set torch version to what is installed during axolotl install (#1234)

* Make sure test_dataset are used and treat val_set_size.

* Add test_datasets docs.

* Apply suggestions from code review

---------

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

.github/FUNDING.yml

@@ -1,6 +1,6 @@
 # These are supported funding model platforms
 
-github: OpenAccess-AI-Collective # Replace with up to 4 GitHub Sponsors-enabled usernames e.g., [user1, user2]
+github: [winglian, OpenAccess-AI-Collective] # Replace with up to 4 GitHub Sponsors-enabled usernames e.g., [user1, user2]
 patreon: # Replace with a single Patreon username
 open_collective: # Replace with a single Open Collective username
 ko_fi: axolotl_ai # Replace with a single Ko-fi username

.github/workflows/base.yml

@@ -1,10 +1,7 @@
 name: ci-cd-base
 
 on:
-  push:
-    branches:
-      - "main-base"
-      - "dev-base"
+  workflow_dispatch:
 
 jobs:
   build-base:
@@ -15,11 +12,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: "118"
-            cuda_version: 11.8.0
-            python_version: "3.9"
-            pytorch: 2.0.1
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
           - cuda: "118"
             cuda_version: 11.8.0
             python_version: "3.10"
@@ -28,12 +20,17 @@ jobs:
           - cuda: "118"
             cuda_version: 11.8.0
             python_version: "3.10"
-            pytorch: 2.1.1
+            pytorch: 2.1.2
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
           - cuda: "121"
             cuda_version: 12.1.0
             python_version: "3.10"
-            pytorch: 2.1.1
+            pytorch: 2.1.2
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
+          - cuda: "121"
+            cuda_version: 12.1.0
+            python_version: "3.11"
+            pytorch: 2.1.2
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
     steps:
       - name: Checkout
@@ -56,7 +53,7 @@ jobs:
           context: .
           file: ./docker/Dockerfile-base
           push: ${{ github.event_name != 'pull_request' }}
-          tags: ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+          tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
           labels: ${{ steps.metadata.outputs.labels }}
           build-args: |
             CUDA_VERSION=${{ matrix.cuda_version }}

.github/workflows/main.yml

@@ -4,6 +4,7 @@ on:
   push:
     branches:
       - "main"
+  workflow_dispatch:
 
 jobs:
   build-axolotl:
@@ -15,24 +16,24 @@ jobs:
         include:
           - cuda: 118
             cuda_version: 11.8.0
-            python_version: "3.9"
+            python_version: "3.10"
             pytorch: 2.0.1
             axolotl_extras:
           - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.10"
-            pytorch: 2.0.1
+            pytorch: 2.1.2
             axolotl_extras:
             is_latest: true
-          - cuda: 118
-            cuda_version: 11.8.0
-            python_version: "3.10"
-            pytorch: 2.1.1
-            axolotl_extras:
           - cuda: 121
             cuda_version: 12.1.0
             python_version: "3.10"
-            pytorch: 2.1.1
+            pytorch: 2.1.2
+            axolotl_extras:
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.11"
+            pytorch: 2.1.2
             axolotl_extras:
     runs-on: [self-hosted, gpu, docker]
     steps:
@@ -86,24 +87,24 @@ jobs:
         include:
           - cuda: 118
             cuda_version: 11.8.0
-            python_version: "3.9"
+            python_version: "3.10"
             pytorch: 2.0.1
             axolotl_extras:
           - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.10"
-            pytorch: 2.0.1
+            pytorch: 2.1.2
             axolotl_extras:
             is_latest: true
-          - cuda: 118
-            cuda_version: 11.8.0
-            python_version: "3.10"
-            pytorch: 2.1.1
-            axolotl_extras:
           - cuda: 121
             cuda_version: 12.1.0
             python_version: "3.10"
-            pytorch: 2.1.1
+            pytorch: 2.1.2
+            axolotl_extras:
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.11"
+            pytorch: 2.1.2
             axolotl_extras:
     runs-on: [self-hosted, gpu, docker]
     steps:

.github/workflows/tests.yml

@@ -73,7 +73,7 @@ jobs:
           - cuda: 121
             cuda_version: 12.1.0
             python_version: "3.10"
-            pytorch: 2.1.1
+            pytorch: 2.1.2
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -106,3 +106,7 @@ jobs:
       - name: GPU Unit Tests monkeypatched w docker image
         run: |
           docker run --privileged --gpus "all" --env WANDB_DISABLED=true --rm ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }} pytest /workspace/axolotl/tests/e2e/patched/
+      - name: Prune image from docker
+        if: github.ref != 'refs/heads/main'
+        run: |
+          docker rmi -f ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}

README.md

@@ -607,12 +607,25 @@ datasets:
       # For `completion` datsets only, uses the provided field instead of `text` column
       field:
 
+# A list of one or more datasets to eval the model with.
+# You can use either test_datasets, or val_set_size, but not both.
+test_datasets:
+  - path: /workspace/data/eval.jsonl
+    ds_type: json
+    # You need to specify a split. For "json" datasets the default split is called "train".
+    split: train
+    type: completion
+    data_files:
+      - /workspace/data/eval.jsonl
+
 # use RL training: dpo, ipo, kto_pair
 rl:
 
