fix the torch dtype check

.github/workflows/base.yml

@@ -30,12 +30,7 @@ jobs:
           - cuda: "121"
             cuda_version: 12.1.0
             python_version: "3.11"
-            pytorch: 2.2.2
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-          - cuda: "121"
-            cuda_version: 12.1.0
-            python_version: "3.11"
-            pytorch: 2.3.0
+            pytorch: 2.2.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
     steps:
       - name: Checkout

.github/workflows/lint.yml

@@ -7,7 +7,6 @@ on:
        - 'requirements.txt'
        - '.github/workflows/*.yml'
        - "*.md"
-       - "examples/**/*.y[a]?ml"
   workflow_dispatch:
 
 jobs:

.github/workflows/main.yml

@@ -28,12 +28,7 @@ jobs:
           - cuda: 121
             cuda_version: 12.1.0
             python_version: "3.11"
-            pytorch: 2.2.2
-            axolotl_extras:
-          - cuda: 121
-            cuda_version: 12.1.0
-            python_version: "3.11"
-            pytorch: 2.3.0
+            pytorch: 2.2.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -89,12 +84,7 @@ jobs:
           - cuda: 121
             cuda_version: 12.1.0
             python_version: "3.11"
-            pytorch: 2.2.2
-            axolotl_extras:
-          - cuda: 121
-            cuda_version: 12.1.0
-            python_version: "3.11"
-            pytorch: 2.3.0
+            pytorch: 2.2.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -125,45 +115,3 @@ jobs:
              ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
              ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
           labels: ${{ steps.metadata.outputs.labels }}
-
-  build-axolotl-cloud-no-tmux:
-    needs: build-axolotl
-    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]]') && github.repository_owner == 'OpenAccess-AI-Collective' }}
-    # this job needs to be run on self-hosted GPU runners...
-    strategy:
-      matrix:
-        include:
-          - cuda: 121
-            cuda_version: 12.1.0
-            python_version: "3.11"
-            pytorch: 2.3.0
-            axolotl_extras:
-    runs-on: axolotl-gpu-runner
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-      - name: Docker metadata
-        id: metadata
-        uses: docker/metadata-action@v5
-        with:
-          images: winglian/axolotl-cloud-term
-      - name: Login to Docker Hub
-        uses: docker/login-action@v3
-        with:
-          username: ${{ secrets.DOCKERHUB_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_TOKEN }}
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
-      - name: Build
-        uses: docker/build-push-action@v5
-        with:
-          context: .
-          build-args: |
-            BASE_TAG=${{ github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
-            CUDA=${{ matrix.cuda }}
-          file: ./docker/Dockerfile-cloud-no-tmux
-          push: ${{ github.event_name != 'pull_request' }}
-          tags: |
-             ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
-             ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
-          labels: ${{ steps.metadata.outputs.labels }}

.github/workflows/nightlies.yml

@@ -27,12 +27,7 @@ jobs:
           - cuda: 121
             cuda_version: 12.1.0
             python_version: "3.11"
-            pytorch: 2.2.2
-            axolotl_extras:
-          - cuda: 121
-            cuda_version: 12.1.0
-            python_version: "3.11"
-            pytorch: 2.3.0
+            pytorch: 2.2.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -89,12 +84,7 @@ jobs:
           - cuda: 121
             cuda_version: 12.1.0
             python_version: "3.11"
-            pytorch: 2.2.2
-            axolotl_extras:
-          - cuda: 121
-            cuda_version: 12.1.0
-            python_version: "3.11"
-            pytorch: 2.3.0
+            pytorch: 2.2.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:

.github/workflows/tests.yml

@@ -82,12 +82,7 @@ jobs:
           - cuda: 121
             cuda_version: 12.1.0
             python_version: "3.11"
-            pytorch: 2.2.2
-            num_gpus: 1
-          - cuda: 121
-            cuda_version: 12.1.0
-            python_version: "3.11"
-            pytorch: 2.3.0
+            pytorch: 2.2.1
             num_gpus: 1
     steps:
       - name: Checkout

.gitignore

@@ -133,7 +133,6 @@ venv/
 ENV/
 env.bak/
 venv.bak/
-venv3.10/
 
 # Spyder project settings
 .spyderproject
@@ -176,9 +175,3 @@ qlora-out/*
 mlruns/*
 
 /.quarto/
-prepared-datasets/
-submit.sh
-*.out*
-
-typings/
-out/

README.md

@@ -34,7 +34,6 @@ Features:
   - [Mac](#mac)
   - [Google Colab](#google-colab)
   - [Launching on public clouds via SkyPilot](#launching-on-public-clouds-via-skypilot)
-  - [Launching on public clouds via dstack](#launching-on-public-clouds-via-dstack)
 - [Dataset](#dataset)
 - [Config](#config)
   - [Train](#train)
@@ -45,7 +44,6 @@ Features:
 - Advanced Topics
   - [Multipack](./docs/multipack.qmd)<svg width="24" height="24" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><path d="M17 13.5v6H5v-12h6m3-3h6v6m0-6-9 9" class="icon_svg-stroke" stroke="#666" stroke-width="1.5" fill="none" fill-rule="evenodd" stroke-linecap="round" stroke-linejoin="round"></path></svg>
   - [RLHF & DPO](./docs/rlhf.qmd)<svg width="24" height="24" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><path d="M17 13.5v6H5v-12h6m3-3h6v6m0-6-9 9" class="icon_svg-stroke" stroke="#666" stroke-width="1.5" fill="none" fill-rule="evenodd" stroke-linecap="round" stroke-linejoin="round"></path></svg>
-  - [Dataset Pre-Processing](./docs/dataset_preprocessing.qmd)<svg width="24" height="24" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><path d="M17 13.5v6H5v-12h6m3-3h6v6m0-6-9 9" class="icon_svg-stroke" stroke="#666" stroke-width="1.5" fill="none" fill-rule="evenodd" stroke-linecap="round" stroke-linejoin="round"></path></svg>
 - [Common Errors](#common-errors-)
   - [Tokenization Mismatch b/w Training & Inference](#tokenization-mismatch-bw-inference--training)
 - [Debugging Axolotl](#debugging-axolotl)
@@ -83,7 +81,6 @@ Features:
 | llama       | ✅         | ✅    | ✅     | ✅             | ✅                 | ✅          | ✅            |
 | Mistral     | ✅         | ✅    | ✅     | ✅             | ✅                 | ✅          | ✅            |
 | Mixtral-MoE | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
-| Mixtral8X22 | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
 | Pythia      | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
 | cerebras    | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
 | btlm        | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
@@ -124,11 +121,11 @@ accelerate launch -m axolotl.cli.train examples/openllama-3b/lora.yml
 
 # inference
 accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
-    --lora_model_dir="./outputs/lora-out"
+    --lora_model_dir="./lora-out"
 
 # gradio
 accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
-    --lora_model_dir="./outputs/lora-out" --gradio
+    --lora_model_dir="./lora-out" --gradio
 
 # remote yaml files - the yaml config can be hosted on a public URL
 # Note: the yaml config must directly link to the **raw** yaml
@@ -293,42 +290,6 @@ HF_TOKEN=xx sky launch axolotl.yaml --env HF_TOKEN
 HF_TOKEN=xx BUCKET=<unique-name> sky spot launch axolotl-spot.yaml --env HF_TOKEN --env BUCKET
 ```
 
-#### Launching on public clouds via dstack
-To launch on GPU instance (both on-demand and spot instances) on public clouds (GCP, AWS, Azure, Lambda Labs, TensorDock, Vast.ai, and CUDO), you can use [dstack](https://dstack.ai/).
-
-Write a job description in YAML as below:
-
-```yaml
-# dstack.yaml
-type: task
-
-image: winglian/axolotl-cloud:main-20240429-py3.11-cu121-2.2.2
-
-env:
-  - HUGGING_FACE_HUB_TOKEN
-  - WANDB_API_KEY
-
-commands:
-  - accelerate launch -m axolotl.cli.train config.yaml
-
-ports:
-  - 6006
-
-resources:
-  gpu:
-    memory: 24GB..
-    count: 2
-```
-
-then, simply run the job with `dstack run` command. Append `--spot` option if you want spot instance. `dstack run` command will show you the instance with cheapest price across multi cloud services:
-
-```bash
-pip install dstack
-HUGGING_FACE_HUB_TOKEN=xxx WANDB_API_KEY=xxx dstack run . -f dstack.yaml # --spot
-```
-
-For further and fine-grained use cases, please refer to the official [dstack documents](https://dstack.ai/docs/) and the detailed description of [axolotl example](https://github.com/dstackai/dstack/tree/master/examples/fine-tuning/axolotl) on the official repository.
-
 ### Dataset
 
 Axolotl supports a variety of dataset formats.  It is recommended to use a JSONL.  The schema of the JSONL depends upon the task and the prompt template you wish to use.  Instead of a JSONL, you can also use a HuggingFace dataset with columns for each JSONL field.
@@ -464,7 +425,7 @@ deepspeed: deepspeed_configs/zero1.json
 ```
 
 ```shell
-accelerate launch -m axolotl.cli.train examples/llama-2/config.yml --deepspeed deepspeed_configs/zero1.json
+accelerate launch -m axolotl.cli.train examples/llama-2/config.py --deepspeed deepspeed_configs/zero1.json
 ```
 
 ##### FSDP
@@ -609,7 +570,7 @@ If you decode a prompt constructed by axolotl, you might see spaces between toke
 3. Make sure the inference string from #2 looks **exactly** like the data you fine tuned on from #1, including spaces and new lines.  If they aren't the same, adjust your inference server accordingly.
 4. As an additional troubleshooting step, you can look at the token ids between 1 and 2 to make sure they are identical.
 
-Having misalignment between your prompts during training and inference can cause models to perform very poorly, so it is worth checking this.  See [this blog post](https://hamel.dev/notes/llm/finetuning/05_tokenizer_gotchas.html) for a concrete example.
+Having misalignment between your prompts during training and inference can cause models to perform very poorly, so it is worth checking this.  See [this blog post](https://hamel.dev/notes/llm/05_tokenizer_gotchas.html) for a concrete example.
 
