use uint8 dtype for qlora

fix error with lambda and pickling

make sure things are in float instead of np.float

.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

@@ -18,96 +18,9 @@ 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

@@ -44,104 +44,12 @@ 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"]
@@ -151,20 +59,14 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-#      - name: Restore HF cache
+      - name: Restore HF cache
-#        id: hf-cache-restore
+        id: hf-cache-restore
-#        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v4
-#        with:
+        with:
-#          path: |
+          path: |
-#            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/datasets--*
-#            /home/runner/.cache/huggingface/hub/models--*
+            /home/runner/.cache/huggingface/hub/models--*
-#          key: ${{ runner.os }}-hf-hub-cache-v2
+          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
@@ -219,12 +121,21 @@ 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"]
@@ -234,20 +145,14 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-#      - name: Restore HF cache
+      - name: Restore HF cache
-#        id: hf-cache-restore
+        id: hf-cache-restore
-#        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v4
-#        with:
+        with:
-#          path: |
+          path: |
-#            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/datasets--*
-#            /home/runner/.cache/huggingface/hub/models--*
+            /home/runner/.cache/huggingface/hub/models--*
-#          key: ${{ runner.os }}-hf-hub-cache-v2
+          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
@@ -294,6 +199,15 @@ 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: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
@@ -353,6 +267,12 @@ jobs:
             pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras: llmcompressor
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.4.1
+            num_gpus: 1
+            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -389,43 +309,3 @@ jobs:
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
-
-  docker-e2e-cleanup:
-    runs-on: [self-hosted, modal]
-    timeout-minutes: 90
-    needs: [docker-e2e-tests]
-
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            num_gpus: 1
-            axolotl_extras: vllm
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-      - name: Install Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.11"
-      - name: Install Modal
-        run: |
-          python -m pip install --upgrade pip
-          pip install modal==0.71.8 jinja2
-      - name: Update env vars
-        run: |
-          echo "BASE_TAG=main-base-py${{ 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

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

_quarto.yml

@@ -48,23 +48,8 @@ quartodoc:
       contents:
         - core.trainers.base
         - core.trainers.trl
-        - core.trainers.mamba
-        - core.trainers.relora
         - core.trainers.dpo.trainer
         - core.trainers.grpo.trainer
-        - core.trainers.grpo.sampler
-        - core.trainers.utils
-    - title: Mixins
-      desc: Mixin classes for augmenting trainers
-      contents:
-        - core.trainers.mixins.optimizer
-        - core.trainers.mixins.rng_state_loader
-        - core.trainers.mixins.scheduler
-        - core.trainers.mixins.sequence_parallel
-    - title: Context Managers
-      desc: Context managers for altering trainer behaviors
-      contents:
-        - utils.ctx_managers.sequence_parallel
     - title: Prompt Strategies
       desc: Prompt formatting strategies
       contents:
@@ -101,7 +86,7 @@ quartodoc:
         - kernels.swiglu
         - kernels.quantize
         - kernels.utils
-    - title: Monkey Patches
+    - title: MonkeyPatches
       desc: Runtime patches for model optimizations
       contents:
         - monkeypatch.llama_attn_hijack_flash

cicd/cicd.sh

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

cicd/cleanup.py

@@ -1,19 +0,0 @@
-"""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

@@ -1,6 +0,0 @@
-#!/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

@@ -1,12 +1,75 @@
 """Modal app to run axolotl GPU tests"""
 
-from .single_gpu import GPU_CONFIG, VOLUME_CONFIG, app, cicd_image, run_cmd
+# 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
 
 
 @app.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=90 * 60,  # 90 min
+    timeout=60 * 60,
     cpu=8.0,
     memory=131072,
     volumes=VOLUME_CONFIG,

cicd/single_gpu.py

@@ -1,66 +0,0 @@
-"""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: true
+        only_pulls: false
         flags: null
         paths: null
     patch:

docs/config.qmd

@@ -32,8 +32,6 @@ 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:
@@ -75,12 +73,11 @@ load_in_8bit: true
 load_in_4bit:
 
 # Use CUDA bf16
-bf16: true # bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection. require >=ampere
+bf16: true # bool or 'full' for `bf16_full_eval`. 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
@@ -187,8 +184,8 @@ datasets:
     # adding a system turn with empty content.
     drop_system_message:
 
-    # Optional[bool]. (for Qwen3 template only) Whether to split the assistant content based on a reasoning trace inside delimited tags
+    # Optional[bool]. Whether to split the assistant turn based on a reasoning trace inside delimited tags
-    # See example at `docs/dataset-formats/conversation.qmd`
+    # defaults to False
     split_thinking:
 
     # IMPORTANT: The following fields determine which parts of the conversation to train on.
@@ -505,7 +502,6 @@ 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
@@ -551,7 +547,7 @@ gradient_checkpointing: false
 early_stopping_patience: 3
 
 # Specify a scheduler and kwargs to use with the optimizer
-lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | '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: # 'one_cycle' | 'rex' | 'log_sweep' | 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)
@@ -613,7 +609,6 @@ 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/dataset-formats/conversation.qmd

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

docs/sequence_parallelism.qmd

@@ -3,6 +3,8 @@ title: Sequence Parallelism
 description: Train with long sequences split across multiple GPUs.
 ---
 
