Add kd coefficient scheduler

.github/workflows/base.yml

@@ -40,12 +40,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-          - cuda: "128"
-            cuda_version: 12.8.1
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: nightly
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -67,7 +61,7 @@ jobs:
         uses: docker/build-push-action@v4
         with:
           context: .
-          file: ${{ matrix.pytorch == 'nightly' && './docker/Dockerfile-base-nightly' || './docker/Dockerfile-base' }}
+          file: ./docker/Dockerfile-base
           push: ${{ github.event_name != 'pull_request' }}
           tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
           labels: ${{ steps.metadata.outputs.labels }}

.github/workflows/docs.yml

@@ -20,11 +20,9 @@ jobs:
           uses: actions/setup-python@v5
           with:
             python-version: '3.11'
-        - name: Install dependencies
+        - name: install dependencies
           run: |
-            python3 -m pip install jupyter quartodoc
+            python3 -m pip install jupyter
-        - name: Build autodoc
-          run: quartodoc build
         - name: Publish to GitHub Pages (and render)
           uses: quarto-dev/quarto-actions/publish@v2
           with:

.github/workflows/precommit-autoupdate.yml

@@ -1,49 +0,0 @@
-name: Pre-commit auto-update
-
-on:
-  schedule:
-    - cron: '0 0 * * 0'  # Run weekly
-  workflow_dispatch:  # Manual kickoff
-
-jobs:
-  auto-update:
-    runs-on: ubuntu-latest
-    permissions:
-      contents: write
-      pull-requests: write
-    steps:
-      - uses: actions/checkout@v4
-
-      - uses: actions/setup-python@v5
-        with:
-          python-version: '3.11'
-
-      - name: Update pre-commit hooks
-        id: update
-        run: |
-          pip install pre-commit
-          pre-commit autoupdate
-          if [[ -n $(git status --porcelain) ]]; then
-            echo "changes=true" >> $GITHUB_OUTPUT
-            git diff .pre-commit-config.yaml > pre-commit-update.diff
-          fi
-
-      - name: Create Pull Request
-        if: steps.update.outputs.changes == 'true'
-        uses: peter-evans/create-pull-request@v6
-        with:
-          token: ${{ secrets.GITHUB_TOKEN }}
-          branch: update/pre-commit-hooks
-          delete-branch: true
-          title: "chore: update pre-commit hooks"
-          commit-message: "chore: update pre-commit hooks"
-          body: |
-            Automated PR to update pre-commit hooks to their latest versions.
-
-            <details>
-            <summary>Changes:</summary>
-
-            ```diff
-            ${{ steps.update.outputs.diff }}
-            ```
-            </details>

.github/workflows/pypi.yml

@@ -40,7 +40,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip3 install wheel packaging==23.2
+          pip3 install wheel packaging
           pip3 install --no-build-isolation -e .
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 

.github/workflows/tests-nightly.yml

@@ -42,7 +42,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+          pip3 install --upgrade packaging setuptools wheel
 
       - name: Install PyTorch
         run: |
@@ -59,7 +59,7 @@ jobs:
       - name: Install dependencies
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2
+          pip3 install --upgrade packaging
           pip3 install --no-build-isolation -U -e .
           python scripts/unsloth_install.py | sh
           python scripts/cutcrossentropy_install.py | sh

.github/workflows/tests.yml

@@ -74,7 +74,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+          pip3 install --upgrade packaging setuptools wheel
 
       - name: Install PyTorch
         run: |
@@ -98,9 +98,8 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ tests/
           pytest -v tests/patched/
-          pytest -v tests/cli/
 
       - name: cleanup pip cache
         run: |
@@ -148,7 +147,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 setuptools_scm build wheel
+          pip3 install --upgrade packaging setuptools setuptools_scm build wheel
 
       - name: Install PyTorch
         run: |
@@ -173,9 +172,8 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ tests/
           pytest -v tests/patched/
-          pytest -v tests/cli/
 
       - name: cleanup pip cache
         run: |

.gitignore

@@ -181,10 +181,6 @@ prepared-datasets/
 submit.sh
 *.out*
 
-# Quartodoc generated files
-objects.json
-site_libs/
-
 typings/
 out/
 

.pre-commit-config.yaml

@@ -3,7 +3,7 @@ default_language_version:
 
 repos:
 -   repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v5.0.0
+    rev: v4.4.0
     hooks:
     -   id: check-yaml
     -   id: end-of-file-fixer
@@ -11,23 +11,23 @@ repos:
     -   id: no-commit-to-branch
         args: ['--branch', 'main']
 -   repo: https://github.com/psf/black
-    rev: 25.1.0
+    rev: 23.3.0
     hooks:
     -   id: black
 -   repo: https://github.com/pycqa/isort
-    rev: 6.0.1
+    rev: 5.12.0
     hooks:
       - id: isort
 -   repo: https://github.com/PyCQA/flake8
-    rev: 7.1.2
+    rev: 6.1.0
     hooks:
     - id: flake8
--   repo: https://github.com/pylint-dev/pylint
+-   repo: https://github.com/PyCQA/pylint
-    rev: v3.3.6
+    rev: v3.3.0
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.15.0
+    rev: v1.3.0
     hooks:
     - id: mypy
       additional_dependencies:
@@ -36,7 +36,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.8.3
+    rev: 1.7.5
     hooks:
     -   id: bandit
         args: [

README.md

@@ -55,7 +55,6 @@ Features:
 ### Installation
 
 ```bash
-pip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 
 # Download example axolotl configs, deepspeed configs
@@ -97,7 +96,6 @@ That's it! Check out our [Getting Started Guide](https://axolotl-ai-cloud.github
 - [Multi-GPU Training](https://axolotl-ai-cloud.github.io/axolotl/docs/multi-gpu.html)
 - [Multi-Node Training](https://axolotl-ai-cloud.github.io/axolotl/docs/multi-node.html)
 - [Multipacking](https://axolotl-ai-cloud.github.io/axolotl/docs/multipack.html)
-- [API Reference](https://axolotl-ai-cloud.github.io/axolotl/docs/api/) - Auto-generated code documentation
 - [FAQ](https://axolotl-ai-cloud.github.io/axolotl/docs/faq.html) - Frequently asked questions
 
 ## 🤝 Getting Help

_quarto.yml

@@ -1,178 +1,6 @@
 project:
   type: website
 
-quartodoc:
-  dir: docs/api
-  package: axolotl
-  title: API Reference
-  parser: google
-
-  sections:
-    - title: Core
-      desc: Core functionality for training
-      contents:
-        - train
-        - evaluate
-        - datasets
-        - convert
-        - prompt_tokenizers
-        - logging_config
-        - core.trainer_builder
-        - core.training_args
-        - core.chat.messages
-        - core.chat.format.chatml
-        - core.chat.format.llama3x
-        - core.chat.format.shared
-        - core.datasets.chat
-        - core.datasets.transforms.chat_builder
-    - title: CLI
-      desc: Command-line interface
-      contents:
-        - cli.main
-        - cli.train
-        - cli.evaluate
-        - cli.args
-        - cli.checks
-        - cli.config
-        - cli.inference
-        - cli.merge_lora
-        - cli.merge_sharded_fsdp_weights
-        - cli.preprocess
-        - cli.sweeps
-        - cli.utils
-        - cli.cloud.base
-        - cli.cloud.modal_
-    - title: Trainers
-      desc: Training implementations
-      contents:
-        - core.trainers.base
-        - core.trainers.trl
-        - core.trainers.dpo.trainer
-        - core.trainers.grpo.trainer
-    - title: Prompt Strategies
-      desc: Prompt formatting strategies
-      contents:
-        - prompt_strategies.base
-        - prompt_strategies.chat_template
-        - prompt_strategies.alpaca_chat
-        - prompt_strategies.alpaca_instruct
-        - prompt_strategies.alpaca_w_system
-        - prompt_strategies.user_defined
-        - prompt_strategies.llama2_chat
-        - prompt_strategies.completion
-        - prompt_strategies.input_output
-        - prompt_strategies.stepwise_supervised
-        - prompt_strategies.metharme
-        - prompt_strategies.orcamini
-        - prompt_strategies.pygmalion
-        - prompt_strategies.messages.chat
-        - prompt_strategies.dpo.chat_template
-        - prompt_strategies.dpo.llama3
-        - prompt_strategies.dpo.chatml
-        - prompt_strategies.dpo.zephyr
-        - prompt_strategies.dpo.user_defined
-        - prompt_strategies.dpo.passthrough
-        - prompt_strategies.kto.llama3
-        - prompt_strategies.kto.chatml
-        - prompt_strategies.kto.user_defined
-        - prompt_strategies.orpo.chat_template
-        - prompt_strategies.bradley_terry.llama3
-    - title: Kernels
-      desc: Low-level performance optimizations
-      contents:
-        - kernels.lora
-        - kernels.geglu
-        - kernels.swiglu
-        - kernels.quantize
-        - kernels.utils
-    - title: MonkeyPatches
-      desc: Runtime patches for model optimizations
-      contents:
-        - monkeypatch.llama_attn_hijack_flash
-        - monkeypatch.llama_attn_hijack_xformers
-        - monkeypatch.mistral_attn_hijack_flash
-        - monkeypatch.multipack
-        - monkeypatch.relora
-        - monkeypatch.llama_expand_mask
-        - monkeypatch.lora_kernels
-        - monkeypatch.utils
-        - monkeypatch.btlm_attn_hijack_flash
-        - monkeypatch.llama_patch_multipack
-        - monkeypatch.stablelm_attn_hijack_flash
-        - monkeypatch.trainer_fsdp_optim
-        - monkeypatch.transformers_fa_utils
-        - monkeypatch.unsloth_
-        - monkeypatch.attention.mllama
-        - monkeypatch.data.batch_dataset_fetcher
-        - monkeypatch.mixtral
-    - title: Utils
-      desc: Utility functions
-      contents:
-        - utils.models
-        - utils.tokenization
-        - utils.chat_templates
-        - utils.lora
-        - utils.lora_embeddings
-        - utils.model_shard_quant
-        - utils.bench
-        - utils.freeze
-        - utils.trainer
-        - utils.schedulers
-        - utils.distributed
-        - utils.dict
-        - utils.optimizers.adopt
-        - utils.data.pretraining
-        - utils.data.sft
-        - utils.gradient_checkpointing.unsloth
-    - title: Schemas
-      desc: Pydantic data models for Axolotl config
-      contents:
-        - utils.schemas.config
-        - utils.schemas.model
-        - utils.schemas.training
-        - utils.schemas.datasets
-        - utils.schemas.peft
-        - utils.schemas.trl
-        - utils.schemas.integrations
-        - utils.schemas.enums
-        - utils.schemas.utils
-    - title: Integrations
-      desc: Third-party integrations and extensions
-      contents:
-        - integrations.base
-        - integrations.cut_cross_entropy.args
-        - integrations.grokfast.optimizer
-        - integrations.kd.trainer
-        - integrations.liger.args
-        - integrations.lm_eval.args
-        - integrations.spectrum.args
-    - title: Common
-      desc: Common utilities and shared functionality
-      contents:
-        - common.architectures
-        - common.const
-        - common.datasets
-    - title: Models
-      desc: Custom model implementations
-      contents:
-        - models.mamba.modeling_mamba
-    - title: Data Processing
-      desc: Data processing utilities
-      contents:
-        - utils.collators.core
-        - utils.collators.batching
-        - utils.collators.mamba
-        - utils.collators.mm_chat
-        - utils.samplers.multipack
-    - title: Callbacks
-      desc: Training callbacks
-      contents:
-        - utils.callbacks.perplexity
-        - utils.callbacks.profiler
-        - utils.callbacks.lisa
-        - utils.callbacks.mlflow_
-        - utils.callbacks.comet_
-
 website:
   title: "Axolotl"
   description: "We make fine-tuning accessible, scalable, and fun"
@@ -204,11 +32,8 @@ website:
           contents:
             - docs/getting-started.qmd
             - docs/installation.qmd
-            - docs/inference.qmd
             - docs/cli.qmd
-            - docs/config.qmd
+            - docs/inference.qmd
-            - text: "API Reference"
-              href: docs/api
 
         - section: "Dataset Formats"
           contents: docs/dataset-formats/*
@@ -249,27 +74,12 @@ website:
             - docs/debugging.qmd
             - docs/nccl.qmd
 
+        - section: "Reference"
+          contents:
+            - docs/config.qmd
+
 format:
   html:
     theme: darkly
     css: styles.css
     toc: true
-    # Enable better handling of line breaks in markdown
-    preserve-tabs: true
-    html-math-method: mathjax
-    # Improved markdown processing options
-    md-extensions:
-      - markdown_it
-      - def_list
-      - attr_list
-      - fenced_divs
-      - tables
-      - html_admonition
-      - lineblocks
-      - fancy_lists
-    # Control whitespace handling
-    whitespace: preserve
-    # Process newlines in paragraphs
-    wrap: preserve
-    # Better line break handling
-    preserve-linebreaks: true

cicd/Dockerfile.jinja

@@ -31,11 +31,10 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
-RUN pip install packaging==23.2 setuptools==75.8.0
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
     fi
 
 RUN python scripts/unsloth_install.py | sh

cicd/cicd.sh

@@ -3,10 +3,9 @@ set -e
 
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
-pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli /workspace/axolotl/tests/
+pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ /workspace/axolotl/tests/
 pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/lora_kernels  # running these with the other patches causes a failure
 pytest -v --durations=10 --ignore=tests/e2e/patched/lora_kernels /workspace/axolotl/tests/e2e/patched
 pytest -v --durations=10 -n1 /workspace/axolotl/tests/e2e/solo/
 pytest -v --durations=10 /workspace/axolotl/tests/e2e/integrations/
-pytest -v --durations=10 /workspace/axolotl/tests/cli
+pytest -v --durations=10 --ignore=tests/e2e/solo/ --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/
-pytest -v --durations=10 --ignore=tests/e2e/solo/ --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ --ignore=tests/cli /workspace/axolotl/tests/e2e/

cicd/multigpu.py

@@ -1,7 +1,6 @@
 """
  modal application to run axolotl gpu tests in Modal
  """
-
 # pylint: disable=duplicate-code
 
 import os

cicd/tests.py

@@ -1,5 +1,4 @@
 """Modal app to run axolotl GPU tests"""
-
 # pylint: disable=duplicate-code
 
 import os

docker/Dockerfile-base

@@ -28,7 +28,7 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
 
 WORKDIR /workspace
 
-RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
+RUN python3 -m pip install --upgrade pip && pip3 install packaging && \
     python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
     python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
     python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"

docker/Dockerfile-base-nightly

@@ -1,39 +0,0 @@
-ARG CUDA_VERSION="12.8.1"
-ARG CUDNN_VERSION="8"
-ARG UBUNTU_VERSION="22.04"
-ARG MAX_JOBS=4
-
-FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION AS base-builder
-
-ENV PATH="/root/miniconda3/bin:${PATH}"
-
-ARG PYTHON_VERSION="3.11"
-ARG PYTORCH_VERSION="nightly"
-ARG CUDA="128"
-ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
-
-ENV PYTHON_VERSION=$PYTHON_VERSION
-ENV TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST
-
-RUN apt-get update \
-    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev pkg-config && rm -rf /var/lib/apt/lists/* \
-    && wget \
-    https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
-    && mkdir /root/.conda \
-    && bash Miniconda3-latest-Linux-x86_64.sh -b \
-    && rm -f Miniconda3-latest-Linux-x86_64.sh \
-    && conda create -n "py${PYTHON_VERSION}" python="${PYTHON_VERSION}"
-
-ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
-
-WORKDIR /workspace
-
-RUN python3 -m pip install --upgrade pip && pip3 install packaging && \
-    python3 -m pip install --no-cache-dir -U torch --extra-index-url https://download.pytorch.org/whl/nightly/cu$CUDA && \
-    python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
-    python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"
-
-RUN git lfs install --skip-repo && \
-    pip3 install awscli && \
-    # The base image ships with `pydantic==1.8.2` which is not working
-    pip3 install -U --no-cache-dir pydantic==1.10.10

docs/.gitignore

@@ -1,4 +1,2 @@
 /.quarto/
 _site/
-/api/*.qmd
-/api/*.html

docs/cli.qmd

@@ -1,5 +1,5 @@
 ---
-title: "Command Line Interface (CLI)"
+title: "CLI Reference"
 format:
   html:
     toc: true

docs/config.qmd

@@ -1,5 +1,5 @@
 ---
-title: Config Reference
+title: Config options
 description: A complete list of all configuration options.
 ---
 
@@ -30,11 +30,6 @@ tokenizer_legacy:
 # Resize the model embeddings when new tokens are added to multiples of 32
 # This is reported to improve training speed on some models
 resize_token_embeddings_to_32x:
-# Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.
-shrink_embeddings:
-# Whether to load the model with randomly initialized weights. Useful for
-# pre-training a model from scratch or debugging purposes.
-random_init_weights:
 
 # (Internal use only)
 # Used to identify which the model is based on
@@ -88,12 +83,6 @@ gpu_memory_limit: 20GiB
 # Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge
 lora_on_cpu: true
 
-# List[str]. Add plugins to extend the pipeline.
-# See `src/axolotl/integrations` for the available plugins or doc below for more details.
-# https://axolotl-ai-cloud.github.io/axolotl/docs/custom_integrations.html
-plugins:
-  # - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
 # A list of one or more datasets to finetune the model with
 datasets:
   # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
@@ -216,46 +205,10 @@ test_datasets:
     data_files:
       - /workspace/data/eval.jsonl
 
-# use RL training: 'dpo', 'ipo', 'kto', 'simpo', 'orpo', 'grpo'
+# use RL training: 'dpo', 'ipo', 'kto'
 rl:
-rl_beta:  # Optional[float]. The beta parameter for the RL training.
+# whether to perform weighting if doing DPO training. Boolean.
-
+dpo_use_weighting:
-# dpo
-dpo_use_weighting:  # Optional[bool]. Whether to perform weighting.
-rpo_alpha: # Optional[float]. Weighting of NLL term in loss from RPO paper.
-
-# orpo
-orpo_alpha: 0.1  # Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to `beta` in `ORPOConfig` due to trl mapping.
-
-# kto
-kto_desirable_weight: # Optional[float]. Factor for desirable loss term in KTO loss.
-kto_undesirable_weight: # Optional[float]. Factor for undesirable loss term in KTO loss.
-
-# simpo
-cpo_alpha: 1.0  # Weight of the BC regularizer
-simpo_gamma: 0.5  # Target reward margin for the SimPO loss
-
-# grpo
-trl:
-  use_vllm: # Optional[bool]. Whether to use VLLM for RL training.
-  vllm_device: # Optional[str]. Device to use for VLLM.
-  vllm_gpu_memory_utilization: # Optional[float]. GPU memory utilization for VLLM.
-  vllm_max_model_len: # Optional[int]. Maximum length of the model for VLLM.
-  vllm_dtype: # Optional[str]. Data type for VLLM.
-
-  beta: # Optional[float]. Beta parameter for the RL training. Same as `rl_beta`. Use
-  max_completion_length: # Optional[int]. Maximum length of the completion for RL training.
-
-  reward_funcs: # Optional[list[str]]. List of reward functions to load. Paths must be importable from current dir.
-  reward_weights: # Optional[list[float]]. List of reward weights for the reward functions.
-
-  num_generations: # Optional[int]. Number of generations to sample.
-  log_completions: # Optional[bool]. Whether to log completions.
-
-  sync_ref_model: # Optional[bool]. Whether to sync the reference model.
-  ref_model_mixup_alpha: # Optional[float]. Mixup alpha for the reference model.
-  ref_model_sync_steps: # Optional[int]. Sync steps for the reference model.
-
 
 # reward modelling: `True` or `False`
 reward_model:
@@ -279,7 +232,7 @@ default_system_message: You are a helpful assistant. Please give a long and deta
 # subsequent training attempts load faster, relative path
 dataset_prepared_path: data/last_run_prepared
 # Push prepared dataset to hub
-push_dataset_to_hub: # Optional[str] repo_org/repo_name
+push_dataset_to_hub: # repo path
 # The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`
 # if not set.
 dataset_processes: # defaults to os.cpu_count() if not set
@@ -620,14 +573,6 @@ ddp_timeout:
 ddp_bucket_cap_mb:
 ddp_broadcast_buffers:
 
-# Sequence parallelism
-# Set to a divisor of the number of GPUs available to split sequences into chunks of equal size.
-# Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM.
-# E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized
-# subsequences, or set to 4 to split into four equal-sized subsequences.
-# See https://axolotl-ai-cloud.github.io/axolotl/docs/sequence_parallelism.html for more details.
-sequence_parallel_degree:
-
 # Path to torch distx for optim 'adamw_anyprecision'
 torchdistx_path:
 

docs/custom_integrations.qmd

@@ -55,47 +55,3 @@ sections = [
 for section_name, folder_name in sections:
     print(print_section(section_name, folder_name))
 ```
-
-## Adding a new integration
-
-Plugins can be used to customize the behavior of the training pipeline through [hooks](https://en.wikipedia.org/wiki/Hooking). See [`axolotl.integrations.BasePlugin`](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/integrations/base.py) for the possible hooks.
-
-To add a new integration, please follow these steps:
-
-1. Create a new folder in the `src/axolotl/integrations` directory.
-2. Add any relevant files (`LICENSE`, `README.md`, `ACKNOWLEDGEMENTS.md`, etc.) to the new folder.
-3. Add `__init__.py` and `args.py` files to the new folder.
-  - `__init__.py` should import the integration and hook into the appropriate functions.
-  - `args.py` should define the arguments for the integration.
-4. (If applicable) Add CPU tests under `tests/integrations` or GPU tests under `tests/e2e/integrations`.
-
-::: {.callout-tip}
-
-See [src/axolotl/integrations/cut_cross_entropy](https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/cut_cross_entropy) for a minimal integration example.
-
-:::
-
-::: {.callout-warning}
-
-If you could not load your integration, please ensure you are pip installing in editable mode.
-
-```bash
-pip install -e .
-```
-
-and correctly spelled the integration name in the config file.
-
-```yaml
-plugins:
-  - axolotl.integrations.your_integration_name.YourIntegrationPlugin
-```
-
-:::
-
-::: {.callout-note}
-
-It is not necessary to place your integration in the `integrations` folder. It can be in any location, so long as it's installed in a package in your python env.
-
-See this repo for an example: [https://github.com/axolotl-ai-cloud/diff-transformer](https://github.com/axolotl-ai-cloud/diff-transformer)
-
-:::

