set v0.9.2 version for tag

* offload activations to disk instead of CPU RAM

* add prefetch

* Disco :dance:

* include offload_disk in e2e test for AC

* document and make sure to cleanup

* fix annotation to match docs

* fix docs build

* address PR feedback

.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
@@ -199,15 +294,6 @@ 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:
     if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
@@ -261,18 +347,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            num_gpus: 1
-            axolotl_extras: llmcompressor
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            num_gpus: 1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -309,3 +383,43 @@ jobs:
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
+
+  docker-e2e-cleanup:
+    runs-on: [self-hosted, modal]
+    timeout-minutes: 90
+    needs: [docker-e2e-tests]
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            num_gpus: 1
+            axolotl_extras: vllm
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Install Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Install Modal
+        run: |
+          python -m pip install --upgrade pip
+          pip install modal==0.71.8 jinja2
+      - name: Update env vars
+        run: |
+          echo "BASE_TAG=main-base-py${{ 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

@@ -124,7 +124,8 @@ 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/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:

docs/custom_integrations.qmd

@@ -49,8 +49,7 @@ sections = [
     ("Knowledge Distillation (KD)", "kd"),
     ("Liger Kernels", "liger"),
     ("Language Model Evaluation Harness (LM Eval)", "lm_eval"),
-    ("Spectrum", "spectrum"),
-    ("LLMCompressor", "llm_compressor")
+    ("Spectrum", "spectrum")
 ]
 
 for section_name, folder_name in sections:

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}

examples/llama-3/sparse-finetuning.yaml

@@ -1,77 +0,0 @@
-base_model: neuralmagic/Sparse-Llama-3.1-8B-2of4
-
-plugins:
-  - axolotl.integrations.llm_compressor.LLMCompressorPlugin
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.05
-output_dir: ./outputs/out
-
-sequence_len: 4096
-sample_packing: true
-pad_to_sequence_len: true
-eval_sample_packing: false
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 8
-micro_batch_size: 1
-num_epochs: 1
-optimizer: paged_adamw_8bit
-lr_scheduler: cosine
-learning_rate: 2e-5
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 100
-evals_per_epoch: 2
-eval_table_size:
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-  pad_token: <|end_of_text|>
-
-llmcompressor:
-  recipe:
-    finetuning_stage:
-      finetuning_modifiers:
-        ConstantPruningModifier:
-          targets: [
-            're:.*q_proj.weight',
-            're:.*k_proj.weight',
-            're:.*v_proj.weight',
-            're:.*o_proj.weight',
-            're:.*gate_proj.weight',
-            're:.*up_proj.weight',
-            're:.*down_proj.weight',
-          ]
-          start: 0
-  save_compressed: true

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/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>

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]",
@@ -149,9 +150,6 @@ extras_require = {
     "vllm": [
         "vllm==0.7.2",
     ],
-    "llmcompressor": [
-        "llmcompressor==0.5.1",
-    ],
 }
 
 install_requires, dependency_links, extras_require_build = parse_requirements(

src/axolotl/__init__.py

@@ -4,4 +4,4 @@ import pkgutil
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.10.0.dev0"
+__version__ = "0.9.2"

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/art.py

@@ -16,8 +16,15 @@ AXOLOTL_LOGO = """
     @@@@  @@@@@@@@@@@@@@@@
 """
 
+HAS_PRINTED_LOGO = False
+
 
 def print_axolotl_text_art():
     """Prints axolotl ASCII art."""
+
+    global HAS_PRINTED_LOGO  # pylint: disable=global-statement
+    if HAS_PRINTED_LOGO:
+        return
     if is_main_process():
+        HAS_PRINTED_LOGO = True
         print(AXOLOTL_LOGO)

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/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

@@ -48,6 +48,7 @@ def load_datasets(
     *,
     cfg: DictDefault,
     cli_args: PreprocessCliArgs | TrainerCliArgs | None = None,
+    debug: bool = False,
 ) -> TrainDatasetMeta:
     """
     Loads one or more training or evaluation datasets, calling
@@ -56,6 +57,7 @@ def load_datasets(
     Args:
         cfg: Dictionary mapping `axolotl` config keys to values.
         cli_args: Command-specific CLI arguments.
+        debug: Whether to print out tokenization of sample
 
     Returns:
         Dataclass with fields for training and evaluation datasets and the computed
@@ -77,20 +79,25 @@ def load_datasets(
         preprocess_iterable=preprocess_iterable,
     )
 
-    if cli_args and (
-        cli_args.debug
-        or cfg.debug
-        or cli_args.debug_text_only
-        or int(cli_args.debug_num_examples) > 0
-    ):
+    if (  # pylint: disable=too-many-boolean-expressions
+        cli_args
+        and (
+            cli_args.debug
+            or cfg.debug
+            or cli_args.debug_text_only
+            or int(cli_args.debug_num_examples) > 0
+        )
+    ) or debug:
         LOG.info("check_dataset_labels...")
 
-        train_samples = sample_dataset(train_dataset, cli_args.debug_num_examples)
+        num_examples = cli_args.debug_num_examples if cli_args else 1
+        text_only = cli_args.debug_text_only if cli_args else False
+        train_samples = sample_dataset(train_dataset, num_examples)
         check_dataset_labels(
             train_samples,
             tokenizer,
-            num_examples=cli_args.debug_num_examples,
-            text_only=cli_args.debug_text_only,
+            num_examples=num_examples,
+            text_only=text_only,
         )
 
         LOG.info("printing prompters...")

src/axolotl/core/trainer_builder.py

@@ -21,6 +21,7 @@ import importlib.util
 import inspect
 import logging
 import math
+import os
 import sys
 from abc import abstractmethod
 from pathlib import Path
@@ -72,6 +73,7 @@ from axolotl.utils.callbacks import (
     SaveBetterTransformerModelCallback,
     bench_eval_callback_factory,
     causal_lm_bench_eval_callback_factory,
+    colab_inference_post_train_callback,
     log_prediction_callback_factory,
 )
 from axolotl.utils.callbacks.lisa import lisa_callback_factory
@@ -168,6 +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)
@@ -249,9 +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):
@@ -293,6 +295,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
             callbacks.append(lisa_callback_factory(trainer))
 
+        if any("COLAB_" in key for key in os.environ):
+            ColabCallback = colab_inference_post_train_callback(trainer)
+            callbacks.append(ColabCallback(self.cfg))
+
         callbacks.extend(super().get_post_trainer_create_callbacks(trainer=trainer))
         return callbacks
 
@@ -702,6 +708,20 @@ 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_beta2", 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:
@@ -1037,6 +1057,8 @@ 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
 
@@ -1166,6 +1188,10 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         else:
             raise ValueError(f"Unsupported RL: {self.cfg.rl}")
 
+        if self.cfg.plugins:
+            plugin_manager = PluginManager.get_instance()
+            trainer_cls = plugin_manager.get_trainer_cls(self.cfg)
+
         sig = inspect.signature(trainer_cls)
         if "tokenizer" in sig.parameters.keys():
             dpo_trainer_kwargs["tokenizer"] = self.tokenizer

src/axolotl/core/trainers/base.py

@@ -114,6 +114,8 @@ class AxolotlTrainer(
             packing_efficiency_estimate=self.args.sample_packing_efficiency,
             batch_max_len=batch_max_len,
             batch_size=batch_size,
+            group_size=self.args.sample_packing_group_size,
+            bin_size=self.args.sample_packing_bin_size,
             sequential=self.args.sample_packing_sequentially,
             drop_last=True,
         )
@@ -610,15 +612,3 @@ class AxolotlTrainer(
         output_dir = os.path.join(run_dir, checkpoint_folder)
         os.makedirs(output_dir, exist_ok=True)
         return super()._save_checkpoint(model, trial, **kwargs)
-
-    def compute_loss_context_manager(self):
-        from contextlib import ExitStack
-
-        from torchtune.training import OffloadActivations
-
-        stack = ExitStack()
-
-        stack.enter_context(super().compute_loss_context_manager())
-        stack.enter_context(OffloadActivations())
-
-        return stack

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

@@ -247,7 +247,9 @@ class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
                 )
 
         # Base evaluation
-        initial_output = super().evaluation_loop(
+        initial_output = super(  # pylint: disable=bad-super-call
+            DPOTrainer, self
+        ).evaluation_loop(
             dataloader,
             description,
             prediction_loss_only,

src/axolotl/integrations/base.py

@@ -26,6 +26,8 @@ from typing import OrderedDict
 import torch
 from torch.optim.lr_scheduler import LRScheduler
 
+from axolotl.utils.dict import DictDefault
+
 
 class BasePlugin:
     """
@@ -36,11 +38,13 @@ class BasePlugin:
 
     Methods:
     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.
@@ -63,20 +67,32 @@ class BasePlugin:
         None
         """
 
-    def get_input_args(self):
+    def get_input_args(self) -> str | None:
         """
         Returns a pydantic model for the plugin's input arguments.
         """
 
+    def load_datasets(self, cfg: DictDefault, preprocess: bool = False):
+        """
+        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:
+            dataset_meta: The metadata for the training dataset.
+        """
+
     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.
+        Args:
+            cfg (dict): The configuration for the plugin.
 
         Returns:
-        None
+            None
         """
 
     def post_model_build(self, cfg, model):  # pylint: disable=unused-argument
@@ -91,59 +107,71 @@ class BasePlugin:
         """
         Performs actions after the model is loaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg (dict): The configuration for the plugin.
+            model (object): The loaded model.
 
         Returns:
-        None
+            None
         """
 
     def pre_lora_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions before LoRA weights are loaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg (dict): The configuration for the plugin.
+            model (object): The loaded model.
 
         Returns:
-        None
+            None
         """
 
     def post_lora_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions after LoRA weights are loaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg (dict): The configuration for the plugin.
+            model (object): The loaded model.
 
