_zero3_enabled,
+)
+
+from axolotl.common.architectures import MOE_ARCH_BLOCK
+from axolotl.integrations.base import PluginManager
+from axolotl.loaders.adapter import load_adapter, load_lora
+from axolotl.loaders.constants import MULTIMODAL_AUTO_MODEL_MAPPING
+from axolotl.loaders.patch_manager import PatchManager
+from axolotl.loaders.utils import (
+    get_linear_embedding_layers,
+    get_module_class_from_name,
+    load_model_config,
+)
+from axolotl.models.mamba import fix_mamba_attn_for_loss
+from axolotl.utils.bench import log_gpu_memory_usage
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.distributed import (
+    get_device_count,
+    get_device_type,
+)
+from axolotl.utils.model_shard_quant import load_sharded_model_quant
+from axolotl.utils.schemas.enums import RLType
+
+LOG = logging.getLogger(__name__)
+PLUGIN_MANAGER = PluginManager.get_instance()
+
+
+class ModelLoader:
+    """Manages model configuration, initialization and application of patches during
+    model loading.
+
+    This class orchestrates the entire process of loading a model from configuration to
+    final preparation. It handles device mapping, quantization, attention mechanisms,
+    adapter integration, and various optimizations.
+
+    The loading process includes:
+        - Loading and validating model configuration
+        - Applying monkey patches for optimizations / fixes
+        - Setting up device mapping (including multi-GPU configurations)
+        - Configuring quantization
+        - Setting attention mechanisms (Flash Attention, SDPA, etc.)
+        - Loading and initializing the model
+        - Applying adapters (LoRA, QLoRA, etc.)
+
+    Attributes:
+        model: The loaded model instance (available after load() is called).
+        model_kwargs: Dictionary of keyword arguments passed to model initialization.
+        base_model: Name or path of the base model to load.
+        model_type: Type of model to load (e.g., `AutoModelForCausalLM`).
+        model_config: Configuration object for the model.
+        auto_model_loader: class used for loading the model (default:
+            `AutoModelForCausalLM`).
+    """
+
+    def __init__(
+        self,
+        cfg: DictDefault,
+        tokenizer: PreTrainedTokenizerBase,
+        *,
+        inference: bool = False,
+        reference_model: bool = False,
+        **kwargs,  # pylint: disable=unused-argument
+    ):
+        """Initializes the ModelLoader.
+
+        Args:
+            cfg: Configuration dictionary with model and training settings.
+            tokenizer: Tokenizer instance associated with the model.
+            processor: Optional processor for multimodal models. Defaults to None.
+            inference: Whether the model is being loaded for inference mode. Defaults
+                to False.
+            reference_model: Whether this is a reference model (used in setups like DPO
+                training). Defaults to False.
+            **kwargs: Additional keyword arguments (ignored).
+        """
+        self.cfg = cfg
+        self.tokenizer = tokenizer
+        self.inference: bool = inference
+        self.reference_model: bool = reference_model
+
+        # Init model kwargs
+        self.model_kwargs: dict[str, Any] = {}
+        if cfg.overrides_of_model_kwargs:
+            for key, val in cfg.overrides_of_model_kwargs.items():
+                self.model_kwargs[key] = val
+
+        # Init model
+        self.model: PreTrainedModel | PeftModel | PeftMixedModel
+        self.base_model = cfg.base_model
+        self.model_type = cfg.type_of_model
+
+        # Init model config
+        self.model_config = load_model_config(cfg)
+        self.auto_model_loader = AutoModelForCausalLM  # pylint: disable=invalid-name
+
+        # Initialize the patch manager
+        self.patch_manager = PatchManager(
+            cfg=cfg,
+            model_config=self.model_config,
+            inference=inference,
+        )
+
+    @cached_property
+    def has_flash_attn(self) -> bool:
+        """Check if flash attention is installed."""
+        return find_spec("flash_attn") is not None
+
+    @cached_property
+    def qlora_fsdp(self):
+        """Property that determines if FSDP with QLoRA is enabled."""
+        return self.cfg.fsdp and self.cfg.adapter == "qlora"
+
+    def load(self) -> tuple[PreTrainedModel, PeftConfig | None]:
+        """Load and prepare the model with all configurations and patches.
+
+        Returns:
+            A tuple with the loaded model and its LoRA configuration (if applicable).
+        """
+        # Initial setup and patches
+        self.patch_manager.apply_pre_model_load_patches()
+        self._apply_pre_model_load_setup()
+
+        # Build the model
+        PLUGIN_MANAGER.pre_model_load(self.cfg)
+        skip_move_to_device = self._build_model()
+        PLUGIN_MANAGER.post_model_build(self.cfg, self.model)
+
+        # Post-build model configuration
+        self._apply_post_model_load_setup()
+
+        # Load adapters (LoRA, etc.)
+        PLUGIN_MANAGER.pre_lora_load(self.cfg, self.model)
+        lora_config = self._load_adapters()
+        PLUGIN_MANAGER.post_lora_load(self.cfg, self.model)
+
+        # Apply remaining patches and finalize
+        self._apply_post_lora_load_setup(skip_move_to_device)
+        self.patch_manager.apply_post_model_load_patches(self.model)
+        PLUGIN_MANAGER.post_model_load(self.cfg, self.model)
+
+        return self.model, lora_config
+
+    def _apply_pre_model_load_setup(self):
+        """Apply patches and setup configurations before model loading."""
+        self._set_auto_model_loader()
+        self._set_device_map_config()
+        if self.cfg.revision_of_model:
+            self.model_kwargs["revision"] = self.cfg.revision_of_model
+        self._set_quantization_config()
+        self._set_attention_config()
+
+    def _apply_post_model_load_setup(self):
+        """Configure the model after it has been loaded."""
+        # Handle PeftModel if needed
+        if (
+            isinstance(self.model, (peft.PeftModel, peft.PeftModelForCausalLM))
+            and not self.qlora_fsdp
+        ):
+            self.model = self.model.merge_and_unload()
+
+        self._resize_token_embeddings()
+        self._adjust_model_config()
+        self._log_memory_usage()
+        self._configure_embedding_dtypes()
+
+    def _resize_token_embeddings(self):
+        """Resize token embeddings if needed."""
+        embeddings_len = (
+            math.ceil(len(self.tokenizer) / 32) * 32
+            if self.cfg.resize_token_embeddings_to_32x
+            else len(self.tokenizer)
+        )
+        if hasattr(self.model, "get_input_embeddings") and (
+            self.model.get_input_embeddings().num_embeddings < embeddings_len
+            or (
+                self.model.get_input_embeddings().num_embeddings > embeddings_len
+                and self.cfg.shrink_embeddings
+            )
+        ):
+            resize_kwargs = {}
+            if self.cfg.mean_resizing_embeddings is not None and (
+                self.model_config.model_type != "llava"
+            ):
+                resize_kwargs["mean_resizing"] = self.cfg.mean_resizing_embeddings
+            self.model.resize_token_embeddings(embeddings_len, **resize_kwargs)
+        else:
+            self.model.tie_weights()
+
+    def _adjust_model_config(self):
+        if (
+            hasattr(self.model, "config")
+            and hasattr(self.model.config, "max_position_embeddings")
+            and self.model.config.max_position_embeddings
+            and self.cfg.sequence_len > self.model.config.max_position_embeddings
+        ):
+            LOG.warning(
+                "increasing model.config.max_position_embeddings from "
+                f"{self.model.config.max_position_embeddings} to {self.cfg.sequence_len}"
+            )
+            self.model.config.max_position_embeddings = self.cfg.sequence_len
+
+        if (
+            hasattr(self.model, "config")
+            and hasattr(self.model.config, "bos_token_id")
+            and self.model.config.bos_token_id
+            and self.model.config.bos_token_id != self.tokenizer.bos_token_id
+        ):
+            self.model.config.bos_token_id = self.tokenizer.bos_token_id
+
+        if (
+            hasattr(self.model, "config")
+            and hasattr(self.model.config, "eos_token_id")
+            and self.model.config.eos_token_id
+            and self.model.config.eos_token_id != self.tokenizer.eos_token_id
+        ):
+            self.model.config.eos_token_id = self.tokenizer.eos_token_id
+
+    def _log_memory_usage(self):
+        """Log device memory usage after model load."""
+        if hasattr(self.model, "device") and self.model.device.type in (
+            "cuda",
+            "mps",
+            "npu",
+        ):
+            log_gpu_memory_usage(LOG, "after model load", self.model.device)
+
+    def _configure_embedding_dtypes(self):
+        """Configure embedding module dtypes."""
+        # Get embedding modules
+        embedding_modules = get_linear_embedding_layers(self.cfg.model_config_type)
+
+        # Initial dtype conversion
+        if not self.cfg.fsdp:
+            # We don't run this during FSDP because this will leave mixed and bfloat16
+            # dtypes in the model which FSDP doesn't like
+            if self.cfg.load_in_4bit and self.cfg.embeddings_skip_upcast:
+                embedding_modules = []
+            self._convert_embedding_modules_dtype(
+                embedding_modules,
+                dist_dtype=torch.float32,
+                before_kbit_train_or_finetune=True,
+            )
+
+        # Handle DeepSpeed Zero3
+        if is_deepspeed_zero3_enabled():
+            self._set_z3_leaf_modules()
+
+        # Apply gradient checkpointing if needed
+        needs_fa2_dtype = self.cfg.adapter or self.cfg.fsdp
+        if self.cfg.adapter in ["lora", "qlora"]:
+            needs_fa2_dtype = True
+            if self.cfg.gradient_checkpointing:
+                self.model.gradient_checkpointing_enable(
+                    gradient_checkpointing_kwargs=self.cfg.gradient_checkpointing_kwargs
+                )
+
+        self._prepare_model_for_quantization()
+
+        # Convert dtypes if needed
+        should_convert = (
+            # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so
+            # we need to convert them back to fp16/bf16 for flash-attn compatibility.
+            (
+                (needs_fa2_dtype or self.cfg.flash_attention or self.cfg.flex_attention)
+                and not self.qlora_fsdp
+            )
+            # CCE requires embedding layers to be in fp16/bf16 for backward pass
+            or self.cfg.cut_cross_entropy
+        )
+
+        if should_convert:
+            LOG.info("Converting modules to %s", self.cfg.torch_dtype)
+            self._convert_embedding_modules_dtype(
+                embedding_modules=embedding_modules,
+                dist_dtype=self.cfg.torch_dtype,
+                before_kbit_train_or_finetune=False,
+            )
+
+    def _load_adapters(self) -> PeftConfig | None:
+        """Load LoRA or other adapters."""
+        # Load LoRA or adapter
+        lora_config = None
+        if not self.reference_model or self.cfg.lora_model_dir:
+            # If we're not loading the reference model, then we're loading the model
+            # for training. Then, the DPO trainer doesn't want the PEFT model loaded
+            # over it, it just wants the LoRA / PEFT config.
+            if (
+                self.cfg.adapter
+                and self.cfg.rl in [RLType.DPO, RLType.IPO, RLType.KTO]
+                and not self.cfg.merge_lora
+            ):
+                _, lora_config = load_lora(
+                    self.model, self.cfg, inference=False, config_only=True
+                )
+            else:
+                self.model, lora_config = load_adapter(
+                    self.model, self.cfg, self.cfg.adapter
+                )
+
+        return lora_config
+
+    def _apply_post_lora_load_setup(self, skip_move_to_device: bool):
+        """Apply final optimizations and patches."""
+        # Place model on accelerator
+        if (
+            self.cfg.ddp
+            and not self.cfg.load_in_8bit
+            and not (self.cfg.rl and self.cfg.load_in_4bit)
+            and not skip_move_to_device
+        ):
+            # TODO: validate this conditional
+            self.model.to(f"{str(get_device_type())}:{self.cfg.local_rank}")
+
+        if get_device_count() > 1 and int(os.getenv("WORLD_SIZE", "1")) == 1:
+            self.model.is_parallelizable = True
+            self.model.model_parallel = True
+
+        if not any(
+            param.requires_grad
+            for _, param in self.model.named_parameters(recurse=True)
+        ):
+            LOG.warning("There are no parameters that require gradient updates")
+
+        if self.cfg.flash_optimum:
+            from optimum.bettertransformer import BetterTransformer
+
+            self.model = BetterTransformer.transform(self.model)
+
+        if self.cfg.adapter is not None:
+            log_gpu_memory_usage(LOG, "after adapters", self.model.device)
+
+        for _ in range(3):
+            gc.collect()
+            torch.cuda.empty_cache()
+
+    def _set_auto_model_loader(self):
+        """Set `self.auto_model_loader`. Defaults to `transformers.AutoModelForCausalLM`
+        (set at `__init__`). When using a multimodal model, `self.auto_model_loader`
+        should be set according to the type of the model.
+        """
+        if self.cfg.is_multimodal:
+            self.auto_model_loader = MULTIMODAL_AUTO_MODEL_MAPPING.get(
+                self.model_config.model_type, AutoModelForVision2Seq
+            )
+
+    def _set_device_map_config(self):
+        """Setup `device_map` according to config"""
+        device_map = self.cfg.device_map
+        max_memory = self.cfg.max_memory
+
+        if self.cfg.gpu_memory_limit:
+            gpu_memory_limit = (
+                str(self.cfg.gpu_memory_limit) + "GiB"
+                if isinstance(self.cfg.gpu_memory_limit, int)
+                else self.cfg.gpu_memory_limit
+            )
+
+            max_memory = {}
+            num_device = get_device_count()
+            for i in range(num_device):
+                max_memory[i] = gpu_memory_limit
+            max_memory["cpu"] = "256GiB"  # something sufficiently large to fit anything
+
+        if max_memory is not None:
+            # Based on https://github.com/togethercomputer/OpenChatKit/blob/main/inference/bot.py
+            from accelerate import infer_auto_device_map
+
+            with init_empty_weights():
+                model_canvas = self.auto_model_loader.from_config(
+                    self.model_config,
+                    trust_remote_code=self.cfg.trust_remote_code or False,
+                )
+            model_canvas.tie_weights()
+            device_map = infer_auto_device_map(
+                model_canvas,
+                max_memory=max_memory,
+                dtype=self.cfg.torch_dtype,
+            )
+            # We can discard max_memory now as we have a device map set up
+            max_memory = None
+
+        self.model_kwargs["torch_dtype"] = self.cfg.torch_dtype
+
+        if not is_deepspeed_zero3_enabled():
+            self.model_kwargs["device_map"] = device_map
+
+            cur_device = get_device_type()
+            if "mps" in str(cur_device):
+                self.model_kwargs["device_map"] = "mps:0"
+            elif "npu" in str(cur_device):
+                self.model_kwargs["device_map"] = "npu:0"
+
+        # TODO: can we put the reference model on it's own gpu? I think we have to move
+        # logits around to calculate loss
+        # if cfg.rl:
+        #     if torch.cuda.device_count() > 1:
+        #         if reference_model:
+        #             model_kwargs["device_map"] = "cuda:" + str(
+        #                 torch.cuda.current_device() + 1
+        #             )
+        #         else:
+        #             model_kwargs["device_map"] = "cuda:" + str(torch.cuda.current_device())
+
+    def _set_quantization_config(self):
+        """Set up quantization config (bitsandbytes, awq, gptq, etc.)"""
+        self.model_kwargs["load_in_8bit"] = self.cfg.load_in_8bit
+        self.model_kwargs["load_in_4bit"] = self.cfg.load_in_4bit
+
+        if self.cfg.gptq:
+            if not hasattr(self.model_config, "quantization_config"):
+                LOG.warning(
+                    "model config does not contain quantization_config information"
+                )
+            else:
+                if self.cfg.gptq_disable_exllama is not None:
+                    self.model_config.quantization_config["disable_exllama"] = (
+                        self.cfg.gptq_disable_exllama
+                    )
+                self.model_kwargs["quantization_config"] = GPTQConfig(
+                    **self.model_config.quantization_config
+                )
+        if (
+            self.cfg.adapter in ["qlora", "lora"]
+            and hasattr(self.model_config, "quantization_config")
+            and self.model_config.quantization_config["quant_method"]
+            in ["gptq", "awq", "bitsandbytes"]
+        ):
+            if self.model_config.quantization_config["quant_method"] == "gptq":
+                self.model_kwargs["quantization_config"] = GPTQConfig(
+                    **self.model_config.quantization_config
+                )
+            elif self.model_config.quantization_config["quant_method"] == "awq":
+                self.model_kwargs["quantization_config"] = AwqConfig(
+                    **self.model_config.quantization_config
+                )
+            elif (
+                self.model_config.quantization_config["quant_method"] == "bitsandbytes"
+            ):
+                self.model_kwargs["quantization_config"] = BitsAndBytesConfig(
+                    **self.model_config.quantization_config
+                )
+        elif self.cfg.adapter == "qlora" and self.model_kwargs["load_in_4bit"]:
+            bnb_config = {
+                "load_in_4bit": True,
+                "llm_int8_threshold": 6.0,
+                "llm_int8_has_fp16_weight": False,
+                "bnb_4bit_compute_dtype": self.cfg.torch_dtype,
+                "bnb_4bit_use_double_quant": True,
+                "bnb_4bit_quant_type": "nf4",
+                "bnb_4bit_quant_storage": torch.bfloat16,
+            }
+            if self.cfg.model_config_type in ["jamba", "qwen2_moe"] and not (
+                self.cfg.deepspeed or self.cfg.fsdp
+            ):
+                # for some reason, this causes the loss to be off by an order of magnitude
+                # but deepspeed needs this still in bfloat16
+                bnb_config["bnb_4bit_quant_storage"] = torch.float32
+
+            if self.cfg.bnb_config_kwargs:
+                bnb_config.update(self.cfg.bnb_config_kwargs)
+
+            self.model_kwargs["quantization_config"] = BitsAndBytesConfig(
+                **bnb_config,
+            )
+        elif self.cfg.adapter == "lora" and self.model_kwargs["load_in_8bit"]:
+            bnb_config = {
+                "load_in_8bit": True,
+            }
+            # Exclude mamba blocks from int8 quantization for jamba
+            if self.cfg.model_config_type == "jamba":
+                bnb_config["llm_int8_skip_modules"] = ["mamba"]
+            self.model_kwargs["quantization_config"] = BitsAndBytesConfig(
+                **bnb_config,
+            )
+
+        # no longer needed per https://github.com/huggingface/transformers/pull/26610
+        if "quantization_config" in self.model_kwargs or self.cfg.gptq:
+            self.model_kwargs.pop("load_in_8bit", None)
+            self.model_kwargs.pop("load_in_4bit", None)
+
+    def _set_attention_config(self):
+        """Sample packing uses custom FA2 patch"""
+        if self.cfg.flex_attention:
+            self.model_kwargs["attn_implementation"] = "flex_attention"
+            self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                "flex_attention"
+            )
+
+        elif self.cfg.flash_attention:
+            if not self.cfg.sample_packing and self.cfg.s2_attention:
+                pass
+            self.model_kwargs["attn_implementation"] = "flash_attention_2"
+            self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                "flash_attention_2"
+            )
+        elif self.cfg.sdp_attention:
+            self.model_kwargs["attn_implementation"] = "sdpa"
+            self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                "sdpa"
+            )
+        elif self.cfg.eager_attention:
+            self.model_kwargs["attn_implementation"] = "eager"
+            self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                "eager"
+            )
+
+        if self.cfg.low_cpu_mem_usage:
+            self.model_kwargs["low_cpu_mem_usage"] = True
+
+    def _configure_zero3_memory_efficient_loading(self):
+        """Set the deepspeed config to load the model into RAM first before moving
+        to VRAM.
+
+        We need to return `hf_ds_cfg` as it needs to exist before model loading.
+        """
+        hf_ds_cfg = None
+
+        if os.getenv("ACCELERATE_DEEPSPEED_ZERO_STAGE") == "3":
+            hf_ds_cfg = HfTrainerDeepSpeedConfig(self.cfg.deepspeed)
+            hf_ds_cfg.fill_match(
+                "train_micro_batch_size_per_gpu", self.cfg.micro_batch_size
+            )
+            hf_ds_cfg.fill_match(
+                "gradient_accumulation_steps", self.cfg.gradient_accumulation_steps
+            )
+            hf_ds_cfg.fill_match(
+                "train_batch_size",
+                int(os.getenv("WORLD_SIZE", "1"))
+                * self.cfg.micro_batch_size
+                * self.cfg.gradient_accumulation_steps,
+            )
+            if "device_map" in self.model_kwargs:
+                del self.model_kwargs["device_map"]
+
+            transformers.modeling_utils.is_deepspeed_zero3_enabled = lambda: True
+            transformers.integrations.deepspeed.is_deepspeed_zero3_enabled = (
+                lambda: True
+            )
+
+        return hf_ds_cfg
+
+    def _build_model(self) -> bool:
+        """Load model, with load strategy depending on config."""
+        skip_move_to_device = False
+        if (
+            self.qlora_fsdp
+            and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading
+            and (
+                self.cfg.model_config_type == "dbrx"
+                or self.cfg.qlora_sharded_model_loading
+            )
+        ):
+            quant_storage = self.cfg.torch_dtype
+            quantization_config = getattr(
+                self.model_config, "quantization_config", None
+            )
+            quantization_config = (
+                quantization_config or self.model_kwargs["quantization_config"]
+            )
+            self.model = load_sharded_model_quant(
+                self.base_model,
+                self.model_config,
+                self.cfg,
+                quant_storage=quant_storage,
+                quantization_config=quantization_config,
+            )
+            skip_move_to_device = True
+        elif (
+            self.model_config.model_type in ["llama", "llama4"]
+            and not self.cfg.trust_remote_code
+            and not self.cfg.gptq
+        ):
+            # TODO: Do we need to open this up for all models?
+            if self.cfg.fsdp and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading:
+                skip_move_to_device = True
+                if "device_map" in self.model_kwargs:
+                    del self.model_kwargs["device_map"]
+
+            self._configure_zero3_memory_efficient_loading()
+
+            # Load model with random initialization if specified
+            if self.cfg.random_init_weights:
+                # AutoModel classes support the from_config method
+                if self.auto_model_loader in [
+                    AutoModelForCausalLM,
+                    AutoModelForVision2Seq,
+                ]:
+                    self.model = self.auto_model_loader.from_config(
+                        config=self.model_config,
+                    )
+                else:
+                    self.model = self.auto_model_loader(config=self.model_config)
+            else:
+                self.model = self.auto_model_loader.from_pretrained(
+                    self.base_model,
+                    config=self.model_config,
+                    **self.model_kwargs,
+                )
+        elif self.model_type == "MambaLMHeadModel":
+            # FIXME this is janky at best and hacked together to make it work
+            MambaLMHeadModel = fix_mamba_attn_for_loss()  # pylint: disable=invalid-name
+
+            self.model_kwargs["dtype"] = self.model_kwargs["torch_dtype"]
+            self.model_kwargs["device"] = torch.cuda.current_device()
+            self.model_kwargs.pop("torch_dtype", None)
+            self.model_kwargs.pop("device_map", None)
+
+            self.model = MambaLMHeadModel.from_pretrained(
+                self.base_model,
+                **self.model_kwargs,
+            )
+        elif (
+            self.model_type
+            and self.model_type != "AutoModelForCausalLM"
+            and not self.cfg.trust_remote_code
+        ):
+            if self.cfg.gptq:
+                self.model = self.auto_model_loader.from_pretrained(
+                    self.base_model,
+                    config=self.model_config,
+                    trust_remote_code=self.cfg.trust_remote_code or False,
+                    **self.model_kwargs,
+                )
+            else:
+                self.model = getattr(transformers, self.model_type).from_pretrained(
+                    self.base_model,
+                    config=self.model_config,
+                    trust_remote_code=self.cfg.trust_remote_code or False,
+                    **self.model_kwargs,
+                )
+        else:
+            if self.cfg.gptq:
+                self.model = self.auto_model_loader.from_pretrained(
+                    self.base_model,
+                    config=self.model_config,
+                    trust_remote_code=self.cfg.trust_remote_code or False,
+                    **self.model_kwargs,
+                )
+            else:
+                if (
+                    self.cfg.fsdp
+                    and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading
+                ):
+                    # disabling either of these two still leads to VRAM spike before setting back down
+                    skip_move_to_device = True
+                    if "device_map" in self.model_kwargs:
+                        del self.model_kwargs["device_map"]
+
+                self._configure_zero3_memory_efficient_loading()
+
+                self.model = self.auto_model_loader.from_pretrained(
+                    self.base_model,
+                    config=self.model_config,
+                    trust_remote_code=self.cfg.trust_remote_code or False,
+                    **self.model_kwargs,
+                )
+        if is_deepspeed_zero3_enabled():
+            skip_move_to_device = True
+
+        return skip_move_to_device
+
+    def _set_z3_leaf_modules(self):
+        from deepspeed.utils import set_z3_leaf_modules
+
+        if self.cfg.model_config_type in MOE_ARCH_BLOCK:
+            moe_blocks = MOE_ARCH_BLOCK[self.cfg.model_config_type]
+            moe_blocks = [moe_blocks] if isinstance(moe_blocks, str) else moe_blocks
+            set_z3_leaf_modules(
+                self.model,
+                [
+                    get_module_class_from_name(self.model, module_name)
+                    for module_name in moe_blocks
+                ],
+            )
+
+    def _prepare_model_for_quantization(self):
+        """Prepare loaded model for quantization."""
+        skip_prepare_model_for_kbit_training = False
+        if self.cfg.model_config_type == "qwen" and self.cfg.adapter == "lora":
+            # Qwen doesn't play nicely with LoRA if this is enabled
+            skip_prepare_model_for_kbit_training = True
+
+        loftq_bits = (
+            self.cfg.peft
+            and self.cfg.peft.loftq_config
+            and self.cfg.peft.loftq_config.loftq_bits
+        )
+        if self.cfg.adapter == "lora" and loftq_bits:
+            skip_prepare_model_for_kbit_training = True
+
+        if (
+            self.qlora_fsdp
+            or (self.cfg.fsdp and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading)
+            or is_deepspeed_zero3_enabled()
+        ):
+            # Make sure everything is in the same dtype
+            skip_prepare_model_for_kbit_training = True
+
+        if (
+            not skip_prepare_model_for_kbit_training
+            and self.cfg.adapter in ["lora", "qlora"]
+            and (self.cfg.load_in_8bit or self.cfg.load_in_4bit)
+        ):
+            LOG.info("converting PEFT model w/ prepare_model_for_kbit_training")
+            self.model = prepare_model_for_kbit_training(
+                self.model, use_gradient_checkpointing=self.cfg.gradient_checkpointing
+            )
+
+    def _convert_embedding_modules_dtype(
+        self,
+        embedding_modules: list[str],
+        dist_dtype: torch.dtype,
+        before_kbit_train_or_finetune: bool,
+    ):
+        for name, module in self.model.named_modules():
+            if "norm" in name:
+                module.to(dist_dtype)
+            if before_kbit_train_or_finetune:
+                if name.endswith(".gate"):
+                    module.to(dist_dtype)
+                if self.model_config.model_type == "btlm":
+                    # don't upcast lm_head for btlm
+                    continue
+            if any(m in name for m in embedding_modules) and hasattr(module, "weight"):
+                module.to(dist_dtype)

