remove direct dependency on fused dense lib (#2027)

* remove fastchat and sharegpt

* remove imports

* remove more fastchat imports

* chore: remove unused functions

* feat: remove sharegpt and deprecate from docs

* chore: remove unused sharegpt checks

* fix: remove sharegpt type from tests

* feat: add sharegpt deprecation error

* feat: update readme

---------

Co-authored-by: NanoCode012 <nano@axolotl.ai>

.github/workflows/base.yml

@@ -40,7 +40,7 @@ jobs:
             cuda_version: 12.4.1
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
     steps:
       - name: Checkout

.github/workflows/main.yml

@@ -32,7 +32,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -94,7 +94,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:

.github/workflows/multi-gpu-e2e.yml

@@ -31,7 +31,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             axolotl_extras:
             num_gpus: 2
             nightly_build: "true"

.github/workflows/nightlies.yml

@@ -31,7 +31,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -93,7 +93,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:

.github/workflows/pypi.yml

@@ -4,6 +4,7 @@ on:
   push:
     tags:
       - '*'
+  workflow_dispatch:
 
 jobs:
   pypi-publish:

.github/workflows/tests-nightly.yml

@@ -25,7 +25,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.10", "3.11"]
-        pytorch_version: ["2.3.1", "2.4.1", "2.5.0"]
+        pytorch_version: ["2.3.1", "2.4.1", "2.5.1"]
     timeout-minutes: 20
 
     steps:
@@ -82,13 +82,6 @@ jobs:
             num_gpus: 1
             axolotl_extras: mamba-ssm
             nightly_build: "true"
-          - cuda: 121
-            cuda_version: 12.1.1
-            python_version: "3.11"
-            pytorch: 2.3.1
-            num_gpus: 1
-            axolotl_extras: mamba-ssm
-            nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -99,7 +92,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"

.github/workflows/tests.yml

@@ -36,7 +36,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.10", "3.11"]
-        pytorch_version: ["2.3.1", "2.4.1", "2.5.0"]
+        pytorch_version: ["2.3.1", "2.4.1", "2.5.1"]
     timeout-minutes: 20
 
     steps:
@@ -72,7 +72,7 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
-  docker-e2e-tests:
+  docker-e2e-tests-1st:
     if: github.repository_owner == 'axolotl-ai-cloud'
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
@@ -83,30 +83,58 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 121
-            cuda_version: 12.1.1
-            python_version: "3.10"
-            pytorch: 2.3.1
-            num_gpus: 1
-            axolotl_extras: mamba-ssm
-          - cuda: 121
-            cuda_version: 12.1.1
-            python_version: "3.11"
-            pytorch: 2.3.1
-            num_gpus: 1
-            axolotl_extras: mamba-ssm
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.4.1
             num_gpus: 1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.0
-            num_gpus: 1
-            axolotl_extras:
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Install Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+      - name: Install Modal
+        run: |
+          python -m pip install --upgrade pip
+          pip install modal==0.63.64 jinja2
+      - name: Update env vars
+        run: |
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
+          echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
+          echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
+          echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
+          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+      - name: Run tests job on Modal
+        run: |
+          modal run cicd.tests
+
+  docker-e2e-tests:
+    if: github.repository_owner == 'axolotl-ai-cloud'
+    # this job needs to be run on self-hosted GPU runners...
+    runs-on: [self-hosted, modal]
+    timeout-minutes: 90
+    needs: [pre-commit, pytest, docker-e2e-tests-1st]
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 121
+            cuda_version: 12.1.1
+            python_version: "3.10"
+            pytorch: 2.3.1
+            num_gpus: 1
+            axolotl_extras: mamba-ssm
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.5.1
+            num_gpus: 1
+            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4

README.md

@@ -383,11 +383,10 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
         - typescript
       type: ... # unimplemented custom format
 
-      # fastchat conversation (deprecation soon, use chat_template https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html#chat_template)
-      # See 'conversation' options: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+      # chat_template https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html#chat_template
     - path: ...
-      type: sharegpt
-      conversation: chatml # default: vicuna_v1.1
+      type: chat_template
+      chat_template: chatml # defaults to tokenizer's chat_template
 
       # local
     - path: data.jsonl # or json
@@ -562,7 +561,8 @@ plugins:
   - axolotl.integrations.liger.LigerPlugin
 liger_rope: true
 liger_rms_norm: true
-liger_swiglu: true
+liger_glu_activation: true
+liger_layer_norm: true
 liger_fused_linear_cross_entropy: true
 ```
 

devtools/dev_chat_template.yml

@@ -1,4 +1,4 @@
-# Example config for debugging the sharegpt prompt format
+# Example config for debugging the chat_template prompt format
 base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
@@ -7,8 +7,8 @@ load_in_8bit: true
 load_in_4bit: false
 
 datasets:
-  - path: philschmid/guanaco-sharegpt-style
-    type: sharegpt
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
     shards: 10
 val_set_size: 0
 output_dir: temp_debug/axolotl_outputs/model

docker/Dockerfile-base

@@ -35,3 +35,7 @@ RUN git lfs install --skip-repo && \
     pip3 install awscli && \
     # The base image ships with `pydantic==1.8.2` which is not working
     pip3 install -U --no-cache-dir pydantic==1.10.10
+
+RUN if [ "$PYTHON_VERSION" != "2.5.1" ] ; then \
+        pip3 install flash-attn==2.6.3; \
+    fi

docs/config.qmd

@@ -83,7 +83,7 @@ lora_on_cpu: true
 datasets:
   # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
   - path: vicgalle/alpaca-gpt4
-    # The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection]
+    # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]
     type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
     ds_type: # Optional[str] (json|arrow|parquet|text|csv) defines the datatype when path is a file
     data_files: # Optional[str] path to source data files
@@ -92,15 +92,6 @@ datasets:
     train_on_split: train # Optional[str] name of dataset split to load from
     revision: # Optional[str] The specific revision of the dataset to use when loading from the Hugging Face Hub. This can be a commit hash, tag, or branch name. If not specified, the latest version will be used. This parameter is ignored for local datasets.
 
-    # Optional[str] fastchat conversation type, only used with type: sharegpt
-    conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
-    field_human: # Optional[str]. Human key to use for conversation.
-    field_model: # Optional[str]. Assistant key to use for conversation.
-    # Add additional keys from your dataset as input or output roles
-    roles:
-      input: # Optional[List[str]]. These will be masked based on train_on_input
-      output: # Optional[List[str]].
-
   # Custom user instruction prompt
   - path: repo
     type:
@@ -183,6 +174,8 @@ test_datasets:
 
 # use RL training: 'dpo', 'ipo', 'kto'
 rl:
+# whether to perform weighting if doing DPO training. Boolean.
+dpo_use_weighting:
 
 # The name of the chat template to use for training, following values are supported:
 # - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.

docs/dataset-formats/conversation.qmd

@@ -6,33 +6,8 @@ order: 3
 
 ## sharegpt
 
-UPDATE: ShareGPT is being deprecated in the next release. Please see `chat_template` section below.
+IMPORTANT: ShareGPT is deprecated!. Please see `chat_template` section below.
 
-conversations where `from` is `human`/`gpt`. (optional: first row with role `system` to override default system prompt)
-
-```{.json filename="data.jsonl"}
-{"conversations": [{"from": "...", "value": "..."}]}
-```
-
-Note: `type: sharegpt` opens special configs:
-- `conversation`: enables conversions to many Conversation types. Refer to the 'name' [here](https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py) for options.
-- `roles`: allows you to specify the roles for input and output. This is useful for datasets with custom roles such as `tool` etc to support masking.
-- `field_human`: specify the key to use instead of `human` in the conversation.
-- `field_model`: specify the key to use instead of `gpt` in the conversation.
-
-```yaml
-datasets:
-    path: ...
-    type: sharegpt
-
-    conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
-    field_human: # Optional[str]. Human key to use for conversation.
-    field_model: # Optional[str]. Assistant key to use for conversation.
-    # Add additional keys from your dataset as input or output roles
-    roles:
-      input: # Optional[List[str]]. These will be masked based on train_on_input
-      output: # Optional[List[str]].
-```
 
 ## pygmalion
 
@@ -40,38 +15,6 @@ datasets:
 {"conversations": [{"role": "...", "value": "..."}]}
 ```
 
-## sharegpt.load_role
-
-conversations where `role` is used instead of `from`
-
-```{.json filename="data.jsonl"}
-{"conversations": [{"role": "...", "value": "..."}]}
-```
-
-## sharegpt.load_guanaco
-
-conversations where `from` is `prompter` `assistant` instead of default sharegpt
-
-```{.json filename="data.jsonl"}
-{"conversations": [{"from": "...", "value": "..."}]}
-```
-
-## sharegpt.load_ultrachat
-
-conversations where the turns field is 'messages', human is 'user' and gpt is 'assistant'.
-
-```{.json filename="data.jsonl"}
-{"messages": [{"user": "...", "assistant": "..."}]}
-```
-
-## sharegpt_jokes
-
-creates a chat where bot is asked to tell a joke, then explain why the joke is funny
-
-```{.json filename="data.jsonl"}
-{"conversations": [{"title": "...", "text": "...", "explanation": "..."}]}
-```
-
 
 ## chat_template
 

docs/debugging.qmd

@@ -51,12 +51,12 @@ While debugging it's helpful to simplify your test scenario as much as possible.
 
 ### Background
 
-The below example shows how to configure VSCode to debug data preprocessing of the `sharegpt` format.  This is the format used when you have the following in your axolotl config:
+The below example shows how to configure VSCode to debug data preprocessing of the `chat_template` format.  This is the format used when you have the following in your axolotl config:
 
 ```yaml
 datasets:
-  - path: <path to your sharegpt formatted dataset> # example on HF Hub: philschmid/guanaco-sharegpt-style
-    type: sharegpt
+  - path: <path to your chat_template formatted dataset> # example on HF Hub: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
 ```
 
 >[!Important]
@@ -83,7 +83,7 @@ If you developing on a remote host, you can easily use VSCode to debug remotely.
 
 The easiest way to get started is to modify the [.vscode/launch.json](../.vscode/launch.json) file in this project.  This is just an example configuration, so you may need to modify or copy it to suit your needs.
 