 ## Debugging Axolotl
 

cicd/cicd.sh

@@ -1,5 +1,4 @@
 #!/bin/bash
-set -e
 
 pytest --ignore=tests/e2e/ /workspace/axolotl/tests/
 pytest /workspace/axolotl/tests/e2e/patched/

deepspeed_configs/zero3_bf16_cpuoffload_all.json

@@ -1,6 +1,4 @@
 {
-  "zero_force_ds_cpu_optimizer": false,
-  "zero_allow_untested_optimizer": true,
   "zero_optimization": {
     "stage": 3,
     "offload_optimizer": {

deepspeed_configs/zero3_bf16_cpuoffload_params.json

@@ -1,6 +1,4 @@
 {
-  "zero_force_ds_cpu_optimizer": false,
-  "zero_allow_untested_optimizer": true,
   "zero_optimization": {
     "stage": 3,
     "offload_param": {

docker/Dockerfile

@@ -11,7 +11,7 @@ ARG PYTORCH_VERSION="2.1.2"
 ENV PYTORCH_VERSION=$PYTORCH_VERSION
 
 RUN apt-get update && \
-    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev rsync s3fs
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev
 
 WORKDIR /workspace
 

docker/Dockerfile-cloud-no-tmux

@@ -1,27 +0,0 @@
-ARG BASE_TAG=main
-FROM winglian/axolotl:$BASE_TAG
-
-ENV HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets"
-ENV HUGGINGFACE_HUB_CACHE="/workspace/data/huggingface-cache/hub"
-ENV TRANSFORMERS_CACHE="/workspace/data/huggingface-cache/hub"
-ENV HF_HOME="/workspace/data/huggingface-cache/hub"
-ENV HF_HUB_ENABLE_HF_TRANSFER="1"
-
-EXPOSE 8888
-EXPOSE 22
-
-COPY scripts/cloud-entrypoint-term.sh /root/cloud-entrypoint.sh
-COPY scripts/motd /etc/motd
-
-RUN pip install jupyterlab notebook ipywidgets && \
-    jupyter lab clean
-RUN apt install --yes --no-install-recommends openssh-server tmux sudo && \
-    pip3 install -U --no-cache-dir grpcio ray[default]==2.9.3 && \
-    mkdir -p ~/.ssh && \
-    chmod 700 ~/.ssh && \
-    printf "[ ! -z \"\$TERM\" -a -r /etc/motd ] && cat /etc/motd\n" >> ~/.bashrc && \
-    chmod +x /workspace/axolotl/scripts/cloud-entrypoint.sh && \
-    chmod +x /root/cloud-entrypoint.sh
-
-ENTRYPOINT ["/root/cloud-entrypoint.sh"]
-CMD ["sleep", "infinity"]

docs/config.qmd

@@ -186,11 +186,6 @@ eval_sample_packing:
 # The trainer will provide recommended values for these values.
 sample_packing_eff_est:
 total_num_tokens:
-# Increasing the following values helps with packing, but usually only slightly (<%1.)
-# The number of samples packed at a time.
-sample_packing_group_size: 100000
-# The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.
-sample_packing_bin_size: 200
 
 # Passed through to transformers when loading the model when launched without accelerate
 # Use `sequential` when training w/ model parallelism to limit memory
@@ -232,12 +227,6 @@ lora_modules_to_save:
 
 lora_fan_in_fan_out: false
 
-# LoRA+ hyperparameters
-# For more details about the following options, see:
-# https://arxiv.org/abs/2402.12354  and `src/axolotl/core/train_builder.py`
-loraplus_lr_ratio: # loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4.
-loraplus_lr_embedding: #  loraplus learning rate for lora embedding layers. Default value is 1e-6.
-
 peft:
   # Configuration options for loftq initialization for LoRA
   # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization
@@ -279,7 +268,6 @@ torch_compile_backend:  # Optional[str]
 # If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.
 gradient_accumulation_steps: 1
 # The number of samples to include in each batch. This is the number of samples sent to each GPU.
-# Batch size per gpu = micro_batch_size * gradient_accumulation_steps
 micro_batch_size: 2
 eval_batch_size:
 num_epochs: 4
@@ -290,7 +278,7 @@ lr_quadratic_warmup:
 logging_steps:
 eval_steps: # Leave empty to eval at each epoch, integers for every N steps. decimal for fraction of total steps
 evals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps
-save_strategy: # Set to `"no"` to skip checkpoint saves
+save_strategy: # Set to `no` to skip checkpoint saves
 save_steps: # Leave empty to save at each epoch
 saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps
 save_total_limit: # Checkpoints saved at a time
@@ -424,7 +412,6 @@ special_tokens:
   # bos_token: "<s>"
   # eos_token: "</s>"
   # unk_token: "<unk>"
-  # pad_token: "[PAD]"
 
 # Add extra tokens.
 tokens:

docs/dataset_preprocessing.qmd

@@ -1,35 +0,0 @@
----
-title: Dataset Preprocessing
-description: How datasets are processed
----
-
-Dataset pre-processing is the step where Axolotl takes each dataset you've configured alongside
-the (dataset format)[../dataset-formats/] and prompt strategies to:
- - parse the dataset based on the *dataset format*
- - transform the dataset to how you would interact with the model based on the *prompt strategy*
- - tokenize the dataset based on the configured model & tokenizer
- - shuffle and merge multiple datasets together if using more than one
-
-The processing of the datasets can happen one of two ways:
-
-1. Before kicking off training by calling `python -m axolotl.cli.preprocess /path/to/your.yaml --debug`
-2. When training is started
-
-What are the benefits of pre-processing? When training interactively or for sweeps
-(e.g. you are restarting the trainer often), processing the datasets can oftentimes be frustratingly
-slow. Pre-processing will cache the tokenized/formatted datasets according to a hash of dependent
-training parameters so that it will intelligently pull from its cache when possible.
-
-The path of the cache is controlled by `dataset_prepared_path:` and is often left blank in example
-YAMLs as this leads to a more robust solution that prevents unexpectedly reusing cached data.
-
-If `dataset_prepared_path:` is left empty, when training, the processed dataset will be cached in a
-default path of `./last_run_prepared/`, but will ignore anything already cached there. By explicitly
-setting `dataset_prepared_path: ./last_run_prepared`, the trainer will use whatever pre-processed
-data is in the cache.
-
-What are the edge cases? Let's say you are writing a custom prompt strategy or using a user-defined
-prompt template. Because the trainer cannot readily detect these changes, we cannot change the
-calculated hash value for the pre-processed dataset. If you have `dataset_prepared_path: ...` set
-and change your prompt templating logic, it may not pick up the changes you made and you will be
-training over the old prompt.

