fix broken link

* adding pre-commit auto-update GH action and bumping plugin versions

* running updated pre-commit plugins

* sorry to revert, but pylint complained

* Update .pre-commit-config.yaml

Co-authored-by: Wing Lian <wing.lian@gmail.com>

---------

Co-authored-by: Dan Saunders <dan@axolotl.ai>
Co-authored-by: Wing Lian <wing.lian@gmail.com>

.github/workflows/precommit-autoupdate.yml

@@ -0,0 +1,49 @@
+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>

.gitignore

@@ -181,6 +181,10 @@ 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: v4.4.0
+    rev: v5.0.0
     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: 23.3.0
+    rev: 25.1.0
     hooks:
     -   id: black
 -   repo: https://github.com/pycqa/isort
-    rev: 5.12.0
+    rev: 6.0.1
     hooks:
       - id: isort
 -   repo: https://github.com/PyCQA/flake8
-    rev: 6.1.0
+    rev: 7.1.2
     hooks:
     - id: flake8
 -   repo: https://github.com/pylint-dev/pylint
-    rev: c8c96d20cde3552a79858c7456bb1483bf83d633
+    rev: v3.3.6
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.3.0
+    rev: v1.15.0
     hooks:
     - id: mypy
       additional_dependencies:
@@ -36,7 +36,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.7.5
+    rev: 1.8.3
     hooks:
     -   id: bandit
         args: [

README.md

@@ -97,6 +97,7 @@ 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/api/) - Auto-generated code documentation
 - [FAQ](https://axolotl-ai-cloud.github.io/axolotl/docs/faq.html) - Frequently asked questions
 
 ## 🤝 Getting Help

_quarto.yml

@@ -1,6 +1,166 @@
 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: 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"
@@ -36,6 +196,12 @@ website:
             - docs/cli.qmd
             - docs/config.qmd
 
+        - section: "Reference"
+          contents:
+            - docs/config.qmd
+            - text: "API Reference"
+              href: docs/api/index.qmd
+
         - section: "Dataset Formats"
           contents: docs/dataset-formats/*
 
@@ -80,3 +246,22 @@ format:
     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/multigpu.py

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

cicd/tests.py

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

docs/.gitignore

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

docs/cli.qmd

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

docs/config.qmd

@@ -85,6 +85,12 @@ 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

docs/custom_integrations.qmd

@@ -55,3 +55,47 @@ 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](docs/dataset-formats) and prompt strategies to:
+the [dataset format](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/installation.qmd

@@ -79,6 +79,7 @@ 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}
 

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

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

scripts/chat_datasets.py

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

scripts/cutcrossentropy_install.py

@@ -1,4 +1,5 @@
 """Script to output the correct installation command for cut-cross-entropy."""
+
 import importlib.util
 import sys
 
@@ -17,12 +18,12 @@ if v < V("2.4.0"):
 
 cce_spec = importlib.util.find_spec("cut_cross_entropy")
 
-uninstall_prefix = ""
+UNINSTALL_PREFIX = ""
 if cce_spec:
     if not importlib.util.find_spec("cut_cross_entropy.transformers"):
-        uninstall_prefix = "pip uninstall -y cut-cross-entropy && "
+        UNINSTALL_PREFIX = "pip uninstall -y cut-cross-entropy && "
 
 print(
-    uninstall_prefix
+    UNINSTALL_PREFIX
     + 'pip install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@24fbe4b5dab9a6c250a014573613c1890190536c"'
 )

src/axolotl/cli/cloud/__init__.py

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

src/axolotl/cli/cloud/base.py

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

src/axolotl/cli/cloud/modal_.py

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

src/axolotl/cli/main.py

@@ -1,4 +1,5 @@
 """Click CLI definitions for various axolotl commands."""
+
 # pylint: disable=redefined-outer-name
 
 import logging

src/axolotl/cli/utils.py

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

src/axolotl/convert.py

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

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

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

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

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

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

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

src/axolotl/core/chat/messages.py

@@ -1,6 +1,7 @@
 """
 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

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

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

@@ -1,6 +1,7 @@
 """
 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

@@ -332,9 +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"
-            ] = self.cfg.include_tokens_per_second
+            training_arguments_kwargs["include_tokens_per_second"] = (
+                self.cfg.include_tokens_per_second
+            )
 
         if self.cfg.bf16 == "full":
             training_arguments_kwargs["bf16_full_eval"] = True
@@ -351,13 +351,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["seed"] = self.cfg.seed
 
         if self.cfg.gradient_checkpointing:
-            training_arguments_kwargs[
-                "gradient_checkpointing"
-            ] = self.cfg.gradient_checkpointing
+            training_arguments_kwargs["gradient_checkpointing"] = (
+                self.cfg.gradient_checkpointing
+            )
             if self.cfg.gradient_checkpointing_kwargs is not None:
-                training_arguments_kwargs[
-                    "gradient_checkpointing_kwargs"
-                ] = self.cfg.gradient_checkpointing_kwargs
+                training_arguments_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:
@@ -373,9 +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"
-            ] = self.cfg.lr_quadratic_warmup
+            training_arguments_kwargs["lr_quadratic_warmup"] = (
+                self.cfg.lr_quadratic_warmup
+            )
 
