refactor build() into smaller fns

* lean mistral ft tests, remove e2e torch 2.4.1 test

* make sure to pass save_only_model for RL

* more tests to make ci leaner, add cleanup to modal ci

* fix module for import in e2e tests

* use mp spawn to prevent deadlocks with packing

* make sure cleanup shell script is executable when cloned out

.github/workflows/tests-nightly.yml

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

.github/workflows/tests.yml

@@ -44,12 +44,98 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
-  pytest:
-    name: PyTest
+  preload-cache:
+    name: Preload HF cache
     runs-on: ubuntu-latest
     strategy:
       fail-fast: false
-      max-parallel: 2
+      matrix:
+        python_version: ["3.11"]
+        pytorch_version: ["2.6.0"]
+    timeout-minutes: 20
+
+    env:
+      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
+
+    steps:
+      - name: Check out repository code
+        uses: actions/checkout@v4
+
+      - name: Restore HF cache
+        id: hf-cache-restore
+        uses: actions/cache/restore@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ runner.os }}-hf-hub-cache-v2
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python_version }}
+          cache: 'pip' # caching pip dependencies
+
+      - name: upgrade pip
+        run: |
+          pip3 install --upgrade pip
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+
+      - name: Install PyTorch
+        run: |
+          pip3 install torch==${{ matrix.pytorch_version }}
+
+      - name: Install dependencies
+        run: |
+          pip3 show torch
+          pip3 install --no-build-isolation -U -e .
+          python scripts/unsloth_install.py | sh
+          python scripts/cutcrossentropy_install.py | sh
+          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+
+      - name: Make sure PyTorch version wasn't clobbered
+        run: |
+          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+
+      - name: Ensure axolotl CLI was installed
+        run: |
+          axolotl --help
+
+      - name: Pre-Download dataset fixture
+        run: |
+          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+
+      - name: Run tests
+        run: |
+          pytest -v tests/conftest.py
+
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v5
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          files: ./coverage.xml
+          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+          fail_ci_if_error: false
+
+      - name: cleanup pip cache
+        run: |
+          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+
+      - name: Save HF cache
+        id: hf-cache
+        uses: actions/cache/save@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+
+  pytest:
+    name: PyTest
+    runs-on: ubuntu-latest
+    needs: [preload-cache]
+    strategy:
+      fail-fast: false
       matrix:
         python_version: ["3.11"]
         pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
@@ -121,21 +207,12 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
-      - name: Save HF cache
-        id: hf-cache
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
-
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
+    needs: [preload-cache]
     strategy:
       fail-fast: false
-      max-parallel: 1
       matrix:
         python_version: ["3.11"]
         pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
@@ -199,15 +276,6 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
-      - name: Save HF cache
-        id: hf-cache
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
-
   docker-e2e-tests-1st:
     if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
@@ -267,12 +335,6 @@ jobs:
             pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras: llmcompressor
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            num_gpus: 1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -309,3 +371,43 @@ jobs:
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
+
+  docker-e2e-cleanup:
+    runs-on: [self-hosted, modal]
+    timeout-minutes: 90
+    needs: [docker-e2e-tests]
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            num_gpus: 1
+            axolotl_extras: vllm
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Install Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Install Modal
+        run: |
+          python -m pip install --upgrade pip
+          pip install modal==0.71.8 jinja2
+      - name: Update env vars
+        run: |
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
+          echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
+          echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
+          echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
+          echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
+          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
+      - name: Run tests job on Modal
+        run: |
+          modal run cicd.cleanup

cicd/cicd.sh

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

cicd/cleanup.py

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

cicd/cleanup.sh

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

cicd/e2e_tests.py

@@ -1,69 +1,6 @@
 """Modal app to run axolotl GPU tests"""
 
-# pylint: disable=duplicate-code
-
-import os
-import pathlib
-import tempfile
-
-import jinja2
-import modal
-from jinja2 import select_autoescape
-from modal import App, Image
-
-cicd_path = pathlib.Path(__file__).parent.resolve()
-
-template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
-template_env = jinja2.Environment(
-    loader=template_loader, autoescape=select_autoescape()
-)
-df_template = template_env.get_template("Dockerfile.jinja")
-
-df_args = {
-    "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
-    "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
-    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.4.1"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu121-2.4.1"),
-    "CUDA": os.environ.get("CUDA", "121"),
-    "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
-    "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
-    "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
-    "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
-    "HF_HOME": "/workspace/data/huggingface-cache/hub",
-}
-
-dockerfile_contents = df_template.render(**df_args)
-
-temp_dir = tempfile.mkdtemp()
-with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
-    f.write(dockerfile_contents)
-
-cicd_image = Image.from_dockerfile(
-    pathlib.Path(temp_dir) / "Dockerfile",
-    context_mount=None,
-    force_build=True,
-    gpu="A10G",
-).env(df_args)
-
-app = App("Axolotl CI/CD", secrets=[])
-
-hf_cache_volume = modal.Volume.from_name(
-    "axolotl-ci-hf-hub-cache", create_if_missing=True
-)
-VOLUME_CONFIG = {
-    "/workspace/data/huggingface-cache/hub": hf_cache_volume,
-}
-
-N_GPUS = int(os.environ.get("N_GPUS", 1))
-GPU_CONFIG = modal.gpu.L40S(count=N_GPUS)
-
-
-def run_cmd(cmd: str, run_folder: str):
-    import subprocess  # nosec
-
-    # Propagate errors from subprocess.
-    if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
-        exit(exit_code)  # pylint: disable=consider-using-sys-exit
+from .single_gpu import GPU_CONFIG, VOLUME_CONFIG, app, cicd_image, run_cmd
 
 
 @app.function(

cicd/single_gpu.py

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

docs/config.qmd

@@ -32,6 +32,8 @@ tokenizer_legacy:
 resize_token_embeddings_to_32x:
 # Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.
 shrink_embeddings:
+# Optional[bool] Don't upcast the embeddings to float32 when using PEFT. Useful for low-VRAM GPUs
+embeddings_skip_upcast:
 # Whether to load the model with randomly initialized weights. Useful for
 # pre-training a model from scratch or debugging purposes.
 random_init_weights:
@@ -73,11 +75,12 @@ load_in_8bit: true
 load_in_4bit:
 
 # Use CUDA bf16
-bf16: true # bool or 'full' for `bf16_full_eval`. require >=ampere
+bf16: true # bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection. require >=ampere
 # Use CUDA fp16
 fp16: true
 # Use CUDA tf32
 tf32: true # require >=ampere
+# Note: if bf16 is set to 'auto', and fp16 is set to true, we will prefer the explict fp16 setting
 
 # No AMP (automatic mixed precision)
 bfloat16: true # require >=ampere
@@ -184,8 +187,8 @@ datasets:
     # adding a system turn with empty content.
     drop_system_message:
 
-    # Optional[bool]. Whether to split the assistant turn based on a reasoning trace inside delimited tags
-    # defaults to False
+    # Optional[bool]. (for Qwen3 template only) Whether to split the assistant content based on a reasoning trace inside delimited tags
+    # See example at `docs/dataset-formats/conversation.qmd`
     split_thinking:
 
     # IMPORTANT: The following fields determine which parts of the conversation to train on.
@@ -547,7 +550,7 @@ gradient_checkpointing: false
 early_stopping_patience: 3
 
 # Specify a scheduler and kwargs to use with the optimizer
-lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine
+lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | 'linear' | 'cosine_with_restarts' | 'polynomial' | 'constant' | 'constant_with_warmup' | 'inverse_sqrt' | 'reduce_lr_on_plateau' | 'cosine_with_min_lr' | 'warmup_stable_decay' | empty for cosine
 lr_scheduler_kwargs:
 cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr
 cosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)
@@ -609,6 +612,7 @@ lr_div_factor: # Learning rate div factor
 # - optimi_adamw
 # - ao_adamw_8bit
 # - ao_adamw_fp8
+# - came_pytorch
 optimizer:
 # Dictionary of arguments to pass to the optimizer
 optim_args:

docs/dataset-formats/conversation.qmd

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

examples/llama-4/README.md

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

examples/orpheus/README.md

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

examples/orpheus/finetune.yml

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

requirements.txt

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

setup.py

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

src/axolotl/cli/art.py

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

src/axolotl/cli/preprocess.py

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

src/axolotl/cli/train.py

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

src/axolotl/common/datasets.py

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

src/axolotl/core/trainers/base.py

@@ -114,6 +114,8 @@ class AxolotlTrainer(
             packing_efficiency_estimate=self.args.sample_packing_efficiency,
             batch_max_len=batch_max_len,
             batch_size=batch_size,
+            group_size=self.args.sample_packing_group_size,
+            bin_size=self.args.sample_packing_bin_size,
             sequential=self.args.sample_packing_sequentially,
             drop_last=True,
         )

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

@@ -0,0 +1,21 @@
+# Copyright 2024 Axolotl AI. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Init for axolotl.core.trainers.builders"""
+
+# pylint: disable=unused-import
+# flake8: noqa
+
+from .causal import HFCausalTrainerBuilder
+from .rl import HFRLTrainerBuilder

