Fix(debug): Use space delimiter for debug_text_only also

* Update tests-docker.yml

* Update tests-docker.yml

* run ci tests on ci yaml updates

---------

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

.github/workflows/base.yml

@@ -28,7 +28,12 @@ jobs:
           - cuda: "118"
             cuda_version: 11.8.0
             python_version: "3.10"
-            pytorch: 2.1.0
+            pytorch: 2.1.1
+            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
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
     steps:
       - name: Checkout

.github/workflows/main.yml

@@ -27,38 +27,56 @@ jobs:
           - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.10"
-            pytorch: 2.1.0
+            pytorch: 2.1.1
+            axolotl_extras:
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.10"
+            pytorch: 2.1.1
             axolotl_extras:
     runs-on: [self-hosted, gpu, docker]
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Docker metadata
         id: metadata
-        uses: docker/metadata-action@v3
+        uses: docker/metadata-action@v5
         with:
           images: winglian/axolotl
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
       - name: Login to Docker Hub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
-      - name: Build
-        uses: docker/build-push-action@v4
+      # guidance for testing before pushing: https://docs.docker.com/build/ci/github-actions/test-before-push/
+      - name: Build and export to Docker
+        uses: docker/build-push-action@v5
         with:
           context: .
+          load: true
           build-args: |
             BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
             CUDA=${{ matrix.cuda }}
             PYTORCH_VERSION=${{ matrix.pytorch }}
           file: ./docker/Dockerfile
-          push: ${{ github.event_name != 'pull_request' }}
           tags: |
             ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
             ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
           labels: ${{ steps.metadata.outputs.labels }}
+      - name: Unit Tests
+        run: |
+          docker run --rm ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }} pytest --ignore=tests/e2e/ /workspace/axolotl/tests/
+      - name: Push to Docker Hub
+        if: github.event_name != 'pull_request'
+        run: |
+          docker push ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+          latest_tag=${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
+          if [ -n "$latest_tag" ]; then
+            docker push "$latest_tag"
+          fi
+
   build-axolotl-runpod:
     needs: build-axolotl
     if: github.repository_owner == 'OpenAccess-AI-Collective'
@@ -80,26 +98,31 @@ jobs:
           - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.10"
-            pytorch: 2.1.0
+            pytorch: 2.1.1
+            axolotl_extras:
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.10"
+            pytorch: 2.1.1
             axolotl_extras:
     runs-on: [self-hosted, gpu, docker]
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Docker metadata
         id: metadata
-        uses: docker/metadata-action@v3
+        uses: docker/metadata-action@v5
         with:
           images: winglian/axolotl-runpod
       - name: Login to Docker Hub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v2
       - name: Build
-        uses: docker/build-push-action@v4
+        uses: docker/build-push-action@v5
         with:
           context: .
           build-args: |

.github/workflows/tests-docker.yml

@@ -0,0 +1,46 @@
+name: e2e-docker-tests
+
+on:
+  pull_request:
+    paths:
+      - '**.py'
+      - 'requirements.txt'
+      - '.github/workflows/*.yml'
+  workflow_dispatch:
+
+jobs:
+  build-axolotl:
+    if: github.repository_owner == 'OpenAccess-AI-Collective'
+    # this job needs to be run on self-hosted GPU runners...
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 118
+            cuda_version: 11.8.0
+            python_version: "3.10"
+            pytorch: 2.0.1
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.10"
+            pytorch: 2.1.1
+    runs-on: [self-hosted, gpu, docker]
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Build Docker image
+        run: |
+          # Set up build arguments
+          BASE_TAG="main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}"
+          CUDA="${{ matrix.cuda }}"
+          PYTORCH_VERSION="${{ matrix.pytorch }}"
+          # Build the Docker image
+          docker build . \
+            --file ./docker/Dockerfile \
+            --build-arg BASE_TAG=$BASE_TAG \
+            --build-arg CUDA=$CUDA \
+            --build-arg PYTORCH_VERSION=$PYTORCH_VERSION \
+            --tag test-axolotl
+      - name: Unit Tests w docker image
+        run: |
+          docker run --rm test-axolotl pytest --ignore=tests/e2e/ /workspace/axolotl/tests/

.github/workflows/tests.yml

@@ -73,7 +73,7 @@ jobs:
         run: |
           pip3 install --extra-index-url https://download.pytorch.org/whl/cu118 -U torch==2.0.1
           pip3 uninstall -y transformers accelerate
-          pip3 install -U -e .[flash-attn]
+          pip3 install -U -e .[flash-attn,mamba-ssm]
           pip3 install -r requirements-tests.txt
 
       - name: Run e2e tests

.mypy.ini

@@ -8,6 +8,9 @@ ignore_missing_imports = True
 [mypy-axolotl.monkeypatch.*]
 ignore_errors = True
 
+[mypy-axolotl.models.mixtral.*]
+ignore_errors = True
+
 [mypy-axolotl.models.phi.*]
 ignore_errors = True
 

README.md

@@ -36,7 +36,9 @@ Features:
   - [Train](#train)
   - [Inference](#inference)
   - [Merge LORA to Base](#merge-lora-to-base)
+  - [Special Tokens](#special-tokens)
 - [Common Errors](#common-errors-)
+  - [Tokenization Mismatch b/w Training & Inference](#tokenization-mismatch-bw-inference--training)
 - [Need Help?](#need-help-)
 - [Badge](#badge-)
 - [Community Showcase](#community-showcase)
@@ -65,19 +67,21 @@ Features:
 
 ## Axolotl supports
 
-|          | fp16/fp32 | lora | qlora | gptq | gptq w/flash attn | flash attn | xformers attn |
-|----------|:----------|:-----|-------|------|-------------------|------------|--------------|
-| llama    | ✅         | ✅    | ✅     | ✅             | ✅                 | ✅          | ✅            |
-| Pythia   | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
-| cerebras | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
-| btlm     | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
-| mpt      | ✅         | ❌    | ❓     | ❌             | ❌                 | ❌          | ❓            |
-| falcon   | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
-| gpt-j    | ✅         | ✅    | ✅     | ❌             | ❌                 | ❓          | ❓            |
-| XGen     | ✅         | ❓    | ✅     | ❓             | ❓                 | ❓          | ✅            |
-| phi      | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
-| RWKV     | ✅         | ❓    | ❓     | ❓             | ❓                 | ❓          | ❓            |
-| Qwen     | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
+|             | fp16/fp32 | lora | qlora | gptq | gptq w/flash attn | flash attn | xformers attn |
+|-------------|:----------|:-----|-------|------|-------------------|------------|--------------|
+| llama       | ✅         | ✅    | ✅     | ✅             | ✅                 | ✅          | ✅            |
+| Mistral     | ✅         | ✅    | ✅     | ✅             | ✅                 | ✅          | ✅            |
+| Mixtral-MoE | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
+| Pythia      | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
+| cerebras    | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
+| btlm        | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
+| mpt         | ✅         | ❌    | ❓     | ❌             | ❌                 | ❌          | ❓            |
+| falcon      | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
+| gpt-j       | ✅         | ✅    | ✅     | ❌             | ❌                 | ❓          | ❓            |
+| XGen        | ✅         | ❓    | ✅     | ❓             | ❓                 | ❓          | ✅            |
+| phi         | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
+| RWKV        | ✅         | ❓    | ❓     | ❓             | ❓                 | ❓          | ❓            |
+| Qwen        | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
 
 
 ## Quickstart ⚡
@@ -245,10 +249,17 @@ Have dataset(s) in one of the following format (JSONL recommended):
   ```json
   {"instruction": "...", "input": "...", "output": "..."}
   ```
-- `sharegpt`: conversations where `from` is `human`/`gpt`
+- `sharegpt`: conversations where `from` is `human`/`gpt`. (optional: `system` to override default system prompt)
   ```json
   {"conversations": [{"from": "...", "value": "..."}]}
   ```
+- `llama-2`: the json is the same format as `sharegpt` above, with the following config (see the [config section](#config) for more details)
+    ```yml
+    datasets:
+      - path: <your-path>
+        type: sharegpt
+        conversation: llama-2
+    ```
 - `completion`: raw corpus
   ```json
   {"text": "..."}
@@ -509,6 +520,14 @@ model_config:
     type: # linear | dynamic
     factor: # float
 
+# optional overrides to the bnb 4bit quantization configuration
+# https://huggingface.co/docs/transformers/main/main_classes/quantization#transformers.BitsAndBytesConfig
+bnb_config_kwargs:
+  # These are default values
+  llm_int8_has_fp16_weight: false
+  bnb_4bit_quant_type: nf4
+  bnb_4bit_use_double_quant: true
+
 
 # Whether you are training a 4-bit GPTQ quantized model
 gptq: true
@@ -531,6 +550,11 @@ tf32: true # require >=ampere
 bfloat16: true # require >=ampere
 float16: true
 
+# Limit the memory for all available GPUs to this amount (if an integer, expressed in gigabytes); default: unset
+gpu_memory_limit: 20GiB
+# Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge
+lora_on_cpu: true
+
 # A list of one or more datasets to finetune the model with
 datasets:
   # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
@@ -570,6 +594,9 @@ datasets:
       # For `completion` datsets only, uses the provided field instead of `text` column
       field:
 
+# Saves the desired chat template to the tokenizer_config.json for easier inferencing
+# Currently supports chatml and inst (mistral/mixtral)
+chat_template: 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
@@ -612,10 +639,17 @@ eval_sample_packing:
 sample_packing_eff_est:
 total_num_tokens:
 
+# Passed through to transformers when loading the model when launched without accelerate
+# Use `sequential` when training w/ model parallelism to limit memory
+device_map:
+# Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model.
+max_memory:
+
 # If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model
 adapter: lora
 # If you already have a lora model trained that you want to load, put that here.
-# This means after training, if you want to test the model, you should set this to the value of `lora_out_dir`.
+# This means after training, if you want to test the model, you should set this to the value of `output_dir`.
+# Note that if you merge an adapter to the base model, a new subdirectory `merged` will be created under the `output_dir`.
 lora_model_dir:
 
 # LoRA hyperparameters
@@ -642,10 +676,6 @@ lora_modules_to_save:
 #  - embed_tokens
 #  - lm_head
 
-# Once you complete training, the model will be saved to the following directory.
-# If you merge the adapter to the base model, a subdirectory `merged` will be created under this directory.
-# Make sure `lora_model_dir` points to this directory if you want to use the trained model.
-lora_out_dir:
 lora_fan_in_fan_out: false
 
 # ReLoRA configuration
@@ -655,11 +685,13 @@ relora_warmup_steps: # Number of per-restart warmup steps
 relora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings
 
 # wandb configuration if you're using it
+# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.
 wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
 wandb_project: # Your wandb project name
 wandb_entity: # A wandb Team name if using a Team
 wandb_watch:
-wandb_run_id: # Set the name of your wandb run
+wandb_name: # Set the name of your wandb run
+wandb_run_id: # Set the ID of your wandb run
 wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only at the end of training
 
 # Where to save the full-finetuned model to
@@ -682,9 +714,11 @@ warmup_ratio: 0.05  # cannot use with warmup_steps
 learning_rate: 0.00003
 lr_quadratic_warmup:
 logging_steps:
+eval_steps: # Leave empty to eval at each epoch, integers for every N steps. decimal for fraction of total steps
+evals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps
 save_strategy: # Set to `no` to skip checkpoint saves
 save_steps: # Leave empty to save at each epoch
-eval_steps: # Leave empty to eval at each epoch, integers for every N steps. decimal for fraction of total steps
+saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps
 save_total_limit: # Checkpoints saved at a time
 # Maximum number of iterations to train for. It precedes num_epochs which means that
 # if both are set, num_epochs will not be guaranteed.
@@ -694,6 +728,9 @@ max_steps:
 eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
 eval_table_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
 
+loss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)
+loss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)
+
 # Save model as safetensors (require safetensors package)
 save_safetensors:
 
@@ -706,6 +743,9 @@ group_by_length: false
 
 # Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 gradient_checkpointing: false
+# additional kwargs to pass to the trainer for gradient checkpointing
+# gradient_checkpointing_kwargs:
+#   use_reentrant: false
 
 # Stop training after this many evaluation losses have increased in a row
 # https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback
@@ -760,7 +800,7 @@ max_grad_norm:
 # Augmentation techniques
 # NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings
 # currently only supported on Llama and Mistral
-noisy_embedding_alpha:
+neftune_noise_alpha:
 
 # Whether to bettertransformers
 flash_optimum:
@@ -775,11 +815,6 @@ flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
 # Whether to use scaled-dot-product attention
 # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
 sdp_attention:
-# Landmark attention (only llama)
-landmark_attention:
-# xpos RoPE see https://github.com/kaiokendev/cutoff-len-is-context-len/blob/main/util/xpos_rope_llama_monkey_patch.py
-# LLaMA only
-xpos_rope:
 
 # Resume from a specific checkpoint dir
 resume_from_checkpoint:
@@ -902,8 +937,9 @@ accelerate launch -m axolotl.cli.train your_config.yml
 You can optionally pre-tokenize dataset with the following before finetuning.
 This is recommended for large datasets.
 
-- Set `push_dataset_to_hub: hf_user/repo` to push it to Huggingface.
-- Use `--debug` to see preprocessed examples.
+- Set `dataset_prepared_path:` to a local folder for saving and loading pre-tokenized dataset.
+- (Optional): Set `push_dataset_to_hub: hf_user/repo` to push it to Huggingface.
+- (Optional): Use `--debug` to see preprocessed examples.
 