src/axolotl/loaders/patch_manager.py

@@ -0,0 +1,380 @@
+"""Patch manager class implementation to complement `axolotl.loaders.ModelLoader`.
+
+Applies pre- and post-model load patches for various fixes and optimizations.
+"""
+
+import importlib.util
+import logging
+from functools import cached_property
+
+import addict
+import transformers
+from transformers import PretrainedConfig, PreTrainedModel
+
+from axolotl.integrations.base import PluginManager
+from axolotl.monkeypatch.multipack import (
+    SUPPORTED_MULTIPACK_MODEL_TYPES,
+    patch_for_multipack,
+)
+from axolotl.utils.dict import DictDefault
+
+LOG = logging.getLogger(__name__)
+PLUGIN_MANAGER = PluginManager.get_instance()
+
+
+class PatchManager:
+    """Manages the application of patches during the model loading process."""
+
+    def __init__(
+        self,
+        cfg: DictDefault,
+        model_config: PretrainedConfig | addict.Dict,
+        inference: bool = False,
+    ):
+        """Initialize the `PatchManager`.
+
+        Args:
+            cfg: Configuration dictionary with model and training settings.
+            model_config: Configuration object for the model.
+            inference: Whether the model is being loaded for inference mode.
+        """
+        self.cfg = cfg
+        self.model_config = model_config
+        self.inference = inference
+
+    @cached_property
+    def has_flash_attn(self) -> bool:
+        """Check if flash attention is installed."""
+        return importlib.util.find_spec("flash_attn") is not None
+
+    def apply_pre_model_load_patches(self):
+        """Apply pre-model load patches based on config."""
+        self._apply_flash_attention_patches()
+        self._apply_fsdp_patches()
+        self._apply_adapter_patches()
+        self._apply_flex_attention_patches()
+        self._apply_model_specific_patches()
+        self._apply_fp8_patches()
+        self._apply_flash_attention_peft_patches()
+        self._apply_gradient_checkpointing_patches()
+        self._patch_attention()
+        self._apply_multipack_patches()
+        self._patch_llama_derived_model()
+        self._apply_mistral_cross_entropy_patch()
+        self._apply_unsloth_self_attention_patch()
+
+    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)
+
+    def _apply_flash_attention_patches(self):
+        """Apply patches related to Flash Attention."""
+        if self.cfg.xformers_attention and self.cfg.sample_packing:
+            from axolotl.monkeypatch.attention import patch_xformers_attn_over_fa2
+
+            patch_xformers_attn_over_fa2()
+            self.cfg.flash_attention = True
+
+    def _apply_fsdp_patches(self):
+        """Apply patches for FSDP configurations."""
+        if self.cfg.fsdp_config and str(self.cfg.fsdp_config.fsdp_version) == "2":
+            from axolotl.monkeypatch.accelerate.fsdp2 import patch_accelerate_fsdp_utils
+
+            patch_accelerate_fsdp_utils()
+
+    def _apply_adapter_patches(self):
+        """Apply patches for adapter configurations."""
+        if self.cfg.adapter and self.cfg.embeddings_skip_upcast:
+            from axolotl.monkeypatch.peft.utils import patch_peft_prep_code
+
+            patch_peft_prep_code()
+
+    def _apply_flex_attention_patches(self):
+        """Apply patches for flexible attention."""
+        if self.cfg.flex_attention:
+            from axolotl.monkeypatch.attention.flex_attn import (
+                patch_flex_make_mask,
+                patch_flex_wrapper,
+            )
+
+            flex_attn_compile_kwargs = self.cfg.flex_attn_compile_kwargs or {}
+            patch_flex_wrapper(**flex_attn_compile_kwargs)
+            patch_flex_make_mask()
+
+    def _apply_model_specific_patches(self):
+        """Apply patches specific to model architectures."""
+        if (
+            self.cfg.model_config_type == "llama4"
+            and self.cfg.llama4_linearized_experts
+        ):
+            from axolotl.monkeypatch.models.llama4.modeling import (
+                patch_llama4_linearized_modeling,
+            )
+
+            patch_llama4_linearized_modeling()
+
+        if self.cfg.model_config_type == "gemma3":
+            from axolotl.monkeypatch.gemma3 import (
+                patch_gemma3conditionalgeneration_forward,
+            )
+
+            patch_gemma3conditionalgeneration_forward()
+
+    def _apply_fp8_patches(self):
+        """Apply patches for FP8 support."""
+        if self.cfg.fp8:
+            from axolotl.monkeypatch.trainer_accelerator_args import (
+                patch_create_accelerate_code_for_fp8,
+            )
+
+            patch_create_accelerate_code_for_fp8()
+
+    def _apply_flash_attention_peft_patches(self):
+        """Apply patches for Flash Attention with PEFT."""
+        if self.cfg.adapter:
+            from axolotl.monkeypatch.transformers_fa_utils import (
+                patch_fa_peft_integration,
+            )
+
+            patch_fa_peft_integration()
+
+    def _apply_gradient_checkpointing_patches(self):
+        """Apply patches for gradient checkpointing."""
+        if self.cfg.gradient_checkpointing in ["unsloth", "offload"]:
+            from axolotl.monkeypatch.gradient_checkpointing import (
+                hf_grad_checkpoint_offload_wrapper,
+            )
+
+            transformers.modeling_utils.checkpoint = hf_grad_checkpoint_offload_wrapper
+        if self.cfg.gradient_checkpointing == "offload_disk":
+            from axolotl.monkeypatch.gradient_checkpointing import (
+                hf_grad_checkpoint_disk_offload_wrapper,
+            )
+
+            transformers.modeling_utils.checkpoint = (
+                hf_grad_checkpoint_disk_offload_wrapper
+            )
+
+    def _apply_mistral_cross_entropy_patch(self):
+        """Apply Mistral cross entropy patch if configured."""
+        if (
+            self.cfg.model_config_type == "mistral"
+            and self.cfg.flash_attn_cross_entropy_loss
+        ):
+            from axolotl.monkeypatch.mistral_attn_hijack_flash import (
+                patch_mistral_cross_entropy,
+            )
+
+            patch_mistral_cross_entropy()
+
+    def _apply_unsloth_self_attention_patch(self):
+        """Apply Unsloth self-attention patches if configured."""
+        if self.cfg.unsloth_lora_qkv or self.cfg.unsloth_lora_o:
+            from axolotl.monkeypatch.lora_kernels import patch_self_attn_lora
+
+            patch_self_attn_lora(self.cfg)
+
+    def _apply_multipack_patches(self):
+        """Apply multipack patches if necessary."""
+        if (
+            self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES
+            and (self.cfg.flash_attention or self.cfg.flex_attention)
+            and self.cfg.sample_packing
+        ):
+            # Get automap config if it exists
+            auto_map_config = None
+            if isinstance(self.model_config, dict) and "auto_map" in self.model_config:
+                auto_map_config = self.model_config["auto_map"]
+            elif hasattr(self.model_config, "auto_map"):
+                auto_map_config = self.model_config.auto_map
+
+            # Determine if the model has remote code
+            if auto_map_config is not None:
+                has_remote_code = "AutoModelForCausalLM" in auto_map_config
+            else:
+                has_remote_code = False
+
+            if has_remote_code and self.cfg.trust_remote_code is False:
+                # If explicitly set in YAML, prefer that
+                has_remote_code = self.cfg.trust_remote_code
+
+            patch_for_multipack(
+                self.cfg.model_config_type,
+                model_name=self.cfg.base_model,
+                has_remote_code=has_remote_code,
+            )
+
+            if self.cfg.is_llama_derived_model:
+                self._patch_loss_llama()
+
+    def _patch_attention(self):
+        """Apply attention-specific patches based on model type."""
+        if not (self.cfg.flash_attention and hasattr(self.model_config, "model_type")):
+            return
+
+        if self.model_config.model_type == "mllama" and self.cfg.flash_attention:
+            from axolotl.monkeypatch.attention.mllama import patch_mllama
+
+            patch_mllama()
+
+        if self.model_config.model_type == "btlm":
+            from axolotl.monkeypatch.btlm_attn_hijack_flash import (
+                replace_btlm_attn_with_flash_attn,
+            )
+
+            replace_btlm_attn_with_flash_attn(self.cfg.base_model)
+
+        if self.model_config.model_type == "stablelm_epoch" and self.cfg.sample_packing:
+            from axolotl.monkeypatch.stablelm_attn_hijack_flash import (
+                replace_stablelm_attn_with_flash_attn,
+            )
+
+            replace_stablelm_attn_with_flash_attn(self.cfg.base_model)
+
+    def _patch_loss_llama(self):
+        """Patch loss functions and other optimizations for LLaMA models."""
+        if self.cfg.flash_attn_cross_entropy and self.has_flash_attn:
+            from axolotl.monkeypatch.llama_attn_hijack_flash import (
+                patch_fa_llama_cross_entropy,
+            )
+
+            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, packed=False):
+        """Apply Flash Attention patches for LLaMA models."""
+        from axolotl.monkeypatch.llama_attn_hijack_flash import (
+            replace_llama_attn_with_flash_attn,
+        )
+
+        if packed:
+            if self.cfg.device not in ["mps", "cpu"] and not self.inference:
+                LOG.info("patching with flash attention for sample packing")
+                replace_llama_attn_with_flash_attn(
+                    packed=True,
+                    cross_entropy=self.cfg.flash_attn_cross_entropy,
+                    rms_norm=self.cfg.flash_attn_rms_norm,
+                )
+        elif self.cfg.s2_attention:
+            LOG.info("patching w/ flash-enabled, shifted-sparse attention")
+            replace_llama_attn_with_flash_attn(
+                packed=False,
+                cross_entropy=self.cfg.flash_attn_cross_entropy,
+                rms_norm=self.cfg.flash_attn_rms_norm,
+                use_shifted_sparse_attn=True,
+            )
+        elif self.cfg.flash_attn_cross_entropy or self.cfg.flash_attn_rms_norm:
+            replace_llama_attn_with_flash_attn(
+                packed=False,
+                cross_entropy=self.cfg.flash_attn_cross_entropy,
+                rms_norm=self.cfg.flash_attn_rms_norm,
+            )
+
+    def _patch_llama_xformers_attention(self):
+        """Apply xformers attention patches for LLaMA models."""
+        from axolotl.monkeypatch.llama_attn_hijack_xformers import (
+            hijack_llama_attention,
+        )
+
+        LOG.info("Patching with xformers attention...")
+        hijack_llama_attention()
+
+    def _patch_llama_sample_packing(self):
+        """Apply sample packing patches for LLaMA models."""
+        from axolotl.monkeypatch.llama_patch_multipack import (
+            hijack_llama_prepare_4d_mask,
+        )
+
+        LOG.info("Patching llama _prepare_4d_causal_attention_mask*...")
+        hijack_llama_prepare_4d_mask()
+
+    def _patch_llama_derived_model(self):
+        """Modify all llama derived models in one block."""
+        if self.cfg.is_llama_derived_model and not (
+            self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES
+            and (self.cfg.flash_attention or self.cfg.flex_attention)
+            and self.cfg.sample_packing
+        ):
+            self._patch_loss_llama()
+
+            if self.cfg.flash_attention:
+                self._patch_llama_flash_attention(packed=self.cfg.sample_packing)
+            elif self.cfg.xformers_attention:
+                self._patch_llama_xformers_attention()
+            elif self.cfg.sample_packing:
+                self._patch_llama_sample_packing()
+            elif self.cfg.s2_attention:
+                raise NotImplementedError(
+                    "Shifted-sparse attention not currently implemented without flash attention."
+                )
+
+    def _apply_llama_flash_attn_patches(self, model):
+        """Apply LLaMA-specific flash attention patches."""
+        if (
+            self.model_config.model_type in ["llama", "llama4"]
+            and not self.cfg.trust_remote_code
+            and not self.cfg.gptq
+            and self.cfg.flash_attention
+            and not self.inference
+        ):
+            # TODO(MengqingCao): split these patches seperately
+            from axolotl.monkeypatch.llama_attn_hijack_flash import (
+                is_xformers_swiglu_available,
+                replace_llama_mlp_with_swiglu,
+                replace_llama_qkv_with_fused,
+            )
+
+            if self.cfg.flash_attn_fuse_mlp and is_xformers_swiglu_available():
+                LOG.info("Patching with SwiGLU...")
+                replace_llama_mlp_with_swiglu(model)
+
+            if self.cfg.flash_attn_fuse_qkv:
+                LOG.info("Patching with fused QKV...")
+                replace_llama_qkv_with_fused(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 (
+            self.cfg.lora_mlp_kernel
+            or self.cfg.lora_qkv_kernel
+            or self.cfg.lora_o_kernel
+        ):
+            from axolotl.monkeypatch.lora_kernels import apply_lora_kernel_patches
+
+            apply_lora_kernel_patches(model=model, cfg=self.cfg)