src/axolotl/core/trainers/builders/base.py

@@ -0,0 +1,331 @@
+"""Base class trainer / training args builder implementation"""
+
+import abc
+from typing import Any
+
+from torch import Type
+from transformers import TrainerCallback
+from transformers.training_args import TrainingArguments
+
+from axolotl.integrations.base import PluginManager
+from axolotl.monkeypatch.trainer.lr import patch_trainer_get_lr
+from axolotl.utils import is_comet_available, is_mlflow_available
+from axolotl.utils.callbacks import GCCallback, SaveAxolotlConfigtoWandBCallback
+from axolotl.utils.callbacks.profiler import PytorchProfilerCallback
+
+PLUGIN_MANAGER = PluginManager.get_instance()
+
+
+class TrainerBuilderBase(abc.ABC):
+    """Base class for trainer builder."""
+
+    _train_dataset = None
+    _eval_dataset = None
+    _model_ref = None
+    _peft_config = None
+
+    def __init__(self, cfg, model, tokenizer, processor=None):
+        self.cfg = cfg
+        self.model = model
+        self.tokenizer = tokenizer
+        self.processor = processor
+
+        # If the model supports tagging, add the axolotl tag.
+        # This makes sure the tag is correctly pushed even if a user calls
+        # model.push_to_hub instead of trainer.push_to_hub.
+        if hasattr(model, "add_model_tags"):
+            model.add_model_tags(["axolotl"])
+
+        patch_trainer_get_lr()
+
+    @property
+    def model_ref(self):
+        return self._model_ref
+
+    @model_ref.setter
+    def model_ref(self, model):
+        self._model_ref = model
+
+    @property
+    def train_dataset(self):
+        return self._train_dataset
+
+    @train_dataset.setter
+    def train_dataset(self, dataset):
+        self._train_dataset = dataset
+
+    @property
+    def eval_dataset(self):
+        return self._eval_dataset
+
+    @eval_dataset.setter
+    def eval_dataset(self, dataset):
+        self._eval_dataset = dataset
+
+    @property
+    def peft_config(self):
+        return self._peft_config
+
+    @peft_config.setter
+    def peft_config(self, peft_config):
+        self._peft_config = peft_config
+
+    @abc.abstractmethod
+    def build(self, total_num_steps):
+        pass
+
+    def get_common_training_args_kwargs(
+        self, total_num_steps: int | None = None
+    ) -> dict[str, Any]:
+        """Get common training arguments kwargs used across different trainer types."""
+        training_args_kwargs = {}
+
+        # Common parameters
+        for arg in [
+            "adam_beta1",
+            "adam_beta2",
+            "adam_epsilon",
+            "max_grad_norm",
+            "dataloader_num_workers",
+            "dataloader_pin_memory",
+            "dataloader_prefetch_factor",
+            "dataloader_drop_last",
+            "remove_unused_columns",
+        ]:
+            if hasattr(self.cfg, arg) and getattr(self.cfg, arg) is not None:
+                training_args_kwargs[arg] = getattr(self.cfg, arg)
+
+        # Add Hub integration arguments if needed
+        if self.cfg.hub_model_id:
+            training_args_kwargs["hub_model_id"] = self.cfg.hub_model_id
+            training_args_kwargs["push_to_hub"] = True
+            training_args_kwargs["hub_private_repo"] = True
+            training_args_kwargs["hub_always_push"] = True
+
+            if self.cfg.hub_strategy:
+                training_args_kwargs["hub_strategy"] = self.cfg.hub_strategy
+
+        # BF16/FP16 settings
+        if hasattr(self.cfg, "bf16") and self.cfg.bf16:
+            if self.cfg.bf16 == "full":
+                training_args_kwargs["bf16_full_eval"] = True
+            else:
+                training_args_kwargs["bf16"] = self.cfg.bf16
+        elif hasattr(self.cfg, "bfloat16") and self.cfg.bfloat16:
+            training_args_kwargs["bf16"] = True
+
+        if hasattr(self.cfg, "fp16"):
+            training_args_kwargs["fp16"] = (
+                getattr(self.cfg, "fp16", False)
+                and not getattr(self.cfg, "bf16", False)
+            ) or False
+
+        # Set save_strategy and save_steps
+        if self.cfg.save_steps:
+            training_args_kwargs["save_strategy"] = "steps"
+            training_args_kwargs["save_steps"] = self.cfg.save_steps
+        elif self.cfg.save_strategy:
+            training_args_kwargs["save_strategy"] = self.cfg.save_strategy
+        else:
+            # default to saving each epoch if not defined
+            training_args_kwargs["save_strategy"] = "epoch"
+
+        # Handle safetensors
+        if self.cfg.save_safetensors is not None:
+            training_args_kwargs["save_safetensors"] = self.cfg.save_safetensors
+
+        # Handle gradient checkpointing
+        if 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
+                )
+
+        # Common optimizer and LR scheduler settings
+        training_args_kwargs["optim"] = self.cfg.optimizer
+        if hasattr(self.cfg, "lr_scheduler") and self.cfg.lr_scheduler:
+            training_args_kwargs["lr_scheduler_type"] = self.cfg.lr_scheduler
+        else:
+            training_args_kwargs["lr_scheduler_type"] = "cosine"
+
+        if hasattr(self.cfg, "lr_scheduler_kwargs") and self.cfg.lr_scheduler_kwargs:
+            training_args_kwargs["lr_scheduler_kwargs"] = self.cfg.lr_scheduler_kwargs
+        else:
+            training_args_kwargs["lr_scheduler_kwargs"] = {}
+
+        # LoRA+ specific settings
+        if hasattr(self.cfg, "loraplus_lr_ratio"):
+            training_args_kwargs["loraplus_lr_ratio"] = self.cfg.loraplus_lr_ratio
+        if hasattr(self.cfg, "loraplus_lr_embedding"):
+            training_args_kwargs["loraplus_lr_embedding"] = (
+                self.cfg.loraplus_lr_embedding
+            )
+
+        # Reporting tools
+        report_to = []
+        if self.cfg.use_wandb:
+            report_to.append("wandb")
+            if self.cfg.wandb_name:
+                training_args_kwargs["run_name"] = self.cfg.wandb_name
+        if self.cfg.use_mlflow:
+            report_to.append("mlflow")
+        if self.cfg.use_tensorboard:
+            report_to.append("tensorboard")
+        if self.cfg.use_comet:
+            report_to.append("comet_ml")
+
+        if report_to:
+            training_args_kwargs["report_to"] = report_to
+
+        # Basic training settings
+        if hasattr(self.cfg, "sequence_len"):
+            training_args_kwargs["max_length"] = self.cfg.sequence_len
+
+        training_args_kwargs["save_only_model"] = getattr(
+            self.cfg, "save_only_model", False
+        )
+        training_args_kwargs["save_total_limit"] = getattr(
+            self.cfg, "save_total_limit", 5
+        )
+
+        # Compute warmup steps
+        if hasattr(self.cfg, "warmup_steps") and self.cfg.warmup_steps is not None:
+            training_args_kwargs["warmup_steps"] = self.cfg.warmup_steps
+        elif (
+            total_num_steps
+            and hasattr(self.cfg, "warmup_ratio")
+            and self.cfg.warmup_ratio is not None
+        ):
+            training_args_kwargs["warmup_steps"] = max(
+                int(self.cfg.warmup_ratio * total_num_steps), 0
+            )
+        elif total_num_steps:
+            training_args_kwargs["warmup_steps"] = min(int(0.03 * total_num_steps), 100)
+
+        return training_args_kwargs
+
+    def create_training_args(
+        self,
+        args_cls: Type[TrainingArguments],
+        total_num_steps: int | None = None,
+        **additional_kwargs,
+    ) -> TrainingArguments:
+        """Create training arguments with common logic."""
+        # Get common trainings args and update with trainer-specific args
+        training_args_kwargs = self.get_common_training_args_kwargs(total_num_steps)
+        training_args_kwargs.update(additional_kwargs)
+
+        # Create training args with pre- and post-creation hooks
+        training_args_kwargs = self.hook_pre_create_training_args(training_args_kwargs)
+        training_args = args_cls(**training_args_kwargs)
+        training_args = self.hook_post_create_training_args(training_args)
+
+        # Unset run_name so wandb sets up experiment names properly
+        if self.cfg.use_wandb and training_args.run_name == training_args.output_dir:
+            training_args.run_name = None
+
+        return training_args
+
+    def create_trainer(
+        self, trainer_cls, training_args, trainer_args=None, trainer_kwargs=None
+    ):
+        """Create trainer with common logic."""
+        if trainer_args is None:
+            trainer_args = []
+        if trainer_kwargs is None:
+            trainer_kwargs = {}
+
+        # Create trainer with pre- and post- creation hooks
+        trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
+            trainer_kwargs, trainer_cls
+        )
+        trainer = trainer_cls(
+            *trainer_args,
+            args=training_args,
+            train_dataset=self.train_dataset,
+            eval_dataset=self.eval_dataset,
+            callbacks=self.get_callbacks(),
+            **trainer_kwargs,
+        )
+        trainer = self.hook_post_create_trainer(trainer)
+
+        # Add post-creation callbacks
+        for callback in self.get_post_trainer_create_callbacks(trainer):
+            trainer.add_callback(callback)
+
+        return trainer
+
+    def get_callbacks(self) -> list[TrainerCallback]:
+        callbacks = []
+        callbacks.extend(
+            PLUGIN_MANAGER.add_callbacks_pre_trainer(cfg=self.cfg, model=self.model)
+        )
+
+        if self.cfg.profiler_steps:
+            callbacks.append(
+                PytorchProfilerCallback(
+                    steps_to_profile=self.cfg.profiler_steps,
+                )
+            )
+
+        if self.cfg.gc_steps:
+            callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
+
+        if self.cfg.use_wandb:
+            callbacks.append(
+                SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
+            )
+        if self.cfg.use_mlflow and is_mlflow_available():
+            from axolotl.utils.callbacks.mlflow_ import (
+                SaveAxolotlConfigtoMlflowCallback,
+            )
+
+            callbacks.extend(
+                [
+                    SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path),
+                ]
+            )
+        if self.cfg.use_comet and is_comet_available():
+            from axolotl.utils.callbacks.comet_ import SaveAxolotlConfigtoCometCallback
+
+            callbacks.append(
+                SaveAxolotlConfigtoCometCallback(self.cfg.axolotl_config_path)
+            )
+
+        return callbacks
+
+    def get_post_trainer_create_callbacks(self, trainer):
+        """Callbacks added after the trainer is created, usually because these need
+        access to the trainer.
+        """
+        callbacks = []
+        if self.cfg.plugins:
+            callbacks.extend(
+                [
+                    cb
+                    for cb in PLUGIN_MANAGER.add_callbacks_post_trainer(
+                        self.cfg, trainer
+                    )
+                    if cb
+                ]
+            )
+        return callbacks
+
+    def hook_pre_create_training_args(self, training_arguments_kwargs):
+        # TODO
+        return training_arguments_kwargs
+
+    def hook_post_create_training_args(self, training_arguments):
+        # TODO
+        return training_arguments
+
+    def hook_pre_create_trainer(self, trainer_kwargs, trainer_cls):
+        # TODO
+        return trainer_kwargs, trainer_cls
+
+    def hook_post_create_trainer(self, trainer):
+        # TODO
+        return trainer