 ```bash
 python -m axolotl.cli.preprocess your_config.yml
@@ -946,19 +982,40 @@ fsdp_config:
 
 ##### Weights & Biases Logging
 
+Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.
+
 - wandb options
 ```yaml
 wandb_mode:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 ```
 
-### Inference
+##### Special Tokens
 
-Pass the appropriate flag to the train command:
+It is important to have special tokens like delimiters, end-of-sequence, beginning-of-sequence in your tokenizer's vocabulary.  This will help you avoid tokenization issues and help your model train better.  You can do this in axolotl like this:
+
+```yml
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"
+tokens: # these are delimiters
+  - "<|im_start|>"
+  - "<|im_end|>"
+```
+
+When you include these tokens in your axolotl config, axolotl adds these tokens to the tokenizer's vocabulary.
+
+### Inference Playground
+
+Axolotl allows you to load your model in an interactive terminal playground for quick experimentation.
+The config file is the same config file used for training.
+
+Pass the appropriate flag to the inference command, depending upon what kind of model was trained:
 
 - Pretrained LORA:
   ```bash
@@ -984,18 +1041,20 @@ Please use `--sample_packing False` if you have it on and receive the error simi
 
 ### Merge LORA to base
 
-Add below flag to train command above
+The following command will merge your LORA adapater with your base model.  You can optionally pass the argument `--lora_model_dir` to specify the directory where your LORA adapter was saved, otherwhise, this will be inferred from `output_dir` in your axolotl config file.  The merged model is saved in the sub-directory `{lora_model_dir}/merged`.
 
 ```bash
-python3 -m axolotl.cli.merge_lora examples/your_config.yml --lora_model_dir="./completed-model" --load_in_8bit=False --load_in_4bit=False
+python3 -m axolotl.cli.merge_lora your_config.yml --lora_model_dir="./completed-model"
 ```
 
-If you run out of CUDA memory, you can try to merge in system RAM with
+You may need to use the `gpu_memory_limit` and/or `lora_on_cpu` config options to avoid running out of memory. If you still run out of CUDA memory, you can try to merge in system RAM with
 
 ```bash
 CUDA_VISIBLE_DEVICES="" python3 -m axolotl.cli.merge_lora ...
 ```
 
+although this will be very slow, and using the config options above are recommended instead.
+
 ## Common Errors 🧰
 
 See also the [FAQ's](./docs/faq.md).
@@ -1008,6 +1067,10 @@ Please reduce any below
   - `gradient_accumulation_steps`
   - `sequence_len`
 
+If it does not help, try running without deepspeed and without accelerate (replace "accelerate launch" with "python") in the command.
+
+Using adamw_bnb_8bit might also save you some memory.
+
 > `failed (exitcode: -9)`
 
 Usually means your system has run out of system memory.
@@ -1030,6 +1093,20 @@ It's safe to ignore it.
 
 See the [NCCL](docs/nccl.md) guide.
 
+
+### Tokenization Mismatch b/w Inference & Training
+
+For many formats, Axolotl constructs prompts by concatenating token ids _after_ tokenizing strings.  The reason for concatenating token ids rather than operating on strings is to maintain precise accounting for attention masks.
+
+If you decode a prompt constructed by axolotl, you might see spaces between tokens (or lack thereof) that you do not expect, especially around delimiters and special tokens.  When you are starting out with a new format, you should always do the following:
+
+1. Materialize some data using `python -m axolotl.cli.preprocess your_config.yml --debug`, and then decode the first few rows with your model's tokenizer.
+2. During inference, right before you pass a tensor of token ids to your model, decode these tokens back into a string.
+3. Make sure the inference string from #2 looks **exactly** like the data you fine tuned on from #1, including spaces and new lines.  If they aren't the same adjust your inference server accordingly.
+4. As an additional troubleshooting step, you can look look at the token ids between 1 and 2 to make sure they are identical.
+
+Having misalignment between your prompts during training and inference can cause models to perform very poorly, so it is worth checking this.  See [this blog post](https://hamel.dev/notes/llm/05_tokenizer_gotchas.html) for a concrete example.
+
 ## Need help? 🙋♂️
 
 Join our [Discord server](https://discord.gg/HhrNrHJPRb) where we can help you

deepspeed/zero1.json

@@ -24,16 +24,6 @@
       "weight_decay": "auto"
     }
   },
-  "scheduler": {
-    "type": "WarmupDecayLR",
-    "params": {
-      "warmup_min_lr": "auto",
-      "warmup_max_lr": "auto",
-      "warmup_num_steps": "auto",
-      "warmup_type": "linear",
-      "total_num_steps": "auto"
-    }
-  },
   "gradient_accumulation_steps": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",

deepspeed/zero2.json

@@ -28,16 +28,6 @@
       "weight_decay": "auto"
     }
   },
-  "scheduler": {
-    "type": "WarmupDecayLR",
-    "params": {
-      "warmup_min_lr": "auto",
-      "warmup_max_lr": "auto",
-      "warmup_num_steps": "auto",
-      "warmup_type": "linear",
-      "total_num_steps": "auto"
-    }
-  },
   "gradient_accumulation_steps": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",

deepspeed/zero3.json

@@ -32,16 +32,6 @@
       "weight_decay": "auto"
     }
   },
-  "scheduler": {
-    "type": "WarmupDecayLR",
-    "params": {
-      "warmup_min_lr": "auto",
-      "warmup_max_lr": "auto",
-      "warmup_num_steps": "auto",
-      "warmup_type": "linear",
-      "total_num_steps": "auto"
-    }
-  },
   "gradient_accumulation_steps": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",

deepspeed/zero3_bf16.json

@@ -0,0 +1,39 @@
+{
+  "zero_optimization": {
+    "stage": 3,
+    "overlap_comm": true,
+    "contiguous_gradients": true,
+    "sub_group_size": 0,
+    "reduce_bucket_size": "auto",
+    "stage3_prefetch_bucket_size": "auto",
+    "stage3_param_persistence_threshold": "auto",
+    "stage3_max_live_parameters": 0,
+    "stage3_max_reuse_distance": 0,
+    "stage3_gather_16bit_weights_on_model_save": true
+  },
+  "bf16": {
+    "enabled": true
+  },
+  "fp16": {
+    "enabled": "auto",
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 32,
+    "loss_scale_window": 1000,
+    "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",
+  "wall_clock_breakdown": false
+}

docker/Dockerfile

@@ -10,7 +10,7 @@ ARG PYTORCH_VERSION="2.0.1"
 ENV PYTORCH_VERSION=$PYTORCH_VERSION
 
 RUN apt-get update && \
-    apt-get install -y vim curl
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev
 
 WORKDIR /workspace
 
@@ -19,13 +19,15 @@ RUN git clone --depth=1 https://github.com/OpenAccess-AI-Collective/axolotl.git
 WORKDIR /workspace/axolotl
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
-RUN sed -i "s/torch==.*/torch==$PYTORCH_VERSION/" requirements.txt
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install -e .[deepspeed,flash-attn,$AXOLOTL_EXTRAS]; \
     else \
         pip install -e .[deepspeed,flash-attn]; \
     fi
 
+# So we can test the Docker image
+RUN pip install pytest
+
 # fix so that git fetch/pull from remote works
 RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \
     git config --get remote.origin.fetch

docker/Dockerfile-runpod

@@ -4,6 +4,7 @@ FROM winglian/axolotl:$BASE_TAG
 ENV HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets"
 ENV HUGGINGFACE_HUB_CACHE="/workspace/data/huggingface-cache/hub"
 ENV TRANSFORMERS_CACHE="/workspace/data/huggingface-cache/hub"
+ENV HF_HOME="/workspace/data/huggingface-cache/hub"
 
 COPY scripts/runpod-entrypoint.sh /root/runpod-entrypoint.sh
 

docs/rlhf.md

@@ -0,0 +1,35 @@
+# RLHF (Beta)
+
+### Overview
+
+Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human
+feedback. Various methods include, but not limited to:
+
+- Proximal Policy Optimization (PPO) (not yet supported in axolotl)
+- Direct Preference Optimization (DPO)
+- Identity Preference Optimization (IPO)
+
+
+### RLHF using Axolotl
+
+[!IMPORTANT]
+This is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.
+
+The various RL training methods are implemented in trl and wrapped via axolotl. Below are various examples with how you can use various preference datasets to train models that use ChatML
+
+#### DPO
+```yaml
+rl: true
+datasets:
+  - path: Intel/orca_dpo_pairs
+    split: train
+    type: intel_apply_chatml
+  - path: argilla/ultrafeedback-binarized-preferences
+    split: train
+    type: argilla_apply_chatml
+```
+
+#### IPO
+```yaml
+rl: ipo
+```

examples/cerebras/btlm-ft.yml

@@ -35,7 +35,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 output_dir: btlm-out
@@ -72,8 +72,8 @@ gptq_groupsize:
 gptq_model_v1:
 
 warmup_steps: 32
-eval_steps:
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 save_total_limit:
 
 debug:

examples/cerebras/qlora.yml

@@ -24,7 +24,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./qlora-out
 batch_size: 4
@@ -49,8 +49,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

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

@@ -29,7 +29,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -54,8 +54,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -31,7 +31,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -56,8 +56,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -29,7 +29,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -54,8 +54,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -31,7 +31,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -56,8 +56,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -29,7 +29,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -54,8 +54,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -31,7 +31,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -56,8 +56,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -26,7 +26,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./falcon-7b
 batch_size: 2
@@ -51,8 +51,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 40
-eval_steps: 5
-save_steps: 43
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -40,7 +40,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./qlora-out
 
@@ -80,8 +80,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 5
-save_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.000001

examples/falcon/config-7b.yml

@@ -26,7 +26,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./falcon-7b
 batch_size: 2
@@ -51,8 +51,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 40
-eval_steps: 5
-save_steps: 43
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/gptj/qlora.yml

@@ -21,7 +21,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./qlora-out
 gradient_accumulation_steps: 2
@@ -46,8 +46,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/jeopardy-bot/config.yml

@@ -19,7 +19,7 @@ lora_fan_in_fan_out: false
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./jeopardy-bot-7b
 gradient_accumulation_steps: 1
@@ -42,8 +42,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 110
-save_steps: 660
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/llama-2/fft_optimized.yml

@@ -29,7 +29,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 1
@@ -58,9 +58,9 @@ flash_attn_fuse_qkv: false
 flash_attn_fuse_mlp: true
 
 warmup_steps: 100
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed: #deepspeed/zero2.json # multi-gpu only
 weight_decay: 0.1

examples/llama-2/gptq-lora.yml

@@ -32,7 +32,7 @@ lora_target_linear:
 lora_fan_in_fan_out:
 wandb_project:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./model-out
 gradient_accumulation_steps: 1
@@ -62,8 +62,8 @@ flash_attention:
 sdp_attention:
 flash_optimum:
 warmup_steps: 100
-eval_steps:
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/llama-2/lora.yml

@@ -29,7 +29,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -54,10 +54,10 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/llama-2/qlora.yml

@@ -31,7 +31,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -56,9 +56,9 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/llama-2/relora.yml

@@ -35,7 +35,7 @@ relora_cpu_offload: false
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -60,8 +60,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps: 50
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/mamba/config.yml

@@ -0,0 +1,61 @@
+base_model: state-spaces/mamba-2.8b
+model_type: MambaLMHeadModel
+tokenizer_type: AutoTokenizer
+tokenizer_config: EleutherAI/gpt-neox-20b
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./out
+
+sequence_len: 2048
+sample_packing: false
+pad_to_sequence_len: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 2
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 5e-5
+
+train_on_inputs: false
+group_by_length: true
+
+bf16: true
+fp16: false
+tf32: true
+
+gradient_checkpointing: false
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_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:
+tokens:
+save_safetensors: False

examples/mistral/config.yml

@@ -17,11 +17,12 @@ output_dir: ./out
 sequence_len: 8192
 sample_packing: true
 pad_to_sequence_len: true
+eval_sample_packing: false
 
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -46,10 +47,10 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/mistral/mixtral.yml

@@ -0,0 +1,91 @@
+base_model: mistralai/Mixtral-8x7B-v0.1
+model_type: AutoModelForCausalLM
+tokenizer_type: LlamaTokenizer
+trust_remote_code: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./qlora-out
+
+## You can optionally freeze the entire model and unfreeze a subset of parameters
+unfrozen_parameters:
+#  - lm_head.*
+#  - model.embed_tokens.*
+#  - model.layers.2[0-9]+.block_sparse_moe.gate.*
+#  - model.layers.2[0-9]+.block_sparse_moe.experts.*
+#  - model.layers.3[0-9]+.block_sparse_moe.gate.*
+#  - model.layers.3[0-9]+.block_sparse_moe.experts.*
+
+model_config:
+  output_router_logits: true
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+#lora_target_modules:
+#  - gate
+#  - q_proj
+#  - k_proj
+#  - v_proj
+#  - o_proj
+#  - w1
+#  - w2
+#  - w3
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_steps: 10
+evals_per_epoch: 4
+eval_table_size:
+eval_table_max_new_tokens: 128
+saves_per_epoch: 1
+debug:
+deepspeed: deepspeed/zero2.json
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

examples/mistral/qlora.yml

@@ -11,7 +11,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path: last_run_prepared
-val_set_size: 0.05
+val_set_size: 0.1
 output_dir: ./qlora-out
 
 adapter: qlora
@@ -38,7 +38,7 @@ lora_target_modules:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -62,11 +62,14 @@ logging_steps: 1
 xformers_attention:
 flash_attention: true
 
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/mpt-7b/config.yml

@@ -21,7 +21,7 @@ lora_fan_in_fan_out: false
 wandb_project: mpt-alpaca-7b
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./mpt-alpaca-7b
 gradient_accumulation_steps: 1
@@ -44,8 +44,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 110
-save_steps: 660
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0001

examples/openllama-3b/config.yml

@@ -23,7 +23,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./openllama-out
 gradient_accumulation_steps: 1
@@ -49,8 +49,8 @@ flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/openllama-3b/lora.yml

@@ -29,7 +29,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./lora-out
 gradient_accumulation_steps: 1
@@ -54,8 +54,8 @@ flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/openllama-3b/qlora.yml

@@ -23,7 +23,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./qlora-out
 gradient_accumulation_steps: 1
@@ -48,8 +48,8 @@ flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/phi/phi-ft.yml

@@ -31,7 +31,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 1
@@ -59,8 +59,8 @@ xformers_attention:
 flash_attention:
 
 warmup_steps: 100
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/phi/phi-qlora.yml

@@ -31,7 +31,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 1
@@ -59,8 +59,8 @@ xformers_attention:
 flash_attention:
 
 warmup_steps: 100
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/pythia-12b/config.yml

@@ -24,7 +24,7 @@ lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./pythia-12b
 gradient_accumulation_steps: 1

examples/pythia/lora.yml

@@ -18,7 +18,7 @@ lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./lora-alpaca-pythia
 gradient_accumulation_steps: 1
@@ -33,5 +33,5 @@ early_stopping_patience:
 resume_from_checkpoint:
 local_rank:
 weight_decay: 0.1
-eval_steps: 0.05
+evals_per_epoch: 4
 logging_steps: 1

examples/qwen/lora.yml

@@ -31,7 +31,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -56,10 +56,10 @@ xformers_attention:
 flash_attention:
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/qwen/qlora.yml