src/axolotl/loaders/processor.py

@@ -0,0 +1,56 @@
+"""Processor loading functionality for multi-modal models"""
+
+import logging
+from typing import Any
+
+import transformers
+from transformers import (
+    AutoProcessor,
+    PreTrainedTokenizerBase,
+)
+
+from axolotl.utils.dict import DictDefault
+
+LOG = logging.getLogger(__name__)
+
+
+def load_processor(cfg: DictDefault, tokenizer: PreTrainedTokenizerBase):
+    processor_kwargs: dict[str, Any] = {}  # Do we actually need this?
+
+    processor_cls = AutoProcessor
+    if cfg.processor_type:
+        processor_cls = getattr(transformers, cfg.processor_type)
+
+    processor = processor_cls.from_pretrained(
+        cfg.processor_config,
+        trust_remote_code=cfg.trust_remote_code or False,
+        tokenizer=tokenizer,
+        **processor_kwargs,
+    )
+
+    # Attempt to load image size from processor if available
+    if (
+        cfg.image_size is None
+        and hasattr(processor, "size")
+        and any(dim in processor.size for dim in ["width", "height"])
+    ):
+        im_width = None
+        im_height = None
+        if "width" in processor.size:
+            im_width = processor.size["width"]
+        if "height" in processor.size:
+            im_height = processor.size["height"]
+
+        # If both width and height are set, use a tuple
+        if im_width is not None and im_height is not None:
+            cfg.image_size = (im_width, im_height)
+        # If only width is set, use as integer
+        elif im_width is not None:
+            cfg.image_size = im_width
+        # If only height is set, use as integer
+        elif im_height is not None:
+            cfg.image_size = im_height
+
+        LOG.debug(f"Loaded image size: {cfg.image_size} from processor")
+
+    return processor

src/axolotl/loaders/tokenizer.py

@@ -0,0 +1,281 @@
+"""Tokenizer loading functionality and associated utils"""
+
+import json
+import logging
+import os
+
+import transformers
+from transformers import (
+    AddedToken,
+    AutoTokenizer,
+)
+
+from axolotl.integrations.base import PluginManager
+from axolotl.loaders.utils import get_linear_embedding_layers, load_model_config
+from axolotl.prompt_tokenizers import LLAMA_DEFAULT_EOS_TOKEN
+from axolotl.utils.chat_templates import get_chat_template_from_config
+from axolotl.utils.distributed import (
+    barrier,
+    is_local_main_process,
+    is_main_process,
+)
+
+LOG = logging.getLogger(__name__)
+PLUGIN_MANAGER = PluginManager.get_instance()
+
+
+def modify_tokenizer_files(
+    tokenizer_path: str, token_mappings: dict[int, str], output_dir: str
+) -> str:
+    """
+    Modify tokenizer files to replace added_tokens strings, save to output directory,
+    and return the path to the modified tokenizer.
+
+    This only works with reserved tokens that were added to the tokenizer, not tokens
+    already part of the vocab.
+
+    Args:
+        tokenizer_path: Path or name of the original tokenizer
+        token_mappings: Dict mapping {token_id (int): new_token_string}
+        output_dir: Directory to save the modified tokenizer
+
+    Returns:
+        Path to the modified tokenizer directory
+
+    Ref: https://github.com/huggingface/transformers/issues/27974#issuecomment-1854188941
+    """
+    # Create the tokenizer directory in output_dir if it doesn't exist
+    tokenizer_dir = os.path.join(output_dir, "tokenizer")
+    os.makedirs(tokenizer_dir, exist_ok=True)
+
+    if is_local_main_process():  # pylint: disable=too-many-nested-blocks
+        # Load the tokenizer
+        temp_tokenizer = AutoTokenizer.from_pretrained(tokenizer_path, use_fast=True)
+
+        # Save the tokenizer to the output directory
+        temp_tokenizer.save_pretrained(tokenizer_dir)
+
+        # Get the token IDs and map them to their new values
+        token_id_mappings = {
+            int(token_id): new_value for token_id, new_value in token_mappings.items()
+        }
+
+        # 1. Update tokenizer_config.json - added_tokens_decoder
+        config_path = os.path.join(tokenizer_dir, "tokenizer_config.json")
+        if os.path.exists(config_path):
+            with open(config_path, "r", encoding="utf-8") as f:
+                config_data = json.load(f)
+
+            # Update added_tokens_decoder
+            if "added_tokens_decoder" in config_data:
+                for token_id, new_value in token_id_mappings.items():
+                    token_id_str = str(token_id)
+                    if token_id_str in config_data["added_tokens_decoder"]:
+                        config_data["added_tokens_decoder"][token_id_str][
+                            "content"
+                        ] = new_value
+                    else:
+                        raise ValueError(
+                            f"Token ID {token_id_str} not found in added_tokens_decoder"
+                        )
+
+            # Write the updated config back
+            with open(config_path, "w", encoding="utf-8") as f:
+                json.dump(config_data, f, indent=2)
+
+        # 2. Update tokenizer.json - added_tokens
+        tokenizer_path = os.path.join(tokenizer_dir, "tokenizer.json")
+        if os.path.exists(tokenizer_path):
+            with open(tokenizer_path, "r", encoding="utf-8") as f:
+                tokenizer_data = json.load(f)
+
+            # Update added_tokens
+            if "added_tokens" in tokenizer_data:
+                for token_id, new_value in token_id_mappings.items():
+                    for i, token_entry in enumerate(tokenizer_data["added_tokens"]):
+                        if token_entry["id"] == token_id:
+                            tokenizer_data["added_tokens"][i]["content"] = new_value
+                            break
+                    else:
+                        # Reaching this section means the token_id was not found in tokenizer.json added_tokens
+                        raise ValueError(
+                            f"Token ID {token_id} not found in added_tokens"
+                        )
+            if "model" in tokenizer_data and "vocab" in tokenizer_data["model"]:
+                for token_id, new_value in token_id_mappings.items():
+                    for entry_val, entry_id in tokenizer_data["model"]["vocab"].items():
+                        if entry_id == token_id:
+                            del tokenizer_data["model"]["vocab"][entry_val]
+                            tokenizer_data["model"]["vocab"][new_value] = token_id
+                            break
+
+            # Write the updated tokenizer data back
+            with open(tokenizer_path, "w", encoding="utf-8") as f:
+                json.dump(tokenizer_data, f, indent=2)
+
+    barrier()
+    return tokenizer_dir
+
+
+def load_tokenizer(cfg):
+    """Load and configure the tokenizer based on the provided config."""
+    model_config = load_model_config(cfg)
+    tokenizer_kwargs = {}
+    use_fast = True  # this is the default
+
+    if cfg.tokenizer_use_fast is not None:
+        use_fast = cfg.tokenizer_use_fast
+    if cfg.tokenizer_legacy is not None:
+        # True is the default w/ https://github.com/huggingface/transformers/pull/25224
+        tokenizer_kwargs["legacy"] = cfg.tokenizer_legacy
+
+    tokenizer_cls = AutoTokenizer
+    if cfg.tokenizer_type:
+        tokenizer_cls = getattr(transformers, cfg.tokenizer_type)
+
+    # Set base tokenizer path
+    tokenizer_path = cfg.tokenizer_config
+
+    # Apply token string overrides if specified
+    if cfg.added_tokens_overrides:
+        # Modify tokenizer files and get path to modified tokenizer
+        tokenizer_path = modify_tokenizer_files(
+            tokenizer_path, cfg.added_tokens_overrides, output_dir=cfg.output_dir
+        )
+
+    tokenizer = tokenizer_cls.from_pretrained(
+        tokenizer_path,
+        trust_remote_code=cfg.trust_remote_code or False,
+        use_fast=use_fast,
+        **tokenizer_kwargs,
+    )
+
+    if (
+        tokenizer.__class__.__name__
+        in [
+            "LlamaTokenizer",
+            "LlamaTokenizerFast",
+            "CodeLlamaTokenizer",
+            "CodeLlamaTokenizerFast",
+        ]
+        and hasattr(tokenizer, "pad_token")
+        and not tokenizer.pad_token
+    ):
+        # set a pad_token, but use eos_token so we don't add a new token
+        tokenizer.pad_token = LLAMA_DEFAULT_EOS_TOKEN
+
+    if tokenizer.__class__.__name__ == "GPTNeoXTokenizerFast":
+        tokenizer.add_special_tokens({"pad_token": "[PAD]"})
+        os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
+    # Mistral's official FA implementation requires left padding
+    if cfg.is_mistral_derived_model and cfg.flash_attention and not cfg.sample_packing:
+        tokenizer.padding_side = "left"
+
+    # Qwen base only has single token, so we need to set the special tokens
+    if cfg.is_qwen_derived_model:
+        token_ids = ["bos_token_id", "eos_token_id", "pad_token_id", "unk_token_id"]
+        for attr_name in token_ids:
+            if getattr(tokenizer, attr_name) is None:
+                setattr(tokenizer, attr_name, tokenizer.eod_id)
+
+        token_names = ["bos_token", "eos_token", "pad_token", "unk_token"]
+        for attr_name in token_names:
+            if getattr(tokenizer, attr_name) is None:
+                setattr(tokenizer, attr_name, "<|endoftext|>")
+
+    additional_special_tokens = None
+    if cfg.special_tokens:
+        special_tokens = cfg.special_tokens.to_dict()
+        additional_special_tokens = special_tokens.pop(
+            "additional_special_tokens", None
+        )
+        lora_modules_to_save = get_linear_embedding_layers(model_config.model_type)
+        for k, val in special_tokens.items():
+            # check if new special token is not already in tokenizer and
+            # is adapter training to make sure lora_modules_to_save is set
+            # pylint: disable=too-many-boolean-expressions
+            if (
+                (getattr(tokenizer, k) is None or getattr(tokenizer, k) != val)
+                and (len(tokenizer.encode(val, add_special_tokens=False)) > 2)
+                and cfg.adapter
+                and (
+                    not cfg.lora_modules_to_save
+                    or not all(
+                        x in cfg.lora_modules_to_save for x in lora_modules_to_save
+                    )
+                )
+                and k != "pad_token"
+            ):
+                lora_modules_to_save = ", ".join(
+                    [f"`{x}`" for x in lora_modules_to_save]
+                )
+                raise ValueError(
+                    f"Please set lora_modules_to_save to [{lora_modules_to_save}] when using an adapter and changing the special tokens."
+                )
+
+            tokenizer.add_special_tokens(
+                {k: AddedToken(val, rstrip=False, lstrip=False, normalized=False)}
+            )
+
+        # If we add bos_token and eos_token, we need to update the post processor to
+        # handle them correctly.
+        # https://github.com/huggingface/transformers/pull/24132
+        bos_or_eos_in_special_tokens = (
+            "bos_token" in cfg.special_tokens and "eos_token" in cfg.special_tokens
+        )
+        if (
+            tokenizer.__class__.__name__
+            in (
+                "LlamaTokenizerFast",
+                "CodeLlamaTokenizerFast",
+            )
+            and bos_or_eos_in_special_tokens
+        ):
+            tokenizer.update_post_processor()
+
+    if cfg.tokens:
+        tokenizer.add_tokens(
+            [
+                AddedToken(token, rstrip=False, lstrip=False, normalized=False)
+                for token in cfg.tokens
+            ]
+        )
+
+    # Additional special tokens are a List, and need to be treated differently than regular special
+    # tokens. We add them after we have called `add_tokens` in case these additional special tokens
+    # are new tokens.
+    #
+    # Usage:
+    #
+    # ```py
+    # special_tokens:
+    #   additional_special_tokens: ["<|im_start|>", "<|im_end|>"]
+    # ```
+    if additional_special_tokens is not None:
+        tokenizer.add_special_tokens(
+            {"additional_special_tokens": additional_special_tokens}
+        )
+
+    if is_main_process(use_environ=True):
+        LOG.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
+        LOG.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
+        LOG.debug(f"PAD: {tokenizer.pad_token_id} / {tokenizer.pad_token}")
+        LOG.debug(f"UNK: {tokenizer.unk_token_id} / {tokenizer.unk_token}")
+
+    if cfg.chat_template:
+        chat_template_string = get_chat_template_from_config(
+            cfg=cfg,
+            tokenizer=tokenizer,
+        )
+        if cfg.default_system_message and cfg.chat_template == "chatml":
+            chat_template_string = chat_template_string.replace(
+                "You are a helpful assistant.", cfg.default_system_message
+            )
+
+        tokenizer.chat_template = chat_template_string
+    else:
+        LOG.info(
+            "No Chat template selected. Consider adding a chat template for easier inference."
+        )
+    return tokenizer

src/axolotl/loaders/utils.py

@@ -0,0 +1,211 @@
+"""Utilities for axolotl.loaders module"""
+
+import contextlib
+import logging
+from typing import Type
+
+import addict
+import torch
+from transformers import AutoConfig, PretrainedConfig, PreTrainedModel
+
+from axolotl.utils.dict import DictDefault
+
+LOG = logging.getLogger(__name__)
+
+
+def get_module_class_from_name(
+    module: torch.nn.Module, name: str
+) -> Type[torch.nn.Module] | None:
+    """Gets a class from a module by its name. Copied from `accelerate.utils.dataclasses`
+    (https://github.com/huggingface/accelerate/blob/main/src/accelerate/utils/dataclasses.py#L2805).
+
+    Args:
+        module: The module to get the class from.
+        name: The name of the class.
+
+    Returns:
+        The class type of the matching module, or `None` if no match is found.
+    """
+    modules_children = list(module.children())
+    if module.__class__.__name__ == name:
+        return module.__class__
+
+    if len(modules_children) == 0:
+        return None
+
+    for child_module in modules_children:
+        module_class = get_module_class_from_name(child_module, name)
+        if module_class is not None:
+            return module_class
+
+    return None
+
+
+def check_model_config(cfg: DictDefault, model_config: PretrainedConfig):
+    """Validates and adjusts model config based on `axolotl` config.
+
+    This function performs several important checks and adjustments:
+        - Disables model caching for better memory efficiency
+        - Handles multimodal model-specific configurations
+        - Validates quantization settings
+        - Ensures proper LoRA configuration when using adapters with new tokens
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+        model_config: The model's configuration object from `transformers`.
+
+    Raises:
+        ValueError: If a multimodal model lacks text configuration, if GPTQ settings
+            are inconsistent, or if LoRA `modules_to_save` is improperly configured
+            with new tokens.
+    """
+    if hasattr(model_config, "use_cache"):
+        model_config.use_cache = False
+
+    if cfg.is_multimodal:
+        # For multimodal configs, use_cache is set in the text_config
+        if hasattr(model_config, "get_text_config"):
+            text_config = model_config.get_text_config()
+            if hasattr(text_config, "use_cache"):
+                text_config.use_cache = False
+        else:
+            raise ValueError(
+                "No text config found for multimodal model. Please raise an Issue with model details."
+            )
+
+        # Check if image_size is not set and load image size from model config if available
+        if (
+            cfg.image_size is None
+            and hasattr(model_config, "vision_config")
+            and hasattr(model_config.vision_config, "image_size")
+        ):
+            cfg.image_size = model_config.vision_config.image_size
+            LOG.debug(f"Loaded image size: {cfg.image_size} from model config")
+
+    quant_config_exists = (
+        hasattr(model_config, "quantization_config")
+        and model_config.quantization_config
+    )
+
+    # Detect compressed-tensors config
+    is_compressed_tensors_config = (
+        quant_config_exists
+        and model_config.quantization_config.get("quant_method") == "compressed-tensors"
+    )
+
+    if is_compressed_tensors_config:
+        if model_config.quantization_config.get("config_groups"):
+            LOG.warning(
+                "Found `config_groups` in a compressed-tensors config. "
+                "QAT integration with llmcompressor is not tested."
+            )
+        # Skip further quant checks for compressed-tensors
+        return
+
+    quant_config_method_is_gptq = (
+        quant_config_exists
+        and "quant_method" in model_config.quantization_config
+        and model_config.quantization_config["quant_method"] == "gptq"
+    )
+
+    if cfg.gptq and not quant_config_method_is_gptq:
+        raise ValueError(
+            "model_config.quantization_config is not set or quant_method is not set to gptq. "
+            "Please make sure to point to a GPTQ model."
+        )
+
+    lora_modules_to_save = get_linear_embedding_layers(model_config.model_type)
+    if (
+        cfg.adapter
+        and cfg.tokens
+        and (
+            not cfg.lora_modules_to_save
+            or not all(x in cfg.lora_modules_to_save for x in lora_modules_to_save)
+        )
+    ):
+        lora_modules_to_save_joined = ", ".join(
+            map(lambda x: f"`{x}`", lora_modules_to_save)
+        )
+        raise ValueError(
+            "`lora_modules_to_save` not properly set when adding new tokens. "
+            f"Please include [{lora_modules_to_save_joined}] in `lora_modules_to_save`."
+        )
+
+
+def load_model_config(cfg: DictDefault) -> PretrainedConfig | addict.Dict:
+    """Loads and configures a model configuration from HuggingFace or local sources.
+
+    This function determines the appropriate model config source, loads it, applies any
+    necessary overrides, and validates it for compatibility with the `axolotl` config.
+
+    Args:
+        cfg: Dictionary mapping `axolotl` config keys to values.
+
+    Returns:
+        A configured model configuration object (`AutoConfig` instance), or a simple
+            dictionary configuration for special cases like Mamba models.
+
+    Raises:
+        ValueError: If configuration loading fails for reasons other than special cases
+            that are handled (e.g., Mamba models).
+    """
+    model_config_name = cfg.base_model_config or cfg.base_model
+    if not model_config_name and cfg.tokenizer_config:
+        model_config_name = cfg.tokenizer_config
+    trust_remote_code = cfg.trust_remote_code is True
+    config_kwargs = {}
+    if cfg.revision_of_model:
+        config_kwargs["revision"] = cfg.revision_of_model
+    if cfg.num_labels:
+        # num_labels is used to initialize classifier models
+        config_kwargs["num_labels"] = cfg.num_labels
+    try:
+        model_config = AutoConfig.from_pretrained(
+            model_config_name,
+            trust_remote_code=trust_remote_code,
+            **config_kwargs,
+        )
+    except ValueError as error:
+        if "mamba" in model_config_name:
+            return addict.Dict(
+                {
+                    "model_type": "mamba",
+                }
+            )
+        raise error
+
+    if cfg.overrides_of_model_config:
+        for key, val in cfg.overrides_of_model_config.items():
+            setattr(model_config, key, val)
+
+    check_model_config(cfg, model_config)
+
+    return model_config
+
+
+def ensure_dtype(model: PreTrainedModel, dtype: torch.dtype = torch.bfloat16):
+    """Ensures all modules in the model are converted to the specified data type."""
+    for name, module in model.named_modules():
+        weight_mismatch = False
+        with contextlib.suppress(AttributeError):
+            weight_mismatch = module.weight.dtype != dtype
+
+        bias_mismatch = False
+        with contextlib.suppress(AttributeError):
+            bias_mismatch = module.bias.dtype != dtype
+
+        if weight_mismatch:
+            print(f"Converting module {name}.weight: {module.weight.dtype} -> {dtype}")
+        if bias_mismatch:
+            print(f"Converting module {name}.bias: {module.bias.dtype} -> {dtype}")
+        if weight_mismatch or bias_mismatch:
+            module.to(dtype)
+
+
+def get_linear_embedding_layers(model_type: str) -> list[str]:
+    """Returns layer names of linear embeddings needed for LoRA based on model type."""
+    if model_type == "gpt_neox":
+        return ["embed_in", "embed_out"]
+    if model_type == "falcon":
+        return ["word_embeddings", "lm_head"]
+    return ["embed_tokens", "lm_head"]