-For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_sharegpt.yml`, you would use the below configuration[^1].  Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to `devtools` and set the `env` variable `HF_HOME` to a temporary folder that is later partially deleted.  This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.
+For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_chat_template.yml`, you would use the below configuration[^1].  Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to `devtools` and set the `env` variable `HF_HOME` to a temporary folder that is later partially deleted.  This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.
 
 ```jsonc
 // .vscode/launch.json
@@ -91,12 +91,12 @@ For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 acceler
     "version": "0.2.0",
     "configurations": [
         {
-            "name": "Debug axolotl prompt - sharegpt",
+            "name": "Debug axolotl prompt - chat_template",
             "type": "python",
             "module": "accelerate.commands.launch",
             "request": "launch",
             "args": [
-                "-m", "axolotl.cli.train", "dev_sharegpt.yml",
+                "-m", "axolotl.cli.train", "dev_chat_template.yml",
                 // The flags below simplify debugging by overriding the axolotl config
                 // with the debugging tips above.  Modify as needed.
                 "--dataset_processes=1",      // limits data preprocessing to one process
@@ -240,6 +240,6 @@ style="border-radius: 10px; display: block; margin: auto;" width="560" height="3
 </div>
 <br>
 
-[^1]: The config actually mimics the command `CUDA_VISIBLE_DEVICES=0 python -m accelerate.commands.launch -m axolotl.cli.train devtools/sharegpt.yml`, but this is the same thing.
+[^1]: The config actually mimics the command `CUDA_VISIBLE_DEVICES=0 python -m accelerate.commands.launch -m axolotl.cli.train devtools/chat_template.yml`, but this is the same thing.
 
 [^2]: Many of the below flags are recommended best practices by Nvidia when using nvidia-container-toolkit.  You can read more about these flags [here](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html).

examples/deepseek-v2/qlora-fsdp-2_5.yaml

@@ -9,14 +9,17 @@ strict: false
 plugins:
   - axolotl.integrations.liger.LigerPlugin
 liger_rms_norm: true
-liger_swiglu: true
+liger_glu_activation: true
 liger_fused_linear_cross_entropy: true
 
 chat_template: deepseek_v2
 datasets:
   - path: mlabonne/FineTome-100k
     type: chat_template
-    split: train
+    split: train[:20%]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0

examples/gemma2/qlora.yml

@@ -11,8 +11,11 @@ chat_template: gemma
 datasets:
   - path: cgato/SlimOrcaDedupCleaned
     type: chat_template
-    chat_template: gemma
     drop_system_message: true
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+
 val_set_size: 0.0
 output_dir: ./outputs/out
 

examples/jamba/qlora_fsdp_large.yaml

@@ -4,11 +4,15 @@ tokenizer_type: AutoTokenizer
 load_in_4bit: true
 strict: false
 use_tensorboard: true
+chat_template: jamba
 datasets:
   - path: cgato/SlimOrcaDedupCleaned
     type: chat_template
-    chat_template: jamba
     drop_system_message: true
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: jamba-large-fsdp-qlora-ft

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

@@ -4,7 +4,7 @@ plugins:
   - axolotl.integrations.liger.LigerPlugin
 liger_rope: true
 liger_rms_norm: true
-liger_swiglu: true
+liger_glu_activation: true
 liger_fused_linear_cross_entropy: true
 
 strict: false
@@ -14,6 +14,10 @@ datasets:
   - path: mlabonne/FineTome-100k
     type: chat_template
     split: train[:20%]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.02
 output_dir: ./outputs/out

examples/phi/lora-3.5.yaml

@@ -10,7 +10,6 @@ chat_template: phi_3
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
-    chat_template: phi_3
     field_messages: messages
     message_field_role: role
     message_field_content: content

requirements.txt

@@ -1,10 +1,10 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
 peft==0.13.2
-transformers==4.46.0
+transformers==4.46.1
 tokenizers>=0.20.1
 bitsandbytes==0.44.1
-accelerate==1.0.1
+accelerate==1.1.0
 datasets==3.0.1
 deepspeed==0.15.3
 pydantic==2.6.3
@@ -28,13 +28,12 @@ scipy
 scikit-learn==1.4.2
 pynvml
 art
-fschat @ git+https://github.com/lm-sys/FastChat.git@27a05b04a35510afb1d767ae7e5990cbd278f8fe
 gradio==3.50.2
 tensorboard
 python-dotenv==1.0.1
 autoawq>=0.2.5
 triton>=2.3.0
-liger-kernel==0.3.0
+liger-kernel==0.4.0
 
 mamba-ssm==1.2.0.post1
 
@@ -43,7 +42,7 @@ s3fs>=2024.5.0
 gcsfs>=2024.5.0
 # adlfs
 
-trl @ git+https://github.com/huggingface/trl.git@31d02cfb795284591a084416b9dcb7bef5d08924
+trl==0.12.0
 zstandard==0.22.0
 fastcore
 

setup.py

@@ -89,7 +89,7 @@ install_requires, dependency_links = parse_requirements()
 
 setup(
     name="axolotl",
-    version="0.4.1",
+    version="0.5.0",
     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"},
@@ -100,9 +100,6 @@ setup(
         "flash-attn": [
             "flash-attn==2.6.3",
         ],
-        "fused-dense-lib": [
-            "fused-dense-lib  @ git+https://github.com/Dao-AILab/flash-attention@v2.6.2#subdirectory=csrc/fused_dense_lib",
-        ],
         "deepspeed": [
             "deepspeed==0.14.4",
             "deepspeed-kernels",

src/axolotl/cli/__init__.py

@@ -272,7 +272,7 @@ def do_inference_gradio(
             importlib.import_module("axolotl.prompters"), prompter
         )
     elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template)
+        chat_template_str = get_chat_template(cfg.chat_template, tokenizer=tokenizer)
 
     model = model.to(cfg.device, dtype=cfg.torch_dtype)
 

src/axolotl/cli/preprocess.py

@@ -23,10 +23,6 @@ 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.utils.trainer import disable_datasets_caching
 
 LOG = logging.getLogger("axolotl.cli.preprocess")
@@ -44,23 +40,6 @@ 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 not parsed_cfg.dataset_prepared_path:
         msg = (
             Fore.RED

src/axolotl/cli/train.py

@@ -19,10 +19,6 @@ from axolotl.cli import (
 )
 from axolotl.common.cli import TrainerCliArgs
 from axolotl.integrations.base import PluginManager
-from axolotl.prompt_strategies.sharegpt import (
-    register_chatml_template,
-    register_llama3_template,
-)
 from axolotl.train import train
 
 LOG = logging.getLogger("axolotl.cli.train")
@@ -42,21 +38,6 @@ def do_train(cfg, cli_args) -> None:
     print_axolotl_text_art()
     check_accelerate_default_config()
     check_user_token()
-    if cfg.chat_template == "chatml" and cfg.default_system_message:
-        LOG.info(
-            f"ChatML set. Adding default system message: {cfg.default_system_message}"
-        )
-        register_chatml_template(cfg.default_system_message)
-    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":
         dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)

src/axolotl/core/trainer_builder.py

@@ -48,6 +48,7 @@ from trl import (
 )
 from trl.trainer.utils import RewardDataCollatorWithPadding, pad_to_length
 
+from axolotl.integrations.base import PluginManager
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
 from axolotl.utils import is_comet_available, is_mlflow_available
@@ -895,13 +896,13 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         for key, value in metrics.items():
             self._stored_metrics[train_eval][key].append(value)
 
-    def _save_checkpoint(self, model, trial, metrics=None):
+    def _save_checkpoint(self, model, trial, **kwargs):
         # make sure the checkpoint dir exists, since trainer is flakey
         checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
         run_dir = self._get_output_dir(trial=trial)
         output_dir = os.path.join(run_dir, checkpoint_folder)
         os.makedirs(output_dir, exist_ok=True)
-        return super()._save_checkpoint(model, trial, metrics=metrics)
+        return super()._save_checkpoint(model, trial, **kwargs)
 
 
 class AxolotlMambaTrainer(AxolotlTrainer):
@@ -1147,6 +1148,12 @@ class TrainerBuilderBase(abc.ABC):
 
     def get_callbacks(self) -> List[TrainerCallback]:
         callbacks = []
+
+        plugin_manager = PluginManager.get_instance()
+        callbacks.extend(
+            plugin_manager.add_callbacks_pre_trainer(cfg=self.cfg, model=self.model)
+        )
+
         if self.cfg.use_wandb:
             callbacks.append(
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
@@ -1173,11 +1180,17 @@ class TrainerBuilderBase(abc.ABC):
 
         return callbacks
 
-    @abstractmethod
     def get_post_trainer_create_callbacks(self, trainer):
         """
         Callbacks added after the trainer is created, usually b/c these need access to the trainer
         """
+        callbacks = []
+
+        plugin_manager = PluginManager.get_instance()
+        callbacks.extend(
+            plugin_manager.add_callbacks_post_trainer(cfg=self.cfg, trainer=trainer)
+        )
+        return callbacks
 
     def hook_pre_create_training_args(self, training_arguments_kwargs):
         # TODO
@@ -1223,7 +1236,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = []
+        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
         if self.cfg.use_wandb and self.cfg.eval_table_size > 0:
             LogPredictionCallback = log_prediction_callback_factory(
                 trainer, self.tokenizer, "wandb"
@@ -1595,7 +1608,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["pretraining"] = bool(self.cfg.pretraining_dataset)
         if self.cfg.chat_template:
             training_arguments_kwargs["chat_template"] = get_chat_template(
-                self.cfg.chat_template
+                self.cfg.chat_template,
+                tokenizer=self.tokenizer,
             )
 
         if self.cfg.rl == "orpo":
@@ -1790,7 +1804,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = []
+        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
         return callbacks
 
     def build_training_arguments(self, total_num_steps):
@@ -1876,17 +1890,18 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             # default to saving each epoch if not defined
             training_args_kwargs["save_strategy"] = "epoch"
 
+        training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
+
         if self.cfg.rl_beta:
             training_args_kwargs["beta"] = self.cfg.rl_beta
         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_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
-        training_args_cls = AxolotlDPOConfig
         if self.cfg.rpo_alpha is not None:
             training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
 
+        training_args_cls = None
         if self.cfg.rl == "simpo":
             training_args_cls = AxolotlCPOConfig
             training_args_kwargs["loss_type"] = "simpo"
@@ -1895,13 +1910,13 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.cpo_alpha is not None:
                 training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
 
-        if self.cfg.rl == "orpo":
+        elif self.cfg.rl == "orpo":
             training_args_cls = AxolotlORPOConfig
             training_args_kwargs["max_length"] = self.cfg.sequence_len
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
-        if self.cfg.rl == "kto":
+        elif self.cfg.rl == "kto":
             training_args_cls = AxolotlKTOConfig
 
             training_args_kwargs["desirable_weight"] = (
@@ -1916,6 +1931,11 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
+        else:
+            training_args_cls = AxolotlDPOConfig
+            if self.cfg.dpo_use_weighting is not None:
+                training_args_kwargs["use_weighting"] = self.cfg.dpo_use_weighting
+
         training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
             output_dir=self.cfg.output_dir,
             per_device_train_batch_size=self.cfg.micro_batch_size,
@@ -1999,11 +2019,11 @@ class HFPPOTrainerBuilder(TrainerBuilderBase):
     """
 
     def get_callbacks(self):
-        callbacks = []
+        callbacks = super().get_callbacks()
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = []
+        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
         return callbacks
 
     def build(self, total_num_steps):

src/axolotl/integrations/base.py

@@ -18,9 +18,10 @@ Plugins can be used to integrate third-party models, modify the training process
 
 To create a new plugin, you need to inherit from the BasePlugin class and implement the required methods.
 """
+import collections
 import importlib
 import logging
-from typing import List
+from typing import OrderedDict
 
 
 class BasePlugin:
@@ -47,7 +48,7 @@ class BasePlugin:
         Initializes the BasePlugin.
         """
 
-    def register(self, cfg):
+    def register(self, cfg):  # pylint: disable=unused-argument
         """
         Registers the plugin with the given configuration.
 
@@ -63,7 +64,7 @@ class BasePlugin:
         Returns a pydantic model for the plugin's input arguments.
         """
 
-    def pre_model_load(self, cfg):
+    def pre_model_load(self, cfg):  # pylint: disable=unused-argument
         """
         Performs actions before the model is loaded.
 
@@ -74,7 +75,7 @@ class BasePlugin:
         None
         """
 
-    def post_model_load(self, cfg, model):
+    def post_model_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions after the model is loaded.
 
@@ -86,7 +87,7 @@ class BasePlugin:
         None
         """
 
-    def pre_lora_load(self, cfg, model):
+    def pre_lora_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions before LoRA weights are loaded.
 
@@ -98,7 +99,7 @@ class BasePlugin:
         None
         """
 
-    def post_lora_load(self, cfg, model):
+    def post_lora_load(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions after LoRA weights are loaded.
 
@@ -110,7 +111,7 @@ class BasePlugin:
         None
         """
 
-    def create_optimizer(self, cfg, trainer):
+    def create_optimizer(self, cfg, trainer):  # pylint: disable=unused-argument
         """
         Creates and returns an optimizer for training.
 
@@ -122,7 +123,9 @@ class BasePlugin:
         object: The created optimizer.
         """
 
-    def create_lr_scheduler(self, cfg, trainer, optimizer):
+    def create_lr_scheduler(
+        self, cfg, trainer, optimizer
+    ):  # pylint: disable=unused-argument
         """
         Creates and returns a learning rate scheduler.
 
@@ -135,7 +138,7 @@ class BasePlugin:
         object: The created learning rate scheduler.
         """
 
-    def add_callbacks_pre_trainer(self, cfg, model):
+    def add_callbacks_pre_trainer(self, cfg, model):  # pylint: disable=unused-argument
         """
         Adds callbacks to the trainer before training.
 
@@ -146,8 +149,11 @@ class BasePlugin:
         Returns:
         List[callable]: A list of callback functions to be added to the TrainingArgs
         """
+        return []
 
-    def add_callbacks_post_trainer(self, cfg, trainer):
+    def add_callbacks_post_trainer(
+        self, cfg, trainer
+    ):  # pylint: disable=unused-argument
         """
         Adds callbacks to the trainer after training.
 
@@ -158,8 +164,9 @@ class BasePlugin:
         Returns:
         List[callable]: A list of callback functions to be added to the TrainingArgs
         """
+        return []
 
-    def post_train(self, cfg, model):
+    def post_train(self, cfg, model):  # pylint: disable=unused-argument
         """
         Performs actions after training is complete.
 
@@ -171,7 +178,7 @@ class BasePlugin:
         None
         """
 
-    def post_train_unload(self, cfg):
+    def post_train_unload(self, cfg):  # pylint: disable=unused-argument
         """
         Performs actions after training is complete and the model is unloaded.
 
@@ -227,7 +234,7 @@ class PluginManager:
     pre_model_load(cfg): Calls the pre_model_load method of all registered plugins.
     """
 
-    plugins: List[BasePlugin] = []
+    plugins: OrderedDict[str, BasePlugin] = collections.OrderedDict()
 
     _instance = None
 
@@ -237,7 +244,7 @@ class PluginManager:
         """
         if cls._instance is None:
             cls._instance = super(PluginManager, cls).__new__(cls)
-            cls._instance.plugins: List[BasePlugin] = []
+            cls._instance.plugins = collections.OrderedDict()
         return cls._instance
 
     @staticmethod
@@ -265,7 +272,7 @@ class PluginManager:
         """
         try:
             plugin = load_plugin(plugin_name)
-            self.plugins.append(plugin)
+            self.plugins[plugin_name] = plugin
         except ImportError:
             logging.error(f"Failed to load plugin: {plugin_name}")
 
@@ -277,7 +284,7 @@ class PluginManager:
         list[str]: A list of Pydantic classes for all registered plugins' input arguments.'
         """
         input_args = []
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             input_args_from_plugin = plugin.get_input_args()