         Returns:
-        None
+            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.
+        Args:
+            cfg (dict): The global axolotl configuration.
 
         Returns:
-        class: The class for the trainer.
+            class: The class for the trainer.
+        """
+
+    def post_trainer_create(self, cfg, trainer):  # pylint: disable=unused-argument
+        """
+        Performs actions after the trainer is created.
+
+        Args:
+            cfg (dict): The configuration for the plugin.
+            trainer (object): The trainer object for training.
+
+        Returns:
+            None
         """
 
     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.
+        Args:
+            cfg (dict): The configuration for the plugin.
+            trainer (object): The trainer object for training.
 
         Returns:
-        object: The created optimizer.
+            object: The created optimizer.
         """
 
     def create_lr_scheduler(
@@ -152,26 +180,26 @@ class BasePlugin:
         """
         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 (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
 
         Returns:
-        object (LRScheduler): The created learning rate scheduler.
+            object (LRScheduler): The created learning rate scheduler.
         """
 
     def add_callbacks_pre_trainer(self, cfg, model):  # pylint: disable=unused-argument
         """
         setup callbacks before creating the trainer.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg (dict): The configuration for the plugin.
+            model (object): The loaded model.
 
         Returns:
-        List[callable]: A list of callback functions to be added to the TrainingArgs
+            List[callable]: A list of callback functions to be added to the TrainingArgs
         """
         return []
 
@@ -182,12 +210,12 @@ class BasePlugin:
         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 (dict): The configuration for the plugin.
+            trainer (object): The trainer object for training.
 
         Returns:
-        List[callable]: A list of callback functions to be added
+            List[callable]: A list of callback functions to be added
         """
         return []
 
@@ -195,23 +223,23 @@ class BasePlugin:
         """
         Performs actions after training is complete.
 
-        Parameters:
-        cfg (dict): The axolotl configuration
-        model (object): The loaded model.
+        Args:
+            cfg (dict): The axolotl configuration
+            model (object): The loaded model.
 
         Returns:
-        None
+            None
         """
 
     def post_train_unload(self, cfg):  # pylint: disable=unused-argument
         """
         Performs actions after training is complete and the model is unloaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
+        Args:
+            cfg (dict): The configuration for the plugin.
 
         Returns:
-        None
+            None
         """
 
 
@@ -338,6 +366,27 @@ class PluginManager:
                 input_args.append(input_args_from_plugin)
         return input_args
 
+    def load_datasets(self, cfg, preprocess: bool = False):
+        """
+        Calls the load_datasets method of each registered plugin.
+
+        Args:
+            cfg: The configuration for the plugins.
+            preprocess : Whether this is preprocess step of the datasets.
+
+        Returns:
+            dataset_meta: 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):
         """
         Calls the pre_model_load method of all registered plugins.
@@ -422,6 +471,20 @@ class PluginManager:
                 return trainer_cls
         return None
 
+    def post_trainer_create(self, cfg, trainer):
+        """
+        Calls the post_trainer_create method of all registered plugins.
+
+        Parameters:
+        cfg (dict): The configuration for the plugins.
+        trainer (object): The trainer object for training.
+
+        Returns:
+        None
+        """
+        for plugin in self.plugins.values():
+            plugin.post_trainer_create(cfg, trainer)
+
     def create_optimizer(self, trainer):
         """
         Calls the create_optimizer method of all registered plugins and returns the first non-None optimizer.

src/axolotl/integrations/llm_compressor/README.md

@@ -1,108 +0,0 @@
-# LLMCompressor Integration
-
-Fine-tune sparsified models in Axolotl using Neural Magic's [LLMCompressor](https://github.com/vllm-project/llm-compressor).
-
-This integration enables fine-tuning of models sparsified using LLMCompressor within the Axolotl training framework. By combining LLMCompressor's model compression capabilities with Axolotl's distributed training pipelines, users can efficiently fine-tune sparse models at scale.
-
-It uses Axolotl’s plugin system to hook into the fine-tuning flows while maintaining sparsity throughout training.
-
----
-
-## Requirements
-
-- Axolotl with `llmcompressor` extras:
-
-  ```bash
-  pip install "axolotl[llmcompressor]"
-  ```
-
-- Requires `llmcompressor >= 0.5.1`
-
-This will install all necessary dependencies to fine-tune sparsified models using the integration.
-
----
-
-## Usage
-
-To enable sparse fine-tuning with this integration, include the plugin in your Axolotl config:
-
-```yaml
-plugins:
-  - axolotl.integrations.llm_compressor.LLMCompressorPlugin
-
-llmcompressor:
-  recipe:
-    finetuning_stage:
-      finetuning_modifiers:
-        ConstantPruningModifier:
-          targets: [
-            're:.*q_proj.weight',
-            're:.*k_proj.weight',
-            're:.*v_proj.weight',
-            're:.*o_proj.weight',
-            're:.*gate_proj.weight',
-            're:.*up_proj.weight',
-            're:.*down_proj.weight',
-          ]
-          start: 0
-  save_compressed: true
-# ... (other training arguments)
-```
-
-This plugin **does not apply pruning or sparsification itself** — it is intended for **fine-tuning models that have already been sparsified**.
-
-Pre-sparsified checkpoints can be:
-- Generated using [LLMCompressor](https://github.com/vllm-project/llm-compressor)
-- Downloaded from [Neural Magic's Hugging Face page](https://huggingface.co/neuralmagic)
-- Any custom LLM with compatible sparsity patterns that you've created yourself
-
-To learn more about writing and customizing LLMCompressor recipes, refer to the official documentation:
-[https://github.com/vllm-project/llm-compressor/blob/main/README.md](https://github.com/vllm-project/llm-compressor/blob/main/README.md)
-
-### Storage Optimization with save_compressed
-
-Setting `save_compressed: true` in your configuration enables saving models in a compressed format, which:
-- Reduces disk space usage by approximately 40%
-- Maintains compatibility with vLLM for accelerated inference
-- Maintains compatibility with llmcompressor for further optimization (example: quantization)
-
-This option is highly recommended when working with sparse models to maximize the benefits of model compression.
-
-### Example Config
-
-See [`examples/llama-3/sparse-finetuning.yaml`](examples/llama-3/sparse-finetuning.yaml) for a complete example.
-
----
-
-## Inference with vLLM
-
-After fine-tuning your sparse model, you can leverage vLLM for efficient inference.
-You can also use LLMCompressor to apply additional quantization to your fine-tuned
-sparse model before inference for even greater performance benefits.:
-
-```python
-from vllm import LLM, SamplingParams
-
-prompts = [
-    "Hello, my name is",
-    "The president of the United States is",
-    "The capital of France is",
-    "The future of AI is",
-]
-sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
-llm = LLM("path/to/your/sparse/model")
-outputs = llm.generate(prompts, sampling_params)
-
-for output in outputs:
-    prompt = output.prompt
-    generated_text = output.outputs[0].text
-    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
-```
-
-For more details on vLLM's capabilities and advanced configuration options, see the [official vLLM documentation](https://docs.vllm.ai/).
-
-## Learn More
-
-For details on available sparsity and quantization schemes, fine-tuning recipes, and usage examples, visit the official LLMCompressor repository:
-
-[https://github.com/vllm-project/llm-compressor](https://github.com/vllm-project/llm-compressor)

src/axolotl/integrations/llm_compressor/__init__.py

@@ -1,5 +0,0 @@
-"""Integration entry point for the LLMCompressor plugin."""
-
-from .plugin import LLMCompressorPlugin
-
-__all__ = ["LLMCompressorPlugin"]

src/axolotl/integrations/llm_compressor/args.py

@@ -1,40 +0,0 @@
-"""
-LLMCompressor and Sparse Finetuning config models.
-"""
-
-from typing import Any
-
-from pydantic import BaseModel, Field
-from typing_extensions import Annotated
-
-
-class CompressionArgs(BaseModel):
-    """Sparse Finetuning config for LLMCompressor."""
-
-    # Typing for recipe is set to Any due to:
-    # https://github.com/vllm-project/llm-compressor/issues/1319
-    recipe: Annotated[
-        Any,
-        Field(
-            description="The recipe containing the compression algorithms and hyperparameters to apply."
-        ),
-    ]
-
-    save_compressed: Annotated[
-        bool,
-        Field(
-            default=False,
-            description="Whether to save the compressed model after training.",
-        ),
-    ]
-
-
-class LLMCompressorArgs(BaseModel):
-    """LLMCompressor configuration BaseModel."""
-
-    llmcompressor: Annotated[
-        CompressionArgs,
-        Field(
-            description="Arguments enabling compression pathways through the LLM Compressor plugins"
-        ),
-    ]