docs/rlhf.qmd

@@ -49,7 +49,7 @@ remove_unused_columns: false
 chat_template: chatml
 datasets:
   - path: argilla/ultrafeedback-binarized-preferences-cleaned
-    type: chat_template.argilla
+    type: orpo.chat_template
 ```
 
 #### Using local dataset files

examples/cerebras/btlm-ft.yml

@@ -38,7 +38,7 @@ wandb_watch:
 wandb_name:
 wandb_log_model:
 
-output_dir: ./outputs/btlm-out
+output_dir: btlm-out
 gradient_accumulation_steps: 1
 micro_batch_size: 1
 num_epochs: 1

examples/cerebras/qlora.yml

@@ -25,7 +25,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 batch_size: 4
 micro_batch_size: 4
 num_epochs: 2

examples/code-llama/13b/lora.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 
 sequence_len: 4096
 sample_packing: true

examples/code-llama/13b/qlora.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 
 adapter: qlora
 lora_model_dir:

examples/code-llama/34b/lora.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 
 sequence_len: 4096
 sample_packing: true

examples/code-llama/34b/qlora.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 
 adapter: qlora
 lora_model_dir:

examples/code-llama/7b/lora.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 
 sequence_len: 4096
 sample_packing: true

examples/code-llama/7b/qlora.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 
 adapter: qlora
 lora_model_dir:

examples/colab-notebooks/colab-axolotl-example.ipynb

@@ -1,223 +1,216 @@
 {
- "cells": [
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "AKjdG7tbTb-n"
-   },
-   "source": [
-    "# Example notebook for running Axolotl on google colab"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "id": "RcbNpOgWRcii"
-   },
-   "outputs": [],
-   "source": [
-    "import torch\n",
-    "# Check so there is a gpu available, a T4(free tier) is enough to run this notebook\n",
-    "assert (torch.cuda.is_available()==True)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "h3nLav8oTRA5"
-   },
-   "source": [
-    "## Install Axolotl and dependencies"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "colab": {
-     "base_uri": "https://localhost:8080/"
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "AKjdG7tbTb-n"
+      },
+      "source": [
+        "# Example notebook for running Axolotl on google colab"
+      ]
     },
-    "id": "3c3yGAwnOIdi",
-    "outputId": "e3777b5a-40ef-424f-e181-62dfecd1dd01"
-   },
-   "outputs": [],
-   "source": [
-    "!pip install torch==\"2.1.2\"\n",
-    "!pip install -e git+https://github.com/OpenAccess-AI-Collective/axolotl#egg=axolotl\n",
-    "!pip install flash-attn==\"2.5.0\"\n",
-    "!pip install deepspeed==\"0.13.1\"!pip install mlflow==\"2.13.0\""
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "BW2MFr7HTjub"
-   },
-   "source": [
-    "## Create an yaml config file"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "id": "9pkF2dSoQEUN"
-   },
-   "outputs": [],
-   "source": [
-    "import yaml\n",
-    "\n",
-    "# Your YAML string\n",
-    "yaml_string = \"\"\"\n",
-    "base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T\n",
-    "model_type: LlamaForCausalLM\n",
-    "tokenizer_type: LlamaTokenizer\n",
-    "\n",
-    "load_in_8bit: false\n",
-    "load_in_4bit: true\n",
-    "strict: false\n",
-    "\n",
-    "datasets:\n",
-    "  - path: mhenrichsen/alpaca_2k_test\n",
-    "    type: alpaca\n",
-    "dataset_prepared_path:\n",
-    "val_set_size: 0.05\n",
-    "output_dir: ./outputs/qlora-out\n",
-    "\n",
-    "adapter: qlora\n",
-    "lora_model_dir:\n",
-    "\n",
-    "sequence_len: 4096\n",
-    "sample_packing: true\n",
-    "eval_sample_packing: false\n",
-    "pad_to_sequence_len: true\n",
-    "\n",
-    "lora_r: 32\n",
-    "lora_alpha: 16\n",
-    "lora_dropout: 0.05\n",
-    "lora_target_modules:\n",
-    "lora_target_linear: true\n",
-    "lora_fan_in_fan_out:\n",
-    "\n",
-    "wandb_project:\n",
-    "wandb_entity:\n",
-    "wandb_watch:\n",
-    "wandb_name:\n",
-    "wandb_log_model:\n",
-    "\n",
-    "gradient_accumulation_steps: 4\n",
-    "micro_batch_size: 2\n",
-    "num_epochs: 4\n",
-    "optimizer: paged_adamw_32bit\n",
-    "lr_scheduler: cosine\n",
-    "learning_rate: 0.0002\n",
-    "\n",
-    "train_on_inputs: false\n",
-    "group_by_length: false\n",
-    "bf16: auto\n",
-    "fp16:\n",
-    "tf32: false\n",
-    "\n",
-    "gradient_checkpointing: true\n",
-    "early_stopping_patience:\n",
-    "resume_from_checkpoint:\n",
-    "local_rank:\n",
-    "logging_steps: 1\n",
-    "xformers_attention:\n",
-    "flash_attention: true\n",
-    "\n",
-    "warmup_steps: 10\n",
-    "evals_per_epoch: 4\n",
-    "saves_per_epoch: 1\n",
-    "debug:\n",
-    "deepspeed:\n",
-    "weight_decay: 0.0\n",
-    "fsdp:\n",
-    "fsdp_config:\n",
-    "special_tokens:\n",
-    "\n",
-    "\"\"\"\n",
-    "\n",
-    "# Convert the YAML string to a Python dictionary\n",
-    "yaml_dict = yaml.safe_load(yaml_string)\n",
-    "\n",
-    "# Specify your file path\n",
-    "file_path = 'test_axolotl.yaml'\n",
-    "\n",
-    "# Write the YAML file\n",
-    "with open(file_path, 'w') as file:\n",
-    "    yaml.dump(yaml_dict, file)\n"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "bidoj8YLTusD"
-   },
-   "source": [
-    "## Launch the training"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "colab": {
-     "base_uri": "https://localhost:8080/"
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "RcbNpOgWRcii"
+      },
+      "outputs": [],
+      "source": [
+        "import torch\n",
+        "# Check so there is a gpu available, a T4(free tier) is enough to run this notebook\n",
+        "assert (torch.cuda.is_available()==True)"
+      ]
     },
-    "id": "ydTI2Jk2RStU",
-    "outputId": "d6d0df17-4b53-439c-c802-22c0456d301b"
-   },
-   "outputs": [],
-   "source": [
-    "# Buy using the ! the comand will be executed as a bash command\n",
-    "!accelerate launch -m axolotl.cli.train /content/test_axolotl.yaml"
-   ]
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "h3nLav8oTRA5"
+      },
+      "source": [
+        "## Install Axolotl and dependencies"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "id": "3c3yGAwnOIdi",
+        "outputId": "e3777b5a-40ef-424f-e181-62dfecd1dd01"
+      },
+      "outputs": [],
+      "source": [
+        "!pip install torch==\"2.1.2\"\n",
+        "!pip install -e git+https://github.com/OpenAccess-AI-Collective/axolotl#egg=axolotl\n",
+        "!pip install flash-attn==\"2.5.0\"\n",
+        "!pip install deepspeed==\"0.13.1\""
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "BW2MFr7HTjub"
+      },
+      "source": [
+        "## Create an yaml config file"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "9pkF2dSoQEUN"
+      },
+      "outputs": [],
+      "source": [
+        "import yaml\n",
+        "\n",
+        "# Your YAML string\n",
+        "yaml_string = \"\"\"\n",
+        "base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T\n",
+        "model_type: LlamaForCausalLM\n",
+        "tokenizer_type: LlamaTokenizer\n",
+        "is_llama_derived_model: true\n",
+        "\n",
+        "load_in_8bit: false\n",
+        "load_in_4bit: true\n",
+        "strict: false\n",
+        "\n",
+        "datasets:\n",
+        "  - path: mhenrichsen/alpaca_2k_test\n",
+        "    type: alpaca\n",
+        "dataset_prepared_path:\n",
+        "val_set_size: 0.05\n",
+        "output_dir: ./qlora-out\n",
+        "\n",
+        "adapter: qlora\n",
+        "lora_model_dir:\n",
+        "\n",
+        "sequence_len: 1096\n",
+        "sample_packing: true\n",
+        "pad_to_sequence_len: true\n",
+        "\n",
+        "lora_r: 32\n",
+        "lora_alpha: 16\n",
+        "lora_dropout: 0.05\n",
+        "lora_target_modules:\n",
+        "lora_target_linear: true\n",
+        "lora_fan_in_fan_out:\n",
+        "\n",
+        "wandb_project:\n",
+        "wandb_entity:\n",
+        "wandb_watch:\n",
+        "wandb_name:\n",
+        "wandb_log_model:\n",
+        "\n",
+        "mlflow_experiment_name: colab-example\n",
+        "\n",
+        "gradient_accumulation_steps: 1\n",
+        "micro_batch_size: 1\n",
+        "num_epochs: 4\n",
+        "max_steps: 20\n",
+        "optimizer: paged_adamw_32bit\n",
+        "lr_scheduler: cosine\n",
+        "learning_rate: 0.0002\n",
+        "\n",
+        "train_on_inputs: false\n",
+        "group_by_length: false\n",
+        "bf16: false\n",
+        "fp16: true\n",
+        "tf32: false\n",
+        "\n",
+        "gradient_checkpointing: true\n",
+        "early_stopping_patience:\n",
+        "resume_from_checkpoint:\n",
+        "local_rank:\n",
+        "logging_steps: 1\n",
+        "xformers_attention:\n",
+        "flash_attention: false\n",
+        "\n",
+        "warmup_steps: 10\n",
+        "evals_per_epoch:\n",
+        "saves_per_epoch:\n",
+        "debug:\n",
+        "deepspeed:\n",
+        "weight_decay: 0.0\n",
+        "fsdp:\n",
+        "fsdp_config:\n",
+        "special_tokens:\n",
+        "\n",
+        "\"\"\"\n",
+        "\n",
+        "# Convert the YAML string to a Python dictionary\n",
+        "yaml_dict = yaml.safe_load(yaml_string)\n",
+        "\n",
+        "# Specify your file path\n",
+        "file_path = 'test_axolotl.yaml'\n",
+        "\n",
+        "# Write the YAML file\n",
+        "with open(file_path, 'w') as file:\n",
+        "    yaml.dump(yaml_dict, file)\n"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "bidoj8YLTusD"
+      },
+      "source": [
+        "## Launch the training"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "id": "ydTI2Jk2RStU",
+        "outputId": "d6d0df17-4b53-439c-c802-22c0456d301b"
+      },
+      "outputs": [],
+      "source": [
+        "# Buy using the ! the comand will be executed as a bash command\n",
+        "!accelerate launch -m axolotl.cli.train /content/test_axolotl.yaml"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {},
+      "source": [
+        "## Play with inference"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {},
+      "outputs": [],
+      "source": [
+        "# Buy using the ! the comand will be executed as a bash command\n",
+        "!accelerate launch -m axolotl.cli.inference /content/test_axolotl.yaml \\\n",
+        "    --qlora_model_dir=\"./qlora-out\" --gradio"
+      ]
+    }
+  ],
+  "metadata": {
+    "accelerator": "GPU",
+    "colab": {
+      "gpuType": "T4",
+      "provenance": []
+    },
+    "kernelspec": {
+      "display_name": "Python 3",
+      "name": "python3"
+    },
+    "language_info": {
+      "name": "python"
+    }
   },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Play with inference"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# Buy using the ! the comand will be executed as a bash command\n",
-    "!accelerate launch -m axolotl.cli.inference /content/test_axolotl.yaml \\\n",
-    "    --qlora_model_dir=\"./qlora-out\" --gradio"
-   ]
-  }
- ],
- "metadata": {
-  "accelerator": "GPU",
-  "colab": {
-   "gpuType": "T4",
-   "provenance": []
-  },
-  "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
-   "language": "python",
-   "name": "python3"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.12.1"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 4
+  "nbformat": 4,
+  "nbformat_minor": 0
 }

examples/dbrx/16bit-lora.yaml

@@ -1,81 +0,0 @@
-base_model: LnL-AI/dbrx-base-converted-v2
-trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.0
-output_dir: ./outputs/out
-
-sequence_len: 512
-sample_packing: false
-pad_to_sequence_len: false
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-adapter: lora
-lora_model_dir:
-lora_r: 8
-lora_alpha: 16
-lora_dropout: 0.05
-# w1, w2, & v1 will hang the trainer
-lora_target_modules:
-  - q_proj # attn
-  - k_proj # attn
-  - v_proj # attn
-  - out_proj # attn
-  - layer # router
-#  - w1
-#  - w2
-#  - v1
-
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 1
-optimizer: paged_adamw_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: false  # don't use with fsdp_activation_checkpointing
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 10
-evals_per_epoch:
-saves_per_epoch: 1
-debug:
-weight_decay: 0.0
-fsdp:
-  - full_shard
-  - auto_wrap
-fsdp_config:
-  fsdp_limit_all_gathers: true
-  fsdp_sync_module_states: true