             if input_args_from_plugin is not None:
                 input_args.append(input_args_from_plugin)
@@ -293,7 +300,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             plugin.pre_model_load(cfg)
 
     def post_model_load(self, cfg, model):
@@ -307,7 +314,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             plugin.post_model_load(cfg, model)
 
     def pre_lora_load(self, cfg, model):
@@ -321,7 +328,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             plugin.pre_lora_load(cfg, model)
 
     def post_lora_load(self, cfg, model):
@@ -335,7 +342,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             plugin.post_lora_load(cfg, model)
 
     def create_optimizer(self, cfg, trainer):
@@ -349,7 +356,7 @@ class PluginManager:
         Returns:
         object: The created optimizer, or None if none was found.
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             optimizer = plugin.create_optimizer(cfg, trainer)
             if optimizer is not None:
                 return optimizer
@@ -367,7 +374,7 @@ class PluginManager:
         Returns:
         object: The created learning rate scheduler, or None if none was found.
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             scheduler = plugin.create_lr_scheduler(cfg, trainer, optimizer)
             if scheduler is not None:
                 return scheduler
@@ -385,7 +392,7 @@ class PluginManager:
         List[callable]: A list of callback functions to be added to the TrainingArgs.
         """
         callbacks = []
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             callbacks.extend(plugin.add_callbacks_pre_trainer(cfg, model))
         return callbacks
 
@@ -401,7 +408,7 @@ class PluginManager:
         List[callable]: A list of callback functions to be added to the TrainingArgs.
         """
         callbacks = []
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             callbacks.extend(plugin.add_callbacks_post_trainer(cfg, trainer))
         return callbacks
 
@@ -416,5 +423,5 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins:
+        for plugin in self.plugins.values():
             plugin.post_train_unload(cfg)

src/axolotl/integrations/liger/__init__.py

@@ -18,20 +18,23 @@ Module for the Plugin for LIGER integraton with Axolotl.
 Liger Kernel is the collection of Triton-native kernels for LLM Training.
 It is designed to be performant, correct, and light-weight.
 """
+import inspect
 import logging
 import sys
-from functools import partial
 
 from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss
-from liger_kernel.transformers.geglu import LigerGEGLUMLP
+from liger_kernel.transformers.monkey_patch import MODEL_TYPE_TO_APPLY_LIGER_FN
 from liger_kernel.transformers.rms_norm import LigerRMSNorm
 from liger_kernel.transformers.rope import liger_rotary_pos_emb
 from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
 
 from axolotl.integrations.base import BasePlugin
 
+from ...utils.distributed import zero_only
 from .args import LigerArgs  # pylint: disable=unused-import. # noqa: F401
 
+LOG = logging.getLogger("axolotl.integrations.liger")
+
 
 class LigerPlugin(BasePlugin):
     """
@@ -42,59 +45,31 @@ class LigerPlugin(BasePlugin):
         return "axolotl.integrations.liger.LigerArgs"
 
     def pre_model_load(self, cfg):
-        if cfg.model_config_type == "llama":
-            from liger_kernel.transformers.model.llama import (
-                lce_forward as llama_lce_forward,
-            )
-            from transformers.models.llama import modeling_llama
-
-            if cfg.liger_rope:
-                modeling_llama.apply_rotary_pos_emb = liger_rotary_pos_emb
-            if cfg.liger_rms_norm:
-                modeling_llama.LlamaRMSNorm = LigerRMSNorm
-            if cfg.liger_swiglu:
-                modeling_llama.LlamaMLP = LigerSwiGLUMLP
-            if cfg.liger_cross_entropy:
-                modeling_llama.CrossEntropyLoss = LigerCrossEntropyLoss
-            elif cfg.liger_fused_linear_cross_entropy:
-                modeling_llama.LlamaForCausalLM.forward = llama_lce_forward
-
-        elif cfg.model_config_type == "mistral":
-            from liger_kernel.transformers.model.mistral import (
-                lce_forward as mistral_lce_forward,
-            )
-            from transformers.models.mistral import modeling_mistral
-
-            if cfg.liger_rope:
-                modeling_mistral.apply_rotary_pos_emb = liger_rotary_pos_emb
-            if cfg.liger_rms_norm:
-                modeling_mistral.MistralRMSNorm = LigerRMSNorm
-            if cfg.liger_swiglu:
-                modeling_mistral.MistralMLP = LigerSwiGLUMLP
-            if cfg.liger_cross_entropy:
-                modeling_mistral.CrossEntropyLoss = LigerCrossEntropyLoss
-            if cfg.liger_fused_linear_cross_entropy:
-                modeling_mistral.MistralForCausalLM.forward = mistral_lce_forward
-
-        elif cfg.model_config_type == "gemma":
-            from liger_kernel.transformers.model.gemma import (
-                lce_forward as gemma_lce_forward,
-            )
-            from transformers.models.gemma import modeling_gemma
-
-            if cfg.liger_rope:
-                modeling_gemma.apply_rotary_pos_emb = liger_rotary_pos_emb
-            if cfg.liger_rms_norm:
-                modeling_gemma.GemmaRMSNorm = partial(
-                    LigerRMSNorm, offset=1.0, init_fn="zeros", casting_mode="gemma"
+        if cfg.model_config_type in MODEL_TYPE_TO_APPLY_LIGER_FN:
+            apply_liger_fn = MODEL_TYPE_TO_APPLY_LIGER_FN[cfg.model_config_type]
+            liger_fn_sig = inspect.signature(apply_liger_fn)
+            kwargs = {}
+            if "rope" in liger_fn_sig.parameters:
+                kwargs["rope"] = cfg.liger_rope
+            if "cross_entropy" in liger_fn_sig.parameters:
+                kwargs["cross_entropy"] = cfg.liger_cross_entropy
+            if "fused_linear_cross_entropy" in liger_fn_sig.parameters:
+                kwargs[
+                    "fused_linear_cross_entropy"
+                ] = cfg.liger_fused_linear_cross_entropy
+            if "rms_norm" in liger_fn_sig.parameters:
+                kwargs["rms_norm"] = cfg.liger_rms_norm
+            if "layer_norm" in liger_fn_sig.parameters:
+                kwargs["layer_norm"] = cfg.liger_layer_norm
+            if "geglu" in liger_fn_sig.parameters:
+                kwargs["geglu"] = cfg.liger_glu_activation
+            elif "swiglu" in liger_fn_sig.parameters:
+                kwargs["swiglu"] = cfg.liger_glu_activation
+            with zero_only():
+                LOG.info(
+                    f"Applying LIGER to {cfg.model_config_type} with kwargs: {kwargs}"
                 )
-            if cfg.liger_swiglu:
-                modeling_gemma.GemmaMLP = LigerGEGLUMLP
-            if cfg.liger_cross_entropy:
-                modeling_gemma.CrossEntropyLoss = LigerCrossEntropyLoss
-            if cfg.liger_fused_linear_cross_entropy:
-                modeling_gemma.GemmaForCausalLM.forward = gemma_lce_forward
-
+            apply_liger_fn(**kwargs)
         elif cfg.model_config_type == "jamba":
             from transformers.models.jamba import modeling_jamba
 
@@ -104,30 +79,12 @@ class LigerPlugin(BasePlugin):
                 modeling_jamba.apply_rotary_pos_emb = liger_rotary_pos_emb
             if cfg.liger_rms_norm:
                 modeling_jamba.JambaRMSNorm = LigerRMSNorm
-            if cfg.liger_swiglu:
+            if cfg.liger_glu_activation:
                 modeling_jamba.JambaMLP = LigerSwiGLUMLP
             if cfg.liger_cross_entropy:
                 modeling_jamba.CrossEntropyLoss = LigerCrossEntropyLoss
             if cfg.liger_fused_linear_cross_entropy:
                 modeling_jamba.JambaForCausalLM.forward = jamba_lce_forward
-
-        elif cfg.model_config_type == "qwen2":
-            from liger_kernel.transformers.model.qwen2 import (
-                lce_forward as qwen2_lce_forward,
-            )
-            from transformers.models.qwen2 import modeling_qwen2
-
-            if cfg.liger_rope:
-                modeling_qwen2.apply_rotary_pos_emb = liger_rotary_pos_emb
-            if cfg.liger_rms_norm:
-                modeling_qwen2.Qwen2RMSNorm = LigerRMSNorm
-            if cfg.liger_swiglu:
-                modeling_qwen2.Qwen2MLP = LigerSwiGLUMLP
-            if cfg.liger_cross_entropy:
-                modeling_qwen2.CrossEntropyLoss = LigerCrossEntropyLoss
-            if cfg.liger_fused_linear_cross_entropy:
-                modeling_qwen2.Qwen2ForCausalLM.forward = qwen2_lce_forward
-
         elif cfg.model_config_type == "deepseek_v2":
             from accelerate import init_empty_weights
             from transformers import AutoModelForCausalLM
@@ -146,44 +103,9 @@ class LigerPlugin(BasePlugin):
                 logging.warning("Fused liger_rope is not supported for DeepseekV2.")
             if cfg.liger_rms_norm:
                 modeling_mod.DeepseekV2RMSNorm = LigerRMSNorm
-            if cfg.liger_swiglu:
+            if cfg.liger_glu_activation:
                 modeling_mod.DeepseekV2MLP.forward = LigerSwiGLUMLP.forward
             if cfg.liger_cross_entropy:
                 modeling_mod.CrossEntropyLoss = LigerCrossEntropyLoss
             if cfg.liger_fused_linear_cross_entropy:
                 modeling_mod.DeepseekV2ForCausalLM.forward = deepseekv2_lce_forward
-
-        elif cfg.model_config_type == "gemma2":
-            from transformers.models.gemma2 import modeling_gemma2
-
-            if cfg.liger_rope:
-                modeling_gemma2.apply_rotary_pos_emb = liger_rotary_pos_emb
-            if cfg.liger_rms_norm:
-                modeling_gemma2.Gemma2RMSNorm = partial(
-                    LigerRMSNorm, offset=1.0, init_fn="zeros", casting_mode="gemma"
-                )
-            if cfg.liger_swiglu:
-                modeling_gemma2.Gemma2MLP = LigerGEGLUMLP
-            if cfg.liger_cross_entropy:
-                modeling_gemma2.CrossEntropyLoss = LigerCrossEntropyLoss
-            if cfg.liger_fused_linear_cross_entropy:
-                logging.warning(
-                    "Fused linear cross entropy is not supported for Gemma 2."
-                )
-
-        elif cfg.model_config_type == "phi3":
-            from liger_kernel.transformers.model.phi3 import (
-                lce_forward as phi3_lce_forward,
-            )
-            from transformers.models.phi3 import modeling_phi3
-
-            if cfg.liger_rope:
-                modeling_phi3.apply_rotary_pos_emb = liger_rotary_pos_emb
-            if cfg.liger_rms_norm:
-                modeling_phi3.Phi3RMSNorm = LigerRMSNorm
-            if cfg.liger_swiglu:
-                modeling_phi3.Phi3MLP = LigerSwiGLUMLP
-            if cfg.liger_cross_entropy:
-                modeling_phi3.CrossEntropyLoss = LigerCrossEntropyLoss
-            if cfg.liger_fused_linear_cross_entropy:
-                modeling_phi3.Phi3ForCausalLM.forward = phi3_lce_forward

src/axolotl/integrations/liger/args.py

@@ -15,9 +15,12 @@
 """
 Module for handling LIGER input arguments.
 """
+import logging
 from typing import Optional
 
-from pydantic import BaseModel
+from pydantic import BaseModel, model_validator
+
+LOG = logging.getLogger("axolotl.integrations.liger.args")
 
 
 class LigerArgs(BaseModel):
@@ -27,6 +30,24 @@ class LigerArgs(BaseModel):
 
     liger_rope: Optional[bool] = None
     liger_rms_norm: Optional[bool] = None
+    liger_layer_norm: Optional[bool] = None
     liger_swiglu: Optional[bool] = None
+    liger_glu_activation: Optional[bool] = None
     liger_cross_entropy: Optional[bool] = None
     liger_fused_linear_cross_entropy: Optional[bool] = None
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_deprecated_swiglu(cls, data):
+        if data.get("liger_swiglu") is not None:
+            if data.get("liger_glu_activation") is not None:
+                raise ValueError(
+                    "You cannot have both `liger_swiglu` and `liger_glu_activation` set."
+                )
+
+            LOG.warning(
+                "The 'liger_swiglu' argument is deprecated and will be removed in a future release. "
+                "Please use 'liger_glu_activation' instead."
+            )
+            data["liger_glu_activation"] = data.pop("liger_swiglu")
+        return data

src/axolotl/monkeypatch/fastchat_conversation_turns.py

@@ -1,231 +0,0 @@
-"""
-monkeypatch to add a get_turns method
-"""
-
-import logging
-from typing import Generator, Tuple
-
-from fastchat.conversation import SeparatorStyle
-
-LOG = logging.getLogger("axolotl.monkeypatch.fastchat_conversation_turns")
-
-
-def get_prompt(self) -> str:
-    ret = ""
-    for role, msg in self.get_turns():
-        ret += role + msg
-    return ret
-
-
-def get_turns(  # pylint: disable=too-many-return-statements
-    self,
-) -> Generator[Tuple[str, str], None, None]:
-    """Get the prompt for generation."""
-    system_prompt = self.system_template.format(system_message=self.system_message)
-    if self.sep_style == SeparatorStyle.ADD_COLON_SINGLE:
-        yield "", system_prompt + self.sep
-        for role, message in self.messages:
-            if message:
-                yield role + ": ", message + self.sep
-            else:
-                yield role + ":", ""
-        return
-    if self.sep_style == SeparatorStyle.ADD_COLON_TWO:
-        seps = [self.sep, self.sep2]
-        yield "", system_prompt + seps[0]
-        for i, (role, message) in enumerate(self.messages):
-            if message:
-                yield role + ": ", message + seps[i % 2]
-            else:
-                yield role + ":", ""
-        return
-    if self.sep_style == SeparatorStyle.ADD_COLON_SPACE_SINGLE:
-        yield "", system_prompt + self.sep
-        for role, message in self.messages:
-            if message:
-                yield role + ": ", message + self.sep
-            else:
-                yield role + ": ", ""  # must be end with a space
-        return
-    if self.sep_style == SeparatorStyle.ADD_NEW_LINE_SINGLE:
-        yield "", "" if system_prompt == "" else system_prompt + self.sep
-        for role, message in self.messages:
-            if message:
-                yield role + "\n", message + self.sep
-            else:
-                yield role + "\n", ""
-        return
-    if self.sep_style == SeparatorStyle.NO_COLON_SINGLE:
-        yield "", system_prompt
-        for role, message in self.messages:
-            if message:
-                yield role, message + self.sep
-            else:
-                yield role, ""
-        return
-    if self.sep_style == SeparatorStyle.NO_COLON_TWO:
-        seps = [self.sep, self.sep2]
-        yield "", system_prompt
-        for i, (role, message) in enumerate(self.messages):
-            if message:
-                yield role, message + seps[i % 2]
-            else:
-                yield role, ""
-        return
-    if self.sep_style == SeparatorStyle.RWKV:
-        yield "", system_prompt
-        for i, (role, message) in enumerate(self.messages):
-            if message:
-                yield role + ": ", message.replace("\r\n", "\n").replace(
-                    "\n\n", "\n"
-                ) + "\n\n"
-            else:
-                yield role + ":", ""
-        return
-    if self.sep_style == SeparatorStyle.LLAMA2 and self.name != "mistral":
-        if self.system_message:
-            if self.messages:
-                # For llama, the system message is incorporated into the first human instruction
-                first_role, first_msg = self.messages[0]
-                if first_role == self.roles[0]:
-                    system_prompt += first_msg
-                    self.messages.pop(0)
-            yield "", system_prompt
-        for i, (role, message) in enumerate(self.messages):
-            if message:
-                if (i % 2 == 0 and not self.system_message) or (
-                    i % 2 != 0 and self.system_message
-                ):
-                    role = "<s> " + role
-                yield role + " ", message
-            else:
-                yield role, ""
-        return
-    if self.sep_style == SeparatorStyle.LLAMA2 and self.name == "mistral":
-        contains_sys_msg = False
-        if self.system_message:
-            contains_sys_msg = True
-            if self.messages:
-                # There is no clear guidance on how to handle system messages in Mistral so we just prepend it to the first human instruction separated by a newline
-                first_role, first_msg = self.messages[0]
-                if first_role == self.roles[0]:
-                    system_prompt = self.system_template.format(
-                        system_message=" " + self.system_message
-                    )
-                    system_prompt += first_msg
-                    self.messages.pop(0)
-            yield "", system_prompt
-        for i, (role, message) in enumerate(self.messages):