docs/dataset_preprocessing.qmd

@@ -6,7 +6,7 @@ description: How datasets are processed
 ## Overview
 
 Dataset pre-processing is the step where Axolotl takes each dataset you've configured alongside
-the [dataset format](dataset-formats) and prompt strategies to:
+the [dataset format](docs/dataset-formats) and prompt strategies to:
 
  - parse the dataset based on the *dataset format*
  - transform the dataset to how you would interact with the model based on the *prompt strategy*

docs/faq.qmd

@@ -27,20 +27,6 @@ description: Frequently asked questions
 
 > A: This is usually an issue with the GPU. This can be resolved through setting the os environment variable `CUDA_VISIBLE_DEVICES=0`. If you are on runpod, this is usually a pod issue. Starting a new pod should take care of it.
 
-**Q: Received mismatch error on merge adapters / loading adapters between torch.Size of checkpoint and model.**
-
-> A: This is likely due to vocab size mismatch. By default, Axolotl expands the model's embeddings if the tokenizer has more tokens than the model. Please use the `axolotl merge-lora` command to merge the adapters instead of using your own scripts.
-
-> On the other hand, if the model has more tokens than the tokenizer, Axolotl does not shrink the model's embeddings unless `shrink_embeddings: true` is set in the config.
-
-**Q: How to call Axolotl via custom python scripts?**
-
-> A: Yes, since Axolotl is just Python, please see `src/axolotl/cli/main.py` on how each command is called.
-
-**Q: How to know the value to use for `fsdp_transformer_layer_cls_to_wrap`?**
-
-> A: This is the class name of the transformer layer to wrap with FSDP. For example, for `LlamaForCausalLM`, the value is `LlamaDecoderLayer`. To find this for a specific model, check the model's `PreTrainedModel` definition and look for `_no_split_modules` variable in the `modeling_<model_name>.py` file within `transformers` library.
-
 ### Chat templates
 
 **Q: `jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____`**

docs/getting-started.qmd

@@ -36,9 +36,7 @@ The YAML configuration file controls everything about your training. Here's what
 
 ```yaml
 base_model: NousResearch/Llama-3.2-1B
-
+# hub_model_id: username/custom_model_name
-load_in_8bit: true
-adapter: lora
 
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
@@ -46,15 +44,11 @@ datasets:
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.1
 output_dir: ./outputs/lora-out
+
+adapter: lora
+lora_model_dir:
 ```
 
-::: {.callout-tip}
-`load_in_8bit: true` and `adapter: lora` enables LoRA adapter finetuning.
-
-- To perform Full finetuning, remove these two lines.
-- To perform QLoRA finetuning, replace with `load_in_4bit: true` and `adapter: qlora`.
-:::
-
 See our [Config options](config.qmd) for more details.
 
 ### Training {#sec-training}
@@ -62,7 +56,7 @@ See our [Config options](config.qmd) for more details.
 When you run `axolotl train`, Axolotl:
 
 1. Downloads the base model
-2. (If specified) applies QLoRA/LoRA adapter layers
+2. (If specified) applies LoRA adapter layers
 3. Loads and processes the dataset
 4. Runs the training loop
 5. Saves the trained model and / or LoRA weights
@@ -75,8 +69,6 @@ Let's modify the example for your own data:
 
 ```yaml
 base_model: NousResearch/Nous-Hermes-llama-1b-v1
-
-load_in_8bit: true
 adapter: lora
 
 # Training settings
@@ -112,6 +104,8 @@ format):
 {"instruction": "Classify this text", "input": "Not good at all", "output": "negative"}
 ```
 
+Please consult the supported [Dataset Formats](dataset-formats/) for more details.
+
 3. Run the training:
 
 ```bash

docs/inference.qmd

@@ -1,5 +1,5 @@
 ---
-title: "Inference and Merging"
+title: "Inference"
 format:
   html:
     toc: true
@@ -9,14 +9,10 @@ execute:
   enabled: false
 ---
 
-This guide covers how to use your trained models for inference, including model loading, interactive testing, merging adapters, and common troubleshooting steps.
+This guide covers how to use your trained models for inference, including model loading, interactive testing, and common troubleshooting steps.
 
 ## Quick Start {#sec-quickstart}
 
-::: {.callout-tip}
-Use the same config used for training on inference/merging.
-:::
-
 ### Basic Inference {#sec-basic}
 
 ::: {.panel-tabset}

docs/installation.qmd

@@ -22,7 +22,6 @@ This guide covers all the ways you can install and set up Axolotl for your envir
 ### PyPI Installation (Recommended) {#sec-pypi}
 
 ```{.bash}
-pip3 install -U packaging setuptools wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 ```
 
@@ -38,7 +37,7 @@ For the latest features between releases:
 ```{.bash}
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
-pip3 install -U packaging setuptools wheel ninja
+pip3 install packaging ninja
 pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
 ```
 
@@ -79,7 +78,6 @@ For providers supporting Docker:
   - [Latitude.sh](https://latitude.sh/blueprint/989e0e79-3bf6-41ea-a46b-1f246e309d5c)
   - [JarvisLabs.ai](https://jarvislabs.ai/templates/axolotl)
   - [RunPod](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
-  - [Novita](https://novita.ai/gpus-console?templateId=311)
 
 ### Google Colab {#sec-colab}
 
@@ -109,7 +107,7 @@ We recommend using WSL2 (Windows Subsystem for Linux) or Docker.
 2. Install PyTorch: https://pytorch.org/get-started/locally/
 3. Install Axolotl:
    ```{.bash}
-   pip3 install -U packaging setuptools wheel ninja
+   pip3 install packaging
    pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
    ```
 4. (Optional) Login to Hugging Face:

docs/lora_optims.qmd

@@ -66,10 +66,6 @@ logic to be compatible with more of them.
 
 </details>
 
-::: {.callout-tip}
-Check out our [LoRA optimizations blog](https://axolotlai.substack.com/p/accelerating-lora-fine-tuning-with).
-:::
-
 ## Usage
 
 These optimizations can be enabled in your Axolotl config YAML file. The

docs/reward_modelling.qmd

@@ -41,10 +41,6 @@ Bradley-Terry chat templates expect single-turn conversations in the following f
 
 ### Process Reward Models (PRM)
 
-::: {.callout-tip}
-Check out our [PRM blog](https://axolotlai.substack.com/p/process-reward-models).
-:::
-
 Process reward models are trained using data which contains preference annotations for each step in a series of interactions. Typically, PRMs are trained to provide reward signals over each step of a reasoning trace and are used for downstream reinforcement learning.
 ```yaml
 base_model: Qwen/Qwen2.5-3B

docs/rlhf.qmd

@@ -298,7 +298,7 @@ The input format is a simple JSON input with customizable fields based on the ab
 
 ### IPO
 
-As IPO is just DPO with a different loss function, all supported dataset formats for [DPO](#dpo) are also supported for IPO.
+As IPO is just DPO with a different loss function, all supported options for DPO works here.
 
 ```yaml
 rl: ipo
@@ -344,9 +344,8 @@ ORPO supports the following types with the following dataset format:
 
 ```yaml
 rl: kto
-rl_beta: 0.1  # default
+rl_beta: 0.5
-kto_desirable_weight: 1.0  # default
+kto_desirable_weight: 0.2
-kto_undesirable_weight: 1.0  # default
 
 remove_unused_columns: false
 
@@ -498,10 +497,6 @@ The input format is a simple JSON input with customizable fields based on the ab
 
 ### GRPO
 
-::: {.callout-tip}
-Check out our [GRPO cookbook](https://github.com/axolotl-ai-cloud/axolotl-cookbook/tree/main/grpo#training-an-r1-style-large-language-model-using-grpo).
-:::
-
 GRPO uses custom reward functions and transformations. Please have them ready locally.
 
 For ex, to load OpenAI's GSM8K and use a random reward for completions:
@@ -545,19 +540,6 @@ To see other examples of custom reward functions, please see [TRL GRPO Docs](htt
 
 To see description of the configs, please see [TRLConfig](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/utils/config/models/input/v0_4_1/trl.py).
 
-### SimPO
-
-SimPO uses [CPOTrainer](https://huggingface.co/docs/trl/main/en/cpo_trainer) but with alternative loss function.
-
-```yaml
-rl: simpo
-rl_beta: 0.1  # default in CPOTrainer
-cpo_alpha: 1.0  # default in CPOTrainer
-simpo_gamma: 0.5  # default in CPOTrainer
-```
-
-This method uses the same dataset format as [DPO](#dpo).
-
 ### Using local dataset files
 
 ```yaml

docs/sequence_parallelism.qmd

@@ -1,90 +0,0 @@
----
-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
-through a ring communication pattern.
-
-## When to Use Sequence Parallelism
-
-Use sequence parallelism when:
-
-- You need to train with sequence lengths that don't fit into a single GPU's memory
-- You have multiple GPUs available
-- You're experiencing OOM (Out Of Memory) errors with long sequences
-
-## Configuration
-
-To enable sequence parallelism, add the following to your configuration file:
-
-```yaml
-# Set to a divisor (> 1) of the number of GPUs available
-sequence_parallel_degree: 4  # Split sequences across 4 GPUs
-```
-
-The `sequence_parallel_degree` should be a divisor of the total number of GPUs. For example:
-
-- With 8 GPUs, valid values would be 2, 4, or 8
-- With 4 GPUs, valid values would be 2 or 4
-
-## Implementation Details
-
-When sequence parallelism is enabled:
-
-1. Each sequence is divided into equal chunks across the GPUs in a sequence parallel group
-2. The data collator handles the chunking of input_ids, attention_mask, labels, and position_ids
-3. Position IDs are adjusted to maintain proper relative positions, especially for packed sequences
-4. The trainer uses special ring communication patterns for attention operations
-
-## Requirements
-
-To use sequence parallelism, you need:
-
-- Multiple GPUs (at least 2)
-- The `ring-flash-attn` package. Install with:
-  - `pip install axolotl[ring-flash-attn]` (preferred)
-  - `pip install ring-flash-attn>=0.1.4`
-
-## Limitations
-
-- Flash attention must be enabled for this to work (`flash_attention: true` in config YAML)
-- May have a small performance overhead due to communication between GPUs
-
-## Example
-
-```yaml
-# Example config with sequence parallelism
-base_model: meta-llama/Llama-3-8B-Instruct
-sequence_len: 8192
-sequence_parallel_degree: 2  # Split each sequence into 4 parts
-flash_attention: true  # Required with sequence parallelism
-...
-```
-
-This will train the Llama 3 8B model with 8K context length, with each sequence split
-into 2 subsequences of length 4096 across 2 GPUs.
-
-## Sample Packing with Sequence Parallelism
-
-Sequence parallelism is compatible with Axolotl's sample packing functionality. When using both features together:
-
-1. Samples are first packed together
-2. The packed sequences are then divided across GPUs in the sequence parallel group
-3. Position IDs are automatically adjusted to maintain proper relative positions
-
-## Effect on Batch Size
-
-When using sequence parallelism, your effective global batch size is **divided** by the `sequence_parallel_degree`. This happens because:
-
-- Each group of `sequence_parallel_degree` GPUs works on the same batch (just different parts of each sequence)
-- The number of batches processed per step decreases
-
-For example:
-- With 8 GPUs and no sequence parallelism: 8 different batches processed per step
-- With 8 GPUs and `sequence_parallel_degree=4`: Only 2 different batches processed per step (each split across 4 GPUs)
-- If your per-GPU `micro_batch_size` is 2, the global batch size decreases from 16 to 4

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

@@ -55,7 +55,7 @@ tf32: true
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
-  use_reentrant: false
+  use_reentrant: true
 early_stopping_patience:
 resume_from_checkpoint:
 local_rank:

pyproject.toml

@@ -1,5 +1,5 @@
 [build-system]
-requires = ["setuptools>=64", "wheel", "setuptools_scm>=8", "packaging==23.2"]
+requires = ["setuptools>=64", "wheel", "setuptools_scm>=8"]
 build-backend = "setuptools.build_meta"
 
 [project]
@@ -8,7 +8,6 @@ dynamic = ["version", "dependencies", "optional-dependencies"]
 description = "LLM Trainer"
 readme = "README.md"
 requires-python = ">=3.10"
-# license = "Apache-2.0"
 