src/axolotl/integrations/llm_compressor/plugin.py

@@ -1,171 +0,0 @@
-"""
-Sparse Finetuning plugin for Axolotl — enables handling of sparse neural networks
-by maintaining masks for zero weights during training.
-"""
-
-import logging
-from functools import wraps
-from typing import Any, Callable, Concatenate, ParamSpec, TypeVar
-
-from llmcompressor import active_session, create_session
-from llmcompressor.core import callbacks as session_callbacks
-from llmcompressor.recipe import Recipe
-from torch.nn import Module
-from transformers.trainer import Trainer
-from transformers.trainer_callback import TrainerCallback, TrainerControl, TrainerState
-from transformers.training_args import TrainingArguments
-
-from axolotl.integrations.base import BasePlugin
-
-P = ParamSpec("P")  # Params for generic function signatures
-R = TypeVar("R")  # Return type for generic function signatures
-
-LOG = logging.getLogger("axolotl.integrations.llm_compressor")
-
-
-class LLMCompressorCallbackHandler(TrainerCallback):
-    """
-    Trainer callback for Sparse Finetuning.
-    Maintains sparsity patterns during training by applying masks after optimization steps,
-    ensuring zero-weight updates are canceled out.
-    """
-
-    def __init__(self, trainer: Trainer, recipe: Any):
-        """
-        Initialize the Sparse Finetuning callback handler.
-
-        Args:
-            trainer (Trainer): Huggingface Trainer instance.
-            recipe (Recipe | dict): Sparse finetuning recipe to apply.
-        """
-        super().__init__()
-        self.trainer = trainer
-        self.recipe = (
-            Recipe.model_validate(recipe) if not isinstance(recipe, Recipe) else recipe
-        )
-        self.original_compute_loss = trainer.compute_loss
-        self.trainer.compute_loss = compute_loss_wrapper(self.trainer.compute_loss)
-        create_session()
-
-    def on_train_begin(
-        self,
-        args: TrainingArguments,
-        state: TrainerState,
-        control: TrainerControl,
-        **kwargs,
-    ) -> None:
-        """
-        Called at the beginning of training. Initializes the compression session.
-
-        Args:
-            args (TrainingArguments): Training arguments.
-            state (TrainerState): Trainer state.
-            control (TrainerControl): Trainer control.
-        """
-        super().on_train_begin(args, state, control, **kwargs)
-        self.trainer.accelerator.wait_for_everyone()
-        active_session().initialize(
-            model=self.trainer.model,
-            optimizer=self.trainer.optimizer,
-            start=state.epoch,
-            recipe=self.recipe,
-        )
-        self.trainer.accelerator.wait_for_everyone()
-
-    def on_step_begin(
-        self,
-        args: TrainingArguments,
-        state: TrainerState,
-        control: TrainerControl,
-        **kwargs,
-    ) -> None:
-        """
-        Called at the beginning of a training step. Triggers batch_start callback.
-        """
-        super().on_step_begin(args, state, control, **kwargs)
-        session_callbacks.batch_start()
-
-    def on_step_end(
-        self,
-        args: TrainingArguments,
-        state: TrainerState,
-        control: TrainerControl,
-        **kwargs,
-    ) -> None:
-        """
-        Called at the end of a training step. Triggers optimizer and batch_end callbacks.
-        """
-        super().on_step_end(args, state, control, **kwargs)
-        session_callbacks.optim_pre_step()
-        session_callbacks.optim_post_step()
-        session_callbacks.batch_end()
-
-    def on_train_end(
-        self,
-        args: TrainingArguments,
-        state: TrainerState,
-        control: TrainerControl,
-        **kwargs,
-    ) -> None:
-        """
-        Called at the end of training. Finalizes the compression session.
-        """
-        super().on_train_end(args, state, control, **kwargs)
-        active_session().finalize()
-        self.trainer.compute_loss_func = self.original_compute_loss
-
-
-class LLMCompressorPlugin(BasePlugin):
-    """
-    Sparse Finetuning plugin for Axolotl integration.
-    """
-
-    def get_input_args(self) -> str:
-        """
-        Returns the path to the plugin's argument definition.
-
-        Returns:
-            str: Dotted path to the LLMCompressorArgs class.
-        """
-        return "axolotl.integrations.llm_compressor.args.LLMCompressorArgs"
-
-    def add_callbacks_post_trainer(self, cfg: Any, trainer: Trainer) -> list:
-        """
-        Adds Sparse Finetuning callback to the Trainer instance.
-
-        Args:
-            cfg (Any): Configuration object containing the sparse recipe.
-            trainer (Trainer): Huggingface Trainer instance.
-
-        Returns:
-            list: List containing the configured callback instances.
-        """
-        LOG.info("Adding Sparse Finetuning callback to the trainer")
-        callback = LLMCompressorCallbackHandler(
-            trainer=trainer,
-            recipe=cfg.llmcompressor.recipe,
-        )
-        return [callback]
-
-
-def compute_loss_wrapper(
-    compute_loss_func: Callable[Concatenate[Module, P], R],
-) -> Callable[Concatenate[Module, P], R]:
-    """
-    Wraps the loss computation function to trigger the loss_calculated callback.
-
-    Args:
-        compute_loss_func (Callable): Original loss computation function.
-
-    Returns:
-        Callable: Wrapped function that also invokes the loss_calculated callback.
-    """
-
-    @wraps(compute_loss_func)
-    def compute_and_notify(model: Module, *args: P.args, **kwargs: P.kwargs) -> R:
-        loss = compute_loss_func(model, *args, **kwargs)
-        if active_session().lifecycle.initialized_ and model.training:
-            session_callbacks.loss_calculated(loss=loss)
-        return loss
-
-    return compute_and_notify

src/axolotl/integrations/llm_compressor/utils.py

@@ -1,40 +0,0 @@
-"""Utilities for llmcompressor integration with axolotl."""
-
-from typing import Union
-
-from llmcompressor.transformers.sparsification.compressed_tensors_utils import (
-    modify_save_pretrained,
-)
-from transformers import PreTrainedModel, Trainer
-
-
-def save_compressed_model(
-    model: PreTrainedModel,
-    output_dir: Union[str, bytes],
-    trainer: Trainer,
-    safe_serialization: bool = False,
-    save_compressed: bool = False,
-) -> None:
-    """
-    Synchronize processes, apply compression hooks, and save the model.
-
-    Args:
-        model (PreTrainedModel): The model to be saved.
-        output_dir (str or bytes): Path where the model files will be written.
-        trainer (Trainer): Hugging Face Trainer for process synchronization.
-        safe_serialization (bool): Use safe serialization if True.
-        save_compressed (bool): Write compressed tensors if True.
-    """
-    trainer.accelerator.wait_for_everyone()
-
-    # Only the main process writes the files
-    if not trainer.accelerator.is_main_process:
-        return
-
-    modify_save_pretrained(model)
-    model.save_pretrained(
-        output_dir,
-        safe_serialization=safe_serialization,
-        save_compressed=save_compressed,
-        skip_sparsity_compression_stats=not save_compressed,
-    )

src/axolotl/monkeypatch/attention/__init__.py

@@ -0,0 +1,19 @@
+"""
+attention module for attention monkeypatches
+"""
+
+from transformers.integrations.flash_attention import flash_attention_forward
+
+
+def patch_xformers_attn_over_fa2():
+    from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+
+    from .xformers import xformers_attention_forward
+
+    ALL_ATTENTION_FUNCTIONS["flash_attention_2"] = xformers_attention_forward
+
+
+def unpatch_xformers_attn_over_fa2():
+    from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
+
+    ALL_ATTENTION_FUNCTIONS["flash_attention_2"] = flash_attention_forward()

src/axolotl/monkeypatch/attention/xformers.py

@@ -0,0 +1,160 @@
+"""
+xformers attention implementation for packing
+"""
+
+from typing import Optional
+
+import torch
+import xformers
+import xformers.ops.fmha
+from transformers.modeling_flash_attention_utils import (
+    _upad_input,
+)
+
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+
+xformers_attention = xformers.ops.fmha.memory_efficient_attention
+
+
+def xformers_attention_forward(
+    module: torch.nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    dropout: float = 0.0,  # pylint: disable=unused-argument
+    scaling: Optional[float] = None,  # pylint: disable=unused-argument
+    sliding_window: Optional[int] = None,  # pylint: disable=unused-argument
+    softcap: Optional[float] = None,  # pylint: disable=unused-argument
+    cu_seq_lens_q: Optional[torch.LongTensor] = None,
+    cu_seq_lens_k: Optional[torch.LongTensor] = None,
+    max_length_q: Optional[int] = None,
+    max_length_k: Optional[int] = None,  # pylint: disable=unused-argument
+    **kwargs,  # pylint: disable=unused-argument
+):
+    # Get dimensions
+    # query: [batch, heads, seq_len, hidden_dim]
+    batch_size = query.size(0)
+    query_length = query.shape[2]
+    key_length = key.shape[2]
+
+    # Default causal mask
+    attn_bias = xformers.ops.LowerTriangularMask()
+
+    # Check if we have sliding window attention
+    has_sliding_window = sliding_window is not None and sliding_window < query_length
+
+    # Transpose dimensions for xformers (Q: [b, h, s, d] -> [b, s, h, d])
+    query = query.transpose(1, 2)
+    key = key.transpose(1, 2)
+    value = value.transpose(1, 2)
+
+    # Get GQA parameters
+    num_attention_heads = module.config.num_attention_heads
+    num_key_value_heads = module.config.num_key_value_heads
+    head_dim = query.size(-1)
+    is_gqa = num_attention_heads != num_key_value_heads
+    n_groups = num_attention_heads // num_key_value_heads if is_gqa else 1
+
+    # If position_ids is provided and check all examples do not contain only 1 sequence, If tensor in increasing
+    # then we probably have one sequence, otherwise it is packed. Additionally check we are in pre-fill/training stage.
+    # Use `flash_attn_varlen_func` to prevent cross-example attention and also allow padding free approach
+    if position_ids is not None and (
+        max_length_q is not None
+        or (query_length != 1 and not (torch.diff(position_ids, dim=-1) >= 0).all())
+    ):
+        if cu_seq_lens_q is None or cu_seq_lens_k is None:
+            cu_seq_lens_q = get_cu_seqlens_from_pos_ids(position_ids)[0]
+            cu_seq_lens_q = cu_seq_lens_q.squeeze()
+            seq_lengths = cu_seq_lens_q[1:] - cu_seq_lens_q[:-1]
+            attn_bias = (
+                xformers.ops.fmha.attn_bias.BlockDiagonalCausalMask.from_seqlens(
+                    q_seqlen=seq_lengths.tolist(),
+                )
+            )
+        else:
+            query = query.reshape(-1, query.size(-2), query.size(-1))
+            key = key.reshape(-1, key.size(-2), key.size(-1))
+            value = value.reshape(-1, value.size(-2), value.size(-1))
+
+        # Handle GQA
+        if is_gqa:
+            key = key.repeat_interleave(n_groups, dim=2)
+            value = value.repeat_interleave(n_groups, dim=2)
+
+    elif attention_mask is not None:
+        query, key, value, _, cu_seq_lens, _ = _upad_input(
+            query, key, value, attention_mask, query_length
+        )
+        cu_seq_lens_q, cu_seq_lens_k = cu_seq_lens
+        seq_lengths = []
+        for i in range(len(cu_seq_lens_q) - 1):
+            seq_lengths.append(cu_seq_lens_q[i + 1] - cu_seq_lens_q[i])
+        attn_bias = xformers.ops.fmha.attn_bias.BlockDiagonalCausalMask.from_seqlens(
+            q_seqlen=seq_lengths,
+            kv_seqlen=seq_lengths,
+        )
+
+        # Handle GQA
+        if is_gqa:
+            key = key.repeat_interleave(n_groups, dim=2)
+            value = value.repeat_interleave(n_groups, dim=2)
+    else:
+        # Handle Group Query Attention (GQA) using view/expand approach from reference
+        key = key.view(batch_size, key_length, num_key_value_heads, 1, head_dim)
+        value = value.view(batch_size, key_length, num_key_value_heads, 1, head_dim)
+        key = key.expand(
+            batch_size, key_length, num_key_value_heads, n_groups, head_dim
+        )
+        value = value.expand(
+            batch_size, key_length, num_key_value_heads, n_groups, head_dim
+        )
+
+        if module.training:
+            key = key.reshape(batch_size, key_length, num_attention_heads, head_dim)
+            value = value.reshape(batch_size, key_length, num_attention_heads, head_dim)
+
+            if has_sliding_window:
+                query = query.view(
+                    1, batch_size * query_length, num_attention_heads, head_dim
+                )
+                key = key.view(
+                    1, batch_size * key_length, num_attention_heads, head_dim
+                )
+                value = value.view(
+                    1, batch_size * key_length, num_attention_heads, head_dim
+                )
+        else:
+            query = query.view(
+                batch_size, query_length, num_key_value_heads, n_groups, head_dim
+            )
+
+            # If we need a sliding window attention
+            if has_sliding_window:
+                query = query.view(
+                    1,
+                    batch_size * query_length,
+                    num_key_value_heads,
+                    n_groups,
+                    head_dim,
+                )
+                key = key.view(
+                    1, batch_size * key_length, num_key_value_heads, n_groups, head_dim
+                )
+                value = value.view(
+                    1, batch_size * key_length, num_key_value_heads, n_groups, head_dim
+                )
+
+    # Run the xformers attention
+    attn_output = xformers_attention(
+        query,
+        key,
+        value,
+        attn_bias=attn_bias,
+    )
+
+    attn_output = attn_output.view(
+        batch_size, -1, attn_output.size(-2), attn_output.size(-1)
+    )
+    return attn_output, None