-            if message and i == 0 and not contains_sys_msg:
-                yield "", system_prompt.strip() + " " + message  # if there is no system message, we need to make sure there is the a `<s> [INST]` at the beginning of the first instruction.
-            elif message:
-                yield role + " ", message
-            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
-        round_add_n = 1 if self.name == "chatglm2" else 0
-        if system_prompt:
-            yield "", system_prompt + self.sep
-
-        for i, (role, message) in enumerate(self.messages):
-            if i % 2 == 0:
-                yield "", f"[Round {i//2 + round_add_n}]{self.sep}"
-
-            if message:
-                yield f"{role}：", f"{message}{self.sep}"
-            else:
-                yield f"{role}：", ""
-        return
-    if self.sep_style == SeparatorStyle.CHATML:
-        yield "", "" if system_prompt == "" else system_prompt + self.sep + "\n"
-        for role, message in self.messages:
-            if message:
-                yield role + "\n", message + self.sep + "\n"
-            else:
-                yield role + "\n", ""
-        return
-    if self.sep_style == SeparatorStyle.CHATGLM3:
-        if self.system_message:
-            yield "", system_prompt
-        for role, message in self.messages:
-            if message:
-                yield role + "\n", " " + message
-            else:
-                yield role
-        return
-    if self.sep_style == SeparatorStyle.CHATINTERN:
-        # source: https://huggingface.co/internlm/internlm-chat-7b-8k/blob/bd546fa984b4b0b86958f56bf37f94aa75ab8831/modeling_internlm.py#L771
-        seps = [self.sep, self.sep2]
-        yield "", system_prompt
-        for i, (role, message) in enumerate(self.messages):
-            prefix = "<s>" if i % 2 == 0 else ""
-            if message:
-                yield prefix + role + ":", message + seps[i % 2] + "\n"
-            else:
-                yield role + ":", ""
-        return
-    if self.sep_style == SeparatorStyle.DOLLY:
-        seps = [self.sep, self.sep2]
-        yield "", system_prompt
-        for i, (role, message) in enumerate(self.messages):
-            if message:
-                suffix = "\n\n" if i % 2 == 1 else ""
-                yield role + ":\n", message + seps[i % 2] + suffix
-            else:
-                yield role + ":\n", ""
-        return
-    if self.sep_style == SeparatorStyle.PHOENIX:
-        yield "", system_prompt
-        for role, message in self.messages:
-            if message:
-                yield role + ": ", "<s>" + message + "</s>"
-            else:
-                yield role + ": " + "<s>", ""
-        return
-    if self.sep_style == SeparatorStyle.ROBIN:
-        yield "", system_prompt + self.sep
-        for role, message in self.messages:
-            if message:
-                yield role + ":\n", message + self.sep
-            else:
-                yield role + ":\n", ""
-        return
-    if self.sep_style == SeparatorStyle.FALCON_CHAT:
-        if self.system_message:
-            yield "", system_prompt + self.sep
-        for role, message in self.messages:
-            if message:
-                yield role + ": ", message + self.sep
-            else:
-                yield role + ":", ""
-    else:
-        raise ValueError(f"Invalid style: {self.sep_style}")
-
-
-def add_get_turns_to_conversation():
-    import fastchat.conversation
-
-    fastchat.conversation.Conversation.get_turns = get_turns
-    fastchat.conversation.Conversation.get_prompt = get_prompt

src/axolotl/prompt_strategies/instruct.py

@@ -1,33 +0,0 @@
-"""Module containing the InstructShareGPTPromptTokenizingStrategy class"""
-from typing import Any, Dict, Optional
-
-from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
-from axolotl.prompters import ShareGPTPrompterV2
-
-
-def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
-    conversation = (
-        ds_cfg["conversation"] if ds_cfg and "conversation" in ds_cfg else None
-    )
-    strategy = InstructShareGPTPromptTokenizingStrategy(
-        # pylint: disable=duplicate-code
-        ShareGPTPrompterV2(
-            conversation=conversation,
-        ),
-        tokenizer,
-        cfg.train_on_inputs,
-        cfg.sequence_len,
-    )
-    return strategy
-
-
-class InstructShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
-    """
-    basic sharegpt strategy to grab conversations from the sample row
-    """
-
-    def get_conversation_thread(self, prompt):
-        return [
-            {"from": "human", "value": prompt["instruction"]},
-            {"from": "gpt", "value": prompt["output"]},
-        ]

src/axolotl/prompt_strategies/llama2_chat.py

@@ -29,7 +29,7 @@ from dataclasses import dataclass, field
 from typing import Generator, List, Sequence
 
 from axolotl.prompt_tokenizers import PromptTokenizingStrategy
-from axolotl.prompters import IGNORE_TOKEN_ID, SHAREGPT_ASSERTION_FAILED_ROLE
+from axolotl.prompters import ALTERNATING_ASSERTION_FAILED_ROLE, IGNORE_TOKEN_ID
 
 
 @dataclass
@@ -75,7 +75,7 @@ class Llama2ChatConversation:
 
 class LLama2ChatTokenizingStrategy(PromptTokenizingStrategy):
     """
-    Tokenizing strategy for ShareGPT prompts.
+    Tokenizing strategy for Llama2 prompts.
     adapted from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/train.py
     """
 
@@ -191,7 +191,7 @@ class Llama2ChatPrompter:  # pylint: disable=too-few-public-methods
         conv.messages = []  # pylint: disable=R0801
         for j, sentence in enumerate(source):
             role = roles[sentence["from"]]
-            assert role == conv.roles[j % 2], SHAREGPT_ASSERTION_FAILED_ROLE
+            assert role == conv.roles[j % 2], ALTERNATING_ASSERTION_FAILED_ROLE
             if sentence["value"]:
                 conv.append_message(role, sentence["value"])
         yield conv

src/axolotl/prompt_strategies/sharegpt.py

@@ -1,223 +0,0 @@
-"""Module containing the SimpleShareGPTPromptTokenizingStrategy class"""
-
-import logging
-from typing import Any, Dict, Optional, Type
-
-from fastchat.conversation import Conversation, SeparatorStyle, register_conv_template
-
-from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
-from axolotl.prompters import ShareGPTPrompterV2
-from axolotl.utils.tokenization import (
-    chatml_to_conversation,
-    merge_consecutive_messages,
-)
-
-LOG = logging.getLogger("axolotl")
-
-
-def register_chatml_template(system_message=None):
-    system_message = system_message or "You are a helpful assistant."
-    register_conv_template(
-        Conversation(
-            name="chatml",
-            system_template="<|im_start|>system\n{system_message}",
-            system_message=system_message,
-            roles=("<|im_start|>user", "<|im_start|>assistant"),
-            sep_style=SeparatorStyle.CHATML,
-            sep="<|im_end|>",
-        )
-    )
-    register_conv_template(
-        Conversation(
-            name="chatml_glaive",
-            system_template="<|im_start|>system\n{system_message}",
-            system_message=system_message,
-            roles=("<|im_start|>user", "<|im_start|>assistant", "<|im_start|>tool"),
-            sep_style=SeparatorStyle.CHATML,
-            sep="<|im_end|>",
-        )
-    )
-
-
-def register_llama3_template(system_message=None):
-    system_message = system_message or "You are a helpful assistant."
-    register_conv_template(
-        Conversation(
-            name="llama3",
-            system_template="<|start_header_id|>system<|end_header_id|>\n\n{system_message}<|eot_id|>",
-            system_message=system_message,
-            roles=("user", "assistant"),
-            sep_style=SeparatorStyle.LLAMA3,
-            sep="",
-            stop_str="<|eot_id|>",
-            stop_token_ids=[128001, 128009],
-        )
-    )
-
-
-def build_loader(
-    tokenization_strategy_cls: Type["ShareGPTPromptTokenizingStrategy"],
-    prompter_cls: Type["ShareGPTPrompterV2"],
-    default_conversation: Optional[str] = None,
-):
-    def _load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
-        LOG.warning(
-            "sharegpt type support will be deprecated in the next release of Axolotl. Please use chat_template instead. https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html#chat_template",
-        )
-        conversation = (
-            ds_cfg["conversation"]
-            if ds_cfg and "conversation" in ds_cfg
-            else default_conversation
-        )
-        field_human = (
-            ds_cfg["field_human"] if ds_cfg and "field_human" in ds_cfg else None
-        )
-        field_model = (
-            ds_cfg["field_model"] if ds_cfg and "field_model" in ds_cfg else None
-        )
-        roles = ds_cfg["roles"].to_dict() if ds_cfg and "roles" in ds_cfg else None
-        strategy = tokenization_strategy_cls(
-            prompter_cls(
-                conversation=conversation,
-                role_key_model=field_model,
-                role_key_human=field_human,
-                roles=roles,
-            ),
-            tokenizer,
-            cfg.train_on_inputs,
-            cfg.sequence_len,
-        )
-        if ds_cfg and "strict" in ds_cfg and hasattr(strategy, "strict"):
-            strategy.strict = ds_cfg["strict"]
-        if ds_cfg and "field_messages" in ds_cfg and hasattr(strategy, "messages"):
-            strategy.messages = ds_cfg["field_messages"]
-        return strategy
-
-    return _load
-
-
-class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
-    """
-    basic sharegpt strategy to grab conversations from the sample row
-    """
-
-    _strict = False
-    _messages = "conversations"
-
-    @property
-    def strict(self):
-        return self._strict
-
-    @strict.setter
-    def strict(self, strict):
-        self._strict = strict
-
-    @property
-    def messages(self):
-        return self._messages
-
-    @messages.setter
-    def messages(self, messages):
-        self._messages = messages
-
-    def get_conversation_thread(self, prompt):
-        conversations = prompt[self.messages]
-        if self.strict:
-            return conversations
-        role_key = "from"
-        if "role" in conversations[0].keys():
-            role_key = "role"
-        value_key = "value"
-        if "text" in conversations[0].keys():
-            value_key = "text"
-        elif "content" in conversations[0].keys():
-            value_key = "content"
-        # remap roles - allow for assistant turn"
-        role_map = {
-            "user": "human",
-            "human": "human",
-            "assistant": "gpt",
-            "gpt": "gpt",
-            "system": "system",
-        }
-        turns = [
-            {
-                "from": (
-                    role_map[t[role_key]] if t[role_key] in role_map else t[role_key]
-                ),
-                "value": t[value_key],
-                "weight": 1
-                if "weight" not in t or t["weight"] is None
-                else t["weight"],
-            }
-            for t in conversations
-        ]
-        return turns
-
-
-class SimpleRoleShareGPTPromptTokenizingStrategy(
-    SimpleShareGPTPromptTokenizingStrategy
-):
-    """
-    basic sharegpt strategy to grab conversations from the sample row, but uses role instead of from
-    """
-
-    def get_conversation_thread(self, prompt):
-        conversations = prompt["conversations"]
-        # remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
-        turns = [{"from": t["role"], "value": t["value"]} for t in conversations]
-        return turns
-
-
-class GuanacoShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
-    """
-    sharegpt strategy that remaps oasst data to sharegpt format
-    """
-
-    def get_conversation_thread(self, prompt):
-        conversations = prompt["conversations"]
-        # remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
-        role_map = {"prompter": "human", "assistant": "gpt"}
-        turns = [
-            {"from": role_map[t["role"]], "value": t["text"]} for t in conversations
-        ]
-        return turns
-
-
-class UltrachatShareGPTPromptTokenizingStrategy(SimpleShareGPTPromptTokenizingStrategy):
-    """
-    sharegpt strategy that remaps ultrachat data to sharegpt format
-    """
-
-    def get_conversation_thread(self, prompt):
-        conversations = prompt["messages"]
-        role_map = {"user": "human", "assistant": "gpt"}
-        turns = [
-            {"from": role_map[t["role"]], "value": t["content"]} for t in conversations
-        ]
-        return turns
-
-
-class GlaiveShareGPTPromptTokenizingStrategy(SimpleShareGPTPromptTokenizingStrategy):
-    """
-    sharegpt strategy that remaps glaive data to sharegpt format
-    """
-
-    def get_conversation_thread(self, prompt):
-        conversation = chatml_to_conversation(prompt)
-        conversation = merge_consecutive_messages(conversation)
-
-        return conversation
-
-
-load = build_loader(SimpleShareGPTPromptTokenizingStrategy, ShareGPTPrompterV2)
-load_role = build_loader(SimpleRoleShareGPTPromptTokenizingStrategy, ShareGPTPrompterV2)
-load_ultrachat = build_loader(
-    UltrachatShareGPTPromptTokenizingStrategy, ShareGPTPrompterV2
-)
-load_guanaco = build_loader(GuanacoShareGPTPromptTokenizingStrategy, ShareGPTPrompterV2)
-load_glaive = build_loader(
-    GlaiveShareGPTPromptTokenizingStrategy,
-    ShareGPTPrompterV2,
-    default_conversation="chatml_glaive",
-)

src/axolotl/prompt_strategies/sharegpt_jokes.py

@@ -1,28 +0,0 @@
-"""Module for Jokes prompts using sharegpt style """
-from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
-from axolotl.prompters import ShareGPTPrompterV2
-
-
-def load(tokenizer, cfg):
-    return SimpleJokesShareGPTPromptTokenizingStrategy(
-        ShareGPTPrompterV2(),
-        tokenizer,
-        cfg.train_on_inputs,
-        cfg.sequence_len,
-    )
-
-
-class SimpleJokesShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
-    """
-    Tokenization strategy for asking bot to tell a joke and then explain why its funny
-    """
-
-    # title, text, explanation
-    def get_conversation_thread(self, prompt):
-        title = "" if not prompt["title"] else prompt["title"] + " "
-        return [
-            {"from": "human", "value": "Tell me a joke."},
-            {"from": "gpt", "value": title + prompt["text"]},
-            {"from": "human", "value": "Why is that joke funny?"},
-            {"from": "gpt", "value": prompt["explanation"]},
-        ]

src/axolotl/prompt_tokenizers.py

@@ -1,17 +1,12 @@
 """Module containing PromptTokenizingStrategy and Prompter classes"""
 
 import abc
-import copy
 import logging
 from typing import Dict, List, Tuple, Union
 
-from fastchat.conversation import Conversation
 from transformers import BatchEncoding, PreTrainedTokenizer
 
-from axolotl.monkeypatch.fastchat_conversation_turns import (
-    add_get_turns_to_conversation,
-)
-from axolotl.prompters import IGNORE_TOKEN_ID, Prompter
+from axolotl.prompters import Prompter
 
 LOG = logging.getLogger("axolotl")
 
@@ -21,8 +16,6 @@ LLAMA_DEFAULT_EOS_TOKEN = "</s>"  # nosec
 LLAMA_DEFAULT_BOS_TOKEN = "<s>"  # nosec
 LLAMA_DEFAULT_UNK_TOKEN = "<unk>"  # nosec
 
-add_get_turns_to_conversation()
-
 
 class InvalidDataException(Exception):
     """
@@ -331,154 +324,6 @@ class AlpacaReflectionPTStrategy(ReflectionPromptTokenizingStrategy):
         )
 
 
-class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
-    """
-    Tokenizing strategy for ShareGPT prompts.
-    """
-
-    def get_conversation_thread(self, prompt):
-        return prompt["conversations"]
-
-    def tokenize_prompt(self, prompt):
-        # Initial values. We will append to these as we go through the conversation.
-        result, current_len = tokenize_prompt_default()
-        conversation: Conversation = (
-            self.prompter._conversation.copy()  # pylint: disable=protected-access
-        )
-
-        input_roles = {conversation.roles[0]}
-        output_roles = {conversation.roles[1]}
-
-        if len(conversation.roles) == 3:
-            tool_role_label = conversation.roles[2]
-            input_roles.add(tool_role_label)
-
-        # Add roles from the config
-        if self.prompter.roles:
-            if "input" in self.prompter.roles and self.prompter.roles["input"]:
-                for role in self.prompter.roles["input"]:
-                    input_roles.add(role)
-
-            if "output" in self.prompter.roles and self.prompter.roles["output"]:
-                for role in self.prompter.roles["output"]:
-                    output_roles.add(role)
-
-        # support for custom roles from the dataset, only useful for vicuna style prompts/roles
-        role_remap = []
-        if (
-            conversation.name == "vicuna_v1.1"
-            and "roles" in prompt
-            and len(prompt["roles"]) >= 2
-        ):
-            role_remap = [
-                {"from": conversation.roles[0], "to": prompt["roles"][0]},