 [project.scripts]
 axolotl = "axolotl.cli.main:main"

requirements-dev.txt

@@ -2,5 +2,3 @@ pre-commit
 black
 mypy
 types-requests
-quartodoc
-jupyter

requirements.txt

@@ -1,9 +1,10 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 
 # START section of dependencies that don't install on Darwin/MacOS
-bitsandbytes==0.45.3
+bitsandbytes==0.45.2
 triton>=3.0.0
 mamba-ssm==1.2.0.post1
+flash-attn==2.7.4.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
 liger-kernel==0.5.3
@@ -11,12 +12,12 @@ liger-kernel==0.5.3
 
 packaging==23.2
 
-peft==0.15.0
+peft==0.14.0
 transformers==4.49.0
-tokenizers>=0.21.1
+tokenizers>=0.21.0
-accelerate==1.5.2
+accelerate==1.3.0
-datasets==3.4.1
+datasets==3.2.0
-deepspeed==0.16.4
+deepspeed==0.16.1
 trl==0.15.1
 
 optimum==1.16.2
@@ -35,7 +36,6 @@ einops
 colorama
 numba
 numpy>=1.24.4,<=2.0.1
-
 # qlora things
 evaluate==0.4.1
 scipy

scripts/chat_datasets.py

@@ -1,7 +1,6 @@
 """
 helper script to parse chat datasets into a usable yaml
 """
-
 import click
 import yaml
 from datasets import load_dataset

scripts/cutcrossentropy_install.py

@@ -1,5 +1,4 @@
 """Script to output the correct installation command for cut-cross-entropy."""
-
 import importlib.util
 import sys
 

setup.py

@@ -17,7 +17,11 @@ def parse_requirements():
         lines = [r.strip() for r in requirements_file.readlines()]
         for line in lines:
             is_extras = (
-                "deepspeed" in line or "mamba-ssm" in line or "lion-pytorch" in line
+                "flash-attn" in line
+                or "flash-attention" in line
+                or "deepspeed" in line
+                or "mamba-ssm" in line
+                or "lion-pytorch" in line
             )
             if line.startswith("--extra-index-url"):
                 # Handle custom index URLs
@@ -35,6 +39,7 @@ def parse_requirements():
                 "bitsandbytes",
                 "triton",
                 "mamba-ssm",
+                "flash-attn",
                 "xformers",
                 "autoawq",
                 "liger-kernel",
@@ -119,10 +124,11 @@ setup(
         ],
     },
     extras_require={
-        "flash-attn": ["flash-attn==2.7.4.post1"],
+        "flash-attn": [
-        "ring-flash-attn": ["ring-flash-attn>=0.1.4", "yunchang==0.6.0"],
+            "flash-attn==2.7.4.post1",
+        ],
         "deepspeed": [
-            "deepspeed==0.16.4",
+            "deepspeed==0.16.1",
             "deepspeed-kernels",
         ],
         "mamba-ssm": [

src/axolotl/cli/cloud/__init__.py

@@ -1,7 +1,6 @@
 """
 launch axolotl in supported cloud platforms
 """
-
 from pathlib import Path
 from typing import Union
 

src/axolotl/cli/cloud/base.py

@@ -1,7 +1,6 @@
 """
 base class for cloud platforms from cli
 """
-
 from abc import ABC, abstractmethod
 
 

src/axolotl/cli/cloud/modal_.py

@@ -1,7 +1,6 @@
 """
 Modal Cloud support from CLI
 """
-
 import copy
 import json
 import os

src/axolotl/cli/main.py

@@ -1,5 +1,4 @@
 """Click CLI definitions for various axolotl commands."""
-
 # pylint: disable=redefined-outer-name
 
 import logging
@@ -25,7 +24,7 @@ from axolotl.cli.utils import (
 )
 from axolotl.integrations.lm_eval.cli import lm_eval
 from axolotl.utils import set_pytorch_cuda_alloc_conf
-from axolotl.utils.schemas.config import AxolotlInputConfig
+from axolotl.utils.config.models.input.v0_4_1 import AxolotlInputConfig
 
 
 @click.group()

src/axolotl/cli/train.py

@@ -23,7 +23,7 @@ from axolotl.utils.dict import DictDefault
 LOG = logging.getLogger(__name__)
 
 
-def do_train(cfg: DictDefault, cli_args: TrainerCliArgs):
+def do_train(cfg: DictDefault, cli_args: TrainerCliArgs) -> None:
     """
     Trains a `transformers` model by first loading the dataset(s) specified in the
     `axolotl` config, and then calling `axolotl.train.train`. Also runs the plugin
@@ -44,13 +44,16 @@ def do_train(cfg: DictDefault, cli_args: TrainerCliArgs):
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
     model, tokenizer, trainer = train(cfg=cfg, dataset_meta=dataset_meta)
-    del model, tokenizer, trainer
-
     plugin_manager = PluginManager.get_instance()
+
+    del model
+    del tokenizer
+    del trainer
+
     plugin_manager.post_train_unload(cfg)
 
 
-def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
+def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
     """
     Parses `axolotl` config, CLI args, and calls `do_train`.
 

src/axolotl/cli/utils.py

@@ -5,6 +5,7 @@ import dataclasses
 import hashlib
 import json
 import logging
+import typing
 from functools import wraps
 from pathlib import Path
 from types import NoneType
@@ -23,7 +24,7 @@ configure_logging()
 LOG = logging.getLogger(__name__)
 
 
-def strip_optional_type(field_type: type | str | None):
+def strip_optional_type(field_type: type | typing._SpecialForm | None):
     """
     Extracts the non-`None` type from an `Optional` / `Union` type.
 

src/axolotl/convert.py

@@ -1,5 +1,6 @@
 """Module containing File Reader, File Writer, Json Parser, and Jsonl Serializer classes"""
 
+
 import json
 import sys
 

src/axolotl/core/chat/format/chatml.py

@@ -1,7 +1,6 @@
 """
 ChatML transformation functions for MessageContents
 """
-
 from typing import Optional
 
 from ..messages import MessageContents, Messages

src/axolotl/core/chat/format/llama3x.py

@@ -1,7 +1,6 @@
 """
 Llama 3.x chat formatting functions for MessageContents
 """
-
 from typing import Optional
 
 from ..messages import MessageContents, Messages

src/axolotl/core/chat/format/shared.py

@@ -1,7 +1,6 @@
 """
 shared functions for format transforms
 """
-
 from axolotl.core.chat.messages import MessageContents, Messages
 
 

src/axolotl/core/chat/messages.py

@@ -1,7 +1,6 @@
 """
 internal message representations of chat messages
 """
-
 import json
 from enum import Enum
 from typing import Any, Callable, List, Optional, Union

src/axolotl/core/datasets/chat.py

@@ -1,7 +1,6 @@
 """
 chat dataset module
 """
-
 import os
 from typing import Callable, Optional, Union
 

src/axolotl/core/datasets/transforms/chat_builder.py

@@ -1,7 +1,6 @@
 """
 This module contains a function that builds a transform that takes a row from the dataset and converts it to a Chat.
 """
-
 from typing import Any, Mapping, Union
 
 

src/axolotl/core/trainer_builder.py

@@ -13,7 +13,9 @@
 # limitations under the License.
 
 # pylint: disable=too-many-lines
-"""Builder for the training args and trainer"""
+"""
+Builder for the training args and trainer
+"""
 
 import abc
 import importlib
@@ -36,7 +38,7 @@ from transformers import (
 from transformers.training_args import OptimizerNames
 from trl.trainer.utils import RewardDataCollatorWithPadding
 
-from axolotl.core.trainers import (
+from axolotl.core.trainers.base import (
     AxolotlCPOTrainer,
     AxolotlKTOTrainer,
     AxolotlMambaTrainer,
@@ -83,8 +85,8 @@ from axolotl.utils.collators import (
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
 from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
+from axolotl.utils.config.models.input.v0_4_1 import CustomSupportedOptimizers
 from axolotl.utils.models import ensure_dtype
-from axolotl.utils.schemas.enums import CustomSupportedOptimizers
 
 try:
     import torch._dynamo  # pylint: disable=ungrouped-imports
@@ -330,9 +332,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs = {}
 
         if self.cfg.include_tokens_per_second is not None:
-            training_arguments_kwargs["include_tokens_per_second"] = (
+            training_arguments_kwargs[
-                self.cfg.include_tokens_per_second
+                "include_tokens_per_second"
-            )
+            ] = self.cfg.include_tokens_per_second
 
         if self.cfg.bf16 == "full":
             training_arguments_kwargs["bf16_full_eval"] = True
@@ -349,13 +351,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["seed"] = self.cfg.seed
 
         if self.cfg.gradient_checkpointing:
-            training_arguments_kwargs["gradient_checkpointing"] = (
+            training_arguments_kwargs[
-                self.cfg.gradient_checkpointing
+                "gradient_checkpointing"
-            )
+            ] = self.cfg.gradient_checkpointing
             if self.cfg.gradient_checkpointing_kwargs is not None:
-                training_arguments_kwargs["gradient_checkpointing_kwargs"] = (
+                training_arguments_kwargs[
-                    self.cfg.gradient_checkpointing_kwargs
+                    "gradient_checkpointing_kwargs"
-                )
+                ] = self.cfg.gradient_checkpointing_kwargs
         if self.cfg.fsdp:
             training_arguments_kwargs["fsdp"] = self.cfg.fsdp
             if self.cfg.fsdp_config:
@@ -371,9 +373,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["deepspeed"] = self.cfg.deepspeed
 
         if self.cfg.lr_quadratic_warmup is not None:
-            training_arguments_kwargs["lr_quadratic_warmup"] = (
+            training_arguments_kwargs[
-                self.cfg.lr_quadratic_warmup
+                "lr_quadratic_warmup"
-            )
+            ] = self.cfg.lr_quadratic_warmup
 
         if self.cfg.adam_beta1:
             training_arguments_kwargs["adam_beta1"] = self.cfg.adam_beta1
@@ -397,28 +399,28 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["save_safetensors"] = self.cfg.save_safetensors
 
         if self.cfg.dataloader_pin_memory is not None:
-            training_arguments_kwargs["dataloader_pin_memory"] = (
+            training_arguments_kwargs[
-                self.cfg.dataloader_pin_memory
+                "dataloader_pin_memory"
-            )
+            ] = self.cfg.dataloader_pin_memory
         if self.cfg.dataloader_num_workers is not None:
-            training_arguments_kwargs["dataloader_num_workers"] = (
+            training_arguments_kwargs[
-                self.cfg.dataloader_num_workers
+                "dataloader_num_workers"
-            )
+            ] = self.cfg.dataloader_num_workers
         if self.cfg.dataloader_prefetch_factor is not None:
-            training_arguments_kwargs["dataloader_prefetch_factor"] = (
+            training_arguments_kwargs[
-                self.cfg.dataloader_prefetch_factor
+                "dataloader_prefetch_factor"
-            )
+            ] = self.cfg.dataloader_prefetch_factor
         if self.cfg.dataloader_drop_last is not None:
-            training_arguments_kwargs["dataloader_drop_last"] = (
+            training_arguments_kwargs[
-                self.cfg.dataloader_drop_last
+                "dataloader_drop_last"
-            )
+            ] = self.cfg.dataloader_drop_last
         elif self.cfg.sample_packing and self.cfg.eval_sample_packing is False:
             training_arguments_kwargs["dataloader_drop_last"] = True
 
         if self.cfg.remove_unused_columns is not None:
-            training_arguments_kwargs["remove_unused_columns"] = (
+            training_arguments_kwargs[
-                self.cfg.remove_unused_columns
+                "remove_unused_columns"
-            )
+            ] = self.cfg.remove_unused_columns
 
         if not self.cfg.test_datasets and self.cfg.val_set_size == 0:
             # no eval set, so don't eval
@@ -450,9 +452,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.do_causal_lm_eval:
             training_arguments_kwargs["do_causal_lm_eval"] = self.cfg.do_causal_lm_eval
         if self.cfg.metric_for_best_model:
-            training_arguments_kwargs["metric_for_best_model"] = (
+            training_arguments_kwargs[
-                self.cfg.metric_for_best_model
+                "metric_for_best_model"
-            )
+            ] = self.cfg.metric_for_best_model
         if self.cfg.greater_is_better:
             training_arguments_kwargs["greater_is_better"] = self.cfg.greater_is_better
 
@@ -465,13 +467,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 )
                 training_arguments_kwargs["torch_compile"] = self.cfg.torch_compile
                 if self.cfg.torch_compile_backend:
-                    training_arguments_kwargs["torch_compile_backend"] = (
+                    training_arguments_kwargs[
-                        self.cfg.torch_compile_backend
+                        "torch_compile_backend"
-                    )
+                    ] = self.cfg.torch_compile_backend
                 if self.cfg.torch_compile_mode:
-                    training_arguments_kwargs["torch_compile_mode"] = (
+                    training_arguments_kwargs[
-                        self.cfg.torch_compile_mode
+                        "torch_compile_mode"
-                    )
+                    ] = self.cfg.torch_compile_mode
 
         # DDP Config
         if self.cfg.ddp_timeout:
@@ -480,32 +482,32 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.ddp_bucket_cap_mb:
             training_arguments_kwargs["ddp_bucket_cap_mb"] = self.cfg.ddp_bucket_cap_mb
         if self.cfg.ddp_broadcast_buffers is not None:
-            training_arguments_kwargs["ddp_broadcast_buffers"] = (
+            training_arguments_kwargs[
-                self.cfg.ddp_broadcast_buffers
+                "ddp_broadcast_buffers"
-            )
+            ] = self.cfg.ddp_broadcast_buffers
 
         # these are all the "standard" kwargs that are def used
         training_arguments_kwargs["max_steps"] = (
             total_num_steps if self.cfg.max_steps else -1
         )
         training_arguments_kwargs["max_seq_length"] = self.cfg.sequence_len
-        training_arguments_kwargs["per_device_train_batch_size"] = (
+        training_arguments_kwargs[
-            self.cfg.micro_batch_size
+            "per_device_train_batch_size"
-        )
+        ] = self.cfg.micro_batch_size
         if self.cfg.eval_batch_size:
-            training_arguments_kwargs["per_device_eval_batch_size"] = (
+            training_arguments_kwargs[
-                self.cfg.eval_batch_size
+                "per_device_eval_batch_size"
-            )
+            ] = self.cfg.eval_batch_size
         if self.cfg.auto_find_batch_size is not None:
-            training_arguments_kwargs["auto_find_batch_size"] = (
+            training_arguments_kwargs[
-                self.cfg.auto_find_batch_size
+                "auto_find_batch_size"
-            )
+            ] = self.cfg.auto_find_batch_size
-        training_arguments_kwargs["gradient_accumulation_steps"] = (
+        training_arguments_kwargs[
-            self.cfg.gradient_accumulation_steps
+            "gradient_accumulation_steps"
-        )
+        ] = self.cfg.gradient_accumulation_steps
-        training_arguments_kwargs["eval_accumulation_steps"] = (
+        training_arguments_kwargs[
-            self.cfg.gradient_accumulation_steps
+            "eval_accumulation_steps"
-        )
+        ] = self.cfg.gradient_accumulation_steps
         training_arguments_kwargs["num_train_epochs"] = self.cfg.num_epochs
         training_arguments_kwargs["learning_rate"] = self.cfg.learning_rate
         training_arguments_kwargs["output_dir"] = self.cfg.output_dir
@@ -552,9 +554,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
         if self.cfg.lr_scheduler in ["one_cycle", "rex", "log_sweep"]:
             training_arguments_kwargs["lr_scheduler_type"] = "cosine"
-            training_arguments_kwargs["alternate_lr_scheduler_type"] = (
+            training_arguments_kwargs[
-                self.cfg.lr_scheduler
+                "alternate_lr_scheduler_type"
-            )
+            ] = self.cfg.lr_scheduler
         else:
             training_arguments_kwargs["lr_scheduler_type"] = (
                 self.cfg.lr_scheduler if self.cfg.lr_scheduler else "cosine"
@@ -563,9 +565,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             self.cfg.lr_scheduler_kwargs if self.cfg.lr_scheduler_kwargs else {}
         )
         training_arguments_kwargs["cosine_min_lr_ratio"] = self.cfg.cosine_min_lr_ratio
-        training_arguments_kwargs["cosine_constant_lr_ratio"] = (
+        training_arguments_kwargs[
-            self.cfg.cosine_constant_lr_ratio
+            "cosine_constant_lr_ratio"
-        )
+        ] = self.cfg.cosine_constant_lr_ratio
         training_arguments_kwargs["weight_decay"] = (
             self.cfg.weight_decay if self.cfg.weight_decay is not None else 0.0
         )
@@ -578,40 +580,40 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             self.cfg.eval_sample_packing
         )
         if self.cfg.sample_packing_bin_size is not None:
-            training_arguments_kwargs["sample_packing_bin_size"] = (
+            training_arguments_kwargs[
-                self.cfg.sample_packing_bin_size
+                "sample_packing_bin_size"
-            )
+            ] = self.cfg.sample_packing_bin_size
         if self.cfg.sample_packing_group_size is not None:
-            training_arguments_kwargs["sample_packing_group_size"] = (
+            training_arguments_kwargs[
-                self.cfg.sample_packing_group_size
+                "sample_packing_group_size"
-            )
+            ] = self.cfg.sample_packing_group_size
         if self.cfg.sample_packing_eff_est:
-            training_arguments_kwargs["sample_packing_efficiency"] = (
+            training_arguments_kwargs[
-                self.cfg.sample_packing_eff_est
+                "sample_packing_efficiency"
-            )
+            ] = self.cfg.sample_packing_eff_est
 
         if self.cfg.relora_steps:
             training_arguments_kwargs["relora_steps"] = self.cfg.relora_steps
-            training_arguments_kwargs["relora_warmup_steps"] = (
+            training_arguments_kwargs[
-                self.cfg.relora_warmup_steps
+                "relora_warmup_steps"
-            )
+            ] = self.cfg.relora_warmup_steps
             if self.cfg.relora_anneal_steps:
-                training_arguments_kwargs["relora_anneal_steps"] = (
+                training_arguments_kwargs[
-                    self.cfg.relora_anneal_steps
+                    "relora_anneal_steps"
-                )
+                ] = self.cfg.relora_anneal_steps
             if self.cfg.relora_prune_ratio:
-                training_arguments_kwargs["relora_prune_ratio"] = (
+                training_arguments_kwargs[
-                    self.cfg.relora_prune_ratio
+                    "relora_prune_ratio"
-                )
+                ] = self.cfg.relora_prune_ratio
 
         if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
             training_arguments_kwargs["lisa_n_layers"] = self.cfg.lisa_n_layers
-            training_arguments_kwargs["lisa_step_interval"] = (
+            training_arguments_kwargs[
-                self.cfg.lisa_step_interval
+                "lisa_step_interval"
-            )
+            ] = self.cfg.lisa_step_interval
-            training_arguments_kwargs["lisa_layers_attribute"] = (
+            training_arguments_kwargs[
-                self.cfg.lisa_layers_attribute
+                "lisa_layers_attribute"
-            )
+            ] = self.cfg.lisa_layers_attribute
 
         training_arguments_kwargs = self.hook_pre_create_training_args(
             training_arguments_kwargs
@@ -625,9 +627,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             )
 