src/axolotl/monkeypatch/peft/utils.py

@@ -0,0 +1,78 @@
+"""
+Patch prepare_model_for_kbit_training to not upcast everything
+"""
+
+import inspect
+import logging
+
+import peft
+
+import axolotl
+from axolotl.monkeypatch.utils import detab_code
+
+LOG = logging.getLogger(__name__)
+
+ORIGINAL_PREPARE_CODE = """
+        for param in model.parameters():
+            if (
+                (param.dtype == torch.float16) or (param.dtype == torch.bfloat16)
+            ) and param.__class__.__name__ != "Params4bit":
+                param.data = param.data.to(torch.float32)
+"""
+
+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 all(embed_name not in name for embed_name in ["embed_tokens", "lm_head"]):
+                param.data = param.data.to(torch.float32)
+"""
+
+
+def get_peft_prep_code() -> str:
+    prepare = inspect.getsource(peft.utils.other.prepare_model_for_kbit_training)
+    return prepare
+
+
+def check_peft_prep_code_is_patchable() -> bool:
+    prep_code = get_peft_prep_code()
+    prep_code, _ = detab_code(prep_code)
+    return ORIGINAL_PREPARE_CODE in prep_code
+
+
+def patch_peft_prep_code():
+    """
+    monkeypatch create_accelerator_and_postprocess so it checks for additional kwargs
+    """
+
+    try:
+        prep_code = get_peft_prep_code()
+    except OSError:
+        return
+    peft.utils.other._original_create_accelerator_and_postprocess = (  # pylint: disable=protected-access
+        prep_code
+    )
+    prep_code, _ = detab_code(prep_code)
+    if ORIGINAL_PREPARE_CODE not in prep_code:
+        return
+
+    prep_code = prep_code.replace(ORIGINAL_PREPARE_CODE, PATCHED_PREPARE_CODE)
+    prep_code = prep_code.replace(
+        "def prepare_model_for_kbit_training(",
+        "def fixed_prepare_model_for_kbit_training(",
+        1,
+    )
+
+    items_to_import = []
+    for item in dir(peft.utils.other):
+        if item in prep_code:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from peft.utils.other import (" + ", ".join(x for x in items_to_import) + ")",
+        globals(),
+    )
+    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

src/axolotl/train.py

@@ -2,6 +2,7 @@
 
 import importlib
 import inspect
+import logging
 import os
 import signal
 import sys
@@ -12,7 +13,6 @@ 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
@@ -21,6 +21,7 @@ from transformers import PreTrainedModel, PreTrainedTokenizer, ProcessorMixin
 from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
 from transformers.trainer import Trainer
 
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.common.datasets import TrainDatasetMeta
 from axolotl.contribs.lgpl import (  # pylint: disable = no-name-in-module
     fix_untrained_tokens,
@@ -41,7 +42,7 @@ try:
 except ImportError:
     BetterTransformer = None
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def setup_model_and_tokenizer(
@@ -62,7 +63,6 @@ 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)
 
@@ -294,23 +294,8 @@ def save_trained_model(
             trainer.model.save_pretrained(
                 cfg.output_dir, safe_serialization=safe_serialization
             )
-
         model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
 
-    if hasattr(cfg, "llmcompressor") and cfg.llmcompressor:
-        # TODO: add integration support so this can be implemented completely within the plugin
-        from axolotl.integrations.llm_compressor.utils import (
-            save_compressed_model,
-        )
-
-        save_compressed_model(
-            model=model,
-            output_dir=cfg.output_dir,
-            trainer=trainer,
-            safe_serialization=safe_serialization,
-            save_compressed=cfg.llmcompressor.save_compressed,
-        )
-
 
 def create_model_card(cfg: DictDefault, trainer: Trainer):
     """
@@ -516,6 +501,8 @@ def train(
     Returns:
         Tuple of (model, tokenizer) after training
     """
+    print_axolotl_text_art()
+
     # Setup model, tokenizer, (causal or RLHF) trainer, etc.
     (
         trainer,
@@ -525,6 +512,9 @@ 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
@@ -547,7 +537,6 @@ 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

@@ -868,3 +868,28 @@ class GCCallback(TrainerCallback):
     ):
         torch.cuda.empty_cache()
         gc.collect()
+
+
+def colab_inference_post_train_callback(trainer: Trainer):
+    class ColabCallback(TrainerCallback):
+        """Callback to prep model for inference on Google Colab"""
+
+        def __init__(self, cfg):
+            self.gpu_name = torch.cuda.get_device_name(0)
+            self.cfg = cfg
+
+        def on_train_end(
+            self, args, state, control, **kwargs
+        ):  # pylint: disable=unused-argument
+            """
+            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.config._attn_implementation = (  # pylint: disable=protected-access
+                    "eager"
+                )
+            trainer.model.gradient_checkpointing_disable()
+            trainer.model.config.use_cache = True
+            trainer.model.eval()
+
+    return ColabCallback

src/axolotl/utils/config/__init__.py

@@ -70,6 +70,9 @@ def resolve_dtype(cfg):
             if cfg.fp16 is None and not cfg.float16:
                 cfg.fp16 = True
 
+    if cfg.fp16 and cfg.bf16 == "auto":
+        cfg.bf16 = False
+
     if cfg.device == "mps":
         cfg.load_in_8bit = False
         cfg.tf32 = False

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

@@ -1,23 +1,40 @@
 """custom checkpointing utils"""
 
+import importlib
 from functools import partial
 
-import torch
-from torch.utils.checkpoint import (
-    CheckpointPolicy,
-    checkpoint,
-    create_selective_checkpoint_contexts,
+from packaging import version
+
+from axolotl.utils.gradient_checkpointing.offload_cpu import (
+    CPU_Offloaded_Gradient_Checkpointer,
+)
+from axolotl.utils.gradient_checkpointing.offload_disk import (
+    Disco,
 )
 
-from axolotl.utils.gradient_checkpointing.unsloth import (
-    Unsloth_Offloaded_Gradient_Checkpointer,
-)
+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
-    return Unsloth_Offloaded_Gradient_Checkpointer.apply(
+    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)
@@ -27,30 +44,20 @@ def hf_grad_checkpoint_offload_wrapper(
     )
 
 
-aten = torch.ops.aten
-compute_intensive_ops = [
-    aten.mm.default,
-    aten.bmm.default,
-    aten.addmm.default,
-]
-
-
-def policy_fn(ctx, op, *args, **kwargs):
-    if op in compute_intensive_ops:
-        return CheckpointPolicy.MUST_SAVE
-    else:
-        return CheckpointPolicy.PREFER_RECOMPUTE
-
-
-context_fn = partial(create_selective_checkpoint_contexts, policy_fn)
-
-
-def checkpoint_w_policy(
+def hf_grad_checkpoint_disk_offload_wrapper(
     decoder_layer, *args, use_reentrant=None
 ):  # pylint: disable=unused-argument
-    return checkpoint(
-        decoder_layer,
+    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,
-        use_reentrant=use_reentrant,
-        context_fn=context_fn,
     )

src/axolotl/utils/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/utils/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/utils/models.py