-  fsdp_offload_params: false
-  fsdp_use_orig_params: false
-  fsdp_cpu_ram_efficient_loading: true
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  fsdp_transformer_layer_cls_to_wrap: DbrxBlock
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_activation_checkpointing: true

examples/dbrx/8bit-lora.yaml

@@ -1,81 +0,0 @@
-base_model: LnL-AI/dbrx-base-converted-v2
-trust_remote_code: true
-
-load_in_8bit: true
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.0
-output_dir: ./outputs/out
-
-sequence_len: 512
-sample_packing: false
-pad_to_sequence_len: false
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-adapter: lora
-lora_model_dir:
-lora_r: 8
-lora_alpha: 16
-lora_dropout: 0.05
-# w1, w2, & v1 will hang the trainer
-lora_target_modules:
-  - q_proj # attn
-  - k_proj # attn
-  - v_proj # attn
-  - out_proj # attn
-  - layer # router
-#  - w1
-#  - w2
-#  - v1
-
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 1
-optimizer: paged_adamw_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: false  # don't use with fsdp_activation_checkpointing
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 10
-evals_per_epoch:
-saves_per_epoch: 1
-debug:
-weight_decay: 0.0
-fsdp:
-  - full_shard
-  - auto_wrap
-fsdp_config:
-  fsdp_limit_all_gathers: true
-  fsdp_sync_module_states: true
-  fsdp_offload_params: false
-  fsdp_use_orig_params: false
-  fsdp_cpu_ram_efficient_loading: true
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  fsdp_transformer_layer_cls_to_wrap: DbrxBlock
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_activation_checkpointing: true

examples/dbrx/README.md

@@ -1,26 +0,0 @@
-# DBRX MoE
-
-Currently, for LoRA, only the `q_proj`, `k_proj`, `v_proj` `out_proj` and `layer` Linear layers are trainable.
-
-We are using the "converted" base models based on [this issue](https://huggingface.co/databricks/dbrx-instruct/discussions/10)
-where the Experts are fused as an `nn.Parameter` rather than a `nn.Linear` layer. However, the implementation
-is still a bit buggy and attempting to train a LoRA adapter over those `w1`, `w2` and `v1` layers
-results in the trainer hanging.
-
-
-### FSDP
-We've tested using the [`LnL-AI/dbrx-base-converted-v2`](https://huggingface.co/LnL-AI/dbrx-base-converted-v2) model as the base model for FSDP.
-
-The high memory usage seen w/ FSDP is due to FSDP not supporting 8bit optimizers.
-
-- 16-bit LoRA w/ FSDP
-  - ✅ w/o CPU Offload - 8x80GB uses ~80GiB/gpu
-  - ❌ w/ CPU Offload - `paged_adamw_8bit` optimizer errors from being on cpu
-- ✅ 8-bit LoRA w/ FSDP
-- ❌ 4-bit QLoRA w/ FSDP - errors w/: `Error an illegal memory access was encountered at line 90 in file /src/csrc/ops.cu`
-- ✅ bf16 full finetune w/ FSDP, freezing all but first 8 layers (8x80GB uses ~78GiB/gpu)
-
-
-### Deepspeed
-
-WIP

examples/dbrx/fft-ds-zero3.yaml

@@ -1,56 +0,0 @@
-base_model: LnL-AI/dbrx-base-converted-v2
-trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.0
-output_dir: ./outputs/out
-
-sequence_len: 512
-sample_packing: false
-pad_to_sequence_len: false
-
-unfrozen_parameters:
-  - transformer.blocks.[0-7].
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 1
-optimizer: paged_adamw_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 10
-evals_per_epoch:
-saves_per_epoch: 1
-debug:
-weight_decay: 0.0
-deepspeed: deepspeed_configs/zero3_bf16.json

examples/falcon/config-7b-lora.yml

@@ -28,7 +28,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/falcon-7b
+output_dir: ./falcon-7b
 batch_size: 2
 micro_batch_size: 1
 num_epochs: 4

examples/falcon/config-7b-qlora.yml

@@ -42,7 +42,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 
 # QLoRA paper Table 9
 # - 16 for 7b & 13b

examples/falcon/config-7b.yml

@@ -28,7 +28,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/falcon-7b
+output_dir: ./falcon-7b
 batch_size: 2
 micro_batch_size: 1
 num_epochs: 4

examples/gemma/qlora.yml

@@ -12,7 +12,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 val_set_size: 0.1
-output_dir: ./outputs/out
+output_dir: ./out
 
 adapter: qlora
 lora_r: 32

examples/gptj/qlora.yml

@@ -23,7 +23,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 gradient_accumulation_steps: 2
 micro_batch_size: 2
 num_epochs: 2

examples/jamba/qlora.yaml

@@ -10,7 +10,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.0
-output_dir: ./outputs/out
+output_dir: ./out
 
 sequence_len: 4096
 sample_packing: false

examples/jamba/qlora_deepspeed.yaml

@@ -10,7 +10,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.0
-output_dir: ./outputs/out
+output_dir: ./out
 
 sequence_len: 4096
 sample_packing: false

examples/jeopardy-bot/config.yml

@@ -21,7 +21,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/jeopardy-bot-7b
+output_dir: ./jeopardy-bot-7b
 gradient_accumulation_steps: 1
 micro_batch_size: 1
 num_epochs: 4

examples/llama-2/fft_optimized.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.05
-output_dir: ./outputs/out
+output_dir: ./out
 
 sequence_len: 4096
 sample_packing: true

examples/llama-2/gptq-lora.yml

@@ -33,7 +33,7 @@ wandb_project:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/model-out
+output_dir: ./model-out
 gradient_accumulation_steps: 1
 micro_batch_size: 1
 num_epochs: 4

examples/llama-2/lisa.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.05
-output_dir: ./outputs/lisa-out
+output_dir: ./lisa-out
 
 sequence_len: 4096
 sample_packing: true

examples/llama-2/loftq.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 
 sequence_len: 4096
 sample_packing: true

examples/llama-2/lora.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 
 sequence_len: 4096
 sample_packing: true

examples/llama-2/qlora-fsdp.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.05
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 
 adapter: qlora
 lora_model_dir:
@@ -65,14 +65,12 @@ deepspeed:
 weight_decay: 0.0
 fsdp:
   - full_shard
-  - auto_wrap
 fsdp_config:
   fsdp_limit_all_gathers: true
   fsdp_sync_module_states: true
   fsdp_offload_params: true
   fsdp_use_orig_params: false
   fsdp_cpu_ram_efficient_loading: true
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
-  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_state_dict_type: SHARDED_STATE_DICT
 special_tokens:

examples/llama-2/qlora.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 
 adapter: qlora
 lora_model_dir:

examples/llama-2/relora.yml

@@ -12,7 +12,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/relora-out
+output_dir: ./relora-out
 
 adapter: qlora
 lora_model_dir:

examples/llama-3/README.md

@@ -1,13 +0,0 @@
-# Llama-3
-
-https://llama.meta.com/llama3/
-
-[8B Base Model](https://huggingface.co/meta-llama/Meta-Llama-3-8B)
- - [Full Fine Tune](./fft-8b.yaml)
-   - Single GPU @ 48GB VRAM
- - [LoRA](./lora-8b.yml)
-   - Single GPU @ 11GB VRAM
-
-[70B Base Model](https://huggingface.co/meta-llama/Meta-Llama-3-70B)
- - [QLORA+FSDP](./qlora-fsdp-70b.yaml)
-   - Dual GPU @ 21GB VRAM

examples/llama-3/fft-8b.yaml

@@ -1,58 +0,0 @@
-base_model: meta-llama/Meta-Llama-3-8B
-model_type: LlamaForCausalLM
-tokenizer_type: AutoTokenizer
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.05
-output_dir: ./outputs/out
-
-sequence_len: 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: 1
-num_epochs: 1
-optimizer: paged_adamw_8bit
-lr_scheduler: cosine
-learning_rate: 2e-5
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 100
-evals_per_epoch: 2
-eval_table_size:
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-  pad_token: <|end_of_text|>

examples/llama-3/instruct-lora-8b.yml

@@ -1,76 +0,0 @@
-base_model: meta-llama/Meta-Llama-3-8B-Instruct
-model_type: LlamaForCausalLM
-tokenizer_type: AutoTokenizer
-
-load_in_8bit: true
-load_in_4bit: false
-strict: false
-
-chat_template: llama3
-datasets:
-  - path: fozziethebeat/alpaca_messages_2k_test
-    type: chat_template
-    chat_template: llama3
-    field_messages: messages
-    message_field_role: role
-    message_field_content: content
-    roles:
-      user:
-        - user
-      assistant:
-        - assistant
-
-dataset_prepared_path:
-val_set_size: 0.05
-output_dir: ./outputs/lora-out
-
-sequence_len: 4096
-sample_packing: false
-pad_to_sequence_len: true
-
-adapter: lora
-lora_model_dir:
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_fan_in_fan_out:
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 4
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-s2_attention:
-
-warmup_steps: 10
-evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:

examples/llama-3/lora-8b.yml

@@ -1,70 +0,0 @@
-base_model: meta-llama/Meta-Llama-3-8B
-model_type: LlamaForCausalLM
-tokenizer_type: AutoTokenizer
-
-load_in_8bit: true
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: mhenrichsen/alpaca_2k_test
-    type: alpaca
-dataset_prepared_path:
-val_set_size: 0.05
-output_dir: ./outputs/lora-out
-
-sequence_len: 4096
-sample_packing: true
-pad_to_sequence_len: true
-
-adapter: lora
-lora_model_dir:
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_fan_in_fan_out:
-lora_modules_to_save:
-  - embed_tokens
-  - lm_head
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 4
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-s2_attention:
-
-warmup_steps: 10
-evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-   pad_token: <|end_of_text|>

examples/llama-3/qlora-fsdp-70b.yaml

@@ -1,80 +0,0 @@
-base_model: casperhansen/llama-3-70b-fp16
-model_type: LlamaForCausalLM
-tokenizer_type: AutoTokenizer  # PreTrainedTokenizerFast
-
-load_in_8bit: false
-load_in_4bit: true
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.05
-output_dir: ./outputs/out/qlora-llama3-70b
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 512
-sample_packing: false
-pad_to_sequence_len: true
-
-lora_r: 8
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules:
-lora_target_linear: true
-lora_fan_in_fan_out:
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 1
-num_epochs: 4
-optimizer: adamw_torch
-lr_scheduler: cosine
-learning_rate: 0.00001
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 10
-evals_per_epoch: 4
-eval_table_size:
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-  - full_shard
-  - auto_wrap
-fsdp_config:
-  fsdp_limit_all_gathers: true
-  fsdp_sync_module_states: true
-  fsdp_offload_params: true
-  fsdp_use_orig_params: false
-  fsdp_cpu_ram_efficient_loading: true
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_sharding_strategy: FULL_SHARD
-special_tokens:
-  pad_token: <|end_of_text|>

examples/llama-3/qlora.yml

@@ -1,67 +0,0 @@
-base_model: meta-llama/Meta-Llama-3-8B
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
-
-load_in_8bit: false
-load_in_4bit: true
-strict: false
-
-datasets:
-  - path: aaditya/alpaca_subset_1
-    type: alpaca
-dataset_prepared_path:
-val_set_size: 0
-output_dir: ./outputs/qlora-out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 4096
-sample_packing: true
-pad_to_sequence_len: true
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules:
-lora_target_linear: true
-lora_fan_in_fan_out:
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 4
-optimizer: paged_adamw_32bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 10
-evals_per_epoch: 4
-eval_table_size:
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-  pad_token: "<|end_of_text|>"

examples/mamba/config.yml

@@ -12,7 +12,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.0
-output_dir: ./outputs/out
+output_dir: ./out
 
 sequence_len: 2048
 sample_packing: false

examples/mistral/bigstral-ds-zero3.yaml

@@ -1,63 +0,0 @@
-base_model: mistral-community/Mixtral-8x22B-v0.1
-model_type: AutoModelForCausalLM
-tokenizer_type: LlamaTokenizer
-trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-unfrozen_parameters:
-  - ^lm_head.weight$
-  - ^model.embed_tokens.weight$
-  - model.layers.4[4-9]+.block_sparse_moe.gate
-  - model.layers.4[4-9]+.block_sparse_moe.experts
-  - model.layers.5[0-5]+.block_sparse_moe.gate
-  - model.layers.5[0-5]+.block_sparse_moe.experts
-
-model_config:
-  output_router_logits: true
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.05
-output_dir: ./outputs/out
-
-sequence_len: 2048
-sample_packing: true
-pad_to_sequence_len: true
-
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 3
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0001