 # Saves the desired chat template to the tokenizer_config.json for easier inferencing
 # Currently supports chatml and inst (mistral/mixtral)
 chat_template: chatml
+# Changes the default system message
+default_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.
 # Axolotl attempts to save the dataset as an arrow after packing the data together so
 # subsequent training attempts load faster, relative path
 dataset_prepared_path: data/last_run_prepared
@@ -694,6 +707,12 @@ lora_modules_to_save:
 
 lora_fan_in_fan_out: false
 
+peft:
+  # Configuration options for loftq initialization for LoRA
+  # https://huggingface.co/docs/peft/developer_guides/quantization#loftq-initialization
+  loftq_config:
+    loftq_bits:  # typically 4 bits
+
 # ReLoRA configuration
 # Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
 relora_steps: # Number of steps per ReLoRA restart

deepspeed_configs/zero1.json

@@ -15,15 +15,6 @@
     "hysteresis": 2,
     "min_loss_scale": 1
   },
-  "optimizer": {
-    "type": "AdamW",
-    "params": {
-      "lr": "auto",
-      "betas": "auto",
-      "eps": "auto",
-      "weight_decay": "auto"
-    }
-  },
   "gradient_accumulation_steps": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",

deepspeed_configs/zero2.json

@@ -19,15 +19,6 @@
     "hysteresis": 2,
     "min_loss_scale": 1
   },
-  "optimizer": {
-    "type": "AdamW",
-    "params": {
-      "lr": "auto",
-      "betas": "auto",
-      "eps": "auto",
-      "weight_decay": "auto"
-    }
-  },
   "gradient_accumulation_steps": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",

deepspeed_configs/zero3.json

@@ -23,15 +23,6 @@
     "hysteresis": 2,
     "min_loss_scale": 1
   },
-  "optimizer": {
-    "type": "AdamW",
-    "params": {
-      "lr": "auto",
-      "betas": "auto",
-      "eps": "auto",
-      "weight_decay": "auto"
-    }
-  },
   "gradient_accumulation_steps": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",

deepspeed_configs/zero3_bf16.json

@@ -23,15 +23,6 @@
     "hysteresis": 2,
     "min_loss_scale": 1
   },
-  "optimizer": {
-    "type": "AdamW",
-    "params": {
-      "lr": "auto",
-      "betas": "auto",
-      "eps": "auto",
-      "weight_decay": "auto"
-    }
-  },
   "gradient_accumulation_steps": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",

examples/cerebras/qlora.yml

@@ -11,7 +11,6 @@ val_set_size: 0.05
 adapter: qlora
 lora_model_dir:
 sequence_len: 2048
-max_packed_sequence_len: 2048
 lora_r: 16
 lora_alpha: 32
 lora_dropout: 0.05

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

@@ -0,0 +1,197 @@
+{
+  "cells": [
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "AKjdG7tbTb-n"
+      },
+      "source": [
+        "# Example notebook for running Axolotl on google colab"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "RcbNpOgWRcii"
+      },
+      "outputs": [],
+      "source": [
+        "import torch\n",
+        "# Check so there is a gpu available, a T4(free tier) is enough to run this notebook\n",
+        "assert (torch.cuda.is_available()==True)"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "h3nLav8oTRA5"
+      },
+      "source": [
+        "## Install Axolotl and dependencies"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "id": "3c3yGAwnOIdi",
+        "outputId": "e3777b5a-40ef-424f-e181-62dfecd1dd01"
+      },
+      "outputs": [],
+      "source": [
+        "!pip install -e git+https://github.com/OpenAccess-AI-Collective/axolotl#egg=axolotl\n",
+        "!pip install flash-attn==\"2.5.0\"\n",
+        "!pip install deepspeed==\"0.13.1\""
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "BW2MFr7HTjub"
+      },
+      "source": [
+        "## Create an yaml config file"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "9pkF2dSoQEUN"
+      },
+      "outputs": [],
+      "source": [
+        "import yaml\n",
+        "\n",
+        "# Your YAML string\n",
+        "yaml_string = \"\"\"\n",
+        "base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T\n",
+        "model_type: LlamaForCausalLM\n",
+        "tokenizer_type: LlamaTokenizer\n",
+        "is_llama_derived_model: true\n",
+        "\n",
+        "load_in_8bit: false\n",
+        "load_in_4bit: true\n",
+        "strict: false\n",
+        "\n",
+        "datasets:\n",
+        "  - path: mhenrichsen/alpaca_2k_test\n",
+        "    type: alpaca\n",
+        "dataset_prepared_path:\n",
+        "val_set_size: 0.05\n",
+        "output_dir: ./qlora-out\n",
+        "\n",
+        "adapter: qlora\n",
+        "lora_model_dir:\n",
+        "\n",
+        "sequence_len: 1096\n",
+        "sample_packing: true\n",
+        "pad_to_sequence_len: true\n",
+        "\n",
+        "lora_r: 32\n",
+        "lora_alpha: 16\n",
+        "lora_dropout: 0.05\n",
+        "lora_target_modules:\n",
+        "lora_target_linear: true\n",
+        "lora_fan_in_fan_out:\n",
+        "\n",
+        "wandb_project:\n",
+        "wandb_entity:\n",
+        "wandb_watch:\n",
+        "wandb_name:\n",
+        "wandb_log_model:\n",
+        "\n",
+        "mlflow_experiment_name: colab-example\n",
+        "\n",
+        "gradient_accumulation_steps: 1\n",
+        "micro_batch_size: 1\n",
+        "num_epochs: 4\n",
+        "max_steps: 20\n",
+        "optimizer: paged_adamw_32bit\n",
+        "lr_scheduler: cosine\n",
+        "learning_rate: 0.0002\n",
+        "\n",
+        "train_on_inputs: false\n",
+        "group_by_length: false\n",
+        "bf16: false\n",
+        "fp16: true\n",
+        "tf32: false\n",
+        "\n",
+        "gradient_checkpointing: true\n",
+        "early_stopping_patience:\n",
+        "resume_from_checkpoint:\n",
+        "local_rank:\n",
+        "logging_steps: 1\n",
+        "xformers_attention:\n",
+        "flash_attention: false\n",
+        "\n",
+        "warmup_steps: 10\n",
+        "evals_per_epoch:\n",
+        "saves_per_epoch:\n",
+        "debug:\n",
+        "deepspeed:\n",
+        "weight_decay: 0.0\n",
+        "fsdp:\n",
+        "fsdp_config:\n",
+        "special_tokens:\n",
+        "\n",
+        "\"\"\"\n",
+        "\n",
+        "# Convert the YAML string to a Python dictionary\n",
+        "yaml_dict = yaml.safe_load(yaml_string)\n",
+        "\n",
+        "# Specify your file path\n",
+        "file_path = 'test_axolotl.yaml'\n",
+        "\n",
+        "# Write the YAML file\n",
+        "with open(file_path, 'w') as file:\n",
+        "    yaml.dump(yaml_dict, file)\n"
+      ]
+    },
+    {
+      "cell_type": "markdown",
+      "metadata": {
+        "id": "bidoj8YLTusD"
+      },
+      "source": [
+        "## Launch the training"
+      ]
+    },
+    {
+      "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "id": "ydTI2Jk2RStU",
+        "outputId": "d6d0df17-4b53-439c-c802-22c0456d301b"
+      },
+      "outputs": [],
+      "source": [
+        "# Buy using the ! the comand will be executed as a bash command\n",
+        "!accelerate launch -m axolotl.cli.train /content/test_axolotl.yaml"
+      ]
+    }
+  ],
+  "metadata": {
+    "accelerator": "GPU",
+    "colab": {
+      "gpuType": "T4",
+      "provenance": []
+    },
+    "kernelspec": {
+      "display_name": "Python 3",
+      "name": "python3"
+    },
+    "language_info": {
+      "name": "python"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 0
+}