         if self.cfg.neftune_noise_alpha is not None:
-            training_arguments_kwargs["neftune_noise_alpha"] = (
+            training_arguments_kwargs[
-                self.cfg.neftune_noise_alpha
+                "neftune_noise_alpha"
-            )
+            ] = self.cfg.neftune_noise_alpha
 
         trainer_kwargs = {}
 
@@ -729,42 +731,42 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 importlib.import_module("torchdistx")
 
         if self.cfg.optim_target_modules:
-            training_arguments_kwargs["optim_target_modules"] = (
+            training_arguments_kwargs[
-                self.cfg.optim_target_modules
+                "optim_target_modules"
-            )
+            ] = self.cfg.optim_target_modules
 
         training_arguments_kwargs["embedding_lr"] = self.cfg.embedding_lr
         training_arguments_kwargs["embedding_lr_scale"] = self.cfg.embedding_lr_scale
 
         training_arguments_kwargs["loraplus_lr_ratio"] = self.cfg.loraplus_lr_ratio
-        training_arguments_kwargs["loraplus_lr_embedding"] = (
+        training_arguments_kwargs[
-            self.cfg.loraplus_lr_embedding
+            "loraplus_lr_embedding"
-        )
+        ] = self.cfg.loraplus_lr_embedding
         training_arguments_kwargs["lr_groups"] = self.cfg.lr_groups
 
         if self.cfg.accelerator_config:
-            training_arguments_kwargs["accelerator_config"] = (
+            training_arguments_kwargs[
-                self.cfg.accelerator_config
+                "accelerator_config"
-            )
+            ] = self.cfg.accelerator_config
 
         if self.cfg.kd_ce_alpha is not None:
             training_arguments_kwargs["kd_ce_alpha"] = self.cfg.kd_ce_alpha
+            if self.cfg.kd_ce_alpha_end is not None:
+                training_arguments_kwargs["kd_ce_alpha_end"] = self.cfg.kd_ce_alpha_end
         if self.cfg.kd_alpha is not None:
             training_arguments_kwargs["kd_alpha"] = self.cfg.kd_alpha
+            if self.cfg.kd_alpha_end is not None:
+                training_arguments_kwargs["kd_alpha_end"] = self.cfg.kd_alpha_end
         if self.cfg.kd_temperature is not None:
             training_arguments_kwargs["kd_temperature"] = self.cfg.kd_temperature
         if self.cfg.kd_zscore_base_temp is not None:
-            training_arguments_kwargs["kd_zscore_base_temp"] = (
+            training_arguments_kwargs[
-                self.cfg.kd_zscore_base_temp
+                "kd_zscore_base_temp"
-            )
+            ] = self.cfg.kd_zscore_base_temp
         if self.cfg.kd_top_k_before_softmax is not None:
-            training_arguments_kwargs["kd_top_k_before_softmax"] = (
+            training_arguments_kwargs[
-                self.cfg.kd_top_k_before_softmax
+                "kd_top_k_before_softmax"
-            )
+            ] = self.cfg.kd_top_k_before_softmax
-
-        training_arguments_kwargs["sequence_parallel_degree"] = (
-            self.cfg.sequence_parallel_degree
-        )
 
         if self.cfg.reward_model:
             training_args_cls = AxolotlRewardConfig
@@ -849,10 +851,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         self, training_args: AxolotlTrainingArguments, is_eval=False, **kwargs
     ):
         if training_args.pretraining:
-            if (
+            if self.cfg.pretraining_sample_concatenation is False:
-                self.cfg.pretraining_sample_concatenation is False
+                return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
-                or self.cfg.micro_batch_size > 1
+            if self.cfg.micro_batch_size > 1:
-            ):
                 return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
             return None
 
@@ -880,7 +881,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             if "max_length" in kwargs:
                 kwargs.pop("max_length")
         elif use_batch_sampler_collator:
-            if self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES or (
+            if self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES:
+                collator = V2BatchSamplerDataCollatorForSeq2Seq
+            elif (
                 self.cfg.model_config_type in ["llama"]
                 and self.cfg.flash_attention is not True
             ):
@@ -911,8 +914,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 collator = DataCollatorForSeq2Seq
 
         kwargs["return_tensors"] = "pt"
-        if issubclass(collator, DataCollatorForSeq2Seq):
-            kwargs["sequence_parallel_degree"] = training_args.sequence_parallel_degree
 
         return collator(
             *collator_args,
@@ -975,32 +976,32 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             self.cfg.lr_scheduler_kwargs if self.cfg.lr_scheduler_kwargs else {}
         )
         if self.cfg.remove_unused_columns is not None:
-            training_args_kwargs["remove_unused_columns"] = (
+            training_args_kwargs[
-                self.cfg.remove_unused_columns
+                "remove_unused_columns"
-            )
+            ] = self.cfg.remove_unused_columns
         else:
             training_args_kwargs["remove_unused_columns"] = False
 
         if self.cfg.dataloader_pin_memory is not None:
-            training_args_kwargs["dataloader_pin_memory"] = (
+            training_args_kwargs[
-                self.cfg.dataloader_pin_memory
+                "dataloader_pin_memory"
-            )
+            ] = self.cfg.dataloader_pin_memory
         if self.cfg.dataloader_num_workers is not None:
-            training_args_kwargs["dataloader_num_workers"] = (
+            training_args_kwargs[
-                self.cfg.dataloader_num_workers
+                "dataloader_num_workers"
-            )
+            ] = self.cfg.dataloader_num_workers
         if self.cfg.dataloader_prefetch_factor is not None:
-            training_args_kwargs["dataloader_prefetch_factor"] = (
+            training_args_kwargs[
-                self.cfg.dataloader_prefetch_factor
+                "dataloader_prefetch_factor"
-            )
+            ] = self.cfg.dataloader_prefetch_factor
         if self.cfg.gradient_checkpointing:
-            training_args_kwargs["gradient_checkpointing"] = (
+            training_args_kwargs[
-                self.cfg.gradient_checkpointing
+                "gradient_checkpointing"
-            )
+            ] = self.cfg.gradient_checkpointing
             if self.cfg.gradient_checkpointing_kwargs is not None:
-                training_args_kwargs["gradient_checkpointing_kwargs"] = (
+                training_args_kwargs[
-                    self.cfg.gradient_checkpointing_kwargs
+                    "gradient_checkpointing_kwargs"
-                )
+                ] = self.cfg.gradient_checkpointing_kwargs
             else:
                 training_args_kwargs["gradient_checkpointing_kwargs"] = {
                     "use_reentrant": False
@@ -1074,9 +1075,9 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.dpo_use_weighting is not None:
                 training_args_kwargs["use_weighting"] = self.cfg.dpo_use_weighting
             if self.cfg.dpo_use_logits_to_keep is not None:
-                training_args_kwargs["use_logits_to_keep"] = (
+                training_args_kwargs[
-                    self.cfg.dpo_use_logits_to_keep
+                    "use_logits_to_keep"
-                )
+                ] = self.cfg.dpo_use_logits_to_keep
 
         for blocklist_key in blocklist_args_kwargs:
             if blocklist_key in training_args_kwargs:
@@ -1111,9 +1112,9 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.cfg.adapter and self.peft_config:
             dpo_trainer_kwargs["peft_config"] = self.peft_config
         if self.cfg.precompute_ref_log_probs is not None:
-            dpo_trainer_kwargs["precompute_ref_log_probs"] = (
+            dpo_trainer_kwargs[
-                self.cfg.precompute_ref_log_probs
+                "precompute_ref_log_probs"
-            )
+            ] = self.cfg.precompute_ref_log_probs
         if self.cfg.rl == "grpo":
             trainer_cls = GRPOStrategy.get_trainer_class()
             trainer_cls_args = [self.model]

src/axolotl/core/trainers/__init__.py

@@ -1,18 +0,0 @@
-"""Init for axolotl.core.trainers"""
-
-# pylint: disable=unused-import
-# flake8: noqa
-
-from .base import AxolotlTrainer
-from .dpo.trainer import AxolotlDPOTrainer
-from .grpo.trainer import AxolotlGRPOTrainer
-from .mamba import AxolotlMambaTrainer
-from .relora import ReLoRATrainer
-from .trl import (
-    AxolotlCPOTrainer,
-    AxolotlKTOTrainer,
-    AxolotlORPOTrainer,
-    AxolotlPRMTrainer,
-    AxolotlRewardTrainer,
-    TRLPPOTrainer,
-)

src/axolotl/core/trainers/base.py

@@ -1,47 +1,365 @@
-"""Module for customized trainers"""
+"""
-
+module for customized trainers
-# pylint: disable=too-many-lines
+"""
 
 from __future__ import annotations
 
+# pylint: disable=too-many-lines
 import logging
 import os
 from collections import defaultdict
 from functools import wraps
-from typing import Any, Literal
+from typing import Dict, Literal, Optional
 
-import datasets
 import torch
 from datasets import Dataset
+from peft.optimizers import create_loraplus_optimizer
 from torch import nn
-from torch.utils.data import (
+from torch.optim.lr_scheduler import OneCycleLR
-    BatchSampler,
+from torch.utils.data import BatchSampler, DataLoader, RandomSampler, SequentialSampler
-    DataLoader,
-    RandomSampler,
-    Sampler,
-    SequentialSampler,
-)
 from transformers import Trainer
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, seed_worker
+from transformers.utils import is_sagemaker_mp_enabled
+from trl import CPOTrainer, KTOTrainer, ORPOTrainer, PRMTrainer, RewardTrainer
 from trl.trainer.utils import pad_to_length
-from typing_extensions import override
 
-from axolotl.core.trainers.mixins import (
+from axolotl.integrations.base import BaseOptimizerFactory
-    OptimizerMixin,
+from axolotl.monkeypatch.relora import ReLoRAScheduler
-    SchedulerMixin,
-    SequenceParallelMixin,
-)
-from axolotl.core.trainers.utils import (
-    sanitize_kwargs_for_ds_tagging,
-    sanitize_kwargs_for_tagging,
-)
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
+from axolotl.utils.schedulers import (
+    RexLR,
+    get_cosine_schedule_with_min_lr,
+    get_cosine_schedule_with_quadratic_warmup,
+    get_cosine_schedule_with_warmup_decay_constant,
+)
 
-LOG = logging.getLogger(__name__)
+if is_sagemaker_mp_enabled():
+    import smdistributed.modelparallel.torch as smp
+
+LOG = logging.getLogger("axolotl.core.trainer_builder")
 
 
-class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trainer):
+def _sanitize_kwargs_for_tagging(tag_names, kwargs=None):
-    """Extend the base Trainer for axolotl helpers"""
+    if isinstance(tag_names, str):
+        tag_names = [tag_names]
+
+    if kwargs is not None:
+        if "tags" not in kwargs:
+            kwargs["tags"] = tag_names
+        elif "tags" in kwargs and isinstance(kwargs["tags"], list):
+            kwargs["tags"].extend(tag_names)
+        elif "tags" in kwargs and isinstance(kwargs["tags"], str):
+            tag_names.append(kwargs["tags"])
+            kwargs["tags"] = tag_names
+
+    return kwargs
+
+
+def _sanitize_kwargs_for_ds_tagging(dataset_tags, kwargs=None):
+    if isinstance(dataset_tags, str):
+        dataset_tags = [dataset_tags]
+
+    if (dataset_tags is not None) and (kwargs is not None):
+        if "dataset_tags" not in kwargs:
+            kwargs["dataset_tags"] = dataset_tags
+        elif "dataset_tags" in kwargs and isinstance(kwargs["dataset_tags"], list):
+            kwargs["dataset_tags"].extend(dataset_tags)
+        elif "dataset_tags" in kwargs and isinstance(kwargs["dataset_tags"], str):
+            dataset_tags.append(kwargs["dataset_tags"])
+            kwargs["dataset_tags"] = dataset_tags
+
+    return kwargs
+
+
+class SchedulerMixin(Trainer):
+    """
+    Mixin class for scheduler setup in CausalTrainer.
+    """
+
+    args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
+
+    def create_scheduler(
+        self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
+    ):
+        """
+        Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or
+        passed as an argument.
+
+        Args:
+            num_training_steps (int): The number of training steps to do.
+            optimizer (torch.optim.Optimizer): The training optimizer
+        """
+        use_cosine_quadratic = (
+            self.args.lr_scheduler_type == "cosine"
+            and self.args.lr_quadratic_warmup is True
+        )
+
+        use_cosine_min_lr = (
+            self.args.lr_scheduler_type == "cosine"
+            and self.args.cosine_min_lr_ratio is not None
+        )
+
+        # fmt: off
+        if self.lr_scheduler is None:  # type: ignore  # pylint: disable=access-member-before-definition
+            # fmt: on
+            if self.args.alternate_lr_scheduler_type == "one_cycle":
+                num_warmup_steps = self.args.get_warmup_steps(num_training_steps)
+                pct_start = num_warmup_steps / num_training_steps
+                extra_lr_kwargs = {}
+                if "pct_start" not in self.args.lr_scheduler_kwargs:
+                    extra_lr_kwargs["pct_start"] = pct_start
+                if "anneal_strategy" not in self.args.lr_scheduler_kwargs:
+                    extra_lr_kwargs["anneal_strategy"] = "cos"
+
+                self.lr_scheduler = OneCycleLR(
+                    optimizer,
+                    max_lr=self.args.learning_rate,
+                    total_steps=num_training_steps,
+                    **extra_lr_kwargs,
+                    **self.args.lr_scheduler_kwargs,
+                )
+            elif self.args.alternate_lr_scheduler_type == "rex":
+                if use_cosine_min_lr:
+                    assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
+
+                self.lr_scheduler = RexLR(
+                    optimizer=optimizer,
+                    max_lr=self.args.learning_rate,
+                    min_lr=0 if not use_cosine_min_lr else (self.args.learning_rate * self.args.cosine_min_lr_ratio),
+                    total_steps=num_training_steps,
+                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
+                )
+            elif use_cosine_quadratic:
+                if use_cosine_min_lr:
+                    LOG.warning("Both cosine quadratic warmup and min lr detected. Using quadratic warmup.")
+
+                self.lr_scheduler = get_cosine_schedule_with_quadratic_warmup(  # pylint: disable=attribute-defined-outside-init
+                    optimizer,
+                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
+                    num_training_steps=num_training_steps,
+                )
+            elif self.args.cosine_min_lr_ratio and self.args.cosine_constant_lr_ratio and use_cosine_min_lr:
+                assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
+                assert 0 <= self.args.cosine_constant_lr_ratio <= 1.0, "cosine_constant_lr_ratio must be between 0.0 and 1.0"
+                self.lr_scheduler = get_cosine_schedule_with_warmup_decay_constant(  # pylint: disable=attribute-defined-outside-init
+                    optimizer,
+                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
+                    num_training_steps=num_training_steps,
+                    min_lr_ratio=self.args.cosine_min_lr_ratio,
+                    constant_lr_ratio=self.args.cosine_constant_lr_ratio,
+                )
+            elif self.args.cosine_min_lr_ratio and use_cosine_min_lr:
+                assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
+                self.lr_scheduler = get_cosine_schedule_with_min_lr(  # pylint: disable=attribute-defined-outside-init
+                    optimizer,
+                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
+                    num_training_steps=num_training_steps,
+                    min_lr_ratio=self.args.cosine_min_lr_ratio,
+                )
+            else:
+                return super().create_scheduler(num_training_steps, optimizer=optimizer)
+        else:
+            if use_cosine_quadratic:
+                LOG.warning("axolotl's cosine scheduler with quadratic warmup not used (e.g., because of deepspeed).")
+
+            if use_cosine_min_lr:
+                LOG.warning("axolotl's cosine scheduler with min lr not used (e.g., because of deepspeed).")
+
+        return self.lr_scheduler
+
+
+class OptimizerMixin(Trainer):
+    """
+    Mixin class for shared handling of building custom optimizers
+    """
+
+    args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
+
+    def create_optimizer_grouped_parameters(
+        self, opt_model, optimizer_kwargs
+    ) -> list[dict]:
+        decay_parameters = self.get_decay_parameter_names(opt_model)
+        params: dict = {
+            "to_weight_decay": {},  # LayerNorm and bias
+            "embeddings": {},  # lm_head, embed_tokens,
+            "no_weight_decay": {},
+        }
+        lr_groups_lookup = {}
+        lr_groups_learning_rates = {}
+        if self.args.lr_groups:
+            for lr_group in self.args.lr_groups:
+                group_name = lr_group["name"]
+                group_modules = lr_group["modules"]
+                for module in group_modules:
+                    lr_groups_lookup[module] = group_name
+                lr_groups_learning_rates[group_name] = lr_group["lr"]
+                params[f"to_weight_decay_{group_name}"] = {}
+
+        for name, param in opt_model.named_parameters():
+            if not param.requires_grad:
+                continue
+            if name.endswith("modules_to_save.default.weight") or any(
+                embed_name in name for embed_name in ["embed_tokens", "lm_head"]
+            ):
+                params["embeddings"][name] = param
+            elif name in decay_parameters:
+                lr_group_modules = [
+                    group_modules
+                    for group_modules in lr_groups_lookup
+                    if group_modules in name
+                ]
+                if lr_groups_lookup and any(lr_group_modules):
+                    lr_group_module = lr_group_modules[0]
+                    group_name = lr_groups_lookup[lr_group_module]
+                    params[f"to_weight_decay_{group_name}"][name] = param
+                else:
+                    params["to_weight_decay"][name] = param
+            else:
+                params["no_weight_decay"][name] = param
+        optimizer_grouped_parameters = []
+        if params["to_weight_decay"]:
+            optimizer_grouped_parameters.append(
+                {
+                    "params": list(params["to_weight_decay"].values()),
+                    "weight_decay": self.args.weight_decay,
+                    "lr": optimizer_kwargs["lr"],
+                }
+            )
+        if params["embeddings"]:
+            lr = optimizer_kwargs["lr"]  # pylint: disable=invalid-name
+            if self.args.embedding_lr_scale:
+                lr *= self.args.embedding_lr_scale  # pylint: disable=invalid-name
+            elif self.args.embedding_lr:
+                lr = self.args.embedding_lr  # pylint: disable=invalid-name
+            optimizer_grouped_parameters.append(
+                {
+                    "params": list(params["embeddings"].values()),
+                    "weight_decay": 0.0,
+                    "lr": lr,
+                }
+            )
+        if params["no_weight_decay"]:
+            optimizer_grouped_parameters.append(
+                {
+                    "params": list(params["no_weight_decay"].values()),
+                    "weight_decay": 0.0,
+                    "lr": optimizer_kwargs["lr"],
+                }
+            )
+        for group_name, group_lr in lr_groups_learning_rates.items():
+            if params[f"to_weight_decay_{group_name}"]:
+                optimizer_grouped_parameters.append(
+                    {
+                        "params": list(
+                            params[f"to_weight_decay_{group_name}"].values()
+                        ),
+                        "weight_decay": self.args.weight_decay,
+                        "lr": group_lr,
+                    }
+                )
+
+        return optimizer_grouped_parameters
+
+    def create_optimizer(self):
+        if (
+            self.args.loraplus_lr_ratio is None
+            and self.args.embedding_lr_scale is None
+            and self.args.embedding_lr is None
+            and self.args.lr_groups is None
+            and self.optimizer_cls_and_kwargs is None
+        ):
+            return super().create_optimizer()
+
+        opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
+
+        if (
+            not self.optimizer
+            and self.optimizer_cls_and_kwargs is not None
+            and issubclass(self.optimizer_cls_and_kwargs[0], BaseOptimizerFactory)
+        ):
+            optimizer_factory_cls, optimizer_kwargs = self.optimizer_cls_and_kwargs
+            self.optimizer = optimizer_factory_cls()(
+                opt_model, self.args, **optimizer_kwargs
+            )
+
+        if not self.optimizer:
+            if self.optimizer_cls_and_kwargs is not None:
+                optimizer_cls, optimizer_kwargs = self.optimizer_cls_and_kwargs
+            else:
+                optimizer_cls, optimizer_kwargs = self.get_optimizer_cls_and_kwargs(
+                    self.args, opt_model
+                )
+
+            optimizer_grouped_parameters = self.create_optimizer_grouped_parameters(
+                opt_model, optimizer_kwargs
+            )
+
+            if self.args.loraplus_lr_ratio is not None:
+                loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
+                loraplus_lr_embedding = getattr(
+                    self.args, "loraplus_lr_embedding", 1e-6
+                )
+                self.optimizer = create_loraplus_optimizer(  # pylint: disable=attribute-defined-outside-init
+                    opt_model,
+                    optimizer_cls,
+                    loraplus_lr_ratio=loraplus_lr_ratio,
+                    loraplus_lr_embedding=loraplus_lr_embedding,
+                    **optimizer_kwargs,
+                )
+            else:
+                # Overwrite `params` in case it's created by `get_optimizer_cls_and_kwargs`
+                # e.g. for GaLore optimizer.
+                if "params" in optimizer_kwargs:
+                    optimizer_grouped_parameters = optimizer_kwargs.pop("params")
+
+                # Overwrite `model` in case it's created by `get_optimizer_cls_and_kwargs`
+                # e.g. for LOMO optimizer.
+                if "model" in optimizer_kwargs:
+                    optimizer_grouped_parameters = optimizer_kwargs.pop("model")
+
+                # For layer-wise dummy optimizers we overwrite optimizer_grouped_parameters with `optimizer_dict`
+                # to avoid arguments conflicts.
+                if "optimizer_dict" in optimizer_kwargs:
+                    optimizer_grouped_parameters = optimizer_kwargs.pop(
+                        "optimizer_dict"
+                    )
+
+                self.optimizer = optimizer_cls(
+                    optimizer_grouped_parameters, **optimizer_kwargs
+                )
+
+            if optimizer_cls.__name__ == "Adam8bit":
+                import bitsandbytes
+
+                manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
+
+                skipped = 0
+                for module in opt_model.modules():
+                    if isinstance(module, nn.Embedding):
+                        skipped += sum(
+                            {
+                                p.data_ptr(): p.numel() for p in module.parameters()
+                            }.values()
+                        )
+                        LOG.info(f"skipped {module}: {skipped/2**20}M params")
+                        manager.register_module_override(
+                            module, "weight", {"optim_bits": 32}
+                        )
+                        LOG.debug(f"bitsandbytes: will optimize {module} in fp32")
+                LOG.info(f"skipped: {skipped/2**20}M params")
+
+        if is_sagemaker_mp_enabled():
+            self.optimizer = smp.DistributedOptimizer(  # pylint: disable=attribute-defined-outside-init
+                self.optimizer
+            )
+
+        return self.optimizer
+
+
+class AxolotlTrainer(SchedulerMixin, OptimizerMixin, Trainer):
+    """
+    Extend the base Trainer for axolotl helpers
+    """
 