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-save_total_limit: 1
-save_steps:
-debug:
-deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_params.json
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-  eos_token: "<|im_end|>"
-tokens:
-  - "<|im_start|>"

examples/mistral/config.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/out
+output_dir: ./out
 
 sequence_len: 8192
 sample_packing: true

examples/mistral/lora-mps.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path: last_run_prepared
 val_set_size: 0
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 eval_sample_packing: false
 
 adapter: lora

examples/mistral/lora.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.1
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 
 adapter: lora
 lora_model_dir:

examples/mistral/mistral-qlora-fsdp.yml

@@ -1,82 +0,0 @@
-base_model: mistralai/Mixtral-8x7B-v0.1
-model_type: AutoModelForCausalLM
-tokenizer_type: LlamaTokenizer
-trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: true
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.02
-output_dir: ./outputs/qlora-out
-
-model_config:
-  output_router_logits: true
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 1024
-sample_packing: false
-pad_to_sequence_len: false
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_fan_in_fan_out:
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: paged_adamw_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-loss_watchdog_threshold: 5.0
-loss_watchdog_patience: 3
-
-warmup_steps: 10
-evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
-saves_per_epoch: 1
-debug:
-weight_decay: 0.0
-fsdp:
-  - full_shard
-  - auto_wrap
-fsdp_config:
-  fsdp_limit_all_gathers: true
-  fsdp_sync_module_states: true
-  fsdp_offload_params: false
-  fsdp_use_orig_params: false
-  fsdp_cpu_ram_efficient_loading: false
-  fsdp_transformer_layer_cls_to_wrap: MistralDecoderLayer
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-special_tokens:

examples/mistral/mistral-qlora-orpo.yml

@@ -1,82 +0,0 @@
-base_model: mistralai/Mistral-7B-v0.1
-model_type: MistralForCausalLM
-tokenizer_type: LlamaTokenizer
-
-load_in_8bit: false
-load_in_4bit: true
-strict: false
-
-rl: orpo
-orpo_alpha: 0.1
-remove_unused_columns: false
-
-chat_template: chatml
-datasets:
-  - path: argilla/ultrafeedback-binarized-preferences-cleaned
-    type: chat_template.argilla
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.1
-output_dir: ./outputs/mistral-qlora-orpo-out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 4096
-sample_packing: false
-pad_to_sequence_len: true
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_fan_in_fan_out:
-lora_target_modules:
-  - gate_proj
-  - down_proj
-  - up_proj
-  - q_proj
-  - v_proj
-  - k_proj
-  - o_proj
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-loss_watchdog_threshold: 5.0
-loss_watchdog_patience: 3
-
-warmup_steps: 10
-evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:

examples/mistral/mixtral-8x22b-qlora-fsdp.yml

@@ -1,81 +0,0 @@
-base_model: mistral-community/Mixtral-8x22B-v0.1
-model_type: AutoModelForCausalLM
-tokenizer_type: LlamaTokenizer
-
-load_in_8bit: false
-load_in_4bit: true
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.02
-output_dir: ./outputs/qlora-out
-
-model_config:
-  output_router_logits: true
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 1024
-sample_packing: false
-pad_to_sequence_len: false
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_fan_in_fan_out:
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_torch
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: true
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-loss_watchdog_threshold: 5.0
-loss_watchdog_patience: 3
-
-warmup_steps: 10
-evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
-saves_per_epoch: 1
-debug:
-weight_decay: 0.0
-fsdp:
-  - full_shard
-  - auto_wrap
-fsdp_config:
-  fsdp_limit_all_gathers: true
-  fsdp_sync_module_states: true
-  fsdp_offload_params: true
-  fsdp_use_orig_params: false
-  fsdp_cpu_ram_efficient_loading: true
-  fsdp_transformer_layer_cls_to_wrap: MixtralSparseMoeBlock
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-special_tokens:

examples/mistral/mixtral-qlora-fsdp.yml

@@ -12,7 +12,7 @@ datasets:
     type: alpaca
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.02
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 
 model_config:
   output_router_logits: true
@@ -39,7 +39,7 @@ wandb_log_model:
 gradient_accumulation_steps: 4
 micro_batch_size: 2
 num_epochs: 1
-optimizer: adamw_torch
+optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
@@ -47,7 +47,7 @@ train_on_inputs: false
 group_by_length: false
 bf16: auto
 fp16:
-tf32: true
+tf32: false
 
 gradient_checkpointing: true
 early_stopping_patience:
@@ -69,17 +69,6 @@ debug:
 weight_decay: 0.0
 fsdp:
   - full_shard
-  - auto_wrap
 fsdp_config:
-  fsdp_limit_all_gathers: true
-  fsdp_sync_module_states: true
-  fsdp_offload_params: true
-  fsdp_use_orig_params: false
-  fsdp_cpu_ram_efficient_loading: true
   fsdp_transformer_layer_cls_to_wrap: MixtralSparseMoeBlock
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  fsdp_sharding_strategy: FULL_SHARD
-  fsdp_forward_prefetch: false
-  fsdp_backward_prefetch: BACKWARD_PRE
 special_tokens:

examples/mistral/mixtral.yml

@@ -12,7 +12,7 @@ datasets:
     type: alpaca
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 
 ## You can optionally freeze the entire model and unfreeze a subset of parameters
 unfrozen_parameters:

examples/mistral/mixtral_22.yml

@@ -1,61 +0,0 @@
-base_model: mistral-community/Mixtral-8x22B-v0.1
-model_type: AutoModelForCausalLM
-tokenizer_type: LlamaTokenizer
-trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-unfrozen_parameters:
-  - ^lm_head.weight$
-  - ^model.embed_tokens.weight$
-  - model.layers.4[4-9]+.block_sparse_moe.gate
-  - model.layers.4[4-9]+.block_sparse_moe.experts
-  - model.layers.5[0-5]+.block_sparse_moe.gate
-  - model.layers.5[0-5]+.block_sparse_moe.experts
-
-model_config:
-  output_router_logits: true
-
-datasets:
-  - path: yahma/alpaca-cleaned
-    type: alpaca
-output_dir: ./outputs/out
-
-sequence_len: 8000
-sample_packing: true
-pad_to_sequence_len: true
-
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 3
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0001
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-save_total_limit: 1
-save_steps:
-debug:
-deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_all.json
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-  eos_token: "<|im_end|>"
-tokens:
-  - "<|im_start|>"

examples/mistral/qlora.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.1
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 
 adapter: qlora
 lora_model_dir:

examples/mpt-7b/config.yml

@@ -23,7 +23,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/mpt-alpaca-7b
+output_dir: ./mpt-alpaca-7b
 gradient_accumulation_steps: 1
 micro_batch_size: 1
 num_epochs: 4

examples/openllama-3b/config.yml

@@ -25,7 +25,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/openllama-out
+output_dir: ./openllama-out
 gradient_accumulation_steps: 1
 micro_batch_size: 1
 num_epochs: 4

examples/openllama-3b/lora.yml

@@ -31,7 +31,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 gradient_accumulation_steps: 1
 micro_batch_size: 2
 num_epochs: 4

examples/openllama-3b/qlora.yml

@@ -25,7 +25,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 gradient_accumulation_steps: 1
 micro_batch_size: 2
 num_epochs: 4

examples/phi/phi-ft.yml

@@ -12,7 +12,7 @@ datasets:
 
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/phi-sft-out
+output_dir: ./phi-sft-out
 
 sequence_len: 2048
 sample_packing: true

examples/phi/phi-qlora.yml

@@ -12,7 +12,7 @@ datasets:
 
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/phi-sft-out
+output_dir: ./phi-sft-out
 
 sequence_len: 2048
 sample_packing: true

examples/phi/phi2-ft.yml

@@ -12,7 +12,7 @@ datasets:
 
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/phi-sft-out
+output_dir: ./phi-sft-out
 
 sequence_len: 2048
 sample_packing: true

examples/phi/phi3-ft-fsdp.yml

@@ -1,83 +0,0 @@
-base_model: microsoft/Phi-3-mini-4k-instruct
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: mhenrichsen/alpaca_2k_test
-    type: alpaca
-
-dataset_prepared_path:
-val_set_size: 0
-output_dir: ./phi-sft-out
-
-sequence_len: 4096
-sample_packing: true
-pad_to_sequence_len: true
-trust_remote_code: true
-
-adapter:
-lora_model_dir:
-lora_r:
-lora_alpha:
-lora_dropout:
-lora_target_linear:
-lora_fan_in_fan_out:
-
-wandb_project: phi3
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 2
-micro_batch_size: 12
-num_epochs: 2
-optimizer: adamw_torch
-adam_beta2: 0.95
-adam_epsilon: 0.00001
-max_grad_norm: 1.0
-lr_scheduler: cosine
-learning_rate: 0.000003
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: true
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 100
-evals_per_epoch: 4
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.1
-fsdp:
-  - full_shard
-  - auto_wrap
-fsdp_config:
-  fsdp_limit_all_gathers: true
-  fsdp_sync_module_states: true
-  fsdp_offload_params: true
-  fsdp_use_orig_params: false
-  fsdp_cpu_ram_efficient_loading: true
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  fsdp_transformer_layer_cls_to_wrap: Phi3DecoderLayer
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_sharding_strategy: FULL_SHARD
-resize_token_embeddings_to_32x: true
-special_tokens:
-  pad_token: "<|endoftext|>"

examples/phi/phi3-ft.yml

@@ -1,64 +0,0 @@
-base_model: microsoft/Phi-3-mini-4k-instruct
-trust_remote_code: true
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
-chat_template: phi_3
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: garage-bAInd/Open-Platypus
-    type: alpaca:phi
-
-dataset_prepared_path:
-val_set_size: 0.01
-output_dir: ./out
-
-sequence_len: 4096
-sample_packing: true
-pad_to_sequence_len: true
-
-adapter: lora
-lora_model_dir:
-lora_r: 64
-lora_alpha: 32
-lora_dropout: 0.05
-lora_target_linear: true
-lora_fan_in_fan_out:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_torch
-adam_beta2: 0.95
-adam_epsilon: 0.00001
-max_grad_norm: 1.0
-lr_scheduler: cosine
-learning_rate: 5.0e-6
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: True
-early_stopping_patience: 3
-logging_steps: 1
-flash_attention: true
-
-eval_steps: 1000
-save_steps: 5000
-eval_table_size: 2
-eval_batch_size: 2
-eval_sample_packing: false
-eval_max_new_tokens: 32
-eval_causal_lm_metrics: ["perplexity"]
-do_causal_lm_eval: true
-
-warmup_ratio: 0.2
-debug: true
-weight_decay: 0.1
-resize_token_embeddings_to_32x: true

examples/pythia-12b/config.yml

@@ -26,7 +26,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/pythia-12b
+output_dir: ./pythia-12b
 gradient_accumulation_steps: 1
 micro_batch_size: 1
 num_epochs: 5

examples/pythia/lora.yml

@@ -20,7 +20,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/lora-alpaca-pythia
+output_dir: ./lora-alpaca-pythia
 gradient_accumulation_steps: 1
 micro_batch_size: 4
 num_epochs: 4

examples/qwen/lora.yml

@@ -13,7 +13,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 
 sequence_len: 2048  # supports up to 8192
 sample_packing: false

examples/qwen/qlora.yml

@@ -13,7 +13,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 
 sequence_len: 2048  # supports up to 8192
 sample_packing: false

examples/qwen/qwen2-moe-lora.yaml

@@ -10,7 +10,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/out
+output_dir: ./out
 
 sequence_len: 1024  # supports up to 32k
 sample_packing: false

examples/qwen/qwen2-moe-qlora.yaml

@@ -10,7 +10,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/out
+output_dir: ./out
 
 sequence_len: 1024  # supports up to 32k
 sample_packing: false

examples/qwen2/qlora-fsdp.yaml

@@ -1,75 +0,0 @@
-base_model: Qwen/Qwen2-7B
-trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: true
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-dataset_prepared_path:
-val_set_size: 0.05
-output_dir: ./outputs/out
-
-sequence_len: 2048
-sample_packing: true
-eval_sample_packing: true
-pad_to_sequence_len: true
-
-adapter: qlora
-lora_model_dir:
-lora_r: 32
-lora_alpha: 64
-lora_dropout: 0.05
-lora_target_linear: true
-lora_fan_in_fan_out:
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 1