+# Sequence Parallelism
+
 Sequence parallelism is a technique that splits sequences across multiple GPUs,
 allowing you to train with very long sequences that wouldn't fit on a single GPU. Each
 GPU processes a different portion of the sequence, and the results are aggregated
@@ -25,7 +27,7 @@ To enable sequence parallelism, add the following to your configuration file:
 sequence_parallel_degree: 4  # Split sequences across 4 GPUs
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
-# Optional; one of "varlen_llama3" or "batch_ring". Defaults to
+# Optional; one of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to
 # "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
 ring_attn_func:
 ```

examples/llama-4/README.md

@@ -34,5 +34,3 @@ 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

@@ -1,341 +0,0 @@
-# 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

@@ -1,52 +0,0 @@
-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

@@ -6,17 +6,16 @@ triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.9
+liger-kernel==0.5.8
 # 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==3.5.1
+datasets==3.5.0
 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.post1"]
+                extras_require_map["vllm"] = ["vllm==0.8.5"]
             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.post1"]
+                extras_require_map["vllm"] = ["vllm==0.8.5"]
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
@@ -142,7 +142,6 @@ extras_require = {
         "apollo-torch",
         "lomo-optim==0.1.1",
         "torch-optimi==0.2.1",
-        "came_pytorch==0.1.3",
     ],
     "ray": [
         "ray[train]",

src/axolotl/cli/args.py

@@ -82,12 +82,6 @@ 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/main.py

@@ -342,13 +342,6 @@ def delinearize_llama4(model: str, output: str) -> None:
     do_delinearize_llama4(model, output)
 
 
-@cli.command()
-def wizard():
-    from axolotl.cli.wizard import do_wizard
-
-    do_wizard()
-
-
 cli.add_command(lm_eval)
 
 

src/axolotl/cli/preprocess.py

@@ -18,7 +18,6 @@ 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
 
@@ -48,10 +47,7 @@ def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:
         cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
     with disable_datasets_caching():
-        plugin_manager = PluginManager.get_instance()
+        if cfg.rl:
-        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

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

@@ -6,6 +6,7 @@ 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
 
@@ -27,9 +28,6 @@ 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/cli/wizard.py

@@ -1,429 +0,0 @@
-"""Wizard for creating yaml configs."""
-
-import click
-import torch
-import yaml
-from packaging import version
-from transformers.training_args import OptimizerNames
-
-from axolotl.cli.art import print_axolotl_text_art
-from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import load_model_config
-from axolotl.utils.schemas.enums import CustomSupportedOptimizers
-
-
-def do_wizard():
-    print_axolotl_text_art()
-
-    # Ask where to save the config
-    cfg = DictDefault({})
-    config_path = click.prompt(
-        "Where do you want to save the config?", type=str, default="config.yaml"
-    )
-
-    # Ask base model
-    base_model = click.prompt("What base model do you want to use?", type=str)
-    cfg["base_model"] = base_model.strip()
-
-    # Ask whether want to enable Vision model
-    # TODO: check if model has vision layers instead of asking user
-    train_vision_model = click.confirm(
-        "If this model has vision layers, do you want to train them?", default=False
-    )
-
-    if train_vision_model:
-        cfg["processor_type"] = "AutoProcessor"
-        cfg["skip_prepare_dataset"] = True
-        cfg["remove_unused_columns"] = False
-        cfg["sample_packing"] = False
-
-    # Ask whether they want to set any advanced model features (custom tokenizer, custom config, etc)
-    advanced_model_features = click.confirm(
-        "Do you want to set any advanced model features? (custom tokenizer, custom config, remote code etc)",
-        default=False,
-    )
-
-    if advanced_model_features:
-        # Ask whether they want to use a custom config
-        base_model_config = click.prompt(
-            "What model config do you want to use? (leave blank for default)",
-            type=str,
-            default="",
-        )
-
-        if base_model_config:
-            cfg["base_model_config"] = base_model_config
-
-        # Ask whether they want to use a specific revision of the model
-        revision_of_model = click.prompt(
-            "What revision of the model do you want to use? (leave blank for default)",
-            type=str,
-            default="",
-        )
-
-        if revision_of_model:
-            cfg["revision_of_model"] = revision_of_model
-
-        # Ask whether they want to use a custom tokenizer
-        tokenizer_config = click.prompt(
-            "What tokenizer do you want to use? (leave blank for default)",
-            type=str,
-            default="",
-        )
-
-        if tokenizer_config:
-            cfg["tokenizer_config"] = tokenizer_config
-
-        # Ask whether they want to use remote code
-        trust_remote_code = click.confirm(
-            "Do you want to use remote code?", default=False
-        )
-
-        if trust_remote_code:
-            cfg["trust_remote_code"] = trust_remote_code
-
-        # Whether to resize token embeddings
-        resize_token_embeddings_to_32x = click.confirm(
-            "Do you want to resize token embeddings to 32x?", default=False
-        )
-
-        if resize_token_embeddings_to_32x:
-            cfg["resize_token_embeddings_to_32x"] = resize_token_embeddings_to_32x
-
-        # Whether to shrink embeddings to len(tokenizer)
-        shrink_embeddings = click.confirm(
-            "Do you want to shrink embeddings to len(tokenizer)?", default=False
-        )
-
-        if shrink_embeddings:
-            cfg["shrink_embeddings"] = shrink_embeddings
-
-        # Whether to skip upcast embeddings
-        embeddings_skip_upcast = click.confirm(
-            "Do you want to skip upcast embeddings?", default=False
-        )
-
-        if embeddings_skip_upcast:
-            cfg["embeddings_skip_upcast"] = embeddings_skip_upcast
-
-        # Whether to random init weights
-        random_init_weights = click.confirm(
-            "Do you want to random init weights?", default=False
-        )
-
-        if random_init_weights:
-            cfg["random_init_weights"] = random_init_weights
-
-    # Get model type
-    config = load_model_config(cfg)
-    model_type = config.model_type
-
-    # Ask sequence length
-    sequence_length = click.prompt("What sequence length do you want to use?", type=int)
-    cfg["sequence_length"] = sequence_length
-
-    # Whether to turn on sample packing
-    if cfg["sample_packing"] is None:
-        cfg["sample_packing"] = click.confirm(
-            "Do you want to turn on sample packing? This will speed up training by packing multiple samples into a single batch.",
-            default=True,
-        )
-
-        if cfg["sample_packing"]:
-            cfg["pad_to_sequence_len"] = True
-
-        # Whether to turn off eval sample packing
-        no_eval_sample_packing = click.confirm(
-            "Do you want to turn off eval sample packing? This will slow down evaluation but is recommended if you are using a small validation set.",
-            default=False,
-        )
-
-        if no_eval_sample_packing:
-            cfg["eval_sample_packing"] = False
-
-    # Hardware check
-    try:
-        is_ampere_or_newer = torch.cuda.get_device_capability()[0] >= 8
-    except RuntimeError:
-        is_ampere_or_newer = False
-    except AssertionError:  # this is raised if no cuda is available
-        is_ampere_or_newer = False
-
-    # Get num gpus
-    try:
-        num_gpus = torch.cuda.device_count()
-    except RuntimeError:
-        num_gpus = 0
-
-    # Get torch version
-    torch_version = str(torch.__version__).split("+", maxsplit=1)[0]
-
-    is_torch_2_6_or_newer = version.parse(torch_version) >= version.parse("2.6.0")
-
-    # Whether to turn on attention
-    opt = ["xformers", "sdp"]
-
-    if is_ampere_or_newer:
-        opt.append("flash")
-
-    if is_torch_2_6_or_newer:
-        opt.append("flex")
-
-    if cfg["sample_packing"]:
-        if "flash" in opt:
-            default_opt = "flash"
-        elif "flex" in opt:
-            default_opt = "flex"
-        else:
-            default_opt = opt[0]
-
-        attention = click.prompt(
-            "Which attention backend do you want to use? Sample packing requires an attention backend to be set.",
-            type=click.Choice(opt),
-            default=default_opt,
-        )
-    else:
-        # non-sample packing supports no attention and S2
-        opt.extend(["none", "s2"])
-
-        attention = click.prompt(
-            "Which attention backend do you want to use?",
-            type=click.Choice(opt),
-            default="none",
-        )
-
-        if attention == "none":
-            attention = None
-
-    # TODO: if xformers, check if FA is installed
-    # TODO: flex doc mentioned requiring seq len to be divisible by 128. Unclear if limitation still exists
-
-    # TODO: requires #2489
-    cfg["attention"] = attention
-
-    # Whether to turn on gradient checkpointing
-    # TODO: need to wait for offload_disk PR to be merged
-    gradient_checkpointing = click.prompt(
-        "Which gradient checkpointing strategy do you want to use?",
-        type=click.Choice(["none", "true", "offload", "offload_disk"]),
-        default="true",
-    )
-
-    if gradient_checkpointing == "none":
-        gradient_checkpointing = False
-    elif gradient_checkpointing == "true":
-        gradient_checkpointing = True
-
-    # Ask whether to set use_reentrant
-    # TODO: get correct defaults based on SFT/RL mode and single/multigpu
-    # use_reentrant = click.confirm(
-    #     "Do you want to set use_reentrant?",
-    #     default=True,
-    # )
-
-    # if use_reentrant:
-    #     cfg["use_reentrant"] = use_reentrant
-
-    # Optimizer
-    cfg["optimizer"] = click.prompt(
-        "Which optimizer do you want to use?",
-        type=click.Choice((OptimizerNames | CustomSupportedOptimizers)),
-        default=OptimizerNames.ADAMW_TORCH_FUSED,
-    )
-
-    cfg["lr_scheduler"] = click.prompt(
-        "Which learning rate scheduler do you want to use?",
-        type=click.Choice(
-            [
-                "cosine",
-                "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",
-            ]
-        ),
-        default="cosine",
-    )
-
-    # Plugins
-
-    cfg["plugins"] = []
-
-    # Whether to turn on cut cross entropy
-    if is_ampere_or_newer:
-        # Note: This may error if users don't have CCE installed
-        from axolotl.integrations.cut_cross_entropy.monkeypatch.patch import (
-            CUT_CROSS_ENTROPY_MODEL_MAPPING,
-        )
-
-        if model_type in CUT_CROSS_ENTROPY_MODEL_MAPPING:
-            cut_cross_entropy = click.confirm(
-                "Do you want to turn on cut cross entropy? This will save VRAM if the model has a large vocab size.",
-                default=True,
-            )
-
-            if cut_cross_entropy:
-                cfg["plugins"].append(
-                    "axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin"
-                )
-
-                cfg["cut_cross_entropy"] = True
-
-    use_liger_kernel = click.confirm(
-        "Do you want to use the liger kernel? This will speed up training and save VRAM.",
-        default=True,
-    )
-
-    if use_liger_kernel:
-        cfg["plugins"].append("axolotl.integrations.liger.LigerPlugin")
-
-        cfg["liger_rope"] = click.confirm(
-            "Do you want to enable liger rope?",
-            default=True,
-        )
-
-        cfg["liger_rms_norm"] = click.confirm(
-            "Do you want to enable liger rms norm?",
-            default=True,
-        )
-
-        cfg["liger_glu_activation"] = click.confirm(
-            "Do you want to enable liger glu activation?",
-            default=True,
-        )
-
-        cfg["liger_layer_norm"] = click.confirm(
-            "Do you want to enable liger layer norm?",
-            default=True,
-        )
-
-        if cfg["cut_cross_entropy"] is not True:
-            cfg["liger_fused_linear_cross_entropy"] = click.confirm(
-                "Do you want to enable liger fused linear cross entropy?",
-                default=True,
-            )
-
-    # TODO: lora kernels (but they auto enable via validator already)
-
-    # TODO: is there incompat between torch compile and liger?
-    cfg["torch_compile"] = click.confirm(
-        "Do you want to enable torch compile?",
-        default=True,
-    )
-
-    # Multi-gpu
-    if num_gpus > 1:
-        # Ask whether to use DDP/Deepspeed/FSDP
-        multi_gpu_mode = click.prompt(
-            "Which multi-gpu mode do you want to use?",
-            type=click.Choice(["ddp", "deepspeed", "fsdp"]),
-            default="ddp",
-        )
-
-        if multi_gpu_mode == "deepspeed":
-            # Ask which deepspeed config to use
-            cfg["deepspeed"] = click.prompt(
-                "Which deepspeed config do you want to use? The higher the number, the more VRAM you will save, but the slower it will run.",
-                type=click.Choice(
-                    [
-                        "zero1.json",
-                        "zero1_torch_compile.json",
-                        "zero2.json",
-                        "zero3.json",
-                        "zero3_bf16.json",
-                        "zero3_bf16_cpuoffload_all.json",
-                        "zero3_bf16_cpuoffload_params.json",
-                    ]
-                ),
-                default="zero1.json",
-            )
-        elif multi_gpu_mode == "fsdp":
-            fsdp_version = click.prompt(
-                "Which fsdp version do you want to use?",
-                type=click.Choice([1, 2]),
-                default=2,
-            )
-
-            # TODO: Handle FSDP config
-
-            if fsdp_version == 1:
-                cfg["fsdp"] = ["full_shard", "auto_wrap"]
-
-                # Ask which state dict type to use
-                fsdp_state_dict_type = click.prompt(
-                    "Which fsdp state dict type do you want to use?",
-                    type=click.Choice(["FULL_STATE_DICT", "SHARDED_STATE_DICT"]),
-                    default="FULL_STATE_DICT",
-                )
-
-                fsdp_offload_params = click.confirm(
-                    "Do you want to offload parameters?",
-                    default=True,
-                )
-
-                # TODO: can we load the model class and auto pull a default for this?
-                fsdp_transformer_layer_cls_to_wrap = click.prompt(
-                    "Which transformer layer class to wrap? It is usually the Decoder layer class.",
-                    type=str,
-                )
-
-                # TODO: add other options
-
-                cfg["fsdp_config"] = {
-                    "state_dict_type": fsdp_state_dict_type,
-                    "offload_params": fsdp_offload_params,
-                    "transformer_layer_cls_to_wrap": fsdp_transformer_layer_cls_to_wrap,
-                }
-
-            elif fsdp_version == 2:
-                raise NotImplementedError()
-
-    # Training mode (sft or rl)
-    training_mode = click.prompt(
-        "Which training mode do you want to use?",
-        type=click.Choice(["sft", "rl"]),
-        default="sft",
-    )
-
-    if training_mode == "rl":
-        cfg["rl"] = click.prompt(
-            "Which rl mode do you want to use?",
-            type=click.Choice(["dpo", "ipo", "orpo", "kto", "grpo", "simpo"]),
-        )
-
-        # TODO: handle RL options
-
-    # Whether to use adapter
-
-    # Get batch/grad accu
-
-    # Get learning rate
-
-    # Get weight decay
-
-    # Get max grad norm
-
-    # Get num train epochs
-
-    # Get warmup ratio
-
-    # Get save ratio
-
-    # Get eval ratio
-
-    # Get dataset config
-
-    # Load metric tracker
-
-    # Save config to yaml
-    # TODO: improve output yaml formatting. Need to add comments to help separate sections
-    with open(config_path, "w", encoding="utf-8") as f:
-        yaml.dump(cfg.to_dict(), f, sort_keys=False)

src/axolotl/common/datasets.py

@@ -14,7 +14,6 @@ from axolotl.utils.data import prepare_dataset
 from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_processor, load_tokenizer
-from axolotl.utils.schemas.enums import RLType
 from axolotl.utils.tokenization import check_dataset_labels
 
 LOG = logging.getLogger(__name__)
@@ -134,7 +133,7 @@ def load_preference_datasets(
     total_num_steps: Optional[int] = int(
         math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
     )
-    if cfg.rl is RLType.GRPO:
+    if cfg.rl == "grpo":
         total_num_steps = None
 
     if cli_args.debug or cfg.debug:

src/axolotl/core/trainer_builder.py

@@ -21,7 +21,6 @@ import importlib.util
 import inspect
 import logging
 import math
-import os
 import sys
 from abc import abstractmethod
 from pathlib import Path
@@ -73,7 +72,6 @@ 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
@@ -87,7 +85,7 @@ from axolotl.utils.collators import (
 )
 from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