     args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
     tag_names = ["axolotl"]
@@ -58,18 +376,12 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trai
         self.eval_data_collator = eval_data_collator
         self.dataset_tags = dataset_tags
         self._signature_columns = None  # workaround for pylint
-
         super().__init__(*_args, **kwargs)
-
         self.train_data_collator = self.data_collator
         self._stored_metrics = defaultdict(lambda: defaultdict(list))
         if self.args.orpo_alpha:
             self.loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
 
-        # Initialize sequence parallelism if enabled
-        if self.args.sequence_parallel_degree > 1:
-            self._setup_sequence_parallel()
-
     def _wrap_model(self, model, training=True, dataloader=None):
         if self.args.torch_compile:
             torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
@@ -82,20 +394,8 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trai
             )
         return super()._wrap_model(model, training=training, dataloader=dataloader)
 
-    def _create_multipack_sampler(
+    def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
-        self, base_sampler: Sampler, dataset: Dataset
+        if self.args.sample_packing and not self.args.pretraining:
-    ) -> MultipackBatchSampler:
-        """
-        Helper method to create a `MultipackBatchSampler` for multipacking sequences
-        for training.
-
-        Args:
-            base_sampler: Sampler to wrap with `MultipackBatchSampler`.
-            dataset: Dataset to sample from.
-
-        Returns:
-            Multipack (sample packing) batch sampler.
-        """
             if self.args.multipack_real_batches:
                 batch_size = self.args.per_device_train_batch_size
                 batch_max_len = self.args.max_seq_length
@@ -106,223 +406,130 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trai
                 )
                 batch_max_len = train_batch_size * self.args.max_seq_length
 
+            if self.args.curriculum_sampling:
+                sampler = SequentialSampler(self.train_dataset)
+            else:
+                sampler = RandomSampler(self.train_dataset)
+
             return MultipackBatchSampler(
-            base_sampler,
+                sampler,
-            lengths=get_dataset_lengths(dataset),
+                lengths=get_dataset_lengths(self.train_dataset),
                 packing_efficiency_estimate=self.args.sample_packing_efficiency,
                 batch_max_len=batch_max_len,
                 batch_size=batch_size,
+                group_size=self.args.sample_packing_group_size,
+                bin_size=self.args.sample_packing_bin_size,
                 drop_last=True,
             )
-
+        if self.args.curriculum_sampling:
-    def _get_train_sampler(self) -> Sampler | None:
+            return SequentialSampler(self.train_dataset)
-        """
-        Helper method to get the sampler for training. Handles cases for sequence
-        parallelism, sample packing, and curriculum sampling (sequential).
-
-        Returns:
-            If the dataset is non-empty, a sampler is returned, the type of which
-                depends on the passed training args.
-        """
-        use_sample_packing = self.args.sample_packing and not self.args.pretraining
-
-        # Determine the base sampler first
-        if self.args.sequence_parallel_degree > 1:
-            base_sampler = self._sp_get_train_sampler(self.train_dataset)
-        elif self.args.curriculum_sampling:
-            base_sampler = SequentialSampler(self.train_dataset)
-        elif use_sample_packing:
-            base_sampler = RandomSampler(self.train_dataset)
-        else:
-            # Default to parent class implementation for standard random sampling
         return super()._get_train_sampler()
 
-        # Apply multipack wrapper if needed
+    def _get_eval_sampler(
-        if use_sample_packing:
+        self, eval_dataset: Dataset
-            return self._create_multipack_sampler(
+    ) -> Optional[torch.utils.data.Sampler]:
-                base_sampler=base_sampler,
+        if self.args.sample_packing and self.args.eval_sample_packing is not False:
-                dataset=self.train_dataset,
+            if self.args.multipack_real_batches:
-            )
+                batch_size = self.args.per_device_eval_batch_size
-
+                batch_max_len = self.args.max_seq_length
-        return base_sampler
-
-    def _get_eval_sampler(self, eval_dataset: Dataset | None = None) -> Sampler | None:
-        """
-        Helper method to get the sampler for evaluation. Handles sequence parallelism
-        and sample packing cases.
-
-        Returns:
-            If the dataset is non-empty, a sampler is returned, the type of which
-                depends on the passed training args.
-        """
-        eval_dataset = eval_dataset if eval_dataset is not None else self.eval_dataset
-
-        # Multipacking enabled if training is enabled and eval is not explicitly disabled
-        use_multipack = (
-            self.args.sample_packing and self.args.eval_sample_packing is not False
-        )
-
-        # Determine the base sampler
-        if self.args.sequence_parallel_degree > 1:
-            base_sampler = self._sp_get_eval_sampler(eval_dataset)
-        elif use_multipack:
-            base_sampler = SequentialSampler(eval_dataset)
             else:
+                batch_size = 1
+                batch_max_len = (
+                    self.args.per_device_eval_batch_size * self.args.max_seq_length
+                )
+            return MultipackBatchSampler(
+                SequentialSampler(eval_dataset),
+                lengths=get_dataset_lengths(self.eval_dataset),
+                packing_efficiency_estimate=self.args.sample_packing_efficiency,
+                batch_max_len=batch_max_len,
+                batch_size=batch_size,
+                group_size=self.args.sample_packing_group_size,
+                bin_size=self.args.sample_packing_bin_size,
+                drop_last=True,
+            )
         return super()._get_eval_sampler(eval_dataset)
 
-        # Apply multipack wrapper if needed
+    def get_train_dataloader(self) -> DataLoader:
-        if use_multipack:
+        if self.args.sample_packing and not self.args.pretraining:
-            return self._create_multipack_sampler(
+            train_dataset = self.train_dataset
-                base_sampler=base_sampler,
+            if "length" in train_dataset.features.keys():
-                dataset=eval_dataset,
+                train_dataset = train_dataset.remove_columns(["length"])
-            )
+            data_collator = self.data_collator
-
+            dataloader_params = {
-        return base_sampler
+                "batch_size": self._train_batch_size,
-
+                "collate_fn": data_collator,
-    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
+                dataloader_params[
+                    "prefetch_factor"
+                ] = self.args.dataloader_prefetch_factor
 
-        return params
+            sampler = self._get_train_sampler()
-
-    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 self.accelerator.prepare_data_loader(
-        # Return unprepared dataloader if using sequence parallelism
+                DataLoader(train_dataset, **dataloader_params)
-        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
-        data_collator = self.data_collator  # type: ignore
-
-        # Handle dataset preprocessing
-        if isinstance(train_dataset, datasets.Dataset):
-            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",
             )
+        return super().get_train_dataloader()
 
-        # Get sampler and create dataloader
+    def get_eval_dataloader(self, eval_dataset: Optional[Dataset] = None) -> DataLoader:
-        sampler = self._get_train_sampler()
-        return self._prepare_dataloader(train_dataset, sampler, is_eval=False)
-
-    def get_eval_dataloader(self, eval_dataset: Dataset | None = None) -> DataLoader:
-        """Get dataloader for evaluation"""
-        eval_dataset = eval_dataset if eval_dataset is not None else self.eval_dataset
-
-        # Handle special case: sample packing is enabled but eval_sample_packing is False
         if self.args.sample_packing and self.args.eval_sample_packing is False:
             self.data_collator = (  # pylint: disable=attribute-defined-outside-init
                 self.eval_data_collator
             )
-            if "length" in eval_dataset.column_names:
+            if eval_dataset:
                 eval_dataset = eval_dataset.remove_columns(["length"])
             dataloader = super().get_eval_dataloader(eval_dataset)
             self.data_collator = (  # pylint: disable=attribute-defined-outside-init
                 self.train_data_collator
             )
-
             return dataloader
 
-        # Handle sample packing or sequence parallelism
+        if self.args.sample_packing and self.args.eval_sample_packing is not False:
-        if (
+            eval_dataset = (
-            self.args.sample_packing
+                eval_dataset if eval_dataset is not None else self.eval_dataset
-            and self.args.eval_sample_packing is not False
-            or self.args.sequence_parallel_degree > 1
-        ):
-            # Get appropriate data collator
-            self.data_collator = (  # pylint: disable=attribute-defined-outside-init
-                self.eval_data_collator
-                if hasattr(self, "eval_data_collator") and self.eval_data_collator
-                else self.data_collator
             )
-            if "length" in eval_dataset.column_names:
+
+            eval_sampler = self._get_eval_sampler(eval_dataset)
             eval_dataset = eval_dataset.remove_columns(["length"])
+            data_collator = self.data_collator
+            dataloader_params = {
+                "batch_size": self.args.eval_batch_size,
+                "collate_fn": data_collator,
+                "num_workers": self.args.dataloader_num_workers,
+                "pin_memory": self.args.dataloader_pin_memory,
+            }
+            if self.args.dataloader_prefetch_factor:
+                dataloader_params[
+                    "prefetch_factor"
+                ] = self.args.dataloader_prefetch_factor
 
-            # Handle dataset preprocessing for SP
+            if isinstance(eval_sampler, BatchSampler):
-            if self.args.sequence_parallel_degree > 1:
+                dataloader_params["batch_sampler"] = eval_sampler
-                if isinstance(eval_dataset, datasets.Dataset):
+                del dataloader_params["batch_size"]
-                    eval_dataset = self._remove_unused_columns(
-                        eval_dataset, description="evaluation"
-                    )
             else:
-                    self.data_collator = self._get_collator_with_removed_columns(  # pylint: disable=attribute-defined-outside-init
+                dataloader_params["sampler"] = eval_sampler
-                        self.data_collator, description="evaluation"
+                dataloader_params["drop_last"] = self.args.dataloader_drop_last
-                    )
 
-            # Use eval_batch_size for sample packing, per_device_eval_batch_size otherwise
+            self.accelerator.even_batches = False
-            batch_size = (
+            return self.accelerator.prepare_data_loader(
-                self.args.eval_batch_size
+                DataLoader(eval_dataset, **dataloader_params)
-                if self.args.sample_packing
-                else self.args.per_device_eval_batch_size
             )
-            sampler = self._get_eval_sampler(eval_dataset)
-            dataloader = self._prepare_dataloader(
-                eval_dataset, sampler, is_eval=True, custom_batch_size=batch_size
-            )
-
-            return dataloader
 
         return super().get_eval_dataloader(eval_dataset)
 
     def _get_bench_sampler(
         self, bench_dataset: Dataset
-    ) -> torch.utils.data.Sampler | None:
+    ) -> Optional[torch.utils.data.Sampler]:
         if self.args.world_size <= 1:
             return SequentialSampler(bench_dataset)
         return None
@@ -347,7 +554,6 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trai
         return DataLoader(bench_dataset, **dataloader_params)
         # return self.accelerator.prepare(DataLoader(bench_dataset, **dataloader_params))
 
-    @override
     def compute_loss(
         self, model, inputs, return_outputs=False, num_items_in_batch=None
     ):
@@ -364,7 +570,6 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trai
                 return_outputs=return_outputs,
                 num_items_in_batch=num_items_in_batch,
             )
-
         return super().compute_loss(
             model,
             inputs,
@@ -539,10 +744,10 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trai
         Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
         model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
         """
-        kwargs = sanitize_kwargs_for_ds_tagging(
+        kwargs = _sanitize_kwargs_for_ds_tagging(
             dataset_tags=self.dataset_tags, kwargs=kwargs
         )
-        kwargs = sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
+        kwargs = _sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
 
         return super().push_to_hub(*args, **kwargs)
 
@@ -559,13 +764,15 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trai
 
         return res
 
-    def log(self, logs: dict[str, float], start_time: float | None = None) -> None:
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
         """
         Log `logs` on the various objects watching training, including stored metrics.
 