@@ -70,7 +70,10 @@ from axolotl.utils.distributed import (
     is_local_main_process,
     is_main_process,
 )
-from axolotl.utils.gradient_checkpointing import hf_grad_checkpoint_offload_wrapper
+from axolotl.utils.gradient_checkpointing import (
+    hf_grad_checkpoint_disk_offload_wrapper,
+    hf_grad_checkpoint_offload_wrapper,
+)
 from axolotl.utils.lora_embeddings import get_linear_embedding_layers
 from axolotl.utils.model_shard_quant import load_sharded_model, load_sharded_model_quant
 
@@ -141,22 +144,6 @@ def check_model_config(cfg: DictDefault, model_config: PretrainedConfig):
         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
@@ -556,11 +543,21 @@ class ModelLoader:
         self.auto_model_loader = AutoModelForCausalLM  # pylint: disable=invalid-name
 
     def apply_patches(self) -> None:
+        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
         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()
 
+        if self.cfg.adapter and self.cfg.embeddings_skip_upcast:
+            from axolotl.monkeypatch.peft.utils import patch_peft_prep_code
+
+            patch_peft_prep_code()
+
         if self.cfg.flex_attention:
             from axolotl.monkeypatch.attention.flex_attn import (
                 patch_flex_make_mask,
@@ -609,6 +606,10 @@ class ModelLoader:
 
         if self.cfg.gradient_checkpointing in ["unsloth", "offload"]:
             transformers.modeling_utils.checkpoint = hf_grad_checkpoint_offload_wrapper
+        if self.cfg.gradient_checkpointing == "offload_disk":
+            transformers.modeling_utils.checkpoint = (
+                hf_grad_checkpoint_disk_offload_wrapper
+            )
 
         if self.cfg.flash_attention:
             self.patch_attention()
@@ -1180,7 +1181,7 @@ class ModelLoader:
                 ],
             )
 
-    def prepare_model(self, qlora_fsdp) -> None:
+    def prepare_model(self, qlora_fsdp: bool) -> None:
         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
@@ -1310,7 +1311,10 @@ class ModelLoader:
         # make sure these are fp32 per Ramesh et al. (2021)
         embedding_modules = get_linear_embedding_layers(self.cfg.model_config_type)
         if not self.cfg.fsdp:
-            # FSDP doesn't like mixed Float and BFloat16
+            # we don't run this during FSDP because this will leave mixed
+            # float 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,

src/axolotl/utils/samplers/multipack.py

@@ -1,10 +1,13 @@
-# pylint: skip-file
 """
-Multipack Batch Sampler
+Multipack Batch Sampler - An efficient batch sampler for packing variable-length sequences
+into fixed-capacity batches to optimize memory usage and training throughput.
 """
+
 import logging
 import math
-from typing import Any, Iterable, List, Union
+from concurrent.futures import ProcessPoolExecutor
+from multiprocessing import cpu_count, get_context
+from typing import Iterable, Union
 
 import numba
 import numpy as np
@@ -13,26 +16,39 @@ from torch.utils.data import BatchSampler, Sampler, SequentialSampler
 from axolotl.utils.distributed import reduce_and_broadcast
 
 LOG = logging.getLogger(__name__)
-
 LOG.setLevel(logging.INFO)
 
 
 @numba.njit
-def ffd_check(a: np.ndarray, c: int, n: int):
-    # First-fit-decreasing bin packing
-    # Check if a[] could fit in n bins with capacity c
-    # https://en.wikipedia.org/wiki/First-fit-decreasing_bin_packing
+def ffd_check(sequence_lengths: np.ndarray, bin_capacity: int, num_bins: int):
+    """
+    First-fit-decreasing bin packing algorithm check
 
-    a = np.sort(a)[::-1]
-    bins = np.full((n,), c, dtype=a.dtype)
-    for size in a:
+    Checks if sequences with the given lengths could fit in the specified number of bins
+
+    Args:
+        sequence_lengths: Array of sequence lengths
+        bin_capacity: Maximum capacity of each bin
+        num_bins: Number of bins available
+
+    Returns:
+        True if all sequences can be packed, False otherwise
+    """
+    # Sort sequence lengths in descending order for optimal packing
+    sequence_lengths = np.sort(sequence_lengths)[::-1]
+    # Initialize all bins with full capacity
+    bins = np.full((num_bins,), bin_capacity, dtype=sequence_lengths.dtype)
+
+    # Try to place each sequence in the first bin it fits
+    for size in sequence_lengths:
         not_found = True
-        for idx in range(n):
+        for idx in range(num_bins):
             if bins[idx] >= size:
                 bins[idx] -= size
                 not_found = False
                 break
 
+        # If no bin could fit this sequence, packing failed
         if not_found:
             return False
 
@@ -40,86 +56,155 @@ def ffd_check(a: np.ndarray, c: int, n: int):
 
 
 @numba.njit
-def ffd_with_result(a: np.ndarray, c: int, start_index: int):
-    # First-fit-decreasing bin packing (with result return)
+def pack_group(
+    sequence_lengths: np.ndarray,
+    group_offset: int,
+    bin_capacity: int,
+    max_bins: int,
+    bin_size: int,
+    safe_mode: bool = True,
+):
+    """
+    Pack a group of sequences into bins using First-Fit Decreasing algorithm
 
-    indices = np.argsort(a)[::-1]
-    a = a[indices]
+    Args:
+        sequence_lengths: Array of sequence lengths
+        group_offset: Offset to apply to indices when returning results
+        bin_capacity: Maximum capacity of each bin
+        max_bins: Maximum number of bins to use
+        bin_size: Maximum number of sequences per bin
+        safe_mode: If True, use a more conservative packing approach
 
-    bins: List[Any] = []
-    bins_result: List[Any] = []
-    for a_id, size in enumerate(a):
-        add_new = True
-        for idx in range(len(bins)):
-            if bins[idx] >= size:
-                bins[idx] -= size
-                bins_result[idx].append(indices[a_id] + start_index)
-                add_new = False
+    Returns:
+        List of bins, where each bin contains indices of sequences assigned to it
+    """
+    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(sequence_lengths):
+        global_idx = seq_id + group_offset
+
+        # Try to place sequence in existing bins
+        add_new_bin = True
+        for bin_idx, _ in enumerate(bins_remaining_space):
+            if (
+                bins_remaining_space[bin_idx] >= size
+                and len(bins_assigned_sequences[bin_idx]) < bin_size
+            ):
+                bins_remaining_space[bin_idx] -= size
+                bins_assigned_sequences[bin_idx].append(global_idx)
+                add_new_bin = False
                 break
 
-        if add_new:
-            bins.append(c - size)
-            bins_result.append([indices[a_id] + start_index])
+        # Create a new bin if needed and if we haven't reached the limit
+        if add_new_bin:
+            if len(bins_remaining_space) >= max_bins and safe_mode:
+                # In safe mode, skip items that would exceed max_bins
+                continue
+            bins_remaining_space.append(bin_capacity - size)
+            bins_assigned_sequences.append([global_idx])
 
-    return bins_result
+            # Safety check to avoid infinite bins
+            if len(bins_remaining_space) > len(sequence_lengths):
+                break
+
+    return bins_assigned_sequences
 
 
-@numba.njit
-def allocate(
-    lengths: np.ndarray, lengths_cumsum: np.ndarray, rank: int, c: int, n: int
+# Define a standalone function for multiprocessing
+def _process_group(args):
+    group_lengths, start_idx, bin_capacity, max_bins, bin_size, safe_mode = args
+    return pack_group(
+        group_lengths, start_idx, bin_capacity, max_bins, bin_size, safe_mode
+    )
+
+
+def pack_parallel(
+    sequence_lengths: np.ndarray,
+    bin_capacity: int,
+    group_size: int,
+    bin_size: int,
+    num_processes: int | None = None,
+    safe_mode: bool = True,
+    mp_start_method: str | None = "spawn",
 ):
-    # Dynamic batch allocator, similar to Multifit
-    # https://en.wikipedia.org/wiki/Multifit_algorithm
-    # ~99.5% efficiency on OpenChat training set (12 * 2048 ctx len)
+    """
+    Pack sequences into bins using parallel processing
 
-    s = 0
-    start_index = 0
-    result = []
+    Args:
+        sequence_lengths: Array of sequence lengths
+        bin_capacity: Maximum capacity of each bin as total number of tokens
+        group_size: Number of sequences to process in each group
+        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
+    """
+    num_items = len(sequence_lengths)
+    if num_processes is None:
+        num_processes = max(1, min(num_items // group_size, cpu_count()))
 
-    while True:
-        # binary search [l, r)
-        left = 1
-        right = 1 + np.searchsorted(lengths_cumsum[start_index:], s + c * n, "right")
+    # Create tasks for parallel processing
+    tasks = []
+    for i in range(0, num_items, group_size):
+        group_lengths = sequence_lengths[i : i + group_size]
+        max_bins = len(group_lengths)  # Allow as many bins as items in the group
+        tasks.append((group_lengths, i, bin_capacity, max_bins, bin_size, safe_mode))
 
-        while right - left > 1:
-            mid = (left + right) // 2
-            if ffd_check(lengths[start_index : start_index + mid], c, n):
-                left = mid
-            else:
-                right = mid
+    # Process groups in parallel
+    all_bins = []
 
-        # use length l
-        batch = ffd_with_result(
-            lengths[start_index : start_index + left], c, start_index
-        )
-        assert len(batch) <= n
-        if len(batch) < n:
-            break
+    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
+            )
 
-        start_index += left
-        s = lengths_cumsum[start_index - 1]
+    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)
 
-        # add local rank
-        result.append(batch[rank])
-
-    return result, s, len(result) * c * n
+    return all_bins
 