-num_epochs: 4
-optimizer: adamw_torch
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: true
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 10
-evals_per_epoch: 4
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-  - full_shard
-  - auto_wrap
-fsdp_config:
-  fsdp_limit_all_gathers: true
-  fsdp_sync_module_states: true
-  fsdp_offload_params: true
-  fsdp_use_orig_params: false
-  fsdp_cpu_ram_efficient_loading: true
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
-  fsdp_state_dict_type: FULL_STATE_DICT
-special_tokens:

examples/redpajama/config-3b.yml

@@ -24,7 +24,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/redpajama-alpaca-3b
+output_dir: ./redpajama-alpaca-3b
 batch_size: 4
 micro_batch_size: 1
 num_epochs: 4

examples/replit-3b/config-lora.yml

@@ -23,7 +23,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/lora-replit
+output_dir: ./lora-replit
 batch_size: 8
 micro_batch_size: 1
 num_epochs: 4

examples/stablelm-2/1.6b/fft.yml

@@ -12,7 +12,7 @@ datasets:
     type: alpaca
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.05
-output_dir: ./outputs/out
+output_dir: ./out
 
 sequence_len: 4096
 sample_packing: true

examples/stablelm-2/1.6b/lora.yml

@@ -12,7 +12,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 
 sequence_len: 4096
 sample_packing: true

examples/starcoder2/qlora.yml

@@ -11,7 +11,7 @@ datasets:
 
 dataset_prepared_path:
 val_set_size: 0.2
-output_dir: ./outputs/qlora
+output_dir: ./qlora
 
 adapter: qlora
 lora_model_dir:

examples/tiny-llama/lora-mps.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 
 sequence_len: 4096
 sample_packing: true

examples/tiny-llama/lora.yml

@@ -11,7 +11,7 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/lora-out
+output_dir: ./lora-out
 
 sequence_len: 4096
 sample_packing: true

examples/tiny-llama/pretrain.yml

@@ -14,7 +14,7 @@ pretraining_dataset:
   type: pretrain
 dataset_prepared_path:
 val_set_size: 0.0
-output_dir: ./outputs/model-out
+output_dir: ./model-out
 
 sequence_len: 2048
 sample_packing: true

examples/tiny-llama/qlora.yml

@@ -11,14 +11,13 @@ datasets:
     type: alpaca
 dataset_prepared_path:
 val_set_size: 0.05
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 
 adapter: qlora
 lora_model_dir:
 
 sequence_len: 4096
 sample_packing: true
-eval_sample_packing: false
 pad_to_sequence_len: true
 
 lora_r: 32

examples/xgen-7b/xgen-7b-8k-qlora.yml

@@ -40,7 +40,7 @@ wandb_entity:
 wandb_watch:
 wandb_name:
 wandb_log_model:
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 
 # QLoRA paper Table 9
 # - 16 for 7b & 13b

examples/yi-34B-chat/qlora.yml

@@ -33,7 +33,7 @@ eval_sample_packing: false
 eval_batch_size: 1
 
 # LoRA
-output_dir: ./outputs/qlora-out
+output_dir: ./qlora-out
 adapter: qlora
 lora_model_dir:
 lora_r: 32

requirements.txt

@@ -1,22 +1,22 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
-peft==0.11.1
-transformers==4.41.1
-tokenizers==0.19.1
-bitsandbytes==0.43.1
-accelerate==0.30.1
-deepspeed @ git+https://github.com/microsoft/DeepSpeed.git@bc48371c5e1fb8fd70fc79285e66201dbb65679b
+peft==0.10.0
+transformers @ git+https://github.com/huggingface/transformers.git@43d17c18360ac9c3d3491389328e2fe55fe8f9ce
+tokenizers==0.15.0
+bitsandbytes==0.43.0
+accelerate==0.28.0
+deepspeed==0.13.1
 pydantic==2.6.3
 addict
 fire
 PyYAML>=6.0
 requests
-datasets==2.19.1
-flash-attn==2.5.8
+datasets>=2.15.0
+flash-attn==2.5.5
 sentencepiece
 wandb
 einops
-xformers==0.0.26.post1
+xformers==0.0.22
 optimum==1.16.2
 hf_transfer
 colorama
@@ -28,7 +28,7 @@ scipy
 scikit-learn==1.2.2
 pynvml
 art
-fschat @ git+https://github.com/lm-sys/FastChat.git@27a05b04a35510afb1d767ae7e5990cbd278f8fe
+fschat==0.2.36
 gradio==3.50.2
 tensorboard
 
@@ -39,6 +39,5 @@ s3fs
 gcsfs
 # adlfs
 
-trl @ git+https://github.com/huggingface/trl.git@f18253bf2d747f68acc9cd89da95c85ebf59dbb9
+trl @ git+https://github.com/huggingface/trl.git@0ee349dcd43b0f4b3169449f16751c38ac4a609f
 zstandard==0.22.0
-fastcore

scripts/cloud-entrypoint-term.sh

@@ -1,82 +0,0 @@
-#!/bin/bash
-
-# Export specific ENV variables to /etc/rp_environment
-echo "Exporting environment variables..."
-printenv | grep -E '^RUNPOD_|^PATH=|^_=' | sed 's/^\(.*\)=\(.*\)$/export \1="\2"/' >> /etc/rp_environment
-conda init
-# this needs to come after conda init
-echo 'source /etc/rp_environment' >> ~/.bashrc
-
-add_keys_to_authorized() {
-    local key_value=$1
-
-    # Create the ~/.ssh directory and set permissions
-    mkdir -p ~/.ssh
-    chmod 700 ~/.ssh
-
-    # Create the authorized_keys file if it doesn't exist
-    touch ~/.ssh/authorized_keys
-
-    # Initialize an empty key variable
-    local key=""
-
-    # Read the key variable word by word
-    for word in $key_value; do
-        # Check if the word looks like the start of a key
-        if [[ $word == ssh-* ]]; then
-            # If there's a key being built, add it to the authorized_keys file
-            if [[ -n $key ]]; then
-                echo $key >> ~/.ssh/authorized_keys
-            fi
-            # Start a new key
-            key=$word
-        else
-            # Append the word to the current key
-            key="$key $word"
-        fi
-    done
-
-    # Add the last key to the authorized_keys file
-    if [[ -n $key ]]; then
-        echo $key >> ~/.ssh/authorized_keys
-    fi
-
-    # Set the correct permissions
-    chmod 600 ~/.ssh/authorized_keys
-    chmod 700 -R ~/.ssh
-}
-
-if [[ $PUBLIC_KEY ]]; then
-    # runpod
-    add_keys_to_authorized "$PUBLIC_KEY"
-    # Start the SSH service in the background
-    service ssh start
-elif [[ $SSH_KEY ]]; then
-    # latitude.sh
-    add_keys_to_authorized "$SSH_KEY"
-    # Start the SSH service in the background
-    service ssh start
-else
-    echo "No PUBLIC_KEY or SSH_KEY environment variable provided, not starting openSSH daemon"
-fi
-
-# Check if JUPYTER_PASSWORD is set and not empty
-if [ -n "$JUPYTER_PASSWORD" ]; then
-    # Set JUPYTER_TOKEN to the value of JUPYTER_PASSWORD
-    export JUPYTER_TOKEN="$JUPYTER_PASSWORD"
-fi
-
-if [ "$JUPYTER_DISABLE" != "1" ]; then
-    # Run Jupyter Lab in the background
-    jupyter lab --port=8888 --ip=* --allow-root --ServerApp.allow_origin=* &
-fi
-
-if [ ! -d "/workspace/data/axolotl-artifacts" ]; then
-    mkdir -p /workspace/data/axolotl-artifacts
-fi
-if [ ! -L "/workspace/axolotl/outputs" ]; then
-    ln -sf /workspace/data/axolotl-artifacts /workspace/axolotl/outputs
-fi
-
-# Execute the passed arguments (CMD)
-exec "$@"

scripts/cloud-entrypoint.sh

@@ -5,53 +5,20 @@ echo "Exporting environment variables..."
 printenv | grep -E '^RUNPOD_|^PATH=|^_=' | sed 's/^\(.*\)=\(.*\)$/export \1="\2"/' >> /etc/rp_environment
 echo 'source /etc/rp_environment' >> ~/.bashrc
 
-add_keys_to_authorized() {
-    local key_value=$1
-
-    # Create the ~/.ssh directory and set permissions
-    mkdir -p ~/.ssh
-    chmod 700 ~/.ssh
-
-    # Create the authorized_keys file if it doesn't exist
-    touch ~/.ssh/authorized_keys
-
-    # Initialize an empty key variable
-    local key=""
-
-    # Read the key variable word by word
-    for word in $key_value; do
-        # Check if the word looks like the start of a key
-        if [[ $word == ssh-* ]]; then
-            # If there's a key being built, add it to the authorized_keys file
-            if [[ -n $key ]]; then
-                echo $key >> ~/.ssh/authorized_keys
-            fi
-            # Start a new key
-            key=$word
-        else
-            # Append the word to the current key
-            key="$key $word"
-        fi
-    done
-
-    # Add the last key to the authorized_keys file
-    if [[ -n $key ]]; then
-        echo $key >> ~/.ssh/authorized_keys
-    fi
-
-    # Set the correct permissions
-    chmod 600 ~/.ssh/authorized_keys
-    chmod 700 -R ~/.ssh
-}
-
 if [[ $PUBLIC_KEY ]]; then
     # runpod
-    add_keys_to_authorized "$PUBLIC_KEY"
+    mkdir -p ~/.ssh
+    chmod 700 ~/.ssh
+    echo $PUBLIC_KEY >> ~/.ssh/authorized_keys
+    chmod 700 -R ~/.ssh
     # Start the SSH service in the background
     service ssh start
-elif [[ $SSH_KEY ]]; then
+elif [ -n "$SSH_KEY" ]; then
     # latitude.sh
-    add_keys_to_authorized "$SSH_KEY"
+    mkdir -p ~/.ssh
+    chmod 700 ~/.ssh
+    echo $SSH_KEY >> ~/.ssh/authorized_keys
+    chmod 700 -R ~/.ssh
     # Start the SSH service in the background
     service ssh start
 else
@@ -66,14 +33,7 @@ fi
 
 if [ "$JUPYTER_DISABLE" != "1" ]; then
     # Run Jupyter Lab in the background
-    jupyter lab --port=8888 --ip=* --allow-root --ServerApp.allow_origin=* &
-fi
-
-if [ ! -d "/workspace/data/axolotl-artifacts" ]; then
-    mkdir -p /workspace/data/axolotl-artifacts
-fi
-if [ ! -L "/workspace/axolotl/outputs" ]; then
-    ln -sf /workspace/data/axolotl-artifacts /workspace/axolotl/outputs
+    jupyter lab --port=8888 --ip=* --allow-root --ServerApp.allow_origin=* --ServerApp.preferred_dir=/workspace &
 fi
 
 # Execute the passed arguments (CMD)

setup.py

@@ -30,11 +30,8 @@ def parse_requirements():
 
     try:
         if "Darwin" in platform.system():
-            # don't install xformers on MacOS
-            _install_requires.pop(_install_requires.index("xformers==0.0.26.post1"))
+            _install_requires.pop(_install_requires.index("xformers==0.0.22"))
         else:
-            # detect the version of torch already installed
-            # and set it so dependencies don't clobber the torch version
             torch_version = version("torch")
             _install_requires.append(f"torch=={torch_version}")
 
@@ -48,15 +45,9 @@ def parse_requirements():
             else:
                 raise ValueError("Invalid version format")
 
-            if (major, minor) >= (2, 3):
-                pass
-            elif (major, minor) >= (2, 2):
-                _install_requires.pop(_install_requires.index("xformers==0.0.26.post1"))
-                _install_requires.append("xformers>=0.0.25.post1")
-            else:
-                _install_requires.pop(_install_requires.index("xformers==0.0.26.post1"))
-                _install_requires.append("xformers>=0.0.23.post1")
-
+            if (major, minor) >= (2, 1):
+                _install_requires.pop(_install_requires.index("xformers==0.0.22"))
+                _install_requires.append("xformers>=0.0.23")
     except PackageNotFoundError:
         pass
 
@@ -68,7 +59,7 @@ install_requires, dependency_links = parse_requirements()
 
 setup(
     name="axolotl",
-    version="0.4.1",
+    version="0.4.0",
     description="LLM Trainer",
     long_description="Axolotl is a tool designed to streamline the fine-tuning of various AI models, offering support for multiple configurations and architectures.",
     package_dir={"": "src"},
@@ -77,13 +68,13 @@ setup(
     dependency_links=dependency_links,
     extras_require={
         "flash-attn": [
-            "flash-attn==2.5.8",
+            "flash-attn==2.5.5",
         ],
         "fused-dense-lib": [
-            "fused-dense-lib  @ git+https://github.com/Dao-AILab/flash-attention@v2.5.8#subdirectory=csrc/fused_dense_lib",
+            "fused-dense-lib  @ git+https://github.com/Dao-AILab/flash-attention@v2.3.3#subdirectory=csrc/fused_dense_lib",
         ],
         "deepspeed": [
-            "deepspeed @ git+https://github.com/microsoft/DeepSpeed.git@bc48371c5e1fb8fd70fc79285e66201dbb65679b",
+            "deepspeed==0.13.1",
             "deepspeed-kernels",
         ],
         "mamba-ssm": [

src/axolotl/cli/__init__.py

@@ -264,8 +264,8 @@ def do_inference_gradio(
         with torch.no_grad():
             generation_config = GenerationConfig(
                 repetition_penalty=1.1,
-                max_new_tokens=cfg.get("gradio_max_new_tokens", 1024),
-                temperature=cfg.get("gradio_temperature", 0.9),
+                max_new_tokens=1024,
+                temperature=0.9,
                 top_p=0.95,
                 top_k=40,
                 bos_token_id=tokenizer.bos_token_id,
@@ -300,13 +300,7 @@ def do_inference_gradio(
         outputs="text",
         title=cfg.get("gradio_title", "Axolotl Gradio Interface"),
     )
-
-    demo.queue().launch(
-        show_api=False,
-        share=cfg.get("gradio_share", True),
-        server_name=cfg.get("gradio_server_name", "127.0.0.1"),