         Args:
-            logs: The values to log.
+            logs (`Dict[str, float]`):
-            start_time: The start of training.
+                The values to log.
+            start_time (`Optional[float]`):
+                The start of training.
         """
         # logs either has 'loss' or 'eval_loss'
         train_eval = "train" if "loss" in logs else "eval"
@@ -577,7 +784,7 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trai
         return super().log(logs, start_time)
 
     def store_metrics(
-        self, metrics: dict[str, float], train_eval: Literal["train", "eval"] = "train"
+        self, metrics: Dict[str, float], train_eval: Literal["train", "eval"] = "train"
     ) -> None:
         for key, value in metrics.items():
             self._stored_metrics[train_eval][key].append(value)
@@ -590,26 +797,110 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trai
         os.makedirs(output_dir, exist_ok=True)
         return super()._save_checkpoint(model, trial, **kwargs)
 
-    def training_step(
+
+class AxolotlMambaTrainer(AxolotlTrainer):
+    """
+    Mamba specific trainer to handle loss calculation
+    """
+
+    tag_names = ["axolotl", "mamba"]
+
+    def compute_loss(
         self,
-        model: nn.Module,
+        model,
-        inputs: dict[str, torch.Tensor | Any],
+        inputs,
-        num_items_in_batch: int | None = None,
+        return_outputs=False,  # pylint: disable=unused-argument
-    ) -> torch.Tensor:
+        num_items_in_batch=None,  # pylint: disable=unused-argument
+    ):
+        input_ids = inputs.pop("input_ids")
+        lm_logits = model(input_ids).logits
+
+        labels = input_ids.to(lm_logits.device)
+        shift_logits = lm_logits[:, :-1, :].contiguous()
+        labels = labels[:, 1:].contiguous()
+
+        loss_fct = torch.nn.CrossEntropyLoss()
+        lm_loss = loss_fct(
+            shift_logits.view(-1, shift_logits.size(-1)), labels.view(-1)
+        )
+
+        return lm_loss
+
+
+class ReLoRATrainer(AxolotlTrainer):
     """
-        Perform a training step on a batch of inputs. Overrides the
+    Trainer subclass that uses the OneCycleLR scheduler
-        `transformers.trainer.Trainer` method to handle sequence parallelism if
-        enabled.
-
-        Args:
-            model: Model to perform training step for.
-            inputs: Dictionary mapping.
     """
-        # Set up sequence parallelism for this step if enabled
-        if self.args.sequence_parallel_degree > 1:
-            self._update_ring_flash_attn_params(inputs)
 
-        # Proceed with normal training step
+    tag_names = ["axolotl", "relora"]
-        loss = super().training_step(model, inputs, num_items_in_batch)
 
-        return loss
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.lr_scheduler = None
+
+    def create_scheduler(
+        self,
+        num_training_steps: int,
+        optimizer: Optional[torch.optim.Optimizer] = None,
+    ):
+        optimizer = self.optimizer if optimizer is None else optimizer
+        lr_scheduler = super().create_scheduler(num_training_steps, optimizer)
+
+        if self.args.relora_steps:
+            warmup_steps = (
+                self.args.relora_warmup_steps if self.args.relora_warmup_steps else 10
+            )
+            anneal_steps = (
+                self.args.relora_anneal_steps if self.args.relora_anneal_steps else 1
+            )
+            self.lr_scheduler = ReLoRAScheduler(
+                optimizer,
+                lr_scheduler,
+                self.args.relora_steps,
+                anneal_steps,
+                warmup_steps,
+            )
+        else:
+            self.lr_scheduler = lr_scheduler
+
+        return self.lr_scheduler
+
+
+class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
+    """
+    Extend the base ORPOTrainer for axolotl helpers
+    """
+
+    tag_names = ["axolotl", "orpo"]
+
+
+class AxolotlKTOTrainer(SchedulerMixin, KTOTrainer):
+    """
+    Extend the base KTOTrainer for axolotl helpers
+    """
+
+    tag_names = ["axolotl", "kto"]
+
+
+class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
+    """
+    Extend the base CPOTrainer for axolotl helpers
+    """
+
+    tag_names = ["axolotl", "cpo"]
+
+
+class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
+    """
+    Extend the base RewardTrainer for axolotl helpers
+    """
+
+    tag_names = ["axolotl", "reward"]
+
+
+class AxolotlPRMTrainer(SchedulerMixin, PRMTrainer):
+    """
+    Extend the base trl.PRMTrainer for axolotl helpers
+    """
+
+    tag_names = ["axolotl", "prm"]

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

@@ -1,7 +1,6 @@
 """
 DPO Specific Strategy for training
 """
-
 from axolotl.core.trainers.dpo.trainer import AxolotlDPOTrainer
 
 

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

@@ -1,7 +1,6 @@
 """
 Axolotl specific DPO args
 """
-
 from dataclasses import dataclass
 
 from trl import DPOConfig

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

@@ -1,7 +1,6 @@
 """
 DPO trainer for axolotl
 """
-
 import gc
 from functools import wraps
 from typing import Any, Dict, Union
@@ -13,10 +12,10 @@ from transformers import Trainer
 from transformers.utils import is_sagemaker_mp_enabled
 from trl import DPOTrainer
 
-from axolotl.core.trainers.mixins import SchedulerMixin
+from axolotl.core.trainers.base import (
-from axolotl.core.trainers.utils import (
+    SchedulerMixin,
-    sanitize_kwargs_for_ds_tagging,
+    _sanitize_kwargs_for_ds_tagging,
-    sanitize_kwargs_for_tagging,
+    _sanitize_kwargs_for_tagging,
 )
 
 if is_sagemaker_mp_enabled():
@@ -74,10 +73,10 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
         Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
         model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
         """
-        kwargs = sanitize_kwargs_for_ds_tagging(
+        kwargs = _sanitize_kwargs_for_ds_tagging(
             dataset_tags=self.dataset_tags, kwargs=kwargs
         )
-        kwargs = sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
+        kwargs = _sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
 
         return super().push_to_hub(*args, **kwargs)
 

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

@@ -9,7 +9,7 @@ import logging
 from trl.trainer.grpo_trainer import RewardFunc
 
 from axolotl.core.trainers.grpo.trainer import AxolotlGRPOTrainer
-from axolotl.utils.schemas.trl import TRLConfig
+from axolotl.utils.config.models.input.v0_4_1.trl import TRLConfig
 
 LOG = logging.getLogger("axolotl")
 
@@ -45,9 +45,9 @@ class GRPOStrategy:
             )
 
             if trl.vllm_gpu_memory_utilization:
-                grpo_args_kwargs["vllm_gpu_memory_utilization"] = (
+                grpo_args_kwargs[
-                    trl.vllm_gpu_memory_utilization
+                    "vllm_gpu_memory_utilization"
-                )
+                ] = trl.vllm_gpu_memory_utilization
 
             if trl.vllm_max_model_len:
                 grpo_args_kwargs["vllm_max_model_len"] = trl.vllm_max_model_len
@@ -86,9 +86,9 @@ class GRPOStrategy:
     def set_trainer_kwargs(cls, cfg):
         trainer_kwargs = {}
         if cfg.trl and cfg.trl.reward_processing_classes:
-            trainer_kwargs["reward_processing_classes"] = (
+            trainer_kwargs[
-                cfg.trl.reward_processing_classes
+                "reward_processing_classes"
-            )
+            ] = cfg.trl.reward_processing_classes
         return trainer_kwargs
 
     @classmethod

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

@@ -1,7 +1,6 @@
 """
 Axolotl Specific Training Args
 """
-
 from dataclasses import dataclass
 
 from trl import GRPOConfig

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

@@ -1,7 +1,6 @@
 """
 Axolotl GRPO trainer
 """
-
 from accelerate.utils import is_peft_model
 from accelerate.utils.other import is_compiled_module
 from transformers import PreTrainedModel

src/axolotl/core/trainers/mamba.py

@@ -1,32 +0,0 @@
-"""Module for mamba trainer"""
-
-import torch
-
-from axolotl.core.trainers.base import AxolotlTrainer
-
-
-class AxolotlMambaTrainer(AxolotlTrainer):
-    """Mamba specific trainer to handle loss calculation"""
-
-    tag_names = ["axolotl", "mamba"]
-
-    def compute_loss(
-        self,
-        model,
-        inputs,
-        return_outputs=False,  # pylint: disable=unused-argument
-        num_items_in_batch=None,  # pylint: disable=unused-argument
-    ):
-        input_ids = inputs.pop("input_ids")
-        lm_logits = model(input_ids).logits
-
-        labels = input_ids.to(lm_logits.device)
-        shift_logits = lm_logits[:, :-1, :].contiguous()
-        labels = labels[:, 1:].contiguous()
-
-        loss_fct = torch.nn.CrossEntropyLoss()
-        lm_loss = loss_fct(
-            shift_logits.view(-1, shift_logits.size(-1)), labels.view(-1)
-        )
-
-        return lm_loss

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

@@ -1,8 +0,0 @@
-"""Init for axolotl.core.trainers.mixins"""
-
-# pylint: disable=unused-import
-# flake8: noqa
-
-from .optimizer import OptimizerMixin
-from .scheduler import SchedulerMixin
-from .sequence_parallel import SequenceParallelMixin

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

@@ -1,201 +0,0 @@
-"""Module for Axolotl trainer optimizer mixin"""
-
-import logging
-
-from peft.optimizers import create_loraplus_optimizer
-from torch import nn
-from transformers.trainer import Trainer
-from transformers.utils import is_sagemaker_mp_enabled
-
-from axolotl.integrations.base import BaseOptimizerFactory
-
-if is_sagemaker_mp_enabled():
-    import smdistributed.modelparallel.torch as smp
-
-LOG = logging.getLogger(__name__)
-
-
-class OptimizerMixin(Trainer):
-    """Mixin class for shared handling of building custom optimizers"""
-
-    args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
-
-    def create_optimizer_grouped_parameters(
-        self, opt_model, optimizer_kwargs
-    ) -> list[dict]:
-        decay_parameters = self.get_decay_parameter_names(opt_model)
-        params: dict = {
-            "to_weight_decay": {},  # LayerNorm and bias
-            "embeddings": {},  # lm_head, embed_tokens,
-            "no_weight_decay": {},
-        }
-        lr_groups_lookup = {}
-        lr_groups_learning_rates = {}
-        if self.args.lr_groups:
-            for lr_group in self.args.lr_groups:
-                group_name = lr_group["name"]
-                group_modules = lr_group["modules"]
-                for module in group_modules:
-                    lr_groups_lookup[module] = group_name
-                lr_groups_learning_rates[group_name] = lr_group["lr"]
-                params[f"to_weight_decay_{group_name}"] = {}
-
-        for name, param in opt_model.named_parameters():
-            if not param.requires_grad:
-                continue
-            if name.endswith("modules_to_save.default.weight") or any(
-                embed_name in name for embed_name in ["embed_tokens", "lm_head"]
-            ):
-                params["embeddings"][name] = param
-            elif name in decay_parameters:
-                lr_group_modules = [
-                    group_modules
-                    for group_modules in lr_groups_lookup
-                    if group_modules in name
-                ]
-                if lr_groups_lookup and any(lr_group_modules):
-                    lr_group_module = lr_group_modules[0]
-                    group_name = lr_groups_lookup[lr_group_module]
-                    params[f"to_weight_decay_{group_name}"][name] = param
-                else:
-                    params["to_weight_decay"][name] = param
-            else:
-                params["no_weight_decay"][name] = param
-        optimizer_grouped_parameters = []
-        if params["to_weight_decay"]:
-            optimizer_grouped_parameters.append(
-                {
-                    "params": list(params["to_weight_decay"].values()),
-                    "weight_decay": self.args.weight_decay,
-                    "lr": optimizer_kwargs["lr"],
-                }
-            )
-        if params["embeddings"]:
-            lr = optimizer_kwargs["lr"]  # pylint: disable=invalid-name
-            if self.args.embedding_lr_scale:
-                lr *= self.args.embedding_lr_scale  # pylint: disable=invalid-name
-            elif self.args.embedding_lr:
-                lr = self.args.embedding_lr  # pylint: disable=invalid-name
-            optimizer_grouped_parameters.append(
-                {
-                    "params": list(params["embeddings"].values()),
-                    "weight_decay": 0.0,
-                    "lr": lr,
-                }
-            )
-        if params["no_weight_decay"]:
-            optimizer_grouped_parameters.append(
-                {
-                    "params": list(params["no_weight_decay"].values()),
-                    "weight_decay": 0.0,
-                    "lr": optimizer_kwargs["lr"],
-                }
-            )
-        for group_name, group_lr in lr_groups_learning_rates.items():
-            if params[f"to_weight_decay_{group_name}"]:
-                optimizer_grouped_parameters.append(
-                    {
-                        "params": list(
-                            params[f"to_weight_decay_{group_name}"].values()
-                        ),
-                        "weight_decay": self.args.weight_decay,
-                        "lr": group_lr,
-                    }
-                )
-
-        return optimizer_grouped_parameters
-
-    def create_optimizer(self):
-        if (
-            self.args.loraplus_lr_ratio is None
-            and self.args.embedding_lr_scale is None
-            and self.args.embedding_lr is None
-            and self.args.lr_groups is None
-            and self.optimizer_cls_and_kwargs is None
-        ):
-            return super().create_optimizer()
-
-        opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
-
-        if (
-            not self.optimizer
-            and self.optimizer_cls_and_kwargs is not None
-            and issubclass(self.optimizer_cls_and_kwargs[0], BaseOptimizerFactory)
-        ):
-            optimizer_factory_cls, optimizer_kwargs = self.optimizer_cls_and_kwargs
-            self.optimizer = optimizer_factory_cls()(
-                opt_model, self.args, **optimizer_kwargs
-            )
-
-        if not self.optimizer:
-            if self.optimizer_cls_and_kwargs is not None:
-                optimizer_cls, optimizer_kwargs = self.optimizer_cls_and_kwargs
-            else:
-                optimizer_cls, optimizer_kwargs = self.get_optimizer_cls_and_kwargs(
-                    self.args, opt_model
-                )
-
-            optimizer_grouped_parameters = self.create_optimizer_grouped_parameters(
-                opt_model, optimizer_kwargs
-            )
-
-            if self.args.loraplus_lr_ratio is not None:
-                loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
-                loraplus_lr_embedding = getattr(
-                    self.args, "loraplus_lr_embedding", 1e-6
-                )
-                self.optimizer = create_loraplus_optimizer(  # pylint: disable=attribute-defined-outside-init
-                    opt_model,
-                    optimizer_cls,
-                    loraplus_lr_ratio=loraplus_lr_ratio,
-                    loraplus_lr_embedding=loraplus_lr_embedding,
-                    **optimizer_kwargs,
-                )
-            else:
-                # Overwrite `params` in case it's created by `get_optimizer_cls_and_kwargs`
-                # e.g. for GaLore optimizer.
-                if "params" in optimizer_kwargs:
-                    optimizer_grouped_parameters = optimizer_kwargs.pop("params")
-
-                # Overwrite `model` in case it's created by `get_optimizer_cls_and_kwargs`
-                # e.g. for LOMO optimizer.
-                if "model" in optimizer_kwargs:
-                    optimizer_grouped_parameters = optimizer_kwargs.pop("model")
-
-                # For layer-wise dummy optimizers we overwrite optimizer_grouped_parameters with `optimizer_dict`
-                # to avoid arguments conflicts.
-                if "optimizer_dict" in optimizer_kwargs:
-                    optimizer_grouped_parameters = optimizer_kwargs.pop(
-                        "optimizer_dict"
-                    )
-
-                self.optimizer = optimizer_cls(
-                    optimizer_grouped_parameters, **optimizer_kwargs
-                )
-
-            if optimizer_cls.__name__ == "Adam8bit":
-                import bitsandbytes
-
-                manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
-
-                skipped = 0
-                for module in opt_model.modules():
-                    if isinstance(module, nn.Embedding):
-                        skipped += sum(
-                            {
-                                p.data_ptr(): p.numel() for p in module.parameters()
-                            }.values()
-                        )
-                        LOG.info(f"skipped {module}: {skipped/2**20}M params")
-                        manager.register_module_override(
-                            module, "weight", {"optim_bits": 32}
-                        )
-                        LOG.debug(f"bitsandbytes: will optimize {module} in fp32")
-                LOG.info(f"skipped: {skipped/2**20}M params")
-
-        if is_sagemaker_mp_enabled():
-            self.optimizer = smp.DistributedOptimizer(  # pylint: disable=attribute-defined-outside-init
-                self.optimizer
-            )
-
-        return self.optimizer

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

@@ -1,113 +0,0 @@
-"""Module for Axolotl trainer scheduler mixin"""
-
-import logging
-
-import torch
-from torch.optim.lr_scheduler import OneCycleLR
-from transformers.trainer import Trainer
-
-from axolotl.utils.schedulers import (
-    RexLR,
-    get_cosine_schedule_with_min_lr,
-    get_cosine_schedule_with_quadratic_warmup,
-    get_cosine_schedule_with_warmup_decay_constant,
-)
-
-LOG = logging.getLogger(__name__)
-
-
-class SchedulerMixin(Trainer):
-    """
-    Mixin class for scheduler setup in CausalTrainer.
-    """
-
-    args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
-
-    def create_scheduler(
-        self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
-    ):
-        """
-        Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or
-        passed as an argument.
-
-        Args:
-            num_training_steps (int): The number of training steps to do.
-            optimizer (torch.optim.Optimizer): The training optimizer
-        """
-        use_cosine_quadratic = (
-            self.args.lr_scheduler_type == "cosine"
-            and self.args.lr_quadratic_warmup is True
-        )
-
-        use_cosine_min_lr = (
-            self.args.lr_scheduler_type == "cosine"
-            and self.args.cosine_min_lr_ratio is not None
-        )
-
-        # fmt: off
-        if self.lr_scheduler is None:  # type: ignore  # pylint: disable=access-member-before-definition
-            # fmt: on
-            if self.args.alternate_lr_scheduler_type == "one_cycle":
-                num_warmup_steps = self.args.get_warmup_steps(num_training_steps)
-                pct_start = num_warmup_steps / num_training_steps
-                extra_lr_kwargs = {}
-                if "pct_start" not in self.args.lr_scheduler_kwargs:
-                    extra_lr_kwargs["pct_start"] = pct_start
-                if "anneal_strategy" not in self.args.lr_scheduler_kwargs:
-                    extra_lr_kwargs["anneal_strategy"] = "cos"
-
-                self.lr_scheduler = OneCycleLR(
-                    optimizer,
-                    max_lr=self.args.learning_rate,
-                    total_steps=num_training_steps,
-                    **extra_lr_kwargs,
-                    **self.args.lr_scheduler_kwargs,
-                )
-            elif self.args.alternate_lr_scheduler_type == "rex":
-                if use_cosine_min_lr:
-                    assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
-
-                self.lr_scheduler = RexLR(
-                    optimizer=optimizer,
-                    max_lr=self.args.learning_rate,
-                    min_lr=0 if not use_cosine_min_lr else (self.args.learning_rate * self.args.cosine_min_lr_ratio),
-                    total_steps=num_training_steps,
-                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
-                )
-            elif use_cosine_quadratic:
-                if use_cosine_min_lr:
-                    LOG.warning("Both cosine quadratic warmup and min lr detected. Using quadratic warmup.")
-
-                self.lr_scheduler = get_cosine_schedule_with_quadratic_warmup(  # pylint: disable=attribute-defined-outside-init
-                    optimizer,
-                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
-                    num_training_steps=num_training_steps,
-                )
-            elif self.args.cosine_min_lr_ratio and self.args.cosine_constant_lr_ratio and use_cosine_min_lr:
-                assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
-                assert 0 <= self.args.cosine_constant_lr_ratio <= 1.0, "cosine_constant_lr_ratio must be between 0.0 and 1.0"
-                self.lr_scheduler = get_cosine_schedule_with_warmup_decay_constant(  # pylint: disable=attribute-defined-outside-init
-                    optimizer,
-                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
-                    num_training_steps=num_training_steps,
-                    min_lr_ratio=self.args.cosine_min_lr_ratio,
-                    constant_lr_ratio=self.args.cosine_constant_lr_ratio,
-                )
-            elif self.args.cosine_min_lr_ratio and use_cosine_min_lr:
-                assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
-                self.lr_scheduler = get_cosine_schedule_with_min_lr(  # pylint: disable=attribute-defined-outside-init
-                    optimizer,
-                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
-                    num_training_steps=num_training_steps,
-                    min_lr_ratio=self.args.cosine_min_lr_ratio,
-                )
-            else:
-                return super().create_scheduler(num_training_steps, optimizer=optimizer)
-        else:
-            if use_cosine_quadratic:
-                LOG.warning("axolotl's cosine scheduler with quadratic warmup not used (e.g., because of deepspeed).")
-
-            if use_cosine_min_lr:
-                LOG.warning("axolotl's cosine scheduler with min lr not used (e.g., because of deepspeed).")
-
-        return self.lr_scheduler