 
 @numba.njit
-def allocate_sequentially(lengths: np.ndarray, rank: int, c: int, n: int):
+def allocate_sequentially(
+    sequence_lengths: np.ndarray, rank: int, bin_capacity: int, num_ranks: int
+):
     """
     Sequential allocator that preserves example order
 
-    Parameters:
-    - lengths: The lengths of all examples
-    - rank: The current rank (for distributed training)
-    - c: The capacity of each bin (maximum sequence length)
-    - n: Number of ranks
+    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)
+        num_ranks: Number of ranks (processes/GPUs)
 
     Returns:
-    - result: List of batches for the current rank
-    - total_used: Number of actual example tokens
-    - total_slots: Maximum theoretical number of example tokens (number of bins * bin capacity)
+        rank_batches: List of batches for the current rank
+        total_tokens_used: Number of actual example tokens
+        total_token_slots: Maximum theoretical number of example tokens (number of bins * bin capacity)
     """
     result = []
     total_used = 0
@@ -127,9 +212,9 @@ def allocate_sequentially(lengths: np.ndarray, rank: int, c: int, n: int):
     # First, do sequential packing into bins
     all_bins = []
     current_bin = [0 for i in range(0)]  # numba hint
-    remaining_capacity = c
+    remaining_capacity = bin_capacity
 
-    for idx, size in enumerate(lengths):
+    for idx, size in enumerate(sequence_lengths):
         if size <= remaining_capacity:
             # Example fits in current bin
             current_bin.append(idx)
@@ -140,7 +225,7 @@ def allocate_sequentially(lengths: np.ndarray, rank: int, c: int, n: int):
             if current_bin:  # Add non-empty bin to all_bins
                 all_bins.append(current_bin)
             current_bin = [idx]
-            remaining_capacity = c - size
+            remaining_capacity = bin_capacity - size
             total_used += size
 
     # Add the last bin if not empty
@@ -148,132 +233,227 @@ def allocate_sequentially(lengths: np.ndarray, rank: int, c: int, n: int):
         all_bins.append(current_bin)
 
     # Assign bins to ranks - each rank gets every n-th bin
-    for bin_idx in range(rank, len(all_bins), n):
+    for bin_idx in range(rank, len(all_bins), num_ranks):
         result.append(all_bins[bin_idx])
 
-    return result, total_used, len(all_bins) * c
+    return result, total_used, len(all_bins) * bin_capacity
 
 
 class MultipackBatchSampler(BatchSampler):
-    """Batch sampler class for multipack"""
+    """
+    Batch sampler class for efficient packing of variable-length sequences
+
+    This sampler packs sequences into fixed-capacity bins (batches) to maximize
+    GPU memory utilization and training throughput by reducing padding.
+
+    It supports both parallel packing (using FFD algorithm) and
+    sequential packing (preserving original sequence order).
+    """
 
     def __init__(
         self,
         sampler: Union[Sampler[int], Iterable[int]],
-        batch_size: int,
-        batch_max_len: int,
-        lengths: np.ndarray,
-        packing_efficiency_estimate: float = 1.0,
-        drop_last: bool = False,
-        num_count_samples: int = 16,
-        sequential: bool = False,
-        **kwargs,
+        batch_size: int,  # Number of bins per batch
+        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 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
+        num_processes: int | None = None,  # Number of processes for parallel packing
+        safe_mode: bool = True,  # Conservative packing to prevent training instability
+        **kwargs,  # pylint: disable=unused-argument
     ):
         super().__init__(sampler, batch_size, drop_last)
         self.batch_size = batch_size
         self.batch_max_len = batch_max_len
-        self.lengths: np.ndarray = lengths
+        self.lengths = np.array(lengths, dtype=np.int32)
         self.packing_efficiency_estimate = packing_efficiency_estimate or 1.0
         self.sequential = sequential
+        self.group_size = group_size
+        self.bin_size = bin_size
+        self.num_processes = num_processes
+        self.safe_mode = safe_mode
 
         assert isinstance(self.lengths, np.ndarray)
 
         self.epoch = 0
 
-        # statistics
-        self.eff_total_used = 0
-        self.eff_total_slots = 0
+        # Efficiency statistics tracking
+        self.total_tokens_used = 0
+        self.total_token_slots = 0
 
-        # The number of times to calculate the batches to determine the minimum packed dataset length for the local rank
+        # The number of times to calculate batches to determine minimum packed dataset length
         self.num_count_samples = num_count_samples
-        # the minimum packed dataset length across all ranks determined by a gather/broadcast
+        # Minimum packed dataset length across all ranks (determined by gather/broadcast)
         self.len_across_ranks = None
 
+        # Cache for batches
+        self._batches = None
+
         if self.sequential and not isinstance(sampler, SequentialSampler):
             LOG.warning(
                 "using sequential sample packing with non-sequential sampler, did you want to also enable curriculum_sampling?"
             )
 
     def set_epoch(self, epoch: int):
+        """Set the epoch number, used for reproducible shuffling across epochs"""
         self.epoch = epoch
+        self._batches = None  # Invalidate batch cache
 
     def generate_batches(self, set_stats=False):
-        indices = [idx for idx in self.sampler]
+        """
+        Generate packed batches for training
 
-        lengths = self.lengths[indices]
-        lengths_cumsum = np.cumsum(lengths)
+        Args:
+            set_stats: Whether to update efficiency statistics
 
-        if self.sequential:
-            batches, total_used, total_slots = allocate_sequentially(
-                lengths=lengths,
-                rank=0,
-                c=self.batch_max_len,
-                n=1,
-            )
-        else:
-            batches, total_used, total_slots = allocate(
-                lengths=lengths,
-                lengths_cumsum=lengths_cumsum,
-                rank=0,
-                c=self.batch_max_len,
-                n=1,
-            )
+        Returns:
+            List of batches, where each batch contains multiple bins,
+            and each bin contains multiple sequence indices
+        """
+        if self._batches is not None:
+            return self._batches
 
-        batches = [
-            [
-                [indices[b_idx] for b_idx in batch]
-                for batch in batches[i : i + self.batch_size]
-            ]
-            for i in range(0, len(batches), self.batch_size)
+        # Get indices from the sampler
+        indices = [  # pylint: disable=unnecessary-comprehension
+            idx for idx in self.sampler
         ]
 
-        # statistics
-        if set_stats:
-            self.eff_total_used += total_used
-            self.eff_total_slots += total_slots
+        # Get lengths of the selected sequences
+        lengths = self.lengths[indices]
 
+        # Pack sequences into bins using either sequential or parallel packing
+        if self.sequential:
+            bins, total_used, total_slots = allocate_sequentially(
+                lengths,
+                rank=0,
+                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(
+                lengths,
+                bin_capacity=self.batch_max_len,
+                group_size=self.group_size,
+                bin_size=self.bin_size,
+                num_processes=self.num_processes,
+                safe_mode=self.safe_mode,
+            )
+
+            # Map bin indices back to original indices
+            bins = [
+                [indices[b_idx] for b_idx in bin_indices] for bin_indices in all_bins
+            ]
+
+            # Calculate efficiency statistics
+            total_used = lengths.sum()
+            total_slots = len(all_bins) * self.batch_max_len
+
+        # Group bins into batches (each batch contains batch_size bins)
+        batches = [
+            bins[i : i + self.batch_size] for i in range(0, len(bins), self.batch_size)
+        ]
+
+        # Drop last batch if requested and it's incomplete
+        if self.drop_last and len(batches[-1]) < self.batch_size:
+            batches = batches[:-1]
+            # Adjust total_slots if we dropped a batch
+            if not self.sequential:
+                total_slots -= (self.batch_size - len(batches[-1])) * self.batch_max_len
+
+        # Update statistics if requested
+        if set_stats:
+            self.total_tokens_used += total_used
+            self.total_token_slots += total_slots
+
+        self._batches = batches
         return batches
 
     def __iter__(self):
+        """
+        Return an iterator over batches
+
+        The batches are truncated to match the minimum number of batches across all ranks
+        to ensure distributed training balance
+        """
         batches = self.generate_batches(set_stats=True)
         if self.len_across_ranks:
-            # make sure the batches we iterate over is truncated to the same min length across all ranks
+            # Truncate batches to ensure all ranks have the same number of batches
             batches = batches[: self.len_across_ranks]
         return iter(batches)
 
-    def num_batches(self):
-        batches = self.generate_batches(set_stats=True)
-        return len(batches)
-
     def efficiency(self):
-        return self.eff_total_used / self.eff_total_slots
+        """
+        Calculate the packing efficiency (ratio of tokens used to total token slots)
+        Higher is better - 1.0 would mean perfect packing with no wasted space
+        """
+        if self.total_token_slots == 0:
+            self.generate_batches(set_stats=True)
+        if self.total_token_slots == 0:
+            return 0.0
+        # Return a Python float instead of potentially a numpy float
+        return float(self.total_tokens_used / self.total_token_slots)
 
     def gather_efficiency(self):
-        def calc_sample_packing_eff_est(estimates: List[float]):
-            LOG.debug(f"sample_packing_eff_est across ranks: {repr(estimates)}")
-            return math.floor(0.997 * max(estimates))
+        """
+        Gather and synchronize packing efficiency estimates across all distributed ranks
+        Returns a conservative efficiency estimate based on the measurements
+        """
 
+        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)
+            return math.floor(0.997 * max_eff)
+
+        # Gather efficiency from all ranks and apply the calculation function
         sample_packing_actual_eff_all = reduce_and_broadcast(
-            lambda: self.efficiency(),  # pylint: disable=unnecessary-lambda
+            lambda: float(self.efficiency()),  # pylint: disable=unnecessary-lambda
             calc_sample_packing_eff_est,
         )
+
+        # Quantize to 0.5% intervals for stability
         sample_packing_eff_est = (
             math.ceil(sample_packing_actual_eff_all * 200.0) / 200.0
         )
         return sample_packing_eff_est
 
     def gather_len_batches(self, num):
+        """
+        Gather and synchronize batch counts across all distributed ranks
+        Returns the minimum number of batches available on any rank
+        """
+
         def calc_min_len(estimates: list[(int, float)]):
             LOG.info(f"gather_len_batches: {repr(estimates)}")
             return math.floor(min(estimates))
 