         if self.cfg.adam_beta1:
             training_arguments_kwargs["adam_beta1"] = self.cfg.adam_beta1
@@ -399,28 +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"
-            ] = self.cfg.dataloader_pin_memory
+            training_arguments_kwargs["dataloader_pin_memory"] = (
+                self.cfg.dataloader_pin_memory
+            )
         if self.cfg.dataloader_num_workers is not None:
-            training_arguments_kwargs[
-                "dataloader_num_workers"
-            ] = self.cfg.dataloader_num_workers
+            training_arguments_kwargs["dataloader_num_workers"] = (
+                self.cfg.dataloader_num_workers
+            )
         if self.cfg.dataloader_prefetch_factor is not None:
-            training_arguments_kwargs[
-                "dataloader_prefetch_factor"
-            ] = self.cfg.dataloader_prefetch_factor
+            training_arguments_kwargs["dataloader_prefetch_factor"] = (
+                self.cfg.dataloader_prefetch_factor
+            )
         if self.cfg.dataloader_drop_last is not None:
-            training_arguments_kwargs[
-                "dataloader_drop_last"
-            ] = self.cfg.dataloader_drop_last
+            training_arguments_kwargs["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"
-            ] = self.cfg.remove_unused_columns
+            training_arguments_kwargs["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
@@ -452,9 +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"
-            ] = self.cfg.metric_for_best_model
+            training_arguments_kwargs["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
 
@@ -467,13 +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"
-                    ] = self.cfg.torch_compile_backend
+                    training_arguments_kwargs["torch_compile_backend"] = (
+                        self.cfg.torch_compile_backend
+                    )
                 if self.cfg.torch_compile_mode:
-                    training_arguments_kwargs[
-                        "torch_compile_mode"
-                    ] = self.cfg.torch_compile_mode
+                    training_arguments_kwargs["torch_compile_mode"] = (
+                        self.cfg.torch_compile_mode
+                    )
 
         # DDP Config
         if self.cfg.ddp_timeout:
@@ -482,32 +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"
-            ] = self.cfg.ddp_broadcast_buffers
+            training_arguments_kwargs["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"
-        ] = self.cfg.micro_batch_size
+        training_arguments_kwargs["per_device_train_batch_size"] = (
+            self.cfg.micro_batch_size
+        )
         if self.cfg.eval_batch_size:
-            training_arguments_kwargs[
-                "per_device_eval_batch_size"
-            ] = self.cfg.eval_batch_size
+            training_arguments_kwargs["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"
-            ] = self.cfg.auto_find_batch_size
-        training_arguments_kwargs[
-            "gradient_accumulation_steps"
-        ] = self.cfg.gradient_accumulation_steps
-        training_arguments_kwargs[
-            "eval_accumulation_steps"
-        ] = self.cfg.gradient_accumulation_steps
+            training_arguments_kwargs["auto_find_batch_size"] = (
+                self.cfg.auto_find_batch_size
+            )
+        training_arguments_kwargs["gradient_accumulation_steps"] = (
+            self.cfg.gradient_accumulation_steps
+        )
+        training_arguments_kwargs["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
@@ -554,9 +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"
-            ] = self.cfg.lr_scheduler
+            training_arguments_kwargs["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"
@@ -565,9 +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"
-        ] = self.cfg.cosine_constant_lr_ratio
+        training_arguments_kwargs["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
         )
@@ -580,40 +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"
-            ] = self.cfg.sample_packing_bin_size
+            training_arguments_kwargs["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"
-            ] = self.cfg.sample_packing_group_size
+            training_arguments_kwargs["sample_packing_group_size"] = (
+                self.cfg.sample_packing_group_size
+            )
         if self.cfg.sample_packing_eff_est:
-            training_arguments_kwargs[
-                "sample_packing_efficiency"
-            ] = self.cfg.sample_packing_eff_est
+            training_arguments_kwargs["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"
-            ] = self.cfg.relora_warmup_steps
+            training_arguments_kwargs["relora_warmup_steps"] = (
+                self.cfg.relora_warmup_steps
+            )
             if self.cfg.relora_anneal_steps:
-                training_arguments_kwargs[
-                    "relora_anneal_steps"
-                ] = self.cfg.relora_anneal_steps
+                training_arguments_kwargs["relora_anneal_steps"] = (
+                    self.cfg.relora_anneal_steps
+                )
             if self.cfg.relora_prune_ratio:
-                training_arguments_kwargs[
-                    "relora_prune_ratio"
-                ] = self.cfg.relora_prune_ratio
+                training_arguments_kwargs["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"
-            ] = self.cfg.lisa_step_interval
-            training_arguments_kwargs[
-                "lisa_layers_attribute"
-            ] = self.cfg.lisa_layers_attribute
+            training_arguments_kwargs["lisa_step_interval"] = (
+                self.cfg.lisa_step_interval
+            )
+            training_arguments_kwargs["lisa_layers_attribute"] = (
+                self.cfg.lisa_layers_attribute
+            )
 
         training_arguments_kwargs = self.hook_pre_create_training_args(
             training_arguments_kwargs
@@ -627,9 +627,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             )
 
         if self.cfg.neftune_noise_alpha is not None:
-            training_arguments_kwargs[
-                "neftune_noise_alpha"
-            ] = self.cfg.neftune_noise_alpha
+            training_arguments_kwargs["neftune_noise_alpha"] = (
+                self.cfg.neftune_noise_alpha
+            )
 
         trainer_kwargs = {}
 
@@ -731,23 +731,23 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 importlib.import_module("torchdistx")
 
         if self.cfg.optim_target_modules:
-            training_arguments_kwargs[
-                "optim_target_modules"
-            ] = self.cfg.optim_target_modules
+            training_arguments_kwargs["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"
-        ] = self.cfg.loraplus_lr_embedding
+        training_arguments_kwargs["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"
-            ] = self.cfg.accelerator_config
+            training_arguments_kwargs["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
@@ -756,13 +756,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         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"
-            ] = self.cfg.kd_zscore_base_temp
+            training_arguments_kwargs["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"
-            ] = self.cfg.kd_top_k_before_softmax
+            training_arguments_kwargs["kd_top_k_before_softmax"] = (
+                self.cfg.kd_top_k_before_softmax
+            )
 