src/axolotl/core/trainers/builders/causal.py

@@ -0,0 +1,619 @@
+"""Causal trainer / training args builder implementation"""
+
+import importlib
+import inspect
+import logging
+import math
+import os
+import sys
+from pathlib import Path
+from typing import Type
+
+import transformers
+from transformers import (
+    DataCollatorWithFlattening,
+    EarlyStoppingCallback,
+)
+from transformers.training_args import OptimizerNames
+from trl.trainer.utils import RewardDataCollatorWithPadding
+
+from axolotl.core.trainers.base import AxolotlTrainer
+from axolotl.core.trainers.builders.base import TrainerBuilderBase
+from axolotl.core.trainers.mamba import AxolotlMambaTrainer
+from axolotl.core.trainers.relora import ReLoRATrainer
+from axolotl.core.trainers.trl import AxolotlPRMTrainer, AxolotlRewardTrainer
+from axolotl.core.training_args import (
+    AxolotlPRMConfig,
+    AxolotlRewardConfig,
+    AxolotlTrainingArguments,
+)
+from axolotl.integrations.base import PluginManager
+from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
+from axolotl.monkeypatch.relora import ReLoRACallback
+from axolotl.processing_strategies import get_processing_strategy
+from axolotl.utils import is_comet_available, is_mlflow_available
+from axolotl.utils.callbacks import (
+    EvalFirstStepCallback,
+    GPUStatsCallback,
+    LossWatchDogCallback,
+    SaveBetterTransformerModelCallback,
+    bench_eval_callback_factory,
+    causal_lm_bench_eval_callback_factory,
+    colab_inference_post_train_callback,
+    log_prediction_callback_factory,
+)
+from axolotl.utils.callbacks.lisa import lisa_callback_factory
+from axolotl.utils.chat_templates import get_chat_template_from_config
+from axolotl.utils.collators.batching import (
+    BatchSamplerDataCollatorForSeq2Seq,
+    DataCollatorForSeq2Seq,
+    V2BatchSamplerDataCollatorForSeq2Seq,
+)
+from axolotl.utils.collators.mamba import MambaDataCollator
+from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
+from axolotl.utils.schemas.enums import CustomSupportedOptimizers
+
+LOG = logging.getLogger(__name__)
+PLUGIN_MANAGER = PluginManager.get_instance()
+
+
+class HFCausalTrainerBuilder(TrainerBuilderBase):
+    """Build the HuggingFace training args / trainer for causal models and reward
+    modeling using TRL.
+    """
+
+    def get_callbacks(self):
+        callbacks = super().get_callbacks()
+        callbacks.append(GPUStatsCallback(self.cfg))
+        callbacks.append(EvalFirstStepCallback())
+
+        if self.cfg.relora_steps:
+            callbacks.append(ReLoRACallback(self.cfg))
+
+        if (
+            hasattr(self.model, "use_bettertransformer")
+            and self.model.use_bettertransformer is True
+        ):
+            callbacks.append(SaveBetterTransformerModelCallback())
+
+        if self.cfg.loss_watchdog_threshold is not None:
+            callbacks.append(LossWatchDogCallback(self.cfg))
+
+        return callbacks
+
+    def get_post_trainer_create_callbacks(self, trainer):
+        callbacks = []
+        if self.cfg.use_wandb and self.cfg.eval_table_size > 0:
+            LogPredictionCallback = log_prediction_callback_factory(
+                trainer, self.tokenizer, "wandb"
+            )
+            callbacks.append(LogPredictionCallback(self.cfg))
+        if (
+            self.cfg.use_mlflow
+            and is_mlflow_available()
+            and self.cfg.eval_table_size > 0
+        ):
+            LogPredictionCallback = log_prediction_callback_factory(
+                trainer, self.tokenizer, "mlflow"
+            )
+            callbacks.append(LogPredictionCallback(self.cfg))
+        if self.cfg.use_comet and is_comet_available() and self.cfg.eval_table_size > 0:
+            LogPredictionCallback = log_prediction_callback_factory(
+                trainer, self.tokenizer, "comet_ml"
+            )
+            callbacks.append(LogPredictionCallback(self.cfg))
+
+        if self.cfg.do_bench_eval:
+            callbacks.append(bench_eval_callback_factory(trainer, self.tokenizer))
+        if self.cfg.do_causal_lm_eval:
+            CausalLMBenchEvalCallback = causal_lm_bench_eval_callback_factory(
+                trainer, self.tokenizer
+            )
+            callbacks.append(CausalLMBenchEvalCallback(self.cfg))
+
+        if self.cfg.early_stopping_patience:
+            early_stop_cb = EarlyStoppingCallback(
+                self.cfg.early_stopping_patience,
+            )
+            callbacks.append(early_stop_cb)
+
+        if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
+            callbacks.append(lisa_callback_factory(trainer))
+
+        if any("COLAB_" in key for key in os.environ):
+            ColabCallback = colab_inference_post_train_callback(trainer)
+            callbacks.append(ColabCallback(self.cfg))
+
+        callbacks.extend(super().get_post_trainer_create_callbacks(trainer=trainer))
+        return callbacks
+
+    def _get_trainer_cls(self):
+        if self.cfg.plugins:
+            trainer_cls = PLUGIN_MANAGER.get_trainer_cls(self.cfg)
+            if trainer_cls:
+                return trainer_cls
+        if self.cfg.relora_steps:
+            return ReLoRATrainer
+        if self.cfg.model_config_type == "mamba":
+            return AxolotlMambaTrainer
+        if self.cfg.reward_model:
+            return AxolotlRewardTrainer
+        if self.cfg.process_reward_model:
+            return AxolotlPRMTrainer
+
+        return AxolotlTrainer
+
+    def build(self, total_num_steps):
+        """Build and return a causal trainer instance using the refactored base class."""
+        # Get trainer class
+        trainer_cls = self._get_trainer_cls()
+
+        # Prepare training arguments
+        training_args = self._prepare_training_args(total_num_steps)
+
+        # Prepare data collators
+        data_collator_kwargs = self._prepare_data_collator_kwargs()
+
+        # Prepare trainer kwargs
+        trainer_kwargs = self._prepare_trainer_kwargs(
+            trainer_cls=trainer_cls,
+            data_collator_kwargs=data_collator_kwargs,
+            training_args=training_args,
+        )
+
+        # Create the trainer
+        trainer = self.create_trainer(
+            trainer_cls=trainer_cls,
+            training_args=training_args,
+            trainer_kwargs={
+                "model": self.model,
+                "data_collator": self.build_collator(
+                    training_args, **data_collator_kwargs
+                ),
+                **trainer_kwargs,
+            },
+        )
+
+        # Handle DeepSpeed config for sample packing if needed
+        if self.cfg.deepspeed and self.cfg.sample_packing:
+            trainer.accelerator.state.deepspeed_plugin.deepspeed_config[
+                "train_micro_batch_size_per_gpu"
+            ] = self.cfg.micro_batch_size
+
+        return trainer
+
+    def _prepare_training_args(self, total_num_steps):
+        """Prepare and return training arguments."""
+        # Base training arguments
+        training_args_kwargs = self._get_base_training_args()
+
+        # Add feature configurations
+        self._add_feature_configs(training_args_kwargs)
+
+        # Handle optimizer configuration
+        self._configure_optimizer(training_args_kwargs)
+
+        # Create training args using the base class method
+        training_args_cls = self._get_training_args_cls()
+
+        return self.create_training_args(
+            args_cls=training_args_cls,
+            total_num_steps=total_num_steps,
+            **training_args_kwargs,
+        )
+
+    def _get_base_training_args(self):
+        """Return the base training arguments."""
+        return {
+            "max_steps": self.cfg.max_steps if self.cfg.max_steps else -1,
+            "max_seq_length": self.cfg.sequence_len,
+            "per_device_train_batch_size": self.cfg.micro_batch_size,
+            "gradient_accumulation_steps": self.cfg.gradient_accumulation_steps,
+            "eval_accumulation_steps": self.cfg.gradient_accumulation_steps,
+            "num_train_epochs": self.cfg.num_epochs,
+            "learning_rate": self.cfg.learning_rate,
+            "output_dir": self.cfg.output_dir,
+            "weight_decay": (
+                self.cfg.weight_decay if self.cfg.weight_decay is not None else 0.0
+            ),
+            "model_type": self.cfg.model_config_type,
+            "pretraining": bool(self.cfg.pretraining_dataset),
+            "sequence_parallel_degree": self.cfg.sequence_parallel_degree,
+            "ring_attn_func": self.cfg.ring_attn_func,
+            "embedding_lr": self.cfg.embedding_lr,
+            "embedding_lr_scale": self.cfg.embedding_lr_scale,
+            "loraplus_lr_ratio": self.cfg.loraplus_lr_ratio,
+            "loraplus_lr_embedding": self.cfg.loraplus_lr_embedding,
+            "lr_groups": self.cfg.lr_groups,
+        }
+
+    def _add_feature_configs(self, training_args_kwargs):
+        """Add various feature configurations."""
+        # Sample packing configurations
+        self._add_sample_packing_configs(training_args_kwargs)
+
+        # Batch size configurations
+        if self.cfg.eval_batch_size:
+            training_args_kwargs["per_device_eval_batch_size"] = (
+                self.cfg.eval_batch_size
+            )
+        if self.cfg.auto_find_batch_size is not None:
+            training_args_kwargs["auto_find_batch_size"] = self.cfg.auto_find_batch_size
+
+        # Advanced training techniques (ReLoRA & Lisa)
+        self._add_advanced_training_configs(training_args_kwargs)
+
+        # Model-specific configurations
+        self._add_model_specific_configs(training_args_kwargs)
+
+    def _add_sample_packing_configs(self, training_args_kwargs):
+        """Add sample packing configurations if applicable."""
+        if hasattr(self.cfg, "sample_packing") and self.cfg.sample_packing:
+            training_args_kwargs.update(
+                {
+                    "sample_packing": bool(self.cfg.sample_packing),
+                    "multipack_real_batches": not self.cfg.flash_attention
+                    or self.cfg.multipack_real_batches,
+                    "eval_sample_packing": bool(self.cfg.eval_sample_packing),
+                }
+            )
+
+            if self.cfg.sample_packing_bin_size is not None:
+                training_args_kwargs["sample_packing_bin_size"] = (
+                    self.cfg.sample_packing_bin_size
+                )
+
+            if self.cfg.sample_packing_group_size is not None:
+                training_args_kwargs["sample_packing_group_size"] = (
+                    self.cfg.sample_packing_group_size
+                )
+
+            if self.cfg.sample_packing_eff_est:
+                training_args_kwargs["sample_packing_efficiency"] = (
+                    self.cfg.sample_packing_eff_est
+                )
+
+    def _add_advanced_training_configs(self, training_args_kwargs):
+        """Add advanced training techniques configurations (ReLoRA & Lisa)."""
+        # ReLoRA configurations
+        if self.cfg.relora_steps:
+            training_args_kwargs.update(
+                {
+                    "relora_steps": self.cfg.relora_steps,
+                    "relora_warmup_steps": self.cfg.relora_warmup_steps,
+                }
+            )
+            if self.cfg.relora_anneal_steps:
+                training_args_kwargs["relora_anneal_steps"] = (
+                    self.cfg.relora_anneal_steps
+                )
+            if self.cfg.relora_prune_ratio:
+                training_args_kwargs["relora_prune_ratio"] = self.cfg.relora_prune_ratio
+
+        # Lisa configurations
+        if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
+            training_args_kwargs.update(
+                {
+                    "lisa_n_layers": self.cfg.lisa_n_layers,
+                    "lisa_step_interval": self.cfg.lisa_step_interval,
+                    "lisa_layers_attribute": self.cfg.lisa_layers_attribute,
+                }
+            )
+
+    def _add_model_specific_configs(self, training_args_kwargs):
+        """Add model-specific configurations."""
+        # Chat template
+        if self.cfg.chat_template:
+            training_args_kwargs["chat_template"] = get_chat_template_from_config(
+                cfg=self.cfg,
+                tokenizer=self.tokenizer,
+            )
+
+        # NEFTune
+        if self.cfg.neftune_noise_alpha is not None:
+            training_args_kwargs["neftune_noise_alpha"] = self.cfg.neftune_noise_alpha