 from axolotl.utils.models import ensure_dtype
-from axolotl.utils.schemas.enums import CustomSupportedOptimizers, RLType
+from axolotl.utils.schemas.enums import CustomSupportedOptimizers
 
 try:
     import torch._dynamo  # pylint: disable=ungrouped-imports
@@ -170,9 +168,6 @@ 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)
@@ -254,6 +249,9 @@ 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):
@@ -295,10 +293,6 @@ 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
 
@@ -353,7 +347,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["warmup_steps"] = warmup_steps
         training_arguments_kwargs["logging_steps"] = logging_steps
 
-        if self.cfg.seed is not None:
+        if self.cfg.seed:
             training_arguments_kwargs["seed"] = self.cfg.seed
 
         if self.cfg.gradient_checkpointing:
@@ -547,6 +541,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         report_to = []
         if self.cfg.use_wandb:
             report_to.append("wandb")
+            if self.cfg.wandb_name:
+                training_arguments_kwargs["run_name"] = self.cfg.wandb_name
         if self.cfg.use_mlflow:
             report_to.append("mlflow")
         if self.cfg.use_tensorboard:
@@ -706,20 +702,6 @@ 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:
@@ -819,15 +801,14 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         data_collator_kwargs = {
             "padding": True,  # True/"longest" is the default
         }
-        multiple = 64
         if self.cfg.pad_to_sequence_len:
-            data_collator_kwargs["pad_to_multiple_of"] = multiple * math.ceil(
+            data_collator_kwargs["pad_to_multiple_of"] = 64 * math.ceil(
-                self.cfg.sequence_len / multiple
+                self.cfg.sequence_len / 64
             )
         else:
             # A100 is best at 64, while others at 8. Let's use the larger so we don't have to check
             # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
-            data_collator_kwargs["pad_to_multiple_of"] = multiple
+            data_collator_kwargs["pad_to_multiple_of"] = 64
 
         if self.cfg.reward_model:
             data_collator_kwargs["max_length"] = self.cfg.sequence_len
@@ -1033,10 +1014,6 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             training_args_kwargs["dataloader_prefetch_factor"] = (
                 self.cfg.dataloader_prefetch_factor
             )
-
-        if self.cfg.seed is not None:
-            training_args_kwargs["seed"] = self.cfg.seed
-
         if self.cfg.gradient_checkpointing:
             training_args_kwargs["gradient_checkpointing"] = (
                 self.cfg.gradient_checkpointing
@@ -1060,8 +1037,6 @@ 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
 
@@ -1079,13 +1054,9 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.cfg.use_wandb:
             training_args_kwargs["run_name"] = self.cfg.wandb_name
 
-        training_args_kwargs["sequence_parallel_degree"] = (
-            self.cfg.sequence_parallel_degree
-        )
-
         training_args_cls = None
         blocklist_args_kwargs = []
-        if self.cfg.rl is RLType.SIMPO:
+        if self.cfg.rl == "simpo":
             training_args_cls = AxolotlCPOConfig
             training_args_kwargs["loss_type"] = "simpo"
             training_args_kwargs["max_length"] = self.cfg.sequence_len
@@ -1093,13 +1064,13 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.cpo_alpha is not None:
                 training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
 
-        elif self.cfg.rl is RLType.ORPO:
+        elif self.cfg.rl == "orpo":
             training_args_cls = AxolotlORPOConfig
             training_args_kwargs["max_length"] = self.cfg.sequence_len
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
-        elif self.cfg.rl is RLType.KTO:
+        elif self.cfg.rl == "kto":
             training_args_cls = AxolotlKTOConfig
 
             training_args_kwargs["desirable_weight"] = (
@@ -1113,14 +1084,14 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
-        elif self.cfg.rl is RLType.GRPO:
+        elif self.cfg.rl == "grpo":
             training_args_cls = GRPOStrategy.get_training_args_class()
             training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
             blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
 
         else:
             training_args_cls = AxolotlDPOConfig
-            if self.cfg.rl is RLType.IPO:
+            if self.cfg.rl == "ipo":
                 training_args_kwargs["loss_type"] = "ipo"
             training_args_kwargs["max_length"] = self.cfg.sequence_len
             training_args_kwargs["max_completion_length"] = None
@@ -1163,73 +1134,67 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
 
     def build(self, total_num_steps):
         training_args = self.build_training_arguments(total_num_steps)
-        trainer_kwargs = {}
+        dpo_trainer_kwargs = {}
-        if self.cfg.rl is RLType.IPO:
+        if self.cfg.rl == "ipo":
             if self.cfg.dpo_label_smoothing:
-                trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
+                dpo_trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
         if self.eval_dataset:
-            trainer_kwargs["eval_dataset"] = self.eval_dataset
+            dpo_trainer_kwargs["eval_dataset"] = self.eval_dataset
         if self.cfg.adapter and self.peft_config:
-            trainer_kwargs["peft_config"] = self.peft_config
+            dpo_trainer_kwargs["peft_config"] = self.peft_config
         if self.cfg.precompute_ref_log_probs is not None:
-            trainer_kwargs["precompute_ref_log_probs"] = (
+            dpo_trainer_kwargs["precompute_ref_log_probs"] = (
                 self.cfg.precompute_ref_log_probs
             )
-        if self.cfg.rl is RLType.GRPO:
+        if self.cfg.rl == "grpo":
-            trainer_cls = GRPOStrategy.get_trainer_class(
+            trainer_cls = GRPOStrategy.get_trainer_class()
-                sequence_parallel=self.cfg.sequence_parallel_degree > 1
-            )
             trainer_cls_args = [self.model]
             trainer_cls_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
-            trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
+            dpo_trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
-        elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
+        elif self.cfg.rl in ["dpo", "ipo"]:
             trainer_cls = DPOStrategy.get_trainer_class()
             trainer_cls_args = [self.model, self.model_ref]
-        elif self.cfg.rl is RLType.ORPO:
+        elif self.cfg.rl == "orpo":
             trainer_cls = AxolotlORPOTrainer
             trainer_cls_args = [self.model]
-        elif self.cfg.rl is RLType.KTO:
+        elif self.cfg.rl in ["kto"]:
             trainer_cls = AxolotlKTOTrainer
             trainer_cls_args = [self.model]
-        elif self.cfg.rl is RLType.SIMPO:
+        elif self.cfg.rl in ["simpo"]:
             trainer_cls = AxolotlCPOTrainer
             trainer_cls_args = [self.model]
         else:
             raise ValueError(f"Unsupported RL: {self.cfg.rl}")
 
-        if self.cfg.plugins:
-            plugin_manager = PluginManager.get_instance()
-            trainer_cls = plugin_manager.get_trainer_cls(self.cfg)
-
         sig = inspect.signature(trainer_cls)
         if "tokenizer" in sig.parameters.keys():
-            trainer_kwargs["tokenizer"] = self.tokenizer
+            dpo_trainer_kwargs["tokenizer"] = self.tokenizer
         else:
-            trainer_kwargs["processing_class"] = self.tokenizer
+            dpo_trainer_kwargs["processing_class"] = self.tokenizer
 
         if self.cfg.datasets is not None and (
             trainer_cls is DPOStrategy.get_trainer_class()
         ):
-            trainer_kwargs["dataset_tags"] = [
+            dpo_trainer_kwargs["dataset_tags"] = [
                 d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
             ]
-        trainer = trainer_cls(
+        dpo_trainer = trainer_cls(
             *trainer_cls_args,
             args=training_args,
             train_dataset=self.train_dataset,
             callbacks=self.get_callbacks(),
-            **trainer_kwargs,
+            **dpo_trainer_kwargs,
         )
         if self.cfg.fsdp:
-            ensure_dtype(trainer.model, dtype=self.cfg.torch_dtype)
+            ensure_dtype(dpo_trainer.model, dtype=self.cfg.torch_dtype)
-            if self.cfg.rl in [RLType.DPO, RLType.IPO] and trainer.ref_model:
+            if self.cfg.rl in ["dpo", "ipo"] and dpo_trainer.ref_model:
-                ensure_dtype(trainer.ref_model, dtype=self.cfg.torch_dtype)
+                ensure_dtype(dpo_trainer.ref_model, dtype=self.cfg.torch_dtype)
 
-        trainer = self.hook_post_create_trainer(trainer)
+        dpo_trainer = self.hook_post_create_trainer(dpo_trainer)
-        for callback in self.get_post_trainer_create_callbacks(trainer):
+        for callback in self.get_post_trainer_create_callbacks(dpo_trainer):
-            trainer.add_callback(callback)
+            dpo_trainer.add_callback(callback)
 
-        return trainer
+        return dpo_trainer
 
 
 class HFPPOTrainerBuilder(TrainerBuilderBase):

src/axolotl/core/trainers/__init__.py

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

src/axolotl/core/trainers/base.py

@@ -373,13 +373,15 @@ class AxolotlTrainer(
                 num_items_in_batch=num_items_in_batch,
             )
 
-        return super().compute_loss(
+        loss = super().compute_loss(
             model,
             inputs,
             return_outputs=return_outputs,
             num_items_in_batch=num_items_in_batch,
         )
 
+        return loss
+
     @staticmethod
     def orpo_concatenate_inputs(inputs, label_pad_token=-100, pad_token=0, device=None):
         concatenated_batch = {}

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

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

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

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

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

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

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

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

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

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

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

@@ -1,63 +1,23 @@
-"""Axolotl GRPO trainers (with and without sequence parallelism handling)"""
+"""
+Axolotl GRPO trainer
+"""
 
-# pylint: disable=too-many-lines,duplicate-code,protected-access,no-member
-
-import warnings
 from contextlib import nullcontext
-from typing import Any
 
-import datasets
+from accelerate.utils import is_deepspeed_available, is_peft_model
-import torch
-import torch.distributed as dist
-import torch.utils.data
-from accelerate.utils import (
-    broadcast_object_list,
-    gather,
-    gather_object,
-    is_peft_model,
-)
-from datasets import Dataset, IterableDataset
-from torch import nn
-from torch.utils.data import (
-    BatchSampler,
-    DataLoader,
-    Sampler,
-)
-from transformers import (
-    PreTrainedModel,
-    PreTrainedTokenizerBase,
-    Trainer,
-    TrainerCallback,
-)
-from transformers.trainer_utils import seed_worker
-from transformers.utils import is_peft_available
 from trl import GRPOTrainer
-from trl.data_utils import (
+from trl.extras.profiling import profiling_decorator
-    apply_chat_template,
-    is_conversational,
-    maybe_apply_chat_template,
-)
-from trl.extras.profiling import profiling_context, profiling_decorator
-from trl.import_utils import is_deepspeed_available
-from trl.models import unwrap_model_for_generation
-from trl.trainer.grpo_config import GRPOConfig
-from trl.trainer.grpo_trainer import RewardFunc, nanstd
-from trl.trainer.utils import pad
 
-from axolotl.core.trainers.grpo.sampler import SequenceParallelRepeatRandomSampler
 from axolotl.core.trainers.mixins import RngLoaderMixin, SchedulerMixin
-from axolotl.monkeypatch.attention.ring_attn.patch import get_ring_attn_group
-
-if is_peft_available():
-    # pylint: disable=unused-import
-    from peft import PeftConfig
 
 if is_deepspeed_available():
     import deepspeed
 
 
 class AxolotlGRPOTrainer(RngLoaderMixin, SchedulerMixin, GRPOTrainer):
-    """Extend the base GRPOTrainer for axolotl helpers"""
+    """
+    Extend the base GRPOTrainer for axolotl helpers
+    """
 
     _tag_names = ["trl", "grpo", "axolotl"]
 
@@ -107,600 +67,3 @@ class AxolotlGRPOTrainer(RngLoaderMixin, SchedulerMixin, GRPOTrainer):
         # Reset cache on main process
         if self.accelerator.is_main_process:
             self.vllm_client.reset_prefix_cache()
-
-
-class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
-    """Extend the base GRPOTrainer for sequence parallelism handling"""
-
-    def __init__(
-        self,
-        model: str | PreTrainedModel,
-        reward_funcs: RewardFunc | list[RewardFunc],
-        args: GRPOConfig | None = None,
-        train_dataset: Dataset | IterableDataset | None = None,
-        eval_dataset: (
-            Dataset | IterableDataset | dict[str, Dataset | IterableDataset] | None
-        ) = None,
-        processing_class: PreTrainedTokenizerBase | None = None,
-        reward_processing_classes: (
-            PreTrainedTokenizerBase | list[PreTrainedTokenizerBase] | None
-        ) = None,
-        callbacks: list[TrainerCallback] | None = None,
-        optimizers: tuple[
-            torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None
-        ] = (None, None),
-        peft_config: "PeftConfig | None" = None,
-    ):
-        # First call the superclass constructor with all arguments
-        super().__init__(
-            model=model,
-            reward_funcs=reward_funcs,
-            args=args,
-            train_dataset=train_dataset,
-            eval_dataset=eval_dataset,
-            processing_class=processing_class,
-            reward_processing_classes=reward_processing_classes,
-            callbacks=callbacks,
-            optimizers=optimizers,
-            peft_config=peft_config,
-        )
-
-        # Get number of SP groups (number of processes divided by SP degree)
-        num_processes = self.accelerator.num_processes
-        num_sp_groups = num_processes // self.args.sequence_parallel_degree
-
-        # Calculate batch size per SP group (not per process)
-        sp_group_batch_size = self.args.per_device_train_batch_size * num_sp_groups
-        possible_values = [
-            n_gen
-            for n_gen in range(2, sp_group_batch_size + 1)
-            if (sp_group_batch_size) % n_gen == 0
-        ]
-
-        if self.num_generations not in possible_values:
-            raise ValueError(
-                f"The batch size per SP group ({num_sp_groups} x "
-                f"{self.args.per_device_train_batch_size}) must be evenly divisible by "
-                f"the number of generations per prompt ({self.num_generations}). Given "
-                "the current configuration, the valid values for the number of "
-                f"generations are: {possible_values}."
-            )
-
-        if self.args.eval_strategy != "no":
-            # If sequence parallelism is enabled, calculate batch size per SP group
-            sp_group_eval_batch_size = args.per_device_eval_batch_size * num_sp_groups  # type: ignore[union-attr]
-            possible_values = [
-                n_gen
-                for n_gen in range(2, sp_group_eval_batch_size + 1)
-                if (sp_group_eval_batch_size) % n_gen == 0
-            ]
-
-            if self.num_generations not in possible_values:
-                raise ValueError(
-                    f"With sequence parallelism (degree {self.args.sequence_parallel_degree}), "
-                    f"the eval batch size per SP group ({num_sp_groups} x {self.args.per_device_eval_batch_size}) "
-                    f"must be evenly divisible by the number of generations per prompt "
-                    f"({self.num_generations}). Given the current eval batch size, "
-                    f"the valid values for the number of generations are: {possible_values}."
-                )
-
-        # Initialize the SP group
-        self.sp_group = get_ring_attn_group()
-        self.rank = dist.get_rank()
-        self.world_size = dist.get_world_size()
-        self.local_rank = dist.get_rank(group=self.sp_group)
-        self.local_world_size = dist.get_world_size(group=self.sp_group)
-
-    def _get_train_sampler(self) -> Sampler:
-        effective_batch_size = (
-            self.args.per_device_train_batch_size
-            * self.world_size
-            * self.args.gradient_accumulation_steps
-        )
-
-        return SequenceParallelRepeatRandomSampler(
-            dataset=self.train_dataset,
-            mini_repeat_count=self.num_generations,
-            world_size=self.world_size,
-            rank=self.rank,
-            batch_size=effective_batch_size
-            // self.num_generations
-            // self.args.sequence_parallel_degree,
-            repeat_count=self.num_iterations * self.args.gradient_accumulation_steps,
-            sequence_parallel_degree=self.args.sequence_parallel_degree,
-            shuffle=True,
-            seed=self.args.seed,
-            drop_last=True,
-        )
-
-    def _create_dataloader_params(self, is_eval=False, custom_batch_size=None):