         if self.cfg.reward_model:
             training_args_cls = AxolotlRewardConfig
@@ -972,32 +972,32 @@ 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"
-            ] = self.cfg.remove_unused_columns
+            training_args_kwargs["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"
-            ] = self.cfg.dataloader_pin_memory
+            training_args_kwargs["dataloader_pin_memory"] = (
+                self.cfg.dataloader_pin_memory
+            )
         if self.cfg.dataloader_num_workers is not None:
-            training_args_kwargs[
-                "dataloader_num_workers"
-            ] = self.cfg.dataloader_num_workers
+            training_args_kwargs["dataloader_num_workers"] = (
+                self.cfg.dataloader_num_workers
+            )
         if self.cfg.dataloader_prefetch_factor is not None:
-            training_args_kwargs[
-                "dataloader_prefetch_factor"
-            ] = self.cfg.dataloader_prefetch_factor
+            training_args_kwargs["dataloader_prefetch_factor"] = (
+                self.cfg.dataloader_prefetch_factor
+            )
         if self.cfg.gradient_checkpointing:
-            training_args_kwargs[
-                "gradient_checkpointing"
-            ] = self.cfg.gradient_checkpointing
+            training_args_kwargs["gradient_checkpointing"] = (
+                self.cfg.gradient_checkpointing
+            )
             if self.cfg.gradient_checkpointing_kwargs is not None:
-                training_args_kwargs[
-                    "gradient_checkpointing_kwargs"
-                ] = self.cfg.gradient_checkpointing_kwargs
+                training_args_kwargs["gradient_checkpointing_kwargs"] = (
+                    self.cfg.gradient_checkpointing_kwargs
+                )
             else:
                 training_args_kwargs["gradient_checkpointing_kwargs"] = {
                     "use_reentrant": False
@@ -1071,9 +1071,9 @@ 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"
-                ] = self.cfg.dpo_use_logits_to_keep
+                training_args_kwargs["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:
@@ -1108,9 +1108,9 @@ 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"
-            ] = self.cfg.precompute_ref_log_probs
+            dpo_trainer_kwargs["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/base.py

@@ -462,9 +462,9 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, Trainer):
                 "pin_memory": self.args.dataloader_pin_memory,
             }
             if self.args.dataloader_prefetch_factor:
-                dataloader_params[
-                    "prefetch_factor"
-                ] = self.args.dataloader_prefetch_factor
+                dataloader_params["prefetch_factor"] = (
+                    self.args.dataloader_prefetch_factor
+                )
 
             sampler = self._get_train_sampler()
             if isinstance(sampler, BatchSampler):
@@ -509,9 +509,9 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, Trainer):
                 "pin_memory": self.args.dataloader_pin_memory,
             }
             if self.args.dataloader_prefetch_factor:
-                dataloader_params[
-                    "prefetch_factor"
-                ] = self.args.dataloader_prefetch_factor
+                dataloader_params["prefetch_factor"] = (
+                    self.args.dataloader_prefetch_factor
+                )
 
             if isinstance(eval_sampler, BatchSampler):
                 dataloader_params["batch_sampler"] = eval_sampler

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

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

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

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

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

@@ -1,6 +1,7 @@
 """
 DPO trainer for axolotl
 """
+
 import gc
 from functools import wraps
 from typing import Any, Dict, Union

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

@@ -45,9 +45,9 @@ class GRPOStrategy:
             )
 
             if trl.vllm_gpu_memory_utilization:
-                grpo_args_kwargs[
-                    "vllm_gpu_memory_utilization"
-                ] = trl.vllm_gpu_memory_utilization
+                grpo_args_kwargs["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"
-            ] = cfg.trl.reward_processing_classes
+            trainer_kwargs["reward_processing_classes"] = (
+                cfg.trl.reward_processing_classes
+            )
         return trainer_kwargs
 
     @classmethod

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

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

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

@@ -1,6 +1,7 @@
 """
 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/trl.py

@@ -1,6 +1,7 @@
 """
 module for TRL PPO training
 """
+
 import torch
 from tqdm import tqdm
 from trl import PPOTrainer

src/axolotl/core/training_args.py

@@ -1,6 +1,7 @@
 """
 extra axolotl specific training args
 """
+
 from dataclasses import dataclass, field
 from typing import Optional
 

src/axolotl/evaluate.py

@@ -8,6 +8,8 @@ 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
@@ -25,18 +27,18 @@ LOG = get_logger("axolotl.evaluate")
 
 
 def evaluate_dataset(
-    trainer, dataset, dataset_type: str, flash_optimum: bool = False
+    trainer: Trainer, dataset: Dataset, dataset_type: str, flash_optimum: bool = False
 ) -> Optional[Dict[str, float]]:
-    """Helper function to evaluate a single dataset safely.
+    """Helper function to evaluate a single dataset.
 