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

@@ -1,131 +0,0 @@
-"""Module for Axolotl trainer sequence parallelism mixin"""
-
-import logging
-from typing import Any
-
-import torch
-import torch.distributed as dist
-import torch.nn.functional as F
-from datasets import Dataset
-from torch.utils.data import DistributedSampler, Sampler
-
-from axolotl.monkeypatch.attention.ring_attn import get_ring_attn_group
-
-LOG = logging.getLogger(__name__)
-
-try:
-    from ring_flash_attn import update_ring_flash_attn_params
-except ImportError:
-    # We pass silently here, but raise an ImportError in our Axolotl config validation
-    # if cfg.sequence_parallel_degree > 1 and `ring-flash-attn` is not installed.
-    pass
-
-
-class SequenceParallelMixin:
-    """
-    Mixin class for sequence parallelism support in trainers.
-
-    This mixin provides functionality for handling sequence parallelism,
-    including creating appropriate samplers, managing data partitioning,
-    and updating ring flash attention parameters during training.
-    """
-
-    args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
-
-    def _setup_sequence_parallel(self):
-        """Set up sequence parallelism environment."""
-        self.ring_attn_group = get_ring_attn_group()
-
-    def _create_sequence_parallel_sampler(
-        self,
-        dataset: Dataset,
-        shuffle: bool = True,
-        is_eval: bool = False,
-    ) -> DistributedSampler:
-        """
-        Helper method to create sampler for sequence parallelism (SP).
-
-        We create a distributed sampler with rank equal to the SP group ID, which
-        means that all ranks in the SP group receive the same sample / set of samples
-        per training step. We also set the number of replicas equal to the number of
-        SP groups, which is a bit of a hack / unintended use, but works!
-
-        Args:
-            dataset: Dataset to sample from.
-            shuffle: Whether to shuffle the dataset.
-            is_eval: Whether we are creating a sampler for evaluation or training.
-
-        Returns:
-            Distributed sampler.
-        """
-        num_sp_groups = self.args.world_size // self.args.sequence_parallel_degree
-        sp_group_id = dist.get_rank() // self.args.sequence_parallel_degree
-
-        return DistributedSampler(
-            dataset,
-            num_replicas=num_sp_groups,
-            rank=sp_group_id,
-            seed=self.args.seed if shuffle else None,
-            shuffle=shuffle,
-            drop_last=not is_eval,
-        )
-
-    def _sp_get_train_sampler(self, dataset) -> Sampler | None:
-        """
-        Get a training sampler configured for sequence parallelism.
-
-        Args:
-            dataset: The training dataset
-
-        Returns:
-            Configured sequence parallel sampler.
-        """
-        return self._create_sequence_parallel_sampler(
-            dataset,
-            shuffle=not self.args.curriculum_sampling,
-        )
-
-    def _sp_get_eval_sampler(self, eval_dataset) -> Sampler | None:
-        """
-        Get an evaluation sampler configured for sequence parallelism.
-
-        Args:
-            eval_dataset: The evaluation dataset.
-
-        Returns:
-            Configured sequence parallel sampler.
-        """
-        return self._create_sequence_parallel_sampler(
-            eval_dataset, shuffle=False, is_eval=True
-        )
-
-    def _update_ring_flash_attn_params(self, inputs: dict[str, torch.Tensor | Any]):
-        """
-        Calculate the cu_seqlens for the current forward pass and pass the value to
-        the substituted ring_flash_attn. This is accomplished by using the passed
-        `input_ids`.
-
-        Args:
-            inputs: Current batch of inputs.
-        """
-        # At this point, inputs should already be partitioned by the sequence
-        # parallel data collator
-        batch_size = inputs["input_ids"].shape[0]
-        seq_len = inputs["input_ids"].shape[1]
-        packed_seq_lens = [seq_len] * batch_size
-
-        # Calculate the full sequence length across all GPUs in this SP group
-        total_seq_len = seq_len * self.args.sequence_parallel_degree
-
-        cu_seqlens = torch.cumsum(
-            torch.tensor(
-                packed_seq_lens, device=torch.cuda.current_device(), dtype=torch.int32
-            ),
-            dim=-1,
-            dtype=torch.int32,
-        )
-        cu_seqlens = F.pad(
-            F.pad(cu_seqlens, (1, 0), value=0), (0, 1), value=total_seq_len
-        )
-
-        update_ring_flash_attn_params(cu_seqlens, self.ring_attn_group)

src/axolotl/core/trainers/relora.py

@@ -1,43 +0,0 @@
-"""Module for ReLoRA trainer"""
-
-import torch
-
-from axolotl.core.trainers.base import AxolotlTrainer
-from axolotl.monkeypatch.relora import ReLoRAScheduler
-
-
-class ReLoRATrainer(AxolotlTrainer):
-    """Trainer subclass that uses the `OneCycleLR` scheduler"""
-
-    tag_names = ["axolotl", "relora"]
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.lr_scheduler = None
-
-    def create_scheduler(
-        self,
-        num_training_steps: int,
-        optimizer: torch.optim.Optimizer | None = None,
-    ):
-        optimizer = self.optimizer if optimizer is None else optimizer
-        lr_scheduler = super().create_scheduler(num_training_steps, optimizer)
-
-        if self.args.relora_steps:
-            warmup_steps = (
-                self.args.relora_warmup_steps if self.args.relora_warmup_steps else 10
-            )
-            anneal_steps = (
-                self.args.relora_anneal_steps if self.args.relora_anneal_steps else 1
-            )
-            self.lr_scheduler = ReLoRAScheduler(
-                optimizer,
-                lr_scheduler,
-                self.args.relora_steps,
-                anneal_steps,
-                warmup_steps,
-            )
-        else:
-            self.lr_scheduler = lr_scheduler
-
-        return self.lr_scheduler

src/axolotl/core/trainers/trl.py

@@ -1,23 +1,15 @@
-"""Module for TRL PPO trainer"""
+"""
-
+module for TRL PPO training
+"""
 import torch
 from tqdm import tqdm
-from trl import (
+from trl import PPOTrainer
-    CPOTrainer,
-    KTOTrainer,
-    ORPOTrainer,
-    PPOTrainer,
-    PRMTrainer,
-    RewardTrainer,
-)
-
-from axolotl.core.trainers.mixins.scheduler import SchedulerMixin
 
 
 class TRLPPOTrainer(PPOTrainer):
-    """Wrapper for TRL PPO trainer to handle customizations"""
+    """
-
+    wrapper for ppo trainer to handle customizations
-    tag_names = ["axolotl", "ppo"]
+    """
 
     def train(
         self,
@@ -38,7 +30,9 @@ class TRLPPOTrainer(PPOTrainer):
             "batch_size": 16,
         }
 
-        for _, batch in tqdm(enumerate(self.dataloader)):
+        for epoch, batch in tqdm(  # pylint: disable=unused-variable
+            enumerate(self.dataloader)
+        ):
             query_tensors = batch["input_ids"]
 
             # generate model response
@@ -70,43 +64,3 @@ class TRLPPOTrainer(PPOTrainer):
                 rewards,
                 columns_to_log=["query", "response", "ref_response", "ref_rewards"],
             )
-
-
-class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
-    """
-    Extend the base ORPOTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "orpo"]
-
-
-class AxolotlKTOTrainer(SchedulerMixin, KTOTrainer):
-    """
-    Extend the base KTOTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "kto"]
-
-
-class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
-    """
-    Extend the base CPOTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "cpo"]
-
-
-class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
-    """
-    Extend the base RewardTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "reward"]
-
-
-class AxolotlPRMTrainer(SchedulerMixin, PRMTrainer):
-    """
-    Extend the base trl.PRMTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "prm"]

src/axolotl/core/trainers/utils.py

@@ -1,33 +0,0 @@
-"""Utils for Axolotl trainers"""
-
-
-def sanitize_kwargs_for_tagging(tag_names, kwargs=None):
-    if isinstance(tag_names, str):
-        tag_names = [tag_names]
-
-    if kwargs is not None:
-        if "tags" not in kwargs:
-            kwargs["tags"] = tag_names
-        elif "tags" in kwargs and isinstance(kwargs["tags"], list):
-            kwargs["tags"].extend(tag_names)
-        elif "tags" in kwargs and isinstance(kwargs["tags"], str):
-            tag_names.append(kwargs["tags"])
-            kwargs["tags"] = tag_names
-
-    return kwargs
-
-
-def sanitize_kwargs_for_ds_tagging(dataset_tags, kwargs=None):
-    if isinstance(dataset_tags, str):
-        dataset_tags = [dataset_tags]
-
-    if (dataset_tags is not None) and (kwargs is not None):
-        if "dataset_tags" not in kwargs:
-            kwargs["dataset_tags"] = dataset_tags
-        elif "dataset_tags" in kwargs and isinstance(kwargs["dataset_tags"], list):
-            kwargs["dataset_tags"].extend(dataset_tags)
-        elif "dataset_tags" in kwargs and isinstance(kwargs["dataset_tags"], str):
-            dataset_tags.append(kwargs["dataset_tags"])
-            kwargs["dataset_tags"] = dataset_tags
-
-    return kwargs

src/axolotl/core/training_args.py

@@ -1,7 +1,6 @@
 """
 extra axolotl specific training args
 """
-
 from dataclasses import dataclass, field
 from typing import Optional
 
@@ -207,19 +206,14 @@ class AxolotlTrainingMixins:
         },
     )
 
-    sequence_parallel_degree: Optional[int] = field(
-        default=1,
-        metadata={"help": "The number of workers to use in sequence parallelism"},
-    )
-
 
 @dataclass
 class AxolotlTrainingArguments(AxolotlTrainingMixins, TrainingArguments):
     """
     Training arguments for Causal trainer
 
-    This code is duplicated due to HF TrainingArguments not setting output_dir with a
+    This code is duplicated due to HF TrainingArguments not setting output_dir with a defaujlt value
-    default value so it can't be used as a mixin.
+    so it can't be used as a mixin.
     """
 
 

src/axolotl/evaluate.py

@@ -8,8 +8,6 @@ from typing import Dict, Optional
 
 import torch
 from accelerate.logging import get_logger
-from datasets import Dataset
-from transformers.trainer import Trainer
 
 from axolotl.logging_config import configure_logging
 from axolotl.train import TrainDatasetMeta
@@ -27,18 +25,18 @@ LOG = get_logger("axolotl.evaluate")
 
 
 def evaluate_dataset(
-    trainer: Trainer, dataset: Dataset, dataset_type: str, flash_optimum: bool = False
+    trainer, dataset, dataset_type: str, flash_optimum: bool = False
 ) -> Optional[Dict[str, float]]:
-    """Helper function to evaluate a single dataset.
+    """Helper function to evaluate a single dataset safely.
 
     Args:
-        trainer: The trainer instance.
+        trainer: The trainer instance
-        dataset: Dataset to evaluate.
+        dataset: Dataset to evaluate
-        dataset_type: Type of dataset ('train' or 'eval').
+        dataset_type: Type of dataset ('train' or 'eval')
-        flash_optimum: Whether to use flash optimum.
+        flash_optimum: Whether to use flash optimum
 
     Returns:
-        Dictionary of metrics or None if dataset is None.
+        Dictionary of metrics or None if dataset is None
     """
     if dataset is None:
         return None
@@ -65,14 +63,17 @@ def evaluate_dataset(
 
 def evaluate(*, cfg: DictDefault, dataset_meta: TrainDatasetMeta) -> Dict[str, float]:
     """
-    Evaluate a model on training and validation datasets.
+    Evaluate a model on training and validation datasets
 
     Args:
         cfg: Dictionary mapping `axolotl` config keys to values.
         dataset_meta: Dataset metadata containing training and evaluation datasets.
 
     Returns:
-        Dictionary mapping metric names to their values.
+        Tuple containing:
+        - The model (either PeftModel or PreTrainedModel)
+        - The tokenizer
+        - Dictionary of evaluation metrics
     """
     # pylint: disable=duplicate-code
     # Enable expandable segments for cuda allocation to improve VRAM usage

src/axolotl/integrations/config.py

@@ -11,17 +11,19 @@
 # the License.
 
 """
-Module to handle merging the plugins' input arguments with the base configurations.
+module to handle merging the plugins' input arguments with the base configurations.
 
-This was moved here to prevent circular imports.
+this was moved here to prevent circular imports
 """
 
 from typing import Any, Dict, List
 
-from axolotl.utils.schemas.config import (
+from axolotl.utils.config.models.input.v0_4_1 import (
     AxolotlConfigWCapabilities as AxolotlConfigWCapabilitiesBase,
 )
-from axolotl.utils.schemas.config import AxolotlInputConfig as AxolotlInputConfigBase
+from axolotl.utils.config.models.input.v0_4_1 import (
+    AxolotlInputConfig as AxolotlInputConfigBase,
+)
 
 
 def merge_input_args():

src/axolotl/integrations/grokfast/__init__.py

@@ -1,7 +1,6 @@
 """
 Grokfast plugin for Axolotl
 """
-
 import logging
 
 from transformers.trainer_callback import TrainerCallback

src/axolotl/integrations/grokfast/args.py

@@ -1,7 +1,6 @@
 """
 config args for grokfast plugin
 """
-
 from typing import Optional
 
 from pydantic import BaseModel

src/axolotl/integrations/kd/__init__.py

@@ -34,3 +34,12 @@ class KDPlugin(BasePlugin):
 
             return AxolotlKDTrainer
         return None
+
+    def add_callbacks_post_trainer(self, cfg, trainer):
+        callbacks = []
+        if cfg.kd_trainer:
+            from .callbacks import KDAlphaSchedulerCallback
+
+            callbacks.append(KDAlphaSchedulerCallback())
+
+        return callbacks

src/axolotl/integrations/kd/args.py

@@ -26,12 +26,14 @@ class KDArgs(BaseModel):
     """
 
     kd_trainer: Optional[bool] = None  # whether to use KD trainer
-    kd_ce_alpha: Optional[float] = (
+    kd_ce_alpha: Optional[
-        None  # loss coefficient for cross-entropy loss during KD
+        float
-    )
+    ] = None  # loss coefficient for cross-entropy loss during KD
     kd_alpha: Optional[float] = None  # loss coefficient for KD loss
+    kd_ce_alpha_end: Optional[float] = None  # end value for kd_ce_alpha
+    kd_alpha_end: Optional[float] = None  # end value for kd_alpha
     kd_temperature: Optional[float] = None  # temperature for sampling during KD
     kd_zscore_base_temp: Optional[float] = None  # base temperature for zscore scaling
-    kd_top_k_before_softmax: Optional[bool] = (
+    kd_top_k_before_softmax: Optional[
-        None  # whether to sample top k before softmax during KD
+        bool
-    )
+    ] = None  # whether to sample top k before softmax during KD

src/axolotl/integrations/kd/callbacks.py

@@ -0,0 +1,28 @@
+from transformers import TrainerCallback
+
+
+class KDAlphaSchedulerCallback(TrainerCallback):
+    """Callback to for scheduling KD alpha during training."""
+
+    def on_epoch_begin(
+        self, args, state, control, **kwargs  # pylint: disable=unused-argument
+    ):
+        if int(state.epoch) == 0:
+            state.kd_alpha = args.kd_alpha
+            state.kd_ce_alpha = args.kd_ce_alpha
+        elif int(state.epoch) == state.num_train_epochs - 1:
+            if args.kd_alpha_end is not None:
+                control.kd_alpha = args.kd_alpha_end
+            if args.kd_ce_alpha_end is not None:
+                control.kd_ce_alpha = args.kd_ce_alpha_end
+        else:
+            epoch_steps = state.num_train_epochs - 1
+            scale = int(state.epoch) / epoch_steps
+            if args.kd_alpha_end is not None:
+                control.kd_alpha = (
+                    args.kd_alpha + (args.kd_alpha_end - args.kd_alpha) * scale
+                )
+            if args.kd_ce_alpha_end is not None:
+                control.kd_ce_alpha = (
+                    args.kd_ce_alpha + (args.kd_ce_alpha_end - args.kd_ce_alpha) * scale
+                )

src/axolotl/integrations/kd/chat_template.py

@@ -62,10 +62,16 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         Transform logprobs to target format for KD training
         """
 