@@ -31,7 +31,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -56,10 +56,10 @@ xformers_attention:
 flash_attention:
 
 warmup_steps: 10
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/redpajama/config-3b.yml

@@ -22,7 +22,7 @@ lora_fan_in_fan_out: false
 wandb_project: redpajama-alpaca-3b
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./redpajama-alpaca-3b
 batch_size: 4
@@ -45,8 +45,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 110
-save_steps: 660
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0001

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

@@ -21,7 +21,7 @@ lora_fan_in_fan_out:
 wandb_project: lora-replit
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./lora-replit
 batch_size: 8
@@ -45,8 +45,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 50
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0

examples/tiny-llama/README.md

@@ -0,0 +1,17 @@
+# Overview
+
+This is a simple example of how to finetune TinyLlama1.1B using either lora or qlora:
+
+LoRa:
+
+```
+accelerate launch -m axolotl.cli.train examples/tiny-llama/lora.yml
+```
+
+qLoRa:
+
+```
+accelerate launch -m axolotl.cli.train examples/tiny-llama/qlora.yml
+```
+
+Both take about 10 minutes to complete on a 4090.

examples/tiny-llama/lora.yml

@@ -1,5 +1,4 @@
-base_model: PY007/TinyLlama-1.1B-intermediate-step-715k-1.5T
-
+base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 is_llama_derived_model: true
@@ -17,6 +16,7 @@ output_dir: ./lora-out
 
 sequence_len: 4096
 sample_packing: true
+pad_to_sequence_len: true
 
 adapter: lora
 lora_model_dir:
@@ -29,7 +29,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -54,15 +54,11 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-eval_table_size:
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0
 fsdp:
 fsdp_config:
 special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

examples/tiny-llama/pretrain.yml

@@ -0,0 +1,58 @@
+base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
+
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+max_steps: 200
+pretraining_dataset:
+  path: c4
+  name: en
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./model-out
+
+sequence_len: 2048
+sample_packing: true
+
+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: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch:
+eval_table_size:
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

examples/tiny-llama/qlora.yml

@@ -0,0 +1,66 @@
+base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

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

@@ -38,7 +38,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./qlora-out
 
@@ -78,8 +78,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 50
-save_steps: 50
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/yi-34B-chat/README.md

@@ -0,0 +1,5 @@
+# Overview
+
+This is an example of a Yi-34B-Chat configuration. It demonstrates that it is possible to finetune a 34B model on a GPU with 24GB of VRAM.
+
+Tested on an RTX 4090 with `python -m axolotl.cli.train examples/mistral/qlora.yml`, a single epoch of finetuning on the alpaca dataset using qlora runs in 47 mins, using 97% of available memory.

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

@@ -0,0 +1,76 @@
+base_model: 01-ai/Yi-34B-Chat
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_mistral_derived_model: false
+is_llama_derived_model: true
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+sequence_len: 1024
+bf16: true
+fp16: false
+tf32: false
+flash_attention: true
+special_tokens:
+  bos_token: "<|startoftext|>"
+  eos_token: "<|endoftext|>"
+  unk_token: "<unk>"
+
+# Data
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+warmup_steps: 10
+
+# Iterations
+num_epochs: 1
+
+# Evaluation
+val_set_size: 0.1
+evals_per_epoch: 5
+eval_table_size:
+eval_table_max_new_tokens: 128
+eval_sample_packing: false
+eval_batch_size: 1
+
+# LoRA
+output_dir: ./qlora-out
+adapter: qlora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+lora_target_modules:
+
+# Sampling
+sample_packing: false
+pad_to_sequence_len: false
+
+# Batching
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+gradient_checkpointing: true
+
+# wandb
+wandb_project:
+
+# Optimizer
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+# Misc
+train_on_inputs: false
+group_by_length: false
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+debug:
+deepspeed:
+weight_decay: 0
+fsdp:
+fsdp_config:

requirements.txt

@@ -2,7 +2,7 @@
 auto-gptq==0.5.1
 packaging
 peft==0.6.0
-transformers==4.35.2
+transformers @ git+https://github.com/huggingface/transformers.git@3cefac1d974db5e2825a0cb2b842883a628be7a0
 tokenizers==0.15.0
 bitsandbytes>=0.41.1
 accelerate==0.24.1
@@ -29,7 +29,7 @@ scipy
 scikit-learn==1.2.2
 pynvml
 art
-fschat==0.2.29
+fschat==0.2.34
 gradio==3.50.2
 tensorboard
 
@@ -37,3 +37,5 @@ tensorboard
 s3fs
 gcsfs
 # adlfs
+
+trl @ git+https://github.com/huggingface/trl.git@main

setup.py

@@ -1,5 +1,7 @@
 """setup.py for axolotl"""
 
+from importlib.metadata import PackageNotFoundError, version
+
 from setuptools import find_packages, setup
 
 
@@ -22,12 +24,13 @@ def parse_requirements():
                 # Handle standard packages
                 _install_requires.append(line)
 
-    # TODO(wing) remove once xformers release supports torch 2.1.0
-    if "torch==2.1.0" in _install_requires:
-        _install_requires.pop(_install_requires.index("xformers>=0.0.22"))
-        _install_requires.append(
-            "xformers @ git+https://github.com/facebookresearch/xformers.git@main"
-        )
+    try:
+        torch_version = version("torch")
+        if torch_version.startswith("2.1.1"):
+            _install_requires.pop(_install_requires.index("xformers==0.0.22"))
+            _install_requires.append("xformers==0.0.23")
+    except PackageNotFoundError:
+        pass
 
     return _install_requires, _dependency_links
 
@@ -46,10 +49,13 @@ setup(
     dependency_links=dependency_links,
     extras_require={
         "flash-attn": [
-            "flash-attn>=2.3.0",
+            "flash-attn==2.3.3",
         ],
         "deepspeed": [
             "deepspeed",
         ],
+        "mamba-ssm": [
+            "mamba-ssm==1.0.1",
+        ],
     },
 )

src/axolotl/cli/__init__.py

@@ -2,6 +2,7 @@
 
 import importlib
 import logging
+import math
 import os
 import random
 import sys
@@ -16,6 +17,7 @@ import yaml
 # add src to the pythonpath so we don't need to pip install this
 from accelerate.commands.config import config_args
 from art import text2art
+from datasets import concatenate_datasets, load_dataset
 from huggingface_hub import HfApi
 from huggingface_hub.utils import LocalTokenNotFoundError
 from transformers import GenerationConfig, TextIteratorStreamer, TextStreamer
@@ -71,7 +73,7 @@ def do_merge_lora(
     safe_serialization = cfg.save_safetensors is True
 
     LOG.info("running merge of LoRA with base model")
-    model = model.merge_and_unload()
+    model = model.merge_and_unload(progressbar=True)
     model.to(dtype=cfg.torch_dtype)
 
     if cfg.local_rank == 0:
@@ -79,6 +81,7 @@ def do_merge_lora(
         model.save_pretrained(
             str(Path(cfg.output_dir) / "merged"),
             safe_serialization=safe_serialization,
+            progressbar=True,
         )
         tokenizer.save_pretrained(str(Path(cfg.output_dir) / "merged"))
 
@@ -103,15 +106,7 @@ def do_inference(
             importlib.import_module("axolotl.prompters"), prompter
         )
 
-    if cfg.landmark_attention:
-        from axolotl.monkeypatch.llama_landmark_attn import set_model_mem_id
-
-        set_model_mem_id(model, tokenizer)
-        model.set_mem_cache_args(
-            max_seq_len=255, mem_freq=50, top_k=5, max_cache_size=None
-        )
-
-    model = model.to(cfg.device)
+    model = model.to(cfg.device, dtype=cfg.torch_dtype)
 
     while True:
         print("=" * 80)
@@ -176,15 +171,7 @@ def do_inference_gradio(
             importlib.import_module("axolotl.prompters"), prompter
         )
 
-    if cfg.landmark_attention:
-        from axolotl.monkeypatch.llama_landmark_attn import set_model_mem_id
-
-        set_model_mem_id(model, tokenizer)
-        model.set_mem_cache_args(
-            max_seq_len=255, mem_freq=50, top_k=5, max_cache_size=None
-        )
-
-    model = model.to(cfg.device)
+    model = model.to(cfg.device, dtype=cfg.torch_dtype)
 
     def generate(instruction):
         if not instruction:
@@ -341,6 +328,94 @@ def load_datasets(
     )
 
 
+def load_rl_datasets(
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,  # pylint: disable=unused-argument
+) -> TrainDatasetMeta:
+    train_datasets: List[Any] = []
+    for i, ds_cfg in enumerate(cfg.datasets):
+        train_datasets.insert(i, load_dataset(ds_cfg["path"], split=ds_cfg["split"]))
+    # eval_dataset = load_dataset(
+    #     cfg.test_datasets[0]["path"], split=cfg.test_datasets[0]["split"]
+    # )
+    eval_dataset = None
+
+    def argilla_apply_chatml(sample):  # pylint: disable=possibly-unused-variable
+        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['instruction']}<|im_end|>\n<|im_start|>assistant\n"
+            )
+        else:
+            sample[
+                "prompt"
+            ] = f"<|im_start|>user\n{sample['instruction']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["chosen"] = f"{sample['chosen_response']}<|im_end|>"
+        sample["rejected"] = f"{sample['rejected_response']}<|im_end|>"
+        return sample
+
+    def intel_apply_chatml(sample):  # pylint: disable=possibly-unused-variable
+        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['question']}<|im_end|>\n<|im_start|>assistant\n"
+            )
+        else:
+            sample[
+                "prompt"
+            ] = f"<|im_start|>user\n{sample['question']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["chosen"] = f"{sample['chosen']}<|im_end|>"
+        sample["rejected"] = f"{sample['rejected']}<|im_end|>"
+        return sample
+
+    def apply_chatml(sample):  # pylint: disable=possibly-unused-variable
+        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['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            )
+        else:
+            sample[
+                "prompt"
+            ] = f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["chosen"] = f"{sample['chosen']}<|im_end|>"
+        sample["rejected"] = f"{sample['rejected']}<|im_end|>"
+        return sample
+
+    def ultra_apply_chatml(sample):  # pylint: disable=possibly-unused-variable
+        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['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            )
+        else:
+            sample[
+                "prompt"
+            ] = f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["chosen"] = f"{sample['chosen'][1]['content']}<|im_end|>"
+        sample["rejected"] = f"{sample['rejected'][1]['content']}<|im_end|>"
+        return sample
+
+    for i, data_set in enumerate(train_datasets):
+        _type = cfg.datasets[i]["type"]
+        ds_type_fn = locals()[_type]
+        train_datasets[i] = data_set.map(ds_type_fn)
+    train_dataset = concatenate_datasets(train_datasets)
+
+    # eval_dataset = eval_dataset.map(intel_apply_chatml)
+
+    total_num_steps = int(
+        math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
+    )
+
+    return TrainDatasetMeta(
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        total_num_steps=total_num_steps,
+    )
+
+
 def check_accelerate_default_config():
     if Path(config_args.default_yaml_config_file).exists():
         LOG.warning(

src/axolotl/cli/merge_lora.py

@@ -18,7 +18,22 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
         return_remaining_strings=True
     )
     parsed_cli_args.merge_lora = True
-    parsed_cfg = load_cfg(config, merge_lora=True, **kwargs)
+
+    parsed_cfg = load_cfg(
+        config,
+        merge_lora=True,
+        load_in_8bit=False,
+        load_in_4bit=False,
+        flash_attention=False,
+        **kwargs,
+    )
+
+    if not parsed_cfg.lora_model_dir and parsed_cfg.output_dir:
+        parsed_cfg.lora_model_dir = parsed_cfg.output_dir
+    if not Path(parsed_cfg.lora_model_dir).exists():
+        raise ValueError(
+            f"Target directory for merge: `{parsed_cfg.lora_model_dir}` does not exist."
+        )
 
     do_merge_lora(cfg=parsed_cfg, cli_args=parsed_cli_args)
 

src/axolotl/cli/train.py

@@ -12,6 +12,7 @@ from axolotl.cli import (
     check_user_token,
     load_cfg,
     load_datasets,
+    load_rl_datasets,
     print_axolotl_text_art,
 )
 from axolotl.common.cli import TrainerCliArgs
@@ -22,15 +23,18 @@ LOG = logging.getLogger("axolotl.cli.train")
 
 def do_cli(config: Path = Path("examples/"), **kwargs):
     # pylint: disable=duplicate-code
-    print_axolotl_text_art()
     parsed_cfg = load_cfg(config, **kwargs)
+    print_axolotl_text_art()
     check_accelerate_default_config()
     check_user_token()
     parser = transformers.HfArgumentParser((TrainerCliArgs))
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True
     )
-    dataset_meta = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    if parsed_cfg.rl:
+        dataset_meta = load_rl_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    else:
+        dataset_meta = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
     train(cfg=parsed_cfg, cli_args=parsed_cli_args, dataset_meta=dataset_meta)
 
 

src/axolotl/core/trainer_builder.py

@@ -9,7 +9,7 @@ import math
 import sys
 from abc import abstractmethod
 from dataclasses import dataclass, field
-from functools import partial
+from functools import wraps
 from pathlib import Path
 from typing import Optional
 
@@ -20,17 +20,22 @@ from torch.optim.lr_scheduler import OneCycleLR
 from torch.utils.data import BatchSampler, DataLoader, RandomSampler, SequentialSampler
 from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
 from transformers.trainer_utils import seed_worker
+from trl import DPOTrainer
 
 from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
 from axolotl.utils.callbacks import (
     EvalFirstStepCallback,
     GPUStatsCallback,
+    LossWatchDogCallback,
     SaveAxolotlConfigtoWandBCallback,
     SaveBetterTransformerModelCallback,
     bench_eval_callback_factory,
     log_prediction_callback_factory,
 )
-from axolotl.utils.collators import BatchSamplerDataCollatorForSeq2Seq
+from axolotl.utils.collators import (
+    BatchSamplerDataCollatorForSeq2Seq,
+    MambaDataCollator,
+)
 from axolotl.utils.samplers import MultipackBatchSampler
 from axolotl.utils.schedulers import get_cosine_schedule_with_quadratic_warmup
 