     Args:
-        trainer: The trainer instance
-        dataset: Dataset to evaluate
-        dataset_type: Type of dataset ('train' or 'eval')
-        flash_optimum: Whether to use flash optimum
+        trainer: The trainer instance.
+        dataset: Dataset to evaluate.
+        dataset_type: Type of dataset ('train' or 'eval').
+        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
@@ -63,17 +65,14 @@ 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:
-        Tuple containing:
-        - The model (either PeftModel or PreTrainedModel)
-        - The tokenizer
-        - Dictionary of evaluation metrics
+        Dictionary mapping metric names to their values.
     """
     # pylint: disable=duplicate-code
     # Enable expandable segments for cuda allocation to improve VRAM usage

src/axolotl/integrations/grokfast/__init__.py

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

src/axolotl/integrations/grokfast/args.py

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

src/axolotl/integrations/kd/args.py

@@ -26,12 +26,12 @@ class KDArgs(BaseModel):
     """
 
     kd_trainer: Optional[bool] = None  # whether to use KD trainer
-    kd_ce_alpha: Optional[
-        float
-    ] = None  # loss coefficient for cross-entropy loss during KD
+    kd_ce_alpha: Optional[float] = (
+        None  # loss coefficient for cross-entropy loss during KD
+    )
     kd_alpha: Optional[float] = None  # loss coefficient for KD loss
     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
-    ] = None  # whether to sample top k before softmax during KD
+    kd_top_k_before_softmax: Optional[bool] = (
+        None  # whether to sample top k before softmax during KD
+    )

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"
-                ] = cfg.liger_fused_linear_cross_entropy
+                kwargs["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

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

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

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

src/axolotl/integrations/lm_eval/__init__.py

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

src/axolotl/integrations/lm_eval/args.py

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

src/axolotl/integrations/lm_eval/cli.py

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

src/axolotl/kernels/geglu.py

@@ -5,6 +5,7 @@ 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

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

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

src/axolotl/kernels/swiglu.py

@@ -5,6 +5,7 @@ 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

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

src/axolotl/monkeypatch/attention/mllama.py

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

src/axolotl/monkeypatch/data/batch_dataset_fetcher.py

@@ -1,4 +1,5 @@
 """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

@@ -12,7 +12,9 @@ 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 LlamaAttention
+from transformers.models.llama.modeling_llama import (
+    LlamaAttention,
+)
 from transformers.models.llama.modeling_llama import (
     LlamaDecoderLayer as OriginalLlamaDecoderLayer,
 )
@@ -490,9 +492,11 @@ 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=attention_mask[:, -query_states.size(1) :]
-            if attention_mask is not None
-            else None,
+            query_padding_mask=(
+                attention_mask[:, -query_states.size(1) :]
+                if attention_mask is not None
+                else None
+            ),
         )
         output_unpad = flash_attn_varlen_qkvpacked_func(
             qkv_unpad,
@@ -531,9 +535,11 @@ def flashattn_forward(
                 value_states,
                 kvpacked=True,
                 key_padding_mask=attention_mask,
-                query_padding_mask=attention_mask[:, -query_states.size(1) :]
-                if attention_mask is not None
-                else None,
+                query_padding_mask=(
+                    attention_mask[:, -query_states.size(1) :]
+                    if attention_mask is not 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

@@ -1,6 +1,7 @@
 """
 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

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

src/axolotl/monkeypatch/mixtral/__init__.py

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

src/axolotl/monkeypatch/relora.py

@@ -1,4 +1,5 @@
 """Implements the ReLoRA training procedure from https://arxiv.org/abs/2307.05695, minus the initial full fine-tune."""
+
 import glob
 import json
 import logging
@@ -411,7 +412,10 @@ 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(Path(model_src) / shard_path)
+            in_tensors = torch.load(
+                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/stablelm_attn_hijack_flash.py

@@ -17,7 +17,7 @@
 # https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
 # https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py
 # pylint: disable=duplicate-code
-""" PyTorch StableLM Epoch model. """
+"""PyTorch StableLM Epoch model."""
 import importlib
 import math
 from typing import Optional, Tuple, Union

src/axolotl/monkeypatch/trainer_fsdp_optim.py

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

src/axolotl/monkeypatch/utils.py

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

src/axolotl/monkeypatch/xformers_/__init__.py

@@ -1,6 +1,7 @@
 """
 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

@@ -1,6 +1,7 @@
 """
 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/completion.py

@@ -1,6 +1,7 @@
 """
 Basic completion text
 """
+
 from collections import defaultdict
 from typing import Any, Dict, Generator, Optional, Tuple
 

src/axolotl/prompt_strategies/context_qa.py

@@ -1,4 +1,5 @@
 """Module containing the classes for Context QA Prompt Tokenization Strategies"""
+
 from typing import Tuple
 
 from axolotl.prompt_tokenizers import InstructionPromptTokenizingStrategy

src/axolotl/prompt_strategies/dpo/__init__.py

@@ -1,6 +1,7 @@
 """
 module for DPO style dataset transform strategies
 """
+
 from functools import partial
 
 from ..base import load as load_base

src/axolotl/prompt_strategies/dpo/chatml.py

@@ -33,9 +33,9 @@ def default(
                 f"<|im_start|>user\n{sample[prompt_key]}<|im_end|>\n<|im_start|>assistant\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|im_start|>user\n{sample[prompt_key]}<|im_end|>\n<|im_start|>assistant\n"
+            sample["prompt"] = (
+                f"<|im_start|>user\n{sample[prompt_key]}<|im_end|>\n<|im_start|>assistant\n"
+            )
         sample["chosen"] = f"{sample[chosen_key]}<|im_end|>"
         sample["rejected"] = f"{sample[rejected_key]}<|im_end|>"
         return sample
@@ -52,9 +52,9 @@ def argilla_chat(
     """
 