src/axolotl/monkeypatch/attention/ring_attn/__init__.py

@@ -1,12 +0,0 @@
-"""Init for ring attention monkeypatch module"""
-
-# pylint: disable=unused-import
-# flake8: noqa
-
-from .patch import (
-    RingAttnFunc,
-    get_ring_attn_group,
-    register_ring_attn,
-    set_ring_attn_group,
-    update_ring_attn_params,
-)

src/axolotl/monkeypatch/attention/ring_attn/patch.py

@@ -1,147 +0,0 @@
-"""
-Ring attention group registration and flash attention patching.
-
-Make use of the `ring-flash-attn` (https://github.com/zhuzilin/ring-flash-attention)
-package, specifically the `hf_adapter.substitute_hf_flash_attn` function to patch in
-their sequence parallel version of Flash Attention 2.
-"""
-
-from enum import Enum
-
-import torch
-import torch.distributed as dist
-from accelerate.logging import get_logger
-
-from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
-
-LOG = get_logger(__name__)
-
-
-RING_ATTN_GROUP = None
-
-
-def get_ring_attn_group() -> dist.ProcessGroup:
-    """
-    Getter for ring attention group on this rank.
-
-    Returns:
-        The process group for ring attention for this rank.
-    """
-    return RING_ATTN_GROUP
-
-
-def set_ring_attn_group(ring_attn_group: dist.ProcessGroup | None):
-    """
-    Setter for ring attention group on this rank.
-
-    Args:
-        Process group for ring attention.
-    """
-    global RING_ATTN_GROUP  # pylint: disable=global-statement
-    RING_ATTN_GROUP = ring_attn_group
-
-
-class RingAttnFunc(str, Enum):
-    """Enum class for supported `ring-flash-attn` implementations"""
-
-    # VARLEN_RING = "varlen_ring"
-    # VARLEN_ZIGZAG = "varlen_zigzag"
-    VARLEN_LLAMA3 = "varlen_llama3"
-    BATCH_RING = "batch_ring"
-    BATCH_ZIGZAG = "batch_zigzag"
-    BATCH_STRIPE = "batch_stripe"
-
-
-def register_ring_attn(
-    sequence_parallel_degree: int,
-    heads_k_stride: int | None,
-    ring_attn_func: RingAttnFunc | None,
-):
-    """
-    Create ring attention group and substitute flash attn with ring flash attn.
-
-    Args:
-        sequence_parallel_degree: Sequence parallelism factor.
-        heads_k_stride: Sequence parallelism K head stride size. Passed
-            through to `ring_flash_attn.substitute_hf_flash_attn`.
-        ring_attn_func: `ring_flash_attn` ring attention implemention. If sample
-            packing is enabled, it must be a `varlen` function; otherwise, it must be a
-            `batch` function.
-    """
-    if get_ring_attn_group() is not None:
-        LOG.info("Ring attention already registered, exiting early...")
-        return
-
-    LOG.info(
-        "Enabling ring attention sequence parallelism: "
-        f"each sequence will be processed across {sequence_parallel_degree} GPUs"
-    )
-
-    rank = dist.get_rank()
-    world_size = dist.get_world_size()
-
-    assert sequence_parallel_degree <= world_size, (
-        f"sequence_parallel_degree ({sequence_parallel_degree}) "
-        f"must be less than or equal to world_size ({world_size})"
-    )
-    assert world_size % sequence_parallel_degree == 0, (
-        f"sequence_parallel_degree ({sequence_parallel_degree}) "
-        f"must evenly divide world_size ({world_size})"
-    )
-
-    # Assign ranks to sequence parallel groups
-    group_assignments = {}
-    for i in range(world_size // sequence_parallel_degree):
-        ring_attn_ranks = list(
-            range(
-                i * sequence_parallel_degree,
-                (i + 1) * sequence_parallel_degree,
-            )
-        )
-        group = dist.new_group(ranks=ring_attn_ranks, backend="nccl")
-
-        # Track which GPUs are in which groups
-        for r in ring_attn_ranks:
-            group_assignments[r] = i
-
-        if rank in ring_attn_ranks:
-            set_ring_attn_group(group)
-
-    # Log the GPU group assignments
-    if rank == 0:
-        LOG.info(f"Sequence parallel group assignments: {group_assignments}")
-
-    if ring_attn_func is RingAttnFunc.VARLEN_LLAMA3:
-        from ring_flash_attn import substitute_hf_flash_attn
-
-        substitute_hf_flash_attn(
-            process_group=get_ring_attn_group(), heads_k_stride=heads_k_stride or 1
-        )
-    elif ring_attn_func in [
-        RingAttnFunc.BATCH_RING,
-        RingAttnFunc.BATCH_ZIGZAG,
-        RingAttnFunc.BATCH_STRIPE,
-    ]:
-        from axolotl.monkeypatch.attention.ring_attn.adapters.batch import (
-            substitute_hf_flash_attn,
-        )
-
-        substitute_hf_flash_attn(
-            process_group=get_ring_attn_group(),
-            ring_attn_func=ring_attn_func,
-        )
-
-
-def update_ring_attn_params(position_ids: torch.Tensor | None):
-    """
-    Calculate the cumulative sequence lengths for the current forward pass and pass the
-    value to the substituted `ring_flash_attn`.
-
-    Args:
-        position_ids: Optional tensor of position IDs (for sample packed data).
-    """
-    from ring_flash_attn import update_ring_flash_attn_params
-
-    cu_seqlens, _ = get_cu_seqlens_from_pos_ids(position_ids)
-    cu_seqlens = cu_seqlens.squeeze().to(device=torch.cuda.current_device())
-    update_ring_flash_attn_params(cu_seqlens, get_ring_attn_group())

src/axolotl/monkeypatch/gemma3.py

@@ -7,24 +7,16 @@ from typing import Optional, Tuple, Union
 import torch
 from transformers.cache_utils import Cache
 from transformers.models.gemma3.modeling_gemma3 import (
-    _CONFIG_FOR_DOC,
-    GEMMA3_INPUTS_DOCSTRING,
     Gemma3CausalLMOutputWithPast,
     logger,
 )
 from transformers.utils import (
-    add_start_docstrings_to_model_forward,
     is_torchdynamo_compiling,
-    replace_return_docstrings,
 )
 from transformers.utils.deprecation import deprecate_kwarg
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
-@add_start_docstrings_to_model_forward(GEMMA3_INPUTS_DOCSTRING)
-@replace_return_docstrings(
-    output_type=Gemma3CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
-)
 def new_forward(
     self,
     input_ids: torch.LongTensor = None,

src/axolotl/monkeypatch/gradient_checkpointing/__init__.py

@@ -0,0 +1,63 @@
+"""custom checkpointing utils"""
+
+import importlib
+from functools import partial
+
+from packaging import version
+
+from axolotl.monkeypatch.gradient_checkpointing.offload_cpu import (
+    CPU_Offloaded_Gradient_Checkpointer,
+)
+from axolotl.monkeypatch.gradient_checkpointing.offload_disk import (
+    Disco,
+)
+
+transformers_version = version.parse(importlib.metadata.version("transformers"))
+if transformers_version > version.parse("4.51.3"):
+    from transformers.modeling_layers import GradientCheckpointingLayer
+
+    def uses_gc_layers(decoder_layer):
+        return isinstance(decoder_layer.func.__self__, GradientCheckpointingLayer)
+
+else:
+
+    def uses_gc_layers(_):
+        return False
+
+
+def hf_grad_checkpoint_offload_wrapper(
+    decoder_layer, *args, use_reentrant=None
+):  # pylint: disable=unused-argument
+    if uses_gc_layers(decoder_layer):
+        return CPU_Offloaded_Gradient_Checkpointer.apply(
+            decoder_layer,
+            *args,
+        )
+
+    return CPU_Offloaded_Gradient_Checkpointer.apply(
+        (
+            decoder_layer.func.__self__
+            if isinstance(decoder_layer, partial)
+            else decoder_layer.__self__
+        ),
+        *args,
+    )
+
+
+def hf_grad_checkpoint_disk_offload_wrapper(
+    decoder_layer, *args, use_reentrant=None
+):  # pylint: disable=unused-argument
+    if uses_gc_layers(decoder_layer):
+        return Disco.apply(
+            decoder_layer,
+            *args,
+        )
+
+    return Disco.apply(
+        (
+            decoder_layer.func.__self__
+            if isinstance(decoder_layer, partial)
+            else decoder_layer.__self__
+        ),
+        *args,
+    )

src/axolotl/monkeypatch/gradient_checkpointing/offload_cpu.py

@@ -1,4 +1,4 @@
-"""Unsloth checkpointing"""
+"""CPU offloaded checkpointing"""
 
 # Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
 #
@@ -26,7 +26,7 @@ else:
     torch_cuda_amp_custom_bwd = torch.amp.custom_bwd(device_type="cuda")
 
 
-class Unsloth_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
+class CPU_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
     torch.autograd.Function
 ):
     """