-        server_port=cfg.get("gradio_server_port", None),
-    )
+    demo.queue().launch(show_api=False, share=True)
 
 
 def choose_config(path: Path):
@@ -439,23 +433,6 @@ def load_rl_datasets(
         math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
     )
 
-    if cli_args.debug or cfg.debug:
-        LOG.info("check_dataset_labels...")
-
-        tokenizer = load_tokenizer(cfg)
-        check_dataset_labels(
-            train_dataset.select(
-                [
-                    random.randrange(0, len(train_dataset) - 1)  # nosec
-                    for _ in range(cli_args.debug_num_examples)
-                ]
-            ),
-            tokenizer,
-            num_examples=cli_args.debug_num_examples,
-            text_only=cli_args.debug_text_only,
-            rl_mode=True,
-        )
-
     return TrainDatasetMeta(
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,

src/axolotl/cli/merge_lora.py

@@ -8,6 +8,7 @@ import transformers
 
 from axolotl.cli import do_merge_lora, load_cfg, print_axolotl_text_art
 from axolotl.common.cli import TrainerCliArgs
+from axolotl.utils.dict import DictDefault
 
 
 def do_cli(config: Path = Path("examples/"), **kwargs):
@@ -25,26 +26,28 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
         load_in_8bit=False,
         load_in_4bit=False,
         flash_attention=False,
-        deepspeed=None,
-        fsdp=None,
         **kwargs,
     )
+    cfg = modify_cfg_for_merge(parsed_cfg)
 
-    if not parsed_cfg.lora_model_dir and parsed_cfg.output_dir:
-        parsed_cfg.lora_model_dir = parsed_cfg.output_dir
-    if not Path(parsed_cfg.lora_model_dir).exists():
+    do_merge_lora(cfg=cfg, cli_args=parsed_cli_args)
+
+
+def modify_cfg_for_merge(cfg: DictDefault) -> DictDefault:
+    if not cfg.lora_model_dir and cfg.output_dir:
+        cfg.lora_model_dir = cfg.output_dir
+    if not Path(cfg.lora_model_dir).exists():
         raise ValueError(
-            f"Target directory for merge: `{parsed_cfg.lora_model_dir}` does not exist."
+            f"Target directory for merge: `{cfg.lora_model_dir}` does not exist."
         )
 
-    parsed_cfg.load_in_4bit = False
-    parsed_cfg.load_in_8bit = False
-    parsed_cfg.flash_attention = False
-    parsed_cfg.deepspeed = None
-    parsed_cfg.fsdp = None
-    parsed_cfg.fsdp_config = None
+    cfg.load_in_4bit = False
+    cfg.load_in_8bit = False
+    cfg.flash_attention = False
+    cfg.deepspeed = None
+    cfg.fsdp = None
 
-    do_merge_lora(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    return cfg
 
 
 if __name__ == "__main__":

src/axolotl/cli/preprocess.py

@@ -7,9 +7,7 @@ from typing import Union
 
 import fire
 import transformers
-from accelerate import init_empty_weights
 from colorama import Fore
-from transformers import AutoModelForCausalLM
 
 from axolotl.cli import (
     check_accelerate_default_config,
@@ -21,10 +19,7 @@ from axolotl.cli import (
 )
 from axolotl.common.cli import PreprocessCliArgs
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
-from axolotl.prompt_strategies.sharegpt import (
-    register_chatml_template,
-    register_llama3_template,
-)
+from axolotl.prompt_strategies.sharegpt import register_chatml_template
 
 LOG = logging.getLogger("axolotl.cli.preprocess")
 
@@ -41,22 +36,13 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
         return_remaining_strings=True
     )
 
-    if parsed_cfg.chat_template == "chatml":
-        if parsed_cfg.default_system_message:
-            LOG.info(
-                f"ChatML set. Adding default system message: {parsed_cfg.default_system_message}"
-            )
-            register_chatml_template(parsed_cfg.default_system_message)
-        else:
-            register_chatml_template()
-    elif parsed_cfg.chat_template == "llama3":
-        if parsed_cfg.default_system_message:
-            LOG.info(
-                f"LLaMA-3 set. Adding default system message: {parsed_cfg.default_system_message}"
-            )
-            register_llama3_template(parsed_cfg.default_system_message)
-        else:
-            register_llama3_template()
+    if parsed_cfg.chat_template == "chatml" and parsed_cfg.default_system_message:
+        LOG.info(
+            f"ChatML set. Adding default system message: {parsed_cfg.default_system_message}"
+        )
+        register_chatml_template(parsed_cfg.default_system_message)
+    else:
+        register_chatml_template()
 
     if not parsed_cfg.dataset_prepared_path:
         msg = (
@@ -68,16 +54,11 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
         LOG.warning(msg)
         parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
-    if parsed_cfg.rl:  # and parsed_cfg.rl != "orpo":
+    if parsed_cfg.rl and parsed_cfg.rl != "orpo":
         load_rl_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
     else:
         load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
 
-    if parsed_cli_args.download:
-        model_name = parsed_cfg.base_model
-        with init_empty_weights():
-            AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
-
     LOG.info(
         Fore.GREEN
         + f"Success! Preprocessed data path: `dataset_prepared_path: {parsed_cfg.dataset_prepared_path}`"

src/axolotl/cli/train.py

@@ -19,10 +19,7 @@ from axolotl.cli import (
     print_axolotl_text_art,
 )
 from axolotl.common.cli import TrainerCliArgs
-from axolotl.prompt_strategies.sharegpt import (
-    register_chatml_template,
-    register_llama3_template,
-)
+from axolotl.prompt_strategies.sharegpt import register_chatml_template
 from axolotl.train import train
 
 LOG = logging.getLogger("axolotl.cli.train")
@@ -50,15 +47,7 @@ def do_train(cfg, cli_args) -> Tuple[PreTrainedModel, PreTrainedTokenizer]:
     else:
         register_chatml_template()
 
-    if cfg.chat_template == "llama3" and cfg.default_system_message:
-        LOG.info(
-            f"LLaMA-3 set. Adding default system message: {cfg.default_system_message}"
-        )
-        register_llama3_template(cfg.default_system_message)
-    else:
-        register_llama3_template()
-
-    if cfg.rl:  # and cfg.rl != "orpo":
+    if cfg.rl and cfg.rl != "orpo":
         dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)
     else:
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

src/axolotl/common/cli.py

@@ -40,7 +40,6 @@ class PreprocessCliArgs:
     debug_text_only: bool = field(default=False)
     debug_num_examples: int = field(default=1)
     prompter: Optional[str] = field(default=None)
-    download: Optional[bool] = field(default=True)
 
 
 def load_model_and_tokenizer(

src/axolotl/core/trainer_builder.py

@@ -30,7 +30,7 @@ from transformers import (
 )
 from transformers.trainer_utils import seed_worker
 from transformers.utils import is_sagemaker_mp_enabled
-from trl import DPOConfig, DPOTrainer, KTOConfig, KTOTrainer, ORPOConfig, ORPOTrainer
+from trl import DPOTrainer
 from trl.trainer.utils import pad_to_length
 
 from axolotl.loraplus import create_loraplus_optimizer
@@ -43,7 +43,6 @@ from axolotl.utils.callbacks import (
     LossWatchDogCallback,
     SaveAxolotlConfigtoWandBCallback,
     SaveBetterTransformerModelCallback,
-    SaveModelCallback,
     bench_eval_callback_factory,
     causal_lm_bench_eval_callback_factory,
     log_prediction_callback_factory,
@@ -55,7 +54,6 @@ from axolotl.utils.collators import (
     MambaDataCollator,
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
-from axolotl.utils.models import ensure_dtype
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 from axolotl.utils.schedulers import (
     get_cosine_schedule_with_min_lr,
@@ -91,12 +89,11 @@ def _sanitize_kwargs_for_tagging(tag_names, kwargs=None):
 
 
 @dataclass
-class AxolotlTrainingMixins:
+class AxolotlTrainingArguments(TrainingArguments):
     """
-    Mixin class for the Axolotl training args.
+    Extend the base TrainingArguments for axolotl helpers
     """
 
-    # pylint: disable=duplicate-code
     model_type: Optional[str] = field(
         default=None, metadata={"help": "HF model configuration model_type."}
     )
@@ -126,22 +123,14 @@ class AxolotlTrainingMixins:
         default=1.0,
         metadata={"help": "Sample packing efficiency for calculating batch length."},
     )
-    sample_packing_bin_size: int = field(
-        default=200,
-        metadata={
-            "help": "The max number of samples that packed sample can contain after packing. Increase for better packing."
-        },
-    )
-    sample_packing_group_size: int = field(
-        default=100000,
-        metadata={
-            "help": "The number of samples to group together for packing. Increase for better packing."
-        },
-    )
     max_seq_length: int = field(
         default=2048,
         metadata={"help": "The maximum sequence length the model can handle"},
     )
+    sample_packing_seq_len_multiplier: int = field(
+        default=1,
+        metadata={"help": "the multiplier for the max len for packed sequences"},
+    )
     relora_steps: Optional[int] = field(
         default=None,
         metadata={"help": "how often to reset for ReLoRA"},
@@ -222,41 +211,6 @@ class AxolotlTrainingMixins:
         default=None,
         metadata={"help": "path under the model to access the layers"},
     )
-    curriculum_sampling: Optional[bool] = field(
-        default=None,
-        metadata={"help": "whether to use sequential sampling for curriculum learning"},
-    )
-
-
-@dataclass
-class AxolotlTrainingArguments(AxolotlTrainingMixins, TrainingArguments):
-    """
-    Training arguments for Causal trainer
-
-    This code is duplicated due to HF TrainingArguments not setting output_dir with a defaujlt value
-    so it can't be used as a mixin.
-    """
-
-
-@dataclass
-class AxolotlDPOConfig(AxolotlTrainingMixins, DPOConfig):
-    """
-    DPO config for DPO training
-    """
-
-
-@dataclass
-class AxolotlORPOConfig(AxolotlTrainingMixins, ORPOConfig):
-    """
-    ORPO config for ORPO training
-    """
-
-
-@dataclass
-class AxolotlKTOConfig(AxolotlTrainingMixins, KTOConfig):
-    """
-    KTO config for KTO training
-    """
 
 
 class AxolotlTrainer(Trainer):
@@ -386,16 +340,12 @@ class AxolotlTrainer(Trainer):
                 )
             return MultipackBatchSampler(
                 RandomSampler(self.train_dataset),
+                batch_size=batch_size,
+                drop_last=True,
+                batch_max_len=batch_max_len,
                 lengths=get_dataset_lengths(self.train_dataset),
                 packing_efficiency_estimate=self.args.sample_packing_efficiency,
-                batch_max_len=batch_max_len,
-                batch_size=batch_size,
-                group_size=self.args.sample_packing_group_size,
-                bin_size=self.args.sample_packing_bin_size,
-                drop_last=True,
             )
-        if self.args.curriculum_sampling:
-            return SequentialSampler(self.train_dataset)
         return super()._get_train_sampler()
 
     def _get_eval_sampler(
@@ -412,13 +362,11 @@ class AxolotlTrainer(Trainer):
                 )
             return MultipackBatchSampler(
                 SequentialSampler(eval_dataset),
-                lengths=get_dataset_lengths(self.eval_dataset),
-                packing_efficiency_estimate=self.args.sample_packing_efficiency,
-                batch_max_len=batch_max_len,
                 batch_size=batch_size,
-                group_size=self.args.sample_packing_group_size,
-                bin_size=self.args.sample_packing_bin_size,
                 drop_last=True,
+                batch_max_len=batch_max_len,
+                lengths=get_dataset_lengths(eval_dataset),
+                packing_efficiency_estimate=self.args.sample_packing_efficiency,
             )
         return super()._get_eval_sampler(eval_dataset)
 
@@ -459,8 +407,6 @@ class AxolotlTrainer(Trainer):
             self.data_collator = (  # pylint: disable=attribute-defined-outside-init
                 self.eval_data_collator
             )
-            if eval_dataset:
-                eval_dataset = eval_dataset.remove_columns(["length"])
             dataloader = super().get_eval_dataloader(eval_dataset)
             self.data_collator = (  # pylint: disable=attribute-defined-outside-init
                 self.train_data_collator