+        if "target_logprobs" in sample.keys() and "target_token_ids" in sample.keys():
+            logprobs = sample.pop("target_logprobs")
+            token_ids = sample.pop("target_token_ids")
+        else:
             logprobs = sample.pop(self.logprobs_field)
+            token_ids = [None] * len(logprobs)
+
         target_seq_len = len(logprobs)
         input_seq_len = len(sample["input_ids"])
-        input_padding_len = input_seq_len - target_seq_len
+        target_padding_len = input_seq_len - target_seq_len
         # get non-zero top-k (prune None logprobs from vllm data step)
         top_k_vals = [
             len(logprobs[i])
@@ -82,11 +88,11 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         target_token_ids = []
         target_mask = []
 
-        if input_padding_len < 0:
+        if target_padding_len < 0:
             # logprobs is longer than target_seq_len,
             # so we need to slice from the left/beginning of logprobs
             logprobs = logprobs[:-input_seq_len]
-            input_padding_len = 0
+            target_padding_len = 0
             # target_seq_len = input_seq_len
 
         # truncate the second dimension of the logprobs to top_k
@@ -98,23 +104,24 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         # for causal models, if we start the range at 1, then we don't need to shift in the trainer
         # otherwise, we need to shift in the trainer
         shift = 0
-        for _ in range(shift, input_padding_len):
+        for _ in range(shift, target_padding_len):
             target_logprobs.append([-float("inf")] * top_k)
             target_token_ids.append(list(range(top_k)))
             target_mask.append([0] * top_k)
 
-        for position in range(input_padding_len, input_seq_len):
+        for position in range(target_padding_len, input_seq_len):
             if sample["labels"][position] == -100:
                 target_mask.append([0] * top_k)
             else:
                 target_mask.append([1] * top_k)
 
-        for _, token_pos_logprobs in enumerate(logprobs):
+        for token_pos_logprobs, token_pos_token_ids in zip(logprobs, token_ids):
             # Initialize collections for logprobs and token_ids
             position_logprobs = []
             position_token_ids = []
 
             # Process each token probability entry
+            if token_pos_token_ids is None:
                 for entry in token_pos_logprobs:
                     # Extract logprob value
                     logprob = entry["logprob"]
@@ -125,6 +132,9 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
                     # Append to our collections
                     position_logprobs.append(logprob)
                     position_token_ids.append(token_id)
+            else:
+                position_logprobs = token_pos_logprobs
+                position_token_ids = token_pos_token_ids
 
             # Convert to a tensor for easier manipulation
             position_logprobs_tensor = torch.tensor(
@@ -143,6 +153,7 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
                 teacher_probs_t2 = teacher_probs_t1**exponent
             else:
                 teacher_probs_t2 = teacher_probs_t1
+
             # Re-normalize
             teacher_probs_t2 = teacher_probs_t2 / teacher_probs_t2.sum(
                 dim=0, keepdim=True

src/axolotl/integrations/kd/trainer.py

@@ -16,17 +16,35 @@
 KD trainer
 """
 
+from transformers import TrainerControl
+
 from axolotl.core.trainers.base import AxolotlTrainer
 
 from .topk_logprob.forward_kl import loss as topk_kd_loss
 from .topk_logprob.forward_kl import topk_kd_loss_with_zscore
 
 
+class AxolotlKDTrainerControl(TrainerControl):
+    kd_alpha: float = 1.0
+    kd_ce_alpha: float = 0.0
+
+    def state(self) -> dict:
+        state_val = super().state()
+        state_val["args"]["kd_alpha"] = self.kd_alpha
+        state_val["args"]["kd_ce_alpha"] = self.kd_ce_alpha
+
+
 class AxolotlKDTrainer(AxolotlTrainer):
     """
     Custom trainer subclass for Knowledge Distillation (KD)
     """
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.kd_alpha = self.args.kd_alpha
+        self.kd_ce_alpha = self.args.kd_ce_alpha
+        self.control = AxolotlKDTrainerControl()
+
     def _set_signature_columns_if_needed(self):
         super()._set_signature_columns_if_needed()
         columns_to_add = []
@@ -95,9 +113,8 @@ class AxolotlKDTrainer(AxolotlTrainer):
                 top_k_before_softmax=1 if self.args.kd_top_k_before_softmax else 0,
             )
 
-        if self.args.kd_ce_alpha > 0:
+        if self.kd_ce_alpha > 0:
-            kd_alpha = self.args.kd_alpha
+            loss = self.kd_ce_alpha * outputs["loss"] + self.kd_alpha * loss_kd
-            loss = self.args.kd_ce_alpha * outputs["loss"] + kd_alpha * loss_kd
         else:
             loss = loss_kd
         # Save past state if it exists

src/axolotl/integrations/liger/__init__.py

@@ -55,9 +55,9 @@ class LigerPlugin(BasePlugin):
             if "cross_entropy" in liger_fn_sig.parameters:
                 kwargs["cross_entropy"] = cfg.liger_cross_entropy
             if "fused_linear_cross_entropy" in liger_fn_sig.parameters:
-                kwargs["fused_linear_cross_entropy"] = (
+                kwargs[
-                    cfg.liger_fused_linear_cross_entropy
+                    "fused_linear_cross_entropy"
-                )
+                ] = cfg.liger_fused_linear_cross_entropy
             if "rms_norm" in liger_fn_sig.parameters:
                 kwargs["rms_norm"] = cfg.liger_rms_norm
             if "layer_norm" in liger_fn_sig.parameters:

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

@@ -1,7 +1,6 @@
 """
 DeepseekV2 model with LigerFusedLinearCrossEntropyLoss
 """
-
 # pylint: disable=duplicate-code
 
 from typing import List, Optional, Tuple, Union

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

@@ -1,7 +1,6 @@
 """
 Jamba model with LigerFusedLinearCrossEntropyLoss
 """
-
 # pylint: disable=duplicate-code
 
 from typing import Optional, Tuple, Union

src/axolotl/integrations/lm_eval/__init__.py

@@ -1,7 +1,6 @@
 """
 Module for the Plugin for LM Eval Harness
 """
-
 import subprocess  # nosec
 
 from axolotl.integrations.base import BasePlugin

src/axolotl/integrations/lm_eval/args.py

@@ -1,7 +1,6 @@
 """
 Module for handling lm eval harness input arguments.
 """
-
 from typing import List, Optional
 
 from pydantic import BaseModel

src/axolotl/integrations/lm_eval/cli.py

@@ -1,7 +1,6 @@
 """
 axolotl CLI for running lm_eval tasks
 """
-
 import subprocess  # nosec
 from collections import defaultdict
 from datetime import datetime

src/axolotl/kernels/geglu.py

@@ -5,7 +5,6 @@ See "GLU Variants Improve Transformer" (https://arxiv.org/abs/2002.05202).
 
 Credit to `unsloth` (https://unsloth.ai/) for inspiration for this implementation.
 """
-
 # pylint: disable=invalid-name,unnecessary-lambda-assignment,duplicate-code
 
 import torch

src/axolotl/kernels/lora.py

@@ -6,7 +6,6 @@ See "LoRA: Low-Rank Adaptation of Large Language Models"
 
 Credit to `unsloth` (https://unsloth.ai/) for inspiration for this implementation.
 """
-
 # pylint: disable=invalid-name
 
 from typing import Callable

src/axolotl/kernels/quantize.py

@@ -1,5 +1,4 @@
 """Dequantization utilities for `bitsandbytes` integration."""
-
 # pylint: disable=invalid-name,global-statement
 
 import ctypes

src/axolotl/kernels/swiglu.py

@@ -5,7 +5,6 @@ See "GLU Variants Improve Transformer" (https://arxiv.org/abs/2002.05202).
 
 Credit to `unsloth` (https://unsloth.ai/) for inspiration for this implementation.
 """
-
 import torch
 import triton
 import triton.language as tl

src/axolotl/models/mamba/configuration_mamba.py

@@ -1,7 +1,6 @@
 """
 HF Transformers MambaConfig
 """
-
 from transformers import PretrainedConfig
 
 

src/axolotl/monkeypatch/attention/mllama.py

@@ -1,7 +1,6 @@
 """
 Monkeypatch for Vision Llama for FA2 support
 """
-
 # pylint: disable=duplicate-code
 
 from typing import Optional, Tuple
@@ -221,10 +220,10 @@ def patch_mllama():
         True
     )
     MLLAMA_TEXT_ATTENTION_CLASSES["flash_attention_2"] = MllamaTextSelfFlashAttention2
-    MLLAMA_TEXT_CROSS_ATTENTION_CLASSES["flash_attention_2"] = (
+    MLLAMA_TEXT_CROSS_ATTENTION_CLASSES[
-        MllamaTextCrossFlashAttention2
+        "flash_attention_2"
-    )
+    ] = MllamaTextCrossFlashAttention2
     # fallback to SDPA
-    MLLAMA_VISION_ATTENTION_CLASSES["flash_attention_2"] = (
+    MLLAMA_VISION_ATTENTION_CLASSES[
-        MLLAMA_VISION_ATTENTION_CLASSES["sdpa"]
+        "flash_attention_2"
-    )
+    ] = MLLAMA_VISION_ATTENTION_CLASSES["sdpa"]

src/axolotl/monkeypatch/attention/ring_attn.py

@@ -1,89 +0,0 @@
-"""
-Ring attention group registration and flash attention patching.
-
-Make use of the `ring-flash-attn` (https://github.com/zhuzilin/ring-flash-attention)
-package, specifically the `hf_adapter.substitute_hf_flash_attn` function to patch in
-their sequence parallel version of Flash Attention 2.
-"""
-
-import torch.distributed as dist
-from accelerate.logging import get_logger
-
-from axolotl.logging_config import configure_logging
-
-configure_logging()
-LOG = get_logger(__name__)
-
-RING_ATTN_GROUP = None
-
-
-def get_ring_attn_group() -> dist.ProcessGroup:
-    """
-    Getter for ring attention group on this rank.
-
-    Returns:
-        The process group for ring attention for this rank.
-    """
-    return RING_ATTN_GROUP
-
-
-def set_ring_attn_group(ring_attn_group: dist.ProcessGroup):
-    """
-    Setter for ring attention group on this rank.
-
-    Args:
-        Process group for ring attention.
-    """
-    global RING_ATTN_GROUP  # pylint: disable=global-statement
-    RING_ATTN_GROUP = ring_attn_group
-
-
-def register_ring_attn(sequence_parallel_degree: int):
-    """
-    Create ring attention group and substitute flash attn with ring flash attn.
-
-    Args:
-        sequence_parallel_degree: Sequence parallelism factor.
-    """
-    LOG.info(
-        "Enabling ring attention sequence parallelism: "
-        f"each sequence will be processed across {sequence_parallel_degree} GPUs"
-    )
-
-    world_size = dist.get_world_size()
-    assert sequence_parallel_degree <= world_size, (
-        f"sequence_parallel_degree ({sequence_parallel_degree}) "
-        f"must be less than or equal to world_size ({world_size})"
-    )
-    assert world_size % sequence_parallel_degree == 0, (
-        f"sequence_parallel_degree ({sequence_parallel_degree}) "
-        f"must evenly divide world_size ({world_size})"
-    )
-
-    # Detailed logging of group formation
-    rank = dist.get_rank()
-    group_assignments = {}
-
-    for i in range(world_size // sequence_parallel_degree):
-        ring_attn_ranks = list(
-            range(
-                i * sequence_parallel_degree,
-                (i + 1) * sequence_parallel_degree,
-            )
-        )
-        group = dist.new_group(ranks=ring_attn_ranks, backend="nccl")
-
-        # Track which GPUs are in which groups
-        for r in ring_attn_ranks:
-            group_assignments[r] = i
-
-        if rank in ring_attn_ranks:
-            set_ring_attn_group(group)
-
-    # Log the GPU group assignments
-    if rank == 0:
-        LOG.info(f"Sequence parallel group assignments: {group_assignments}")
-
-    from ring_flash_attn import substitute_hf_flash_attn
-
-    substitute_hf_flash_attn(get_ring_attn_group(), sequence_parallel_degree)

src/axolotl/monkeypatch/data/batch_dataset_fetcher.py

@@ -1,5 +1,4 @@
 """monkey patches for the dataset fetcher to handle batches of packed indexes"""
-
 # pylint: disable=protected-access
 
 import torch

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -12,9 +12,7 @@ import transformers
 from einops import rearrange
 from flash_attn.bert_padding import pad_input, unpad_input
 from transformers.modeling_outputs import BaseModelOutputWithPast
-from transformers.models.llama.modeling_llama import (
+from transformers.models.llama.modeling_llama import LlamaAttention
-    LlamaAttention,
-)
 from transformers.models.llama.modeling_llama import (
     LlamaDecoderLayer as OriginalLlamaDecoderLayer,
 )
@@ -492,11 +490,9 @@ def flashattn_forward(
             # We have disabled _prepare_decoder_attention_mask in LlamaModel
             # the attention_mask should be the same as the key_padding_mask
             key_padding_mask=attention_mask,
-            query_padding_mask=(
+            query_padding_mask=attention_mask[:, -query_states.size(1) :]
-                attention_mask[:, -query_states.size(1) :]
             if attention_mask is not None
-                else None
+            else None,
-            ),
         )
         output_unpad = flash_attn_varlen_qkvpacked_func(
             qkv_unpad,
@@ -535,11 +531,9 @@ def flashattn_forward(
                 value_states,
                 kvpacked=True,
                 key_padding_mask=attention_mask,
-                query_padding_mask=(
+                query_padding_mask=attention_mask[:, -query_states.size(1) :]
-                    attention_mask[:, -query_states.size(1) :]
                 if attention_mask is not None
-                    else None
+                else None,
-                ),
             )
             if q_unpad.dtype != kv_unpad.dtype:
                 kv_unpad = kv_unpad.to(q_unpad.dtype)

src/axolotl/monkeypatch/llama_expand_mask.py

@@ -1,7 +1,6 @@
 """
 expands the binary attention mask per 3.2.2 of https://arxiv.org/pdf/2107.02027.pdf
 """
-
 from typing import Optional
 
 import torch

src/axolotl/monkeypatch/mistral_attn_hijack_flash.py

@@ -1,5 +1,4 @@
 """Flash attention monkey patch for mistral model"""
-
 # pylint: disable=duplicate-code
 
 import logging
@@ -22,10 +21,7 @@ from transformers.models.mistral.modeling_mistral import (
 from transformers.models.mistral.modeling_mistral import (
     MistralDecoderLayer as OriginalMistralDecoderLayer,
 )
-from transformers.models.mistral.modeling_mistral import (
+from transformers.models.mistral.modeling_mistral import apply_rotary_pos_emb, repeat_kv
-    apply_rotary_pos_emb,
-    repeat_kv,
-)
 
 from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
 
@@ -247,11 +243,9 @@ def flashattn_forward(
             # We have disabled _prepare_decoder_attention_mask in LlamaModel
             # the attention_mask should be the same as the key_padding_mask
             key_padding_mask=attention_mask,
-            query_padding_mask=(
+            query_padding_mask=attention_mask[:, -query_states.size(1) :]
-                attention_mask[:, -query_states.size(1) :]
             if attention_mask is not None
-                else None
+            else None,
-            ),
         )
         output_unpad = flash_attn_varlen_qkvpacked_func(
             qkv_unpad,
@@ -292,11 +286,9 @@ def flashattn_forward(
                 value_states,
                 kvpacked=True,
                 key_padding_mask=attention_mask,
-                query_padding_mask=(
+                query_padding_mask=attention_mask[:, -query_states.size(1) :]
-                    attention_mask[:, -query_states.size(1) :]
                 if attention_mask is not None
-                    else None
+                else None,
-                ),
             )
             if q_unpad.dtype != kv_unpad.dtype:
                 kv_unpad = kv_unpad.to(q_unpad.dtype)

src/axolotl/monkeypatch/mixtral/__init__.py

@@ -1,7 +1,6 @@
 """
 Patches to support multipack for mixtral
 """
-
 import torch
 
 

src/axolotl/monkeypatch/relora.py

@@ -1,5 +1,4 @@
 """Implements the ReLoRA training procedure from https://arxiv.org/abs/2307.05695, minus the initial full fine-tune."""
-
 import glob
 import json
 import logging
@@ -412,10 +411,7 @@ def merge_and_save(
         if shard_path.endswith(".safetensors"):
             in_tensors = st.load_file(str(Path(model_src) / shard_path))
         else:
-            in_tensors = torch.load(
+            in_tensors = torch.load(Path(model_src) / shard_path)
-                Path(model_src) / shard_path,
-                weights_only=True,  # to prevent arbitrary code execution
-            )
             if "state_dict" in in_tensors:
                 in_tensors = in_tensors["state_dict"]
 

src/axolotl/monkeypatch/trainer_fsdp_optim.py

@@ -1,7 +1,6 @@
 """
 fix for FSDP optimizer save in trainer w 4.47.0
 """
-
 import inspect
 import logging
 

src/axolotl/monkeypatch/utils.py

@@ -1,7 +1,6 @@
 """
 Shared utils for the monkeypatches
 """
-
 import re
 from typing import Optional, Tuple
 

src/axolotl/monkeypatch/xformers_/__init__.py

@@ -1,7 +1,6 @@
 """
 Fused MLP layer for incrementally improved training efficiency
 """
-
 import torch
 from transformers.models.llama.modeling_llama import LlamaMLP
 from xformers.ops import SwiGLU

src/axolotl/prompt_strategies/alpaca_w_system.py

@@ -1,7 +1,6 @@
 """
 Prompt strategies loader for alpaca instruction datasets with system prompts
 """
-
 from typing import Generator, Tuple, Union
 
 from axolotl.prompt_tokenizers import PromptTokenizingStrategy

src/axolotl/prompt_strategies/chat_template.py

@@ -13,7 +13,7 @@ from axolotl.prompt_strategies.jinja_template_analyzer import JinjaTemplateAnaly
 from axolotl.prompt_tokenizers import PromptTokenizingStrategy
 from axolotl.prompters import IGNORE_TOKEN_ID, Prompter
 from axolotl.utils.chat_templates import get_chat_template_from_config
-from axolotl.utils.schemas.datasets import DatasetConfig
+from axolotl.utils.config.models.input.v0_4_1 import DatasetConfig
 
 # Configure the logger
 LOG = logging.getLogger("axolotl")