examples/llama-2/fft_optimized.yml

@@ -67,6 +67,3 @@ weight_decay: 0.1
 fsdp:
 fsdp_config:
 special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

examples/llama-2/loftq.yml

@@ -0,0 +1,70 @@
+base_model: NousResearch/Llama-2-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./lora-out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+peft:
+  loftq_config:
+    loftq_bits: 4
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+s2_attention:
+
+warmup_steps: 10
+evals_per_epoch: 4
+eval_table_size:
+eval_table_max_new_tokens: 128
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

examples/llama-2/lora.yml

@@ -65,6 +65,3 @@ weight_decay: 0.0
 fsdp:
 fsdp_config:
 special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

examples/llama-2/qlora.yml

@@ -65,6 +65,3 @@ weight_decay: 0.0
 fsdp:
 fsdp_config:
 special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

requirements.txt

@@ -1,6 +1,6 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
-peft==0.7.0
+peft @ git+https://github.com/huggingface/peft.git
 transformers==4.37.0
 tokenizers==0.15.0
 bitsandbytes>=0.41.1
@@ -15,16 +15,14 @@ sentencepiece
 wandb
 einops
 xformers==0.0.22
-optimum==1.13.2
+optimum==1.16.2
 hf_transfer
 colorama
 numba
 numpy>=1.24.4
 mlflow
 # qlora things
-bert-score==0.3.13
 evaluate==0.4.0
-rouge-score==0.1.2
 scipy
 scikit-learn==1.2.2
 pynvml

setup.py

@@ -27,9 +27,10 @@ def parse_requirements():
 
     try:
         torch_version = version("torch")
-        if torch_version.startswith("2.1.1"):
+        _install_requires.append(f"torch=={torch_version}")
+        if torch_version.startswith("2.1."):
             _install_requires.pop(_install_requires.index("xformers==0.0.22"))
-            _install_requires.append("xformers==0.0.23")
+            _install_requires.append("xformers>=0.0.23")
     except PackageNotFoundError:
         pass
 
@@ -50,7 +51,7 @@ setup(
     dependency_links=dependency_links,
     extras_require={
         "flash-attn": [
-            "flash-attn==2.3.3",
+            "flash-attn==2.5.0",
         ],
         "fused-dense-lib": [
             "fused-dense-lib  @ git+https://github.com/Dao-AILab/flash-attention@v2.3.3#subdirectory=csrc/fused_dense_lib",

src/axolotl/cli/preprocess.py

@@ -18,6 +18,7 @@ from axolotl.cli import (
 )
 from axolotl.common.cli import PreprocessCliArgs
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
+from axolotl.prompt_strategies.sharegpt import register_chatml_template
 
 LOG = logging.getLogger("axolotl.cli.preprocess")
 
@@ -34,6 +35,14 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
         return_remaining_strings=True
     )
 