src/axolotl/monkeypatch/gradient_checkpointing/offload_disk.py

@@ -0,0 +1,531 @@
+"""
+DISCO - DIsk-based Storage and Checkpointing with Optimized prefetching
+"""
+
+# Copyright 2025 Axolotl AI. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import atexit
+import concurrent.futures
+import logging
+import os
+import queue
+import shutil
+import tempfile
+import threading
+import time
+import uuid
+from collections import deque
+from concurrent.futures import Future
+from typing import Dict
+
+import torch
+
+torch_cuda_amp_custom_fwd = torch.amp.custom_fwd(device_type="cuda")
+torch_cuda_amp_custom_bwd = torch.amp.custom_bwd(device_type="cuda")
+
+# Setup logger
+logger = logging.getLogger(__name__)
+
+
+class DiskOffloadManager:
+    """
+    Manages offloaded tensors and handles prefetching in a separate thread.
+    Includes synchronization to prevent race conditions.
+    """
+
+    def __init__(
+        self,
+        prefetch_size: int = 3,
+        prefetch_to_gpu: bool = True,
+        save_workers: int = 4,
+    ):
+        """
+        Args:
+            prefetch_size: Maximum number of tensors to prefetch in the background.
+            prefetch_to_gpu: Whether to prefetch tensors directly to GPU memory.
+            save_workers: Maximum number of concurrent save operations.
+        """
+        self.temp_dir = tempfile.mkdtemp(prefix="disco_")
+
+        # Track tensor paths and their status
+        self.tensor_paths: deque = deque()  # Ordered history of tensor paths (LIFO)
+        self.file_locks: Dict[str, threading.Lock] = (
+            {}
+        )  # Maps file_path -> threading.Lock()
+        # Maps file_path -> status ("saving", "ready", "prefetching", "loaded", "deleted")
+        self.file_status: Dict[str, str] = {}
+
+        self.max_prefetch = prefetch_size
+        self.prefetch_to_gpu = prefetch_to_gpu
+
+        # Thread synchronization
+        self.manager_lock = threading.RLock()  # Used for thread-safe operations
+
+        # Prefetch queue and cache
+        self.prefetch_queue: queue.Queue = queue.Queue()
+        self.prefetch_cache: Dict[str, torch.Tensor] = {}  # Maps file_path -> tensor
+
+        # Save queue and thread pool
+        self.save_queue: queue.Queue = queue.Queue()
+        self.save_pool = concurrent.futures.ThreadPoolExecutor(max_workers=save_workers)
+        self.save_futures: Dict[str, Future] = {}
+        self.save_semaphore = threading.Semaphore(
+            save_workers * 2
+        )  # Limit concurrent save operations
+
+        # Start prefetch worker thread
+        self.stop_event = threading.Event()
+        # start multiple threads for prefetching
+        self.prefetch_worker_count = 2
+        self.prefetch_workers = []
+        for _ in range(self.prefetch_worker_count):
+            worker = threading.Thread(target=self._prefetch_worker, daemon=True)
+            worker.start()
+            self.prefetch_workers.append(worker)
+
+        # Start save worker thread
+        self.save_worker = threading.Thread(target=self._save_worker, daemon=True)
+        self.save_worker.start()
+        self.idx = 0
+
+        atexit.register(self.cleanup)
+
+    def _save_worker(self):
+        """Background thread that processes the save queue"""
+        while not self.stop_event.is_set():
+            try:
+                save_item = self.save_queue.get(timeout=0.5)
+                if save_item is None:
+                    continue
+
+                tensor, file_path = save_item
+
+                # Submit the save task to the thread pool
+                future = self.save_pool.submit(
+                    self._save_tensor_to_disk, tensor, file_path
+                )
+                with self.manager_lock:
+                    self.save_futures[file_path] = future
+
+                self.save_queue.task_done()
+
+            except queue.Empty:
+                time.sleep(0.01)  # Small sleep to prevent CPU spinning
+                continue
+
+    def _save_tensor_to_disk(self, tensor: torch.Tensor, file_path: str):
+        """Actually save the tensor to disk"""
+        try:
+            # Save tensor to disk
+            cpu_tensor = tensor.detach().cpu()
+            torch.save(cpu_tensor, file_path)
+            del cpu_tensor
+
+            with self.manager_lock:
+                # Mark file as ready
+                self.file_status[file_path] = "ready"
+
+            # Release semaphore
+            self.save_semaphore.release()
+
+            return True
+        except FileNotFoundError as e:
+            logger.error(f"Error saving tensor to {file_path}: {e}")
+            with self.manager_lock:
+                self.file_status[file_path] = "error"
+
+            # Release semaphore
+            self.save_semaphore.release()
+
+            return False
+
+    def _prefetch_worker(self):
+        """Background thread that loads tensors from disk ahead of time"""
+        while not self.stop_event.is_set():
+            try:
+                file_path = self.prefetch_queue.get(timeout=0.5)
+                if file_path is None:
+                    continue
+
+                # Check if file is available and not already in cache
+                with self.manager_lock:
+                    if (
+                        file_path not in self.file_status
+                        or self.file_status[file_path] == "deleted"
+                    ):
+                        self.prefetch_queue.task_done()
+                    if file_path in self.prefetch_cache:
+                        self.prefetch_queue.task_done()
+                        continue
+
+                    # If file is still being saved, wait for it
+                    if (
+                        self.file_status[file_path] == "saving"
+                        and file_path in self.save_futures
+                    ):
+                        # Re-queue this prefetch request with a little delay
+                        self.prefetch_queue.task_done()
+                        time.sleep(0.1)
+                        self.prefetch_queue.put(file_path)
+                        continue
+
+                    # Mark file as being prefetched
+                    self.file_status[file_path] = "prefetching"
+
+                # Load tensor from disk and store in cache
+                try:
+                    if os.path.exists(file_path):
+                        if self.prefetch_to_gpu:
+                            tensor = torch.load(
+                                file_path,
+                                map_location=torch.device("cuda"),
+                                weights_only=True,
+                            )
+                        else:
+                            tensor = torch.load(file_path, weights_only=True)
+
+                        with self.manager_lock:
+                            self.prefetch_cache[file_path] = tensor
+                            self.file_status[file_path] = "ready"
+                    else:
+                        with self.manager_lock:
+                            if self.file_status.get(file_path) != "deleted":
+                                logger.warning(
+                                    f"Prefetch error: File not found {file_path}"
+                                )
+                                self.file_status[file_path] = "missing"
+
+                except FileNotFoundError as e:
+                    with self.manager_lock:
+                        if self.file_status.get(file_path) != "deleted":
+                            logger.warning(f"Prefetch error for {file_path}: {e}")
+                            self.file_status[file_path] = "error"
+
+                self.prefetch_queue.task_done()
+
+            except queue.Empty:
+                time.sleep(0.01)  # Small sleep to prevent CPU spinning
+                continue
+
+    def save_tensor(self, tensor: torch.Tensor):
+        """Save tensor to disk asynchronously and return file path with thread-safe operations"""
+        # Generate unique file path
+        self.idx += 1
+        file_path: str = os.path.join(
+            self.temp_dir, f"{self.idx:06d}-{uuid.uuid4()}.pt"
+        )
+
+        with self.manager_lock:
+            # Mark file as being saved
+            self.file_locks[file_path] = threading.Lock()
+            self.file_status[file_path] = "saving"
+            # Add to history
+            self.tensor_paths.append(file_path)
+
+        # Acquire semaphore to limit concurrent save operations
+        self.save_semaphore.acquire()  # pylint: disable=consider-using-with
+        # Queue tensor for saving in background
+        self.save_queue.put((tensor.detach(), file_path))
+
+        return file_path
+
+    def wait_for_save(self, file_path, timeout=None) -> None:
+        """Wait for a tensor to be saved to disk"""
+        start_time = time.time()
+        while timeout is None or time.time() - start_time < timeout:
+            with self.manager_lock:
+                if self.file_status.get(file_path) == "ready":
+                    return
+                if self.file_status.get(file_path) in ["error", "missing", "deleted"]:
+                    return
+
+                if file_path in self.save_futures:
+                    future = self.save_futures[file_path]
+                    if future.done():
+                        return
+
+            # Small sleep to prevent CPU spinning
+            time.sleep(0.01)
+
+        # Timeout
+        logger.warning(f"Timeout waiting for tensor to be saved: {file_path}")
+        return
+
+    def load_tensor(self, file_path, target_device="cuda"):
+        """Load tensor from disk or prefetch cache with proper synchronization"""
+        # Wait for tensor to be saved if it's still in progress
+        self.wait_for_save(file_path)
+
+        tensor = None
+
+        # Try to get from cache first
+        with self.manager_lock:
+            # Check if tensor is already in cache
+            if file_path in self.prefetch_cache:
+                tensor = self.prefetch_cache[file_path]
+                del self.prefetch_cache[file_path]
+                self.file_status[file_path] = "loaded"
+
+        if tensor is not None:
+            # Ensure tensor is on correct device
+            if target_device != "cpu" and tensor.device.type == "cpu":
+                tensor = tensor.to(target_device, non_blocking=True)
+            return tensor
+
+        # If not in cache, load directly from disk
+        try:
+            if not os.path.exists(file_path):
+                logger.error(f"File not found for loading: {file_path}")
+                raise FileNotFoundError(f"File not found: {file_path}")
+
+            tensor = torch.load(file_path, weights_only=True)
+
+            with self.manager_lock:
+                self.file_status[file_path] = "loaded"
+
+            if target_device != "cpu":
+                tensor = tensor.to(target_device, non_blocking=True)
+
+            return tensor
+
+        except Exception as e:
+            logger.error(f"Error loading tensor from {file_path}: {e}")
+            raise
+
+    def _safe_delete_file(self, file_path):
+        """Safely delete a file with proper synchronization"""
+        with self.manager_lock:
+            # Make sure any save operation is completed
+            if file_path in self.save_futures:
+                future = self.save_futures[file_path]
+                try:
+                    if not future.done():
+                        future.cancel()
+                    del self.save_futures[file_path]
+                except FileNotFoundError as e:
+                    logger.warning(
+                        f"Error canceling save operation for {file_path}: {e}"
+                    )
+
+            # Only delete if file exists and is not being prefetched
+            status = self.file_status.get(file_path)
+            if status in ["ready", "loaded", "error", "missing"]:
+                try:
+                    if os.path.exists(file_path):
+                        os.remove(file_path)
+                    self.file_status[file_path] = "deleted"
+                    return True
+                except FileNotFoundError as e:
+                    logger.warning(f"Error deleting file {file_path}: {e}")
+            return False
+
+    def trigger_prefetch(self, n=None):
+        """Trigger prefetching of the next N tensors with proper synchronization"""
+        if n is None:
+            n = self.max_prefetch
+
+        prefetch_paths = []
+        with self.manager_lock:
+            # Find files that are ready to be prefetched (not already in cache or being prefetched)
+            for path in reversed(self.tensor_paths):
+                if (
+                    path not in self.prefetch_cache
+                    and self.file_status.get(path) == "ready"
+                ):
+                    prefetch_paths.append(path)
+                    if len(prefetch_paths) >= n:
+                        break
+
+        # Queue files for prefetching
+        for path in prefetch_paths:
+            self.prefetch_queue.put(path)
+
+    def cleanup_tensor(self, file_path: str):
+        """Clean up a specific tensor file after it's been used"""
+        with self.manager_lock:
+            if file_path in self.tensor_paths:
+                self.tensor_paths.remove(file_path)
+
+            # Remove from prefetch cache if present
+            if file_path in self.prefetch_cache:
+                del self.prefetch_cache[file_path]
+
+            # Remove from save futures if present
+            if file_path in self.save_futures:
+                future = self.save_futures[file_path]
+                if not future.done():
+                    future.cancel()
+                del self.save_futures[file_path]
+
+        # Try to delete the file
+        self._safe_delete_file(file_path)
+
+    def cleanup(self):
+        """Clean up all temp files and stop prefetch thread with proper synchronization"""
+        self.stop_event.set()
+
+        # Cancel all pending save operations
+        with self.manager_lock:
+            for _, future in self.save_futures.items():
+                if not future.done():
+                    future.cancel()
+            self.save_futures.clear()
+
+        # Drain the save queue
+        while not self.save_queue.empty():
+            try:
+                self.save_queue.get_nowait()
+                self.save_queue.task_done()
+            except queue.Empty:
+                break
+
+        # Shutdown the save pool
+        self.save_pool.shutdown(wait=False)
+
+        # Join the save worker thread
+        if self.save_worker.is_alive():
+            self.save_worker.join(timeout=2.0)
+
+        # Join the prefetch worker threads
+        for thread in self.prefetch_workers:
+            if thread.is_alive():
+                thread.join(timeout=2.0)
+
+        # Clear cache and remove all temporary files
+        with self.manager_lock:
+            self.prefetch_cache.clear()
+            paths_to_delete = list(self.tensor_paths)
+            self.tensor_paths.clear()
+
+        # Delete all temporary files
+        for path in paths_to_delete:
+            self._safe_delete_file(path)
+
+        # Remove temp directory
+        try:
+            if os.path.exists(self.temp_dir):
+                shutil.rmtree(self.temp_dir, ignore_errors=True)
+        except FileNotFoundError as e:
+            logger.warning(f"Error removing temporary directory {self.temp_dir}: {e}")
+
+
+class Disco(torch.autograd.Function):
+    """
+    Disco: DIsk-based Storage and Checkpointing with Optimized prefetching
+    Advanced disk-based gradient checkpointer with prefetching.
+    """
+
+    # Shared manager instance across all checkpointing operations
+    _manager = None
+
+    @staticmethod
+    def get_instance(prefetch_size=1, prefetch_to_gpu=True, save_workers=4):
+        """Get or create the offload manager"""
+        if Disco._manager is None:
+            Disco._manager = DiskOffloadManager(
+                prefetch_size=prefetch_size,
+                prefetch_to_gpu=prefetch_to_gpu,
+                save_workers=save_workers,
+            )
+        return Disco._manager
+
+    @staticmethod
+    @torch_cuda_amp_custom_fwd
+    def forward(
+        ctx,
+        forward_function,
+        hidden_states,
+        *args,
+        prefetch_size=1,
+        prefetch_to_gpu=True,
+        save_workers=4,
+    ):
+        """Forward pass that offloads activations to disk asynchronously"""
+        # Get or create the manager
+        manager = Disco.get_instance(
+            prefetch_size=prefetch_size,
+            prefetch_to_gpu=prefetch_to_gpu,
+            save_workers=save_workers,
+        )
+
+        # Save tensor to disk asynchronously
+        file_path = manager.save_tensor(hidden_states)
+
+        # Run forward pass immediately without waiting for save to complete
+        with torch.no_grad():
+            output = forward_function(hidden_states, *args)
+
+        # Store what we need for backward
+        ctx.save_for_backward(torch.tensor([0]))  # Dummy tensor
+        ctx.file_path = file_path
+        ctx.forward_function = forward_function
+        ctx.args = args
+
+        return output
+
+    @staticmethod
+    @torch_cuda_amp_custom_bwd
+    def backward(ctx, *grad_outputs):
+        """Backward pass that loads activations from disk with prefetching"""
+        # Get the manager
+        manager = Disco._manager
+
+        # Trigger prefetching for future tensors
+        # This happens at the start of backward, so should have time to complete
+        manager.trigger_prefetch()
+
+        # Load hidden states from disk or prefetch cache
+        file_path = ctx.file_path
+        try:
+            # Ensure the file is saved before we try to load it
+            manager.wait_for_save(file_path)
+
+            hidden_states = manager.load_tensor(file_path)
+            hidden_states.requires_grad = True
+
+            # Compute gradients
+            with torch.enable_grad():
+                output = ctx.forward_function(hidden_states, *ctx.args)
+
+                # Handle tuple outputs properly
+                if isinstance(output, tuple):
+                    if len(grad_outputs) == len(output):
+                        torch.autograd.backward(output, grad_outputs)
+                    else:
+                        torch.autograd.backward(output, grad_outputs[0])
+                else:
+                    torch.autograd.backward(output, grad_outputs[0])
+
+            # Clean up the file after we're done with it
+            manager.cleanup_tensor(file_path)
+
+            return (
+                (
+                    None,  # forward_function
+                    hidden_states.grad,  # hidden_states grad
+                )
+                + (None,) * len(ctx.args)  # for each arg
+                + (
+                    None,  # prefetch_size
+                    None,  # prefetch_to_gpu
+                    None,  # save_workers
+                )
+            )
+
+        except Exception as e:
+            logger.error(f"Error in backward pass: {e}")
+            # Clean up the file even on error
+            manager.cleanup_tensor(file_path)
+            raise

src/axolotl/monkeypatch/loss/chunked.py

@@ -1,134 +0,0 @@
-"""
-chunked ce loss
-"""
-
-from typing import List, Optional
-
-import torch
-import torch.nn.functional as F
-
-
-# copied and modified from torchtune.modules.loss.CEWithChunkedOutputLoss
-class CEWithChunkedOutputLoss(torch.nn.Module):
-    """
-    Cross-entropy with chunked outputs that saves memory by only upcasting one chunk at a time.
-
-    For more details, please refer to: https://github.com/pytorch/torchtune/pull/1390
-    """
-
-    def __init__(self, num_output_chunks: int = 8, ignore_index: int = -100):
-        super().__init__()
-        self.num_output_chunks = num_output_chunks
-        self.ignore_index = ignore_index
-
-    def compute_cross_entropy(
-        self,
-        logits: torch.Tensor,
-        labels: torch.Tensor,
-        normalize: bool = True,  # pylint: disable=unused-argument
-    ) -> torch.Tensor:
-        """
-        Upcast logits to fp32 and compute cross entropy loss.
-        """
-        return F.cross_entropy(
-            logits.float(), labels, ignore_index=self.ignore_index, reduction="sum"
-        )
-
-    def forward(
-        self, logits: List[torch.Tensor], labels: torch.Tensor, reduction="sum"
-    ) -> torch.Tensor:
-        """
-        Args:
-            logits (List[torch.Tensor]): List of chunked logits of length
-                ``self.num_output_chunks``, where each chunk has shape
-                ``(batch_size, num_tokens / num_output_chunks, vocab_size)``.
-            labels (torch.Tensor): Ground truth labels of shape ``(batch_size, num_tokens)``.
-            reduction (str): The reduction to apply to the output.
-
-        Returns:
-            torch.Tensor: Cross entropy loss of shape (1,).
-        """
-
-        total_elements = (labels != self.ignore_index).sum()
-
-        # chunk and reshape labels (bsz, num_tokens, vocab) -> [(bsz*num_tokens/num_chunks, vocab)]
-        labels = [
-            target_chunk.reshape(-1)
-            for target_chunk in labels.chunk(self.num_output_chunks, dim=1)
-        ]
-        # reshape logits [(bsz, num_tokens/num_chunks, vocab)] -> [(bsz*num_tokens/num_chunks, vocab)]
-        logits = [
-            logit_chunk.reshape(-1, logit_chunk.size(-1)) for logit_chunk in logits
-        ]
-
-        # compute one chunk at a time
-        total_loss = 0.0
-        for logits_chunk, labels_chunk in zip(logits, labels):
-            total_loss += self.compute_cross_entropy(logits_chunk, labels_chunk)
-
-        if reduction == "sum":
-            return total_loss
-        return total_loss / total_elements
-
-
-def _build_chunked_ce_loss_fn(num_output_chunks: int = 8, ignore_index: int = -100):
-    loss_fn_ce = CEWithChunkedOutputLoss(num_output_chunks, ignore_index)
-    loss_fn_ce.compute_cross_entropy = torch.compile(
-        loss_fn_ce.compute_cross_entropy, backend="inductor"
-    )
-    return loss_fn_ce
-
-
-def get_causal_lm_loss(num_output_chunks: int = 8, ignore_index: int = -100):
-    loss_fn_ce = _build_chunked_ce_loss_fn(num_output_chunks, ignore_index)
-
-    def chunked_fix_cross_entropy(
-        source,
-        target,
-        num_items_in_batch: int = None,
-        ignore_index: int = -100,
-        **kwargs,
-    ):  # pylint: disable=unused-argument
-        reduction = "sum" if num_items_in_batch is not None else "mean"
-        logit_chunks = [  # pylint: disable=unnecessary-comprehension
-            chunk for chunk in source.chunk(loss_fn_ce.num_output_chunks, dim=1)
-        ]
-        loss = loss_fn_ce(logit_chunks, target, reduction=reduction)
-        if reduction == "sum":
-            loss = loss / num_items_in_batch
-        return loss
-
-    def for_causal_lm_chunked_loss(
-        logits,
-        labels,
-        vocab_size: int = None,  # pylint: disable=unused-argument
-        num_items_in_batch: Optional[int] = None,
-        ignore_index: int = -100,
-        shift_labels: Optional[torch.Tensor] = None,
-        **kwargs,
-    ) -> torch.Tensor:
-        # skip the upcast to float since we handle that in the chunking loss
-        if shift_labels is None:
-            # Shift so that tokens < n predict n
-            labels = F.pad(labels, (0, 1), value=ignore_index)
-            shift_labels = labels[..., 1:].contiguous()
-
-        # Skip Flattening the tokens
-        # Enable model parallelism
-        shift_labels = shift_labels.to(logits.device)
-        loss = chunked_fix_cross_entropy(
-            logits, shift_labels, num_items_in_batch, ignore_index, **kwargs
-        )
-        return loss
-
-    return for_causal_lm_chunked_loss
-
-
-def patch_chunked_ce_loss_fn(num_output_chunks: int = 8, ignore_index: int = -100):
-    import transformers.loss.loss_utils
-
-    for_causal_lm_chunked_loss = get_causal_lm_loss(num_output_chunks, ignore_index)
-    transformers.loss.loss_utils.ForCausalLMLoss = for_causal_lm_chunked_loss
-    transformers.loss.loss_utils.LOSS_MAPPING["ForCausalLM"] = (
-        for_causal_lm_chunked_loss
-    )

src/axolotl/monkeypatch/peft/utils.py

@@ -24,7 +24,7 @@ PATCHED_PREPARE_CODE = """
         for name, param in model.named_parameters():
             if (
                 (param.dtype == torch.float16) or (param.dtype == torch.bfloat16)
-            ) and param.__class__.__name__ != "Params4bit" and "norm" in name:
+            ) and param.__class__.__name__ != "Params4bit" and all(embed_name not in name for embed_name in ["embed_tokens", "lm_head"]):
                 param.data = param.data.to(torch.float32)
 """
 
@@ -75,4 +75,4 @@ def patch_peft_prep_code():
     exec(prep_code, globals())  # pylint: disable=exec-used  # nosec B102
     LOG.info("patching prepare_model_for_kbit_training to allow for overrides")
     peft.utils.other.prepare_model_for_kbit_training = fixed_prepare_model_for_kbit_training  # pylint: disable=protected-access  # pylint: disable=undefined-variable  # noqa: F821
-    axolotl.utils.models.prepare_model_for_kbit_training = fixed_prepare_model_for_kbit_training  # pylint: disable=protected-access  # pylint: disable=undefined-variable  # noqa: F821
+    axolotl.loaders.model.prepare_model_for_kbit_training = fixed_prepare_model_for_kbit_training  # pylint: disable=protected-access  # pylint: disable=undefined-variable  # noqa: F821

src/axolotl/monkeypatch/ring_attn/__init__.py

@@ -0,0 +1,22 @@
+"""Init for ring attention monkeypatch module"""
+
+# pylint: disable=unused-import
+# flake8: noqa
+
+from .patch import (
+    get_ring_attn_group,
+    patch_prepare_data_loader,
+    patch_prepare_device_mesh,
+    register_ring_attn,
+    set_ring_attn_group,
+    update_ring_attn_params,
+)
+
+__all__ = (
+    "get_ring_attn_group",
+    "patch_prepare_data_loader",
+    "patch_prepare_device_mesh",
+    "register_ring_attn",
+    "set_ring_attn_group",
+    "update_ring_attn_params",
+)

src/axolotl/monkeypatch/ring_attn/adapters/batch.py

@@ -16,11 +16,7 @@ import torch
 import torch.distributed as dist
 import transformers
 import transformers.modeling_flash_attention_utils
-from ring_flash_attn import (
-    ring_flash_attn_func,
-    stripe_flash_attn_func,
-    zigzag_ring_flash_attn_func,
-)
+from ring_flash_attn import ring_flash_attn_func
 from ring_flash_attn.adapters.hf_adapter import check_params
 from transformers.modeling_flash_attention_utils import (
     _flash_supports_window_size,
@@ -28,12 +24,12 @@ from transformers.modeling_flash_attention_utils import (
 )
 from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
 
-from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
+from axolotl.utils.schemas.enums import RingAttnFunc
 
 RING_ATTN_FUNC_MAPPING = {
-    RingAttnFunc.BATCH_RING: ring_flash_attn_func,
-    RingAttnFunc.BATCH_ZIGZAG: zigzag_ring_flash_attn_func,
-    RingAttnFunc.BATCH_STRIPE: stripe_flash_attn_func,
+    RingAttnFunc.BATCH_RING: torch.compile(ring_flash_attn_func),
+    # RingAttnFunc.BATCH_ZIGZAG: torch.compile(zigzag_ring_flash_attn_func),
+    # RingAttnFunc.BATCH_STRIPE: torch.compile(stripe_flash_attn_func),
 }
 