@@ -48,10 +53,19 @@ class AxolotlTrainingArguments(TrainingArguments):
     Extend the base TrainingArguments for axolotl helpers
     """
 
+    model_type: Optional[str] = field(
+        default=None, metadata={"help": "HF model configuration model_type."}
+    )
     lr_quadratic_warmup: bool = field(
         default=False,
         metadata={"help": "Use quadratic warmup for cosine scheduling."},
     )
+    pretraining: bool = field(
+        default=False,
+        metadata={
+            "help": "Indicates to trainer whether we are doing continued pretraining."
+        },
+    )
     sample_packing: bool = field(
         default=False,
         metadata={"help": "Use sample packing for efficient training."},
@@ -113,6 +127,7 @@ class AxolotlTrainer(Trainer):
     """
 
     args = None  # type: AxolotlTrainingArguments
+    tag_names = ["axolotl"]
 
     def __init__(self, *args, num_epochs=1, bench_data_collator=None, **kwargs):
         self.num_epochs = num_epochs
@@ -148,7 +163,7 @@ class AxolotlTrainer(Trainer):
         return self.lr_scheduler
 
     def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
-        if self.args.sample_packing:
+        if self.args.sample_packing and not self.args.pretraining:
             return MultipackBatchSampler(
                 RandomSampler(self.train_dataset),
                 self.args.train_batch_size,
@@ -184,7 +199,7 @@ class AxolotlTrainer(Trainer):
         return super()._get_eval_sampler(eval_dataset)
 
     def get_train_dataloader(self) -> DataLoader:
-        if self.args.sample_packing:
+        if self.args.sample_packing and not self.args.pretraining:
             train_dataset = self.train_dataset
             train_dataset = train_dataset.remove_columns(["length"])
             data_collator = self.data_collator
@@ -283,12 +298,69 @@ 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
+        )
+
+        return super().push_to_hub(*args, **kwargs)
+
+
+class AxolotlMambaTrainer(AxolotlTrainer):
+    """
+    Mamba specific trainer to handle loss calculation
+    """
+
+    tag_names = ["axolotl", "mamba"]
+
+    def compute_loss(
+        self,
+        model,
+        inputs,
+        return_outputs=False,  # pylint: disable=unused-argument
+    ):
+        input_ids = inputs.pop("input_ids")
+        lm_logits = model(input_ids).logits
+
+        labels = input_ids.to(lm_logits.device)
+        shift_logits = lm_logits[:, :-1, :].contiguous()
+        labels = labels[:, 1:].contiguous()
+
+        loss_fct = torch.nn.CrossEntropyLoss()
+        lm_loss = loss_fct(
+            shift_logits.view(-1, shift_logits.size(-1)), labels.view(-1)
+        )
+
+        return lm_loss
+
 
 class OneCycleLRSchedulerTrainer(AxolotlTrainer):
     """
     Trainer subclass that uses the OneCycleLR scheduler
     """
 
+    tag_names = ["axolotl", "onecycle"]
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.lr_scheduler = None
@@ -318,6 +390,8 @@ class ReLoRATrainer(AxolotlTrainer):
     Trainer subclass that uses the OneCycleLR scheduler
     """
 
+    tag_names = ["axolotl", "relora"]
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.lr_scheduler = None
@@ -353,12 +427,21 @@ class TrainerBuilderBase(abc.ABC):
 
     _train_dataset = None
     _eval_dataset = None
+    _model_ref = None
 
     def __init__(self, cfg, model, tokenizer):
         self.cfg = cfg
         self.model = model
         self.tokenizer = tokenizer
 
+    @property
+    def model_ref(self):
+        return self._model_ref
+
+    @model_ref.setter
+    def model_ref(self, model):
+        self._model_ref = model
+
     @property
     def train_dataset(self):
         return self._train_dataset
@@ -430,6 +513,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
             )
 
+        if self.cfg.loss_watchdog_threshold is not None:
+            callbacks.append(LossWatchDogCallback(self.cfg))
+
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
@@ -458,6 +544,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             return OneCycleLRSchedulerTrainer
         if self.cfg.relora_steps:
             return ReLoRATrainer
+        if self.cfg.model_config_type == "mamba":
+            return AxolotlMambaTrainer
         return AxolotlTrainer
 
     def build(self, total_num_steps):
@@ -494,6 +582,14 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs[
                 "gradient_checkpointing"
             ] = self.cfg.gradient_checkpointing
+            if self.cfg.gradient_checkpointing_kwargs:
+                training_arguments_kwargs[
+                    "gradient_checkpointing_kwargs"
+                ] = self.cfg.gradient_checkpointing_kwargs
+            else:
+                training_arguments_kwargs["gradient_checkpointing_kwargs"] = {
+                    "use_reentrant": False
+                }
         if self.cfg.fsdp:
             training_arguments_kwargs["fsdp"] = self.cfg.fsdp
             if self.cfg.fsdp_config:
@@ -521,11 +617,12 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["hub_model_id"] = self.cfg.hub_model_id
             training_arguments_kwargs["push_to_hub"] = True
             training_arguments_kwargs["hub_private_repo"] = True
+            training_arguments_kwargs["hub_always_push"] = True
 
             if self.cfg.hub_strategy:
                 training_arguments_kwargs["hub_strategy"] = self.cfg.hub_strategy
 
-        if self.cfg.save_safetensors:
+        if self.cfg.save_safetensors is not None:
             training_arguments_kwargs["save_safetensors"] = self.cfg.save_safetensors
 
         if self.cfg.sample_packing_eff_est:
@@ -643,7 +740,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["group_by_length"] = self.cfg.group_by_length
         training_arguments_kwargs["report_to"] = "wandb" if self.cfg.use_wandb else None
         training_arguments_kwargs["run_name"] = (
-            self.cfg.wandb_run_id if self.cfg.use_wandb else None
+            self.cfg.wandb_name if self.cfg.use_wandb else None
         )
         training_arguments_kwargs["optim"] = (
             self.cfg.optimizer if self.cfg.optimizer else "adamw_hf"
@@ -654,6 +751,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             and self.cfg.lr_scheduler not in ("one_cycle", "log_sweep")
             else "cosine"
         )
+        training_arguments_kwargs["lr_scheduler_kwargs"] = (
+            self.cfg.lr_scheduler_kwargs if self.cfg.lr_scheduler_kwargs else {}
+        )
         training_arguments_kwargs["weight_decay"] = (
             self.cfg.weight_decay if self.cfg.weight_decay is not None else 0.0
         )
@@ -673,6 +773,14 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs = self.hook_pre_create_training_args(
             training_arguments_kwargs
         )
+        training_arguments_kwargs["model_type"] = self.cfg.model_config_type
+        training_arguments_kwargs["pretraining"] = bool(self.cfg.pretraining_dataset)
+
+        if self.cfg.neftune_noise_alpha is not None:
+            training_arguments_kwargs[
+                "neftune_noise_alpha"
+            ] = self.cfg.neftune_noise_alpha
+
         training_args = (
             AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
                 **training_arguments_kwargs,
@@ -698,26 +806,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
             data_collator_kwargs["pad_to_multiple_of"] = 64
 
-        if self.cfg.is_llama_derived_model and self.cfg.landmark_attention:
-            from axolotl.monkeypatch.llama_landmark_attn import (
-                add_mem_tokens,
-                get_mem_id,
-                set_model_mem_id,
-            )
-
-            set_model_mem_id(self.model, self.tokenizer)
-
-            LOG.info("Adding landmark attention tokens to dataset")
-
-            for dataset in [self.train_dataset, self.eval_dataset]:
-                dataset = dataset.map(
-                    partial(
-                        add_mem_tokens, mem_freq=50, mem_id=get_mem_id(self.tokenizer)
-                    ),
-                    batched=False,
-                    num_proc=32,
-                )
-
         trainer_cls = self._get_trainer_cls()
         trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
             trainer_kwargs, trainer_cls
@@ -727,11 +815,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             train_dataset=self.train_dataset,
             eval_dataset=self.eval_dataset,
             args=training_args,
-            data_collator=BatchSamplerDataCollatorForSeq2Seq(
-                self.tokenizer,
-                return_tensors="pt",
-                **data_collator_kwargs,
-            ),
+            data_collator=self.build_collator(training_args, **data_collator_kwargs),
             bench_data_collator=transformers.DataCollatorForSeq2Seq(
                 self.tokenizer,
                 return_tensors="pt",
@@ -751,3 +835,109 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             ] = self.cfg.micro_batch_size
 
         return trainer
+
+    def build_collator(self, training_args: AxolotlTrainingArguments, **kwargs):
+        if training_args.pretraining:
+            return None
+
+        if self.cfg.model_config_type == "mamba":
+            return MambaDataCollator(tokenizer=self.tokenizer)
+
+        return BatchSamplerDataCollatorForSeq2Seq(
+            self.tokenizer,
+            return_tensors="pt",
+            **kwargs,
+        )
+
+
+class HFDPOTrainerBuilder(TrainerBuilderBase):
+    """
+    Trainer factory class for DPO Trainer
+    """
+
+    def get_callbacks(self):
+        callbacks = []
+        return callbacks
+
+    def get_post_trainer_create_callbacks(self, trainer):
+        callbacks = []
+        return callbacks
+
+    def build_training_arguments(self, total_num_steps):
+        training_args_kwargs = {}
+        for arg in [
+            "adam_beta1",
+            "adam_beta2",
+            "adam_epsilon",
+            "dataloader_num_workers",
+            "dataloader_pin_memory",
+        ]:
+            if hasattr(self.cfg, arg) and getattr(self.cfg, arg) is not None:
+                training_args_kwargs[arg] = getattr(self.cfg, arg)
+        training_args = TrainingArguments(
+            per_device_train_batch_size=self.cfg.micro_batch_size,
+            max_steps=total_num_steps,
+            remove_unused_columns=False,
+            gradient_accumulation_steps=self.cfg.gradient_accumulation_steps,
+            learning_rate=self.cfg.learning_rate,
+            evaluation_strategy="no",
+            # eval_steps=self.cfg.eval_steps,
+            save_strategy="steps",
+            save_steps=self.cfg.save_steps,
+            output_dir=self.cfg.output_dir,
+            warmup_steps=self.cfg.warmup_steps,
+            bf16=True,
+            gradient_checkpointing=self.cfg.gradient_checkpointing,
+            gradient_checkpointing_kwargs={"use_reentrant": False},
+            logging_first_step=True,
+            logging_steps=1,
+            optim=self.cfg.optimizer,
+            save_total_limit=self.cfg.save_total_limit or 5,
+            **training_args_kwargs,
+        )
+
+        return training_args
+
+    def build(self, total_num_steps):
+        training_args = self.build_training_arguments(total_num_steps)
+        dpo_trainer_kwargs = {}
+        if self.cfg.rl == "ipo":
+            dpo_trainer_kwargs["loss_type"] = "ipo"
+            if self.cfg.dpo_label_smoothing:
+                dpo_trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
+
+        dpo_trainer = DPOTrainer(
+            self.model,
+            self.model_ref,
+            args=training_args,
+            beta=self.cfg.dpo_beta or 0.1,
+            train_dataset=self.train_dataset,
+            # eval_dataset=self.eval_dataset,
+            eval_dataset=None,
+            tokenizer=self.tokenizer,
+            max_length=self.cfg.sequence_len,
+            max_target_length=None,
+            max_prompt_length=self.cfg.sequence_len,
+            generate_during_eval=True,
+            **dpo_trainer_kwargs,
+        )
+
+        return dpo_trainer
+
+
+class HFPPOTrainerBuilder(TrainerBuilderBase):
+    """
+    HF Factory class for PPO Trainer
+    """
+
+    def get_callbacks(self):
+        callbacks = []
+        return callbacks
+
+    def get_post_trainer_create_callbacks(self, trainer):
+        callbacks = []
+        return callbacks
+
+    def build(self, total_num_steps):
+        # build PPOConfig
+        pass

src/axolotl/core/trainers/trl.py

@@ -0,0 +1,66 @@
+"""
+module for TRL PPO training
+"""
+import torch
+from tqdm import tqdm
+from trl import PPOTrainer
+
+
+class TRLPPOTrainer(PPOTrainer):
+    """
+    wrapper for ppo trainer to handle customizations
+    """
+
+    def train(
+        self,
+        reward_pipe,
+        resume_from_checkpoint=None,  # pylint: disable=unused-argument
+    ):
+        generation_kwargs = {
+            "min_length": -1,
+            "top_k": 0.0,
+            "top_p": 1.0,
+            "do_sample": True,
+            "pad_token_id": self.tokenizer.eos_token_id,
+            "max_new_tokens": 32,
+        }
+        sent_kwargs = {
+            "return_all_scores": True,
+            "function_to_apply": "none",
+            "batch_size": 16,
+        }
+
+        for epoch, batch in tqdm(  # pylint: disable=unused-variable
+            enumerate(self.dataloader)
+        ):
+            query_tensors = batch["input_ids"]
+
+            # generate model response
+            response_tensors, ref_response_tensors = self.generate(
+                query_tensors,
+                return_prompt=False,
+                generate_ref_response=True,
+                **generation_kwargs
+            )
+            batch["response"] = self.tokenizer.batch_decode(response_tensors)
+            batch["ref_response"] = self.tokenizer.batch_decode(ref_response_tensors)
+
+            # Compute sentiment score
+            texts = [q + r for q, r in zip(batch["query"], batch["response"])]
+            pipe_outputs = reward_pipe(texts, **sent_kwargs)
+            rewards = [torch.tensor(output[1]["score"]) for output in pipe_outputs]
+            ref_texts = [q + r for q, r in zip(batch["query"], batch["ref_response"])]
+            ref_pipe_outputs = reward_pipe(ref_texts, **sent_kwargs)
+            ref_rewards = [
+                torch.tensor(output[1]["score"]) for output in ref_pipe_outputs
+            ]
+            batch["ref_rewards"] = ref_rewards
+
+            # Run PPO step
+            stats = self.step(query_tensors, response_tensors, rewards)
+            self.log_stats(
+                stats,
+                batch,
+                rewards,
+                columns_to_log=["query", "response", "ref_response", "ref_rewards"],
+            )