+    if parsed_cfg.chat_template == "chatml" and parsed_cfg.default_system_message:
+        LOG.info(
+            f"ChatML set. Adding default system message: {parsed_cfg.default_system_message}"
+        )
+        register_chatml_template(parsed_cfg.default_system_message)
+    else:
+        register_chatml_template()
+
     if not parsed_cfg.dataset_prepared_path:
         msg = (
             Fore.RED

src/axolotl/cli/train.py

@@ -18,6 +18,7 @@ from axolotl.cli import (
     print_axolotl_text_art,
 )
 from axolotl.common.cli import TrainerCliArgs
+from axolotl.prompt_strategies.sharegpt import register_chatml_template
 from axolotl.train import train
 
 LOG = logging.getLogger("axolotl.cli.train")
@@ -37,6 +38,14 @@ def do_train(cfg, cli_args) -> Tuple[PreTrainedModel, PreTrainedTokenizer]:
     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.rl:
         dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)
     else:

src/axolotl/core/trainer_builder.py

@@ -59,6 +59,22 @@ except ImportError:
 LOG = logging.getLogger("axolotl.core.trainer_builder")
 
 
+def _sanitize_kwargs_for_tagging(tag_names, kwargs=None):
+    if isinstance(tag_names, str):
+        tag_names = [tag_names]
+
+    if kwargs is not None:
+        if "tags" not in kwargs:
+            kwargs["tags"] = tag_names
+        elif "tags" in kwargs and isinstance(kwargs["tags"], list):
+            kwargs["tags"].extend(tag_names)
+        elif "tags" in kwargs and isinstance(kwargs["tags"], str):
+            tag_names.append(kwargs["tags"])
+            kwargs["tags"] = tag_names
+
+    return kwargs
+
+
 @dataclass
 class AxolotlTrainingArguments(TrainingArguments):
     """
@@ -170,24 +186,30 @@ class AxolotlTrainer(Trainer):
             num_training_steps (int): The number of training steps to do.
             optimizer (torch.optim.Optimizer): The training optimizer
         """
+        use_cosine_quadratic = (
+            self.args.lr_scheduler_type == "cosine"
+            and self.args.lr_quadratic_warmup is True
+        )
+
+        use_cosine_min_lr = (
+            self.args.lr_scheduler_type == "cosine"
+            and self.args.cosine_min_lr_ratio is not None
+        )
 
         # fmt: off
         if self.lr_scheduler is None:  # type: ignore  # pylint: disable=access-member-before-definition
             # fmt: on
-            if (
-                self.args.lr_scheduler_type == "cosine"
-                and self.args.lr_quadratic_warmup is True
-            ):
+            if use_cosine_quadratic:
+                if use_cosine_min_lr:
+                    LOG.warning("Both cosine quadratic warmup and min lr detected. Using quadratic warmup.")
+
                 self.lr_scheduler = get_cosine_schedule_with_quadratic_warmup(  # pylint: disable=attribute-defined-outside-init
                     optimizer,
                     num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
                     num_training_steps=num_training_steps,
                 )
-            elif self.args.lr_scheduler_type == "cosine" and self.args.cosine_min_lr_ratio is not None:
+            elif self.args.cosine_min_lr_ratio and use_cosine_min_lr:
                 assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
-                if self.args.deepspeed:
-                    LOG.warning("Using cosine scheduler with deepspeed. This may be ignored if a scheduler is set \
-                                in the deepspeed JSON")
                 self.lr_scheduler = get_cosine_schedule_with_min_lr(  # pylint: disable=attribute-defined-outside-init
                     optimizer,
                     num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
@@ -196,6 +218,13 @@ class AxolotlTrainer(Trainer):
                 )
             else:
                 return super().create_scheduler(num_training_steps, optimizer)
+        else:
+            if use_cosine_quadratic:
+                LOG.warning("axolotl's cosine scheduler with quadratic warmup not used (e.g., because of deepspeed).")
+
+            if use_cosine_min_lr:
+                LOG.warning("axolotl's cosine scheduler with min lr not used (e.g., because of deepspeed).")
+
         return self.lr_scheduler
 
     def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
@@ -336,30 +365,13 @@ class AxolotlTrainer(Trainer):
         #     return (loss, outputs) if return_outputs else loss
         return super().compute_loss(model, inputs, return_outputs=return_outputs)
 
-    def _sanitize_kwargs_for_tagging(self, tag_names, kwargs=None):
-        if isinstance(tag_names, str):
-            tag_names = [tag_names]
-
-        if kwargs is not None:
-            if "tags" not in kwargs:
-                kwargs["tags"] = tag_names
-            elif "tags" in kwargs and isinstance(kwargs["tags"], list):
-                kwargs["tags"].extend(tag_names)
-            elif "tags" in kwargs and isinstance(kwargs["tags"], str):
-                tag_names.append(kwargs["tags"])
-                kwargs["tags"] = tag_names
-
-        return kwargs
-
     @wraps(Trainer.push_to_hub)
     def push_to_hub(self, *args, **kwargs) -> str:
         """
         Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
         model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
         """
-        kwargs = self._sanitize_kwargs_for_tagging(
-            tag_names=self.tag_names, kwargs=kwargs
-        )
+        kwargs = _sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
 
         return super().push_to_hub(*args, **kwargs)
 
@@ -458,6 +470,24 @@ class ReLoRATrainer(AxolotlTrainer):
         return self.lr_scheduler
 