-                {"from": conversation.roles[1], "to": prompt["roles"][1]},
-            ]
-
-        try:
-            for _, part in enumerate(
-                self.prompter.build_prompt(self.get_conversation_thread(prompt))
-            ):
-                if not isinstance(part, tuple):
-                    LOG.warning(f"expected tuple, got {part}")
-                    continue
-
-                if len(part) <= 2:
-                    role, content = part
-                    weight = 1
-                else:
-                    role, content, weight = part
-
-                # Uses "in" because role contains extra characters
-                input_turn = any(r.lower() in role.lower() for r in input_roles)
-                output_turn = any(r.lower() in role.lower() for r in output_roles)
-                empty_role = role.strip() == ""
-
-                if not any([input_turn, output_turn, empty_role]):
-                    LOG.warning(f"unhandled role: {role}")
-                    continue
-
-                if input_turn:
-                    role = (
-                        role.replace(role_remap[0]["from"], role_remap[0]["to"])
-                        if role_remap
-                        else role
-                    )
-                    turn = role + content
-                    # this is still the user query, we should
-                    if not content.strip():
-                        LOG.warning(f"user turn has empty text: {prompt}")
-                    res = self._tokenize(
-                        turn,
-                        add_eos_token=False,
-                        strip_bos_token=True,
-                    )
-                    if self.train_on_inputs and weight == 1:
-                        labels = copy.deepcopy(res["input_ids"])
-                    else:
-                        # everything from this is masked out from the labels
-                        labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
-                elif output_turn:
-                    role = (
-                        role.replace(role_remap[1]["from"], role_remap[1]["to"])
-                        if role_remap
-                        else role
-                    )
-                    turn = role + content
-                    # this should be the assistant response, should end with an eos token
-                    if not content.strip():
-                        LOG.warning(f"assistant turn has empty text: {prompt}")
-                    add_eos_token = not (
-                        conversation.name == "chatml"
-                        and conversation.sep == self.tokenizer.eos_token
-                    )
-                    res = self._tokenize(
-                        turn,
-                        add_eos_token=add_eos_token,
-                        strip_bos_token=True,
-                    )
-                    role_res = self._tokenize(
-                        role.rstrip(),
-                        add_eos_token=False,
-                        strip_bos_token=True,
-                    )
-                    labels = copy.deepcopy(res["input_ids"])
-                    if not self.train_on_inputs:
-                        # mask out role tokens from the labels
-                        len_role = len(role_res["input_ids"])
-                        labels[:len_role] = [IGNORE_TOKEN_ID] * min(
-                            len_role, len(labels)
-                        )
-                    if weight == 0:
-                        # everything from this is masked out from the labels
-                        # (role is masked out too because it makes no sense if contents is masked out)
-                        labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
-
-                elif empty_role:
-                    turn = content
-                    # this is only ever the first part, should include the bos token and the user query
-                    res = self._tokenize(
-                        turn, add_eos_token=False, strip_bos_token=False
-                    )
-                    if self.train_on_inputs and weight == 1:
-                        labels = copy.deepcopy(res["input_ids"])
-                    else:
-                        # everything from this is masked out from the labels
-                        labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
-
-                # pylint: disable=duplicate-code
-                result, current_len = parse_tokenized_to_result(
-                    result,
-                    current_len,
-                    res,
-                    labels,
-                    pad_token_id=self.tokenizer.pad_token_id,
-                )
-            return result
-        except (KeyError, AssertionError, IndexError) as err:
-            raise InvalidDataException(str(err)) from err
-
-
 def tokenize_prompt_default() -> Tuple[Dict[str, List[int]], int]:
     """
     Returns the default values for the tokenize prompt function

src/axolotl/prompters.py

@@ -5,7 +5,6 @@ from enum import Enum
 from typing import Generator, Optional, Union
 
 from colorama import Fore
-from fastchat.conversation import Conversation, get_conv_template
 
 LOG = logging.getLogger("axolotl")
 IGNORE_TOKEN_ID = -100
@@ -262,166 +261,10 @@ class ReflectAlpacaPrompter(Prompter):
         )
 
 
-SHAREGPT_ASSERTION_FAILED_ROLE = (
+ALTERNATING_ASSERTION_FAILED_ROLE = (
     "Role did not alternate between turns (gpt and human). Please check your data."
 )
 
-CONVERSATION_ROLE_FORMAT = {
-    "chatml": "<|im_start|>{ROLE}",
-    "zephyr": "<|{ROLE}|>",
-    "vicuna_v1.1": "{ROLE}",
-    "llama3": "<|start_header_id|>{ROLE}<|end_header_id|>",
-}
-
-
-class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
-    """
-    A prompter that generates prompts for the ShareGPT
-    """
-
-    role_key_human = "human"
-    role_key_model = "gpt"
-    # Optional, only used for tool usage datasets.
-    role_key_tool: Optional[str] = None
-    # Optional, role input/output mapping
-    roles: Optional[dict] = None
-
-    def __init__(
-        self,
-        prompt_style=None,  # pylint: disable=unused-argument
-        conversation: Optional[Union[str, Conversation]] = None,
-        role_key_human: Optional[str] = None,
-        role_key_model: Optional[str] = None,
-        role_key_tool: Optional[str] = None,
-        roles: Optional[dict] = None,
-    ):
-        if conversation:
-            if isinstance(conversation, Conversation):
-                self._conversation = conversation
-            else:
-                self._conversation = get_conv_template(conversation)
-        else:
-            self._conversation = get_conv_template("vicuna_v1.1")
-        if role_key_human:
-            self.role_key_human = role_key_human
-        if role_key_model:
-            self.role_key_model = role_key_model
-        if role_key_tool:
-            self.role_key_tool = role_key_tool
-        if roles:
-            self.roles = roles
-
-    def _build_result(self, source):
-        if len(source) < 2:
-            # If there isn't a back and forth conversation, ignore it
-            # also happens on the data splitting leaving empty conversations
-            raise IndexError(
-                f"A conversation entry has less than 2 messages :\n{source}"
-            )
-
-        conv = self._conversation.copy()
-
-        original_source = source.copy()
-        # Add the conversation system prompt if provided, otherwise use the default one
-        if source[0]["from"] == "system":
-            conv.set_system_message(source[0]["value"])
-            source.pop(0)
-
-        roles = {self.role_key_human: conv.roles[0], self.role_key_model: conv.roles[1]}
-        if self.role_key_tool:
-            roles[self.role_key_tool] = conv.roles[2]
-
-        try:
-            # Apply prompt templates
-            if source[0]["from"] not in roles:
-                # Skip the first one if it is not from human
-                source = source[1:]
-        except IndexError as err:
-            # sometimes there is a bing or system chat
-            raise err
-
-        conv.messages = []
-        for _, sentence in enumerate(source):
-            from_role = sentence["from"]
-            if from_role in roles:
-                role = roles[from_role]
-            else:
-                if self._conversation.name not in CONVERSATION_ROLE_FORMAT:
-                    raise NotImplementedError(
-                        f"Role ({role}) not in default roles, and {self._conversation.name} does not support role remapping yet."
-                        "Please help us by creating an Issue to add support for this conversation type."
-                    )
-
-                if self._conversation.name in ["llama3"]:
-                    role = from_role
-                else:
-                    role = CONVERSATION_ROLE_FORMAT[self._conversation.name].format(
-                        ROLE=from_role
-                    )
-
-            if len(conv.messages) > 0 and ((role == conv.messages[-1][0])):
-                if (
-                    role != "assistant"
-                ):  # back to back assistant calls may be okay for tool calls
-                    LOG.warning(f"{SHAREGPT_ASSERTION_FAILED_ROLE}: {sentence}")
-
-            conv.append_message(role, sentence["value"])
-        turns = list(conv.get_turns())
-        original_source_length = len(original_source)
-        assert len(turns) in [
-            original_source_length - 1,
-            original_source_length,
-            original_source_length + 1,
-        ]
-        if len(turns) == original_source_length + 1:
-            original_source = [{"weight": None}] + original_source
-        elif len(turns) == original_source_length - 1:
-            original_source = original_source[1:]
-        return [
-            (*turn, weight)
-            for turn, weight in zip(
-                turns,
-                [
-                    1 if "weight" not in e or e["weight"] is None else e["weight"]
-                    for e in original_source
-                ],
-            )
-        ]
-
-    def build_prompt(self, source) -> Generator[str, None, None]:
-        turns = self._build_result(source)
-
-        for part in turns:
-            if part[0] and not part[1]:
-                LOG.warning(f"role with empty message: {part[0]}")
-            yield part
-
-    def __repr__(self) -> str:
-        turns = self._build_result([{"from": "{from}", "value": "{value}"}])
-        return "\n".join([REPR_TEMPLATE.format(full_prompt=part) for part in turns])
-
-
-class ShareGPTPrompterV2(ShareGPTPrompter):
-    """
-    A V2 prompter that generates prompts for the ShareGPT
-    """
-
-    def __init__(
-        self,
-        conversation: Optional[Union[str, Conversation]] = None,
-        role_key_human: Optional[str] = None,
-        role_key_model: Optional[str] = None,
-        role_key_tool: Optional[str] = None,
-        roles: Optional[dict] = None,
-    ):
-        super().__init__(
-            conversation=conversation,
-            role_key_human=role_key_human,
-            role_key_model=role_key_model,
-            role_key_tool=role_key_tool,
-            roles=roles,
-        )
-
 
 class UnsupportedPrompter(Prompter):
     """

src/axolotl/utils/config/__init__.py

@@ -215,11 +215,6 @@ def normalize_cfg_datasets(cfg):
     if cfg.chat_template:
         if cfg.datasets:
             for idx, ds_cfg in enumerate(cfg.datasets):
-                if ds_cfg.type == "sharegpt" and not ds_cfg.conversation:
-                    LOG.info(
-                        f"updating dataset {ds_cfg.path} with `conversation: {cfg.chat_template}` to match your chat_template"
-                    )
-                    cfg.datasets[idx].conversation = cfg.chat_template
                 if (
                     ds_cfg.type in ["orpo.chat_template", "chat_template"]
                     and not ds_cfg.chat_template
@@ -461,27 +456,6 @@ def legacy_validate_config(cfg):
                 "`early_stopping_patience` requires that eval_steps should evenly divide save_steps."
             )
 
-    if cfg.datasets:
-        for idx, ds_cfg in enumerate(cfg.datasets):
-            if not ds_cfg.type:
-                continue
-            if ds_cfg.type == "sharegpt:chat":
-                LOG.warning(
-                    PendingDeprecationWarning(
-                        "`type: sharegpt:chat` will soon be deprecated. simply use `type: sharegpt` instead."
-                    )
-                )
-                cfg.datasets[idx].type = "sharegpt"
-            if "sharegpt_simple" in ds_cfg.type:
-                LOG.warning(
-                    PendingDeprecationWarning(
-                        "`type: sharegpt_simple` will soon be deprecated. simply use `type: sharegpt` instead."
-                    )
-                )
-                cfg.datasets[idx].type = cfg.datasets[idx].type.replace(
-                    "sharegpt_simple", "sharegpt"
-                )
-
     if cfg.saves_per_epoch and cfg.save_steps:
         raise ValueError(
             "save_steps and saves_per_epoch are mutually exclusive and cannot be used together."

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

@@ -57,6 +57,7 @@ class ChatTemplate(str, Enum):
     jinja = "jinja"  # pylint: disable=invalid-name
     qwen_25 = "qwen_25"  # pylint: disable=invalid-name
     tokenizer_default = "tokenizer_default"  # pylint: disable=invalid-name
+    exaone = "exaone"  # pylint: disable=invalid-name
 
 
 class DeprecatedParameters(BaseModel):
@@ -587,6 +588,9 @@ class AxolotlInputConfig(
 
     rl: Optional[RLType] = None
     reward_model: Optional[bool] = None
+    dpo_use_weighting: Optional[
+        bool
+    ] = None  # whether to use weighting in DPO trainer. If none, default is false in the trainer.
 
     datasets: Optional[conlist(Union[SFTDataset, DPODataset, KTODataset], min_length=1)] = None  # type: ignore
     test_datasets: Optional[conlist(Union[SFTDataset, DPODataset, KTODataset], min_length=1)] = None  # type: ignore
@@ -779,26 +783,16 @@ class AxolotlInputConfig(
 
     @field_validator("datasets", mode="before")
     @classmethod
-    def fix_sharegpt_datasets(cls, datasets):
-        for idx, ds_cfg in enumerate(datasets):
-            if not ds_cfg["type"]:
+    def deprecate_sharegpt_datasets(cls, datasets):
+        for _, ds_cfg in enumerate(datasets):
+            if not ds_cfg.get("type"):
                 continue
-            if ds_cfg["type"] == "sharegpt:chat":
-                LOG.warning(
-                    PendingDeprecationWarning(
-                        "`type: sharegpt:chat` will soon be deprecated. simply use `type: sharegpt` instead."
-                    )
-                )
-                datasets[idx]["type"] = "sharegpt"
-            if "sharegpt_simple" in ds_cfg["type"]:
-                LOG.warning(
-                    PendingDeprecationWarning(
-                        "`type: sharegpt_simple` will soon be deprecated. simply use `type: sharegpt` instead."
-                    )
-                )
-                datasets[idx]["type"] = datasets[idx]["type"].replace(
-                    "sharegpt_simple", "sharegpt"
+
+            if ds_cfg["type"].startswith("sharegpt"):
+                raise ValueError(
+                    "`type: sharegpt.*` is deprecated. Please use `type: chat_template` instead."
                 )
+
         return datasets
 
     @model_validator(mode="before")

src/axolotl/utils/data/sft.py

@@ -2,9 +2,11 @@
 
 import functools
 import logging
+import time
 from pathlib import Path
 from typing import List, Optional, Tuple, Union
 
+import requests
 from datasets import (
     Dataset,
     DatasetDict,
@@ -53,6 +55,28 @@ from axolotl.utils.trainer import (
 LOG = logging.getLogger("axolotl")
 
 
+def retry_on_request_exceptions(max_retries=3, delay=1):
+    def decorator(func):
+        @functools.wraps(func)
+        def wrapper(*args, **kwargs):  # pylint: disable=inconsistent-return-statements
+            for attempt in range(max_retries):
+                try:
+                    return func(*args, **kwargs)
+                except (
+                    requests.exceptions.ReadTimeout,
+                    requests.exceptions.ConnectionError,
+                ) as exc:
+                    if attempt < max_retries - 1:
+                        time.sleep(delay)
+                    else:
+                        raise exc
+
+        return wrapper
+
+    return decorator
+
+
+@retry_on_request_exceptions(max_retries=3, delay=5)
 def prepare_dataset(cfg, tokenizer, processor=None):
     prompters = []
     if not cfg.pretraining_dataset:

src/axolotl/utils/models.py

@@ -640,9 +640,7 @@ class ModelLoader:
                 self.model_kwargs["quantization_config"] = BitsAndBytesConfig(
                     **self.model_config.quantization_config
                 )
-        elif self.cfg.adapter == "qlora" and (