src/axolotl/models/mamba/__init__.py

@@ -0,0 +1,12 @@
+"""
+Modeling module for Mamba models
+"""
+
+
+def fix_mamba_attn_for_loss():
+    from mamba_ssm.models import mixer_seq_simple
+
+    from .modeling_mamba import MambaLMHeadModel as MambaLMHeadModelFixed
+
+    mixer_seq_simple.MambaLMHeadModel = MambaLMHeadModelFixed
+    return mixer_seq_simple.MambaLMHeadModel  # pylint: disable=invalid-name

src/axolotl/models/mamba/configuration_mamba.py

@@ -0,0 +1,42 @@
+"""
+HF Transformers MambaConfig
+"""
+from transformers import PretrainedConfig
+
+
+class MambaConfig(PretrainedConfig):
+    """
+    modeling configuration for state space model/mamba
+    """
+
+    model_type = "mamba"
+
+    def __init__(
+        self,
+        vocab_size=50280,
+        d_model=2560,
+        n_layer=64,
+        rms_norm=True,
+        residual_in_fp32=True,
+        fused_add_norm=True,
+        pad_vocab_size_multiple=8,
+        pad_token_id=50277,
+        bos_token_id=0,
+        eos_token_id=0,
+        tie_word_embeddings=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.n_layer = n_layer
+        self.rms_norm = rms_norm
+        self.residual_in_fp32 = residual_in_fp32
+        self.fused_add_norm = fused_add_norm
+        self.pad_vocab_size_multiple = pad_vocab_size_multiple
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

src/axolotl/models/mamba/modeling_mamba.py

@@ -0,0 +1,128 @@
+# pylint: skip-file
+import os
+from collections import namedtuple
+from functools import partial
+from typing import Optional, Union
+
+import torch
+from mamba_ssm.models.mixer_seq_simple import MixerModel, _init_weights
+from mamba_ssm.utils.generation import GenerationMixin
+from mamba_ssm.utils.hf import load_config_hf, load_state_dict_hf
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from axolotl.models.mamba.configuration_mamba import MambaConfig
+
+
+class MambaLMHeadModel(nn.Module, GenerationMixin):
+    def __init__(
+        self,
+        d_model: int,
+        n_layer: int,
+        vocab_size: int,
+        initializer_cfg=None,
+        pad_vocab_size_multiple: int = 1,
+        device=None,
+        dtype=None,
+        **backbone_kwargs,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        if vocab_size % pad_vocab_size_multiple != 0:
+            vocab_size += pad_vocab_size_multiple - (
+                vocab_size % pad_vocab_size_multiple
+            )
+        self.config = MambaConfig(
+            vocab_size=vocab_size,
+            d_model=d_model,
+            n_layer=n_layer,
+            pad_vocab_size_multiple=pad_vocab_size_multiple,
+        )
+        self.backbone = MixerModel(
+            d_model=d_model,
+            n_layer=n_layer,
+            vocab_size=vocab_size,
+            initializer_cfg=initializer_cfg,
+            **backbone_kwargs,
+            **factory_kwargs,
+        )
+        self.lm_head = nn.Linear(d_model, vocab_size, bias=False, **factory_kwargs)
+
+        # Initialize weights and apply final processing
+        self.apply(
+            partial(
+                _init_weights,
+                n_layer=n_layer,
+                **(initializer_cfg if initializer_cfg is not None else {}),
+            )
+        )
+        self.tie_weights()
+
+    def tie_weights(self):
+        self.lm_head.weight = self.backbone.embedding.weight
+
+    def allocate_inference_cache(self, batch_size, max_seqlen, dtype=None, **kwargs):
+        return self.backbone.allocate_inference_cache(
+            batch_size, max_seqlen, dtype=dtype, **kwargs
+        )
+
+    def forward(
+        self,
+        input_ids,
+        position_ids=None,
+        inference_params=None,
+        num_last_tokens=0,
+        labels=None,
+        **kwargs,
+    ):
+        """
+        "position_ids" is just to be compatible with Transformer generation. We don't use it.
+        num_last_tokens: if > 0, only return the logits for the last n tokens
+        """
+        hidden_states = self.backbone(input_ids, inference_params=inference_params)
+        if num_last_tokens > 0:
+            hidden_states = hidden_states[:, -num_last_tokens:]
+        lm_logits = self.lm_head(hidden_states)
+
+        CausalLMOutput = namedtuple("CausalLMOutput", ["logits"])
+        return CausalLMOutput(logits=lm_logits)
+
+        loss = None
+        if labels is not None:
+            logits = lm_logits
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+            CausalLMOutput = namedtuple("CausalLMOutput", ["logits", "loss"])
+            print(loss)
+            return CausalLMOutput(logits=lm_logits, loss=loss)
+
+        else:
+            CausalLMOutput = namedtuple("CausalLMOutput", ["logits"])
+            return CausalLMOutput(logits=lm_logits)
+
+    def save_pretrained(
+        self,
+        save_directory: Union[str, os.PathLike],
+        state_dict: Optional[dict] = None,
+        safe_serialization: Optional[bool] = None,  # pylint: disable=unused-argument
+    ):
+        if state_dict is None:
+            state_dict = self.state_dict()
+        torch.save(state_dict, os.path.join(save_directory, "pytorch_model.bin"))
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name, device=None, dtype=None, **kwargs):
+        config = load_config_hf(pretrained_model_name)
+        model = cls(**config, device=device, dtype=dtype, **kwargs)
+        model.load_state_dict(
+            load_state_dict_hf(pretrained_model_name, device={"": device}, dtype=dtype)
+        )
+        return model

src/axolotl/monkeypatch/fastchat_conversation_turns.py

@@ -82,15 +82,44 @@ def get_turns(  # pylint: disable=too-many-return-statements
             else:
                 yield role + ":", ""
         return
-    if self.sep_style == SeparatorStyle.LLAMA2:
-        seps = [self.sep, self.sep2]
+    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
-        else:
-            yield "", "[INST] "
-        for i, (role, message) in enumerate(self.messages[1:]):
+        for i, (role, message) in enumerate(self.messages):
             if message:
-                yield role + " ", message + seps[i % 2]
+                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 seperated 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
@@ -118,6 +147,15 @@ def get_turns(  # pylint: disable=too-many-return-statements
             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]

src/axolotl/monkeypatch/mixtral/__init__.py

@@ -0,0 +1,22 @@
+"""
+Patches to support multipack for mixtral
+"""
+import transformers
+
+
+def replace_mixtral_attn_with_multipack_flash_attn():
+    from .modeling_mixtral import (
+        MixtralMultipackFlashAttention2,
+        mixtral_decoder_layer_forward,
+        mixtral_model_forward,
+    )
+
+    transformers.models.mixtral.modeling_mixtral.MixtralDecoderLayer.forward = (
+        mixtral_decoder_layer_forward
+    )
+    transformers.models.mixtral.modeling_mixtral.MixtralModel.forward = (
+        mixtral_model_forward
+    )
+    transformers.models.mixtral.modeling_mixtral.MIXTRAL_ATTENTION_CLASSES[
+        "flash_attention_2"
+    ] = MixtralMultipackFlashAttention2

src/axolotl/monkeypatch/mixtral/modeling_mixtral.py

@@ -0,0 +1,383 @@
+"""
+Mixtral modeling for multipack
+"""
+# pylint: disable=missing-module-docstring,unused-argument,protected-access,pointless-string-statement,duplicate-code
+import logging
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+from einops import rearrange
+from flash_attn import flash_attn_varlen_qkvpacked_func
+from transformers import Cache, DynamicCache
+from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
+from transformers.modeling_outputs import MoeModelOutputWithPast
+from transformers.models.mixtral.modeling_mixtral import (
+    MixtralFlashAttention2,
+    apply_rotary_pos_emb,
+    repeat_kv,
+)
+
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+
+LOG = logging.getLogger("axolotl.monkeypatch.mixtral")
+
+
+class MixtralMultipackFlashAttention2(MixtralFlashAttention2):
+    """
+    Custom multipack implementation w flash attention 2
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._flash_attn_uses_top_left_mask = True
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(
+            bsz, q_len, self.num_heads, self.head_dim
+        ).transpose(1, 2)
+        key_states = key_states.view(
+            bsz, q_len, self.num_key_value_heads, self.head_dim
+        ).transpose(1, 2)
+        value_states = value_states.view(
+            bsz, q_len, self.num_key_value_heads, self.head_dim
+        ).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin, position_ids
+        )
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs
+            )
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        if cu_seqlens is not None and max_seqlen is not None and cu_seqlens.dim() == 1:
+            # special handling using sample packing
+            qkv = torch.stack(
+                [query_states, key_states, value_states], dim=2
+            )  # [bsz, nh, 3, q_len, hd]
+            qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]
+            qkv = rearrange(qkv, "b s ... -> (b s) ...")
+
+            attn_output = flash_attn_varlen_qkvpacked_func(
+                qkv,
+                cu_seqlens,
+                max_seqlen,
+                dropout_p=self.attention_dropout,
+                softmax_scale=None,
+                causal=True,
+            )
+            attn_output = rearrange(attn_output, "(b s) ... -> b s ...", b=bsz)
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+def mixtral_decoder_layer_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: Optional[bool] = False,
+    output_router_logits: Optional[bool] = False,
+    use_cache: Optional[bool] = False,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    max_seqlen: Optional[torch.Tensor] = None,
+    **kwargs,
+) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+    if "padding_mask" in kwargs:
+        warnings.warn(
+            "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+        )
+    """
+    Args:
+        hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+        attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+            `(batch, sequence_length)` where padding elements are indicated by 0.
+        past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+            returned tensors for more detail.
+        output_router_logits (`bool`, *optional*):
+            Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
+            should not be returned during inference.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+            (see `past_key_values`).
+    """
+
+    residual = hidden_states
+
+    hidden_states = self.input_layernorm(hidden_states)
+
+    # Self Attention
+    hidden_states, self_attn_weights, present_key_value = self.self_attn(
+        hidden_states=hidden_states,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_value=past_key_value,
+        output_attentions=output_attentions,
+        use_cache=use_cache,
+        cu_seqlens=cu_seqlens,
+        max_seqlen=max_seqlen,
+    )
+    hidden_states = residual + hidden_states
+
+    # Fully Connected
+    residual = hidden_states
+    hidden_states = self.post_attention_layernorm(hidden_states)
+    hidden_states, router_logits = self.block_sparse_moe(hidden_states)
+    hidden_states = residual + hidden_states
+
+    outputs = (hidden_states,)
+
+    if output_attentions:
+        outputs += (self_attn_weights,)
+
+    if use_cache:
+        outputs += (present_key_value,)
+
+    if output_router_logits:
+        outputs += (router_logits,)
+
+    return outputs
+
+
+def mixtral_model_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    output_router_logits: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+) -> Union[Tuple, MoeModelOutputWithPast]:
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_router_logits = (
+        output_router_logits
+        if output_router_logits is not None
+        else self.config.output_router_logits
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # retrieve input_ids and inputs_embeds
+    if input_ids is not None and inputs_embeds is not None:
+        raise ValueError(
+            "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
+        )
+    if input_ids is not None:
+        batch_size, seq_length = input_ids.shape
+    elif inputs_embeds is not None:
+        batch_size, seq_length, _ = inputs_embeds.shape
+    else:
+        raise ValueError(
+            "You have to specify either decoder_input_ids or decoder_inputs_embeds"
+        )
+
+    past_key_values_length = 0
+
+    if use_cache:
+        use_legacy_cache = not isinstance(past_key_values, Cache)
+        if use_legacy_cache:
+            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+        past_key_values_length = past_key_values.get_usable_length(seq_length)
+
+    cu_seqlens = None
+    max_seqlen = None
+    if position_ids is None:
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+        position_ids = torch.arange(
+            past_key_values_length,
+            seq_length + past_key_values_length,
+            dtype=torch.long,
+            device=device,
+        )
+        position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+    else:
+        position_ids = position_ids.view(-1, seq_length).long()
+        cu_seqlens, max_seqlen = get_cu_seqlens_from_pos_ids(position_ids)
+        cu_seqlens = cu_seqlens.squeeze()
+
+    if inputs_embeds is None:
+        inputs_embeds = self.embed_tokens(input_ids)
+
+    if (
+        attention_mask is not None
+        and self._attn_implementation == "flash_attention_2"
+        and use_cache
+    ):
+        is_padding_right = attention_mask[:, -1].sum().item() != batch_size
+        if is_padding_right:
+            raise ValueError(
+                "You are attempting to perform batched generation with padding_side='right'"
+                " this may lead to unexpected behaviour for Flash Attention version of Mixtral. Make sure to "
+                " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+            )
+
+    if self._attn_implementation == "flash_attention_2":
+        # 2d mask is passed through the layers
+        attention_mask = (
+            attention_mask
+            if (attention_mask is not None and 0 in attention_mask)
+            else None
+        )
+    else:
+        # 4d mask is passed through the layers
+        attention_mask = _prepare_4d_causal_attention_mask(
+            attention_mask,
+            (batch_size, seq_length),
+            inputs_embeds,
+            past_key_values_length,
+            sliding_window=self.config.sliding_window,
+        )
+
+    hidden_states = inputs_embeds
+
+    if self.gradient_checkpointing and self.training:
+        if use_cache:
+            LOG.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+            )
+            use_cache = False
+
+    # decoder layers
+    all_hidden_states = () if output_hidden_states else None
+    all_self_attns = () if output_attentions else None
+    all_router_logits = () if output_router_logits else None
+    next_decoder_cache = None
+
+    for decoder_layer in self.layers:
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if self.gradient_checkpointing and self.training:
+            layer_outputs = self._gradient_checkpointing_func(
+                decoder_layer.__call__,
+                hidden_states,
+                attention_mask,
+                position_ids,
+                past_key_values,
+                output_attentions,
+                output_router_logits,
+                use_cache,
+                cu_seqlens,
+                max_seqlen,
+            )
+        else:
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_values,
+                output_attentions=output_attentions,
+                output_router_logits=output_router_logits,
+                use_cache=use_cache,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=max_seqlen,
+            )
+
+        hidden_states = layer_outputs[0]
+
+        if use_cache:
+            next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+        if output_attentions:
+            all_self_attns += (layer_outputs[1],)
+
+        if output_router_logits:
+            all_router_logits += (layer_outputs[-1],)
+
+    hidden_states = self.norm(hidden_states)
+
+    # add hidden states from the last decoder layer
+    if output_hidden_states:
+        all_hidden_states += (hidden_states,)
+
+    next_cache = None
+    if use_cache:
+        next_cache = (
+            next_decoder_cache.to_legacy_cache()
+            if use_legacy_cache
+            else next_decoder_cache
+        )
+
+    if not return_dict:
+        return tuple(
+            v
+            for v in [
+                hidden_states,
+                next_cache,
+                all_hidden_states,
+                all_self_attns,
+                all_router_logits,
+            ]
+            if v is not None
+        )
+
+    return MoeModelOutputWithPast(
+        last_hidden_state=hidden_states,
+        past_key_values=next_cache,
+        hidden_states=all_hidden_states,
+        attentions=all_self_attns,
+        router_logits=all_router_logits,
+    )