 
+class AxolotlDPOTrainer(DPOTrainer):
+    """
+    Extend the base DPOTrainer for axolotl helpers
+    """
+
+    tag_names = ["axolotl", "dpo"]
+
+    @wraps(DPOTrainer.push_to_hub)
+    def push_to_hub(self, *args, **kwargs) -> str:
+        """
+        Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
+        model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
+        """
+        kwargs = _sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
+
+        return super().push_to_hub(*args, **kwargs)
+
+
 class TrainerBuilderBase(abc.ABC):
     """
     Base class for trainer builder
@@ -638,7 +668,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs[
                 "gradient_checkpointing"
             ] = self.cfg.gradient_checkpointing
-            if self.cfg.gradient_checkpointing_kwargs:
+            if self.cfg.gradient_checkpointing_kwargs is not None:
                 training_arguments_kwargs[
                     "gradient_checkpointing_kwargs"
                 ] = self.cfg.gradient_checkpointing_kwargs
@@ -705,7 +735,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         elif self.cfg.sample_packing and self.cfg.eval_sample_packing is False:
             training_arguments_kwargs["dataloader_drop_last"] = True
 
-        if self.cfg.val_set_size == 0:
+        if not self.cfg.test_datasets and self.cfg.val_set_size == 0:
             # no eval set, so don't eval
             training_arguments_kwargs["evaluation_strategy"] = "no"
         elif self.cfg.eval_steps:
@@ -792,6 +822,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 self.cfg.load_best_model_at_end is not False
                 or self.cfg.early_stopping_patience
             )
+            and not self.cfg.test_datasets
             and self.cfg.val_set_size > 0
             and self.cfg.save_steps
             and self.cfg.eval_steps
@@ -1015,6 +1046,18 @@ class HFDPOTrainerBuilder(TrainerBuilderBase):
             training_args_kwargs[
                 "dataloader_prefetch_factor"
             ] = self.cfg.dataloader_prefetch_factor
+        if self.cfg.gradient_checkpointing:
+            training_args_kwargs[
+                "gradient_checkpointing"
+            ] = self.cfg.gradient_checkpointing
+            if self.cfg.gradient_checkpointing_kwargs is not None:
+                training_args_kwargs[
+                    "gradient_checkpointing_kwargs"
+                ] = self.cfg.gradient_checkpointing_kwargs
+            else:
+                training_args_kwargs["gradient_checkpointing_kwargs"] = {
+                    "use_reentrant": False
+                }
 
         training_args = TrainingArguments(
             per_device_train_batch_size=self.cfg.micro_batch_size,
@@ -1025,9 +1068,6 @@ class HFDPOTrainerBuilder(TrainerBuilderBase):
             save_steps=self.cfg.save_steps,
             output_dir=self.cfg.output_dir,
             warmup_steps=self.cfg.warmup_steps,
-            gradient_checkpointing=self.cfg.gradient_checkpointing,
-            gradient_checkpointing_kwargs=self.cfg.gradient_checkpointing_kwargs
-            or {"use_reentrant": False},
             logging_first_step=True,
             logging_steps=1,
             optim=self.cfg.optimizer,
@@ -1050,7 +1090,11 @@ class HFDPOTrainerBuilder(TrainerBuilderBase):
             dpo_trainer_kwargs["eval_dataset"] = self.eval_dataset
         if self.cfg.adapter and self.peft_config:
             dpo_trainer_kwargs["peft_config"] = self.peft_config
-        dpo_trainer = DPOTrainer(
+        if self.cfg.precompute_ref_log_probs is not None:
+            dpo_trainer_kwargs[
+                "precompute_ref_log_probs"
+            ] = self.cfg.precompute_ref_log_probs
+        dpo_trainer = AxolotlDPOTrainer(
             self.model,
             self.model_ref,
             args=training_args,

src/axolotl/monkeypatch/mistral_attn_hijack_flash.py

@@ -94,7 +94,7 @@ def _prepare_decoder_attention_mask(
     sliding_window,
 ):  # pylint: disable=unused-argument
     # [bsz, seq_len]
-    if attention_mask is None:
+    if attention_mask is None or sliding_window is None:
         return attention_mask
 
     # NOTE: attention mask and sliding masks are only broadcastable in certain scenarios.
@@ -151,7 +151,7 @@ def flashattn_forward(
     )
 
     use_sliding_windows = (
-        hasattr(self.config, "sliding_window") is not None
+        getattr(self.config, "sliding_window") is not None
         and kv_seq_len > self.config.sliding_window
     )
 

src/axolotl/prompt_strategies/dpo/chatml.py

@@ -23,6 +23,31 @@ def argilla(
     return transform_fn
 
 
+def icr(
+    cfg,
+):  # pylint: disable=possibly-unused-variable,unused-argument
+    """
+    chatml transforms for datasets with system, input, chosen, rejected
+    ex. https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs
+    """
+
+    def transform_fn(sample):
+        if "system" in sample and sample["system"]:
+            sample["prompt"] = (
+                f"<|im_start|>system\n{sample['system']}<|im_end|>\n"
+                f"<|im_start|>user\n{sample['input']}<|im_end|>\n<|im_start|>assistant\n"
+            )
+        else:
+            sample[
+                "prompt"
+            ] = f"<|im_start|>user\n{sample['input']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["chosen"] = f"{sample['chosen']}<|im_end|>"
+        sample["rejected"] = f"{sample['rejected']}<|im_end|>"
+        return sample
+
+    return transform_fn
+
+
 def intel(cfg):  # pylint: disable=possibly-unused-variable,unused-argument
     """
     For Intel Orca DPO Pairs

src/axolotl/prompt_strategies/sharegpt.py

@@ -6,16 +6,19 @@ from fastchat.conversation import Conversation, SeparatorStyle, register_conv_te
 from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
 from axolotl.prompters import ShareGPTPrompterV2
 