 

src/axolotl/monkeypatch/ring_attn/patch.py

@@ -0,0 +1,225 @@
+"""Ring attention group registration and flash attention patching.
+
+Make use of the `ring-flash-attn` (https://github.com/zhuzilin/ring-flash-attention)
+package, specifically the `hf_adapter.substitute_hf_flash_attn` function to patch in
+their sequence parallel version of Flash Attention 2.
+
+We also provide some patches for accelerate functions to prepare the dataloader for
+sequence parallelism training.
+"""
+
+import inspect
+
+import accelerate
+import torch
+import torch.distributed as dist
+from accelerate.logging import get_logger
+
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+from axolotl.utils.schemas.enums import RingAttnFunc
+
+LOG = get_logger(__name__)
+
+
+RING_ATTN_GROUP = None
+
+ORIGINAL_PREPARE_DATALOADER_CODE = """            submesh_fsdp_size = 1
+            submesh_dp_size = 1
+            submesh_tp_size = 1
+            if "tp" in torch_device_mesh.mesh_dim_names:
+                submesh_tp_size = torch_device_mesh["tp"].size()
+            if "dp" in torch_device_mesh.mesh_dim_names:
+                submesh_dp_size = torch_device_mesh["dp"].size()
+            if "fsdp" in torch_device_mesh.mesh_dim_names:
+                submesh_fsdp_size = torch_device_mesh["fsdp"].size()
+            process_index = process_index // submesh_tp_size"""
+
+NEW_PREPARE_DATALOADER_CODE = """            submesh_fsdp_size = 1
+            submesh_dp_size = 1
+            submesh_tp_size = 1
+            submesh_cp_size = 1
+            if "cp" in torch_device_mesh.mesh_dim_names:
+                submesh_cp_size = torch_device_mesh["cp"].size()
+            if "tp" in torch_device_mesh.mesh_dim_names:
+                submesh_tp_size = torch_device_mesh["tp"].size()
+            if "dp" in torch_device_mesh.mesh_dim_names:
+                submesh_dp_size = torch_device_mesh["dp"].size()
+            if "fsdp" in torch_device_mesh.mesh_dim_names:
+                submesh_fsdp_size = torch_device_mesh["fsdp"].size()
+            process_index = process_index // (submesh_tp_size * submesh_cp_size)"""
+
+
+def get_ring_attn_group() -> dist.ProcessGroup:
+    """Getter for ring attention group on this rank."""
+    if RING_ATTN_GROUP is None:
+        raise RuntimeError("register_ring_attn() not yet called")
+    return RING_ATTN_GROUP
+
+
+def set_ring_attn_group(ring_attn_group: dist.ProcessGroup | None):
+    """Setter for ring attention group on this rank."""
+    global RING_ATTN_GROUP  # pylint: disable=global-statement
+    RING_ATTN_GROUP = ring_attn_group
+
+
+def register_ring_attn(
+    sequence_parallel_degree: int,
+    heads_k_stride: int | None,
+    ring_attn_func: RingAttnFunc | None,
+):
+    """Create ring attention group and substitute flash attn with ring flash attn.
+
+    Args:
+        sequence_parallel_degree: Sequence parallelism factor.
+        heads_k_stride: Sequence parallelism K head stride size. Passed through to
+            `varlen_llama3` `ring_flash_attn` implementation.
+        ring_attn_func: `ring_flash_attn` ring attention implemention. If sample
+            packing is enabled, it must be a `varlen` function; otherwise, it must be a
+            `batch` function.
+    """
+    rank = dist.get_rank()
+    world_size = dist.get_world_size()
+
+    if rank == 0:
+        LOG.info(
+            "Enabling ring attention sequence parallelism: "
+            f"each sequence will be processed across {sequence_parallel_degree} GPUs"
+        )
+
+    assert sequence_parallel_degree <= world_size, (
+        f"sequence_parallel_degree ({sequence_parallel_degree}) "
+        f"must be less than or equal to world_size ({world_size})"
+    )
+    assert world_size % sequence_parallel_degree == 0, (
+        f"sequence_parallel_degree ({sequence_parallel_degree}) "
+        f"must evenly divide world_size ({world_size})"
+    )
+
+    # Assign ranks to sequence parallel groups
+    group_assignments = {}
+    for i in range(world_size // sequence_parallel_degree):
+        ring_attn_ranks = list(
+            range(
+                i * sequence_parallel_degree,
+                (i + 1) * sequence_parallel_degree,
+            )
+        )
+        group = dist.new_group(ranks=ring_attn_ranks, backend="nccl")
+
+        # Track which GPUs are in which groups
+        for r in ring_attn_ranks:
+            group_assignments[r] = i
+
+        if rank in ring_attn_ranks:
+            set_ring_attn_group(group)
+
+    # Log the GPU group assignments
+    if rank == 0:
+        LOG.info(f"Sequence parallel group assignments: {group_assignments}")
+
+    if ring_attn_func is RingAttnFunc.VARLEN_LLAMA3:
+        from ring_flash_attn import substitute_hf_flash_attn
+
+        substitute_hf_flash_attn(
+            process_group=get_ring_attn_group(), heads_k_stride=heads_k_stride or 1
+        )
+    elif ring_attn_func is RingAttnFunc.BATCH_RING:
+        from axolotl.monkeypatch.ring_attn.adapters.batch import (
+            substitute_hf_flash_attn,
+        )
+
+        substitute_hf_flash_attn(
+            process_group=get_ring_attn_group(),
+            ring_attn_func=ring_attn_func,
+        )
+
+
+def update_ring_attn_params(position_ids: torch.Tensor | None):
+    """
+    Calculate the cumulative sequence lengths for the current forward pass and pass the
+    value to the substituted `ring_flash_attn`.
+
+    Args:
+        position_ids: Optional tensor of position IDs (for sample packed data).
+    """
+    from ring_flash_attn import update_ring_flash_attn_params
+
+    cu_seqlens, _ = get_cu_seqlens_from_pos_ids(position_ids)
+    cu_seqlens = cu_seqlens.squeeze().to(device=torch.cuda.current_device())
+    update_ring_flash_attn_params(cu_seqlens, get_ring_attn_group())
+
+
+def patch_prepare_data_loader():
+    """Patch `accelerate.data_loader.prepare_data_loader` to respect the SP degree.
+
+    Raies:
+        RuntimeError: If source code to patch does not exist.
+    """
+    original_fn = accelerate.data_loader.prepare_data_loader
+    original_source = inspect.getsource(original_fn)
+
+    if ORIGINAL_PREPARE_DATALOADER_CODE not in original_source:
+        raise RuntimeError(
+            "SP patch failed - target snippet not found. "
+            "Check accelerate's version or update the patch."
+        )
+
+    patched_source = original_source.replace(
+        ORIGINAL_PREPARE_DATALOADER_CODE, NEW_PREPARE_DATALOADER_CODE
+    )
+
+    # Create a new function from the patched source
+    namespace = {}
+    exec(  # pylint: disable=exec-used  # nosec B102
+        patched_source, accelerate.data_loader.__dict__, namespace
+    )
+    patched_function = namespace["prepare_data_loader"]
+
+    accelerate.data_loader.prepare_data_loader = patched_function
+    LOG.info("Patched accelerate.data_loader.prepare_data_loader for SP support")
+
+
+def patch_prepare_device_mesh(sequence_parallel_degree: int):
+    """Patches the `Accelerator._prepare_device_mesh` method to create a device mesh
+    that includes sequence parallelism with the specified degree.
+
+    Args:
+        sequence_parallel_degree (int): The degree of sequence parallelism to use.
+    """
+
+    def _prepare_device_mesh(self):
+        """Prepare the device mesh for distributed training. The dataloader will
+        determine how to load data based on the device mesh.
+        """
+        if self.state.torch_tp_plugin:
+            return self.state.torch_tp_plugin.torch_device_mesh
+        if (
+            self.distributed_type == accelerate.accelerator.DistributedType.DEEPSPEED
+            and hasattr(self.state, "ds_device_mesh")
+        ):
+            return self.state.ds_device_mesh
+
+        # Create device mesh with sequence parallelism
+        world_size = dist.get_world_size()
+        mesh_shape = (
+            world_size // sequence_parallel_degree,
+            sequence_parallel_degree,
+        )
+        device_ids = list(range(world_size))
+
+        # Note that we use "cp" instead of "sp" to match the PyTorch native "context
+        # parallelism" implementation naming
+        return dist.DeviceMesh(
+            "cuda",
+            torch.tensor(device_ids).reshape(mesh_shape),
+            mesh_dim_names=("dp", "cp"),
+        )
+
+    # Replace the original method with our new method
+    # pylint: disable=protected-access
+    accelerate.accelerator.Accelerator._prepare_device_mesh = _prepare_device_mesh
+
+    LOG.info(
+        "Successfully patched Accelerator._prepare_device_mesh "
+        f"with sequence_parallel_degree={sequence_parallel_degree}"
+    )

src/axolotl/prompt_strategies/chat_template.py

@@ -424,6 +424,20 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
 
             LOG.debug(f"Should train: {should_train}")
 
+            # turn not trainable, skip having to find the turn indices
+            # unless last turn and train_on_eos/train_on_eot is all
+            if not should_train and (
+                self.train_on_eos != "all" and self.train_on_eot != "all"
+            ):
+                if index == len(turns) - 1:
+                    LOG.warning(
+                        "Last turn is not trainable, skipping having to find the turn indices. "
+                        "This may cause incorrect last EOT/EOS token to be unmasked."
+                        "This is likely a dataset design issue. Please ensure last turn is trainable."
+                    )
+
+                continue
+
             turn_start_idx, turn_end_idx = self.find_turn(turns=turns, turn_idx=index)
 
             LOG.debug(f"Turn indices: start={turn_start_idx}, end={turn_end_idx}")

src/axolotl/train.py

@@ -2,17 +2,17 @@
 
 import importlib
 import inspect
+import logging
 import os
 import signal
 import sys
 import weakref
-from contextlib import nullcontext
+from contextlib import ExitStack
 from pathlib import Path
 from typing import Any, Dict
 
 import torch
 import transformers.modelcard
-from accelerate.logging import get_logger
 from accelerate.utils import save_fsdp_model
 from datasets import Dataset
 from huggingface_hub.errors import OfflineModeIsEnabled
@@ -27,14 +27,17 @@ from axolotl.contribs.lgpl import (  # pylint: disable = no-name-in-module
     fix_untrained_tokens,
 )
 from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
-from axolotl.core.trainers.mixins.sequence_parallel import (
-    SequenceParallelContextManager,
-)
 from axolotl.integrations.base import PluginManager
+from axolotl.loaders import (
+    ModelLoader,
+    load_processor,
+    load_tokenizer,
+)
+from axolotl.utils.ctx_managers.sequence_parallel import SequenceParallelContextManager
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import cleanup_distributed
 from axolotl.utils.freeze import freeze_layers_except
-from axolotl.utils.models import load_model, load_processor, load_tokenizer
+from axolotl.utils.schemas.enums import RLType
 from axolotl.utils.trainer import setup_trainer
 
 try:
@@ -42,7 +45,7 @@ try:
 except ImportError:
     BetterTransformer = None
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def setup_model_and_tokenizer(
@@ -63,7 +66,6 @@ def setup_model_and_tokenizer(
     # Load tokenizer
     LOG.debug(
         f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}",
-        main_process_only=True,
     )
     tokenizer = load_tokenizer(cfg)
 
@@ -78,7 +80,8 @@ def setup_model_and_tokenizer(
         msg += " and peft_config..."
     LOG.debug(msg)
 
-    model, peft_config = load_model(cfg, tokenizer, processor=processor)
+    model_loader = ModelLoader(cfg, tokenizer, processor=processor)
+    model, peft_config = model_loader.load()
     if model.generation_config is not None:
         model.generation_config.do_sample = True
 
@@ -108,14 +111,15 @@ def setup_reference_model(
         Reference model if needed for RL training, `None` otherwise.
     """
     model_ref = None
-    if cfg.rl and cfg.rl != "orpo":
+    if cfg.rl and cfg.rl != RLType.ORPO:
         if cfg.adapter and not cfg.rl_adapter_ref_model:
             # use built-in trl autounwrap
             LOG.debug("Passing model_ref: None to RL trainer")
             model_ref = None  # explicit setting to None
         else:
             # load the model again for model_ref/baseline
-            model_ref, _ = load_model(cfg, tokenizer, reference_model=True)
+            model_loader = ModelLoader(cfg, tokenizer, reference_model=True)
+            model_ref, _ = model_loader.load()
     return model_ref
 
 
@@ -189,28 +193,33 @@ def execute_training(
         trainer: The configured trainer object.
         resume_from_checkpoint: Path to checkpoint to resume from, if applicable.
     """
-    # Define the context managers to use
-    flash_context = (
-        torch.backends.cuda.sdp_kernel(
-            enable_flash=True,
-            enable_math=True,
-            enable_mem_efficient=True,
-        )
-        if cfg.flash_optimum
-        else nullcontext()
-    )
-    sequence_parallel_context = (
-        SequenceParallelContextManager(
-            model=trainer.model,
-            sequence_parallel_degree=cfg.sequence_parallel_degree,
-            ring_attn_func=cfg.ring_attn_func,
-        )
-        if cfg.sequence_parallel_degree > 1
-        else nullcontext()
-    )
+    with ExitStack() as stack:
+        # Define the context managers to use
+        if cfg.flash_optimum:
+            stack.enter_context(
+                torch.backends.cuda.sdp_kernel(
+                    enable_flash=True,
+                    enable_math=True,
+                    enable_mem_efficient=True,
+                )
+            )
 
-    LOG.info("Starting trainer...")
-    with flash_context, sequence_parallel_context:
+        if cfg.sequence_parallel_degree > 1:
+            models = [trainer.model]
+            if hasattr(trainer, "ref_model"):
+                models.append(trainer.ref_model)
+
+            stack.enter_context(
+                SequenceParallelContextManager(
+                    models=models,
+                    sequence_parallel_degree=cfg.sequence_parallel_degree,
+                    gradient_accumulation_steps=cfg.gradient_accumulation_steps,
+                    ring_attn_func=cfg.ring_attn_func,
+                    heads_k_stride=cfg.heads_k_stride,
+                )
+            )
+
+        LOG.info("Starting trainer...")
         trainer.train(resume_from_checkpoint=resume_from_checkpoint)
 
 
@@ -528,6 +537,9 @@ def train(
         processor,
     ) = setup_model_and_trainer(cfg, dataset_meta)
 
+    plugin_manager = PluginManager.get_instance()
+    plugin_manager.post_trainer_create(cfg, trainer)
+
     # Handle untrained tokens if configured
     safe_serialization = cfg.save_safetensors is True
     train_dataset = dataset_meta.train_dataset
@@ -550,7 +562,6 @@ def train(
     if not cfg.use_ray:
         cleanup_distributed()
 
-    plugin_manager = PluginManager.get_instance()
     plugin_manager.post_train(cfg, model)
 
     return model, tokenizer, trainer

src/axolotl/utils/callbacks/__init__.py

@@ -885,10 +885,9 @@ def colab_inference_post_train_callback(trainer: Trainer):
             handle T4 gpu, we need to convert attention to eager for inference
             """
             if "Tesla T4" in self.gpu_name and self.cfg.xformers_attention:
-                trainer.model.eval()
-            trainer.model.config._attn_implementation = (  # pylint: disable=protected-access
-                "eager"
-            )
+                trainer.model.config._attn_implementation = (  # pylint: disable=protected-access
+                    "eager"
+                )
             trainer.model.gradient_checkpointing_disable()
             trainer.model.config.use_cache = True
             trainer.model.eval()

src/axolotl/utils/callbacks/mlflow_.py

@@ -1,6 +1,7 @@
 """MLFlow module for trainer callbacks"""
 
 import logging
+import os
 from shutil import copyfile
 from tempfile import NamedTemporaryFile
 from typing import TYPE_CHECKING
@@ -16,6 +17,11 @@ if TYPE_CHECKING:
 LOG = logging.getLogger("axolotl.callbacks")
 
 
+def should_log_artifacts() -> bool:
+    truths = ["TRUE", "1", "YES"]
+    return os.getenv("HF_MLFLOW_LOG_ARTIFACTS", "FALSE").upper() in truths
+
+
 class SaveAxolotlConfigtoMlflowCallback(TrainerCallback):
     # pylint: disable=duplicate-code
     """Callback to save axolotl config to mlflow"""
@@ -32,13 +38,18 @@ class SaveAxolotlConfigtoMlflowCallback(TrainerCallback):
     ):
         if is_main_process():
             try:
-                with NamedTemporaryFile(
-                    mode="w", delete=False, suffix=".yml", prefix="axolotl_config_"
-                ) as temp_file:
-                    copyfile(self.axolotl_config_path, temp_file.name)
-                    mlflow.log_artifact(temp_file.name, artifact_path="")
+                if should_log_artifacts():
+                    with NamedTemporaryFile(
+                        mode="w", delete=False, suffix=".yml", prefix="axolotl_config_"
+                    ) as temp_file:
+                        copyfile(self.axolotl_config_path, temp_file.name)
+                        mlflow.log_artifact(temp_file.name, artifact_path="")
+                        LOG.info(
+                            "The Axolotl config has been saved to the MLflow artifacts."
+                        )
+                else:
                     LOG.info(
-                        "The Axolotl config has been saved to the MLflow artifacts."
+                        "Skipping logging artifacts to MLflow (hf_mlflow_log_artifacts is false)"
                     )
             except (FileNotFoundError, ConnectionError) as err:
                 LOG.warning(f"Error while saving Axolotl config to MLflow: {err}")

src/axolotl/utils/config/__init__.py

@@ -11,9 +11,10 @@ from transformers.utils.import_utils import is_torch_npu_available
 
 from axolotl.integrations.base import PluginManager
 from axolotl.integrations.config import merge_input_args
+from axolotl.loaders import MULTIMODAL_AUTO_MODEL_MAPPING
+from axolotl.loaders.utils import load_model_config
 from axolotl.utils.bench import log_gpu_memory_usage
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import MULTIMODAL_AUTO_MODEL_MAPPING, load_model_config
 from axolotl.utils.schemas.config import (
     AxolotlConfigWCapabilities as AxolotlConfigWCapabilitiesBase,
 )

src/axolotl/utils/ctx_managers/__init__.py

@@ -0,0 +1,6 @@
+"""Init for context manager submodule"""
+
+# pylint: disable=unused-import
+# flake8: noqa
+
+from .sequence_parallel import SequenceParallelContextManager