+
+        # Knowledge distillation configurations
+        if self.cfg.kd_ce_alpha is not None:
+            training_args_kwargs["kd_ce_alpha"] = self.cfg.kd_ce_alpha
+        if self.cfg.kd_alpha is not None:
+            training_args_kwargs["kd_alpha"] = self.cfg.kd_alpha
+        if self.cfg.kd_temperature is not None:
+            training_args_kwargs["kd_temperature"] = self.cfg.kd_temperature
+        if self.cfg.kd_zscore_base_temp is not None:
+            training_args_kwargs["kd_zscore_base_temp"] = self.cfg.kd_zscore_base_temp
+        if self.cfg.kd_top_k_before_softmax is not None:
+            training_args_kwargs["kd_top_k_before_softmax"] = (
+                self.cfg.kd_top_k_before_softmax
+            )
+
+        # Image configurations
+        if self.cfg.image_size:
+            training_args_kwargs["image_size"] = self.cfg.image_size
+        if self.cfg.image_resize_algorithm:
+            training_args_kwargs["image_resize_algorithm"] = (
+                self.cfg.image_resize_algorithm
+            )
+
+        # Accelerator configuration
+        if self.cfg.accelerator_config:
+            training_args_kwargs["accelerator_config"] = self.cfg.accelerator_config
+
+    def _configure_optimizer(self, training_args_kwargs):
+        """Configure optimizer settings."""
+        custom_supported_optimizers = [opt.value for opt in CustomSupportedOptimizers]
+
+        if self.cfg.optimizer in custom_supported_optimizers:
+            # Use custom optimizer implementation
+            self._configure_custom_optimizer(training_args_kwargs)
+        else:
+            # Use transformers' optimizer
+            training_args_kwargs["optim"] = self.cfg.optimizer
+            self._add_optimizer_args(training_args_kwargs)
+
+        # Handle optimizer targeting specific modules
+        if self.cfg.optim_target_modules:
+            training_args_kwargs["optim_target_modules"] = self.cfg.optim_target_modules
+
+        # Special case for anyprecision optimizer
+        if self.cfg.optimizer == "adamw_anyprecision":
+            if Path(self.cfg.torchdistx_path).exists():
+                sys.path.append(self.cfg.torchdistx_path)
+                importlib.import_module("torchdistx")
+
+    def _configure_custom_optimizer(self, training_args_kwargs):
+        """Configure custom optimizer settings."""
+        # Common optimizer kwargs
+        optimizer_kwargs = {
+            "lr": training_args_kwargs.get("learning_rate"),
+            "weight_decay": training_args_kwargs.get("weight_decay"),
+        }
+
+        # Add Adam-specific kwargs if available
+        adam_kwargs = self._get_adam_kwargs(training_args_kwargs)
+
+        # Get optimizer class and update kwargs based on optimizer type
+        optimizer_cls = self._get_optimizer_class(
+            training_args_kwargs, optimizer_kwargs, adam_kwargs
+        )
+
+        # Add any additional optimizer args from config
+        self._update_optimizer_kwargs_from_config(optimizer_kwargs)
+
+        training_args_kwargs["optimizer_cls_and_kwargs"] = (
+            optimizer_cls,
+            optimizer_kwargs,
+        )
+
+    def _get_adam_kwargs(self, training_args_kwargs):
+        """Get Adam-specific kwargs if available."""
+        adam_kwargs = {}
+        if training_args_kwargs.get("adam_beta1") and training_args_kwargs.get(
+            "adam_beta2"
+        ):
+            adam_kwargs["betas"] = (
+                training_args_kwargs.get("adam_beta1"),
+                training_args_kwargs.get("adam_beta2"),
+            )
+        if training_args_kwargs.get("adam_epsilon"):
+            adam_kwargs["eps"] = training_args_kwargs.get("adam_epsilon")
+        return adam_kwargs
+
+    def _get_optimizer_class(self, training_args_kwargs, optimizer_kwargs, adam_kwargs):
+        """Get optimizer class based on configuration."""
+        if self.cfg.optimizer == "muon":
+            from axolotl.contribs.mit.muon import MuonOptimizerFactory   # pylint: disable=no-name-in-module
+
+            optimizer_cls = MuonOptimizerFactory
+            optimizer_kwargs.update(adam_kwargs)
+        elif self.cfg.optimizer == "optimi_adamw":
+            from optimi import AdamW
+
+            optimizer_kwargs["foreach"] = False
+            optimizer_cls = AdamW
+            optimizer_kwargs.update(adam_kwargs)
+        elif self.cfg.optimizer == "ao_adamw_4bit":
+            from torchao.prototype.low_bit_optim import AdamW4bit
+
+            optimizer_cls = AdamW4bit
+            optimizer_kwargs.update(adam_kwargs)
+            LOG.warning(
+                f"`ao_adamw_4bit` will be deprecated soon. Please use `{OptimizerNames.ADAMW_TORCH_4BIT}` instead."
+            )
+        elif self.cfg.optimizer == "ao_adamw_8bit":
+            from torchao.prototype.low_bit_optim import AdamW8bit
+
+            optimizer_cls = AdamW8bit
+            optimizer_kwargs.update(adam_kwargs)
+        elif self.cfg.optimizer == "ao_adamw_fp8":
+            from torchao.prototype.low_bit_optim import AdamWFp8
+
+            optimizer_cls = AdamWFp8
+            optimizer_kwargs.update(adam_kwargs)
+        elif self.cfg.optimizer == "adopt_adamw":
+            from axolotl.utils.optimizers.adopt import ADOPT
+
+            optimizer_cls = ADOPT
+            adam_kwargs["decouple"] = True
+            optimizer_kwargs.update(adam_kwargs)
+        elif self.cfg.optimizer == "came_pytorch":
+            from came_pytorch import CAME
+
+            optimizer_cls = CAME
+
+            beta1 = training_args_kwargs.get("adam_beta1", 0.9)
+            beta2 = training_args_kwargs.get("adam_beta2", 0.999)
+            beta3 = training_args_kwargs.get("adam_beta2", 0.9999)
+            eps1 = training_args_kwargs.get("adam_epsilon", 1e-30)
+            eps2 = training_args_kwargs.get("adam_epsilon2", 1e-16)
+
+            adam_kwargs["betas"] = (beta1, beta2, beta3)
+            adam_kwargs["eps"] = (eps1, eps2)
+            optimizer_kwargs.update(adam_kwargs)
+        else:
+            # Default case or unsupported optimizer
+            optimizer_cls = None
+
+        return optimizer_cls
+
+    def _update_optimizer_kwargs_from_config(self, optimizer_kwargs):
+        """Update optimizer kwargs from config."""
+        if self.cfg.optim_args:
+            if isinstance(self.cfg.optim_args, dict):
+                optimizer_kwargs.update(self.cfg.optim_args)
+            else:
+                # Parse string format "key1=value1,key2=value2"
+                for mapping in self.cfg.optim_args.replace(" ", "").split(","):
+                    key, value = mapping.split("=")
+                    optimizer_kwargs[key] = value
+
+    def _add_optimizer_args(self, training_args_kwargs):
+        """Add optimizer arguments if available."""
+        if self.cfg.optim_args:
+            if isinstance(self.cfg.optim_args, dict):
+                optim_args = ",".join(
+                    [f"{key}={value}" for key, value in self.cfg.optim_args.items()]
+                )
+            else:
+                optim_args = self.cfg.optim_args
+            training_args_kwargs["optim_args"] = optim_args
+
+    def _get_training_args_cls(self):
+        """Get the appropriate training arguments class."""
+        if self.cfg.reward_model:
+            return AxolotlRewardConfig
+        if self.cfg.process_reward_model:
+            return AxolotlPRMConfig
+        return AxolotlTrainingArguments
+
+    def _prepare_data_collator_kwargs(self):
+        """Prepare data collator kwargs."""
+        data_collator_kwargs = {"padding": True}  # True/"longest" is the default
+
+        if self.cfg.pad_to_sequence_len:
+            data_collator_kwargs["pad_to_multiple_of"] = 64 * math.ceil(
+                self.cfg.sequence_len / 64
+            )
+        else:
+            data_collator_kwargs["pad_to_multiple_of"] = 64
+
+        if self.cfg.reward_model:
+            data_collator_kwargs["max_length"] = self.cfg.sequence_len
+
+        return data_collator_kwargs
+
+    def _prepare_trainer_kwargs(self, trainer_cls, data_collator_kwargs, training_args):
+        """Prepare trainer kwargs."""
+        trainer_kwargs = {}
+
+        # Handle special data collators for evaluation
+        if eval_data_collator := self.build_collator(
+            training_args, is_eval=True, **data_collator_kwargs
+        ):
+            if not (self.cfg.reward_model or self.cfg.process_reward_model):
+                trainer_kwargs["eval_data_collator"] = eval_data_collator
+
+        # Add bench data collator if needed
+        if not (self.cfg.reward_model or self.cfg.process_reward_model):
+            trainer_kwargs["bench_data_collator"] = transformers.DataCollatorForSeq2Seq(
+                self.tokenizer,
+                return_tensors="pt",
+                **data_collator_kwargs,
+            )
+
+        # Add tokenizer or processing class
+        sig = inspect.signature(trainer_cls)
+        if "processing_class" in sig.parameters.keys():
+            trainer_kwargs["processing_class"] = self.tokenizer
+        else:
+            trainer_kwargs["tokenizer"] = self.tokenizer
+
+        # Add dataset tags if available
+        if (
+            not (trainer_cls in [AxolotlRewardTrainer, AxolotlPRMTrainer])
+            and self.cfg.datasets is not None
+        ):
+            trainer_kwargs["dataset_tags"] = [
+                d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
+            ]
+
+        return trainer_kwargs
+
+    def build_collator(
+        self, training_args: AxolotlTrainingArguments, is_eval=False, **kwargs
+    ):
+        if training_args.pretraining:
+            if (
+                self.cfg.pretraining_sample_concatenation is False
+                or self.cfg.micro_batch_size > 1
+            ):
+                return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
+            return None
+
+        if self.cfg.model_config_type == "mamba":
+            return MambaDataCollator(tokenizer=self.tokenizer)
+
+        use_batch_sampler_collator = False
+        if is_eval is False and training_args.sample_packing:
+            use_batch_sampler_collator = True
+        if is_eval and training_args.eval_sample_packing:
+            use_batch_sampler_collator = True
+
+        collator: Type[
+            V2BatchSamplerDataCollatorForSeq2Seq
+            | BatchSamplerDataCollatorForSeq2Seq
+            | DataCollatorForSeq2Seq
+            | DataCollatorWithFlattening
+            | RewardDataCollatorWithPadding
+        ]
+        collator_args = [self.tokenizer]
+        if self.cfg.reward_model:
+            collator = RewardDataCollatorWithPadding
+            if "max_length" in kwargs:
+                kwargs.pop("max_length")
+        elif use_batch_sampler_collator:
+            if self.cfg.flex_attention:
+                collator = V2BatchSamplerDataCollatorForSeq2Seq
+            elif 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
+            ):
+                collator = V2BatchSamplerDataCollatorForSeq2Seq
+            else:
+                collator = BatchSamplerDataCollatorForSeq2Seq
+        else:
+            if self.cfg.processor_type and self.processor:
+                collator = MultiModalChatDataCollator
+                kwargs["processing_strategy"] = get_processing_strategy(
+                    self.processor,
+                    training_args.chat_template,
+                    self.cfg.chat_template,
+                    image_size=training_args.image_size,
+                    image_resize_algorithm=training_args.image_resize_algorithm,
+                )
+            elif self.cfg.batch_flattening:
+                collator = DataCollatorWithFlattening
+                collator_args.pop(0)
+                kwargs.pop("pad_to_multiple_of", None)
+                kwargs.pop("padding", None)
+            elif self.cfg.kd_trainer:
+                from axolotl.integrations.kd.collator import (
+                    DataCollatorForKD,
+                    KDBatchSamplerDataCollatorForSeq2Seq,
+                )
+
+                if self.cfg.sample_packing:
+                    collator = KDBatchSamplerDataCollatorForSeq2Seq
+                else:
+                    collator = DataCollatorForKD
+            else:
+                collator = DataCollatorForSeq2Seq
+
+        kwargs["return_tensors"] = "pt"
+
+        return collator(
+            *collator_args,
+            **kwargs,
+        )