-register_conv_template(
-    Conversation(
-        name="chatml",
-        system_template="<|im_start|>system\n{system_message}",
-        system_message="You are a helpful assistant.",
-        roles=["<|im_start|>user", "<|im_start|>assistant"],
-        sep_style=SeparatorStyle.CHATML,
-        sep="<|im_end|>",
+
+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|>",
+        )
     )
-)
 
 
 def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):

src/axolotl/train.py

@@ -63,6 +63,8 @@ def train(
         msg += " and peft_config..."
     LOG.debug(msg)
     model, peft_config = load_model(cfg, tokenizer, inference=cli_args.inference)
+    model.generation_config.do_sample = True
+
     model_ref = None
     if cfg.rl:
         if cfg.adapter and not cfg.rl_adapter_ref_model:

src/axolotl/utils/chat_templates.py

@@ -20,7 +20,7 @@ def chat_templates(user_choice: str):
 
     templates = {
         "inst": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",  # I don't know what this one is called. Used by Mistral/Mixtral.
-        "chatml": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+        "chatml": "{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% set system_message = 'You are a helpful assistant.' %}{% endif %}{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in loop_messages %}{% if loop.index0 == 0 %}{{'<|im_start|>system\n' + system_message + '<|im_end|>\n'}}{% endif %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
     }
 
     if user_choice in templates:

src/axolotl/utils/config.py

@@ -95,7 +95,7 @@ def normalize_config(cfg):
         save_steps = 1.0 / (cfg.saves_per_epoch * cfg.num_epochs)
         if save_steps < 1.0:  # prevent saves on every step
             cfg.save_steps = save_steps
-    if cfg.evals_per_epoch:
+    if (cfg.val_set_size or cfg.test_datasets) and cfg.evals_per_epoch:
         eval_steps = 1.0 / (cfg.evals_per_epoch * cfg.num_epochs)
         if eval_steps < 1.0:  # prevent evals on every step
             cfg.eval_steps = eval_steps
@@ -163,6 +163,7 @@ def normalize_config(cfg):
         cfg.gradient_checkpointing
         and cfg.unfrozen_parameters is None
         and cfg.gradient_checkpointing_kwargs is None
+        and cfg.rl is None
     ):
         cfg.gradient_checkpointing_kwargs = {"use_reentrant": True}
 
@@ -231,9 +232,6 @@ def validate_config(cfg):
             "eval_batch_size != micro_batch_size. This can lead to VRAM instability."
         )
 
-    if cfg.load_4bit:
-        raise ValueError("cfg.load_4bit parameter has been deprecated")
-
     if cfg.adapter == "qlora":
         if cfg.merge_lora:
             # can't merge qlora if loaded in 8bit or 4bit
@@ -259,7 +257,8 @@ def validate_config(cfg):
         if cfg.flash_attn_fuse_qkv or cfg.flash_attn_fuse_mlp:
             raise ValueError("Fused modules are not supported with QLoRA")
 
-    if not cfg.load_in_8bit and cfg.adapter == "lora":
+    loftq = cfg.peft and cfg.peft.loftq_config and cfg.peft.loftq_config.loftq_bits
+    if not cfg.load_in_8bit and cfg.adapter == "lora" and not loftq:
         LOG.warning("We recommend setting `load_in_8bit: true` for LORA finetuning")
 
     if cfg.adapter == "lora" and (cfg.flash_attn_fuse_qkv or cfg.flash_attn_fuse_mlp):
@@ -339,6 +338,11 @@ def validate_config(cfg):
             "push_to_hub_model_id is deprecated. Please use hub_model_id instead."
         )
 
+    if cfg.hub_model_id and not (cfg.save_steps or cfg.saves_per_epoch):
+        LOG.warning(
+            "hub_model_id is set without any models being saved. To save a model, set either save_steps or saves_per_epoch."
+        )
+
     if cfg.gptq and cfg.model_revision:
         raise ValueError(
             "model_revision is not supported for GPTQ models. "
@@ -484,35 +488,43 @@ def validate_config(cfg):
             "`use_reentrant` must be false when used with partially frozen model."
         )
 
-    if cfg.flash_attention and cfg.deepspeed and Path(cfg.deepspeed).is_file():
+    if cfg.deepspeed and Path(cfg.deepspeed).is_file():
         with open(cfg.deepspeed, encoding="utf-8") as file:
             contents = file.read()
             deepspeed_cfg: DictDefault = DictDefault(json.loads(contents))
-            if (
-                deepspeed_cfg.zero_optimization
-                and deepspeed_cfg.zero_optimization.stage == 3
-            ):
-                if not (
-                    (
-                        deepspeed_cfg.bf16
-                        and deepspeed_cfg.bf16.enabled  # pylint: disable=no-member
-                        is True
-                    )
-                    or (
-                        deepspeed_cfg.fp16
-                        and deepspeed_cfg.fp16.enabled  # pylint: disable=no-member
-                        is True
-                    )
+            if cfg.flash_attention:
+                if (
+                    deepspeed_cfg.zero_optimization
+                    and deepspeed_cfg.zero_optimization.stage == 3
                 ):
-                    raise ValueError(
-                        "bf16.enabled or fp16.enabled must be set to true when using ZeRO-3 with flash-attention"
-                    )
+                    if not (
+                        (
+                            deepspeed_cfg.bf16
+                            and deepspeed_cfg.bf16.enabled  # pylint: disable=no-member
+                            is True
+                        )
+                        or (
+                            deepspeed_cfg.fp16
+                            and deepspeed_cfg.fp16.enabled  # pylint: disable=no-member
+                            is True
+                        )
+                    ):
+                        raise ValueError(
+                            "bf16.enabled or fp16.enabled must be set to true when using ZeRO-3 with flash-attention"
+                        )
+            if "8bit" in cfg.optimizer and deepspeed_cfg.optimizer:
+                LOG.warning(
+                    f"conflicting optimizer: {cfg.optimizer} used alongside deepspeed optimizer."
+                )
 