+        # Find minimum batch count across ranks to ensure balance
         min_len_batches = reduce_and_broadcast(lambda: num, calc_min_len)
         return min_len_batches
 
     def __len__(self):
-        if not self.len_across_ranks:
-            len_batches = min(
-                [self.num_batches() for _ in range(self.num_count_samples)]
+        """
+        Return the total number of batches that will be yielded by this sampler
+
+        This is calculated as the minimum number of batches available on any rank
+        to ensure balanced distributed training
+        """
+        if self._batches is None:
+            self._batches = self.generate_batches(set_stats=True)
+
+        if self.len_across_ranks is None:
+            # Sample multiple times to get stable estimate
+            len_batches = min(  # pylint: disable=consider-using-generator
+                [len(self._batches) for _ in range(self.num_count_samples)]
             )
+            # Gather minimum across all ranks
             self.len_across_ranks = self.gather_len_batches(len_batches)
+
         return self.len_across_ranks

src/axolotl/utils/schemas/config.py

@@ -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
@@ -435,16 +436,6 @@ class AxolotlInputConfig(
             )
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_sample_packing_w_xformers(cls, data):
-        if data.get("sample_packing") and data.get("xformers_attention"):
-            raise ValueError(
-                "sample_packing not compatible with xformers_attention. Use flash_attention"
-            )
-
-        return data
-
     @model_validator(mode="before")
     @classmethod
     # pylint: disable=duplicate-code
@@ -471,9 +462,10 @@ 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
@@ -1157,16 +1149,28 @@ 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")
@@ -1353,6 +1357,10 @@ 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

@@ -53,4 +53,5 @@ 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

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

tests/e2e/integrations/test_hooks.py

@@ -29,6 +29,12 @@ class LogHooksPlugin(BasePlugin):
         except FileNotFoundError:
             pass
 
+    def post_trainer_create(self, cfg, trainer):  # pylint: disable=unused-argument
+        with open(
+            self.base_dir.joinpath("plugin_hooks.log"), "a", encoding="utf-8"
+        ) as f:
+            f.write("post_trainer_create\n")
+
     def pre_model_load(self, cfg):  # pylint: disable=unused-argument
         with open(
             self.base_dir.joinpath("plugin_hooks.log"), "a", encoding="utf-8"
@@ -165,6 +171,7 @@ class TestPluginHooks:
         ) as f:
             file_contents = f.readlines()
             file_contents = "\n".join(file_contents)
+            assert "post_trainer_create" in file_contents
             assert "pre_model_load" in file_contents
             assert "post_model_build" in file_contents
             assert "pre_lora_load" in file_contents

tests/e2e/integrations/test_kd.py

@@ -90,7 +90,7 @@ class TestKnowledgeDistillation:
         train(cfg=cfg, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "model.safetensors").exists()
         check_tensorboard(
-            temp_dir + "/runs", "train/loss", 1.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/loss", 1.2, "Train Loss (%s) is too high"
         )
 
     @pytest.mark.parametrize(
@@ -121,5 +121,5 @@ class TestKnowledgeDistillation:
         train(cfg=cfg, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "adapter_model.safetensors").exists()
         check_tensorboard(
-            temp_dir + "/runs", "train/loss", 1.0, "Train Loss is too high"
+            temp_dir + "/runs", "train/loss", 1.2, "Train Loss (%s) is too high"
         )

tests/e2e/integrations/test_llm_compressor.py

@@ -1,111 +0,0 @@
-"""
-E2E smoke tests for LLMCompressorPlugin integration
-"""
-
-from pathlib import Path
-
-import pytest
-
-from axolotl.cli.args import TrainerCliArgs
-from axolotl.common.datasets import load_datasets
-from axolotl.train import train
-from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
-from axolotl.utils.dict import DictDefault
-
-from tests.e2e.utils import (
-    check_model_output_exists,
-    require_llmcompressor,
-    require_torch_2_4_1,
-)
-
-MODELS = [
-    "nm-testing/llama2.c-stories42M-pruned2.4-compressed",
-    "nm-testing/llama2.c-stories42M-gsm8k-sparse-only-compressed",
-]
-
-
-@pytest.mark.parametrize(
-    "base_model", MODELS, ids=["no-checkpoint-recipe", "with-checkpoint-recipe"]
-)
-@pytest.mark.parametrize(
-    "save_compressed", [True, False], ids=["save_compressed", "save_uncompressed"]
-)
-class TestLLMCompressorIntegration:
-    """
-    e2e tests for axolotl.integrations.llm_compressor.LLMCompressorPlugin
-    """
-
-    @require_llmcompressor
-    @require_torch_2_4_1
-    def test_llmcompressor_plugin(
-        self, temp_dir, base_model: str, save_compressed: bool
-    ):
-        from llmcompressor import active_session
-
-        # core cfg
-        cfg = DictDefault(
-            {
-                "base_model": base_model,
-                "plugins": ["axolotl.integrations.llm_compressor.LLMCompressorPlugin"],
-                "sequence_len": 1024,
-                "val_set_size": 0.05,
-                "special_tokens": {"pad_token": "<|endoftext|>"},
-                "datasets": [{"path": "mhenrichsen/alpaca_2k_test", "type": "alpaca"}],
-                "num_epochs": 1,
-                "micro_batch_size": 2,
-                "gradient_accumulation_steps": 2,
-                "output_dir": temp_dir,
-                "learning_rate": 1e-5,
-                "optimizer": "adamw_torch_fused",
-                "lr_scheduler": "cosine",
-                "save_safetensors": True,
-                "bf16": "auto",
-                "max_steps": 5,
-                "llmcompressor": {
-                    "recipe": {
-                        "finetuning_stage": {
-                            "finetuning_modifiers": {
-                                "ConstantPruningModifier": {
-                                    "targets": [
-                                        "re:.*q_proj.weight",
-                                        "re:.*k_proj.weight",
-                                        "re:.*v_proj.weight",
-                                        "re:.*o_proj.weight",
-                                        "re:.*gate_proj.weight",
-                                        "re:.*up_proj.weight",
-                                        "re:.*down_proj.weight",
-                                    ],
-                                    "start": 0,
-                                },
-                            },
-                        },
-                    },
-                    "save_compressed": save_compressed,
-                },
-            }
-        )
-
-        prepare_plugins(cfg)
-        cfg = validate_config(cfg)
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        try:
-            train(cfg=cfg, dataset_meta=dataset_meta)
-            check_model_output_exists(temp_dir, cfg)
-            _check_llmcompressor_model_outputs(temp_dir, save_compressed)
-        finally:
-            active_session().reset()
-
-
-def _check_llmcompressor_model_outputs(temp_dir, save_compressed):
-    if save_compressed:
-        assert (Path(temp_dir) / "recipe.yaml").exists()
-
-        from compressed_tensors import ModelCompressor
-        from compressed_tensors.config import Sparse24BitMaskConfig
-
-        compressor = ModelCompressor.from_pretrained(temp_dir)
-        assert compressor is not None
-        assert isinstance(compressor.sparsity_config, Sparse24BitMaskConfig)

tests/e2e/multigpu/solo/test_grpo.py

@@ -166,6 +166,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
 """
             )
 
+    @pytest.mark.skip(reason="flaky test")
     @pytest.mark.parametrize(
         "num_gpus",
         [1, 2],
@@ -227,7 +228,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
 
         current_env = os.environ.copy()
         env = {
-            "NCCL_P2P_LEVEL": "NVL",
+            "NCCL_P2P_LEVEL": "LOC",
             **current_env,
             "CUDA_VISIBLE_DEVICES": "1",
             "VLLM_DISABLE_COMPILE_CACHE": "1",
@@ -257,7 +258,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
                     f"{get_torch_dist_unique_port()}",
                 ],
                 env={
-                    "NCCL_P2P_LEVEL": "NVL",
+                    "NCCL_P2P_LEVEL": "LOC",
                     "NCCL_DEBUG": "INFO",
                     **current_env,
                 },
@@ -265,6 +266,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
         finally:
             recursive_kill(vllm_process)
 
+    @pytest.mark.skip(reason="flaky test")
     @pytest.mark.parametrize(
         "num_gpus",
         [1, 2],
@@ -320,7 +322,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
 
         current_env = os.environ.copy()
         env = {
-            "NCCL_P2P_LEVEL": "NVL",  # nccl can be brittle, assume P2P isn't reliable
+            "NCCL_P2P_LEVEL": "LOC",  # nccl can be brittle, assume P2P isn't reliable
             **current_env,
             "CUDA_VISIBLE_DEVICES": "1",
             "VLLM_DISABLE_COMPILE_CACHE": "1",
@@ -350,7 +352,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
                     f"{get_torch_dist_unique_port()}",
                 ],
                 env={
-                    "NCCL_P2P_LEVEL": "NVL",
+                    "NCCL_P2P_LEVEL": "LOC",
                     "NCCL_DEBUG": "INFO",
                     **current_env,
                 },

tests/e2e/multigpu/test_llama.py

@@ -479,7 +479,7 @@ class TestMultiGPULlama:
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
                 "sequence_len": 2048,
-                "val_set_size": 0.05,
+                "val_set_size": 0.1,
                 "special_tokens": {
                     "pad_token": "<|endoftext|>",
                 },