src/axolotl/monkeypatch/neft_embeddings.py

@@ -1,65 +0,0 @@
-"""
-patches implemented through the trainer hooks to enable NEFT/noisy embeddings per https://arxiv.org/abs/2310.05914
-"""
-import torch
-from peft import PeftModel
-from transformers import PreTrainedModel
-
-
-def patch_neft(alpha, model):
-    embeddings = None
-    if isinstance(model, PreTrainedModel):
-        embeddings = model.get_input_embeddings()
-    if isinstance(model, PeftModel):
-        embeddings = model.base_model.get_input_embeddings()
-    if not embeddings:
-        raise ValueError(f"unhandled model class for neft: {model.__class__.__name__}")
-    embeddings.noisy_embedding_alpha = alpha
-    old_forward = embeddings.forward
-
-    # This hack seems to be needed to properly use a custom forward pass
-    # all credits to: https://discuss.pytorch.org/t/how-can-i-replace-the-forward-method-of-a-predefined-torchvision-model-with-my-customized-forward-function/54224/11
-    bound_method = neft_forward.__get__(  # pylint: disable=no-value-for-parameter
-        embeddings, embeddings.__class__
-    )
-    setattr(embeddings, "forward", bound_method)
-
-    embeddings._old_forward = old_forward  # pylint: disable=protected-access
-    return model
-
-
-def unpatch_neft(model):
-    embeddings = None
-    if isinstance(model, PreTrainedModel):
-        embeddings = model.get_input_embeddings()
-    if isinstance(model, PeftModel):
-        embeddings = model.base_model.get_input_embeddings()
-    if not embeddings:
-        raise ValueError(f"unhandled model class for neft: {model.__class__.__name__}")
-    if hasattr(embeddings, "_old_forward"):
-        embeddings.forward = embeddings._old_forward  # pylint: disable=protected-access
-        del embeddings._old_forward  # pylint: disable=protected-access
-        del embeddings.noisy_embedding_alpha
-
-
-def neft_forward(self, inputs: torch.Tensor):
-    embeddings = self._old_forward(inputs)  # pylint: disable=protected-access
-
-    if self.training:
-        dims = torch.tensor(embeddings.size(1) * embeddings.size(2))
-        mag_norm = self.noisy_embedding_alpha / torch.sqrt(dims)
-        embeddings = embeddings + torch.zeros_like(embeddings).uniform_(
-            -mag_norm, mag_norm
-        )
-
-    return embeddings
-
-
-def pretrain_hook(cfg, trainer):
-    if cfg.noisy_embedding_alpha:
-        trainer.model = patch_neft(cfg.noisy_embedding_alpha, trainer.model)
-
-
-def post_train_hook(cfg, trainer):
-    if cfg.noisy_embedding_alpha:
-        unpatch_neft(trainer.model)

src/axolotl/monkeypatch/xpos_rope_llama_monkey_patch.py

@@ -1,94 +0,0 @@
-# pylint: skip-file
-"""
-Copied from https://github.com/kaiokendev/cutoff-len-is-context-len/blob/main/util/xpos_rope_llama_monkey_patch.py
-"""
-import torch
-import transformers
-import transformers.models.llama.modeling_llama
-from einops import rearrange
-
-
-class XposRotaryEmbedding(torch.nn.Module):
-    def __init__(
-        self,
-        dim,
-        max_position_embeddings=2048,
-        base=10000,
-        device=None,
-        scale_base=2048,
-        use_xpos=True,
-    ):
-        super().__init__()
-        self.max_seq_len_cached = max_position_embeddings
-        self.scale_base = scale_base
-
-        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
-        t = torch.arange(self.max_seq_len_cached, device=device).type_as(inv_freq)
-        freqs = torch.einsum("i , j -> i j", t, inv_freq)
-        freqs = torch.cat((freqs, freqs), dim=-1)
-
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.register_buffer("freqs_cached", freqs, persistent=False)
-
-        if not use_xpos:
-            self.register_buffer("scale", None)
-            self.register_buffer("scale_cached", torch.ones(1))
-            return
-
-        scale = (torch.arange(0, dim, 2) + 0.4 * dim) / (1.4 * dim)
-        power = (t - (self.max_seq_len_cached // 2)) / self.scale_base
-        scale_cached = scale ** rearrange(power, "n -> n 1")
-        scale_cached = torch.cat((scale_cached, scale_cached), dim=-1)
-
-        self.register_buffer("scale", scale, persistent=False)
-        self.register_buffer("scale_cached", scale_cached, persistent=False)
-
-    def forward(
-        self,
-        x,
-        seq_len,
-    ):
-        if seq_len > self.max_seq_len_cached:
-            self.max_seq_len_cached = seq_len
-            t = torch.arange(self.max_seq_len_cached, device=x.device).type_as(
-                self.inv_freq
-            )
-            freqs = torch.einsum("i , j -> i j", t, self.inv_freq)
-            freqs = torch.cat((freqs, freqs), dim=-1).to(dtype=x.dtype)
-
-            self.register_buffer("freqs_cached", freqs)
-
-            if self.scale is None:
-                self.register_buffer(
-                    "scale_cached", torch.ones(1, device=x.device).to(dtype=x.dtype)
-                )
-
-                return self.freqs_cached.to(dtype=x.dtype), self.scale_cached
-
-            power = (t - (seq_len // 2)) / self.scale_base
-            scale = self.scale ** rearrange(power, "n -> n 1")
-            scale = torch.cat((scale, scale), dim=-1).to(dtype=x.dtype)
-            self.register_buffer("scale_cached", scale)
-
-        return self.freqs_cached.to(dtype=x.dtype), self.scale_cached.to(dtype=x.dtype)
-
-
-def rotate_half(x):
-    x1, x2 = x.chunk(2, dim=-1)
-    return torch.cat((-x2, x1), dim=-1)
-
-
-def apply_rotary_pos_emb(q, k, freqs, scale=1, position_ids=None):
-    freqs = freqs[position_ids, :]
-    if scale.shape[-1] != 1:
-        scale = scale[position_ids, :]
-
-    q_embed = (q * freqs.cos() * scale) + (rotate_half(q) * freqs.sin() * scale)
-    k_embed = (k * freqs.cos() * 1 / scale) + (rotate_half(k) * freqs.sin() * 1 / scale)
-
-    return q_embed, k_embed
-
-
-def replace_llama_rope_with_xpos_rope():
-    transformers.models.llama.modeling_llama.LlamaRotaryEmbedding = XposRotaryEmbedding
-    transformers.models.llama.modeling_llama.apply_rotary_pos_emb = apply_rotary_pos_emb

src/axolotl/prompt_strategies/llama2_chat.py

@@ -81,8 +81,9 @@ class LLama2ChatTokenizingStrategy(PromptTokenizingStrategy):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.sequence_len = 4096
-        self.tokenizer.add_special_tokens({"pad_token": "<pad>"})
+        self.tokenizer.add_special_tokens(
+            {"pad_token": getattr(self.tokenizer, "pad_token", "<pad>")}
+        )
         # https://huggingface.co/meta-llama/Llama-2-7b-chat-hf/blob/main/added_tokens.json
 
     def tokenize_prompt(self, prompt):

src/axolotl/prompt_strategies/sharegpt.py

@@ -13,7 +13,7 @@ register_conv_template(
         system_message="You are a helpful assistant.",
         roles=["<|im_start|>user", "<|im_start|>assistant"],
         sep_style=SeparatorStyle.CHATML,
-        sep="<|im_end|>\n",
+        sep="<|im_end|>",
     )
 )
 
@@ -39,6 +39,23 @@ def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
     return strategy
 
 
+def load_ultrachat(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
+    conversation = (
+        ds_cfg["conversation"] if ds_cfg and "conversation" in ds_cfg else None
+    )
+    strategy = UltrachatShareGPTPromptTokenizingStrategy(
+        ShareGPTPrompterV2(
+            conversation=conversation,
+        ),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+    if ds_cfg and "strict" in ds_cfg:
+        strategy.strict = ds_cfg["strict"]
+    return strategy
+
+
 def load_role(tokenizer, cfg):
     return SimpleRoleShareGPTPromptTokenizingStrategy(
         ShareGPTPrompterV2(),
@@ -109,3 +126,17 @@ class GuanacoShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
             {"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

src/axolotl/prompters.py

@@ -33,8 +33,8 @@ class AlpacaPrompter(Prompter):
     Base class for alpaca prompters
     """
 
-    system_prompt = "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n"
-    system_no_input_prompt = "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n"
+    system_prompt = "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request."
+    system_no_input_prompt = "Below is an instruction that describes a task. Write a response that appropriately completes the request."
     system_format: str = "{system}"
     turn_format: str
     turn_no_input_format: str

src/axolotl/train.py

@@ -12,12 +12,13 @@ import transformers.modelcard
 from accelerate.logging import get_logger
 from datasets import Dataset
 from optimum.bettertransformer import BetterTransformer
+from pkg_resources import get_distribution  # type: ignore
 from transformers.deepspeed import is_deepspeed_zero3_enabled
 
 from axolotl.common.cli import TrainerCliArgs
 from axolotl.logging_config import configure_logging
-from axolotl.monkeypatch import neft_embeddings
 from axolotl.utils.dict import DictDefault
+from axolotl.utils.freeze import freeze_parameters_except
 from axolotl.utils.models import load_model, load_tokenizer
 from axolotl.utils.trainer import setup_trainer
 
@@ -60,6 +61,12 @@ def train(
         msg += " and peft_config..."
     LOG.debug(msg)
     model, peft_config = load_model(cfg, tokenizer, inference=cli_args.inference)
+    model_ref = None
+    if cfg.rl:
+        # load the model again for model_ref/baseline
+        model_ref, _ = load_model(
+            cfg, tokenizer, inference=cli_args.inference, reference_model=True
+        )
 
     safe_serialization = cfg.save_safetensors is True
 
@@ -78,11 +85,15 @@ def train(
             )
     resume_from_checkpoint = cfg.resume_from_checkpoint
 
+    if cfg.unfrozen_parameters:
+        freeze_parameters_except(model, cfg.unfrozen_parameters)
+
     trainer = setup_trainer(
-        cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps
+        cfg, train_dataset, eval_dataset, (model, model_ref), tokenizer, total_num_steps
     )
 
-    model.config.use_cache = False
+    if hasattr(model, "config"):
+        model.config.use_cache = False
 
     # go ahead and presave, so we have the adapter config available to inspect
     if peft_config:
@@ -92,7 +103,8 @@ def train(
     if not Path(cfg.output_dir).is_dir():
         os.makedirs(cfg.output_dir, exist_ok=True)
     tokenizer.save_pretrained(str(Path(cfg.output_dir)))
-    model.config.save_pretrained(str(Path(cfg.output_dir)))
+    if hasattr(model, "config"):
+        model.config.save_pretrained(str(Path(cfg.output_dir)))
 
     # In case we want to stop early with ctrl+c, this is a nice to have to save the pretrained model
     if cfg.local_rank == 0:
@@ -110,6 +122,12 @@ def train(
     badge_markdown = """[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)"""
     transformers.modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n{badge_markdown}"
 
+    if getattr(cfg, "axolotl_config_path"):
+        raw_axolotl_cfg = Path(cfg.axolotl_config_path)
+        version = get_distribution("axolotl").version
+        if raw_axolotl_cfg.is_file():
+            transformers.modelcard.AUTOGENERATED_TRAINER_COMMENT += f"\n<details><summary>See axolotl config</summary>\n\naxolotl version: `{version}`\n```yaml\n{raw_axolotl_cfg.read_text(encoding='utf-8')}\n```\n\n</details><br>\n"
+
     LOG.info("Starting trainer...")
     if cfg.group_by_length:
         LOG.info("hang tight... sorting dataset for group_by_length")
@@ -170,25 +188,26 @@ def train(
 
     if not cfg.hub_model_id:
         trainer.create_model_card(model_name=cfg.output_dir.lstrip("./"))
+    elif cfg.hub_model_id:
+        # defensively push to the hub to ensure the model card is updated
+        trainer.push_to_hub()
 
     return model, tokenizer
 
 
-def pretrain_hooks(cfg, trainer):
+def pretrain_hooks(_cfg, _trainer):
     """
     Run hooks right before kicking off the training
     :param cfg:
     :param trainer:
     :return:
     """
-    neft_embeddings.pretrain_hook(cfg, trainer)
 
 
-def post_train_hooks(cfg, trainer):
+def post_train_hooks(_cfg, _trainer):
     """
     Run hooks right after training completes
     :param cfg:
     :param trainer:
     :return:
     """
-    neft_embeddings.post_train_hook(cfg, trainer)

src/axolotl/utils/callbacks.py

@@ -4,6 +4,8 @@ from __future__ import annotations
 
 import logging
 import os
+from shutil import copyfile
+from tempfile import NamedTemporaryFile
 from typing import TYPE_CHECKING, Dict, List
 
 import evaluate
@@ -124,6 +126,36 @@ class GPUStatsCallback(
         return control
 