src/axolotl/core/trainers/builders/rl.py

@@ -0,0 +1,367 @@
+"""RL trainer / training args builder implementation"""
+
+import inspect
+from pathlib import Path
+
+from axolotl.core.trainers.builders.base import TrainerBuilderBase
+from axolotl.core.trainers.dpo import DPOStrategy
+from axolotl.core.trainers.dpo.args import AxolotlDPOConfig
+from axolotl.core.trainers.grpo import GRPOStrategy
+from axolotl.core.trainers.trl import (
+    AxolotlCPOTrainer,
+    AxolotlKTOTrainer,
+    AxolotlORPOTrainer,
+)
+from axolotl.core.training_args import (
+    AxolotlCPOConfig,
+    AxolotlKTOConfig,
+    AxolotlORPOConfig,
+)
+from axolotl.utils.models import ensure_dtype
+
+
+class HFRLTrainerBuilder(TrainerBuilderBase):
+    """Trainer factory class for TRL-based RLHF trainers (e.g. DPO)"""
+
+    def get_callbacks(self):
+        callbacks = super().get_callbacks()
+
+        return callbacks
+
+    def get_post_trainer_create_callbacks(self, trainer):
+        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
+        return callbacks
+
+    def build_training_arguments(self, total_num_steps):
+        training_args_kwargs = {}
+        for arg in [
+            "adam_beta1",
+            "adam_beta2",
+            "adam_epsilon",
+            "dataloader_num_workers",
+            "dataloader_pin_memory",
+        ]:
+            if hasattr(self.cfg, arg) and getattr(self.cfg, arg) is not None:
+                training_args_kwargs[arg] = getattr(self.cfg, arg)
+
+        if self.cfg.hub_model_id:
+            training_args_kwargs["hub_model_id"] = self.cfg.hub_model_id
+            training_args_kwargs["push_to_hub"] = True
+            training_args_kwargs["hub_private_repo"] = True
+            training_args_kwargs["hub_always_push"] = True
+
+            if self.cfg.hub_strategy:
+                training_args_kwargs["hub_strategy"] = self.cfg.hub_strategy
+
+        if self.cfg.save_safetensors is not None:
+            training_args_kwargs["save_safetensors"] = self.cfg.save_safetensors
+
+        if self.eval_dataset:
+            training_args_kwargs["eval_strategy"] = "steps"
+            training_args_kwargs["eval_steps"] = self.cfg.eval_steps
+        else:
+            training_args_kwargs["eval_strategy"] = "no"
+
+        if self.cfg.bf16 or self.cfg.bfloat16:
+            training_args_kwargs["bf16"] = True
+
+        training_args_kwargs["loraplus_lr_ratio"] = self.cfg.loraplus_lr_ratio
+        training_args_kwargs["loraplus_lr_embedding"] = self.cfg.loraplus_lr_embedding
+        training_args_kwargs["lr_scheduler_type"] = (
+            self.cfg.lr_scheduler if self.cfg.lr_scheduler else "cosine"
+        )
+        training_args_kwargs["lr_scheduler_kwargs"] = (
+            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
+            )
+        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
+            )
+        if self.cfg.dataloader_num_workers is not None:
+            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
+            )
+        if 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
+                )
+            else:
+                training_args_kwargs["gradient_checkpointing_kwargs"] = {
+                    "use_reentrant": False
+                }
+
+        # set save_strategy and save_steps
+        if self.cfg.save_steps:
+            training_args_kwargs["save_strategy"] = "steps"
+            training_args_kwargs["save_steps"] = self.cfg.save_steps
+        elif self.cfg.save_strategy:
+            training_args_kwargs["save_strategy"] = self.cfg.save_strategy
+        else:
+            # default to saving each epoch if not defined
+            training_args_kwargs["save_strategy"] = "epoch"
+
+        training_args_kwargs["save_only_model"] = self.cfg.save_only_model
+
+        if self.cfg.dataset_processes:
+            training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
+
+        if self.cfg.trl and self.cfg.trl.beta is not None:
+            training_args_kwargs["beta"] = self.cfg.trl.beta
+        elif self.cfg.rl_beta is not None:
+            training_args_kwargs["beta"] = self.cfg.rl_beta
+        elif self.cfg.orpo_alpha is not None:
+            # trl does some odd mapping of alpha to beta to reuse the beta parameter ???
+            training_args_kwargs["beta"] = self.cfg.orpo_alpha
+
+        if self.cfg.rpo_alpha is not None:
+            training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
+
+        if self.cfg.use_wandb:
+            training_args_kwargs["run_name"] = self.cfg.wandb_name
+
+        training_args_cls = None
+        blocklist_args_kwargs = []
+        if self.cfg.rl == "simpo":
+            training_args_cls = AxolotlCPOConfig
+            training_args_kwargs["loss_type"] = "simpo"
+            training_args_kwargs["max_length"] = self.cfg.sequence_len
+            training_args_kwargs["simpo_gamma"] = self.cfg.simpo_gamma
+            if self.cfg.cpo_alpha is not None:
+                training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
+
+        elif self.cfg.rl == "orpo":
+            training_args_cls = AxolotlORPOConfig
+            training_args_kwargs["max_length"] = self.cfg.sequence_len
+            if self.cfg.max_prompt_len:
+                training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
+
+        elif self.cfg.rl == "kto":
+            training_args_cls = AxolotlKTOConfig
+
+            training_args_kwargs["desirable_weight"] = (
+                self.cfg.kto_desirable_weight or 1.0
+            )
+            training_args_kwargs["undesirable_weight"] = (
+                self.cfg.kto_undesirable_weight or 1.0
+            )
+
+            training_args_kwargs["max_length"] = self.cfg.sequence_len
+            if self.cfg.max_prompt_len:
+                training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
+
+        elif self.cfg.rl == "grpo":
+            training_args_cls = GRPOStrategy.get_training_args_class()
+            training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
+            blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
+
+        else:
+            training_args_cls = AxolotlDPOConfig
+            if self.cfg.rl == "ipo":
+                training_args_kwargs["loss_type"] = "ipo"
+            training_args_kwargs["max_length"] = self.cfg.sequence_len
+            training_args_kwargs["max_completion_length"] = None
+            training_args_kwargs["max_prompt_length"] = self.cfg.sequence_len
+            training_args_kwargs["generate_during_eval"] = self.cfg.use_wandb
+            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
+                )
+
+        for blocklist_key in blocklist_args_kwargs:
+            if blocklist_key in training_args_kwargs:
+                del training_args_kwargs[blocklist_key]
+
+        max_steps = self.cfg.max_steps or total_num_steps or -1
+        training_args_kwargs["num_train_epochs"] = self.cfg.num_epochs
+        training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
+            self.cfg.output_dir,
+            per_device_train_batch_size=self.cfg.micro_batch_size,
+            max_steps=max_steps,
+            gradient_accumulation_steps=self.cfg.gradient_accumulation_steps,
+            learning_rate=self.cfg.learning_rate,
+            warmup_steps=self.cfg.warmup_steps,
+            logging_first_step=True,
+            logging_steps=1,
+            optim=self.cfg.optimizer,
+            save_total_limit=self.cfg.save_total_limit or 5,
+            **training_args_kwargs,
+        )
+
+        # unset run_name so wandb sets up experiment names
+        if self.cfg.use_wandb and training_args.run_name == training_args.output_dir:
+            training_args.run_name = (  # pylint: disable=attribute-defined-outside-init
+                None
+            )
+
+        return training_args
+
+    def build(self, total_num_steps):
+        """Build and return an RL trainer instance"""
+        # Prepare RL-specific training args kwargs
+        training_args_kwargs = {
+            "per_device_train_batch_size": self.cfg.micro_batch_size,
+            "max_steps": self.cfg.max_steps or total_num_steps or -1,
+            "gradient_accumulation_steps": self.cfg.gradient_accumulation_steps,
+            "learning_rate": self.cfg.learning_rate,
+            "warmup_steps": self.cfg.warmup_steps,
+            "logging_first_step": True,
+            "logging_steps": 1,
+            "output_dir": self.cfg.output_dir,
+            "num_train_epochs": self.cfg.num_epochs,
+        }
+
+        # Handle dataset processes
+        if self.cfg.dataset_processes:
+            training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
+
+        # Handle beta/alpha parameters for different RL algorithms
+        if self.cfg.trl and self.cfg.trl.beta is not None:
+            training_args_kwargs["beta"] = self.cfg.trl.beta
+        elif self.cfg.rl_beta is not None:
+            training_args_kwargs["beta"] = self.cfg.rl_beta
+        elif self.cfg.orpo_alpha is not None:
+            # trl does some odd mapping of alpha to beta to reuse the beta parameter
+            training_args_kwargs["beta"] = self.cfg.orpo_alpha
+
+        if self.cfg.rpo_alpha is not None:
+            training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
+
+        # Determine training args class and add RL-specific parameters
+        training_args_cls = None
+        blocklist_args_kwargs = []
+
+        if self.cfg.rl == "simpo":
+            training_args_cls = AxolotlCPOConfig
+            training_args_kwargs["loss_type"] = "simpo"
+            training_args_kwargs["simpo_gamma"] = self.cfg.simpo_gamma
+            if self.cfg.cpo_alpha is not None:
+                training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
+        elif self.cfg.rl == "orpo":
+            training_args_cls = AxolotlORPOConfig
+            if self.cfg.max_prompt_len:
+                training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
+        elif self.cfg.rl == "kto":
+            training_args_cls = AxolotlKTOConfig
+            training_args_kwargs["desirable_weight"] = (
+                self.cfg.kto_desirable_weight or 1.0
+            )
+            training_args_kwargs["undesirable_weight"] = (
+                self.cfg.kto_undesirable_weight or 1.0
+            )
+            if self.cfg.max_prompt_len:
+                training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
+        elif self.cfg.rl == "grpo":
+            training_args_cls = GRPOStrategy.get_training_args_class()
+            training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
+            blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
+        else:  # Default to DPO
+            training_args_cls = AxolotlDPOConfig
+            if self.cfg.rl == "ipo":
+                training_args_kwargs["loss_type"] = "ipo"
+            training_args_kwargs["max_prompt_length"] = self.cfg.sequence_len
+            training_args_kwargs["max_completion_length"] = None
+            training_args_kwargs["generate_during_eval"] = self.cfg.use_wandb
+            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
+                )
+
+        # Remove any blocklisted arguments
+        for blocklist_key in blocklist_args_kwargs:
+            if blocklist_key in training_args_kwargs:
+                del training_args_kwargs[blocklist_key]
+
+        # Create training args using the base class method
+        training_args = self.create_training_args(
+            args_cls=training_args_cls,
+            total_num_steps=total_num_steps,
+            **training_args_kwargs,
+        )
+
+        # Prepare trainer kwargs
+        trainer_kwargs = {}
+        if self.cfg.rl == "ipo" and self.cfg.dpo_label_smoothing:
+            trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
+        if self.eval_dataset:
+            trainer_kwargs["eval_dataset"] = self.eval_dataset
+        if self.cfg.adapter and self.peft_config:
+            trainer_kwargs["peft_config"] = self.peft_config
+        if self.cfg.precompute_ref_log_probs is not None:
+            trainer_kwargs["precompute_ref_log_probs"] = (
+                self.cfg.precompute_ref_log_probs
+            )
+
+        # Determine trainer class and arguments
+        if self.cfg.rl == "grpo":
+            trainer_cls = GRPOStrategy.get_trainer_class()
+            trainer_args = [self.model]
+            trainer_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
+            trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
+        elif self.cfg.rl in ["dpo", "ipo"]:
+            trainer_cls = DPOStrategy.get_trainer_class()
+            trainer_args = [self.model, self.model_ref]
+        elif self.cfg.rl == "orpo":
+            trainer_cls = AxolotlORPOTrainer
+            trainer_args = [self.model]
+        elif self.cfg.rl in ["kto"]:
+            trainer_cls = AxolotlKTOTrainer
+            trainer_args = [self.model]
+        elif self.cfg.rl in ["simpo"]:
+            trainer_cls = AxolotlCPOTrainer
+            trainer_args = [self.model]
+        else:
+            raise ValueError(f"Unsupported RL: {self.cfg.rl}")
+
+        # Add tokenizer or processing class
+        sig = inspect.signature(trainer_cls)
+        if "tokenizer" in sig.parameters.keys():
+            trainer_kwargs["tokenizer"] = self.tokenizer
+        else:
+            trainer_kwargs["processing_class"] = self.tokenizer
+
+        # Add dataset tags if available
+        if self.cfg.datasets is not None and (
+            trainer_cls is DPOStrategy.get_trainer_class()
+        ):
+            trainer_kwargs["dataset_tags"] = [
+                d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
+            ]
+
+        # Create the trainer
+        trainer = self.create_trainer(
+            trainer_cls=trainer_cls,
+            training_args=training_args,
+            trainer_args=trainer_args,
+            trainer_kwargs=trainer_kwargs,
+        )
+
+        # Handle FSDP specific settings
+        if self.cfg.fsdp:
+            ensure_dtype(trainer.model, dtype=self.cfg.torch_dtype)
+            if (
+                self.cfg.rl in ["dpo", "ipo"]
+                and hasattr(trainer, "ref_model")
+                and trainer.ref_model
+            ):
+                ensure_dtype(trainer.ref_model, dtype=self.cfg.torch_dtype)
+
+        return trainer

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

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

src/axolotl/integrations/base.py

@@ -26,6 +26,8 @@ from typing import OrderedDict
 import torch
 from torch.optim.lr_scheduler import LRScheduler
 
+from axolotl.utils.dict import DictDefault
+
 
 class BasePlugin:
     """
@@ -36,11 +38,13 @@ class BasePlugin:
 
     Methods:
     register(cfg): Registers the plugin with the given configuration.
+    load_datasets(cfg): Loads and preprocesses the dataset for training.
     pre_model_load(cfg): Performs actions before the model is loaded.
     post_model_build(cfg, model): Performs actions after the model is loaded, but before LoRA adapters are applied.
     pre_lora_load(cfg, model): Performs actions before LoRA weights are loaded.
     post_lora_load(cfg, model): Performs actions after LoRA weights are loaded.
     post_model_load(cfg, model): Performs actions after the model is loaded, inclusive of any adapters.
+    post_trainer_create(cfg, trainer): Performs actions after the trainer is created.
     create_optimizer(cfg, trainer): Creates and returns an optimizer for training.
     create_lr_scheduler(cfg, trainer, optimizer, num_training_steps): Creates and returns a learning rate scheduler.
     add_callbacks_pre_trainer(cfg, model): Adds callbacks to the trainer before training.
@@ -63,20 +67,32 @@ class BasePlugin:
         None
         """
 
-    def get_input_args(self):
+    def get_input_args(self) -> str | None:
         """
         Returns a pydantic model for the plugin's input arguments.
         """
 
+    def load_datasets(self, cfg: DictDefault, preprocess: bool = False):
+        """
+        Loads and preprocesses the dataset for training.
+
+        Args:
+            cfg: The configuration for the plugin.
+            preprocess: Whether this is the preprocess step of the datasets.
+
+        Returns:
+            dataset_meta: The metadata for the training dataset.
+        """
+
     def pre_model_load(self, cfg):  # pylint: disable=unused-argument
         """
         Performs actions before the model is loaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
+        Args:
+            cfg (dict): The configuration for the plugin.
 
         Returns:
-        None
+            None
         """
 
     def post_model_build(self, cfg, model):  # pylint: disable=unused-argument
@@ -91,59 +107,71 @@ class BasePlugin:
         """
         Performs actions after the model is loaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg (dict): The configuration for the plugin.
+            model (object): The loaded model.
 
         Returns:
-        None
+            None
         """
 
     def pre_lora_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions before LoRA weights are loaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg (dict): The configuration for the plugin.
+            model (object): The loaded model.
 
         Returns:
-        None
+            None
         """
 
     def post_lora_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions after LoRA weights are loaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg (dict): The configuration for the plugin.
+            model (object): The loaded model.
 
         Returns:
-        None
+            None
         """
 
     def get_trainer_cls(self, cfg):  # pylint: disable=unused-argument):
         """
         Returns a custom class for the trainer.
 
-        Parameters:
-        cfg (dict): The global axolotl configuration.
+        Args:
+            cfg (dict): The global axolotl configuration.
 
         Returns:
-        class: The class for the trainer.
+            class: The class for the trainer.
+        """
+
+    def post_trainer_create(self, cfg, trainer):  # pylint: disable=unused-argument
+        """
+        Performs actions after the trainer is created.
+
+        Args:
+            cfg (dict): The configuration for the plugin.
+            trainer (object): The trainer object for training.
+
+        Returns:
+            None
         """
 
     def create_optimizer(self, cfg, trainer):  # pylint: disable=unused-argument
         """
         Creates and returns an optimizer for training.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        trainer (object): The trainer object for training.
+        Args:
+            cfg (dict): The configuration for the plugin.
+            trainer (object): The trainer object for training.
 
         Returns:
-        object: The created optimizer.
+            object: The created optimizer.
         """
 
     def create_lr_scheduler(
@@ -152,26 +180,26 @@ class BasePlugin:
         """
         Creates and returns a learning rate scheduler.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        trainer (object): The trainer object for training.
-        optimizer (object): The optimizer for training.
-        num_training_steps (int): Total number of training steps
+        Args:
+            cfg (dict): The configuration for the plugin.
+            trainer (object): The trainer object for training.
+            optimizer (object): The optimizer for training.
+            num_training_steps (int): Total number of training steps
 
         Returns:
-        object (LRScheduler): The created learning rate scheduler.
+            object (LRScheduler): The created learning rate scheduler.
         """
 
     def add_callbacks_pre_trainer(self, cfg, model):  # pylint: disable=unused-argument
         """
         setup callbacks before creating the trainer.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg (dict): The configuration for the plugin.
+            model (object): The loaded model.
 
         Returns:
-        List[callable]: A list of callback functions to be added to the TrainingArgs
+            List[callable]: A list of callback functions to be added to the TrainingArgs
         """
         return []
 
@@ -182,12 +210,12 @@ class BasePlugin:
         Adds callbacks to the trainer after creating the trainer.
         This is useful for callbacks that require access to the model or trainer.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        trainer (object): The trainer object for training.
+        Args:
+            cfg (dict): The configuration for the plugin.
+            trainer (object): The trainer object for training.
 
         Returns:
-        List[callable]: A list of callback functions to be added
+            List[callable]: A list of callback functions to be added
         """
         return []
 
@@ -195,23 +223,23 @@ class BasePlugin:
         """
         Performs actions after training is complete.
 
-        Parameters:
-        cfg (dict): The axolotl configuration
-        model (object): The loaded model.
+        Args:
+            cfg (dict): The axolotl configuration
+            model (object): The loaded model.
 
         Returns:
-        None
+            None
         """
 
     def post_train_unload(self, cfg):  # pylint: disable=unused-argument
         """
         Performs actions after training is complete and the model is unloaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
+        Args:
+            cfg (dict): The configuration for the plugin.
 