-        """Create common dataloader parameters for train or eval."""
-        batch_size = custom_batch_size or (
-            self.args.eval_batch_size if is_eval else self._train_batch_size
-        )
-
-        params = {
-            "batch_size": batch_size,
-            "collate_fn": self.data_collator,
-            "num_workers": self.args.dataloader_num_workers,
-            "pin_memory": self.args.dataloader_pin_memory,
-        }
-
-        # Add persistent workers only for training
-        if not is_eval and hasattr(self.args, "dataloader_persistent_workers"):
-            params["persistent_workers"] = self.args.dataloader_persistent_workers
-
-        # Add prefetch factor if specified
-        if self.args.dataloader_prefetch_factor:
-            params["prefetch_factor"] = self.args.dataloader_prefetch_factor
-
-        return params
-
-    def _prepare_dataloader(
-        self, dataset, sampler, is_eval=False, custom_batch_size=None
-    ):
-        """Prepare a dataloader with the given dataset and sampler."""
-        # Get base parameters
-        dataloader_params = self._create_dataloader_params(is_eval, custom_batch_size)
-
-        # Add sampler configuration
-        if not isinstance(dataset, torch.utils.data.IterableDataset):
-            if isinstance(sampler, BatchSampler):
-                # batch_size and batch_sampler are mutually exclusive
-                dataloader_params["batch_sampler"] = sampler
-                del dataloader_params["batch_size"]
-            else:
-                dataloader_params["sampler"] = sampler
-                dataloader_params["drop_last"] = self.args.dataloader_drop_last
-
-            if not is_eval:
-                dataloader_params["worker_init_fn"] = seed_worker
-
-        # Create the dataloader
-        dataloader = DataLoader(dataset, **dataloader_params)
-
-        if self.args.sample_packing and (
-            (not is_eval and not self.args.pretraining)
-            or (is_eval and self.args.eval_sample_packing is not False)
-        ):
-            self.accelerator.even_batches = False
-
-        # Return unprepared dataloader if using sequence parallelism
-        # TODO(djsaunde): We might be able to use `accelerate`'s dataloader preparation
-        # if we use `dispatch_batches` and `slice_fn_for_dispatch` properly (i.e.,
-        # slice each batch along the sequence dimension).
-        if self.args.sequence_parallel_degree > 1:
-            return dataloader
-
-        # Otherwise prepare with accelerator
-        return self.accelerator.prepare_data_loader(dataloader)
-
-    def get_train_dataloader(self) -> DataLoader:
-        """Get dataloader for training"""
-        train_dataset = self.train_dataset
-        # pylint: disable=access-member-before-definition
-        data_collator = self.data_collator  # type: ignore
-
-        # Handle dataset preprocessing
-        if isinstance(train_dataset, datasets.Dataset):
-            # Add debug print before any modifications
-            if self.args.sample_packing and not self.args.pretraining:
-                train_dataset = train_dataset.remove_columns(["length"])
-            if not self.args.sample_packing or self.args.pretraining:
-                train_dataset = self._remove_unused_columns(
-                    train_dataset, description="training"
-                )
-        else:
-            self.data_collator = self._get_collator_with_removed_columns(  # pylint: disable=attribute-defined-outside-init
-                data_collator,
-                description="training",
-            )
-
-        # Get sampler and create dataloader
-        sampler = self._get_train_sampler()
-        dataloader = self._prepare_dataloader(train_dataset, sampler, is_eval=False)
-
-        return dataloader
-
-    def _generate_and_score_completions(
-        self, inputs: list[dict[str, torch.Tensor | Any]]
-    ) -> dict[str, torch.Tensor | Any]:
-        device = self.accelerator.device
-        mode = "eval" if self.control.should_evaluate else "train"
-
-        prompts = [x["prompt"] for x in inputs]
-        prompts_text = [
-            maybe_apply_chat_template(example, self.processing_class)["prompt"]
-            for example in inputs
-        ]
-        prompt_inputs = self.processing_class(
-            text=prompts_text,
-            return_tensors="pt",
-            padding=True,
-            padding_side="left",
-            add_special_tokens=False,
-        )
-        prompt_inputs = Trainer._prepare_inputs(self, prompt_inputs)
-        prompt_ids, prompt_mask = (
-            prompt_inputs["input_ids"],
-            prompt_inputs["attention_mask"],
-        )
-
-        if self.max_prompt_length is not None:
-            prompt_ids = prompt_ids[:, -self.max_prompt_length :]
-            prompt_mask = prompt_mask[:, -self.max_prompt_length :]
-
-        # Generate completions using either vLLM or regular generation
-        if self.args.use_vllm:
-            # First, have main process load weights if needed
-            # pylint: disable=access-member-before-definition
-            if self.state.global_step != self._last_loaded_step:  # type: ignore[has-type]
-                self._move_model_to_vllm()
-                # pylint: disable=attribute-defined-outside-init
-                self._last_loaded_step = self.state.global_step
-
-            # Generate completions using vLLM: gather all prompts and use them in a single call in the main process
-            all_prompts_text = gather_object(prompts_text)
-            if self.accelerator.is_main_process:
-                if self.args.sequence_parallel_degree > 1:
-                    # Calculate sequence parallel group information
-                    world_size = self.accelerator.num_processes
-                    sequence_parallel_degree = self.args.sequence_parallel_degree
-                    num_sp_groups = world_size // sequence_parallel_degree
-
-                    # Since processes in the same SP group have the same prompts, we need to ensure
-                    # we only take one copy of each prompt from each SP group
-                    ordered_set_of_prompts = []
-                    for sp_group_id in range(num_sp_groups):
-                        # Get the first process from each SP group (typically the group leader)
-                        group_leader_rank = sp_group_id * sequence_parallel_degree
-
-                        # Extract prompts from this SP group, accounting for num_generations duplicates
-                        # We only need prompts from one rank in each SP group
-                        group_prompts = all_prompts_text[
-                            group_leader_rank
-                            * len(prompts_text) : (group_leader_rank + 1)
-                            * len(prompts_text) : self.num_generations
-                        ]
-
-                        ordered_set_of_prompts.extend(group_prompts)
-                else:
-                    # Since 'prompts' contains 'num_generations' duplicates, we first take unique prompts, and generate
-                    # num_generations outputs for each one. This is faster than generating outputs for each duplicate
-                    # prompt individually.
-                    ordered_set_of_prompts = all_prompts_text[
-                        :: self.num_generations * self.args.sequence_parallel_degree
-                    ]
-
-                with profiling_context(self, "vLLM.generate"):
-                    completion_ids = self.vllm_client.generate(
-                        prompts=ordered_set_of_prompts,
-                        n=self.num_generations,
-                        repetition_penalty=self.repetition_penalty,
-                        temperature=self.temperature,
-                        top_p=self.top_p,
-                        top_k=-1 if self.top_k is None else self.top_k,
-                        min_p=0.0 if self.min_p is None else self.min_p,
-                        max_tokens=self.max_completion_length,
-                        guided_decoding_regex=self.guided_decoding_regex,
-                    )
-            else:
-                completion_ids = [None] * (
-                    len(all_prompts_text) // self.args.sequence_parallel_degree
-                )
-
-            # Broadcast the completions from the main process to all processes
-            completion_ids = broadcast_object_list(completion_ids, from_process=0)
-
-            # Determine the appropriate slice based on sequence parallelism
-            if self.args.sequence_parallel_degree > 1:
-                # Calculate SP group ID (which group of ranks this rank belongs to)
-                sp_group_id = self.accelerator.process_index // self.local_world_size
-
-                # Calculate the start index for this SP group
-                sp_group_start = sp_group_id * len(prompts) * self.local_world_size
-
-                # All ranks in the same SP group get the same data slice
-                process_slice = slice(
-                    sp_group_start,
-                    sp_group_start + len(prompts),
-                )
-                completion_ids = completion_ids[process_slice]
-            else:
-                # Original behavior for non-sequence parallel case
-                process_slice = slice(
-                    self.accelerator.process_index * len(prompts),
-                    (self.accelerator.process_index + 1) * len(prompts),
-                )
-                completion_ids = completion_ids[process_slice]
-
-            # Pad the completions, and concatenate them with the prompts
-            completion_ids = [
-                torch.tensor(ids, device=device) for ids in completion_ids
-            ]
-            completion_ids = pad(
-                completion_ids, padding_value=self.processing_class.pad_token_id
-            )
-            prompt_completion_ids = torch.cat([prompt_ids, completion_ids], dim=1)
-        else:
-            # Regular generation path
-            with unwrap_model_for_generation(
-                self.model_wrapped,
-                self.accelerator,
-                gather_deepspeed3_params=self.args.ds3_gather_for_generation,
-            ) as unwrapped_model:
-                prompt_completion_ids = unwrapped_model.generate(
-                    prompt_ids,
-                    attention_mask=prompt_mask,
-                    generation_config=self.generation_config,
-                )
-
-            # Compute prompt length and extract completion ids
-            prompt_length = prompt_ids.size(1)
-            prompt_ids = prompt_completion_ids[:, :prompt_length]
-            completion_ids = prompt_completion_ids[:, prompt_length:]
-
-        # Mask everything after the first EOS token
-        is_eos = completion_ids == self.processing_class.eos_token_id
-        eos_idx = torch.full(
-            (is_eos.size(0),), is_eos.size(1), dtype=torch.long, device=device
-        )
-        eos_idx[is_eos.any(dim=1)] = is_eos.int().argmax(dim=1)[is_eos.any(dim=1)]
-        sequence_indices = torch.arange(is_eos.size(1), device=device).expand(
-            is_eos.size(0), -1
-        )
-        completion_mask = (sequence_indices <= eos_idx.unsqueeze(1)).int()
-
-        # If mask_truncated_completions is enabled, zero out truncated completions in completion_mask
-        if self.args.mask_truncated_completions:
-            truncated_completions = ~is_eos.any(dim=1)
-            completion_mask = (
-                completion_mask * (~truncated_completions).unsqueeze(1).int()
-            )
-
-        # Concatenate prompt_mask with completion_mask for logit computation
-        attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)  # (B, P+C)
-
-        logits_to_keep = completion_ids.size(
-            1
-        )  # we only need to compute the logits for the completion tokens
-        batch_size = (
-            self.args.per_device_train_batch_size
-            if mode == "train"
-            else self.args.per_device_eval_batch_size
-        )
-
-        with torch.no_grad():
-            # When using num_iterations == 1, old_per_token_logps == per_token_logps, so we can skip it's
-            # computation here, and use per_token_logps.detach() instead.
-            if self.num_iterations > 1:
-                old_per_token_logps = self._get_per_token_logps(
-                    self.model,
-                    prompt_completion_ids,
-                    attention_mask,
-                    logits_to_keep,
-                    batch_size,
-                )
-            else:
-                old_per_token_logps = None
-
-            if self.beta == 0.0:
-                ref_per_token_logps = None
-            elif self.ref_model is not None:
-                ref_per_token_logps = self._get_per_token_logps(
-                    self.ref_model,
-                    prompt_completion_ids,
-                    attention_mask,
-                    logits_to_keep,
-                    batch_size,
-                )
-            else:
-                with self.accelerator.unwrap_model(self.model).disable_adapter():
-                    ref_per_token_logps = self._get_per_token_logps(
-                        self.model,
-                        prompt_completion_ids,
-                        attention_mask,
-                        logits_to_keep,
-                        batch_size,
-                    )
-
-        # Decode the generated completions
-        completions_text = self.processing_class.batch_decode(
-            completion_ids, skip_special_tokens=True
-        )
-        if is_conversational(inputs[0]):
-            completions = []
-            for prompt, completion in zip(prompts, completions_text):
-                bootstrap = (
-                    prompt.pop()["content"] if prompt[-1]["role"] == "assistant" else ""
-                )
-                completions.append(
-                    [{"role": "assistant", "content": bootstrap + completion}]
-                )
-        else:
-            completions = completions_text
-
-        rewards_per_func = torch.zeros(
-            len(prompts), len(self.reward_funcs), device=device
-        )
-        for i, (reward_func, reward_processing_class, reward_func_name) in enumerate(
-            zip(
-                self.reward_funcs,
-                self.reward_processing_classes,
-                self.reward_func_names,
-            )
-        ):
-            with profiling_context(self, reward_func_name):
-                if isinstance(
-                    reward_func, nn.Module
-                ):  # Module instead of PretrainedModel for compat with compiled models
-                    if is_conversational(inputs[0]):
-                        messages = [
-                            {"messages": p + c} for p, c in zip(prompts, completions)
-                        ]
-                        texts = [
-                            apply_chat_template(x, reward_processing_class)["text"]
-                            for x in messages
-                        ]
-                    else:
-                        texts = [p + c for p, c in zip(prompts, completions)]
-                    reward_inputs = reward_processing_class(
-                        text=texts,
-                        return_tensors="pt",
-                        padding=True,
-                        padding_side="right",
-                        add_special_tokens=False,
-                    )
-                    reward_inputs = Trainer._prepare_inputs(self, reward_inputs)
-                    with torch.inference_mode():
-                        rewards_per_func[:, i] = reward_func(**reward_inputs).logits[
-                            :, 0
-                        ]  # Shape (B*G,)
-                else:
-                    # Repeat all input columns (but "prompt" and "completion") to match the number of generations
-                    keys = [
-                        key for key in inputs[0] if key not in ["prompt", "completion"]
-                    ]
-                    reward_kwargs = {
-                        key: [example[key] for example in inputs] for key in keys
-                    }
-                    output_reward_func = reward_func(
-                        prompts=prompts, completions=completions, **reward_kwargs
-                    )
-                    # Convert None values to NaN
-                    output_reward_func = [
-                        reward if reward is not None else torch.nan
-                        for reward in output_reward_func
-                    ]
-
-                    rewards_per_func[:, i] = torch.tensor(
-                        output_reward_func, dtype=torch.float32, device=device
-                    )
-
-        # If all reward functions return None for a given row, issue a detailed warning
-        if torch.isnan(rewards_per_func).all(dim=1).any():
-            nan_row_idx = (
-                torch.isnan(rewards_per_func).all(dim=1).nonzero(as_tuple=True)[0][0]
-            )
-            row_reward_kwargs = {
-                key: value[nan_row_idx] for key, value in reward_kwargs.items()
-            }
-            row_reward_kwargs["prompt"] = prompts[nan_row_idx]
-            row_reward_kwargs["completion"] = completions[nan_row_idx]
-            warnings.warn(
-                f"All reward functions returned None for the following kwargs: {row_reward_kwargs}. "
-                "Please ensure that at least one reward function returns a valid reward."