tests/e2e/patched/lora_kernels/test_lora_kernel_patching.py

@@ -29,12 +29,12 @@ from axolotl.utils.dict import DictDefault
 
 MODEL_CONFIGS = [
     {
-        "name": "openaccess-ai-collective/tiny-mistral",
+        "name": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
     {
-        "name": "Qwen/Qwen2-7B",
+        "name": "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
@@ -44,7 +44,7 @@ MODEL_CONFIGS = [
         "dtype": torch.float32,
     },
     {
-        "name": "mhenrichsen/gemma-2b",
+        "name": "trl-internal-testing/tiny-Gemma2ForCausalLM",
         "expected_activation": apply_lora_mlp_geglu,
         "dtype": torch.float16,
     },
@@ -156,7 +156,9 @@ def test_swiglu_mlp_integration(small_llama_model):
 def test_geglu_model_integration():
     """Test GeGLU activation with Gemma model."""
     model = AutoModelForCausalLM.from_pretrained(
-        "mhenrichsen/gemma-2b", torch_dtype=torch.float16, device_map="cuda:0"
+        "trl-internal-testing/tiny-Gemma2ForCausalLM",
+        torch_dtype=torch.float16,
+        device_map="cuda:0",
     )
     peft_config = get_peft_config(
         {

tests/e2e/patched/test_4d_multipack_llama.py

@@ -57,9 +57,9 @@ class Test4dMultipackLlama(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
-                "save_steps": 10,
-                "eval_steps": 10,
+                "max_steps": 5,
+                "save_steps": 3,
+                "eval_steps": 4,
                 "fp16": True,
             }
         )
@@ -105,9 +105,9 @@ class Test4dMultipackLlama(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
-                "save_steps": 10,
-                "eval_steps": 10,
+                "max_steps": 5,
+                "save_steps": 3,
+                "eval_steps": 4,
                 "fp16": True,
             }
         )

tests/e2e/patched/test_activation_checkpointing.py

@@ -26,10 +26,15 @@ class TestActivationCheckpointing:
     E2E tests for activation checkpointing
     """
 
+    @pytest.mark.parametrize(
+        "gradient_checkpointing",
+        ["offload", "offload_disk"],
+    )
     def test_activation_checkpointing_offload(
         self,
         temp_dir,
         fix_checkpoint_after_test,  # pylint: disable=unused-argument,redefined-outer-name
+        gradient_checkpointing,
     ):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
@@ -64,7 +69,7 @@ class TestActivationCheckpointing:
                 "sample_packing": True,
                 "bf16": True,
                 "save_safetensors": True,
-                "gradient_checkpointing": "offload",
+                "gradient_checkpointing": gradient_checkpointing,
             }
         )
 

tests/e2e/patched/test_falcon_samplepack.py

@@ -6,6 +6,8 @@ import logging
 import os
 import unittest
 
+import pytest
+
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
@@ -23,6 +25,7 @@ class TestFalconPatched(unittest.TestCase):
     Test case for Falcon models
     """
 
+    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_qlora(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -71,6 +74,7 @@ class TestFalconPatched(unittest.TestCase):
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
 
+    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_ft(self, temp_dir):
         # pylint: disable=duplicate-code

tests/e2e/patched/test_mistral_samplepack.py

@@ -28,7 +28,7 @@ class TestMistral(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "openaccess-ai-collective/tiny-mistral",
+                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 1024,
@@ -57,9 +57,9 @@ class TestMistral(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
-                "save_steps": 10,
-                "eval_steps": 10,
+                "max_steps": 5,
+                "save_steps": 3,
+                "eval_steps": 4,
                 "bf16": "auto",
             }
         )
@@ -76,7 +76,7 @@ class TestMistral(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "openaccess-ai-collective/tiny-mistral",
+                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 1024,
@@ -99,9 +99,9 @@ class TestMistral(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
-                "save_steps": 10,
-                "eval_steps": 10,
+                "max_steps": 5,
+                "save_steps": 3,
+                "eval_steps": 4,
                 "bf16": "auto",
             }
         )

tests/e2e/patched/test_mixtral_samplepack.py

@@ -54,9 +54,9 @@ class TestMixtral(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
-                "save_steps": 10,
-                "eval_steps": 10,
+                "max_steps": 5,
+                "save_steps": 3,
+                "eval_steps": 4,
                 "bf16": "auto",
             }
         )
@@ -93,9 +93,9 @@ class TestMixtral(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
-                "save_steps": 10,
-                "eval_steps": 10,
+                "max_steps": 5,
+                "save_steps": 3,
+                "eval_steps": 4,
                 "bf16": "auto",
             }
         )

tests/e2e/patched/test_model_patches.py

@@ -56,7 +56,7 @@ class TestModelPatches(unittest.TestCase):
     def test_mistral_multipack(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "openaccess-ai-collective/tiny-mistral",
+                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,

tests/e2e/patched/test_peft_embeddings.py

@@ -0,0 +1,63 @@
+"""
+Test case for handling embeddings when using peft
+"""
+
+import torch
+
+from axolotl.train import setup_model_and_tokenizer
+from axolotl.utils.config import normalize_config, validate_config
+from axolotl.utils.dict import DictDefault
+
+
+class TestLlamaPeftEmbeddings:
+    """
+    test class for handling embeddings when using peft
+    """
+
+    def test_peft_embeddings_upcast(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_target_linear": True,
+                "trust_remote_code": True,
+                "sequence_len": 512,
+                "val_set_size": 0.01,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 2,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "sample_packing": False,
+                "bf16": "auto",
+                "save_safetensors": True,
+                "embeddings_skip_upcast": True,
+            }
+        )
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+
+        model, _, _, _ = setup_model_and_tokenizer(cfg)
+
+        # Check if the embeddings are upcast correctly
+        # only embed_tokens is a parameter that may be upcast
+        assert model.base_model.model.model.embed_tokens.weight.dtype == torch.bfloat16
+        assert model.base_model.model.lm_head.weight.dtype == torch.bfloat16

tests/e2e/patched/test_phi_multipack.py

@@ -56,9 +56,9 @@ class TestPhiMultipack(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
-                "eval_steps": 10,
-                "save_steps": 10,
+                "max_steps": 5,
+                "eval_steps": 3,
+                "save_steps": 4,
                 "bf16": "auto",
             }
         )
@@ -108,9 +108,9 @@ class TestPhiMultipack(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
-                "eval_steps": 10,
-                "save_steps": 10,
+                "max_steps": 5,
+                "eval_steps": 3,
+                "save_steps": 4,
                 "bf16": "auto",
             }
         )

tests/e2e/patched/test_resume.py

@@ -15,7 +15,7 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_model_output_exists, most_recent_subdir
+from ..utils import check_model_output_exists, most_recent_subdir, require_torch_2_6_0
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -26,6 +26,7 @@ class TestResumeLlama:
     Test case for resuming training of llama models
     """
 
+    @require_torch_2_6_0
     def test_resume_lora_packed(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
@@ -62,6 +63,7 @@ class TestResumeLlama:
                 "save_total_limit": 5,
                 "max_steps": 15,
                 "use_tensorboard": True,
+                "save_safetensors": True,
             }
         )
         if is_torch_bf16_gpu_available():

tests/e2e/test_evaluate.py

@@ -19,14 +19,11 @@ class TestE2eEvaluate:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sequence_len": 1024,
                 "val_set_size": 0.02,
                 "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
+                    "pad_token": "<|endoftext|>",
                 },
                 "datasets": [
                     {

tests/e2e/test_falcon.py

@@ -6,6 +6,8 @@ import logging
 import os
 import unittest
 
+import pytest
+
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
@@ -23,6 +25,7 @@ class TestFalcon(unittest.TestCase):
     Test case for falcon
     """
 
+    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_lora(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -74,6 +77,7 @@ class TestFalcon(unittest.TestCase):
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
 
+    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_lora_added_vocab(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -129,6 +133,7 @@ class TestFalcon(unittest.TestCase):
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
 
+    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_ft(self, temp_dir):
         # pylint: disable=duplicate-code

tests/e2e/test_mistral.py

@@ -30,7 +30,7 @@ class TestMistral(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "openaccess-ai-collective/tiny-mistral",
+                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "load_in_8bit": True,
@@ -77,7 +77,7 @@ class TestMistral(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "openaccess-ai-collective/tiny-mistral",
+                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "val_set_size": 0.02,

tests/e2e/test_optimizers.py

@@ -199,3 +199,50 @@ class TestCustomOptimizers(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+
+    @with_temp_dir
+    def test_came_pytorch(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.1,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "came_pytorch",
+                "adam_beta3": 0.9999,
+                "adam_epsilon2": 1e-16,
+                "max_steps": 5,
+                "lr_scheduler": "cosine",
+            }
+        )
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)

tests/e2e/utils.py

@@ -105,25 +105,7 @@ def require_vllm(test_case):
             return False
 
     return unittest.skipUnless(
-        is_vllm_installed(), "test requires vllm to be installed"
-    )(test_case)
-
-
-def require_llmcompressor(test_case):
-    """
-    Decorator marking a test that requires a llmcompressor to be installed
-    """
-
-    def is_llmcompressor_installed():
-        try:
-            import llmcompressor  # pylint: disable=unused-import  # noqa: F401
-
-            return True
-        except ImportError:
-            return False
-
-    return unittest.skipUnless(
-        is_llmcompressor_installed(), "test requires llmcompressor to be installed"
+        is_vllm_installed(), "test requires a vllm to be installed"
     )(test_case)
 
 

tests/test_datasets.py

@@ -414,7 +414,6 @@ class TestDatasetPreparation:
             snapshot_path = snapshot_download(
                 repo_id="mhenrichsen/alpaca_2k_test",
                 repo_type="dataset",
-                local_dir=tmp_ds_path,
             )
             shutil.copytree(snapshot_path, tmp_ds_path, dirs_exist_ok=True)
 

tests/test_packed_batch_sampler.py

@@ -106,3 +106,4 @@ class TestBatchedSamplerPacking:
 
         original_idxs = set(range(len(train_dataset)))
         assert original_idxs == set(batch_idxs)
+        assert len(batch_idxs) == len(set(batch_idxs))