         Returns:
-        None
+            None
         """
 
 
@@ -338,6 +366,27 @@ class PluginManager:
                 input_args.append(input_args_from_plugin)
         return input_args
 
+    def load_datasets(self, cfg, preprocess: bool = False):
+        """
+        Calls the load_datasets method of each registered plugin.
+
+        Args:
+            cfg: The configuration for the plugins.
+            preprocess : Whether this is preprocess step of the datasets.
+
+        Returns:
+            dataset_meta: The dataset metadata loaded from all registered plugins.
+        """
+        return_ds_meta = None
+        for plugin in self.plugins.values():
+            dataset_meta = plugin.load_datasets(cfg, preprocess)
+            if dataset_meta is not None:
+                if return_ds_meta is None:
+                    return_ds_meta = dataset_meta
+                else:
+                    raise RuntimeError("Multiple plugins loaded datasets")
+        return return_ds_meta
+
     def pre_model_load(self, cfg):
         """
         Calls the pre_model_load method of all registered plugins.
@@ -422,6 +471,20 @@ class PluginManager:
                 return trainer_cls
         return None
 
+    def post_trainer_create(self, cfg, trainer):
+        """
+        Calls the post_trainer_create method of all registered plugins.
+
+        Parameters:
+        cfg (dict): The configuration for the plugins.
+        trainer (object): The trainer object for training.
+
+        Returns:
+        None
+        """
+        for plugin in self.plugins.values():
+            plugin.post_trainer_create(cfg, trainer)
+
     def create_optimizer(self, trainer):
         """
         Calls the create_optimizer method of all registered plugins and returns the first non-None optimizer.

src/axolotl/monkeypatch/attention/__init__.py

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

src/axolotl/monkeypatch/attention/xformers.py

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

src/axolotl/monkeypatch/peft/utils.py

@@ -0,0 +1,78 @@
+"""
+Patch prepare_model_for_kbit_training to not upcast everything
+"""
+
+import inspect
+import logging
+
+import peft
+
+import axolotl
+from axolotl.monkeypatch.utils import detab_code
+
+LOG = logging.getLogger(__name__)
+
+ORIGINAL_PREPARE_CODE = """
+        for param in model.parameters():
+            if (
+                (param.dtype == torch.float16) or (param.dtype == torch.bfloat16)
+            ) and param.__class__.__name__ != "Params4bit":
+                param.data = param.data.to(torch.float32)
+"""
+
+PATCHED_PREPARE_CODE = """
+        for name, param in model.named_parameters():
+            if (
+                (param.dtype == torch.float16) or (param.dtype == torch.bfloat16)
+            ) and param.__class__.__name__ != "Params4bit" and all(embed_name not in name for embed_name in ["embed_tokens", "lm_head"]):
+                param.data = param.data.to(torch.float32)
+"""
+
+
+def get_peft_prep_code() -> str:
+    prepare = inspect.getsource(peft.utils.other.prepare_model_for_kbit_training)
+    return prepare
+
+
+def check_peft_prep_code_is_patchable() -> bool:
+    prep_code = get_peft_prep_code()
+    prep_code, _ = detab_code(prep_code)
+    return ORIGINAL_PREPARE_CODE in prep_code
+
+
+def patch_peft_prep_code():
+    """
+    monkeypatch create_accelerator_and_postprocess so it checks for additional kwargs
+    """
+
+    try:
+        prep_code = get_peft_prep_code()
+    except OSError:
+        return
+    peft.utils.other._original_create_accelerator_and_postprocess = (  # pylint: disable=protected-access
+        prep_code
+    )
+    prep_code, _ = detab_code(prep_code)
+    if ORIGINAL_PREPARE_CODE not in prep_code:
+        return
+
+    prep_code = prep_code.replace(ORIGINAL_PREPARE_CODE, PATCHED_PREPARE_CODE)
+    prep_code = prep_code.replace(
+        "def prepare_model_for_kbit_training(",
+        "def fixed_prepare_model_for_kbit_training(",
+        1,
+    )
+
+    items_to_import = []
+    for item in dir(peft.utils.other):
+        if item in prep_code:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from peft.utils.other import (" + ", ".join(x for x in items_to_import) + ")",
+        globals(),
+    )
+    exec(prep_code, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching prepare_model_for_kbit_training to allow for overrides")
+    peft.utils.other.prepare_model_for_kbit_training = fixed_prepare_model_for_kbit_training  # pylint: disable=protected-access  # pylint: disable=undefined-variable  # noqa: F821
+    axolotl.utils.models.prepare_model_for_kbit_training = fixed_prepare_model_for_kbit_training  # pylint: disable=protected-access  # pylint: disable=undefined-variable  # noqa: F821

src/axolotl/train.py

@@ -2,6 +2,7 @@
 
 import importlib
 import inspect
+import logging
 import os
 import signal
 import sys
@@ -12,7 +13,6 @@ from typing import Any, Dict
 
 import torch
 import transformers.modelcard
-from accelerate.logging import get_logger
 from accelerate.utils import save_fsdp_model
 from datasets import Dataset
 from huggingface_hub.errors import OfflineModeIsEnabled
@@ -21,11 +21,12 @@ from transformers import PreTrainedModel, PreTrainedTokenizer, ProcessorMixin
 from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
 from transformers.trainer import Trainer
 
+from axolotl.cli.art import print_axolotl_text_art
 from axolotl.common.datasets import TrainDatasetMeta
 from axolotl.contribs.lgpl import (  # pylint: disable = no-name-in-module
     fix_untrained_tokens,
 )
-from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
+from axolotl.core.trainers.builders import HFCausalTrainerBuilder, HFRLTrainerBuilder
 from axolotl.core.trainers.mixins.sequence_parallel import (
     SequenceParallelContextManager,
 )
@@ -41,7 +42,7 @@ try:
 except ImportError:
     BetterTransformer = None
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def setup_model_and_tokenizer(
@@ -62,7 +63,6 @@ def setup_model_and_tokenizer(
     # Load tokenizer
     LOG.debug(
         f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}",
-        main_process_only=True,
     )
     tokenizer = load_tokenizer(cfg)
 
@@ -516,6 +516,8 @@ def train(
     Returns:
         Tuple of (model, tokenizer) after training
     """
+    print_axolotl_text_art()
+
     # Setup model, tokenizer, (causal or RLHF) trainer, etc.
     (
         trainer,
@@ -525,6 +527,9 @@ def train(
         processor,
     ) = setup_model_and_trainer(cfg, dataset_meta)
 
+    plugin_manager = PluginManager.get_instance()
+    plugin_manager.post_trainer_create(cfg, trainer)
+
     # Handle untrained tokens if configured
     safe_serialization = cfg.save_safetensors is True
     train_dataset = dataset_meta.train_dataset
@@ -547,7 +552,6 @@ def train(
     if not cfg.use_ray:
         cleanup_distributed()
 
-    plugin_manager = PluginManager.get_instance()
     plugin_manager.post_train(cfg, model)
 
     return model, tokenizer, trainer

src/axolotl/utils/callbacks/__init__.py

@@ -46,11 +46,11 @@ from axolotl.utils.distributed import (
 from axolotl.utils.schemas.config import AxolotlInputConfig
 
 if TYPE_CHECKING:
-    from axolotl.core.trainer_builder import AxolotlTrainingArguments
+    from axolotl.core.training_args import AxolotlTrainingArguments
 
 
 IGNORE_INDEX = -100
-LOG = logging.getLogger("axolotl.callbacks")
+LOG = logging.getLogger(__name__)
 
 
 class EvalFirstStepCallback(
@@ -868,3 +868,28 @@ class GCCallback(TrainerCallback):
     ):
         torch.cuda.empty_cache()
         gc.collect()
+
+
+def colab_inference_post_train_callback(trainer: Trainer):
+    class ColabCallback(TrainerCallback):
+        """Callback to prep model for inference on Google Colab"""
+
+        def __init__(self, cfg):
+            self.gpu_name = torch.cuda.get_device_name(0)
+            self.cfg = cfg
+
+        def on_train_end(
+            self, args, state, control, **kwargs
+        ):  # pylint: disable=unused-argument
+            """
+            handle T4 gpu, we need to convert attention to eager for inference
+            """
+            if "Tesla T4" in self.gpu_name and self.cfg.xformers_attention:
+                trainer.model.config._attn_implementation = (  # pylint: disable=protected-access
+                    "eager"
+                )
+            trainer.model.gradient_checkpointing_disable()
+            trainer.model.config.use_cache = True
+            trainer.model.eval()
+
+    return ColabCallback

src/axolotl/utils/config/__init__.py

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

src/axolotl/utils/data/shared.py

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

src/axolotl/utils/gradient_checkpointing/__init__.py

@@ -1,15 +1,36 @@
 """custom checkpointing utils"""
 
+import importlib
 from functools import partial
 
+from packaging import version
+
 from axolotl.utils.gradient_checkpointing.unsloth import (
     Unsloth_Offloaded_Gradient_Checkpointer,
 )
 
+transformers_version = version.parse(importlib.metadata.version("transformers"))
+if transformers_version > version.parse("4.51.3"):
+    from transformers.modeling_layers import GradientCheckpointingLayer
+
+    def uses_gc_layers(decoder_layer):
+        return isinstance(decoder_layer.func.__self__, GradientCheckpointingLayer)
+
+else:
+
+    def uses_gc_layers(_):
+        return False
+
 
 def hf_grad_checkpoint_offload_wrapper(
     decoder_layer, *args, use_reentrant=None
 ):  # pylint: disable=unused-argument
+    if uses_gc_layers(decoder_layer):
+        return Unsloth_Offloaded_Gradient_Checkpointer.apply(
+            decoder_layer,
+            *args,
+        )
+
     return Unsloth_Offloaded_Gradient_Checkpointer.apply(
         (
             decoder_layer.func.__self__

src/axolotl/utils/models.py

@@ -556,11 +556,21 @@ class ModelLoader:
         self.auto_model_loader = AutoModelForCausalLM  # pylint: disable=invalid-name
 
     def apply_patches(self) -> None:
+        if self.cfg.xformers_attention and self.cfg.sample_packing:
+            from axolotl.monkeypatch.attention import patch_xformers_attn_over_fa2
+
+            patch_xformers_attn_over_fa2()
+            self.cfg.flash_attention = True
         if self.cfg.fsdp_config and str(self.cfg.fsdp_config.fsdp_version) == "2":
             from axolotl.monkeypatch.accelerate.fsdp2 import patch_accelerate_fsdp_utils
 
             patch_accelerate_fsdp_utils()
 
+        if self.cfg.adapter and self.cfg.embeddings_skip_upcast:
+            from axolotl.monkeypatch.peft.utils import patch_peft_prep_code
+
+            patch_peft_prep_code()
+
         if self.cfg.flex_attention:
             from axolotl.monkeypatch.attention.flex_attn import (
                 patch_flex_make_mask,
@@ -1180,7 +1190,7 @@ class ModelLoader:
                 ],
             )
 
-    def prepare_model(self, qlora_fsdp) -> None:
+    def prepare_model(self, qlora_fsdp: bool) -> None:
         skip_prepare_model_for_kbit_training = False
         if self.cfg.model_config_type == "qwen" and self.cfg.adapter == "lora":
             # Qwen doesn't play nicely with LoRA if this is enabled
@@ -1310,7 +1320,10 @@ class ModelLoader:
         # make sure these are fp32 per Ramesh et al. (2021)
         embedding_modules = get_linear_embedding_layers(self.cfg.model_config_type)
         if not self.cfg.fsdp:
-            # FSDP doesn't like mixed Float and BFloat16
+            # we don't run this during FSDP because this will leave mixed
+            # float and bfloat16 dtypes in the model which FSDP doesn't like
+            if self.cfg.load_in_4bit and self.cfg.embeddings_skip_upcast:
+                embedding_modules = []
             self.convert_embedding_modules_dtype(
                 embedding_modules,
                 dist_dtype=torch.float32,

src/axolotl/utils/samplers/multipack.py

@@ -1,10 +1,13 @@
-# pylint: skip-file
 """
-Multipack Batch Sampler
+Multipack Batch Sampler - An efficient batch sampler for packing variable-length sequences
+into fixed-capacity batches to optimize memory usage and training throughput.
 """
+
 import logging
 import math
-from typing import Any, Iterable, List, Union
+from concurrent.futures import ProcessPoolExecutor
+from multiprocessing import cpu_count, get_context
+from typing import Iterable, Union
 
 import numba
 import numpy as np
@@ -13,26 +16,39 @@ from torch.utils.data import BatchSampler, Sampler, SequentialSampler
 from axolotl.utils.distributed import reduce_and_broadcast
 
 LOG = logging.getLogger(__name__)
-
 LOG.setLevel(logging.INFO)
 
 
 @numba.njit
-def ffd_check(a: np.ndarray, c: int, n: int):
-    # First-fit-decreasing bin packing
-    # Check if a[] could fit in n bins with capacity c
-    # https://en.wikipedia.org/wiki/First-fit-decreasing_bin_packing
+def ffd_check(sequence_lengths: np.ndarray, bin_capacity: int, num_bins: int):
+    """
+    First-fit-decreasing bin packing algorithm check
 