     if cfg.test_datasets and cfg.val_set_size:
         raise ValueError(
             "non-zero val_set_size should not be used with test_datasets configuration"
         )
 
+    if cfg.fsdp and "bnb" in cfg.optimizer:
+        raise ValueError(f"FSDP not compatible with {cfg.optimizer}")
+
     # TODO
     # MPT 7b
     # https://github.com/facebookresearch/bitsandbytes/issues/25

src/axolotl/utils/data.py

@@ -16,6 +16,7 @@ from datasets import (
     load_from_disk,
 )
 from huggingface_hub import hf_hub_download
+from huggingface_hub.utils import HFValidationError
 from torch.utils.data import RandomSampler
 from transformers import PreTrainedTokenizerBase
 
@@ -213,7 +214,7 @@ def load_tokenized_prepared_datasets(
                     token=use_auth_token,
                 )
                 ds_from_hub = True
-            except (FileNotFoundError, ConnectionError):
+            except (FileNotFoundError, ConnectionError, HFValidationError):
                 pass
 
             ds_from_cloud = False
@@ -439,7 +440,7 @@ def load_prepare_datasets(
     split="train",
 ) -> Tuple[Dataset, Dataset, List[Prompter]]:
     dataset, prompters = load_tokenized_prepared_datasets(
-        tokenizer, cfg, default_dataset_prepared_path
+        tokenizer, cfg, default_dataset_prepared_path, split=split
     )
 
     if cfg.dataset_shard_num and cfg.dataset_shard_idx is not None:

src/axolotl/utils/models.py

@@ -9,7 +9,7 @@ import bitsandbytes as bnb
 import torch
 import transformers
 from optimum.bettertransformer import BetterTransformer
-from peft import PeftConfig, prepare_model_for_kbit_training
+from peft import LoftQConfig, PeftConfig, prepare_model_for_kbit_training
 from peft.tuners.lora import QuantLinear
 from transformers import (  # noqa: F401
     AddedToken,
@@ -67,7 +67,7 @@ def check_model_config(cfg: DictDefault, model_config: Union[AutoConfig, DictDef
     ):
         lora_modules_to_save = ", ".join(map(lambda x: f"`{x}`", lora_modules_to_save))
         raise ValueError(
-            f"`lora_modules_to_save` not properly set when adding new tokens. Please include {lora_modules_to_save} in `lora_modules_to_save`."
+            f"`lora_modules_to_save` not properly set when adding new tokens. Please include [{lora_modules_to_save}] in `lora_modules_to_save`."
         )
 
 
@@ -182,7 +182,7 @@ def load_tokenizer(cfg):
                     [f"`{x}`" for x in lora_modules_to_save]
                 )
                 raise ValueError(
-                    f"Please set lora_modules_to_save to {lora_modules_to_save} when using an adapter and changing the special tokens."
+                    f"Please set lora_modules_to_save to [{lora_modules_to_save}] when using an adapter and changing the special tokens."
                 )
 
             tokenizer.add_special_tokens(
@@ -219,7 +219,13 @@ def load_tokenizer(cfg):
     LOG.debug(f"UNK: {tokenizer.unk_token_id} / {tokenizer.unk_token}")
 
     if cfg.chat_template:
-        tokenizer.chat_template = chat_templates(cfg.chat_template)
+        chat_template_string = chat_templates(cfg.chat_template)
+        if cfg.default_system_message and cfg.chat_template == "chatml":
+            chat_template_string = chat_template_string.replace(
+                "You are a helpful assistant.", cfg.default_system_message
+            )
+
+        tokenizer.chat_template = chat_template_string
     else:
         LOG.info(
             "No Chat template selected. Consider adding a chat template for easier inference."
@@ -636,15 +642,17 @@ def load_model(
 
     # make sure these are fp32 per Ramesh et al. (2021)
     embedding_modules = get_linear_embedding_layers(cfg.model_config_type)
-    for name, module in model.named_modules():
-        if any(m in name for m in ["norm", "gate"]):
-            module.to(torch.float32)
-        if model_config.model_type == "btlm":
-            # don't upcast lm_head for btlm
-            continue
-        if any(m in name for m in embedding_modules):
-            if hasattr(module, "weight"):
+    if not cfg.fsdp:
+        # FSDP doesn't like mixed Float and BFloat16
+        for name, module in model.named_modules():
+            if any(m in name for m in ["norm", "gate"]):
                 module.to(torch.float32)
+            if model_config.model_type == "btlm":
+                # don't upcast lm_head for btlm
+                continue
+            if any(m in name for m in embedding_modules):
+                if hasattr(module, "weight"):
+                    module.to(torch.float32)
 
     needs_fa2_dtype = cfg.adapter or cfg.fsdp
     skip_prepare_model_for_kbit_training = False
@@ -659,13 +667,17 @@ def load_model(
         # Qwen doesn't play nicely with LoRA if this is enabled
         skip_prepare_model_for_kbit_training = True
 