src/axolotl/utils/ctx_managers/sequence_parallel.py

@@ -0,0 +1,376 @@
+"""Module for Axolotl trainer sequence parallelism manager and utilities"""
+
+import functools
+import inspect
+
+import torch
+import torch.distributed as dist
+from torch import nn
+from torch.utils.hooks import RemovableHandle
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from transformers.utils import ModelOutput
+
+from axolotl.monkeypatch.ring_attn import (
+    get_ring_attn_group,
+    patch_prepare_data_loader,
+    patch_prepare_device_mesh,
+    register_ring_attn,
+    update_ring_attn_params,
+)
+from axolotl.utils.schemas.enums import RingAttnFunc
+
+
+# TODO(djsaunde): implement zigzag, stripe patterns here (and elsewhere) in this
+# module. Currently, we just focus on batch ring and varlen llama3 for simplicity.
+def apply_sequence_parallelism(
+    batch: dict[str, torch.Tensor],
+    local_rank: int,
+    local_world_size: int,
+    gradient_accumulation_steps: int,
+    ring_attn_func: RingAttnFunc,  # pylint: disable=unused-argument
+) -> tuple[dict[str, torch.Tensor], int, int]:
+    """
+    Apply sequence parallelism slicing to a batch.
+
+    Special handling is implemented for integer logits_to_keep, which indicates
+    to only keep the last N tokens in the sequence during generation.
+
+    Args:
+        batch: Batch dictionary (e.g., input_ids, attention_mask, etc.).
+        local_rank: Local rank in the sequence parallel group.
+        local_world_size: World size of the sequence parallel group.
+        gradient_accumulation_steps: Number of steps to accumulate gradients over.
+        ring_attn_func: Which ring attention function to use. Currently unused, but
+            related to above TODO.
+
+    Returns:
+        tuple of:
+            - Batch dictionary with sliced tensors.
+            - The original sequence length before padding.
+            - The number of padding tokens added.
+    """
+    original_seq_len = batch["input_ids"].size(1)
+
+    # Update ring attention params if needed
+    if batch.get("position_ids") is not None:
+        update_ring_attn_params(position_ids=batch["position_ids"])
+    else:
+        # If position_ids aren't already in the batch, create them
+        batch["position_ids"] = torch.arange(
+            0,
+            original_seq_len,
+            dtype=torch.long,
+            device=batch["input_ids"].device,
+        ).expand(batch["input_ids"].size(0), -1)
+
+    if "logits_to_keep" in batch and isinstance(batch["logits_to_keep"], int):
+        logits_to_keep = batch["logits_to_keep"]
+
+        # Calculate which positions in the full sequence contain the last N tokens
+        start_position = max(0, original_seq_len - logits_to_keep)
+        chunk_size = original_seq_len // local_world_size
+        rank_start = local_rank * chunk_size
+        rank_end = rank_start + chunk_size
+
+        # Create a boolean mask tensor for this rank's chunk
+        mask = torch.zeros(
+            chunk_size,
+            dtype=torch.bool,
+            device=batch["input_ids"].device,
+        )
+
+        if rank_end > start_position:
+            # Calculate how many of the last N tokens fall within this rank's range
+            tokens_in_rank = min(rank_end, original_seq_len) - max(
+                rank_start, start_position
+            )
+
+            # Calculate where these tokens start in the local chunk
+            local_start_idx = max(0, start_position - rank_start)
+
+            # Set the appropriate positions in the mask to True
+            mask[local_start_idx : local_start_idx + tokens_in_rank] = True
+
+        # Replace the integer with the boolean mask
+        batch["logits_to_keep"] = mask
+
+    # Add padding to make sequence length divisible by local_world_size
+    total_seq_len = original_seq_len
+    pad_len = 0
+    divisor = min(local_world_size, 64)
+    if total_seq_len % divisor != 0:
+        pad_len = divisor - (total_seq_len % divisor)
+
+        # Apply padding to all relevant tensors
+        for key in batch:
+            if (
+                isinstance(batch[key], torch.Tensor)
+                and batch[key].dim() > 1
+                and batch[key].size(1) == total_seq_len
+            ):
+                # Create padding tensor
+                pad_value = -100 if key == "labels" else 0
+                padding = torch.full(
+                    (batch[key].size(0), pad_len, *batch[key].shape[2:]),
+                    pad_value,
+                    dtype=batch[key].dtype,
+                    device=batch[key].device,
+                )
+
+                # Concatenate padding to the right side of the tensor
+                batch[key] = torch.cat([batch[key], padding], dim=1)
+            if key == "logits_to_keep":
+                # Create padding tensor
+                padding = torch.ones(
+                    1,
+                    dtype=batch[key].dtype,
+                    device=batch[key].device,
+                )
+
+                # Concatenate padding to the right side of the tensor
+                batch[key] = torch.cat([batch[key], padding], dim=0)
+
+        # Update the total sequence length after padding
+        total_seq_len = batch["input_ids"].size(1)
+
+    # Slice batch for sequence parallel
+    for key in batch:
+        if not isinstance(batch[key], torch.Tensor) or batch[key].dim() <= 1:
+            continue
+
+        # Split in sequential fashion and grab this rank's chunk
+        if batch[key].size(1) == total_seq_len:
+            batch[key] = (
+                batch[key].chunk(local_world_size, dim=1)[local_rank].contiguous()
+            )
+        elif key == "logits_to_keep":
+            batch[key] = (
+                batch[key].chunk(local_world_size, dim=0)[local_rank].contiguous()
+            )
+
+        # Handle num_items_in_batch
+        if "num_items_in_batch" in batch:
+            # Approximation; this needed since num_items_in_batch may be counted across
+            # all samples in a gradient accumulated batch, not on a per-step basis.
+            batch["num_items_in_batch"] = (
+                batch["labels"] != -100
+            ).sum() * gradient_accumulation_steps
+
+    return batch, original_seq_len, pad_len
+
+
+class SequenceParallelContextManager:
+    """Context manager for sequence parallelism operations.
+
+    This class provides a context that will automatically apply sequence parallelism
+    during model forward passes using a pre-forward hook, and gather outputs from
+    across the sequence parallelism group using a post-forward hook.
+
+    Args:
+        models: List of models to apply sequence parallelism to pre- and post- forward
+            hooks.
+        sequence_parallel_degree: Number of processes to split sequences over.
+        gradient_accumulation_steps: Number of steps to accumulate gradients over.
+        ring_attn_func: Which ring attention function to use. Currently unused.
+        heads_k_stride: Sequence parallelism K head stride size. Passed through to
+            `varlen_llama3` `ring_flash_attn` implementation.
+    """
+
+    def __init__(
+        self,
+        models: list[nn.Module],
+        sequence_parallel_degree: int,
+        gradient_accumulation_steps: int,
+        ring_attn_func: RingAttnFunc,
+        heads_k_stride: int | None,
+    ):
+        self.models = models
+        self.sequence_parallel_degree = sequence_parallel_degree
+        self.gradient_accumulation_steps = gradient_accumulation_steps
+        self.ring_attn_func = ring_attn_func
+        self.heads_k_stride = heads_k_stride
+        self._register_ring_attn()
+
+        # Set distributed info for local rank
+        self.process_group = get_ring_attn_group()
+        self.local_rank = dist.get_rank(self.process_group)
+        self.local_world_size = dist.get_world_size(self.process_group)
+
+        # Will store hook handles for removal
+        self.hook_handles: list[RemovableHandle] = []
+
+        # Store original sequence length and padding information
+        self.original_seq_len = 0
+        self.pad_len = 0
+
+        # Create a partially applied version of the apply_sequence_parallelism function
+        self.apply_sequence_parallelism = functools.partial(
+            apply_sequence_parallelism,
+            local_rank=self.local_rank,
+            local_world_size=self.local_world_size,
+            gradient_accumulation_steps=self.gradient_accumulation_steps,
+            ring_attn_func=self.ring_attn_func,
+        )
+
+    def __enter__(self):
+        self._register_model_hooks()
+
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        # Remove all hooks
+        for handle in self.hook_handles:
+            handle.remove()
+        self.hook_handles = []
+
+        # TODO(djsaunde): Un-patch attention and accelerate functions (low priority)
+
+    def _register_ring_attn(self):
+        # Initialize ring attn for sequence parallelism
+        register_ring_attn(
+            sequence_parallel_degree=self.sequence_parallel_degree,
+            heads_k_stride=self.heads_k_stride,
+            ring_attn_func=self.ring_attn_func,
+        )
+
+        # Patches for accelerate functionality
+        patch_prepare_data_loader()
+        patch_prepare_device_mesh(
+            sequence_parallel_degree=self.sequence_parallel_degree
+        )
+
+    def _register_model_hooks(self):
+        # Forward pre-hook to apply sequence parallelism
+        def sequence_parallel_pre_hook(_, args, kwargs):
+            # Get parameter names from the model's forward function
+            forward_params = list(
+                inspect.signature(self.models[0].forward).parameters.keys()
+            )
+
+            updated_kwargs = kwargs.copy()
+            for i, arg in enumerate(args):
+                if i < len(forward_params):
+                    updated_kwargs[forward_params[i]] = arg
+
+            # Any excess positional arguments are kept as-is
+            remaining_args = args[len(forward_params) :]
+
+            # Apply sequence parallelism to updated kwargs
+            updated_kwargs, self.original_seq_len, self.pad_len = (
+                self.apply_sequence_parallelism(updated_kwargs)
+            )
+
+            return remaining_args, updated_kwargs
+
+        # Forward post-hook to gather outputs
+        def sequence_parallel_post_hook(_, __, output: ModelOutput) -> ModelOutput:
+            # Gather the sharded outputs
+            output = self._gather_outputs(output)
+
+            # Remove padding if it was added
+            if self.pad_len > 0:
+                for key, value in output.items():
+                    if isinstance(value, torch.Tensor) and value.dim() > 1:
+                        if value.size(1) == self.original_seq_len + self.pad_len:
+                            # Slice to remove padding
+                            output[key] = value[:, : self.original_seq_len].contiguous()
+
+            return output
+
+        # Register both hooks
+        for model in self.models:
+            self.hook_handles.append(
+                model.register_forward_pre_hook(
+                    sequence_parallel_pre_hook, with_kwargs=True
+                )
+            )
+            self.hook_handles.append(
+                model.register_forward_hook(sequence_parallel_post_hook)
+            )
+
+    def _gather_outputs(self, output: CausalLMOutputWithPast) -> CausalLMOutputWithPast:
+        """Gather sharded outputs from all ranks and reconstruct the full tensor."""
+        for key, value in output.items():
+            if isinstance(value, torch.Tensor) and value.dim() > 1:
+                output[key] = AllGatherWithGrad.apply(value, self.process_group)
+
+        return output
+
+
+class AllGatherWithGrad(torch.autograd.Function):
+    """Custom autograd function for all-gather to preserve gradients."""
+
+    @staticmethod
+    def forward(
+        ctx: torch.autograd.function.FunctionCtx,
+        input_tensor: torch.Tensor,
+        group: dist.ProcessGroup,
+    ) -> torch.Tensor:
+        """
+        Forward pass of all-gather of data with sequence dimension.
+
+        Args:
+            ctx: `torch.autograd` function context.
+            input_tensor: Tensor from model output with sequence dimension.
+            group: `torch.distributed` process group.
+
+        Returns:
+            Tensor from gathering the `input_tensor` from across the process group and
+                concatenating along the sequence dimension.
+        """
+        ctx.group = group
+        ctx.rank = dist.get_rank(group)
+        world_size = dist.get_world_size(group)
+
+        # Gather shape metadata
+        local_shape = torch.tensor(list(input_tensor.shape), device=input_tensor.device)
+        all_shapes = [torch.zeros_like(local_shape) for _ in range(world_size)]
+        dist.all_gather(all_shapes, local_shape, group=group)
+
+        # Store sequence lengths for backward pass
+        seq_lens = [int(shape[1].item()) for shape in all_shapes]
+        ctx.seq_lens = seq_lens
+
+        # Perform all_gather operation
+        gathered = [
+            torch.zeros(
+                tuple(shape.tolist()),
+                dtype=input_tensor.dtype,
+                device=input_tensor.device,
+            )
+            for shape in all_shapes
+        ]
+        dist.all_gather(gathered, input_tensor, group=group)
+
+        # Concatenate tensors along sequence dimension
+        result = torch.cat(gathered, dim=1)
+
+        return result
+
+    @staticmethod
+    def backward(
+        ctx: torch.autograd.function.FunctionCtx, grad_output: torch.Tensor
+    ) -> tuple[torch.Tensor, None]:
+        """
+        Backward pass for all-gather operation.
+
+        Extracts the gradient slice corresponding to this rank's original input
+        from the full gradient tensor.
+
+        Args:
+            ctx: `torch.autograd` function context.
+            grad_output: Gradient from subsequent layers with respect to the
+                concatenated output tensor.
+
+        Returns:
+            Tuple containing the gradient slice for this rank's input tensor and `None`
+                for the process group parameter which doesn't require gradients.
+        """
+        rank = ctx.rank
+        seq_lens = ctx.seq_lens
+
+        # Extract gradient for this rank's chunk
+        offset = sum(seq_lens[:rank])
+        grad_slice = grad_output[:, offset : offset + seq_lens[rank]].contiguous()
+
+        return grad_slice, None

src/axolotl/utils/data/rl.py

@@ -10,6 +10,7 @@ import yaml
 from datasets import Dataset, DatasetDict, concatenate_datasets, load_from_disk
 
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
+from axolotl.loaders import load_tokenizer
 from axolotl.prompt_strategies.dpo import load as load_dpo
 from axolotl.prompt_strategies.kto import load as load_kto
 from axolotl.prompt_strategies.orpo import load as load_orpo
@@ -17,9 +18,9 @@ from axolotl.utils.data.shared import datasets_w_name_generator, load_dataset_w_
 from axolotl.utils.data.utils import deduplicate_and_log_datasets, md5
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process, zero_first
-from axolotl.utils.models import load_tokenizer
+from axolotl.utils.schemas.enums import RLType
 
-LOG = logging.getLogger("axolotl")
+LOG = logging.getLogger(__name__)
 
 
 def _get_path(ds_hash, cfg):
@@ -71,6 +72,7 @@ def map_dataset(cfg, data_set, ds_transform_fn, tokenizer, **map_kwargs):
     data_set = data_set.map(
         ds_transform_fn,
         desc="Mapping RL Dataset",
+        num_proc=cfg.dataset_processes,
         **map_kwargs,
     )
 
@@ -80,7 +82,7 @@ def map_dataset(cfg, data_set, ds_transform_fn, tokenizer, **map_kwargs):
 def drop_long_rl_seq(
     sample, rl, tokenizer, sequence_len  # pylint: disable=invalid-name
 ):
-    if rl in ("dpo", "ipo", "orpo", "simpo"):
+    if rl in (RLType.DPO, RLType.IPO, RLType.ORPO, RLType.SIMPO):
         if not (
             sample.get("prompt") and sample.get("chosen") and sample.get("rejected")
         ):
@@ -100,7 +102,7 @@ def drop_long_rl_seq(
             len_prompt + len_rejected
         ) <= sequence_len
 
-    if rl == "kto":
+    if rl is RLType.KTO:
         if not (sample.get("prompt") and sample.get("completion")):
             raise ValueError("Prompt and completion keys are required for KTO datasets")
 
@@ -114,7 +116,7 @@ def drop_long_rl_seq(
 
         return (len_prompt + len_completion) <= sequence_len
 
-    if rl == "grpo":
+    if rl is RLType.GRPO:
         return True
 
     raise ValueError("Unknown RL type")
@@ -137,9 +139,9 @@ def load_prepare_preference_datasets(cfg):
             if _type:
                 if isinstance(_type, DictDefault):
                     _type = "user_defined.default"
-                if _cfg.rl == "orpo":
+                if _cfg.rl is RLType.ORPO:
                     ds_transform_fn = load_orpo(_type, _cfg, dataset_idx=i)
-                elif _cfg.rl == "kto":
+                elif _cfg.rl is RLType.KTO:
                     ds_transform_fn = load_kto(_type, _cfg, dataset_idx=i)
                 else:
                     ds_transform_fn = load_dpo(_type, _cfg, dataset_idx=i)
@@ -150,7 +152,7 @@ def load_prepare_preference_datasets(cfg):
                 split_datasets[i] = map_dataset(
                     cfg, data_set, ds_transform_fn, tokenizer, **map_kwargs
                 )
-            elif _cfg.rl == "kto":
+            elif _cfg.rl is RLType.KTO:
                 ds_transform_fn = load_kto(_type, _cfg, dataset_idx=i)
                 map_kwargs = {}
                 if isinstance(ds_transform_fn, tuple):
@@ -185,7 +187,7 @@ def load_prepare_preference_datasets(cfg):
                     )
 
         combined_datasets = concatenate_datasets(split_datasets)
-        combined_datasets = combined_datasets.shuffle(seed=cfg.seed)
+        combined_datasets = combined_datasets.shuffle(seed=cfg.seed or 42)
 
         return combined_datasets
 
@@ -205,6 +207,8 @@ def load_prepare_preference_datasets(cfg):
                 eval_dataset = load_split(cfg.test_datasets, cfg)
         if not eval_dataset:
             if cfg.val_set_size:
+                seed = cfg.seed if cfg.seed is not None else 42
+
                 # ensure we end up with the same fingerprint by doing rank0 first and being able to cache
                 to_hash_train = (
                     train_dataset._fingerprint  # pylint: disable=protected-access
@@ -213,7 +217,7 @@ def load_prepare_preference_datasets(cfg):
                     + "|"
                     + "train"
                     + "|"
-                    + str(cfg.seed or 42)
+                    + str(seed)
                 )
                 to_hash_test = (
                     train_dataset._fingerprint  # pylint: disable=protected-access
@@ -222,13 +226,13 @@ def load_prepare_preference_datasets(cfg):
                     + "|"
                     + "test"
                     + "|"
-                    + str(cfg.seed or 42)
+                    + str(seed)
                 )
                 train_fingerprint = md5(to_hash_train)
                 test_fingerprint = md5(to_hash_test)
                 ds_w_test_split = train_dataset.train_test_split(
                     test_size=cfg.val_set_size,
-                    seed=cfg.seed,
+                    seed=seed,
                     shuffle=False,
                     train_new_fingerprint=train_fingerprint,
                     test_new_fingerprint=test_fingerprint,

src/axolotl/utils/data/sft.py

@@ -148,7 +148,7 @@ def prepare_dataset(cfg, tokenizer, processor=None, preprocess_iterable=None):
             ds_wrapper_partial,
             max_tokens=cfg.sequence_len,
             batch_size=cfg.micro_batch_size,
-            seed=cfg.seed or 42,
+            seed=cfg.seed if cfg.seed is not None else 42,
             buffer_size=cfg.pretrain_multipack_buffer_size or 10_000,
         )
         # https://discuss.huggingface.co/t/how-to-use-huggingface-trainer-streaming-datasets-without-wrapping-it-with-torchdatas-iterablewrapper/25230
@@ -416,6 +416,8 @@ def load_prepare_datasets(
     )
 
     if split == "train" and val_set_size:
+        seed = cfg.seed if cfg.seed is not None else 42
+
         # ensure we end up with the same fingerprint by doing rank0 first and being able to cache
         to_hash_train = (
             dataset._fingerprint  # pylint: disable=protected-access
@@ -424,7 +426,7 @@ def load_prepare_datasets(
             + "|"
             + "train"
             + "|"
-            + str(cfg.seed or 42)
+            + str(seed)
         )
         to_hash_test = (
             dataset._fingerprint  # pylint: disable=protected-access
@@ -433,7 +435,7 @@ def load_prepare_datasets(
             + "|"
             + "test"
             + "|"
-            + str(cfg.seed or 42)
+            + str(seed)
         )
         train_fingerprint = md5(to_hash_train)
         test_fingerprint = md5(to_hash_test)
@@ -442,7 +444,7 @@ def load_prepare_datasets(
         dataset = dataset.train_test_split(
             test_size=val_set_size,
             shuffle=False,
-            seed=cfg.seed or 42,
+            seed=seed,
             train_new_fingerprint=train_fingerprint,
             test_new_fingerprint=test_fingerprint,
         )
@@ -482,7 +484,7 @@ def get_dataset_wrapper(
     }
 