@@ -844,40 +790,6 @@ class AxolotlDPOTrainer(DPOTrainer):
 
     tag_names = ["axolotl", "dpo"]
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.optimizer = None
-
-    def create_optimizer(self):
-        if self.args.loraplus_lr_ratio is None:
-            return super().create_optimizer()
-
-        opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
-        if self.optimizer is None:  # pylint: disable=access-member-before-definition
-            optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(
-                self.args,
-                opt_model,
-            )
-
-            loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
-            if loraplus_lr_ratio:
-                print("Using lora+")
-            loraplus_lr_embedding = getattr(self.args, "loraplus_lr_embedding", None)
-            self.optimizer = create_loraplus_optimizer(  # pylint: disable=attribute-defined-outside-init
-                opt_model,
-                optimizer_cls,
-                optimizer_kwargs,
-                loraplus_lr_ratio,
-                loraplus_lr_embedding,
-            )
-
-        if is_sagemaker_mp_enabled():
-            self.optimizer = smp.DistributedOptimizer(  # pylint: disable=attribute-defined-outside-init
-                self.optimizer
-            )
-
-        return self.optimizer
-
     @wraps(DPOTrainer.push_to_hub)
     def push_to_hub(self, *args, **kwargs) -> str:
         """
@@ -898,22 +810,6 @@ class AxolotlDPOTrainer(DPOTrainer):
         return res
 
 
-class AxolotlORPOTrainer(ORPOTrainer):
-    """
-    Extend the base ORPOTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "orpo"]
-
-
-class AxolotlKTOTrainer(KTOTrainer):
-    """
-    Extend the base KTOTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "kto"]
-
-
 class TrainerBuilderBase(abc.ABC):
     """
     Base class for trainer builder
@@ -977,14 +873,6 @@ class TrainerBuilderBase(abc.ABC):
             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.append(
-                SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path)
-            )
 
         return callbacks
 
@@ -1030,11 +918,22 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         ):
             callbacks.append(SaveBetterTransformerModelCallback())
 
+        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.append(
+                SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path)
+            )
+
         if self.cfg.loss_watchdog_threshold is not None:
             callbacks.append(LossWatchDogCallback(self.cfg))
 
-        callbacks.append(SaveModelCallback())
-
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
@@ -1159,6 +1058,11 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.save_safetensors is not None:
             training_arguments_kwargs["save_safetensors"] = self.cfg.save_safetensors
 
+        if self.cfg.sample_packing_eff_est:
+            training_arguments_kwargs[
+                "sample_packing_efficiency"
+            ] = self.cfg.sample_packing_eff_est
+
         if self.cfg.dataloader_pin_memory is not None:
             training_arguments_kwargs[
                 "dataloader_pin_memory"
@@ -1206,8 +1110,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             # default to saving each epoch if not defined
             training_arguments_kwargs["save_strategy"] = "epoch"
 
-        training_arguments_kwargs["save_only_model"] = self.cfg.save_only_model
-
         if self.cfg.do_bench_eval:
             training_arguments_kwargs["do_bench_eval"] = self.cfg.do_bench_eval
             if self.cfg.bench_dataset:
@@ -1286,15 +1188,11 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             False if self.cfg.ddp else None
         )
         training_arguments_kwargs["group_by_length"] = self.cfg.group_by_length
-        training_arguments_kwargs["curriculum_sampling"] = self.cfg.curriculum_sampling
-        report_to = []
+        report_to = None
         if self.cfg.use_wandb:
-            report_to.append("wandb")
+            report_to = "wandb"
         if self.cfg.use_mlflow:
-            report_to.append("mlflow")
-        if self.cfg.use_tensorboard:
-            report_to.append("tensorboard")
-
+            report_to = "mlflow"
         training_arguments_kwargs["report_to"] = report_to
         training_arguments_kwargs["run_name"] = (
             self.cfg.wandb_name if self.cfg.use_wandb else None
@@ -1334,27 +1232,20 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["weight_decay"] = (
             self.cfg.weight_decay if self.cfg.weight_decay is not None else 0.0
         )
-
-        training_arguments_kwargs["sample_packing"] = bool(self.cfg.sample_packing)
-        training_arguments_kwargs[
-            "multipack_real_batches"
-        ] = not self.cfg.flash_attention
-        training_arguments_kwargs["eval_sample_packing"] = bool(
-            self.cfg.eval_sample_packing
+        training_arguments_kwargs["sample_packing"] = (
+            self.cfg.sample_packing if self.cfg.sample_packing else False
         )
-        if self.cfg.sample_packing_bin_size is not None:
-            training_arguments_kwargs[
-                "sample_packing_bin_size"
-            ] = self.cfg.sample_packing_bin_size
-        if self.cfg.sample_packing_group_size is not None:
-            training_arguments_kwargs[
-                "sample_packing_group_size"
-            ] = self.cfg.sample_packing_group_size
-        if self.cfg.sample_packing_eff_est:
-            training_arguments_kwargs[
-                "sample_packing_efficiency"
-            ] = self.cfg.sample_packing_eff_est
-
+        training_arguments_kwargs["multipack_real_batches"] = (
+            self.cfg.flash_attention is not True
+        )
+        training_arguments_kwargs["eval_sample_packing"] = (
+            self.cfg.sample_packing
+            if self.cfg.eval_sample_packing is not False
+            else False
+        )
+        training_arguments_kwargs[
+            "sample_packing_seq_len_multiplier"
+        ] = self.cfg.micro_batch_size
         if self.cfg.relora_steps:
             training_arguments_kwargs["relora_steps"] = self.cfg.relora_steps
             training_arguments_kwargs[
@@ -1517,15 +1408,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         )
 
 
-class HFRLTrainerBuilder(TrainerBuilderBase):
+class HFDPOTrainerBuilder(TrainerBuilderBase):
     """
     Trainer factory class for DPO Trainer
     """
 
     def get_callbacks(self):
         callbacks = super().get_callbacks()
-        callbacks.append(SaveModelCallback())
-
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
@@ -1561,12 +1450,9 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             training_args_kwargs["eval_steps"] = self.cfg.eval_steps
         else:
             training_args_kwargs["evaluation_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"
         )
@@ -1615,42 +1501,12 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             # default to saving each epoch if not defined
             training_args_kwargs["save_strategy"] = "epoch"
 
-        if self.cfg.orpo_alpha:
-            # trl does some odd mapping of alpha to beta to reuse the beta parameter ???
-            training_args_kwargs["beta"] = self.cfg.orpo_alpha
-
-        training_args_cls = AxolotlDPOConfig
-        if self.cfg.rpo_alpha is not None:
-            training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
-        if self.cfg.rl == "orpo":
-            training_args_cls = AxolotlORPOConfig
-            training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
-            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
-
-        if self.cfg.rl == "kto":
-            training_args_cls = AxolotlKTOConfig
-
-            training_args_kwargs["beta"] = self.cfg.rl_beta or 0.1
-            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["dataset_num_proc"] = self.cfg.dataset_processes
-            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
-
-        training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
-            output_dir=self.cfg.output_dir,
+        training_args = TrainingArguments(
             per_device_train_batch_size=self.cfg.micro_batch_size,
             max_steps=self.cfg.max_steps or total_num_steps,
             gradient_accumulation_steps=self.cfg.gradient_accumulation_steps,
             learning_rate=self.cfg.learning_rate,
+            output_dir=self.cfg.output_dir,
             warmup_steps=self.cfg.warmup_steps,
             logging_first_step=True,
             logging_steps=1,
@@ -1678,37 +1534,20 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             dpo_trainer_kwargs[
                 "precompute_ref_log_probs"
             ] = self.cfg.precompute_ref_log_probs
-        if self.cfg.rl in ["dpo", "ipo", "kto_pair"]:
-            trainer_cls = AxolotlDPOTrainer
-            dpo_trainer_kwargs["beta"] = self.cfg.rl_beta or 0.1
-            trainer_cls_args = [self.model, self.model_ref]
-
-            # these aren't used for the ORPO trainer
-            dpo_trainer_kwargs["max_length"] = self.cfg.sequence_len
-            dpo_trainer_kwargs["max_target_length"] = None
-            dpo_trainer_kwargs["max_prompt_length"] = self.cfg.sequence_len
-            dpo_trainer_kwargs["generate_during_eval"] = True
-            if self.cfg.rl == "dpo":
-                dpo_trainer_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
-        elif self.cfg.rl == "orpo":
-            trainer_cls = AxolotlORPOTrainer
-            trainer_cls_args = [self.model]
-        elif self.cfg.rl == "kto":
-            trainer_cls = AxolotlKTOTrainer
-            trainer_cls_args = [self.model]
-        else:
-            raise ValueError(f"Unsupported RL: {self.cfg.rl}")
-        dpo_trainer = trainer_cls(
-            *trainer_cls_args,
+        dpo_trainer = AxolotlDPOTrainer(
+            self.model,
+            self.model_ref,
             args=training_args,
+            beta=self.cfg.dpo_beta or 0.1,
             train_dataset=self.train_dataset,
             tokenizer=self.tokenizer,
+            max_length=self.cfg.sequence_len,
+            max_target_length=None,
+            max_prompt_length=self.cfg.sequence_len,
+            generate_during_eval=True,
             callbacks=self.get_callbacks(),
             **dpo_trainer_kwargs,
         )
-        if self.cfg.fsdp:
-            ensure_dtype(dpo_trainer.model, dtype=self.cfg.torch_dtype)
-
         dpo_trainer = self.hook_post_create_trainer(dpo_trainer)
         for callback in self.get_post_trainer_create_callbacks(dpo_trainer):
             dpo_trainer.add_callback(callback)

src/axolotl/monkeypatch/fastchat_conversation_turns.py

@@ -123,25 +123,6 @@ def get_turns(  # pylint: disable=too-many-return-statements
             else:
                 yield role, ""
         return
-    if self.sep_style == SeparatorStyle.LLAMA3:
-        if self.system_message:
-            # For llama3, the system message is NOT incorporated into the first human instruction
-            # All messages follow <|start_header_id|>' + role + '<|end_header_id|>\n\n'+ message + '<|eot_id|>
-            yield "", system_prompt
-        for i, (role, message) in enumerate(self.messages):
-            if message:
-                yield f"<|start_header_id|>{role}<|end_header_id|>\n\n", f"{message.strip()}<|eot_id|>"
-            else:
-                yield f"<|start_header_id|>{role}<|end_header_id|>\n\n", ""
-        return
-    if self.sep_style == SeparatorStyle.GEMMA:
-        if self.system_message:
-            raise ValueError("Gemma chat template does not support system messages")
-        for i, (role, message) in enumerate(self.messages):
-            prefix = "<bos>" if i == 0 else ""
-            message_str = message if message else ""
-            yield prefix + "<start_of_turn>" + role + "\n", message_str + "<end_of_turn>\n"
-        return
     if self.sep_style == SeparatorStyle.CHATGLM:
         # source: https://huggingface.co/THUDM/chatglm-6b/blob/1d240ba371910e9282298d4592532d7f0f3e9f3e/modeling_chatglm.py#L1302-L1308
         # source2: https://huggingface.co/THUDM/chatglm2-6b/blob/e186c891cf64310ac66ef10a87e6635fa6c2a579/modeling_chatglm.py#L926