-    a = np.sort(a)[::-1]
-    bins = np.full((n,), c, dtype=a.dtype)
-    for size in a:
+    Checks if sequences with the given lengths could fit in the specified number of bins
+
+    Args:
+        sequence_lengths: Array of sequence lengths
+        bin_capacity: Maximum capacity of each bin
+        num_bins: Number of bins available
+
+    Returns:
+        True if all sequences can be packed, False otherwise
+    """
+    # Sort sequence lengths in descending order for optimal packing
+    sequence_lengths = np.sort(sequence_lengths)[::-1]
+    # Initialize all bins with full capacity
+    bins = np.full((num_bins,), bin_capacity, dtype=sequence_lengths.dtype)
+
+    # Try to place each sequence in the first bin it fits
+    for size in sequence_lengths:
         not_found = True
-        for idx in range(n):
+        for idx in range(num_bins):
             if bins[idx] >= size:
                 bins[idx] -= size
                 not_found = False
                 break
 
+        # If no bin could fit this sequence, packing failed
         if not_found:
             return False
 
@@ -40,86 +56,155 @@ def ffd_check(a: np.ndarray, c: int, n: int):
 
 
 @numba.njit
-def ffd_with_result(a: np.ndarray, c: int, start_index: int):
-    # First-fit-decreasing bin packing (with result return)
+def pack_group(
+    sequence_lengths: np.ndarray,
+    group_offset: int,
+    bin_capacity: int,
+    max_bins: int,
+    bin_size: int,
+    safe_mode: bool = True,
+):
+    """
+    Pack a group of sequences into bins using First-Fit Decreasing algorithm
 
-    indices = np.argsort(a)[::-1]
-    a = a[indices]
+    Args:
+        sequence_lengths: Array of sequence lengths
+        group_offset: Offset to apply to indices when returning results
+        bin_capacity: Maximum capacity of each bin
+        max_bins: Maximum number of bins to use
+        bin_size: Maximum number of sequences per bin
+        safe_mode: If True, use a more conservative packing approach
 
-    bins: List[Any] = []
-    bins_result: List[Any] = []
-    for a_id, size in enumerate(a):
-        add_new = True
-        for idx in range(len(bins)):
-            if bins[idx] >= size:
-                bins[idx] -= size
-                bins_result[idx].append(indices[a_id] + start_index)
-                add_new = False
+    Returns:
+        List of bins, where each bin contains indices of sequences assigned to it
+    """
+    bins_remaining_space: list = []  # Tracks remaining capacity in each bin
+    bins_assigned_sequences: list = []  # Tracks sequence indices assigned to each bin
+
+    for seq_id, size in enumerate(sequence_lengths):
+        global_idx = seq_id + group_offset
+
+        # Try to place sequence in existing bins
+        add_new_bin = True
+        for bin_idx, _ in enumerate(bins_remaining_space):
+            if (
+                bins_remaining_space[bin_idx] >= size
+                and len(bins_assigned_sequences[bin_idx]) < bin_size
+            ):
+                bins_remaining_space[bin_idx] -= size
+                bins_assigned_sequences[bin_idx].append(global_idx)
+                add_new_bin = False
                 break
 
-        if add_new:
-            bins.append(c - size)
-            bins_result.append([indices[a_id] + start_index])
+        # Create a new bin if needed and if we haven't reached the limit
+        if add_new_bin:
+            if len(bins_remaining_space) >= max_bins and safe_mode:
+                # In safe mode, skip items that would exceed max_bins
+                continue
+            bins_remaining_space.append(bin_capacity - size)
+            bins_assigned_sequences.append([global_idx])
 
-    return bins_result
+            # Safety check to avoid infinite bins
+            if len(bins_remaining_space) > len(sequence_lengths):
+                break
+
+    return bins_assigned_sequences
 
 
-@numba.njit
-def allocate(
-    lengths: np.ndarray, lengths_cumsum: np.ndarray, rank: int, c: int, n: int
+# Define a standalone function for multiprocessing
+def _process_group(args):
+    group_lengths, start_idx, bin_capacity, max_bins, bin_size, safe_mode = args
+    return pack_group(
+        group_lengths, start_idx, bin_capacity, max_bins, bin_size, safe_mode
+    )
+
+
+def pack_parallel(
+    sequence_lengths: np.ndarray,
+    bin_capacity: int,
+    group_size: int,
+    bin_size: int,
+    num_processes: int | None = None,
+    safe_mode: bool = True,
+    mp_start_method: str | None = "spawn",
 ):
-    # Dynamic batch allocator, similar to Multifit
-    # https://en.wikipedia.org/wiki/Multifit_algorithm
-    # ~99.5% efficiency on OpenChat training set (12 * 2048 ctx len)
+    """
+    Pack sequences into bins using parallel processing
 
-    s = 0
-    start_index = 0
-    result = []
+    Args:
+        sequence_lengths: Array of sequence lengths
+        bin_capacity: Maximum capacity of each bin as total number of tokens
+        group_size: Number of sequences to process in each group
+        bin_size: Maximum number of bins to use
+        num_processes: Number of parallel processes to use
+        safe_mode: If True, use a more conservative packing approach
+        mp_start_method: Multiprocessing start method ('fork', 'spawn', 'forkserver').
+                         'spawn' is often safer with Numba/PyTorch.
+                         Set to None to use system default.
+    Returns:
+        List of bins, where each bin contains indices of sequences assigned to it
+    """
+    num_items = len(sequence_lengths)
+    if num_processes is None:
+        num_processes = max(1, min(num_items // group_size, cpu_count()))
 
-    while True:
-        # binary search [l, r)
-        left = 1
-        right = 1 + np.searchsorted(lengths_cumsum[start_index:], s + c * n, "right")
+    # Create tasks for parallel processing
+    tasks = []
+    for i in range(0, num_items, group_size):
+        group_lengths = sequence_lengths[i : i + group_size]
+        max_bins = len(group_lengths)  # Allow as many bins as items in the group
+        tasks.append((group_lengths, i, bin_capacity, max_bins, bin_size, safe_mode))
 
-        while right - left > 1:
-            mid = (left + right) // 2
-            if ffd_check(lengths[start_index : start_index + mid], c, n):
-                left = mid
-            else:
-                right = mid
+    # Process groups in parallel
+    all_bins = []
 
-        # use length l
-        batch = ffd_with_result(
-            lengths[start_index : start_index + left], c, start_index
-        )
-        assert len(batch) <= n
-        if len(batch) < n:
-            break
+    mp_ctx = None
+    if mp_start_method:
+        try:
+            mp_ctx = get_context(mp_start_method)
+        except ValueError:
+            LOG.warning(
+                f"Failed to get multiprocessing context '{mp_start_method}'. "
+                f"Falling back to default. Available: {get_context().get_all_start_methods()}"
+            )
+            mp_ctx = (
+                None  # Fallback to default context if specified one is not available
+            )
 
-        start_index += left
-        s = lengths_cumsum[start_index - 1]
+    if num_processes == 1:
+        LOG.debug("Using single process for pack_parallel, running sequentially.")
+        for task_args in tasks:
+            group_bins = _process_group(task_args)
+            all_bins.extend(group_bins)
+    else:
+        # Use ProcessPoolExecutor only if num_processes > 1
+        # Pass mp_context if available
+        with ProcessPoolExecutor(
+            max_workers=num_processes, mp_context=mp_ctx
+        ) as executor:
+            for group_bins in executor.map(_process_group, tasks):
+                all_bins.extend(group_bins)
 
-        # add local rank
-        result.append(batch[rank])
-
-    return result, s, len(result) * c * n
+    return all_bins
 
 
 @numba.njit
-def allocate_sequentially(lengths: np.ndarray, rank: int, c: int, n: int):
+def allocate_sequentially(
+    sequence_lengths: np.ndarray, rank: int, bin_capacity: int, num_ranks: int
+):
     """
     Sequential allocator that preserves example order
 
-    Parameters:
-    - lengths: The lengths of all examples
-    - rank: The current rank (for distributed training)
-    - c: The capacity of each bin (maximum sequence length)
-    - n: Number of ranks
+    Args:
+        sequence_lengths: The lengths of all examples
+        rank: The current rank (for distributed training)
+        bin_capacity: The capacity of each bin (maximum sequence length)
+        num_ranks: Number of ranks (processes/GPUs)
 
     Returns:
-    - result: List of batches for the current rank
-    - total_used: Number of actual example tokens
-    - total_slots: Maximum theoretical number of example tokens (number of bins * bin capacity)
+        rank_batches: List of batches for the current rank
+        total_tokens_used: Number of actual example tokens
+        total_token_slots: Maximum theoretical number of example tokens (number of bins * bin capacity)
     """
     result = []
     total_used = 0
@@ -127,9 +212,9 @@ def allocate_sequentially(lengths: np.ndarray, rank: int, c: int, n: int):
     # First, do sequential packing into bins
     all_bins = []
     current_bin = [0 for i in range(0)]  # numba hint
-    remaining_capacity = c
+    remaining_capacity = bin_capacity
 
-    for idx, size in enumerate(lengths):
+    for idx, size in enumerate(sequence_lengths):
         if size <= remaining_capacity:
             # Example fits in current bin
             current_bin.append(idx)
@@ -140,7 +225,7 @@ def allocate_sequentially(lengths: np.ndarray, rank: int, c: int, n: int):
             if current_bin:  # Add non-empty bin to all_bins
                 all_bins.append(current_bin)
             current_bin = [idx]
-            remaining_capacity = c - size
+            remaining_capacity = bin_capacity - size
             total_used += size
 
     # Add the last bin if not empty
@@ -148,132 +233,227 @@ def allocate_sequentially(lengths: np.ndarray, rank: int, c: int, n: int):
         all_bins.append(current_bin)
 
     # Assign bins to ranks - each rank gets every n-th bin
-    for bin_idx in range(rank, len(all_bins), n):
+    for bin_idx in range(rank, len(all_bins), num_ranks):
         result.append(all_bins[bin_idx])
 
-    return result, total_used, len(all_bins) * c
+    return result, total_used, len(all_bins) * bin_capacity
 
 
 class MultipackBatchSampler(BatchSampler):
-    """Batch sampler class for multipack"""
+    """
+    Batch sampler class for efficient packing of variable-length sequences
+
+    This sampler packs sequences into fixed-capacity bins (batches) to maximize
+    GPU memory utilization and training throughput by reducing padding.
+
+    It supports both parallel packing (using FFD algorithm) and
+    sequential packing (preserving original sequence order).
+    """
 
     def __init__(
         self,
         sampler: Union[Sampler[int], Iterable[int]],
-        batch_size: int,
-        batch_max_len: int,
-        lengths: np.ndarray,
-        packing_efficiency_estimate: float = 1.0,
-        drop_last: bool = False,
-        num_count_samples: int = 16,
-        sequential: bool = False,
-        **kwargs,
+        batch_size: int,  # Number of bins per batch
+        batch_max_len: int,  # Maximum sequence length (bin capacity)
+        lengths: np.ndarray,  # Sequence lengths
+        packing_efficiency_estimate: float = 1.0,  # Initial efficiency estimate
+        drop_last: bool = False,  # Whether to drop final batches (might be incomplete)
+        num_count_samples: int = 16,  # Number of times to estimate batch count
+        sequential: bool = False,  # Whether to use sequential packing
+        group_size: int = 100_000,  # Size of groups for parallel packing
+        bin_size: int = 200,  # The max number of samples that can be packed in a single bin
+        num_processes: int | None = None,  # Number of processes for parallel packing
+        safe_mode: bool = True,  # Conservative packing to prevent training instability
+        **kwargs,  # pylint: disable=unused-argument
     ):
         super().__init__(sampler, batch_size, drop_last)
         self.batch_size = batch_size
         self.batch_max_len = batch_max_len
-        self.lengths: np.ndarray = lengths
+        self.lengths = np.array(lengths, dtype=np.int32)
         self.packing_efficiency_estimate = packing_efficiency_estimate or 1.0
         self.sequential = sequential
+        self.group_size = group_size
+        self.bin_size = bin_size
+        self.num_processes = num_processes
+        self.safe_mode = safe_mode
 
         assert isinstance(self.lengths, np.ndarray)
 
         self.epoch = 0
 
-        # statistics
-        self.eff_total_used = 0
-        self.eff_total_slots = 0
+        # Efficiency statistics tracking
+        self.total_tokens_used = 0
+        self.total_token_slots = 0
 
-        # The number of times to calculate the batches to determine the minimum packed dataset length for the local rank
+        # The number of times to calculate batches to determine minimum packed dataset length
         self.num_count_samples = num_count_samples
-        # the minimum packed dataset length across all ranks determined by a gather/broadcast
+        # Minimum packed dataset length across all ranks (determined by gather/broadcast)
         self.len_across_ranks = None
 