-            "load_in_4bit" in self.model_kwargs and self.model_kwargs["load_in_4bit"]
-        ):
+        elif self.cfg.adapter == "qlora" and self.model_kwargs["load_in_4bit"]:
             bnb_config = {
                 "load_in_4bit": True,
                 "llm_int8_threshold": 6.0,
@@ -665,9 +663,7 @@ class ModelLoader:
             self.model_kwargs["quantization_config"] = BitsAndBytesConfig(
                 **bnb_config,
             )
-        elif self.cfg.adapter == "lora" and (
-            "load_in_8bit" in self.model_kwargs and self.model_kwargs["load_in_8bit"]
-        ):
+        elif self.cfg.adapter == "lora" and self.model_kwargs["load_in_8bit"]:
             bnb_config = {
                 "load_in_8bit": True,
             }
@@ -680,10 +676,8 @@ class ModelLoader:
 
         # no longer needed per https://github.com/huggingface/transformers/pull/26610
         if "quantization_config" in self.model_kwargs or self.cfg.gptq:
-            if "load_in_8bit" in self.model_kwargs:
-                del self.model_kwargs["load_in_8bit"]
-            if "load_in_4bit" in self.model_kwargs:
-                del self.model_kwargs["load_in_4bit"]
+            self.model_kwargs.pop("load_in_8bit", None)
+            self.model_kwargs.pop("load_in_4bit", None)
 
     def set_attention_config(self) -> None:
         """
@@ -968,17 +962,10 @@ class ModelLoader:
         if is_deepspeed_zero3_enabled():
             skip_prepare_model_for_kbit_training = True
 
-        is_load_in_8bit = (
-            "load_in_8bit" in self.model_kwargs and self.model_kwargs["load_in_8bit"]
-        )
-        is_load_in_4bit = (
-            "load_in_4bit" in self.model_kwargs and self.model_kwargs["load_in_4bit"]
-        )
-
         if (
             not skip_prepare_model_for_kbit_training
             and self.cfg.adapter in ["lora", "qlora"]
-            and (is_load_in_8bit or is_load_in_4bit)
+            and (self.cfg.load_in_8bit or self.cfg.load_in_4bit)
         ):
             LOG.info("converting PEFT model w/ prepare_model_for_kbit_training")
             self.model = prepare_model_for_kbit_training(
@@ -1116,16 +1103,10 @@ class ModelLoader:
         # ---------------------------------------------------------
         #  put model to accelerator
         # ---------------------------------------------------------
-        is_load_in_8bit = (
-            "load_in_8bit" in self.model_kwargs and self.model_kwargs["load_in_8bit"]
-        )
-        is_load_in_4bit = (
-            "load_in_4bit" in self.model_kwargs and self.model_kwargs["load_in_4bit"]
-        )
         if (
             self.cfg.ddp
-            and not is_load_in_8bit
-            and not (self.cfg.rl and is_load_in_4bit)
+            and not self.cfg.load_in_8bit
+            and not (self.cfg.rl and self.cfg.load_in_4bit)
             and not skip_move_to_device
         ):
             # TODO revaldate this conditional

src/axolotl/utils/tokenization.py

@@ -1,8 +1,6 @@
 """Module for tokenization utilities"""
 
 import logging
-import re
-from typing import Dict, List
 
 from termcolor import colored
 
@@ -93,65 +91,3 @@ def check_rl_example_labels(example, tokenizer, text_only=False):
     LOG.info(f"REJECTED RESPONSE: {delimiter.join(colored_rejecteds)}\n\n\n")
 
     return delimiter.join(colored_tokens)
-
-
-GLAIVE_ROLES = ["USER", "ASSISTANT", "FUNCTION RESPONSE"]
-GLAIVE_TO_SHAREGPT_ROLE = {
-    "SYSTEM": "system",
-    "USER": "human",
-    "ASSISTANT": "gpt",
-    "FUNCTION RESPONSE": "tool",
-}
-
-GLAIVE_MSG_REGEX = re.compile(rf"({'|'.join(GLAIVE_ROLES)}): ")
-
-
-def chatml_to_conversation(row: Dict[str, str]) -> List[Dict[str, str]]:
-    """
-    Converts a ChatML formatted row to a list of messages in ShareGPT format.
-    Initially based off https://github.com/lilacai/lilac/blob/main/notebooks/GlaiveToShareGPT.ipynb.
-    """
-
-    system_prompt = row.get("system")
-    if system_prompt:
-        system_prompt = system_prompt.removeprefix("SYSTEM: ")
-
-    chat_str = row["chat"]
-    chat_msgs = [s.strip() for s in GLAIVE_MSG_REGEX.split(chat_str) if s]
-
-    chat_msg_dicts = [
-        {"from": GLAIVE_TO_SHAREGPT_ROLE[role], "value": value}
-        for role, value in zip(chat_msgs[::2], chat_msgs[1::2])
-    ]
-
-    if system_prompt:
-        chat_msg_dicts = [
-            {"from": GLAIVE_TO_SHAREGPT_ROLE["SYSTEM"], "value": system_prompt}
-        ] + chat_msg_dicts
-
-    return chat_msg_dicts
-
-
-def merge_consecutive_messages(messages):
-    """
-    Merge consecutive messages from the same sender into a single message.
-    This can be useful with datasets that contain multiple consecutive tool calls.
-    """
-
-    merged_messages = []
-    current_from = None
-    current_message = ""
-
-    for msg in messages:
-        if current_from == msg["from"]:
-            current_message += msg["value"]
-        else:
-            if current_from is not None:
-                merged_messages.append({"from": current_from, "value": current_message})
-            current_from = msg["from"]
-            current_message = msg["value"]
-
-    if current_from is not None:
-        merged_messages.append({"from": current_from, "value": current_message})
-
-    return merged_messages

tests/e2e/integrations/liger.py

@@ -1,7 +1,6 @@
 """
 Simple end-to-end test for Liger integration
 """
-
 import unittest
 from pathlib import Path
 

tests/e2e/multigpu/test_llama.py

@@ -14,7 +14,7 @@ from huggingface_hub import snapshot_download
 
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import is_hopper, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
 os.environ["WANDB_DISABLED"] = "true"
@@ -59,7 +59,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -116,7 +116,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 50,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -144,6 +144,146 @@ class TestMultiGPULlama(unittest.TestCase):
             ]
         )
 
+    @pytest.mark.skipif(is_hopper(), reason="h100 doesn't support 8-bit lora")
+    @with_temp_dir
+    def test_dpo_lora_ddp(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "TinyLlama/TinyLlama_v1.1",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 2048,
+                "sample_packing": False,
+                "eval_sample_packing": False,
+                "pad_to_sequence_len": True,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.05,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "rl": "dpo",
+                "chat_template": "llama3",
+                "datasets": [
+                    {
+                        "path": "fozziethebeat/alpaca_messages_2k_dpo_test",
+                        "type": "chat_template.default",
+                        "field_messages": "conversation",
+                        "field_chosen": "chosen",
+                        "field_rejected": "rejected",
+                        "message_field_role": "role",
+                        "message_field_content": "content",
+                        "roles": {
+                            "system": ["system"],
+                            "user": ["user"],
+                            "assistant": ["assistant"],
+                        },
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 15,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "warmup_steps": 0,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+            }
+        )
+
+        # write cfg to yaml file
+        Path(temp_dir).mkdir(parents=True, exist_ok=True)
+        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
+            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
+
+        execute_subprocess_async(
+            [
+                "accelerate",
+                "launch",
+                "--num-processes",
+                "2",
+                "-m",
+                "axolotl.cli.train",
+                str(Path(temp_dir) / "config.yaml"),
+            ]
+        )
+
+    @with_temp_dir
+    def test_dpo_qlora_ddp(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM-135M",
+                "sequence_len": 2048,
+                "sample_packing": False,
+                "eval_sample_packing": False,
+                "pad_to_sequence_len": True,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.05,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "rl": "dpo",
+                "chat_template": "chatml",
+                "datasets": [
+                    {
+                        "path": "fozziethebeat/alpaca_messages_2k_dpo_test",
+                        "type": "chat_template.default",
+                        "field_messages": "conversation",
+                        "field_chosen": "chosen",
+                        "field_rejected": "rejected",
+                        "message_field_role": "role",
+                        "message_field_content": "content",
+                        "roles": {
+                            "system": ["system"],
+                            "user": ["user"],
+                            "assistant": ["assistant"],
+                        },
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 15,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "warmup_steps": 0,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+            }
+        )
+
+        # write cfg to yaml file
+        Path(temp_dir).mkdir(parents=True, exist_ok=True)
+        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
+            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
+
+        execute_subprocess_async(
+            [
+                "accelerate",
+                "launch",
+                "--num-processes",
+                "2",
+                "-m",
+                "axolotl.cli.train",
+                str(Path(temp_dir) / "config.yaml"),
+            ]
+        )
+
     @with_temp_dir
     def test_fsdp(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -165,7 +305,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -231,7 +371,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -273,7 +413,6 @@ class TestMultiGPULlama(unittest.TestCase):
             ]
         )
 
-    @pytest.mark.skip("disabled due to upstream issue")
     @with_temp_dir
     def test_fsdp_qlora_prequant_packed(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -282,6 +421,7 @@ class TestMultiGPULlama(unittest.TestCase):
                 "base_model": "axolotl-ai-co/TinyLlama_v1.1-bnb-nf4-bf16",
                 "tokenizer_type": "AutoTokenizer",
                 "adapter": "qlora",
+                "mean_resizing_embeddings": True,
                 "load_in_4bit": True,
                 "lora_r": 8,
                 "lora_alpha": 16,
@@ -297,7 +437,7 @@ class TestMultiGPULlama(unittest.TestCase):
                 "sequence_len": 2048,
                 "val_set_size": 0.05,
                 "special_tokens": {
-                    "pad_token": "<|end_of_text|>",
+                    "pad_token": "</s>",
                 },
                 "datasets": [
                     {
@@ -307,7 +447,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -373,7 +513,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -432,7 +572,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,

tests/e2e/multigpu/test_qwen2.py

@@ -47,7 +47,7 @@ class TestMultiGPUQwen2(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "warmup_steps": 20,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 2,

tests/e2e/patched/test_4d_multipack_llama.py

@@ -13,7 +13,7 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import require_torch_2_1_1, with_temp_dir
+from ..utils import require_torch_2_3_1, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -24,7 +24,7 @@ class Test4dMultipackLlama(unittest.TestCase):
     Test case for Llama models using 4d attention with multipack
     """
 
-    @require_torch_2_1_1
+    @require_torch_2_3_1
     @with_temp_dir
     def test_sdp_lora_packing(self, temp_dir):
         # pylint: disable=duplicate-code

tests/e2e/test_dpo.py

@@ -115,6 +115,51 @@ class TestDPOLlamaLora(unittest.TestCase):
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "checkpoint-20/adapter_model.safetensors").exists()
 
+    @with_temp_dir
+    def test_dpo_use_weighting(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 64,
+                "lora_alpha": 32,
+                "lora_dropout": 0.1,
+                "lora_target_linear": True,
+                "special_tokens": {},
+                "rl": "dpo",
+                "dpo_use_weighting": True,
+                "datasets": [
+                    {
+                        "path": "arcee-ai/distilabel-intel-orca-dpo-pairs-binarized",
+                        "type": "chatml.ultra",
+                        "split": "train",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "paged_adamw_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "warmup_steps": 5,
+                "gradient_checkpointing": True,
+                "gradient_checkpointing_kwargs": {"use_reentrant": True},
+            }
+        )
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "checkpoint-20/adapter_model.safetensors").exists()
+
     @pytest.mark.skip("kto_pair no longer supported in trl")
     @with_temp_dir
     def test_kto_pair_lora(self, temp_dir):

tests/e2e/utils.py

@@ -9,6 +9,8 @@ from functools import wraps
 from importlib.metadata import version
 from pathlib import Path
 
+import torch
+
 
 def with_temp_dir(test_func):
     @wraps(test_func)
@@ -35,13 +37,18 @@ def most_recent_subdir(path):
     return subdir
 
 
-def require_torch_2_1_1(test_case):
+def require_torch_2_3_1(test_case):
     """
-    Decorator marking a test that requires torch >= 2.1.1
+    Decorator marking a test that requires torch >= 2.3.1
     """
 
-    def is_min_2_1_1():
+    def is_min_2_3_1():
         torch_version = version("torch")
-        return torch_version >= "2.1.1"
+        return torch_version >= "2.3.1"
 
-    return unittest.skipUnless(is_min_2_1_1(), "test torch 2.1.1")(test_case)
+    return unittest.skipUnless(is_min_2_3_1(), "test torch 2.3.1")(test_case)
+
+
+def is_hopper():
+    compute_capability = torch.cuda.get_device_capability()
+    return compute_capability == (9, 0)

tests/integrations/liger.py

@@ -0,0 +1,80 @@
+"""
+config validation tests for swiglu args
+"""
+# pylint: disable=duplicate-code
+import logging
+from typing import Optional
+
+import pytest
+
+from axolotl.utils.config import validate_config
+from axolotl.utils.dict import DictDefault
+
+
+@pytest.fixture(name="minimal_base_cfg")
+def fixture_cfg():
+    return DictDefault(
+        {
+            "base_model": "TinyLlama/TinyLlama-1.1B-Chat-v0.6",
+            "learning_rate": 0.000001,
+            "datasets": [
+                {
+                    "path": "mhenrichsen/alpaca_2k_test",
+                    "type": "alpaca",
+                }
+            ],
+            "micro_batch_size": 1,
+            "gradient_accumulation_steps": 1,
+        }
+    )
+
+
+class BaseValidation:
+    """
+    Base validation module to setup the log capture
+    """
+
+    _caplog: Optional[pytest.LogCaptureFixture] = None
+
+    @pytest.fixture(autouse=True)
+    def inject_fixtures(self, caplog):
+        self._caplog = caplog
+
+
+# pylint: disable=too-many-public-methods
+class TestValidation(BaseValidation):
+    """
+    Test the validation module for liger
+    """
+
+    def test_deprecated_swiglu(self, minimal_cfg):
+        test_cfg = DictDefault(
+            {
+                "liger_swiglu": False,
+            }
+            | minimal_cfg
+        )
+
+        with self._caplog.at_level(logging.WARNING):
+            updated_cfg = validate_config(test_cfg)
+            assert (
+                "The 'liger_swiglu' argument is deprecated"
+                in self._caplog.records[0].message
+            )
+            assert updated_cfg.liger_swiglu is None
+            assert updated_cfg.liger_glu_activations is False
+
+    def test_conflict_swiglu_ligergluactivation(self, minimal_cfg):
+        test_cfg = DictDefault(
+            {
+                "liger_swiglu": False,
+                "liger_glu_activations": True,
+            }
+            | minimal_cfg
+        )
+
+        with pytest.raises(
+            ValueError,
+            match=r".*You cannot have both `liger_swiglu` and `liger_glu_activation` set.*",
+        ):
+            validate_config(test_cfg)

tests/prompt_strategies/test_sharegpt.py

@@ -1,500 +0,0 @@
-"""
-Test module for sharegpt integration w chatml
-"""
-
-import pytest
-from datasets import Dataset
-from tokenizers import AddedToken
-from transformers import AutoTokenizer
-
-from axolotl.datasets import TokenizedPromptDataset
-from axolotl.prompt_strategies.sharegpt import (
-    GlaiveShareGPTPromptTokenizingStrategy,
-    SimpleShareGPTPromptTokenizingStrategy,
-    register_chatml_template,
-    register_llama3_template,
-)
-from axolotl.prompters import ShareGPTPrompterV2
-
-register_chatml_template()
-register_llama3_template()
-
-
-@pytest.fixture(name="sharegpt_dataset")