 
+class LossWatchDogCallback(TrainerCallback):
+    """Callback to track loss and stop training if loss is too high"""
+
+    def __init__(self, cfg):
+        self.cfg = cfg
+        self.logged = False
+        self.violations = 0
+        self.threshold = cfg.loss_watchdog_threshold
+        self.patience = cfg.loss_watchdog_patience or 3
+
+    def on_step_end(
+        self,
+        _args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        **_kwargs,
+    ):
+        if len(state.log_history) > 0 and "loss" in state.log_history[-1]:
+            if state.log_history[-1]["loss"] > self.threshold:
+                self.violations += 1
+                if self.violations >= self.patience:
+                    LOG.warning(
+                        "Loss is too high, stopping training (loss_watchdog_threshold)"
+                    )
+                    control.should_training_stop = True
+            else:
+                self.violations = 0
+        return control
+
+
 def bench_eval_callback_factory(trainer, tokenizer):
     accuracy = evaluate.load("accuracy")
     abcd_idx = [
@@ -531,10 +563,15 @@ class SaveAxolotlConfigtoWandBCallback(TrainerCallback):
     ):
         if is_main_process():
             try:
-                artifact = wandb.Artifact(name="axolotl-config", type="config")
-                artifact.add_file(local_path=self.axolotl_config_path)
-                wandb.run.log_artifact(artifact)
-                LOG.info("Axolotl config has been saved to WandB as an artifact.")
+                # sync config to top level in run, cannot delete file right away because wandb schedules it to be synced even w/policy = 'now', so let OS delete it later.
+                with NamedTemporaryFile(
+                    mode="w", delete=False, suffix=".yml", prefix="axolotl_config_"
+                ) as temp_file:
+                    copyfile(self.axolotl_config_path, temp_file.name)
+                    wandb.save(temp_file.name)
+                LOG.info(
+                    "The Axolotl config has been saved to the WandB run under files."
+                )
             except (FileNotFoundError, ConnectionError) as err:
                 LOG.warning(f"Error while saving Axolotl config to WandB: {err}")
         return control

src/axolotl/utils/chat_templates.py

@@ -0,0 +1,29 @@
+"""
+This module provides functionality for selecting chat templates based on user choices.
+These templates are used for formatting messages in a conversation.
+"""
+
+
+def chat_templates(user_choice: str):
+    """
+    Finds the correct chat_template for the tokenizer_config.
+
+    Args:
+        user_choice (str): The user's choice of template.
+
+    Returns:
+        str: The chosen template string.
+
+    Raises:
+        ValueError: If the user_choice is not found in the templates.
+    """
+
+    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 %}",
+    }
+
+    if user_choice in templates:
+        return templates[user_choice]
+
+    raise ValueError(f"Template '{user_choice}' not found.")

src/axolotl/utils/collators.py

@@ -2,12 +2,16 @@
 DataCollator for axolotl to pad labels and position_ids for packed sequences
 """
 from dataclasses import dataclass
-from typing import Any, Optional, Union
+from typing import Any, Dict, Optional, Sequence, Union
 
 import numpy as np
+import torch
+import transformers
 from transformers import PreTrainedTokenizerBase
 from transformers.utils import PaddingStrategy
 
+IGNORE_INDEX = -100
+
 
 @dataclass
 class DataCollatorForSeq2Seq:
@@ -146,3 +150,52 @@ class BatchSamplerDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
                 chunked_data[feature] = np.concatenate(arrays)
         features = [chunked_data]
         return super().__call__(features, return_tensors=return_tensors)
+
+
+@dataclass
+class MambaDataCollator:
+    """
+    Collator for State Space Models (Mamba)
+    """
+
+    tokenizer: transformers.PreTrainedTokenizer
+
+    def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
+        input_ids, labels = tuple(
+            [torch.LongTensor(instance[key]) for instance in instances]
+            for key in ("input_ids", "labels")
+        )
+        input_ids = torch.nn.utils.rnn.pad_sequence(
+            input_ids,
+            batch_first=True,
+            padding_value=self.tokenizer.pad_token_id,
+        )
+        labels = torch.nn.utils.rnn.pad_sequence(
+            labels, batch_first=True, padding_value=IGNORE_INDEX
+        )
+
+        return {
+            "input_ids": input_ids,
+            "labels": labels,
+        }
+
+
+@dataclass
+class PretrainingBatchSamplerDataCollatorForSeq2Seq(DataCollatorForSeq2Seq):
+    """
+    Collator for multipack specific to the using the BatchSampler
+    """
+
+    def __call__(self, features, return_tensors=None):
+        chunked_data = {}
+        for feature in features.keys():
+            if feature == "length":
+                continue
+            if feature == "attention_mask":
+                arrays = [(1) * np.array(item) for item in features[feature]]
+                chunked_data[feature] = np.concatenate(arrays)
+            else:
+                arrays = [np.array(item) for item in features[feature]]
+                chunked_data[feature] = np.concatenate(arrays)
+        features = [chunked_data]
+        return super().__call__(features, return_tensors=return_tensors)

src/axolotl/utils/config.py

@@ -27,7 +27,7 @@ def choose_device(cfg):
 
     cfg.device = get_device()
     if cfg.world_size == 1:
-        cfg.device_map = "auto"
+        cfg.device_map = cfg.device_map or "auto"
     else:
         if cfg.device.startswith("cuda"):
             cfg.device_map = {"": torch.cuda.current_device()}
@@ -77,6 +77,15 @@ def normalize_config(cfg):
     else:
         cfg.torch_dtype = torch.float32
 
+    if cfg.saves_per_epoch:
+        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:
+        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
+
     cfg.dataset_processes = cfg.dataset_processes or os.cpu_count()
 
     if not cfg.base_model_config:
@@ -352,6 +361,27 @@ def validate_config(cfg):
                 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."
+        )
+    if cfg.saves_per_epoch and cfg.save_strategy and cfg.save_strategy != "steps":
+        raise ValueError(
+            "save_strategy must be empty or set to `steps` when used with saves_per_epoch."
+        )
+    if cfg.evals_per_epoch and cfg.eval_steps:
+        raise ValueError(
+            "eval_steps and evals_per_epoch are mutually exclusive and cannot be used together."
+        )
+    if (
+        cfg.evals_per_epoch
+        and cfg.evaluation_strategy
+        and cfg.evaluation_strategy != "steps"
+    ):
+        raise ValueError(
+            "evaluation_strategy must be empty or set to `steps` when used with evals_per_epoch."
+        )
     if cfg.save_strategy and cfg.save_steps and cfg.save_strategy != "steps":
         raise ValueError(
             "save_strategy and save_steps mismatch. Please set save_strategy to 'steps' or remove save_steps."
@@ -392,9 +422,44 @@ def validate_config(cfg):
     if cfg.warmup_steps and cfg.warmup_ratio:
         raise ValueError("warmup_steps and warmup_ratio are mutually exclusive")
 
-    if cfg.is_qwen_derived_model and cfg.gradient_checkpointing:
+    if cfg.wandb_run_id and not cfg.wandb_name:
+        cfg.wandb_name = cfg.wandb_run_id
+
         LOG.warning(
-            "Gradient checkpointing is broken for Qwen models for transformers>=4.35.0, except main branch."
+            "wandb_run_id sets the ID of the run. If you would like to set the name, please use wandb_name instead."
+        )
+
+    if cfg.noisy_embedding_alpha is not None:
+        # Deprecated, use neftune_noise_alpha
+        LOG.warning("noisy_embedding_alpha is deprecated, use neftune_noise_alpha")
+        if cfg.neftune_noise_alpha is None:
+            cfg.neftune_noise_alpha = cfg.noisy_embedding_alpha
+        else:
+            # User is providing both; bail and have them sort out their settings
+            raise ValueError(
+                "noisy_embedding_alpha is deprecated, use neftune_noise_alpha; both are set, please remove the deprecated noisy_embedding_alpha setting"
+            )
+
+    if cfg.neftune_noise_alpha is not None and cfg.neftune_noise_alpha <= 0.0:
+        raise ValueError("neftune_noise_alpha must be > 0.0")
+
+    if (
+        cfg.adapter
+        and cfg.tokens
+        and (
+            not cfg.lora_modules_to_save
+            or not all(
+                x in cfg.lora_modules_to_save for x in ["embed_tokens", "lm_head"]
+            )
+        )
+    ):
+        raise ValueError(
+            "lora_modules_to_save not properly set yet adding new tokens. Please add `embed_tokens` and `lm_head` to `lora_modules_to_save`."
+        )
+
+    if cfg.max_memory is not None and cfg.gpu_memory_limit is not None:
+        raise ValueError(
+            "max_memory and gpu_memory_limit are mutually exclusive and cannot be used together."
         )
 
     # TODO

src/axolotl/utils/data.py

@@ -2,6 +2,7 @@
 import functools
 import hashlib
 import logging
+from collections import defaultdict
 from pathlib import Path
 from typing import Dict, List, Tuple, Union
 
@@ -14,6 +15,7 @@ from datasets import (
     load_from_disk,
 )
 from huggingface_hub import hf_hub_download
+from torch.utils.data import RandomSampler
 from transformers import PreTrainedTokenizerBase
 
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
@@ -39,11 +41,14 @@ from axolotl.prompters import (
     SummarizeTLDRPrompter,
     UnsupportedPrompter,
 )
+from axolotl.utils.collators import PretrainingBatchSamplerDataCollatorForSeq2Seq
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process, zero_first
+from axolotl.utils.samplers.multipack import MultipackBatchSampler
 from axolotl.utils.trainer import (
     calculate_total_num_steps,
     process_datasets_for_packing,
+    process_pretraining_datasets_for_packing,
 )
 
 LOG = logging.getLogger("axolotl")
@@ -64,9 +69,17 @@ def prepare_dataset(cfg, tokenizer):
                 tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
             )
     else:
+        path = cfg.pretraining_dataset
+        name = None
+        if isinstance(cfg.pretraining_dataset, dict):
+            path = cfg.pretraining_dataset["path"]
+            name = cfg.pretraining_dataset["name"]
+
         train_dataset = load_pretraining_dataset(
-            cfg.pretraining_dataset,
+            path,
             tokenizer,
+            cfg,
+            name=name,
             max_tokens=cfg.sequence_len,
             seed=cfg.seed or 42,
         )
@@ -806,9 +819,27 @@ def encode_pretraining(
     return ret
 
 
-def load_pretraining_dataset(path, tokenizer, max_tokens=2048, seed=42):
-    encode = functools.partial(encode_pretraining, tokenizer, max_tokens)
-    dataset = load_dataset(path, streaming=True, split="train")
+def load_pretraining_dataset(path, tokenizer, cfg, name=None, max_tokens=2048, seed=42):
+    if cfg.sample_packing:
+        collate_fn = PretrainingBatchSamplerDataCollatorForSeq2Seq(
+            tokenizer,
+            return_tensors="pt",
+            padding=True,
+            pad_to_multiple_of=max_tokens * cfg.micro_batch_size,
+        )
+        encode = functools.partial(
+            encode_packed_pretraining,
+            tokenizer,
+            collate_fn,
+            max_seq_length=max_tokens,
+            batch_size=cfg.micro_batch_size,
+        )
+        # set this to 1 so downstream data_loader doesn't try to increase the batch again
+        cfg.micro_batch_size = 1
+    else:
+        encode = functools.partial(encode_pretraining, tokenizer, max_tokens)
+
+    dataset = load_dataset(path, streaming=True, split="train", name=name)
     dataset = dataset.shuffle(seed=seed, buffer_size=10_000)
     dataset = dataset.map(
         encode,
@@ -819,3 +850,63 @@ def load_pretraining_dataset(path, tokenizer, max_tokens=2048, seed=42):
         remove_columns=dataset.features.keys(),
     )
     return dataset
+
+
+def encode_packed_pretraining(
+    tokenizer: PreTrainedTokenizerBase,
+    collate_fn,
+    examples: List[str],
+    max_seq_length: int = 2048,
+    batch_size: int = 4,
+) -> Dict[str, List]:
+    # pylint: disable=duplicate-code
+    # tokenize all the examples
+    # rows get split with stride (overlap)
+    res = tokenizer(
+        examples,
+        truncation=True,
+        max_length=max_seq_length - 1,
+        add_special_tokens=True,
+        return_overflowing_tokens=True,
+        stride=256,
+    )
+
+    input_ids = [seq + [tokenizer.eos_token_id] for seq in res["input_ids"]]
+    attention_mask = [seq + [1] for seq in res["attention_mask"]]
+
+    tokenized_examples = {
+        "input_ids": input_ids,
+        "attention_mask": attention_mask,
+    }
+
+    train_dataset = Dataset.from_dict(tokenized_examples)
+    train_dataset = process_pretraining_datasets_for_packing(
+        train_dataset, max_seq_length
+    )
+
+    sampler = MultipackBatchSampler(
+        RandomSampler(train_dataset),
+        batch_size=batch_size,
+        drop_last=True,
+        batch_max_len=batch_size * max_seq_length,
+        lengths=(
+            train_dataset.data.column("position_ids")
+            .to_pandas()
+            .apply(lambda x: x[-1] + 1)
+            .values
+        ),
+    )
+
+    chunked_data = defaultdict(list)
+
+    for data in sampler:
+        features = train_dataset[data]
+        features["labels"] = features["input_ids"].copy()
+        collated_features = collate_fn(features)
+
+        for feature in features.keys():
+            if feature == "length":
+                continue
+            chunked_data[feature].append(collated_features[feature].squeeze(0))
+
+    return chunked_data

src/axolotl/utils/freeze.py

@@ -0,0 +1,38 @@
+"""
+module to freeze/unfreeze parameters by name
+"""
+import logging
+import re
+
+from axolotl.utils.distributed import is_main_process
+
+LOG = logging.getLogger("axolotl.utils.freeze")
+
+
+def freeze_parameters_except(model, regex_patterns):
+    """
+    Freezes all layers of the given model except for the layers that match given regex patterns.
+    Periods in the patterns are treated as literal periods, not as wildcard characters.
+
+    Parameters:
+    - model (nn.Module): The PyTorch model to be modified.
+    - regex_patterns (list of str): List of regex patterns to match layer names to keep unfrozen.
+
+    Returns:
+    None; the model is modified in place.
+    """
+    # Escape periods and compile the regex patterns
+    compiled_patterns = [
+        re.compile(pattern.replace(".", "\\.")) for pattern in regex_patterns
+    ]
+
+    # First, freeze all parameters in the model
+    for param in model.parameters():
+        param.requires_grad = False
+
+    # Unfreeze layers that match the regex patterns
+    for name, param in model.named_parameters():
+        if any(pattern.match(name) for pattern in compiled_patterns):
+            if is_main_process():
+                LOG.debug(f"unfreezing {name}")
+            param.requires_grad = True

src/axolotl/utils/models.py

@@ -2,8 +2,9 @@
 import logging
 import math
 import os
-from typing import Optional, Tuple  # noqa: F401
+from typing import Any, Optional, Tuple  # noqa: F401
 