+        # Cache for batches
+        self._batches = None
+
         if self.sequential and not isinstance(sampler, SequentialSampler):
-            LOG.warn(
+            LOG.warning(
                 "using sequential sample packing with non-sequential sampler, did you want to also enable curriculum_sampling?"
             )
 
     def set_epoch(self, epoch: int):
+        """Set the epoch number, used for reproducible shuffling across epochs"""
         self.epoch = epoch
+        self._batches = None  # Invalidate batch cache
 
     def generate_batches(self, set_stats=False):
-        indices = [idx for idx in self.sampler]
+        """
+        Generate packed batches for training
 
-        lengths = self.lengths[indices]
-        lengths_cumsum = np.cumsum(lengths)
+        Args:
+            set_stats: Whether to update efficiency statistics
 
-        if self.sequential:
-            batches, total_used, total_slots = allocate_sequentially(
-                lengths=lengths,
-                rank=0,
-                c=self.batch_max_len,
-                n=1,
-            )
-        else:
-            batches, total_used, total_slots = allocate(
-                lengths=lengths,
-                lengths_cumsum=lengths_cumsum,
-                rank=0,
-                c=self.batch_max_len,
-                n=1,
-            )
+        Returns:
+            List of batches, where each batch contains multiple bins,
+            and each bin contains multiple sequence indices
+        """
+        if self._batches is not None:
+            return self._batches
 
-        batches = [
-            [
-                [indices[b_idx] for b_idx in batch]
-                for batch in batches[i : i + self.batch_size]
-            ]
-            for i in range(0, len(batches), self.batch_size)
+        # Get indices from the sampler
+        indices = [  # pylint: disable=unnecessary-comprehension
+            idx for idx in self.sampler
         ]
 
-        # statistics
-        if set_stats:
-            self.eff_total_used += total_used
-            self.eff_total_slots += total_slots
+        # Get lengths of the selected sequences
+        lengths = self.lengths[indices]
 
+        # Pack sequences into bins using either sequential or parallel packing
+        if self.sequential:
+            bins, total_used, total_slots = allocate_sequentially(
+                lengths,
+                rank=0,
+                bin_capacity=self.batch_max_len,
+                num_ranks=1,
+            )
+            # Map bin indices back to original indices
+            bins = [[indices[b_idx] for b_idx in bin_indices] for bin_indices in bins]
+        else:
+            # Use parallel packing
+            all_bins = pack_parallel(
+                lengths,
+                bin_capacity=self.batch_max_len,
+                group_size=self.group_size,
+                bin_size=self.bin_size,
+                num_processes=self.num_processes,
+                safe_mode=self.safe_mode,
+            )
+
+            # Map bin indices back to original indices
+            bins = [
+                [indices[b_idx] for b_idx in bin_indices] for bin_indices in all_bins
+            ]
+
+            # Calculate efficiency statistics
+            total_used = lengths.sum()
+            total_slots = len(all_bins) * self.batch_max_len
+
+        # Group bins into batches (each batch contains batch_size bins)
+        batches = [
+            bins[i : i + self.batch_size] for i in range(0, len(bins), self.batch_size)
+        ]
+
+        # Drop last batch if requested and it's incomplete
+        if self.drop_last and len(batches[-1]) < self.batch_size:
+            batches = batches[:-1]
+            # Adjust total_slots if we dropped a batch
+            if not self.sequential:
+                total_slots -= (self.batch_size - len(batches[-1])) * self.batch_max_len
+
+        # Update statistics if requested
+        if set_stats:
+            self.total_tokens_used += total_used
+            self.total_token_slots += total_slots
+
+        self._batches = batches
         return batches
 
     def __iter__(self):
+        """
+        Return an iterator over batches
+
+        The batches are truncated to match the minimum number of batches across all ranks
+        to ensure distributed training balance
+        """
         batches = self.generate_batches(set_stats=True)
         if self.len_across_ranks:
-            # make sure the batches we iterate over is truncated to the same min length across all ranks
+            # Truncate batches to ensure all ranks have the same number of batches
             batches = batches[: self.len_across_ranks]
         return iter(batches)
 
-    def num_batches(self):
-        batches = self.generate_batches(set_stats=True)
-        return len(batches)
-
     def efficiency(self):
-        return self.eff_total_used / self.eff_total_slots
+        """
+        Calculate the packing efficiency (ratio of tokens used to total token slots)
+        Higher is better - 1.0 would mean perfect packing with no wasted space
+        """
+        if self.total_token_slots == 0:
+            self.generate_batches(set_stats=True)
+        if self.total_token_slots == 0:
+            return 0.0
+        # Return a Python float instead of potentially a numpy float
+        return float(self.total_tokens_used / self.total_token_slots)
 
     def gather_efficiency(self):
-        def calc_sample_packing_eff_est(estimates: List[float]):
-            LOG.debug(f"sample_packing_eff_est across ranks: {repr(estimates)}")
-            return math.floor(0.997 * max(estimates))
+        """
+        Gather and synchronize packing efficiency estimates across all distributed ranks
+        Returns a conservative efficiency estimate based on the measurements
+        """
 
+        def calc_sample_packing_eff_est(estimates: list[float]):
+            LOG.debug(f"sample_packing_eff_est across ranks: {repr(estimates)}")
+            # Use 99.7% of max observed efficiency as a safe estimate
+            max_eff = max(float(eff) for eff in estimates)
+            return math.floor(0.997 * max_eff)
+
+        # Gather efficiency from all ranks and apply the calculation function
         sample_packing_actual_eff_all = reduce_and_broadcast(
-            lambda: self.efficiency(),  # pylint: disable=unnecessary-lambda
+            lambda: float(self.efficiency()),  # pylint: disable=unnecessary-lambda
             calc_sample_packing_eff_est,
         )
+
+        # Quantize to 0.5% intervals for stability
         sample_packing_eff_est = (
             math.ceil(sample_packing_actual_eff_all * 200.0) / 200.0
         )
         return sample_packing_eff_est
 
     def gather_len_batches(self, num):
+        """
+        Gather and synchronize batch counts across all distributed ranks
+        Returns the minimum number of batches available on any rank
+        """
+
         def calc_min_len(estimates: list[(int, float)]):
             LOG.info(f"gather_len_batches: {repr(estimates)}")
             return math.floor(min(estimates))
 
+        # Find minimum batch count across ranks to ensure balance
         min_len_batches = reduce_and_broadcast(lambda: num, calc_min_len)
         return min_len_batches
 
     def __len__(self):
-        if not self.len_across_ranks:
-            len_batches = min(
-                [self.num_batches() for _ in range(self.num_count_samples)]
+        """
+        Return the total number of batches that will be yielded by this sampler
+
+        This is calculated as the minimum number of batches available on any rank
+        to ensure balanced distributed training
+        """
+        if self._batches is None:
+            self._batches = self.generate_batches(set_stats=True)
+
+        if self.len_across_ranks is None:
+            # Sample multiple times to get stable estimate
+            len_batches = min(  # pylint: disable=consider-using-generator
+                [len(self._batches) for _ in range(self.num_count_samples)]
             )
+            # Gather minimum across all ranks
             self.len_across_ranks = self.gather_len_batches(len_batches)
+
         return self.len_across_ranks

src/axolotl/utils/schemas/config.py

@@ -82,6 +82,7 @@ class AxolotlInputConfig(
     mean_resizing_embeddings: bool | None = False
     # optionally shrink the embeddings when the tokenizer vocab size is smaller
     shrink_embeddings: bool | None = None
+    embeddings_skip_upcast: bool | None = None
 
     rl: RLType | None = None
     trl: TRLConfig | None = Field(
@@ -435,16 +436,6 @@ class AxolotlInputConfig(
             )
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_sample_packing_w_xformers(cls, data):
-        if data.get("sample_packing") and data.get("xformers_attention"):
-            raise ValueError(
-                "sample_packing not compatible with xformers_attention. Use flash_attention"
-            )
-
-        return data
-
     @model_validator(mode="before")
     @classmethod
     # pylint: disable=duplicate-code
@@ -471,9 +462,10 @@ class AxolotlInputConfig(
             and not data.get("flash_attention")
             and not data.get("sdp_attention")
             and not data.get("flex_attention")
+            and not data.get("xformers_attention")
         ):
             LOG.warning(
-                "sample_packing without flash, sdp or flex attention does not handle cross sample decontamination."
+                "sample_packing without flash, sdp, xformers or flex attention does not handle cross sample decontamination."
             )
 
         return data

src/axolotl/utils/schemas/enums.py

@@ -53,4 +53,5 @@ class CustomSupportedOptimizers(str, Enum):
     ao_adamw_8bit = "ao_adamw_8bit"  # pylint: disable=invalid-name
     ao_adamw_fp8 = "ao_adamw_fp8"  # pylint: disable=invalid-name
     adopt_adamw = "adopt_adamw"  # pylint: disable=invalid-name
+    came_pytorch = "came_pytorch"  # pylint: disable=invalid-name
     muon = "muon"  # pylint: disable=invalid-name

src/axolotl/utils/schemas/training.py

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

src/axolotl/utils/trainer.py

@@ -16,7 +16,7 @@ from datasets import IterableDataset, disable_caching, enable_caching
 from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
 from transformers.utils import is_torch_bf16_gpu_available
 
-from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
+from axolotl.core.trainers.builders import HFCausalTrainerBuilder, HFRLTrainerBuilder
 from axolotl.monkeypatch.trainer_eval_guard import patch_evaluation_loop_for_fsdp2
 from axolotl.utils.distributed import reduce_and_broadcast
 from axolotl.utils.environment import check_cuda_p2p_ib_support
@@ -633,8 +633,7 @@ def setup_trainer(
         peft_config: Optional PEFT (Parameter-Efficient Fine-Tuning) configuration. Default is None.
 
     Returns:
-        A trainer instance (either `HFRLTrainer` or `HFCausalTrainer`) configured based
-            on the provided parameters.
+        A trainer instance configured based on the provided parameters.
     """
     if (
         cfg.torch_compile

tests/conftest.py

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

tests/core/test_trainer_builder.py

@@ -1,10 +1,8 @@
-"""
-unit tests for axolotl.core.trainer_builder
-"""
+"""Unit tests for axolotl.core.trainers.builders"""
 
 import pytest
 
-from axolotl.core.trainer_builder import HFRLTrainerBuilder
+from axolotl.core.trainers.builders import HFRLTrainerBuilder
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_model, load_tokenizer
@@ -53,9 +51,7 @@ def fixture_model(cfg, tokenizer):
 
 
 class TestHFRLTrainerBuilder:
-    """
-    TestCase class for DPO trainer builder
-    """
+    """Test case class for RL trainer builder"""
 
     def test_build_training_arguments(self, cfg, model, tokenizer):
         builder = HFRLTrainerBuilder(cfg, model, tokenizer)

tests/e2e/integrations/test_hooks.py

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

tests/e2e/integrations/test_kd.py

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

tests/e2e/multigpu/test_llama.py

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

tests/e2e/patched/lora_kernels/test_lora_kernel_patching.py

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

tests/e2e/patched/test_4d_multipack_llama.py

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

tests/e2e/patched/test_falcon_samplepack.py

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

tests/e2e/patched/test_mistral_samplepack.py

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

tests/e2e/patched/test_mixtral_samplepack.py

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

tests/e2e/patched/test_model_patches.py

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

tests/e2e/patched/test_peft_embeddings.py

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

tests/e2e/patched/test_phi_multipack.py

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

tests/e2e/patched/test_resume.py

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

tests/e2e/test_evaluate.py

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

tests/e2e/test_falcon.py

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

tests/e2e/test_imports.py

@@ -1,21 +1,21 @@
-"""
-test module to import various submodules that have historically broken due to dependency issues
+"""Test module to import various submodules that have historically broken due to
+dependency issues.
 """
 
 import unittest
 
 
 class TestImports(unittest.TestCase):
-    """
-    Test class to import various submodules that have historically broken due to dependency issues
+    """Test class to import various submodules that have historically broken due to
+    dependency issues.
     """
 
     def test_import_causal_trainer(self):
-        from axolotl.core.trainer_builder import (  # pylint: disable=unused-import  # noqa: F401
+        from axolotl.core.trainers.builders import (  # pylint: disable=unused-import  # noqa: F401
             HFCausalTrainerBuilder,
         )
 
     def test_import_rl_trainer(self):
-        from axolotl.core.trainer_builder import (  # pylint: disable=unused-import  # noqa: F401
+        from axolotl.core.trainers.builders import (  # pylint: disable=unused-import  # noqa: F401
             HFRLTrainerBuilder,
         )

tests/e2e/test_mistral.py

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

tests/e2e/test_optimizers.py

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

tests/test_datasets.py

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

tests/test_packed_batch_sampler.py

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