-def fixture_sharegpt_dataset():
-    return Dataset.from_list(
-        [
-            {
-                "conversations": [
-                    {
-                        "from": "system",
-                        "value": "repeat",
-                    },
-                    {
-                        "from": "human",
-                        "value": "hello",
-                    },
-                    {
-                        "from": "gpt",
-                        "value": "hello",
-                    },
-                    {
-                        "from": "human",
-                        "value": "goodbye",
-                    },
-                    {
-                        "from": "gpt",
-                        "value": "goodbye",
-                    },
-                ]
-            }
-        ]
-    )
-
-
-@pytest.fixture(name="sharegpt_dataset_with_weights")
-def fixture_sharegpt_dataset_with_weights():
-    return Dataset.from_list(
-        [
-            {
-                "conversations": [
-                    {
-                        "from": "system",
-                        "value": "repeat",
-                    },
-                    {
-                        "from": "human",
-                        "value": "hello",
-                        "weight": 1,
-                    },
-                    {
-                        "from": "gpt",
-                        "value": "hello",
-                        "weight": 0,
-                    },
-                    {
-                        "from": "human",
-                        "value": "rehello",
-                        "weight": 0,
-                    },
-                    {
-                        "from": "gpt",
-                        "value": "rehello",
-                        "weight": 1,
-                    },
-                    {
-                        "from": "human",
-                        "value": "goodbye",
-                    },
-                    {
-                        "from": "gpt",
-                        "value": "goodbye",
-                        "weight": 0,
-                    },
-                ]
-            }
-        ]
-    )
-
-
-@pytest.fixture(name="glaive_dataset")
-def fixture_sharegpt_glaive_dataset():
-    return Dataset.from_list(
-        [
-            {
-                "system": "SYSTEM: This is a system prompt",
-                "chat": "USER: Can you book a flight for me from New York to London? ASSISTANT: I'm sorry, but I don't have the capability to book flights.  <|endoftext|>",
-            }
-        ]
-    )
-
-
-@pytest.fixture(name="multi_role_dataset")
-def fixture_multi_role_dataset():
-    return Dataset.from_list(
-        [
-            {
-                "conversations": [
-                    {
-                        "from": "system",
-                        "value": "use get_weather(city) to get the weather for a city",
-                    },
-                    {
-                        "from": "human",
-                        "value": "hello, what's the weather in New York?",
-                    },
-                    {
-                        "from": "gpt",
-                        "value": "let me get that for you",
-                    },
-                    {
-                        "from": "tool",
-                        "value": "get_weather(New York)",
-                    },
-                    {
-                        "from": "gpt",
-                        "value": "the weather in New York is 70 degrees and sunny",
-                    },
-                ]
-            }
-        ]
-    )
-
-
-@pytest.fixture(name="tokenizer")
-def fixture_tokenizer():
-    tokenizer = AutoTokenizer.from_pretrained(
-        "casperhansen/mistral-7b-instruct-v0.1-awq"
-    )
-    tokenizer.add_special_tokens(
-        {
-            "eos_token": AddedToken(
-                "<|im_end|>", rstrip=False, lstrip=False, normalized=False
-            )
-        }
-    )
-    tokenizer.add_tokens(
-        [
-            AddedToken("<|im_start|>", rstrip=False, lstrip=False, normalized=False),
-        ]
-    )
-
-    return tokenizer
-
-
-@pytest.fixture(name="llama3_tokenizer")
-def fixture_llama3_tokenizer():
-    tokenizer = AutoTokenizer.from_pretrained("NousResearch/Meta-Llama-3-8B")
-    tokenizer.eos_token = "<|eot_id|>"
-
-    return tokenizer
-
-
-class TestSharegptLlama3:
-    """Test class for ShareGPT style datasets with llama-3 prompts"""
-
-    def test_tokenization(self, sharegpt_dataset, llama3_tokenizer):
-        strategy = SimpleShareGPTPromptTokenizingStrategy(
-            ShareGPTPrompterV2(
-                conversation="llama3",
-                role_key_model=None,
-                role_key_human=None,
-            ),
-            llama3_tokenizer,
-            False,  # train_on_inputs
-            2048,  # sequence_len
-        )
-
-        dataset_wrapper = TokenizedPromptDataset(
-            strategy, sharegpt_dataset, process_count=1
-        )
-
-        input_ids = dataset_wrapper[0]["input_ids"]
-
-        # fmt: off
-        # pylint: disable=duplicate-code
-        assert input_ids == [
-            128000,  # bos
-            128006, 9125, 128007,  # system header
-            271, 31724, 128009,  # sys prompt, eot
-            128006, 882, 128007,  # user header
-            271, 15339, 128009,  # user prompt eot
-            128006, 78191, 128007,  # assistant header
-            271, 15339, 128009,   # assistant response eot
-            128006, 882, 128007,
-            271, 19045, 29474, 128009,
-            128006, 78191, 128007,
-            271, 19045, 29474, 128009,
-        ]
-        # fmt: on
-
-    def test_tokenization_with_weights(
-        self, sharegpt_dataset_with_weights, llama3_tokenizer
-    ):
-        strategy = SimpleShareGPTPromptTokenizingStrategy(
-            ShareGPTPrompterV2(
-                conversation="llama3",
-                role_key_model=None,
-                role_key_human=None,
-            ),
-            llama3_tokenizer,
-            False,  # train_on_inputs
-            2048,  # sequence_len
-        )
-
-        dataset_wrapper = TokenizedPromptDataset(
-            strategy, sharegpt_dataset_with_weights, process_count=1
-        )
-
-        input_ids = dataset_wrapper[0]["input_ids"]
-
-        # fmt: off
-        # pylint: disable=duplicate-code
-        assert input_ids == [
-            128000,  # bos
-            128006, 9125, 128007,  # system header
-            271, 31724, 128009,  # sys prompt, eot
-            128006, 882, 128007,  # user header
-            271, 15339, 128009,  # user prompt eot
-            128006, 78191, 128007,  # assistant header
-            271, 15339, 128009,   # assistant response eot
-            128006, 882, 128007,
-            271, 11310, 4896, 128009,
-            128006, 78191, 128007,
-            271, 11310, 4896, 128009,
-            128006, 882, 128007,
-            271, 19045, 29474, 128009,
-            128006, 78191, 128007,
-            271, 19045, 29474, 128009,
-        ]
-        # fmt: on
-
-
-class TestSharegptChatML:
-    """
-    Test class for sharegpt prompter
-    """
-
-    def test_no_double_im_end(self, sharegpt_dataset, tokenizer):
-        strategy = SimpleShareGPTPromptTokenizingStrategy(
-            ShareGPTPrompterV2(
-                conversation="chatml",
-                role_key_model=None,
-                role_key_human=None,
-            ),
-            tokenizer,
-            False,  # train_on_inputs
-            2048,  # sequence_len
-        )
-
-        dataset_wrapper = TokenizedPromptDataset(
-            strategy, sharegpt_dataset, process_count=1
-        )
-
-        input_ids = dataset_wrapper[0]["input_ids"]
-        # fmt: off
-        assert input_ids == [
-            #  28705, 13, is " \n"
-            1,   # bos
-            32001, 1587, 13, 25997, 32000, 28705, 13,  # system
-            32001, 2188, 13, 21558, 32000, 28705, 13,  # human
-            32001, 13892, 13, 21558, 32000, 28705, 13,  # gpt
-            32001, 2188, 13, 12684, 17664, 32000, 28705, 13,   # human
-            32001, 13892, 13, 12684, 17664, 32000, 28705, 13,  # gpt
-        ]
-        # fmt: on
-
-    def test_no_double_im_end_with_weights(
-        self, sharegpt_dataset_with_weights, tokenizer
-    ):
-        strategy = SimpleShareGPTPromptTokenizingStrategy(
-            ShareGPTPrompterV2(
-                conversation="chatml",
-                role_key_model=None,
-                role_key_human=None,
-            ),
-            tokenizer,
-            False,  # train_on_inputs
-            2048,  # sequence_len
-        )
-
-        dataset_wrapper = TokenizedPromptDataset(
-            strategy, sharegpt_dataset_with_weights, process_count=1
-        )
-
-        input_ids = dataset_wrapper[0]["input_ids"]
-        # fmt: off
-        assert input_ids == [
-            #  28705, 13, is " \n"
-            1,   # bos
-            32001, 1587, 13, 25997, 32000, 28705, 13,  # system
-            32001, 2188, 13, 21558, 32000, 28705, 13,  # human
-            32001, 13892, 13, 21558, 32000, 28705, 13,  # gpt
-            32001, 2188, 13, 267, 21558, 32000, 28705, 13,  # human
-            32001, 13892, 13, 267, 21558, 32000, 28705, 13,  # gpt
-            32001, 2188, 13, 12684, 17664, 32000, 28705, 13,   # human
-            32001, 13892, 13, 12684, 17664, 32000, 28705, 13,  # gpt
-        ]
-        # fmt: on
-
-    def test_no_train_on_input(self, sharegpt_dataset, tokenizer):
-        strategy = SimpleShareGPTPromptTokenizingStrategy(
-            ShareGPTPrompterV2(
-                conversation="chatml",
-                role_key_model=None,
-                role_key_human=None,
-            ),
-            tokenizer,
-            False,  # train_on_inputs
-            2048,  # sequence_len
-        )
-
-        dataset_wrapper = TokenizedPromptDataset(
-            strategy, sharegpt_dataset, process_count=1
-        )
-
-        labels = dataset_wrapper[0]["labels"]
-        # fmt: off
-        assert labels == [
-            -100,   # bos
-            -100, -100, -100, -100, -100, -100, -100,  # system
-            -100, -100, -100, -100, -100, -100, -100,  # human
-            -100, -100, 13, 21558, 32000, 28705, 13,  # gpt
-            -100, -100, -100, -100, -100, -100, -100, -100,   # human
-            -100, -100, 13, 12684, 17664, 32000, 28705, 13,  # gpt
-        ]
-        # fmt: on
-
-    def test_no_train_on_input_with_weights(
-        self, sharegpt_dataset_with_weights, tokenizer
-    ):
-        strategy = SimpleShareGPTPromptTokenizingStrategy(
-            ShareGPTPrompterV2(
-                conversation="chatml",
-                role_key_model=None,
-                role_key_human=None,
-            ),
-            tokenizer,
-            False,  # train_on_inputs
-            2048,  # sequence_len
-        )
-
-        dataset_wrapper = TokenizedPromptDataset(
-            strategy, sharegpt_dataset_with_weights, process_count=1
-        )
-
-        labels = dataset_wrapper[0]["labels"]
-        # fmt: off
-        assert labels == [
-            -100,   # bos
-            -100, -100, -100, -100, -100, -100, -100,  # system
-            -100, -100, -100, -100, -100, -100, -100,  # human
-            -100, -100, -100, -100, -100, -100, -100,  # gpt with weight zero
-            -100, -100, -100, -100, -100, -100, -100, -100,  # human
-            -100, -100, 13, 267, 21558, 32000, 28705, 13,  # gpt
-            -100, -100, -100, -100, -100, -100, -100, -100,   # human
-            -100, -100, -100, -100, -100, -100, -100, -100  # gpt with weight zero
-        ]
-        # fmt: on
-
-    def test_w_train_on_input(self, sharegpt_dataset, tokenizer):
-        strategy = SimpleShareGPTPromptTokenizingStrategy(
-            ShareGPTPrompterV2(
-                conversation="chatml",
-                role_key_model=None,
-                role_key_human=None,
-            ),
-            tokenizer,
-            True,  # train_on_inputs
-            2048,  # sequence_len
-        )
-
-        dataset_wrapper = TokenizedPromptDataset(
-            strategy, sharegpt_dataset, process_count=1
-        )
-
-        labels = dataset_wrapper[0]["labels"]
-        # fmt: off
-        assert labels == [
-            1,   # bos
-            32001, 1587, 13, 25997, 32000, 28705, 13,  # system
-            32001, 2188, 13, 21558, 32000, 28705, 13,  # human
-            32001, 13892, 13, 21558, 32000, 28705, 13,  # gpt
-            32001, 2188, 13, 12684, 17664, 32000, 28705, 13,   # human
-            32001, 13892, 13, 12684, 17664, 32000, 28705, 13,  # gpt
-        ]
-        # fmt: on
-
-    def test_w_train_on_input_with_weights(
-        self, sharegpt_dataset_with_weights, tokenizer
-    ):
-        strategy = SimpleShareGPTPromptTokenizingStrategy(
-            ShareGPTPrompterV2(
-                conversation="chatml",
-                role_key_model=None,
-                role_key_human=None,
-            ),
-            tokenizer,
-            True,  # train_on_inputs
-            2048,  # sequence_len
-        )
-
-        dataset_wrapper = TokenizedPromptDataset(
-            strategy, sharegpt_dataset_with_weights, process_count=1
-        )
-
-        labels = dataset_wrapper[0]["labels"]
-        # fmt: off
-        assert labels == [
-            1,   # bos
-            32001, 1587, 13, 25997, 32000, 28705, 13,  # system
-            32001, 2188, 13, 21558, 32000, 28705, 13,  # human
-            -100, -100, -100, -100, -100, -100, -100,  # gpt with weight 0
-            -100, -100, -100, -100, -100, -100, -100, -100,  # human with weight 0
-            32001, 13892, 13, 267, 21558, 32000, 28705, 13,  # gpt
-            32001, 2188, 13, 12684, 17664, 32000, 28705, 13,  # human
-            -100, -100, -100, -100, -100, -100, -100, -100  # gpt with weight 0
-        ]
-        # fmt: on
-
-    def test_chatml_glaive(self, glaive_dataset, tokenizer):
-        strategy = GlaiveShareGPTPromptTokenizingStrategy(
-            ShareGPTPrompterV2(
-                conversation="chatml",
-                role_key_model=None,
-                role_key_human=None,
-            ),
-            tokenizer,
-            True,  # train_on_inputs
-            2048,  # sequence_len
-        )
-
-        dataset_wrapper = TokenizedPromptDataset(
-            strategy, glaive_dataset, process_count=1
-        )
-
-        labels = dataset_wrapper[0]["labels"]
-        # fmt: off
-        assert labels == [
-            1,  # bos
-            32001, 1587, 13, 3260, 349, 264, 1587, 11510, 32000, 28705, 13,  # system
-            32001, 2188, 13, 6325, 368, 1820, 264, 9314, 354, 528, 477, 1450, 2726, 298, 4222, 28804, 32000, 28705, 13,  # human
-            32001, 13892, 13, 28737, 28742, 28719, 7371, 28725, 562, 315, 949, 28742, 28707, 506, 272, 21368, 298, 1820, 22447, 28723, 28705, 523, 28766, 416, 1009, 772, 28766, 28767, 32000, 28705, 13  # gpt
-        ]
-        # fmt: on
-
-    def test_multi_role_dataset(self, multi_role_dataset, tokenizer):
-        strategy = SimpleShareGPTPromptTokenizingStrategy(
-            ShareGPTPrompterV2(conversation="chatml", roles={"input": ["tool"]}),
-            tokenizer,
-            False,  # train_on_inputs
-            2048,  # sequence_len
-        )
-
-        dataset_wrapper = TokenizedPromptDataset(
-            strategy, multi_role_dataset, process_count=1
-        )
-
-        input_ids = dataset_wrapper[0]["input_ids"]
-        # fmt: off
-        assert input_ids == [
-            1,   # bos
-            32001, 1587, 13, 1730, 625, 28730, 769, 1223, 28732, 18373, 28731, 298, 625, 272, 8086, 354, 264, 2990, 32000, 28705, 13,  # system
-            32001, 2188, 13, 21558, 28725, 767, 28742, 28713, 272, 8086, 297, 1450, 2726, 28804, 32000, 28705, 13,  # human
-            32001, 13892, 13, 895, 528, 625, 369, 354, 368, 32000, 28705, 13,  # gpt
-            32001, 3921, 13, 527, 28730, 769, 1223, 28732, 2972, 2726, 28731, 32000, 28705, 13,  # tool
-            32001, 13892, 13, 1237, 8086, 297, 1450, 2726, 349, 28705, 28787, 28734, 11182, 304, 4376, 1780, 32000, 28705, 13  # gpt
-        ]
-        # fmt: on
-
-        labels = dataset_wrapper[0]["labels"]
-        # fmt: off
-        assert labels == [
-            -100,  # bos
-            -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,  # system
-            -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,  # human
-            -100, -100, 13, 895, 528, 625, 369, 354, 368, 32000, 28705, 13,  # gpt
-            -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,  # tool
-            -100, -100, 13, 1237, 8086, 297, 1450, 2726, 349, 28705, 28787, 28734, 11182, 304, 4376, 1780, 32000, 28705, 13  # gpt
-        ]
-        # fmt: on

tests/test_datasets.py

@@ -306,6 +306,10 @@ class TestDatasetPreparation(unittest.TestCase):
         """Verify that processing data from the hub works with a specific revision"""
         with tempfile.TemporaryDirectory() as tmp_dir:
             prepared_path = Path(tmp_dir) / "prepared"
+
+            # make sure prepared_path is empty
+            shutil.rmtree(prepared_path, ignore_errors=True)
+
             cfg = DictDefault(
                 {
                     "tokenizer_config": "huggyllama/llama-7b",
@@ -367,43 +371,44 @@ class TestDatasetPreparation(unittest.TestCase):
     def test_load_local_hub_with_revision(self):
         """Verify that a local copy of a hub dataset can be loaded with a specific revision"""
         with tempfile.TemporaryDirectory() as tmp_dir:
-            tmp_ds_path = Path("mhenrichsen/alpaca_2k_test")
-            tmp_ds_path.mkdir(parents=True, exist_ok=True)
-            snapshot_download(
-                repo_id="mhenrichsen/alpaca_2k_test",
-                repo_type="dataset",
-                local_dir=tmp_ds_path,
-                revision="d05c1cb",
-            )
+            with tempfile.TemporaryDirectory() as tmp_dir2:
+                tmp_ds_path = Path(tmp_dir2) / "mhenrichsen/alpaca_2k_test"
+                tmp_ds_path.mkdir(parents=True, exist_ok=True)
+                snapshot_download(
+                    repo_id="mhenrichsen/alpaca_2k_test",
+                    repo_type="dataset",
+                    local_dir=tmp_ds_path,
+                    revision="d05c1cb",
+                )
 
-            prepared_path = Path(tmp_dir) / "prepared"
-            cfg = DictDefault(
-                {
-                    "tokenizer_config": "huggyllama/llama-7b",
-                    "sequence_len": 1024,
-                    "datasets": [
-                        {
-                            "path": "mhenrichsen/alpaca_2k_test",
-                            "ds_type": "parquet",
-                            "type": "alpaca",
-                            "data_files": [
-                                "mhenrichsen/alpaca_2k_test/alpaca_2000.parquet",
-                            ],
-                            "revision": "d05c1cb",
-                        },
-                    ],
-                }
-            )
+                prepared_path = Path(tmp_dir) / "prepared"
+                cfg = DictDefault(
+                    {
+                        "tokenizer_config": "huggyllama/llama-7b",
+                        "sequence_len": 1024,
+                        "datasets": [
+                            {
+                                "path": "mhenrichsen/alpaca_2k_test",
+                                "ds_type": "parquet",
+                                "type": "alpaca",
+                                "data_files": [
+                                    f"{tmp_ds_path}/alpaca_2000.parquet",
+                                ],
+                                "revision": "d05c1cb",
+                            },
+                        ],
+                    }
+                )
 
-            dataset, _ = load_tokenized_prepared_datasets(
-                self.tokenizer, cfg, prepared_path
-            )
+                dataset, _ = load_tokenized_prepared_datasets(
+                    self.tokenizer, cfg, prepared_path
+                )
 
-            assert len(dataset) == 2000
-            assert "input_ids" in dataset.features
-            assert "attention_mask" in dataset.features
-            assert "labels" in dataset.features
-            shutil.rmtree(tmp_ds_path)
+                assert len(dataset) == 2000
+                assert "input_ids" in dataset.features
+                assert "attention_mask" in dataset.features
+                assert "labels" in dataset.features
+                shutil.rmtree(tmp_ds_path)
 
 
 if __name__ == "__main__":

tests/test_normalize_config.py

@@ -39,12 +39,12 @@ class NormalizeConfigTestCase(unittest.TestCase):
                 "datasets": [
                     {
                         "path": "lorem/ipsum",
-                        "type": "sharegpt",
-                        "conversation": "vicuna_v1.1",
+                        "type": "chat_template",
+                        "chat_template": "gemma",
                     },
                     {
                         "path": "sit/amet",
-                        "type": "sharegpt",
+                        "type": "chat_template",
                     },
                 ],
             }
@@ -52,8 +52,8 @@ class NormalizeConfigTestCase(unittest.TestCase):
 
         normalize_cfg_datasets(cfg)
 
-        assert cfg.datasets[0].conversation == "vicuna_v1.1"
-        assert cfg.datasets[1].conversation == "chatml"
+        assert cfg.datasets[0].chat_template == "gemma"
+        assert cfg.datasets[1].chat_template == "chatml"
 
     @patch("axolotl.utils.config.is_torch_bf16_gpu_available")
     def test_bf16_auto_setter_available(self, mock_bf16_avail):

tests/test_prompt_tokenizers.py

@@ -3,7 +3,6 @@
 import json
 import logging
 import unittest
-from copy import deepcopy
 from pathlib import Path
 from typing import Optional
 
@@ -21,12 +20,8 @@ from axolotl.prompt_strategies.llama2_chat import (
     LLama2ChatTokenizingStrategy,
 )
 from axolotl.prompt_strategies.orpo.chat_template import load
-from axolotl.prompt_strategies.sharegpt import GlaiveShareGPTPromptTokenizingStrategy
-from axolotl.prompt_tokenizers import (
-    AlpacaPromptTokenizingStrategy,
-    ShareGPTPromptTokenizingStrategy,
-)
-from axolotl.prompters import AlpacaPrompter, PromptStyle, ShareGPTPrompterV2
+from axolotl.prompt_tokenizers import AlpacaPromptTokenizingStrategy
+from axolotl.prompters import AlpacaPrompter, PromptStyle
 from axolotl.utils.dict import DictDefault
 
 LOG = logging.getLogger("axolotl")
@@ -65,17 +60,6 @@ test_data = {
 }
 
 
-def prompt_strat(conversation, tokenizer):
-    "Helper function to create a prompt strategy for testing."
-    prompter = ShareGPTPrompterV2(conversation=conversation)
-    return ShareGPTPromptTokenizingStrategy(
-        prompter,
-        tokenizer,
-        False,
-        2048,
-    )
-
-
 class TestPromptTokenizationStrategies(unittest.TestCase):
     """
     Test class for prompt tokenization strategies.
@@ -98,196 +82,6 @@ class TestPromptTokenizationStrategies(unittest.TestCase):
             }
         )
 
-    def test_sharegpt_integration(self):
-        with open(
-            Path(__file__).parent / "fixtures/conversation.json", encoding="utf-8"
-        ) as fin:
-            data = fin.read()
-            conversation = json.loads(data)
-        with open(
-            Path(__file__).parent / "fixtures/conversation.tokenized.json",
-            encoding="utf-8",
-        ) as fin:
-            data = fin.read()
-            tokenized_conversation = json.loads(data)
-        prompter = ShareGPTPrompterV2()
-        strat = ShareGPTPromptTokenizingStrategy(
-            prompter,
-            self.tokenizer,
-            False,
-            2048,
-        )
-        example = strat.tokenize_prompt(conversation)
-        for fields in ["input_ids", "attention_mask", "labels"]:
-            self.assertEqual(len(example[fields]), len(tokenized_conversation[fields]))
-            self.assertEqual(example[fields], tokenized_conversation[fields])
-
-    def test_sharegpt_warnings_integration(self):
-        with open(
-            Path(__file__).parent / "fixtures/conversation.missingturns.json",
-            encoding="utf-8",
-        ) as fin:
-            data = fin.read()
-            conversation = json.loads(data)
-        prompter = ShareGPTPrompterV2()
-        strat = ShareGPTPromptTokenizingStrategy(
-            prompter,
-            self.tokenizer,
-            False,
-            2048,
-        )
-        with self._caplog.at_level(logging.WARNING):
-            strat.tokenize_prompt(conversation)
-            assert "assistant turn has empty text" in self._caplog.records[1].message
-
-    def test_sharegpt_warnings_turns(self):
-        conversation = {
-            "conversations": [
-                {"from": "system", "value": "lorem"},
-                {"from": "gpt", "value": "ipsum"},
-                {"from": "human", "value": "dolor"},
-                {"from": "human", "value": "dolor"},
-                {"from": "gpt", "value": "sit"},
-            ]
-        }
-        prompter = ShareGPTPrompterV2()
-        strat = ShareGPTPromptTokenizingStrategy(
-            prompter,
-            self.tokenizer,
-            False,
-            2048,
-        )
-        with self._caplog.at_level(logging.WARNING):
-            strat.tokenize_prompt(conversation)
-            assert (
-                "Role did not alternate between turns (gpt and human)"
-                in self._caplog.records[0].message
-            )
-
-    def test_sharegpt_llama(self):
-        "Make sure the sharegpt/llama is tokenized and formatted correctly."
-        strat = prompt_strat("llama-2", self.tokenizer)
-
-        def tokenize(conv):
-            return strat.tokenize_prompt(deepcopy(conv))["input_ids"]
-
-        def decode(ids):
-            return strat.tokenizer.decode(ids)
-
-        # fmt: off
-        # System message, multi-turn conversations
-        mt_ids = tokenize(test_data['multi_turn_sys'])
-        assert decode(mt_ids) == '<s> [INST] <<SYS>>\nlorem\n<</SYS>>\n\nabc [/INST] ipsum</s><s> [INST] 123 [/INST] sit</s>'
-        assert mt_ids == [1, 518, 25580, 29962, 3532, 14816, 29903, 6778, 13, 29880, 3668, 13, 29966, 829, 14816, 29903, 6778, 13, 13, 10736, 518, 29914, 25580, 29962, 23421, 2, 1, 518, 25580, 29962, 29871, 29896, 29906, 29941, 518, 29914, 25580, 29962, 7845, 2]
-
-        # System message, single-turn conversations
-        st_ids = tokenize(test_data['single_turn_sys'])
-        assert decode(st_ids) == '<s> [INST] <<SYS>>\nlorem\n<</SYS>>\n\nabc [/INST] ipsum</s>'
-        assert st_ids == [1, 518, 25580, 29962, 3532, 14816, 29903, 6778, 13, 29880, 3668, 13, 29966, 829, 14816, 29903, 6778, 13, 13, 10736, 518, 29914, 25580, 29962, 23421, 2]
-
-        # No system message, single-turn
-        ns_ids = tokenize(test_data['single_turn_no_sys'])
-        assert decode(ns_ids) == '<s> [INST] abc [/INST] ipsum</s>'
-        assert ns_ids == [1, 518, 25580, 29962, 25638, 518, 29914, 25580, 29962, 23421, 2]
-
-        # No system message, multi-turn
-        ns_mt_ids = tokenize(test_data['multi_turn_no_sys'])
-        assert decode(ns_mt_ids) == '<s> [INST] abc [/INST] ipsum</s><s> [INST] 123 [/INST] sit</s>'
-        assert ns_mt_ids == [1, 518, 25580, 29962, 25638, 518, 29914, 25580, 29962, 23421, 2, 1, 518, 25580, 29962, 29871, 29896, 29906, 29941, 518, 29914, 25580, 29962, 7845, 2]
-        # fmt: on
-
-    def test_sharegpt_mistral(self):
-        "Make sure the sharegpt/mistral is tokenized and formatted correctly."
-        strat = prompt_strat("mistral", self.tokenizer)
-
-        def tokenize(conv):
-            return strat.tokenize_prompt(deepcopy(conv))["input_ids"]
-
-        def decode(ids):
-            return strat.tokenizer.decode(ids)
-
-        # fmt: off
-        # System message, multi-turn conversations
-        mt_ids = tokenize(test_data['multi_turn_sys'])
-        assert decode(mt_ids) == '<s> [INST]  lorem\nabc [/INST] ipsum</s> [INST] 123 [/INST] sit</s>'
-        assert mt_ids == [1, 518, 25580, 29962, 29871, 301, 3668, 13, 10736, 518, 29914, 25580, 29962, 23421, 2, 518, 25580, 29962, 29871, 29896, 29906, 29941, 518, 29914, 25580, 29962, 7845, 2]
-
-        # System message, single-turn conversations
-        st_ids = tokenize(test_data['single_turn_sys'])
-        assert decode(st_ids) == '<s> [INST]  lorem\nabc [/INST] ipsum</s>'
-        assert st_ids == [1, 518, 25580, 29962, 29871, 301, 3668, 13, 10736, 518, 29914, 25580, 29962, 23421, 2]
-
-        # No system message, single-turn
-        ns_ids = tokenize(test_data['single_turn_no_sys'])
-        assert decode(ns_ids) == '<s> [INST] abc [/INST] ipsum</s>'
-        assert ns_ids == [1, 518, 25580, 29962, 25638, 518, 29914, 25580, 29962, 23421, 2]
-
-        # No system message, multi-turn
-        ns_mt_ids = tokenize(test_data['multi_turn_no_sys'])
-        assert decode(ns_mt_ids) == '<s> [INST] abc [/INST] ipsum</s> [INST] 123 [/INST] sit</s>'
-        assert ns_mt_ids == [1, 518, 25580, 29962, 25638, 518, 29914, 25580, 29962, 23421, 2, 518, 25580, 29962, 29871, 29896, 29906, 29941, 518, 29914, 25580, 29962, 7845, 2]
-        # fmt: on
-
-    def test_sharegpt_changes_roles(self):
-        conversation = {
-            "roles": ["USER", "CHARACTER"],
-            "conversations": [
-                {"from": "system", "value": "lorem"},
-                {"from": "gpt", "value": "ipsum"},
-                {"from": "human", "value": "dolor"},
-                {"from": "gpt", "value": "sit"},
-            ],
-        }
-        prompter = ShareGPTPrompterV2()
-        strat = ShareGPTPromptTokenizingStrategy(
-            prompter,
-            self.tokenizer,
-            False,
-            2048,
-        )
-        with self._caplog.at_level(logging.WARNING):
-            res = strat.tokenize_prompt(conversation)
-            assert "CHARACTER" in self.tokenizer.decode(res["input_ids"])
-
-    def test_sharegpt_assistant_label_ignore(self):
-        conversation = {
-            "roles": ["user", "assistant"],
-            "conversations": [
-                {"from": "system", "value": "lorem"},
-                {"from": "gpt", "value": "ipsum"},
-                {"from": "human", "value": "dolor"},
-                {"from": "gpt", "value": "sit"},
-            ],
-        }
-        prompter = ShareGPTPrompterV2()
-        strat = ShareGPTPromptTokenizingStrategy(
-            prompter,
-            self.tokenizer,
-            False,
-            2048,
-        )
-        with self._caplog.at_level(logging.WARNING):
-            res = strat.tokenize_prompt(conversation)
-            idx = res["input_ids"].index(20255)  # assistant token
-            assert res["labels"][idx] == -100
-
-    def test_glaive_tool_label_ignore(self):
-        conversation = {
-            "system": "SYSTEM: This is a system prompt",
-            "chat": "USER: Can you book a flight for me from New York to London? ASSISTANT: I'm sorry, but I don't have the capability to book flights.  <|endoftext|>",
-        }
-        prompter = ShareGPTPrompterV2()
-        strat = GlaiveShareGPTPromptTokenizingStrategy(
-            prompter,
-            self.tokenizer,
-            False,
-            2048,
-        )
-        with self._caplog.at_level(logging.WARNING):
-            res = strat.tokenize_prompt(conversation)
-            idx = res["input_ids"].index(13566)  # assistant token
-            assert res["labels"][idx] == -100
-
     def test_no_sys_prompt(self):
         """
         tests the interface between the user and assistant parts

tests/test_validation.py

@@ -646,39 +646,6 @@ class TestValidation(BaseValidation):
 
         validate_config(cfg)
 
-    def test_sharegpt_deprecation(self, minimal_cfg):
-        cfg = (
-            DictDefault(
-                {"datasets": [{"path": "lorem/ipsum", "type": "sharegpt:chat"}]}
-            )
-            | minimal_cfg
-        )
-        with self._caplog.at_level(logging.WARNING):
-            new_cfg = validate_config(cfg)
-            assert any(
-                "`type: sharegpt:chat` will soon be deprecated." in record.message
-                for record in self._caplog.records
-            )
-        assert new_cfg.datasets[0].type == "sharegpt"
-
-        cfg = (
-            DictDefault(
-                {
-                    "datasets": [
-                        {"path": "lorem/ipsum", "type": "sharegpt_simple:load_role"}
-                    ]
-                }
-            )
-            | minimal_cfg
-        )
-        with self._caplog.at_level(logging.WARNING):
-            new_cfg = validate_config(cfg)
-            assert any(
-                "`type: sharegpt_simple` will soon be deprecated." in record.message
-                for record in self._caplog.records
-            )
-        assert new_cfg.datasets[0].type == "sharegpt:load_role"
-
     def test_no_conflict_save_strategy(self, minimal_cfg):
         cfg = (
             DictDefault(

tests/test_validation_dataset.py

@@ -48,9 +48,8 @@ class TestValidationCheckDatasetConfig(BaseValidation):
             | {
                 "datasets": [
                     {
-                        "path": "LDJnr/Puffin",
-                        "type": "sharegpt",
-                        "conversation": "chatml",
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
                         "shards": 10,
                     }
                 ]
@@ -62,7 +61,6 @@ class TestValidationCheckDatasetConfig(BaseValidation):
         def _check_config():
             assert checked_cfg.datasets[0].path == cfg.datasets[0].path
             assert checked_cfg.datasets[0].type == cfg.datasets[0].type
-            assert checked_cfg.datasets[0].conversation == cfg.datasets[0].conversation
             assert checked_cfg.datasets[0].shards == cfg.datasets[0].shards
 
         _check_config()