     def transform_fn(sample):
-        sample[
-            "prompt"
-        ] = f"<|im_start|>user\n{sample['chosen'][0]['content']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["prompt"] = (
+            f"<|im_start|>user\n{sample['chosen'][0]['content']}<|im_end|>\n<|im_start|>assistant\n"
+        )
         sample["chosen"] = f"{sample['chosen'][1]['content']}<|im_end|>"
         sample["rejected"] = f"{sample['rejected'][1]['content']}<|im_end|>"
         return sample
@@ -78,9 +78,9 @@ def icr(
                 f"<|im_start|>user\n{sample['input']}<|im_end|>\n<|im_start|>assistant\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|im_start|>user\n{sample['input']}<|im_end|>\n<|im_start|>assistant\n"
+            sample["prompt"] = (
+                f"<|im_start|>user\n{sample['input']}<|im_end|>\n<|im_start|>assistant\n"
+            )
         sample["chosen"] = f"{sample['chosen']}<|im_end|>"
         sample["rejected"] = f"{sample['rejected']}<|im_end|>"
         return sample
@@ -100,9 +100,9 @@ def intel(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-arg
                 f"<|im_start|>user\n{sample['question']}<|im_end|>\n<|im_start|>assistant\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|im_start|>user\n{sample['question']}<|im_end|>\n<|im_start|>assistant\n"
+            sample["prompt"] = (
+                f"<|im_start|>user\n{sample['question']}<|im_end|>\n<|im_start|>assistant\n"
+            )
         sample["chosen"] = f"{sample['chosen']}<|im_end|>"
         sample["rejected"] = f"{sample['rejected']}<|im_end|>"
         return sample
@@ -120,9 +120,9 @@ def prompt_pairs(
                 f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            sample["prompt"] = (
+                f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            )
         sample["chosen"] = f"{sample['chosen']}<|im_end|>"
         sample["rejected"] = f"{sample['rejected']}<|im_end|>"
         return sample
@@ -142,9 +142,9 @@ def ultra(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-arg
                 f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            sample["prompt"] = (
+                f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            )
         sample["chosen"] = f"{sample['chosen'][1]['content']}<|im_end|>"
         sample["rejected"] = f"{sample['rejected'][1]['content']}<|im_end|>"
         return sample

src/axolotl/prompt_strategies/dpo/llama3.py

@@ -34,9 +34,9 @@ def default(
                 f"<|start_header_id|>user<|end_header_id|>\n\n{sample[prompt_key]}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample[prompt_key]}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            sample["prompt"] = (
+                f"<|start_header_id|>user<|end_header_id|>\n\n{sample[prompt_key]}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
         sample["chosen"] = f"{sample[chosen_key]}<|eot_id|>"
         sample["rejected"] = f"{sample[rejected_key]}<|eot_id|>"
         return sample
@@ -53,9 +53,9 @@ def argilla_chat(
     """
 
     def transform_fn(sample):
-        sample[
-            "prompt"
-        ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['chosen'][0]['content']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+        sample["prompt"] = (
+            f"<|start_header_id|>user<|end_header_id|>\n\n{sample['chosen'][0]['content']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+        )
         sample["chosen"] = f"{sample['chosen'][1]['content']}<|eot_id|>"
         sample["rejected"] = f"{sample['rejected'][1]['content']}<|eot_id|>"
         return sample
@@ -79,9 +79,9 @@ def icr(
                 f"<|start_header_id|>user<|end_header_id|>\n\n{sample['input']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['input']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            sample["prompt"] = (
+                f"<|start_header_id|>user<|end_header_id|>\n\n{sample['input']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
         sample["chosen"] = f"{sample['chosen']}<|eot_id|>"
         sample["rejected"] = f"{sample['rejected']}<|eot_id|>"
         return sample
@@ -101,9 +101,9 @@ def intel(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-arg
                 f"<|start_header_id|>user<|end_header_id|>\n\n{sample['question']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['question']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            sample["prompt"] = (
+                f"<|start_header_id|>user<|end_header_id|>\n\n{sample['question']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
         sample["chosen"] = f"{sample['chosen']}<|eot_id|>"
         sample["rejected"] = f"{sample['rejected']}<|eot_id|>"
         return sample
@@ -121,9 +121,9 @@ def prompt_pairs(
                 f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            sample["prompt"] = (
+                f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
         sample["chosen"] = f"{sample['chosen']}<|eot_id|>"
         sample["rejected"] = f"{sample['rejected']}<|eot_id|>"
         return sample
@@ -143,9 +143,9 @@ def ultra(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-arg
                 f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            sample["prompt"] = (
+                f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
         sample["chosen"] = f"{sample['chosen'][1]['content']}<|eot_id|>"
         sample["rejected"] = f"{sample['rejected'][1]['content']}<|eot_id|>"
         return sample

src/axolotl/prompt_strategies/input_output.py

@@ -1,4 +1,5 @@
 """Module for plain input/output prompt pairs"""
+
 from typing import Generator, Tuple
 
 from axolotl.prompt_tokenizers import PromptTokenizingStrategy

src/axolotl/prompt_strategies/jinja_template_analyzer.py

@@ -1,4 +1,5 @@
 """Module for inspect jinja templates for the variables they use"""
+
 from typing import Dict, Optional, Set, TypedDict, Union
 
 from jinja2 import Environment, meta, nodes

src/axolotl/prompt_strategies/kto/chatml.py

@@ -1,6 +1,7 @@
 """
 KTO strategies for chatml
 """
+
 # pylint: disable=duplicate-code
 
 
@@ -15,9 +16,9 @@ def argilla(
                 f"<|im_start|>user\n{sample['instruction']}<|im_end|>\n<|im_start|>assistant\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|im_start|>user\n{sample['instruction']}<|im_end|>\n<|im_start|>assistant\n"
+            sample["prompt"] = (
+                f"<|im_start|>user\n{sample['instruction']}<|im_end|>\n<|im_start|>assistant\n"
+            )
         sample["completion"] = f"{sample['completion']}<|im_end|>"
         return sample
 