-            )
-
-        # Gather the reward per function: this part is crucial, because the rewards are normalized per group and the
-        # completions may be distributed across processes
-        rewards_per_func = gather(rewards_per_func)
-
-        # Apply weights to each reward function's output and sum
-        rewards = (
-            rewards_per_func * self.reward_weights.to(device).unsqueeze(0)
-        ).nansum(dim=1)
-
-        # Compute grouped-wise rewards
-        mean_grouped_rewards = rewards.view(-1, self.num_generations).mean(dim=1)
-        std_grouped_rewards = rewards.view(-1, self.num_generations).std(dim=1)
-
-        # Normalize the rewards to compute the advantages
-        mean_grouped_rewards = mean_grouped_rewards.repeat_interleave(
-            self.num_generations, dim=0
-        )
-        std_grouped_rewards = std_grouped_rewards.repeat_interleave(
-            self.num_generations, dim=0
-        )
-        advantages = rewards - mean_grouped_rewards
-        if self.args.scale_rewards:
-            advantages = advantages / (std_grouped_rewards + 1e-4)
-
-        # Slice to keep only the local part of the data
-        if self.args.sequence_parallel_degree > 1:
-            # Calculate SP group ID (which group of ranks this rank belongs to)
-            sp_group_id = self.accelerator.process_index // self.local_world_size
-
-            # Calculate the start index for this SP group
-            sp_group_start = sp_group_id * len(prompts) * self.local_world_size
-
-            # All ranks in the same SP group get the same data slice
-            process_slice = slice(
-                sp_group_start,
-                sp_group_start + len(prompts),
-            )
-        else:
-            # Original behavior for non-sequence parallel case
-            process_slice = slice(
-                self.accelerator.process_index * len(prompts),
-                (self.accelerator.process_index + 1) * len(prompts),
-            )
-        advantages = advantages[process_slice]
-
-        # Log the metrics
-        if mode == "train":
-            self._total_train_tokens += (
-                self.accelerator.gather_for_metrics(attention_mask.sum()).sum().item()
-            )
-        self._metrics[mode]["num_tokens"] = [self._total_train_tokens]
-
-        # log completion lengths, mean, min, max
-        agg_completion_mask = self.accelerator.gather_for_metrics(
-            completion_mask.sum(1)
-        )
-        self._metrics[mode]["completions/mean_length"].append(
-            agg_completion_mask.float().mean().item()
-        )
-        self._metrics[mode]["completions/min_length"].append(
-            agg_completion_mask.float().min().item()
-        )
-        self._metrics[mode]["completions/max_length"].append(
-            agg_completion_mask.float().max().item()
-        )
-
-        # identify sequences that terminated with EOS and log their lengths
-        agg_terminated_with_eos = self.accelerator.gather_for_metrics(is_eos.any(dim=1))
-        term_completion_mask = agg_completion_mask[agg_terminated_with_eos]
-        clipped_completions_ratio = 1 - len(term_completion_mask) / len(
-            agg_completion_mask
-        )
-        self._metrics[mode]["completions/clipped_ratio"].append(
-            clipped_completions_ratio
-        )
-        if len(term_completion_mask) == 0:
-            # edge case where no completed sequences are found
-            term_completion_mask = torch.zeros(1, device=device)
-        self._metrics[mode]["completions/mean_terminated_length"].append(
-            term_completion_mask.float().mean().item()
-        )
-        self._metrics[mode]["completions/min_terminated_length"].append(
-            term_completion_mask.float().min().item()
-        )
-        self._metrics[mode]["completions/max_terminated_length"].append(
-            term_completion_mask.float().max().item()
-        )
-
-        # Calculate mean reward per function, but only for samples where the function was applied (non-NaN values)
-        for i, reward_func_name in enumerate(self.reward_func_names):
-            mean_rewards = torch.nanmean(rewards_per_func[:, i]).item()
-            self._metrics[mode][f"rewards/{reward_func_name}/mean"].append(mean_rewards)
-            std_rewards = nanstd(rewards_per_func[:, i]).item()
-            self._metrics[mode][f"rewards/{reward_func_name}/std"].append(std_rewards)
-        self._metrics[mode]["reward"].append(mean_grouped_rewards.mean().item())
-        self._metrics[mode]["reward_std"].append(std_grouped_rewards.mean().item())
-
-        # Log prompt and completion texts
-        self._textual_logs["prompt"].extend(gather_object(prompts_text))
-        self._textual_logs["completion"].extend(gather_object(completions_text))
-        for i, name in enumerate(self.reward_func_names):
-            self._textual_logs["rewards"][name].extend(rewards_per_func[:, i].tolist())
-
-        return {
-            "prompt_ids": prompt_ids,
-            "prompt_mask": prompt_mask,
-            "completion_ids": completion_ids,
-            "completion_mask": completion_mask,
-            "advantages": advantages,
-            "old_per_token_logps": old_per_token_logps,
-            "ref_per_token_logps": ref_per_token_logps,
-        }

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

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

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

@@ -1,13 +1,85 @@
-"""Module for Axolotl trainer sequence parallelism mixin"""
+"""
+Module for Axolotl trainer sequence parallelism mixin and training context manager
+"""
 
+import functools
+import logging
+
+import torch
 import torch.distributed as dist
 from datasets import Dataset
+from torch import nn
 from torch.utils.data import DistributedSampler, Sampler
+from torch.utils.hooks import RemovableHandle
 
 from axolotl.monkeypatch.attention.ring_attn import (
+    RingAttnFunc,
     get_ring_attn_group,
+    update_ring_attn_params,
 )
 
+LOG = logging.getLogger(__name__)
+
+
+def apply_sequence_parallelism(
+    batch: dict[str, torch.Tensor],
+    local_rank: int,
+    local_world_size: int,
+    ring_attn_func: RingAttnFunc,
+) -> dict[str, torch.Tensor]:
+    """
+    Apply sequence parallelism slicing to a batch.
+
+    Args:
+        batch: Batch dictionary (e.g., input_ids, attention_mask, etc.)
+        local_rank: Local rank in the sequence parallel group
+        local_world_size: World size of the sequence parallel group
+        ring_attn_func: The ring attention function to use
+
+    Returns:
+        Sliced batch dictionary.
+    """
+    # Update ring attention params if needed
+    if batch.get("position_ids") is not None:
+        update_ring_attn_params(position_ids=batch["position_ids"])
+
+    # Slice batch for sequence parallel processing
+    total_seq_len = batch["input_ids"].size(1)
+    for key in batch:
+        if (
+            key in batch
+            and isinstance(batch[key], torch.Tensor)
+            and batch[key].dim() > 1
+            and batch[key].size(1) == total_seq_len
+        ):
+
+            if ring_attn_func in [
+                RingAttnFunc.VARLEN_LLAMA3,
+                RingAttnFunc.BATCH_RING,
+            ]:
+                # Split in sequential fashion and grab this rank's chunk
+                batch[key] = (
+                    batch[key].chunk(local_world_size, dim=1)[local_rank].contiguous()
+                )
+            elif ring_attn_func is RingAttnFunc.BATCH_ZIGZAG:
+                chunks = batch[key].chunk(2 * local_world_size, dim=1)
+
+                # Take rank's chunk and opposing chunk for zigzag pattern
+                selected_chunks = [
+                    chunks[local_rank],
+                    chunks[2 * local_world_size - local_rank - 1],
+                ]
+                batch[key] = torch.cat(selected_chunks, dim=1).contiguous()
+            elif ring_attn_func is RingAttnFunc.BATCH_STRIPE:
+                # Split into striped data and stack
+                tensor = torch.stack(
+                    batch[key].split(local_world_size, dim=1),
+                    dim=1,
+                ).transpose(1, 2)
+                batch[key] = tensor[:, local_rank].contiguous()
+
+    return batch
+
 
 class SequenceParallelMixin:
     """
@@ -85,3 +157,157 @@ class SequenceParallelMixin:
         return self._create_sequence_parallel_sampler(
             eval_dataset, shuffle=False, is_eval=True
         )
+
+
+class SequenceParallelContextManager:
+    """
+    Context manager for sequence parallelism operations.
+
+    This class provides a context that will automatically apply sequence parallelism
+    during model forward passes using a pre-forward hook, and gather outputs from
+    across the sequence parallelism group using a post-forward hook.
+    """
+
+    def __init__(
+        self,
+        model: nn.Module,
+        sequence_parallel_degree: int,
+        ring_attn_func: RingAttnFunc,
+    ):
+        self.model = model
+        self.sequence_parallel_degree = sequence_parallel_degree
+        self.ring_attn_func = ring_attn_func
+        self.process_group = get_ring_attn_group()
+
+        # Initialize sequence parallel group details
+        self.local_rank = dist.get_rank(self.process_group)
+        self.local_world_size = dist.get_world_size(self.process_group)
+
+        # Will store hook handles for removal
+        self.hook_handles: list[RemovableHandle] = []
+
+        # Create a partially applied version of the apply_sequence_parallelism function
+        # with pre-configured params
+        self.apply_sequence_parallelism = functools.partial(
+            apply_sequence_parallelism,
+            local_rank=self.local_rank,
+            local_world_size=self.local_world_size,
+            ring_attn_func=self.ring_attn_func,
+        )
+
+    def __enter__(self):
+        # Forward pre-hook to apply sequence parallelism
+        def sequence_parallel_pre_hook(_, args, kwargs):
+            # Apply sequence parallelism to kwargs
+            kwargs = self.apply_sequence_parallelism(batch=kwargs)
+            return args, kwargs
+
+        # Forward post-hook to gather outputs
+        def sequence_parallel_post_hook(_, __, output):
+            # Gather the sharded outputs
+            return self.gather_outputs(output)
+
+        # Register both hooks
+        self.hook_handles.append(
+            self.model.register_forward_pre_hook(
+                sequence_parallel_pre_hook, with_kwargs=True
+            )
+        )
+        self.hook_handles.append(
+            self.model.register_forward_hook(sequence_parallel_post_hook)
+        )
+
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        # Remove all hooks
+        for handle in self.hook_handles:
+            handle.remove()
+        self.hook_handles = []
+
+    def gather_outputs(self, output):
+        """Gather sharded outputs from all ranks and reconstruct the full tensor."""
+        # Handle different output formats (dict, tensor, etc.)
+        if isinstance(output, dict):
+            gathered_output = {}
+            for key, value in output.items():
+                if isinstance(value, torch.Tensor) and value.dim() > 1:
+                    # Gather logits or other sequence-sharded tensors
+                    gathered_value = self.gather_tensor(value)
+                    gathered_output[key] = gathered_value
+                else:
+                    gathered_value = value.clone()
+                    dist.all_reduce(
+                        gathered_value, op=dist.ReduceOp.SUM, group=self.process_group
+                    )
+                    gathered_output[key] = gathered_value
+            return gathered_output
+        if isinstance(output, torch.Tensor):
+            return self.gather_tensor(output)
+
+        return output
+
+    def gather_tensor(self, tensor):
+        """Gather a sharded tensor from all ranks."""
+        # Prepare tensors for all_gather
+        world_size = self.local_world_size
+
+        # Create list to store tensors from all ranks
+        gathered_tensors = [torch.zeros_like(tensor) for _ in range(world_size)]
+
+        # All-gather operation
+        dist.all_gather(gathered_tensors, tensor, group=self.process_group)
+
+        # Concatenate along sequence dimension (typically dim=1)
+        if self.ring_attn_func in [RingAttnFunc.VARLEN_LLAMA3, RingAttnFunc.BATCH_RING]:
+            # Simple concatenation for standard sharding
+            return torch.cat(gathered_tensors, dim=1)
+
+        if self.ring_attn_func is RingAttnFunc.BATCH_ZIGZAG:
+            # Each rank has a pattern of (rank, world_size*2-rank-1)
+            reconstituted_tensors = [None] * (world_size * 2)
+
+            # First, split each gathered tensor into its two chunks
+            for rank, gathered_tensor in enumerate(gathered_tensors):
+                # Each tensor contains two chunks in the sequence dimension
+                chunk_size = gathered_tensor.size(1) // 2
+                chunk1, chunk2 = gathered_tensor.split(chunk_size, dim=1)
+
+                # Place chunks in their original positions
+                reconstituted_tensors[rank] = chunk1
+                reconstituted_tensors[world_size * 2 - rank - 1] = chunk2
+
+            # Concatenate the reconstituted tensors in the correct order
+            return torch.cat(reconstituted_tensors, dim=1)
+
+        # Otherwise, RingAttnFunc.BATCH_STRIPE
+        # In striping, each rank has every world_size-th slice
+        batch_size = tensor.size(0)
+        hidden_dim = tensor.size(-1)
+
+        # First, determine the full sequence length
+        total_seq_len = 0
+        for t in gathered_tensors:
+            total_seq_len += t.size(1)
+
+        # Create a tensor to hold the unstriped result
+        result = torch.zeros(
+            batch_size,
+            total_seq_len,
+            hidden_dim,
+            dtype=tensor.dtype,
+            device=tensor.device,
+        )
+
+        # For each rank's tensor, distribute its slices to the correct positions
+        for rank, gathered_tensor in enumerate(gathered_tensors):
+            # The rank's tensor contains every world_size-th slice
+            # starting from its rank position
+            seq_len = gathered_tensor.size(1)
+            for i in range(seq_len):
+                # Calculate the position in the full tensor
+                pos = i * world_size + rank
+                if pos < total_seq_len:
+                    result[:, pos] = gathered_tensor[:, i]
+
+        return result

src/axolotl/core/training_args.py

@@ -9,7 +9,7 @@ from PIL.Image import Resampling
 from transformers import TrainingArguments
 from trl import CPOConfig, KTOConfig, ORPOConfig, PRMConfig, RewardConfig
 
-from axolotl.utils.schemas.enums import RingAttnFunc
+from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
 
 
 @dataclass

src/axolotl/integrations/base.py

@@ -26,8 +26,6 @@ from typing import OrderedDict
 import torch
 from torch.optim.lr_scheduler import LRScheduler
 
-from axolotl.utils.dict import DictDefault
-
 
 class BasePlugin:
     """
@@ -38,13 +36,11 @@ 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.
@@ -67,32 +63,20 @@ class BasePlugin:
         None
         """
 
-    def get_input_args(self) -> str | None:
+    def get_input_args(self):
         """
         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.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
 
         Returns:
-            None
+        None
         """
 
     def post_model_build(self, cfg, model):  # pylint: disable=unused-argument
@@ -107,71 +91,59 @@ class BasePlugin:
         """
         Performs actions after the model is loaded.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            model (object): The loaded model.
+        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.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            model (object): The loaded model.
+        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.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            model (object): The loaded model.
+        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.
 
-        Args:
+        Parameters:
-            cfg (dict): The global axolotl configuration.
+        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.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            trainer (object): The trainer object for training.
+        trainer (object): The trainer object for training.
 
         Returns:
-            object: The created optimizer.
+        object: The created optimizer.
         """
 
     def create_lr_scheduler(
@@ -180,26 +152,26 @@ class BasePlugin:
         """
         Creates and returns a learning rate scheduler.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            trainer (object): The trainer object for training.
+        trainer (object): The trainer object for training.
-            optimizer (object): The optimizer for training.
+        optimizer (object): The optimizer for training.
-            num_training_steps (int): Total number of training steps
+        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.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            model (object): The loaded model.
+        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 []
 
@@ -210,12 +182,12 @@ class BasePlugin:
         Adds callbacks to the trainer after creating the trainer.