-    if (cfg.adapter == "lora" and load_in_8bit) or (
-        cfg.adapter == "qlora" and cfg.load_in_4bit
-    ):
-        LOG.info("converting PEFT model w/ prepare_model_for_kbit_training")
+    loftq_bits = cfg.peft and cfg.peft.loftq_config and cfg.peft.loftq_config.loftq_bits
+    if cfg.adapter == "lora" and loftq_bits:
+        skip_prepare_model_for_kbit_training = True
+
+    if cfg.adapter in ["lora", "qlora"]:
         if cfg.gradient_checkpointing:
             model.gradient_checkpointing_enable()
-        if not skip_prepare_model_for_kbit_training:
+        if (
+            cfg.load_in_8bit or cfg.load_in_4bit
+        ) and not skip_prepare_model_for_kbit_training:
+            LOG.info("converting PEFT model w/ prepare_model_for_kbit_training")
             model = prepare_model_for_kbit_training(
                 model, use_gradient_checkpointing=cfg.gradient_checkpointing
             )
@@ -692,6 +704,7 @@ def load_model(
             model, lora_config = load_adapter(model, cfg, cfg.adapter)
 
     if cfg.ddp and not load_in_8bit and not (cfg.rl and cfg.load_in_4bit):
+        # TODO revaldate this conditional
         model.to(f"cuda:{cfg.local_rank}")
 
     if torch.cuda.device_count() > 1 and int(os.getenv("WORLD_SIZE", "1")) == 1:
@@ -743,7 +756,7 @@ def load_llama_adapter(model, cfg):
     )
 
     if cfg.lora_model_dir:
-        LOG.debug("Loading pretained PEFT - llama_adapter")
+        LOG.debug("Loading pretrained PEFT - llama_adapter")
         model = PeftModel.from_pretrained(
             model,
             cfg.lora_model_dir,
@@ -789,6 +802,12 @@ def load_lora(model, cfg, inference=False, config_only=False):
         LOG.info(f"found linear modules: {repr(linear_names)}")
         lora_target_modules = list(set(lora_target_modules + linear_names))
 
+    lora_config_kwargs = {}
+    loftq_bits = cfg.peft and cfg.peft.loftq_config and cfg.peft.loftq_config.loftq_bits
+    if loftq_bits:
+        lora_config_kwargs["loftq_config"] = LoftQConfig(loftq_bits=loftq_bits)
+        lora_config_kwargs["init_lora_weights"] = "loftq"
+
     lora_config = LoraConfig(
         r=cfg.lora_r,
         lora_alpha=cfg.lora_alpha,
@@ -799,13 +818,14 @@ def load_lora(model, cfg, inference=False, config_only=False):
         modules_to_save=cfg.lora_modules_to_save if cfg.lora_modules_to_save else None,
         bias="none",
         task_type="CAUSAL_LM",
+        **lora_config_kwargs,
     )
 
     if config_only:
         return None, lora_config
 
     if cfg.lora_model_dir:
-        LOG.debug("Loading pretained PEFT - LoRA")
+        LOG.debug("Loading pretrained PEFT - LoRA")
         model_kwargs: Any = {}
         if cfg.lora_on_cpu:
             model_kwargs["max_memory"] = {"cpu": "256GiB"}

tests/prompt_strategies/test_sharegpt.py

@@ -7,9 +7,14 @@ from tokenizers import AddedToken
 from transformers import AutoTokenizer
 
 from axolotl.datasets import TokenizedPromptDataset
-from axolotl.prompt_strategies.sharegpt import SimpleShareGPTPromptTokenizingStrategy
+from axolotl.prompt_strategies.sharegpt import (
+    SimpleShareGPTPromptTokenizingStrategy,
+    register_chatml_template,
+)
 from axolotl.prompters import ShareGPTPrompterV2
 
+register_chatml_template()
+
 
 @pytest.fixture(name="sharegpt_dataset")
 def fixture_sharegpt_dataset():

tests/test_validation.py

@@ -26,21 +26,12 @@ class BaseValidation(unittest.TestCase):
         self._caplog = caplog
 
 
+# pylint: disable=too-many-public-methods
 class ValidationTest(BaseValidation):
     """
     Test the validation module
     """
 
-    def test_load_4bit_deprecate(self):
-        cfg = DictDefault(
-            {
-                "load_4bit": True,
-            }
-        )
-
-        with pytest.raises(ValueError):
-            validate_config(cfg)
-
     def test_batch_size_unused_warning(self):
         cfg = DictDefault(
             {
@@ -698,6 +689,22 @@ class ValidationTest(BaseValidation):
         ):
             validate_config(cfg)
 
+    def test_hub_model_id_save_value_warns(self):
+        cfg = DictDefault({"hub_model_id": "test"})
+
+        with self._caplog.at_level(logging.WARNING):
+            validate_config(cfg)
+            assert (
+                "set without any models being saved" in self._caplog.records[0].message
+            )
+
+    def test_hub_model_id_save_value(self):
+        cfg = DictDefault({"hub_model_id": "test", "saves_per_epoch": 4})
+
+        with self._caplog.at_level(logging.WARNING):
+            validate_config(cfg)
+            assert len(self._caplog.records) == 0
+
 
 class ValidationCheckModelConfig(BaseValidation):
     """