@@ -33,9 +34,9 @@ def argilla_chat(
     """
 
     def transform_fn(sample):
-        sample[
-            "prompt"
-        ] = f"<|im_start|>user\n{sample['chosen'][0]['content']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["prompt"] = (
+            f"<|im_start|>user\n{sample['chosen'][0]['content']}<|im_end|>\n<|im_start|>assistant\n"
+        )
         sample["completion"] = f"{sample['completion'][1]['content']}<|im_end|>"
         return sample
 
@@ -55,9 +56,9 @@ def intel(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-arg
                 f"<|im_start|>user\n{sample['question']}<|im_end|>\n<|im_start|>assistant\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|im_start|>user\n{sample['question']}<|im_end|>\n<|im_start|>assistant\n"
+            sample["prompt"] = (
+                f"<|im_start|>user\n{sample['question']}<|im_end|>\n<|im_start|>assistant\n"
+            )
         sample["completion"] = f"{sample['completion']}<|im_end|>"
         return sample
 
@@ -74,9 +75,9 @@ def prompt_pairs(
                 f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            sample["prompt"] = (
+                f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            )
         sample["completion"] = f"{sample['completion']}<|im_end|>"
         return sample
 
@@ -96,9 +97,9 @@ def ultra(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-arg
                 f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            sample["prompt"] = (
+                f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            )
         sample["completion"] = f"{sample['completion']}<|im_end|>"
         return sample
 

src/axolotl/prompt_strategies/kto/llama3.py

@@ -1,6 +1,7 @@
 """
 KTO strategies for llama-3 chat template
 """
+
 # pylint: disable=duplicate-code
 
 
@@ -15,9 +16,9 @@ def argilla(
                 f"<|start_header_id|>user<|end_header_id|>\n\n{sample['instruction']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['instruction']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            sample["prompt"] = (
+                f"<|start_header_id|>user<|end_header_id|>\n\n{sample['instruction']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
         sample["completion"] = f"{sample['completion']}<|eot_id|>"
         return sample
 
@@ -33,9 +34,9 @@ def argilla_chat(
     """
 
     def transform_fn(sample):
-        sample[
-            "prompt"
-        ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['completion'][0]['content']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+        sample["prompt"] = (
+            f"<|start_header_id|>user<|end_header_id|>\n\n{sample['completion'][0]['content']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+        )
         sample["completion"] = f"{sample['completion'][1]['content']}<|eot_id|>"
         return sample
 
@@ -55,9 +56,9 @@ def intel(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-arg
                 f"<|start_header_id|>user<|end_header_id|>\n\n{sample['question']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['question']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            sample["prompt"] = (
+                f"<|start_header_id|>user<|end_header_id|>\n\n{sample['question']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
         sample["completion"] = f"{sample['completion']}<|eot_id|>"
         return sample
 
@@ -74,9 +75,9 @@ def prompt_pairs(
                 f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            sample["prompt"] = (
+                f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
         sample["completion"] = f"{sample['completion']}<|eot_id|>"
         return sample
 
@@ -96,9 +97,9 @@ def ultra(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-arg
                 f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
             )
         else:
-            sample[
-                "prompt"
-            ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            sample["prompt"] = (
+                f"<|start_header_id|>user<|end_header_id|>\n\n{sample['prompt']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+            )
         sample["completion"] = f"{sample['completion']}<|eot_id|>"
         return sample
 

src/axolotl/prompt_strategies/kto/user_defined.py

@@ -1,6 +1,7 @@
 """
 User-defined KTO strategies
 """
+
 # pylint: disable=duplicate-code
 
 

src/axolotl/prompt_strategies/messages/chat.py

@@ -1,6 +1,7 @@
 """
 Chat dataset wrapping strategy for new internal messages representations
 """
+
 from typing import Any, Callable, Dict, Optional
 
 from axolotl.core.datasets.chat import TokenizedChatDataset

src/axolotl/prompt_strategies/orcamini.py

@@ -9,6 +9,7 @@ this one specifies the system prompt with "### System:".
 
 Not suited/tested for multiple-turn conversations without further adjustments.
 """
+
 from typing import Generator, Union
 
 from axolotl.prompt_strategies.alpaca_w_system import OpenOrcaPromptTokenizingStrategy

src/axolotl/prompt_strategies/orpo/chat_template.py

@@ -1,4 +1,5 @@
 """chatml prompt tokenization strategy for ORPO"""
+
 from typing import Any, Dict, Generator, List, Optional, Tuple
 
 from pydantic import BaseModel

src/axolotl/prompt_strategies/pretrain.py

@@ -1,4 +1,5 @@
 """pretraining prompt strategies"""
+
 from typing import Generator
 
 from transformers import BatchEncoding

src/axolotl/train.py

@@ -406,9 +406,7 @@ def handle_untrained_tokens_fix(
         )
 