         This is useful for callbacks that require access to the model or trainer.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
-            trainer (object): The trainer object for training.
+        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 []
 
@@ -223,23 +195,23 @@ class BasePlugin:
         """
         Performs actions after training is complete.
 
-        Args:
+        Parameters:
-            cfg (dict): The axolotl configuration
+        cfg (dict): The axolotl configuration
-            model (object): The loaded model.
+        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.
 
-        Args:
+        Parameters:
-            cfg (dict): The configuration for the plugin.
+        cfg (dict): The configuration for the plugin.
 
         Returns:
-            None
+        None
         """
 
 
@@ -366,27 +338,6 @@ 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.
@@ -471,20 +422,6 @@ 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/cut_cross_entropy/__init__.py

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

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

@@ -4,6 +4,7 @@
 # flake8: noqa
 
 from .patch import (
+    RingAttnFunc,
     get_ring_attn_group,
     register_ring_attn,
     set_ring_attn_group,

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

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

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

@@ -6,12 +6,13 @@ package, specifically the `hf_adapter.substitute_hf_flash_attn` function to patc
 their sequence parallel version of Flash Attention 2.
 """
 
+from enum import Enum
+
 import torch
 import torch.distributed as dist
 from accelerate.logging import get_logger
 
 from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
-from axolotl.utils.schemas.enums import RingAttnFunc
 
 LOG = get_logger(__name__)
 
@@ -40,6 +41,17 @@ def set_ring_attn_group(ring_attn_group: dist.ProcessGroup | None):
     RING_ATTN_GROUP = ring_attn_group
 
 
+class RingAttnFunc(str, Enum):
+    """Enum class for supported `ring-flash-attn` implementations"""
+
+    # VARLEN_RING = "varlen_ring"
+    # VARLEN_ZIGZAG = "varlen_zigzag"
+    VARLEN_LLAMA3 = "varlen_llama3"
+    BATCH_RING = "batch_ring"
+    BATCH_ZIGZAG = "batch_zigzag"
+    BATCH_STRIPE = "batch_stripe"
+
+
 def register_ring_attn(
     sequence_parallel_degree: int,
     heads_k_stride: int | None,
@@ -105,7 +117,11 @@ def register_ring_attn(
         substitute_hf_flash_attn(
             process_group=get_ring_attn_group(), heads_k_stride=heads_k_stride or 1
         )
-    elif ring_attn_func is RingAttnFunc.BATCH_RING:
+    elif ring_attn_func in [
+        RingAttnFunc.BATCH_RING,
+        RingAttnFunc.BATCH_ZIGZAG,
+        RingAttnFunc.BATCH_STRIPE,
+    ]:
         from axolotl.monkeypatch.attention.ring_attn.adapters.batch import (
             substitute_hf_flash_attn,
         )

src/axolotl/monkeypatch/loss/chunked.py

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

src/axolotl/monkeypatch/peft/utils.py

@@ -24,7 +24,7 @@ PATCHED_PREPARE_CODE = """
         for name, param in model.named_parameters():
             if (
                 (param.dtype == torch.float16) or (param.dtype == torch.bfloat16)
-            ) and param.__class__.__name__ != "Params4bit" and all(embed_name not in name for embed_name in ["embed_tokens", "lm_head"]):
+            ) and param.__class__.__name__ != "Params4bit" and "norm" in name:
                 param.data = param.data.to(torch.float32)
 """
 

src/axolotl/train.py

@@ -2,17 +2,17 @@
 
 import importlib
 import inspect
-import logging
 import os
 import signal
 import sys
 import weakref
-from contextlib import ExitStack
+from contextlib import nullcontext
 from pathlib import Path
 from typing import Any, Dict
 
 import torch
 import transformers.modelcard
+from accelerate.logging import get_logger
 from accelerate.utils import save_fsdp_model
 from datasets import Dataset
 from huggingface_hub.errors import OfflineModeIsEnabled
@@ -27,13 +27,14 @@ from axolotl.contribs.lgpl import (  # pylint: disable = no-name-in-module
     fix_untrained_tokens,
 )
 from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
+from axolotl.core.trainers.mixins.sequence_parallel import (
+    SequenceParallelContextManager,
+)
 from axolotl.integrations.base import PluginManager
-from axolotl.utils.ctx_managers.sequence_parallel import SequenceParallelContextManager
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import cleanup_distributed
 from axolotl.utils.freeze import freeze_layers_except
 from axolotl.utils.models import load_model, load_processor, load_tokenizer
-from axolotl.utils.schemas.enums import RLType
 from axolotl.utils.trainer import setup_trainer
 
 try:
@@ -41,7 +42,7 @@ try:
 except ImportError:
     BetterTransformer = None
 
-LOG = logging.getLogger(__name__)
+LOG = get_logger(__name__)
 
 
 def setup_model_and_tokenizer(
@@ -62,6 +63,7 @@ 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)
 
@@ -106,7 +108,7 @@ def setup_reference_model(
         Reference model if needed for RL training, `None` otherwise.
     """
     model_ref = None
-    if cfg.rl and cfg.rl != RLType.ORPO:
+    if cfg.rl and cfg.rl != "orpo":
         if cfg.adapter and not cfg.rl_adapter_ref_model:
             # use built-in trl autounwrap
             LOG.debug("Passing model_ref: None to RL trainer")
@@ -187,32 +189,28 @@ def execute_training(
         trainer: The configured trainer object.
         resume_from_checkpoint: Path to checkpoint to resume from, if applicable.
     """
-    with ExitStack() as stack:
+    # Define the context managers to use
-        # Define the context managers to use
+    flash_context = (
-        if cfg.flash_optimum:
+        torch.backends.cuda.sdp_kernel(
-            stack.enter_context(
+            enable_flash=True,
-                torch.backends.cuda.sdp_kernel(
+            enable_math=True,
-                    enable_flash=True,
+            enable_mem_efficient=True,
-                    enable_math=True,
+        )
-                    enable_mem_efficient=True,
+        if cfg.flash_optimum
-                )
+        else nullcontext()
-            )
+    )
+    sequence_parallel_context = (
+        SequenceParallelContextManager(
+            model=trainer.model,
+            sequence_parallel_degree=cfg.sequence_parallel_degree,
+            ring_attn_func=cfg.ring_attn_func,
+        )
+        if cfg.sequence_parallel_degree > 1
+        else nullcontext()
+    )
 
-        if cfg.sequence_parallel_degree > 1:
+    LOG.info("Starting trainer...")
-            models = [trainer.model]
+    with flash_context, sequence_parallel_context:
-            if hasattr(trainer, "ref_model"):
-                models.append(trainer.ref_model)
-
-            stack.enter_context(
-                SequenceParallelContextManager(
-                    models=models,
-                    sequence_parallel_degree=cfg.sequence_parallel_degree,
-                    gradient_accumulation_steps=cfg.gradient_accumulation_steps,
-                    ring_attn_func=cfg.ring_attn_func,
-                )
-            )
-
-        LOG.info("Starting trainer...")
         trainer.train(resume_from_checkpoint=resume_from_checkpoint)
 
 
@@ -530,9 +528,6 @@ 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
@@ -555,6 +550,7 @@ 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

@@ -868,28 +868,3 @@ 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/ctx_managers/__init__.py

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

src/axolotl/utils/ctx_managers/sequence_parallel.py

@@ -1,335 +0,0 @@
-"""Module for Axolotl trainer sequence parallelism manager and utilities"""
-
-import functools
-
-import torch
-import torch.distributed as dist
-from torch import nn
-from torch.utils.hooks import RemovableHandle
-from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.utils import ModelOutput
-
-from axolotl.monkeypatch.attention.ring_attn.patch import (
-    get_ring_attn_group,
-    update_ring_attn_params,
-)
-from axolotl.utils.schemas.enums import RingAttnFunc
-
-
-# TODO(djsaunde): implement zigzag, stripe patterns here (and elsewhere) in this
-# module. Currently, we just focus on batch ring and varlen llama3 for simplicity.
-def apply_sequence_parallelism(
-    batch: dict[str, torch.Tensor],
-    local_rank: int,
-    local_world_size: int,
-    gradient_accumulation_steps: int,
-    ring_attn_func: RingAttnFunc,  # pylint: disable=unused-argument
-) -> tuple[dict[str, torch.Tensor], int, int]:
-    """
-    Apply sequence parallelism slicing to a batch.
-
-    Special handling is implemented for integer logits_to_keep, which indicates
-    to only keep the last N tokens in the sequence during generation.
-
-    Args:
-        batch: Batch dictionary (e.g., input_ids, attention_mask, etc.).
-        local_rank: Local rank in the sequence parallel group.
-        local_world_size: World size of the sequence parallel group.
-        gradient_accumulation_steps: Number of steps to accumulate gradients over.
-        ring_attn_func: Which ring attention function to use. Currently unused, but
-            related to above TODO.
-
-    Returns:
-        tuple of:
-            - Batch dictionary with sliced tensors.
-            - The original sequence length before padding.
-            - The number of padding tokens added.
-    """
-    original_seq_len = batch["input_ids"].size(1)
-
-    # Update ring attention params if needed
-    if batch.get("position_ids") is not None:
-        update_ring_attn_params(position_ids=batch["position_ids"])
-    else:
-        # If position_ids aren't already in the batch, create them
-        batch["position_ids"] = torch.arange(
-            0,
-            original_seq_len,
-            dtype=torch.long,
-            device=batch["input_ids"].device,
-        ).expand(batch["input_ids"].size(0), -1)
-
-    if "logits_to_keep" in batch and isinstance(batch["logits_to_keep"], int):
-        logits_to_keep = batch["logits_to_keep"]
-
-        # Calculate which positions in the full sequence contain the last N tokens
-        start_position = max(0, original_seq_len - logits_to_keep)
-        chunk_size = original_seq_len // local_world_size
-        rank_start = local_rank * chunk_size
-        rank_end = rank_start + chunk_size
-
-        # Create a boolean mask tensor for this rank's chunk
-        mask = torch.zeros(
-            chunk_size,
-            dtype=torch.bool,
-            device=batch["input_ids"].device,
-        )
-
-        if rank_end > start_position:
-            # Calculate how many of the last N tokens fall within this rank's range
-            tokens_in_rank = min(rank_end, original_seq_len) - max(
-                rank_start, start_position
-            )
-
-            # Calculate where these tokens start in the local chunk
-            local_start_idx = max(0, start_position - rank_start)
-
-            # Set the appropriate positions in the mask to True
-            mask[local_start_idx : local_start_idx + tokens_in_rank] = True
-
-        # Replace the integer with the boolean mask
-        batch["logits_to_keep"] = mask
-
-    # Add padding to make sequence length divisible by local_world_size
-    total_seq_len = original_seq_len
-    pad_len = 0
-    divisor = min(local_world_size, 64)
-    if total_seq_len % divisor != 0:
-        pad_len = divisor - (total_seq_len % divisor)
-
-        # Apply padding to all relevant tensors
-        for key in batch:
-            if (
-                isinstance(batch[key], torch.Tensor)
-                and batch[key].dim() > 1
-                and batch[key].size(1) == total_seq_len
-            ):
-                # Create padding tensor
-                pad_value = -100 if key == "labels" else 0
-                padding = torch.full(
-                    (batch[key].size(0), pad_len, *batch[key].shape[2:]),
-                    pad_value,
-                    dtype=batch[key].dtype,
-                    device=batch[key].device,
-                )
-
-                # Concatenate padding to the right side of the tensor
-                batch[key] = torch.cat([batch[key], padding], dim=1)
-            if key == "logits_to_keep":
-                # Create padding tensor
-                padding = torch.ones(
-                    1,
-                    dtype=batch[key].dtype,
-                    device=batch[key].device,
-                )
-
-                # Concatenate padding to the right side of the tensor
-                batch[key] = torch.cat([batch[key], padding], dim=0)
-
-        # Update the total sequence length after padding
-        total_seq_len = batch["input_ids"].size(1)
-
-    # Slice batch for sequence parallel
-    for key in batch:
-        if not isinstance(batch[key], torch.Tensor) or batch[key].dim() <= 1:
-            continue
-
-        # Split in sequential fashion and grab this rank's chunk
-        if batch[key].size(1) == total_seq_len:
-            batch[key] = (
-                batch[key].chunk(local_world_size, dim=1)[local_rank].contiguous()
-            )
-        elif key == "logits_to_keep":
-            batch[key] = (
-                batch[key].chunk(local_world_size, dim=0)[local_rank].contiguous()
-            )
-
-        # Handle num_items_in_batch
-        if "num_items_in_batch" in batch:
-            # Approximation; this needed since num_items_in_batch may be counted across
-            # all samples in a gradient accumulated batch, not on a per-step basis.
-            batch["num_items_in_batch"] = (
-                batch["labels"] != -100
-            ).sum() * gradient_accumulation_steps
-
-    return batch, original_seq_len, pad_len
-
-
-class SequenceParallelContextManager:
-    """Context manager for sequence parallelism operations.
-
-    This class provides a context that will automatically apply sequence parallelism
-    during model forward passes using a pre-forward hook, and gather outputs from
-    across the sequence parallelism group using a post-forward hook.
-
-    Args:
-        models: List of models to apply sequence parallelism to pre- and post- forward
-            hooks.
-        sequence_parallel_degree: Number of processes to split sequences over.
-        gradient_accumulation_steps: Number of steps to accumulate gradients over.
-        ring_attn_func: Which ring attention function to use. Currently unused.
-    """
-
-    def __init__(
-        self,
-        models: list[nn.Module],
-        sequence_parallel_degree: int,
-        gradient_accumulation_steps: int,
-        ring_attn_func: RingAttnFunc,
-    ):
-        self.models = models
-        self.sequence_parallel_degree = sequence_parallel_degree
-        self.gradient_accumulation_steps = gradient_accumulation_steps
-        self.ring_attn_func = ring_attn_func
-        self.process_group = get_ring_attn_group()
-
-        # Initialize sequence parallel group details
-        self.local_rank = dist.get_rank(self.process_group)
-        self.local_world_size = dist.get_world_size(self.process_group)
-
-        # Will store hook handles for removal
-        self.hook_handles: list[RemovableHandle] = []
-
-        # Store original sequence length and padding information
-        self.original_seq_len = 0
-        self.pad_len = 0
-
-        # Create a partially applied version of the apply_sequence_parallelism function
-        self.apply_sequence_parallelism = functools.partial(
-            apply_sequence_parallelism,
-            local_rank=self.local_rank,
-            local_world_size=self.local_world_size,
-            gradient_accumulation_steps=self.gradient_accumulation_steps,
-            ring_attn_func=self.ring_attn_func,
-        )
-
-    def __enter__(self):
-        # Forward pre-hook to apply sequence parallelism
-        def sequence_parallel_pre_hook(_, args, kwargs):
-            # Apply sequence parallelism to kwargs and get original sequence length and padding info
-            kwargs, self.original_seq_len, self.pad_len = (
-                self.apply_sequence_parallelism(batch=kwargs)
-            )
-
-            return args, kwargs
-
-        # Forward post-hook to gather outputs
-        def sequence_parallel_post_hook(_, __, output: ModelOutput) -> ModelOutput:
-            # Gather the sharded outputs
-            output = self.gather_outputs(output)
-
-            # Remove padding if it was added
-            if self.pad_len > 0:
-                for key, value in output.items():
-                    if isinstance(value, torch.Tensor) and value.dim() > 1:
-                        if value.size(1) == self.original_seq_len + self.pad_len:
-                            # Slice to remove padding
-                            output[key] = value[:, : self.original_seq_len].contiguous()
-
-            return output
-
-        # Register both hooks
-        for model in self.models:
-            self.hook_handles.append(
-                model.register_forward_pre_hook(
-                    sequence_parallel_pre_hook, with_kwargs=True
-                )
-            )
-            self.hook_handles.append(
-                model.register_forward_hook(sequence_parallel_post_hook)
-            )
-
-        return self
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        # Remove all hooks
-        for handle in self.hook_handles:
-            handle.remove()
-        self.hook_handles = []
-
-    def gather_outputs(self, output: CausalLMOutputWithPast) -> CausalLMOutputWithPast:
-        """Gather sharded outputs from all ranks and reconstruct the full tensor."""
-        for key, value in output.items():
-            if isinstance(value, torch.Tensor) and value.dim() > 1:
-                output[key] = AllGatherWithGrad.apply(value, self.process_group)
-
-        return output
-
-
-class AllGatherWithGrad(torch.autograd.Function):
-    """Custom autograd function for all-gather to preserve gradients."""
-
-    @staticmethod
-    def forward(
-        ctx: torch.autograd.function.FunctionCtx,
-        input_tensor: torch.Tensor,
-        group: dist.ProcessGroup,
-    ) -> torch.Tensor:
-        """
-        Forward pass of all-gather of data with sequence dimension.
-
-        Args:
-            ctx: `torch.autograd` function context.
-            input_tensor: Tensor from model output with sequence dimension.
-            group: `torch.distributed` process group.
-
-        Returns:
-            Tensor from gathering the `input_tensor` from across the process group and
-                concatenating along the sequence dimension.
-        """
-        ctx.group = group
-        ctx.rank = dist.get_rank(group)
-        world_size = dist.get_world_size(group)
-
-        # Gather shape metadata
-        local_shape = torch.tensor(list(input_tensor.shape), device=input_tensor.device)
-        all_shapes = [torch.zeros_like(local_shape) for _ in range(world_size)]
-        dist.all_gather(all_shapes, local_shape, group=group)
-
-        # Store sequence lengths for backward pass
-        seq_lens = [int(shape[1].item()) for shape in all_shapes]
-        ctx.seq_lens = seq_lens
-
-        # Perform all_gather operation
-        gathered = [
-            torch.zeros(
-                tuple(shape.tolist()),
-                dtype=input_tensor.dtype,
-                device=input_tensor.device,
-            )
-            for shape in all_shapes
-        ]
-        dist.all_gather(gathered, input_tensor, group=group)
-
-        # Concatenate tensors along sequence dimension
-        result = torch.cat(gathered, dim=1)
-
-        return result
-
-    @staticmethod
-    def backward(
-        ctx: torch.autograd.function.FunctionCtx, grad_output: torch.Tensor
-    ) -> tuple[torch.Tensor, None]:
-        """
-        Backward pass for all-gather operation.
-
-        Extracts the gradient slice corresponding to this rank's original input
-        from the full gradient tensor.
-
-        Args:
-            ctx: `torch.autograd` function context.
-            grad_output: Gradient from subsequent layers with respect to the
-                concatenated output tensor.
-
-        Returns:
-            Tuple containing the gradient slice for this rank's input tensor and `None`
-                for the process group parameter which doesn't require gradients.
-        """
-        rank = ctx.rank
-        seq_lens = ctx.seq_lens
-
-        # Extract gradient for this rank's chunk
-        offset = sum(seq_lens[:rank])
-        grad_slice = grad_output[:, offset : offset + seq_lens[rank]].contiguous()
-
-        return grad_slice, None

src/axolotl/utils/data/rl.py

@@ -18,9 +18,8 @@ from axolotl.utils.data.utils import deduplicate_and_log_datasets, md5
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process, zero_first
 from axolotl.utils.models import load_tokenizer
-from axolotl.utils.schemas.enums import RLType
 
-LOG = logging.getLogger(__name__)
+LOG = logging.getLogger("axolotl")
 
 
 def _get_path(ds_hash, cfg):
@@ -81,7 +80,7 @@ def map_dataset(cfg, data_set, ds_transform_fn, tokenizer, **map_kwargs):
 def drop_long_rl_seq(
     sample, rl, tokenizer, sequence_len  # pylint: disable=invalid-name
 ):
-    if rl in (RLType.DPO, RLType.IPO, RLType.ORPO, RLType.SIMPO):
+    if rl in ("dpo", "ipo", "orpo", "simpo"):
         if not (
             sample.get("prompt") and sample.get("chosen") and sample.get("rejected")
         ):
@@ -101,7 +100,7 @@ def drop_long_rl_seq(
             len_prompt + len_rejected
         ) <= sequence_len
 
-    if rl is RLType.KTO:
+    if rl == "kto":
         if not (sample.get("prompt") and sample.get("completion")):
             raise ValueError("Prompt and completion keys are required for KTO datasets")
 
@@ -115,7 +114,7 @@ def drop_long_rl_seq(
 
         return (len_prompt + len_completion) <= sequence_len
 
-    if rl is RLType.GRPO:
+    if rl == "grpo":
         return True
 
     raise ValueError("Unknown RL type")
@@ -138,9 +137,9 @@ def load_prepare_preference_datasets(cfg):
             if _type:
                 if isinstance(_type, DictDefault):
                     _type = "user_defined.default"
-                if _cfg.rl is RLType.ORPO:
+                if _cfg.rl == "orpo":
                     ds_transform_fn = load_orpo(_type, _cfg, dataset_idx=i)
-                elif _cfg.rl is RLType.KTO:
+                elif _cfg.rl == "kto":
                     ds_transform_fn = load_kto(_type, _cfg, dataset_idx=i)
                 else:
                     ds_transform_fn = load_dpo(_type, _cfg, dataset_idx=i)
@@ -151,7 +150,7 @@ def load_prepare_preference_datasets(cfg):
                 split_datasets[i] = map_dataset(
                     cfg, data_set, ds_transform_fn, tokenizer, **map_kwargs
                 )
-            elif _cfg.rl is RLType.KTO:
+            elif _cfg.rl == "kto":
                 ds_transform_fn = load_kto(_type, _cfg, dataset_idx=i)
                 map_kwargs = {}
                 if isinstance(ds_transform_fn, tuple):
@@ -186,7 +185,7 @@ def load_prepare_preference_datasets(cfg):
                     )
 
         combined_datasets = concatenate_datasets(split_datasets)
-        combined_datasets = combined_datasets.shuffle(seed=cfg.seed or 42)
+        combined_datasets = combined_datasets.shuffle(seed=cfg.seed)
 
         return combined_datasets
 
@@ -206,8 +205,6 @@ def load_prepare_preference_datasets(cfg):
                 eval_dataset = load_split(cfg.test_datasets, cfg)
         if not eval_dataset:
             if cfg.val_set_size:
-                seed = cfg.seed if cfg.seed is not None else 42
-
                 # ensure we end up with the same fingerprint by doing rank0 first and being able to cache
                 to_hash_train = (
                     train_dataset._fingerprint  # pylint: disable=protected-access
@@ -216,7 +213,7 @@ def load_prepare_preference_datasets(cfg):
                     + "|"
                     + "train"
                     + "|"
-                    + str(seed)
+                    + str(cfg.seed or 42)
                 )
                 to_hash_test = (
                     train_dataset._fingerprint  # pylint: disable=protected-access
@@ -225,13 +222,13 @@ def load_prepare_preference_datasets(cfg):
                     + "|"
                     + "test"
                     + "|"
-                    + str(seed)
+                    + str(cfg.seed or 42)
                 )
                 train_fingerprint = md5(to_hash_train)
                 test_fingerprint = md5(to_hash_test)
                 ds_w_test_split = train_dataset.train_test_split(
                     test_size=cfg.val_set_size,
-                    seed=seed,
+                    seed=cfg.seed,
                     shuffle=False,
                     train_new_fingerprint=train_fingerprint,
                     test_new_fingerprint=test_fingerprint,

src/axolotl/utils/data/sft.py

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

src/axolotl/utils/data/shared.py

@@ -281,10 +281,6 @@ def load_dataset_w_config(
             **load_ds_kwargs,
         )
     if not ds:
-        raise ValueError(
+        raise ValueError("unhandled dataset load")
-            "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

@@ -1,36 +1,15 @@
 """custom checkpointing utils"""
 
-import importlib
 from functools import partial
 
-from packaging import version
-
 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
-    if uses_gc_layers(decoder_layer):
-        return Unsloth_Offloaded_Gradient_Checkpointer.apply(
-            decoder_layer,
-            *args,
-        )
-
     return Unsloth_Offloaded_Gradient_Checkpointer.apply(
         (
             decoder_layer.func.__self__

src/axolotl/utils/models.py

@@ -73,7 +73,6 @@ from axolotl.utils.distributed import (
 from axolotl.utils.gradient_checkpointing import 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
-from axolotl.utils.schemas.enums import RLType
 
 LOG = logging.getLogger(__name__)
 PLUGIN_MANAGER = PluginManager.get_instance()
@@ -562,12 +561,21 @@ class ModelLoader:
 
             patch_xformers_attn_over_fa2()
             self.cfg.flash_attention = True
+
+        if self.cfg.chunked_cross_entropy:
+            from axolotl.monkeypatch.loss.chunked import patch_chunked_ce_loss_fn
+
+            if self.cfg.chunked_cross_entropy_num_chunks:
+                patch_chunked_ce_loss_fn(self.cfg.chunked_cross_entropy_num_chunks)
+            else:
+                patch_chunked_ce_loss_fn()
+
         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:
+        if self.cfg.adapter:
             from axolotl.monkeypatch.peft.utils import patch_peft_prep_code
 
             patch_peft_prep_code()
@@ -904,7 +912,7 @@ class ModelLoader:
                 "bnb_4bit_compute_dtype": self.cfg.torch_dtype,
                 "bnb_4bit_use_double_quant": True,
                 "bnb_4bit_quant_type": "nf4",
-                "bnb_4bit_quant_storage": torch.bfloat16,
+                "bnb_4bit_quant_storage": torch.uint8,
             }
             if self.cfg.model_config_type in ["jamba", "qwen2_moe"] and not (
                 self.cfg.deepspeed or self.cfg.fsdp
@@ -1320,11 +1328,8 @@ 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:
+        if self.cfg.fsdp:
-            # we don't run this during FSDP because this will leave mixed
+            # FSDP doesn't like mixed Float and BFloat16
-            # 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,
@@ -1373,7 +1378,7 @@ class ModelLoader:
             # then the dpo trainer doesn't want the peft model loaded over it, it just wants the lora/peft config
             if (
                 self.cfg.adapter
-                and self.cfg.rl in [RLType.DPO, RLType.IPO, RLType.KTO]
+                and self.cfg.rl in ["dpo", "ipo", "kto"]
                 and not self.cfg.merge_lora
             ):
                 _, lora_config = load_lora(

src/axolotl/utils/samplers/multipack.py

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

src/axolotl/utils/schemas/config.py

@@ -27,7 +27,7 @@ from axolotl.utils.schemas.datasets import (
     StepwiseSupervisedDataset,
 )
 from axolotl.utils.schemas.deprecated import DeprecatedParameters, RemappedParameters
-from axolotl.utils.schemas.enums import ChatTemplate, RingAttnFunc, RLType
+from axolotl.utils.schemas.enums import ChatTemplate, RLType
 from axolotl.utils.schemas.integrations import (
     CometConfig,
     GradioConfig,
@@ -82,8 +82,6 @@ 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
-    random_init_weights: bool | None = None
 
     rl: RLType | None = None
     trl: TRLConfig | None = Field(
@@ -244,6 +242,9 @@ class AxolotlInputConfig(
     unsloth_rms_norm: bool | None = None
     unsloth_rope: bool | None = None
 
+    chunked_cross_entropy: bool | None = None
+    chunked_cross_entropy_num_chunks: int | None = None
+
     lora_mlp_kernel: bool | None = None
     lora_qkv_kernel: bool | None = None
     lora_o_kernel: bool | None = None
@@ -261,7 +262,7 @@ class AxolotlInputConfig(
 
     sequence_parallel_degree: int | None = None
     heads_k_stride: int | None = None
-    ring_attn_func: RingAttnFunc | None = None
+    ring_attn_func: str | None = None
 
     special_tokens: SpecialTokensConfig | None = None
     tokens: list[str] | None = None
@@ -463,10 +464,9 @@ 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, xformers or flex attention does not handle cross sample decontamination."
+                "sample_packing without flash, sdp or flex attention does not handle cross sample decontamination."
             )
 
         return data
@@ -783,7 +783,7 @@ class AxolotlInputConfig(
 
     @model_validator(mode="after")
     def check_simpo_warmup(self):
-        if self.rl is RLType.SIMPO and self.warmup_ratio:
+        if self.rl == "simpo" and self.warmup_ratio:
             raise ValueError(
                 "warmup_ratio is not supported with the simpo trainer. Please use `warmup_steps` instead"
             )
@@ -1150,28 +1150,16 @@ 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_liger_sequence_parallel(cls, data):
+    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("sequence_parallel_degree", 1) > 1
+            and data.get("adapter")
         ):
-            raise ValueError("GRPO + SP + Liger not currently supported")
+            raise ValueError("PEFT + GRPO + Liger is not yet supported")
         return data
 
     @model_validator(mode="after")
@@ -1186,7 +1174,7 @@ class AxolotlInputConfig(
 
             if self.sample_packing and self.micro_batch_size > 1:
                 raise ValueError(
-                    "micro_batch_size must be set to 1 when sample_packing is enabled "
+                    "micro_batch_size must be set to 1 when sample_packing is enabled"
                     "due to a `ring-flash-attn` requirement"
                 )
 
@@ -1218,8 +1206,16 @@ class AxolotlInputConfig(
         if getattr(self, "sequence_parallel_degree", 1) == 1:
             return self
 
+        from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
+
         if self.ring_attn_func is not None:
-            self.ring_attn_func = RingAttnFunc(self.ring_attn_func)
+            valid_funcs = list(RingAttnFunc)
+            if self.ring_attn_func in valid_funcs:
+                self.ring_attn_func = RingAttnFunc(self.ring_attn_func)
+            else:
+                raise ValueError(
+                    f"ring_attn_func: {self.ring_attn_func} must be in {valid_funcs}"
+                )
         else:
             # Default ring attention function selection
             sample_packing = getattr(self, "sample_packing", False)
@@ -1350,10 +1346,6 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
             ):
                 return data
 
-            # Skip if dropout is not 0, as auto enabling it would just disable it during runtime patch checks
-            if data.get("lora_dropout") != 0:
-                return data
-
             # Check multi-GPU compatibility
             capabilities = data.get("capabilities")
             is_multi_gpu = capabilities and capabilities.get("n_gpu", 0) > 1

src/axolotl/utils/schemas/enums.py

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

src/axolotl/utils/schemas/training.py

@@ -75,10 +75,8 @@ 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

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

tests/e2e/integrations/test_hooks.py

@@ -29,12 +29,6 @@ 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"
@@ -171,7 +165,6 @@ 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", 1.2, "Train Loss (%s) is too high"
+            temp_dir + "/runs", "train/loss", 1.0, "Train Loss 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", 1.2, "Train Loss (%s) is too high"
+            temp_dir + "/runs", "train/loss", 1.0, "Train Loss is too high"
         )

tests/e2e/multigpu/patched/test_sp.py

@@ -25,7 +25,6 @@ class TestSequenceParallelism:
         micro_batch_size=1,
         pad_to_sequence_len=True,
         ring_attn_func=None,
-        threshold=2.0,
     ):
         """Helper method to run sequence parallel tests with different configurations"""
         cfg = DictDefault(
@@ -94,22 +93,22 @@ class TestSequenceParallelism:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", threshold, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.6, "Train Loss is too high"
         )
 
     @pytest.mark.parametrize(
-        "sample_packing, micro_batch_size, pad_to_sequence_len, ring_attn_func, threshold",
+        "sample_packing, micro_batch_size, pad_to_sequence_len, ring_attn_func",
         [
-            (True, 1, True, None, 2.5),  # defaults to varlen_llama3 ring_attn_func
+            (True, 1, True, None),  # defaults to varlen_llama3 ring_attn_func
-            (False, 2, True, None, 2.5),  # defaults to batch_ring ring_attn_func
+            (False, 2, True, None),  # defaults to batch_ring ring_attn_func
-            # (False, 2, True, "batch_zigzag", 2.5),
+            (False, 2, True, "batch_zigzag"),
-            (False, 2, False, None, 2.5),  # defaults to batch_ring ring_attn_func
+            # (False, 2, False),  # not yet working
         ],
         ids=[
             "sample_packing, varlen_llama3 ring_attn_func",
-            "no sample_packing, pad_to_sequence_len, batch_ring ring_attn_func",
-            # "no sample_packing, no pad_to_sequence_len, batch_zigzag ring_attn_func",
             "no sample_packing, no pad_to_sequence_len, batch_ring ring_attn_func",
+            "no sample_packing, no pad_to_sequence_len, batch_zigzag ring_attn_func",
+            # "no sample_packing, pad_to_sequence_len",  # not yet working
         ],
     )
     def test_sequence_parallel_training(
@@ -119,7 +118,6 @@ class TestSequenceParallelism:
         micro_batch_size,
         pad_to_sequence_len,
         ring_attn_func,
-        threshold,
     ):
         """Test sequence parallel training with different configurations"""