     LOG.info(
-        f"Loading dataset with base_type: {d_base_type} and prompt_style: {d_prompt_style}"
+        f"Loading dataset: {config_dataset['path']} with base_type: {d_base_type} and prompt_style: {d_prompt_style}"
     )
 
     if (

src/axolotl/utils/data/shared.py

@@ -281,6 +281,10 @@ def load_dataset_w_config(
             **load_ds_kwargs,
         )
     if not ds:
-        raise ValueError("unhandled dataset load")
+        raise ValueError(
+            "The dataset could not be loaded. This could be due to a misconfigured dataset path "
+            f"({config_dataset.path}). Try double-check your path / name / data_files. "
+            "This is not caused by the dataset type."
+        )
 
     return ds

src/axolotl/utils/gradient_checkpointing/__init__.py

@@ -1,20 +0,0 @@
-"""custom checkpointing utils"""
-
-from functools import partial
-
-from axolotl.utils.gradient_checkpointing.unsloth import (
-    Unsloth_Offloaded_Gradient_Checkpointer,
-)
-
-
-def hf_grad_checkpoint_offload_wrapper(
-    decoder_layer, *args, use_reentrant=None
-):  # pylint: disable=unused-argument
-    return Unsloth_Offloaded_Gradient_Checkpointer.apply(
-        (
-            decoder_layer.func.__self__
-            if isinstance(decoder_layer, partial)
-            else decoder_layer.__self__
-        ),
-        *args,
-    )

src/axolotl/utils/lora_embeddings.py

@@ -1,14 +0,0 @@
-"""
-helpers for lora embeddings
-"""
-
-
-def get_linear_embedding_layers(model_type):
-    """
-    returns the linear embedding layers needed for loras, dependent on the model arch
-    """
-    if model_type == "gpt_neox":
-        return ["embed_in", "embed_out"]
-    if model_type == "falcon":
-        return ["word_embeddings", "lm_head"]
-    return ["embed_tokens", "lm_head"]

src/axolotl/utils/samplers/multipack.py

@@ -6,8 +6,8 @@ into fixed-capacity batches to optimize memory usage and training throughput.
 import logging
 import math
 from concurrent.futures import ProcessPoolExecutor
-from multiprocessing import cpu_count
-from typing import Iterable, List, Union
+from multiprocessing import cpu_count, get_context
+from typing import Iterable, Union
 
 import numba
 import numpy as np
@@ -78,15 +78,11 @@ def pack_group(
     Returns:
         List of bins, where each bin contains indices of sequences assigned to it
     """
-    # Get sorting indices and sort lengths in descending order
-    indices = np.argsort(sequence_lengths)[::-1]
-    sorted_lengths = sequence_lengths[indices]
-
     bins_remaining_space: list = []  # Tracks remaining capacity in each bin
     bins_assigned_sequences: list = []  # Tracks sequence indices assigned to each bin
 
-    for seq_id, size in enumerate(sorted_lengths):
-        global_idx = indices[seq_id] + group_offset
+    for seq_id, size in enumerate(sequence_lengths):
+        global_idx = seq_id + group_offset
 
         # Try to place sequence in existing bins
         add_new_bin = True
@@ -130,6 +126,7 @@ def pack_parallel(
     bin_size: int,
     num_processes: int | None = None,
     safe_mode: bool = True,
+    mp_start_method: str | None = "spawn",
 ):
     """
     Pack sequences into bins using parallel processing
@@ -141,7 +138,9 @@ def pack_parallel(
         bin_size: Maximum number of bins to use
         num_processes: Number of parallel processes to use
         safe_mode: If True, use a more conservative packing approach
-
+        mp_start_method: Multiprocessing start method ('fork', 'spawn', 'forkserver').
+                         'spawn' is often safer with Numba/PyTorch.
+                         Set to None to use system default.
     Returns:
         List of bins, where each bin contains indices of sequences assigned to it
     """
@@ -158,9 +157,33 @@ def pack_parallel(
 
     # Process groups in parallel
     all_bins = []
-    with ProcessPoolExecutor(max_workers=num_processes) as executor:
-        for group_bins in executor.map(_process_group, tasks):
+
+    mp_ctx = None
+    if mp_start_method:
+        try:
+            mp_ctx = get_context(mp_start_method)
+        except ValueError:
+            LOG.warning(
+                f"Failed to get multiprocessing context '{mp_start_method}'. "
+                f"Falling back to default. Available: {get_context().get_all_start_methods()}"
+            )
+            mp_ctx = (
+                None  # Fallback to default context if specified one is not available
+            )
+
+    if num_processes == 1:
+        LOG.debug("Using single process for pack_parallel, running sequentially.")
+        for task_args in tasks:
+            group_bins = _process_group(task_args)
             all_bins.extend(group_bins)
+    else:
+        # Use ProcessPoolExecutor only if num_processes > 1
+        # Pass mp_context if available
+        with ProcessPoolExecutor(
+            max_workers=num_processes, mp_context=mp_ctx
+        ) as executor:
+            for group_bins in executor.map(_process_group, tasks):
+                all_bins.extend(group_bins)
 
     return all_bins
 
@@ -172,7 +195,7 @@ def allocate_sequentially(
     """
     Sequential allocator that preserves example order
 
-    Parameters:
+    Args:
         sequence_lengths: The lengths of all examples
         rank: The current rank (for distributed training)
         bin_capacity: The capacity of each bin (maximum sequence length)
@@ -183,38 +206,37 @@ def allocate_sequentially(
         total_tokens_used: Number of actual example tokens
         total_token_slots: Maximum theoretical number of example tokens (number of bins * bin capacity)
     """
-    rank_batches = []
-    total_tokens_used = 0
+    result = []
+    total_used = 0
 
     # First, do sequential packing into bins
     all_bins = []
-    current_bin = []
+    current_bin = [0 for i in range(0)]  # numba hint
     remaining_capacity = bin_capacity
 
-    # Process each sequence in order
     for idx, size in enumerate(sequence_lengths):
         if size <= remaining_capacity:
             # Example fits in current bin
             current_bin.append(idx)
             remaining_capacity -= size
-            total_tokens_used += size
+            total_used += size
         else:
             # Example doesn't fit, start a new bin
             if current_bin:  # Add non-empty bin to all_bins
                 all_bins.append(current_bin)
             current_bin = [idx]
             remaining_capacity = bin_capacity - size
-            total_tokens_used += size
+            total_used += size
 
     # Add the last bin if not empty
     if current_bin:
         all_bins.append(current_bin)
 
-    # Assign bins to ranks - each rank gets every num_ranks-th bin
+    # Assign bins to ranks - each rank gets every n-th bin
     for bin_idx in range(rank, len(all_bins), num_ranks):
-        rank_batches.append(all_bins[bin_idx])
+        result.append(all_bins[bin_idx])
 
-    return rank_batches, total_tokens_used, len(all_bins) * bin_capacity
+    return result, total_used, len(all_bins) * bin_capacity
 
 
 class MultipackBatchSampler(BatchSampler):
@@ -235,8 +257,8 @@ class MultipackBatchSampler(BatchSampler):
         batch_max_len: int,  # Maximum sequence length (bin capacity)
         lengths: np.ndarray,  # Sequence lengths
         packing_efficiency_estimate: float = 1.0,  # Initial efficiency estimate
-        drop_last: bool = False,  # Whether to drop incomplete batches
-        num_count_samples: int = 16,  # Number of samples to estimate batch count
+        drop_last: bool = False,  # Whether to drop final batches (might be incomplete)
+        num_count_samples: int = 16,  # Number of times to estimate batch count
         sequential: bool = False,  # Whether to use sequential packing
         group_size: int = 100_000,  # Size of groups for parallel packing
         bin_size: int = 200,  # The max number of samples that can be packed in a single bin
@@ -311,6 +333,8 @@ class MultipackBatchSampler(BatchSampler):
                 bin_capacity=self.batch_max_len,
                 num_ranks=1,
             )
+            # Map bin indices back to original indices
+            bins = [[indices[b_idx] for b_idx in bin_indices] for bin_indices in bins]
         else:
             # Use parallel packing
             all_bins = pack_parallel(
@@ -382,7 +406,7 @@ class MultipackBatchSampler(BatchSampler):
         Returns a conservative efficiency estimate based on the measurements
         """
 
-        def calc_sample_packing_eff_est(estimates: List[float]):
+        def calc_sample_packing_eff_est(estimates: list[float]):
             LOG.debug(f"sample_packing_eff_est across ranks: {repr(estimates)}")
             # Use 99.7% of max observed efficiency as a safe estimate
             max_eff = max(float(eff) for eff in estimates)

src/axolotl/utils/schemas/config.py

@@ -27,7 +27,7 @@ from axolotl.utils.schemas.datasets import (
     StepwiseSupervisedDataset,
 )
 from axolotl.utils.schemas.deprecated import DeprecatedParameters, RemappedParameters
-from axolotl.utils.schemas.enums import ChatTemplate, RLType
+from axolotl.utils.schemas.enums import ChatTemplate, RingAttnFunc, RLType
 from axolotl.utils.schemas.integrations import (
     CometConfig,
     GradioConfig,
@@ -82,6 +82,7 @@ class AxolotlInputConfig(
     mean_resizing_embeddings: bool | None = False
     # optionally shrink the embeddings when the tokenizer vocab size is smaller
     shrink_embeddings: bool | None = None
+    embeddings_skip_upcast: bool | None = None
 
     rl: RLType | None = None
     trl: TRLConfig | None = Field(
@@ -177,7 +178,7 @@ class AxolotlInputConfig(
 
     # torch_dtype: torch.dtype | None
 
-    gradient_checkpointing: Literal["unsloth", "offload"] | bool | None = Field(
+    gradient_checkpointing: Literal["offload", "offload_disk"] | bool | None = Field(
         default=False
     )
     gradient_checkpointing_kwargs: dict[str, Any] | None = None
@@ -242,9 +243,6 @@ class AxolotlInputConfig(
     unsloth_rms_norm: bool | None = None
     unsloth_rope: bool | None = None
 
-    chunked_cross_entropy: bool | None = None
-    chunked_cross_entropy_num_chunks: int | None = None
-
     lora_mlp_kernel: bool | None = None
     lora_qkv_kernel: bool | None = None
     lora_o_kernel: bool | None = None
@@ -262,7 +260,7 @@ class AxolotlInputConfig(
 
     sequence_parallel_degree: int | None = None
     heads_k_stride: int | None = None
-    ring_attn_func: str | None = None
+    ring_attn_func: RingAttnFunc | None = None
 
     special_tokens: SpecialTokensConfig | None = None
     tokens: list[str] | None = None
@@ -464,13 +462,24 @@ class AxolotlInputConfig(
             and not data.get("flash_attention")
             and not data.get("sdp_attention")
             and not data.get("flex_attention")
+            and not data.get("xformers_attention")
         ):
             LOG.warning(
-                "sample_packing without flash, sdp or flex attention does not handle cross sample decontamination."
+                "sample_packing without flash, sdp, xformers or flex attention does not handle cross sample decontamination."
             )
 
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_sample_packing_with_s2attn(cls, data):
+        if data.get("sample_packing") and data.get("s2_attention"):
+            raise ValueError(
+                "Received `sample_packing=true` and `s2_attention=true`; however, \
+                shifted-sparse attention does not currently support sample packing."
+            )
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_batch_flattening_fa(cls, data):
@@ -783,7 +792,7 @@ class AxolotlInputConfig(
 
     @model_validator(mode="after")
     def check_simpo_warmup(self):
-        if self.rl == "simpo" and self.warmup_ratio:
+        if self.rl is RLType.SIMPO and self.warmup_ratio:
             raise ValueError(
                 "warmup_ratio is not supported with the simpo trainer. Please use `warmup_steps` instead"
             )
@@ -1150,16 +1159,28 @@ class AxolotlInputConfig(
 
         return data
 
+    # @model_validator(mode="before")
+    # @classmethod
+    # def check_grpo_peft_liger(cls, data):
+    #     if (
+    #         data.get("rl") == "grpo"
+    #         and data.get("trl", {})
+    #         and data.get("trl").get("use_liger_loss")
+    #         and data.get("adapter")
+    #     ):
+    #         raise ValueError("PEFT + GRPO + Liger is not yet supported")
+    #     return data
+    #
     @model_validator(mode="before")
     @classmethod
-    def check_grpo_peft_liger(cls, data):
+    def check_grpo_liger_sequence_parallel(cls, data):
         if (
             data.get("rl") == "grpo"
             and data.get("trl", {})
             and data.get("trl").get("use_liger_loss")
-            and data.get("adapter")
+            and data.get("sequence_parallel_degree", 1) > 1
         ):
-            raise ValueError("PEFT + GRPO + Liger is not yet supported")
+            raise ValueError("GRPO + SP + Liger not currently supported")
         return data
 
     @model_validator(mode="after")
@@ -1174,7 +1195,7 @@ class AxolotlInputConfig(
 
             if self.sample_packing and self.micro_batch_size > 1:
                 raise ValueError(
-                    "micro_batch_size must be set to 1 when sample_packing is enabled"
+                    "micro_batch_size must be set to 1 when sample_packing is enabled "
                     "due to a `ring-flash-attn` requirement"
                 )
 
@@ -1206,16 +1227,8 @@ class AxolotlInputConfig(
         if getattr(self, "sequence_parallel_degree", 1) == 1:
             return self
 
-        from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
-
         if self.ring_attn_func is not None:
-            valid_funcs = list(RingAttnFunc)
-            if self.ring_attn_func in valid_funcs:
-                self.ring_attn_func = RingAttnFunc(self.ring_attn_func)
-            else:
-                raise ValueError(
-                    f"ring_attn_func: {self.ring_attn_func} must be in {valid_funcs}"
-                )
+            self.ring_attn_func = RingAttnFunc(self.ring_attn_func)
         else:
             # Default ring attention function selection
             sample_packing = getattr(self, "sample_packing", False)
@@ -1346,6 +1359,10 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
             ):
                 return data
 
+            # Skip if dropout is not 0, as auto enabling it would just disable it during runtime patch checks
+            if data.get("lora_dropout") != 0:
+                return data
+
             # Check multi-GPU compatibility
             capabilities = data.get("capabilities")
             is_multi_gpu = capabilities and capabilities.get("n_gpu", 0) > 1

src/axolotl/utils/schemas/enums.py

@@ -6,12 +6,12 @@ from enum import Enum
 class RLType(str, Enum):
     """RL trainer type configuration subset"""
 
-    dpo = "dpo"  # pylint: disable=invalid-name
-    grpo = "grpo"  # pylint: disable=invalid-name
-    ipo = "ipo"  # pylint: disable=invalid-name
-    orpo = "orpo"  # pylint: disable=invalid-name
-    kto = "kto"  # pylint: disable=invalid-name
-    simpo = "simpo"  # pylint: disable=invalid-name
+    DPO = "dpo"  # pylint: disable=invalid-name
+    GRPO = "grpo"  # pylint: disable=invalid-name
+    IPO = "ipo"  # pylint: disable=invalid-name
+    ORPO = "orpo"  # pylint: disable=invalid-name
+    KTO = "kto"  # pylint: disable=invalid-name
+    SIMPO = "simpo"  # pylint: disable=invalid-name
 
 
 class ChatTemplate(str, Enum):
@@ -53,4 +53,16 @@ class CustomSupportedOptimizers(str, Enum):
     ao_adamw_8bit = "ao_adamw_8bit"  # pylint: disable=invalid-name
     ao_adamw_fp8 = "ao_adamw_fp8"  # pylint: disable=invalid-name
     adopt_adamw = "adopt_adamw"  # pylint: disable=invalid-name
+    came_pytorch = "came_pytorch"  # pylint: disable=invalid-name
     muon = "muon"  # pylint: disable=invalid-name
+
+
+class RingAttnFunc(str, Enum):
+    """Enum class for supported `ring-flash-attn` implementations"""
+
+    # VARLEN_RING = "varlen_ring"
+    # VARLEN_ZIGZAG = "varlen_zigzag"
+    VARLEN_LLAMA3 = "varlen_llama3"
+    BATCH_RING = "batch_ring"
+    # BATCH_ZIGZAG = "batch_zigzag"
+    # BATCH_STRIPE = "batch_stripe"

src/axolotl/utils/schemas/training.py

@@ -75,8 +75,10 @@ class HyperparametersConfig(BaseModel):
     lr_groups: list[LrGroup] | None = None
 
     adam_epsilon: float | None = None
+    adam_epsilon2: float | None = None
     adam_beta1: float | None = None
     adam_beta2: float | None = None
+    adam_beta3: float | None = None
     max_grad_norm: float | None = None
     num_epochs: float = Field(default=1.0)
 

tests/conftest.py

@@ -4,6 +4,7 @@ shared pytest fixtures
 
 import functools
 import importlib
+import os
 import shutil
 import sys
 import tempfile
@@ -529,31 +530,32 @@ def dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff(
 
 
 # # pylint: disable=redefined-outer-name,unused-argument
-# def test_load_fixtures(
-#     download_smollm2_135m_model,
-#     download_llama_68m_random_model,
-#     download_qwen_2_5_half_billion_model,
-#     download_tatsu_lab_alpaca_dataset,
-#     download_mhenrichsen_alpaca_2k_dataset,
-#     download_mhenrichsen_alpaca_2k_w_revision_dataset,
-#     download_mlabonne_finetome_100k_dataset,
-#     download_argilla_distilabel_capybara_dpo_7k_binarized_dataset,
-#     download_argilla_ultrafeedback_binarized_preferences_cleaned_dataset,
-#     download_fozzie_alpaca_dpo_dataset,
-#     download_arcee_ai_distilabel_intel_orca_dpo_pairs_dataset,
-#     download_argilla_dpo_pairs_dataset,
-#     download_tiny_shakespeare_dataset,
-#     download_deepseek_model_fixture,
-#     download_huggyllama_model_fixture,
-#     download_llama_1b_model_fixture,
-#     download_llama3_8b_model_fixture,
-#     download_llama3_8b_instruct_model_fixture,
-#     download_phi_35_mini_model_fixture,
-#     download_phi_3_medium_model_fixture,
-#     download_mistral_7b_model_fixture,
-#     download_gemma_2b_model_fixture,
-#     download_gemma2_9b_model_fixture,
-#     download_mlx_mistral_7b_model_fixture,
-#     download_llama2_model_fixture,
-# ):
-#     pass
+@pytest.mark.skipif(
+    os.environ.get("AXOLOTL_IS_CI_CACHE_PRELOAD", "-1") != "1",
+    reason="Not running in CI cache preload",
+)
+def test_load_fixtures(
+    download_smollm2_135m_model,
+    download_qwen_2_5_half_billion_model,
+    download_tatsu_lab_alpaca_dataset,
+    download_mhenrichsen_alpaca_2k_dataset,
+    download_mhenrichsen_alpaca_2k_w_revision_dataset,
+    download_mlabonne_finetome_100k_dataset,
+    download_argilla_distilabel_capybara_dpo_7k_binarized_dataset,
+    download_arcee_ai_distilabel_intel_orca_dpo_pairs_dataset,
+    download_argilla_dpo_pairs_dataset,
+    download_tiny_shakespeare_dataset,
+    download_deepseek_model_fixture,
+    download_huggyllama_model_fixture,
+    download_llama_1b_model_fixture,
+    download_llama3_8b_model_fixture,
+    download_llama3_8b_instruct_model_fixture,
+    download_phi_35_mini_model_fixture,
+    download_phi_3_medium_model_fixture,
+    download_mistral_7b_model_fixture,
+    download_gemma_2b_model_fixture,
+    download_gemma2_9b_model_fixture,
+    download_mlx_mistral_7b_model_fixture,
+    download_llama2_model_fixture,
+):
+    pass