 
-def setup_model_and_trainer(
-    cfg: DictDefault, dataset_meta: TrainDatasetMeta
-) -> tuple[
+def setup_model_and_trainer(cfg: DictDefault, dataset_meta: TrainDatasetMeta) -> tuple[
     HFRLTrainerBuilder | HFCausalTrainerBuilder,
     PeftModel | PreTrainedModel,
     PreTrainedTokenizer,

src/axolotl/utils/__init__.py

@@ -40,6 +40,6 @@ def set_pytorch_cuda_alloc_conf():
     torch_major, torch_minor = int(torch_version[0]), int(torch_version[1])
     if torch_major == 2 and torch_minor >= 2:
         if os.getenv("PYTORCH_CUDA_ALLOC_CONF") is None:
-            os.environ[
-                "PYTORCH_CUDA_ALLOC_CONF"
-            ] = "expandable_segments:True,roundup_power2_divisions:16"
+            os.environ["PYTORCH_CUDA_ALLOC_CONF"] = (
+                "expandable_segments:True,roundup_power2_divisions:16"
+            )

src/axolotl/utils/bench.py

@@ -1,4 +1,5 @@
 """Benchmarking and measurement utilities"""
+
 import functools
 
 import torch

src/axolotl/utils/callbacks/__init__.py

@@ -343,9 +343,9 @@ def bench_eval_callback_factory(trainer, tokenizer):
                     bench_refs.extend(combined_bench_names[bench_name]["refs"])
                     bench_preds.extend(combined_bench_names[bench_name]["preds"])
                     if not pd.isna(bench_score):
-                        results[
-                            f"{bench_split}_bench_accuracy_{bench_name}"
-                        ] = bench_score
+                        results[f"{bench_split}_bench_accuracy_{bench_name}"] = (
+                            bench_score
+                        )
                         bench_scores.append(bench_score)
                     else:
                         results[f"{bench_split}_bench_accuracy_{bench_name}"] = 0.0

src/axolotl/utils/callbacks/mlflow_.py

@@ -1,4 +1,5 @@
 """MLFlow module for trainer callbacks"""
+
 import logging
 from shutil import copyfile
 from tempfile import NamedTemporaryFile

src/axolotl/utils/callbacks/perplexity.py

@@ -1,4 +1,5 @@
 """callback to calculate perplexity as an evaluation metric."""
+
 from typing import Dict, List, Optional
 
 import torch

src/axolotl/utils/callbacks/profiler.py

@@ -1,6 +1,7 @@
 """
 HF Trainer callback for creating pytorch profiling snapshots
 """
+
 from pathlib import Path
 from pickle import dump  # nosec B403
 

src/axolotl/utils/chat_templates.py

@@ -2,6 +2,7 @@
 This module provides functionality for selecting chat templates based on user choices.
 These templates are used for formatting messages in a conversation.
 """
+
 import logging
 from typing import TYPE_CHECKING, Any, Dict, Optional
 

src/axolotl/utils/collators/__init__.py

@@ -1,6 +1,7 @@
 """
 shared axolotl collators for multipack, mamba, multimodal
 """
+
 from .batching import (  # noqa: F401
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,

src/axolotl/utils/collators/core.py

@@ -1,4 +1,5 @@
 """
 basic shared collator constants
 """
+
 IGNORE_INDEX = -100

src/axolotl/utils/collators/mamba.py

@@ -1,6 +1,7 @@
 """
 collators for Mamba
 """
+
 from dataclasses import dataclass
 from typing import Dict, Sequence
 

src/axolotl/utils/config/__init__.py

@@ -18,7 +18,11 @@ from axolotl.utils.config.models.input.v0_4_1 import (
 from axolotl.utils.config.models.input.v0_4_1 import (
     AxolotlInputConfig as AxolotlInputConfigBase,
 )
-from axolotl.utils.config.models.input.v0_4_1 import DPODataset, KTODataset, SFTDataset
+from axolotl.utils.config.models.input.v0_4_1 import (
+    DPODataset,
+    KTODataset,
+    SFTDataset,
+)
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_model_config
 

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -200,12 +200,12 @@ class SFTDataset(BaseModel):
     field_human: Optional[str] = None
     field_model: Optional[str] = None
     field_messages: Optional[str] = None
-    message_field_role: Optional[
-        str
-    ] = None  # deprecated, use message_property_mappings
-    message_field_content: Optional[
-        str
-    ] = None  # deprecated, use message_property_mappings
+    message_field_role: Optional[str] = (
+        None  # deprecated, use message_property_mappings
+    )
+    message_field_content: Optional[str] = (
+        None  # deprecated, use message_property_mappings
+    )
     message_property_mappings: Optional[Dict[str, str]] = None
     message_field_training: Optional[str] = None
     message_field_training_detail: Optional[str] = None
@@ -505,9 +505,9 @@ class HyperparametersConfig(BaseModel):
     embedding_lr: Optional[float] = None
     embedding_lr_scale: Optional[float] = None
     weight_decay: Optional[float] = 0.0
-    optimizer: Optional[
-        Union[OptimizerNames, CustomSupportedOptimizers]
-    ] = OptimizerNames.ADAMW_TORCH_FUSED
+    optimizer: Optional[Union[OptimizerNames, CustomSupportedOptimizers]] = (
+        OptimizerNames.ADAMW_TORCH_FUSED
+    )
     optim_args: Optional[Union[str, Dict[str, Any]]] = Field(
         default=None,
         json_schema_extra={"description": "Optional arguments to supply to optimizer."},
@@ -699,9 +699,9 @@ class AxolotlInputConfig(
     reward_model: Optional[bool] = None
     process_reward_model: Optional[bool] = None
     num_labels: Optional[int] = None
-    dpo_use_weighting: Optional[
-        bool
-    ] = None  # whether to use weighting in DPO trainer. If none, default is false in the trainer.
+    dpo_use_weighting: Optional[bool] = (
+        None  # whether to use weighting in DPO trainer. If none, default is false in the trainer.
+    )
     dpo_use_logits_to_keep: Optional[bool] = None
 
     datasets: Optional[
@@ -780,9 +780,9 @@ class AxolotlInputConfig(
 
     # torch_dtype: Optional[torch.dtype]
 
-    gradient_checkpointing: Optional[
-        Union[Literal["unsloth", "offload"], bool]
-    ] = Field(default=False)
+    gradient_checkpointing: Optional[Union[Literal["unsloth", "offload"], bool]] = (
+        Field(default=False)
+    )
     gradient_checkpointing_kwargs: Optional[Dict[str, Any]] = None
 
     unfrozen_parameters: Optional[List[str]] = None
@@ -894,9 +894,9 @@ class AxolotlInputConfig(
     kto_undesirable_weight: Optional[float] = None
     rl_beta: Optional[float] = None
 