         self._run_sequence_parallel_test(
@@ -128,5 +126,4 @@ class TestSequenceParallelism:
             micro_batch_size=micro_batch_size,
             pad_to_sequence_len=pad_to_sequence_len,
             ring_attn_func=ring_attn_func,
-            threshold=threshold,
         )

tests/e2e/multigpu/solo/test_grpo.py

@@ -166,7 +166,6 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
 """
             )
 
-    @pytest.mark.skip(reason="flaky test")
     @pytest.mark.parametrize(
         "num_gpus",
         [1, 2],
@@ -228,7 +227,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
 
         current_env = os.environ.copy()
         env = {
-            "NCCL_P2P_LEVEL": "LOC",
+            "NCCL_P2P_LEVEL": "NVL",
             **current_env,
             "CUDA_VISIBLE_DEVICES": "1",
             "VLLM_DISABLE_COMPILE_CACHE": "1",
@@ -258,7 +257,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
                     f"{get_torch_dist_unique_port()}",
                 ],
                 env={
-                    "NCCL_P2P_LEVEL": "LOC",
+                    "NCCL_P2P_LEVEL": "NVL",
                     "NCCL_DEBUG": "INFO",
                     **current_env,
                 },
@@ -266,7 +265,6 @@ 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],
@@ -322,7 +320,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
 
         current_env = os.environ.copy()
         env = {
-            "NCCL_P2P_LEVEL": "LOC",  # nccl can be brittle, assume P2P isn't reliable
+            "NCCL_P2P_LEVEL": "NVL",  # nccl can be brittle, assume P2P isn't reliable
             **current_env,
             "CUDA_VISIBLE_DEVICES": "1",
             "VLLM_DISABLE_COMPILE_CACHE": "1",
@@ -352,7 +350,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
                     f"{get_torch_dist_unique_port()}",
                 ],
                 env={
-                    "NCCL_P2P_LEVEL": "LOC",
+                    "NCCL_P2P_LEVEL": "NVL",
                     "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": 0.1,
+                "val_set_size": 0.05,
                 "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": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+        "name": "openaccess-ai-collective/tiny-mistral",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
     {
-        "name": "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
+        "name": "Qwen/Qwen2-7B",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
@@ -44,7 +44,7 @@ MODEL_CONFIGS = [
         "dtype": torch.float32,
     },
     {
-        "name": "trl-internal-testing/tiny-Gemma2ForCausalLM",
+        "name": "mhenrichsen/gemma-2b",
         "expected_activation": apply_lora_mlp_geglu,
         "dtype": torch.float16,
     },
@@ -156,9 +156,7 @@ def test_swiglu_mlp_integration(small_llama_model):
 def test_geglu_model_integration():
     """Test GeGLU activation with Gemma model."""
     model = AutoModelForCausalLM.from_pretrained(
-        "trl-internal-testing/tiny-Gemma2ForCausalLM",
+        "mhenrichsen/gemma-2b", torch_dtype=torch.float16, device_map="cuda:0"
-        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": 5,
+                "max_steps": 20,
-                "save_steps": 3,
+                "save_steps": 10,
-                "eval_steps": 4,
+                "eval_steps": 10,
                 "fp16": True,
             }
         )
@@ -105,9 +105,9 @@ class Test4dMultipackLlama(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 20,
-                "save_steps": 3,
+                "save_steps": 10,
-                "eval_steps": 4,
+                "eval_steps": 10,
                 "fp16": True,
             }
         )

tests/e2e/patched/test_falcon_samplepack.py

@@ -6,8 +6,6 @@ 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
@@ -25,7 +23,6 @@ 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
@@ -74,7 +71,6 @@ 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": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "openaccess-ai-collective/tiny-mistral",
                 "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": 5,
+                "max_steps": 20,
-                "save_steps": 3,
+                "save_steps": 10,
-                "eval_steps": 4,
+                "eval_steps": 10,
                 "bf16": "auto",
             }
         )
@@ -76,7 +76,7 @@ class TestMistral(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "openaccess-ai-collective/tiny-mistral",
                 "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": 5,
+                "max_steps": 20,
-                "save_steps": 3,
+                "save_steps": 10,
-                "eval_steps": 4,
+                "eval_steps": 10,
                 "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": 5,
+                "max_steps": 20,
-                "save_steps": 3,
+                "save_steps": 10,
-                "eval_steps": 4,
+                "eval_steps": 10,
                 "bf16": "auto",
             }
         )
@@ -93,9 +93,9 @@ class TestMixtral(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 20,
-                "save_steps": 3,
+                "save_steps": 10,
-                "eval_steps": 4,
+                "eval_steps": 10,
                 "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": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "openaccess-ai-collective/tiny-mistral",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,

tests/e2e/patched/test_peft_embeddings.py

@@ -1,63 +0,0 @@
-"""
-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": 5,
+                "max_steps": 20,
-                "eval_steps": 3,
+                "eval_steps": 10,
-                "save_steps": 4,
+                "save_steps": 10,
                 "bf16": "auto",
             }
         )
@@ -108,9 +108,9 @@ class TestPhiMultipack(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 20,
-                "eval_steps": 3,
+                "eval_steps": 10,
-                "save_steps": 4,
+                "save_steps": 10,
                 "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, require_torch_2_6_0
+from ..utils import check_model_output_exists, most_recent_subdir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -26,7 +26,6 @@ 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(
@@ -63,7 +62,6 @@ class TestResumeLlama:
                 "save_total_limit": 5,
                 "max_steps": 15,
                 "use_tensorboard": True,
-                "save_safetensors": True,
             }
         )
         if is_torch_bf16_gpu_available():

tests/e2e/patched/test_sp.py

@@ -10,15 +10,14 @@ import pytest
 import torch
 from accelerate.state import PartialState
 
+from axolotl.core.trainers.mixins.sequence_parallel import apply_sequence_parallelism
 from axolotl.monkeypatch.attention.ring_attn import (
+    RingAttnFunc,
     get_ring_attn_group,
     register_ring_attn,
     set_ring_attn_group,
 )
-from axolotl.utils.ctx_managers.sequence_parallel import apply_sequence_parallelism
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.schemas.enums import RingAttnFunc
-from axolotl.utils.schemas.trl import TRLConfig
 
 
 @pytest.fixture
@@ -63,14 +62,12 @@ def sequence_parallel_batch():
     input_ids = torch.arange(batch_size * seq_len).reshape(batch_size, seq_len)
     attention_mask = torch.ones(batch_size, seq_len)
     position_ids = torch.arange(seq_len).expand(batch_size, seq_len)
-    labels = input_ids.clone()
 
     # Create test batch
     batch = {
         "input_ids": input_ids,
         "attention_mask": attention_mask,
         "position_ids": position_ids,
-        "labels": labels,
     }
 
     return batch
@@ -182,44 +179,12 @@ class TestConfigValidation:
                 False,
                 "micro_batch_size must be set to 1",
             ),
-            # Valid: Basic GRPO config
-            (
-                {
-                    "sequence_parallel_degree": 2,
-                    "flash_attention": True,
-                    "micro_batch_size": 2,
-                    "trl": {"use_liger_loss": True},
-                },
-                {
-                    "sequence_parallel_degree": 2,
-                    "flash_attention": True,
-                    "micro_batch_size": 2,
-                    "trl": TRLConfig(use_liger_loss=True),
-                },
-                True,
-                "GRPO + SP + Liger not currently supported",
-            ),
-            # Invalid: GRPO config with Liger loss
-            (
-                {
-                    "rl": "grpo",
-                    "sequence_parallel_degree": 2,
-                    "flash_attention": True,
-                    "micro_batch_size": 2,
-                    "trl": {"use_liger_loss": True},
-                },
-                None,
-                False,
-                "GRPO + SP + Liger not currently supported",
-            ),
         ],
         ids=[
             "valid_config",
             "default_sp_degree",
             "without_flash_attention",
             "sample_packing_with_large_batch",
-            "valid_grpo",
-            "grpo_with_liger_loss",
         ],
     )
     def test_sequence_parallel_config_validation(
@@ -291,7 +256,7 @@ class TestConfigValidation:
             AxolotlInputConfig(**cfg)
 
         # Verify error message
-        assert "Input should be 'varlen_llama3' or 'batch_ring'" in str(excinfo.value)
+        assert "ring_attn_func: INVALID_FUNC must be in" in str(excinfo.value)
 
 
 class TestApplySequenceParallelism:
@@ -325,11 +290,10 @@ class TestApplySequenceParallelism:
 
     def test_world_size_one(self, sequence_parallel_batch):
         """Test that function returns original batch when world size is 1."""
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=sequence_parallel_batch,
             local_rank=0,
             local_world_size=1,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
@@ -341,11 +305,10 @@ class TestApplySequenceParallelism:
         batch = sequence_parallel_batch
         seq_len = batch["input_ids"].size(1)
 
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=batch,
             local_rank=0,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
@@ -365,59 +328,57 @@ class TestApplySequenceParallelism:
         seq_len = batch["input_ids"].size(1)
         original_input_ids = batch["input_ids"].clone()
 
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=batch,
             local_rank=1,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
         # Verify content: rank 1 should get the second half of the sequence
         assert torch.equal(result["input_ids"], original_input_ids[:, seq_len // 2 :])
 
-    # TODO(djsaunde): add back once implemented.
+    def test_batch_zigzag(self, sequence_parallel_batch):
-    # def test_batch_zigzag(self, sequence_parallel_batch):
+        """Test BATCH_ZIGZAG sharding pattern."""
-    #     """Test BATCH_ZIGZAG sharding pattern."""
+        batch = sequence_parallel_batch
-    #     batch = sequence_parallel_batch
+        original_input_ids = batch["input_ids"].clone()
-    #     original_input_ids = batch["input_ids"].clone()
+        seq_len = batch["input_ids"].size(1)
-    #     seq_len = batch["input_ids"].size(1)
 
-    #     # Test rank 0
+        # Test rank 0
-    #     result_rank0 = apply_sequence_parallelism(
+        result_rank0 = apply_sequence_parallelism(
-    #         batch={k: v.clone() for k, v in batch.items()},
+            batch={k: v.clone() for k, v in batch.items()},
-    #         local_rank=0,
+            local_rank=0,
-    #         local_world_size=2,
+            local_world_size=2,
-    #         ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
+            ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
-    #     )
+        )
 
-    #     # Test rank 1
+        # Test rank 1
-    #     result_rank1 = apply_sequence_parallelism(
+        result_rank1 = apply_sequence_parallelism(
-    #         batch={k: v.clone() for k, v in batch.items()},
+            batch={k: v.clone() for k, v in batch.items()},
-    #         local_rank=1,
+            local_rank=1,
-    #         local_world_size=2,
+            local_world_size=2,
-    #         ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
+            ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
-    #     )
+        )
 
-    #     # Checks for both ranks
+        # Checks for both ranks
-    #     assert result_rank0["input_ids"].shape[1] == seq_len // 2
+        assert result_rank0["input_ids"].shape[1] == seq_len // 2
-    #     assert result_rank1["input_ids"].shape[1] == seq_len // 2
+        assert result_rank1["input_ids"].shape[1] == seq_len // 2
 
-    #     # For a 2-rank system with 8 tokens, check specific zigzag pattern
+        # For a 2-rank system with 8 tokens, check specific zigzag pattern
-    #     # Rank 0 should get chunks [0, 1] and [6, 7]
+        # Rank 0 should get chunks [0, 1] and [6, 7]
-    #     # Rank 1 should get chunks [2, 3] and [4, 5]
+        # Rank 1 should get chunks [2, 3] and [4, 5]
-    #     if seq_len == 8:
+        if seq_len == 8:
-    #         # Create expected tensors for comparison
+            # Create expected tensors for comparison
-    #         rank0_expected = torch.cat(
+            rank0_expected = torch.cat(
-    #             [original_input_ids[:, :2], original_input_ids[:, 6:8]], dim=1
+                [original_input_ids[:, :2], original_input_ids[:, 6:8]], dim=1
-    #         )
+            )
 
-    #         rank1_expected = torch.cat(
+            rank1_expected = torch.cat(
-    #             [original_input_ids[:, 2:4], original_input_ids[:, 4:6]], dim=1
+                [original_input_ids[:, 2:4], original_input_ids[:, 4:6]], dim=1
-    #         )
+            )
 
-    #         assert torch.equal(result_rank0["input_ids"], rank0_expected)
+            assert torch.equal(result_rank0["input_ids"], rank0_expected)
-    #         assert torch.equal(result_rank1["input_ids"], rank1_expected)
+            assert torch.equal(result_rank1["input_ids"], rank1_expected)
 
     def test_partial_application(self, sequence_parallel_batch):
         """Test that we can create a partially applied version of the function."""
@@ -429,12 +390,11 @@ class TestApplySequenceParallelism:
             apply_sequence_parallelism,
             local_rank=0,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
         # Use the partially applied function
-        result, _, _ = rank0_ring_parallel(batch=batch)
+        result = rank0_ring_parallel(batch=batch)
 
         # Verify it works as expected
         assert result["input_ids"].shape[1] == original_input_ids.shape[1] // 2
@@ -452,15 +412,13 @@ class TestApplySequenceParallelism:
         original_input_ids = batch["input_ids"].clone()
 
         # This should run without error even though position_ids is missing
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=batch,
             local_rank=0,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
         # Verification should pass
-        assert "position_ids" in result
+        assert "position_ids" not in result
-        assert result["input_ids"].shape[1] == result["position_ids"].shape[1]
         assert result["input_ids"].shape[1] == original_input_ids.shape[1] // 2

tests/e2e/test_evaluate.py

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

tests/e2e/test_falcon.py

@@ -6,8 +6,6 @@ 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
@@ -25,7 +23,6 @@ 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
@@ -77,7 +74,6 @@ 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
@@ -133,7 +129,6 @@ 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": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "openaccess-ai-collective/tiny-mistral",
                 "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": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "openaccess-ai-collective/tiny-mistral",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "val_set_size": 0.02,

tests/e2e/test_optimizers.py

@@ -199,50 +199,3 @@ 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/test_chunked_xentropy.py

@@ -0,0 +1,40 @@
+"""
+test suite for chunked cross entropy
+"""
+
+import pytest
+import torch
+from torch import nn
+
+from axolotl.monkeypatch.loss.chunked import get_causal_lm_loss
+
+
+@pytest.fixture
+def chunked_fixtures():
+    model_dim = 512
+    vocab_size = 1024 * 256
+    seq_len = 2048
+    batch_size = 1
+
+    lm_head = nn.Linear(model_dim, vocab_size)
+    hidden_state = torch.randn(batch_size, seq_len, model_dim)
+    labels = torch.randint(low=0, high=vocab_size, size=(batch_size, seq_len))
+    return lm_head, hidden_state, labels, vocab_size
+
+
+def test_chunked_forward(chunked_fixtures):  # pylint: disable=redefined-outer-name
+    lm_head, hidden_state, labels, vocab_size = chunked_fixtures
+    lm_loss = get_causal_lm_loss()
+
+    logits = lm_head(hidden_state)
+
+    chunked_lm_loss = lm_loss(logits, labels)
+
+    logits_flattened = logits.view(-1, vocab_size)
+    labels_flattened = labels.view(-1)
+
+    loss = nn.functional.cross_entropy(
+        logits_flattened.float(), labels_flattened, reduction="mean"
+    )
+
+    assert torch.allclose(chunked_lm_loss, loss, atol=1e-2, rtol=1e-2)

tests/test_datasets.py

@@ -414,6 +414,7 @@ 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

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