-    max_memory: Optional[
-        Dict[Union[int, Literal["cpu", "disk"]], Union[int, str]]
-    ] = None
+    max_memory: Optional[Dict[Union[int, Literal["cpu", "disk"]], Union[int, str]]] = (
+        None
+    )
     gpu_memory_limit: Optional[Union[int, str]] = None
     low_cpu_mem_usage: Optional[bool] = None
 
@@ -1679,6 +1679,30 @@ class AxolotlInputConfig(
 
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_rl_config_gradient_checkpointing(cls, data):
+        # TODO: SalmanMohammadi
+        # Distributed RL with QLoRA + gradient checkpointing
+        # and use_reentrant = True is broken upstream in TRL
+        # pylint: disable=too-many-boolean-expressions
+        if (
+            data.get("rl")
+            and data.get("gradient_checkpointing")
+            and data.get("gradient_checkpointing_kwargs")
+            and data.get("gradient_checkpointing_kwargs").get("use_reentrant")
+            and data.get("load_in_4bit")
+            and data.get("adapter") == "qlora"
+            and data.get("capabilities")
+            and data.get("capabilities").get("n_gpu", 1) > 1
+        ):
+            raise ValueError(
+                "The `use_reentrant: True` implementation of gradient checkpointing "
+                "is not supported for distributed RL training with QLoRA. Please set "
+                "`use_reentrant: False` in `gradient_checkpointing_kwargs`."
+            )
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_kto_config(cls, data):
@@ -1689,15 +1713,6 @@ class AxolotlInputConfig(
             if data.get("remove_unused_columns") is not False:
                 raise ValueError("Set `remove_unused_columns: False` when using kto")
 
-            if data.get("gradient_checkpointing") and not (
-                data.get("gradient_checkpointing_kwargs")
-                and isinstance(data.get("gradient_checkpointing_kwargs"), dict)
-                and data["gradient_checkpointing_kwargs"].get("use_reentrant")
-            ):
-                raise ValueError(
-                    "Set `gradient_checkpointing_kwargs: {use_reentrant: true}` for when kto is enabled"
-                )
-
         return data
 
 

src/axolotl/utils/config/models/internals/__init__.py

@@ -1,4 +1,5 @@
 """module for gpu capabilities"""
+
 from typing import Optional
 
 from pydantic import BaseModel, Field

src/axolotl/utils/data/__init__.py

@@ -1,6 +1,7 @@
 """
 Data processing modules
 """
+
 from axolotl.utils.data.pretraining import (  # noqa: F401
     encode_pretraining,
     wrap_pretraining_dataset,

src/axolotl/utils/distributed.py

@@ -1,6 +1,7 @@
 """
 utility helpers for distributed checks
 """
+
 import os
 import pickle  # nosec
 from contextlib import contextmanager

src/axolotl/utils/environment.py

@@ -1,10 +1,13 @@
 """
 utils to get GPU info for the current environment
 """
+
 from accelerate.utils.environment import (
     check_cuda_p2p_ib_support as accelerate_check_cuda_p2p_ib_support,
 )
-from accelerate.utils.environment import get_gpu_info
+from accelerate.utils.environment import (
+    get_gpu_info,
+)
 
 
 def check_cuda_p2p_ib_support():

src/axolotl/utils/freeze.py

@@ -1,6 +1,7 @@
 """
 module to freeze/unfreeze parameters by name
 """
+
 import logging
 import re
 from typing import Callable, List, Tuple, Union

src/axolotl/utils/gradient_checkpointing/__init__.py

@@ -1,4 +1,5 @@
 """custom checkpointing utils"""
+
 from axolotl.utils.gradient_checkpointing.unsloth import (
     Unsloth_Offloaded_Gradient_Checkpointer,
 )

src/axolotl/utils/model_shard_quant.py

@@ -1,6 +1,7 @@
 """
 module to handle loading model on cpu/meta device for FSDP
 """
+
 import os
 import time
 from typing import List, Optional, Type, Union
@@ -45,13 +46,13 @@ def _replace_linear(
 
         if isinstance(module, torch.nn.Linear) and name not in skip_modules:
             if issubclass(linear_replacement, Linear4bit):
-                model._modules[  # pylint: disable=protected-access
-                    name
-                ] = linear_replacement(
-                    module.in_features,
-                    module.out_features,
-                    module.bias is not None,
-                    **kwargs,
+                model._modules[name] = (  # pylint: disable=protected-access
+                    linear_replacement(
+                        module.in_features,
+                        module.out_features,
+                        module.bias is not None,
+                        **kwargs,
+                    )
                 )
             else:
                 raise ValueError(

src/axolotl/utils/models.py

@@ -741,9 +741,9 @@ class ModelLoader:
                 )
             else:
                 if self.cfg.gptq_disable_exllama is not None:
-                    self.model_config.quantization_config[
-                        "disable_exllama"
-                    ] = self.cfg.gptq_disable_exllama
+                    self.model_config.quantization_config["disable_exllama"] = (
+                        self.cfg.gptq_disable_exllama
+                    )
                 self.model_kwargs["quantization_config"] = GPTQConfig(
                     **self.model_config.quantization_config
                 )