+import addict
 import bitsandbytes as bnb
 import torch
 import transformers
@@ -20,24 +21,61 @@ from transformers import (  # noqa: F401
     PreTrainedModel,
     PreTrainedTokenizerBase,
 )
+from transformers.deepspeed import is_deepspeed_zero3_enabled
 
+from axolotl.models.mamba import fix_mamba_attn_for_loss
 from axolotl.prompt_tokenizers import LLAMA_DEFAULT_EOS_TOKEN
 from axolotl.utils.bench import log_gpu_memory_usage
+from axolotl.utils.chat_templates import chat_templates
 from axolotl.utils.dict import DictDefault
 
 LOG = logging.getLogger("axolotl")
 
 
+def check_model_config(cfg: DictDefault, model_config: AutoConfig):
+    quant_config_exists = hasattr(model_config, "quantization_config")
+    quant_config_method_is_gptq = (
+        quant_config_exists
+        and "quant_method" in model_config.quantization_config
+        and model_config.quantization_config["quant_method"] == "gptq"
+    )
+
+    if cfg.gptq and not quant_config_method_is_gptq:
+        raise ValueError(
+            "model_config.quantization_config is not set or quant_method is not set to gptq. "
+            "Please make sure to point to a GPTQ model."
+        )
+
+    if not cfg.gptq and quant_config_exists:
+        raise ValueError(
+            "model_config.quantization_config is set but `gptq` flag is not. "
+            "Please use the `gptq` flag to train quantized model or point to a non-quantized model."
+        )
+
+
 def load_model_config(cfg):
     model_config_name = cfg.base_model_config or cfg.base_model
     trust_remote_code = cfg.trust_remote_code is True
-    model_config = AutoConfig.from_pretrained(
-        model_config_name, trust_remote_code=trust_remote_code
-    )
+
+    try:
+        model_config = AutoConfig.from_pretrained(
+            model_config_name, trust_remote_code=trust_remote_code
+        )
+    except ValueError as err:
+        if "mamba" in model_config_name:
+            return addict.Dict(
+                {
+                    "model_type": "mamba",
+                }
+            )
+        raise err
+
     if cfg.model_config:
         for key, val in cfg.model_config.items():
             setattr(model_config, key, val)
 
+    check_model_config(cfg, model_config)
+
     return model_config
 
 
@@ -69,6 +107,7 @@ def load_tokenizer(cfg):
             "LlamaTokenizer",
             "LlamaTokenizerFast",
             "CodeLlamaTokenizer",
+            "CodeLlamaTokenizerFast",
         ]
         and hasattr(tokenizer, "pad_token")
         and not tokenizer.pad_token
@@ -98,9 +137,43 @@ def load_tokenizer(cfg):
 
     if cfg.special_tokens:
         for k, val in cfg.special_tokens.items():
+            # check if new special token is not already in tokenizer and
+            # is adapter training to make sure lora_modules_to_save is set
+            if (
+                (getattr(tokenizer, k) is None or getattr(tokenizer, k) != val)
+                and cfg.adapter
+                and (
+                    not cfg.lora_modules_to_save
+                    or not all(
+                        x in cfg.lora_modules_to_save
+                        for x in ["embed_tokens", "lm_head"]
+                    )
+                )
+            ):
+                raise ValueError(
+                    "Please set lora_modules_to_save to ['embed_tokens', 'lm_head'] when using an adapter and changing the special tokens."
+                )
+
             tokenizer.add_special_tokens(
                 {k: AddedToken(val, rstrip=False, lstrip=False, normalized=False)}
             )
+
+        # If we add bos_token and eos_token, we need to update the post processor to
+        # handle them correctly.
+        # https://github.com/huggingface/transformers/pull/24132
+        bos_or_eos_in_special_tokens = (
+            "bos_token" in cfg.special_tokens and "eos_token" in cfg.special_tokens
+        )
+        if (
+            tokenizer.__class__.__name__
+            in (
+                "LlamaTokenizerFast",
+                "CodeLlamaTokenizerFast",
+            )
+            and bos_or_eos_in_special_tokens
+        ):
+            tokenizer.update_post_processor()
+
     if cfg.tokens:
         tokenizer.add_tokens(
             [
@@ -114,6 +187,12 @@ def load_tokenizer(cfg):
     LOG.debug(f"PAD: {tokenizer.pad_token_id} / {tokenizer.pad_token}")
     LOG.debug(f"UNK: {tokenizer.unk_token_id} / {tokenizer.unk_token}")
 
+    if cfg.chat_template:
+        tokenizer.chat_template = chat_templates(cfg.chat_template)
+    else:
+        LOG.info(
+            "No Chat template selected. Consider adding a chat template for easier inference."
+        )
     return tokenizer
 
 
@@ -121,6 +200,7 @@ def load_model(
     cfg: DictDefault,
     tokenizer: PreTrainedTokenizerBase,
     inference: bool = False,
+    reference_model: bool = False,
 ) -> Tuple[PreTrainedModel, Optional[PeftConfig]]:
     """
     Load a model for a given configuration and tokenizer.
@@ -175,17 +255,6 @@ def load_model(
 
         LOG.info("patching with sdp attention")
         hijack_llama_sdp_attention()
-    elif cfg.is_llama_derived_model and cfg.landmark_attention:
-        from axolotl.monkeypatch.llama_landmark_attn import (
-            MEM_TOKEN,
-            patch_llama_with_landmark_attn,
-        )
-
-        LOG.info("patching with landmark attention")
-        patch_llama_with_landmark_attn()
-
-        # Note: This might overwrite previous additional_special_tokens
-        tokenizer.add_special_tokens({"additional_special_tokens": [MEM_TOKEN]})
 
     if cfg.is_mistral_derived_model and cfg.flash_attention and cfg.sample_packing:
         from axolotl.monkeypatch.mistral_attn_hijack_flash import (
@@ -195,13 +264,17 @@ def load_model(
         LOG.info("patching with flash attention")
         replace_mistral_attn_with_flash_attn(packed=cfg.sample_packing)
 
-    if cfg.is_llama_derived_model and cfg.xpos_rope:
-        from axolotl.monkeypatch.xpos_rope_llama_monkey_patch import (
-            replace_llama_rope_with_xpos_rope,
+    if (
+        cfg.model_config_type == "mixtral"
+        and cfg.flash_attention
+        and cfg.sample_packing
+    ):
+        from axolotl.monkeypatch.mixtral import (
+            replace_mixtral_attn_with_multipack_flash_attn,
         )
 
-        LOG.info("patching with xpos rope")
-        replace_llama_rope_with_xpos_rope()
+        LOG.info("patching with flash attention")
+        replace_mixtral_attn_with_multipack_flash_attn()
 
     if (
         cfg.is_llama_derived_model
@@ -215,8 +288,50 @@ def load_model(
 
     model_kwargs = {}
 
-    model_kwargs["device_map"] = cfg.device_map
+    max_memory = cfg.max_memory
+    device_map = cfg.device_map
+
+    if cfg.gpu_memory_limit:
+        gpu_memory_limit = (
+            str(cfg.gpu_memory_limit) + "GiB"
+            if isinstance(cfg.gpu_memory_limit, int)
+            else cfg.gpu_memory_limit
+        )
+
+        max_memory = {}
+        for i in range(torch.cuda.device_count()):
+            max_memory[i] = gpu_memory_limit
+        max_memory["cpu"] = "256GiB"  # something sufficiently large to fit anything
+
+    if max_memory is not None:
+        # Based on https://github.com/togethercomputer/OpenChatKit/blob/main/inference/bot.py
+        from accelerate import infer_auto_device_map, init_empty_weights
+
+        with init_empty_weights():
+            model_canvas = AutoModelForCausalLM.from_config(model_config)
+        model_canvas.tie_weights()
+        device_map = infer_auto_device_map(
+            model_canvas,
+            max_memory=max_memory,
+            dtype=cfg.torch_dtype,
+        )
+        # We can discard max_memory now as we have a device map set up for us
+        max_memory = None
+
+    model_kwargs["device_map"] = device_map
     model_kwargs["torch_dtype"] = cfg.torch_dtype
+    # TODO can we put the reference model on it's own gpu? I think we have to move logits around to calculate loss
+    # if cfg.rl:
+    #     if torch.cuda.device_count() > 1:
+    #         if reference_model:
+    #             model_kwargs["device_map"] = "cuda:" + str(
+    #                 torch.cuda.current_device() + 1
+    #             )
+    #         else:
+    #             model_kwargs["device_map"] = "cuda:" + str(torch.cuda.current_device())
+
+    if is_deepspeed_zero3_enabled():
+        del model_kwargs["device_map"]
 
     if cfg.model_revision:
         model_kwargs["revision"] = cfg.model_revision
@@ -232,22 +347,45 @@ def load_model(
                 **model_config.quantization_config
             )
     if cfg.adapter == "qlora" and cfg.load_in_4bit:
+        bnb_config = {
+            "load_in_4bit": True,
+            "llm_int8_threshold": 6.0,
+            "llm_int8_has_fp16_weight": False,
+            "bnb_4bit_compute_dtype": cfg.torch_dtype,
+            "bnb_4bit_use_double_quant": True,
+            "bnb_4bit_quant_type": "nf4",
+        }
+
+        if cfg.bnb_config_kwargs:
+            bnb_config.update(cfg.bnb_config_kwargs)
+
         model_kwargs["quantization_config"] = BitsAndBytesConfig(
-            load_in_4bit=True,
-            llm_int8_threshold=6.0,
-            llm_int8_has_fp16_weight=False,
-            bnb_4bit_compute_dtype=cfg.torch_dtype,
-            bnb_4bit_use_double_quant=True,
-            bnb_4bit_quant_type="nf4",
+            **bnb_config,
         )
     # sample packing uses custom FA2 patch
-    if cfg.flash_attention and not cfg.sample_packing:
-        if (
-            cfg.is_llama_derived_model
-            or cfg.is_falcon_derived_model
-            or cfg.is_mistral_derived_model
-        ):
-            model_kwargs["use_flash_attention_2"] = True
+    if cfg.flash_attention:
+        if not cfg.sample_packing:
+            if (
+                cfg.is_llama_derived_model
+                or cfg.is_falcon_derived_model
+                or cfg.is_mistral_derived_model
+                or model_config.model_type == "mixtral"
+            ):
+                model_kwargs["attn_implementation"] = "flash_attention_2"
+                model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "flash_attention_2"
+                )
+        else:
+            if model_config.model_type == "mixtral":
+                model_kwargs["attn_implementation"] = "flash_attention_2"
+                model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "flash_attention_2"
+                )
+            else:
+                model_kwargs["attn_implementation"] = "eager"
+                model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "eager"
+                )
 
     try:
         if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
@@ -309,6 +447,19 @@ def load_model(
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
                 **model_kwargs,
             )
+        elif model_type == "MambaLMHeadModel":
+            # FIXME this is janky at best and hacked together to make it work
+            MambaLMHeadModel = fix_mamba_attn_for_loss()  # pylint: disable=invalid-name
+
+            model_kwargs["dtype"] = model_kwargs["torch_dtype"]
+            model_kwargs["device"] = torch.cuda.current_device()
+            del model_kwargs["torch_dtype"]
+            del model_kwargs["device_map"]
+
+            model = MambaLMHeadModel.from_pretrained(
+                base_model,
+                **model_kwargs,
+            )
         elif model_type and not cfg.trust_remote_code:
             if cfg.gptq:
                 model = AutoModelForCausalLM.from_pretrained(
@@ -368,13 +519,17 @@ def load_model(
         if cfg.resize_token_embeddings_to_32x
         else len(tokenizer)
     )
-    if model.get_input_embeddings().num_embeddings < embeddings_len:
+    if (
+        hasattr(model, "get_input_embeddings")
+        and model.get_input_embeddings().num_embeddings < embeddings_len
+    ):
         model.resize_token_embeddings(embeddings_len)
     else:
         model.tie_weights()
 
     if (
-        hasattr(model.config, "max_position_embeddings")
+        hasattr(model, "config")
+        and hasattr(model.config, "max_position_embeddings")
         and model.config.max_position_embeddings
         and cfg.sequence_len > model.config.max_position_embeddings
     ):
@@ -384,20 +539,22 @@ def load_model(
         model.config.max_position_embeddings = cfg.sequence_len
 
     if (
-        hasattr(model.config, "bos_token_id")
+        hasattr(model, "config")
+        and hasattr(model.config, "bos_token_id")
         and model.config.bos_token_id
         and model.config.bos_token_id != tokenizer.bos_token_id
     ):
         model.config.bos_token_id = tokenizer.bos_token_id
 
     if (
-        hasattr(model.config, "eos_token_id")
+        hasattr(model, "config")
+        and hasattr(model.config, "eos_token_id")
         and model.config.eos_token_id
         and model.config.eos_token_id != tokenizer.eos_token_id
     ):
         model.config.eos_token_id = tokenizer.eos_token_id
 
-    if model.device.type == "cuda":
+    if hasattr(model, "device") and model.device.type == "cuda":
         log_gpu_memory_usage(LOG, "after model load", model.device)
 
     # make sure these are fp32 per Ramesh et al. (2021)
@@ -441,9 +598,11 @@ def load_model(
                 if hasattr(module, "weight"):
                     module.to(cfg.torch_dtype)
 
-    model, lora_config = load_adapter(model, cfg, cfg.adapter)
+    lora_config = None
+    if not reference_model or cfg.lora_model_dir:
+        model, lora_config = load_adapter(model, cfg, cfg.adapter)
 
-    if cfg.ddp and not load_in_8bit:
+    if cfg.ddp and not load_in_8bit and not (cfg.rl and cfg.load_in_4bit):
         model.to(f"cuda:{cfg.local_rank}")
 
     if torch.cuda.device_count() > 1 and int(os.getenv("WORLD_SIZE", "1")) == 1:
@@ -456,7 +615,8 @@ def load_model(
             requires_grad.append(f"{name}: {param.requires_grad}")
     if len(requires_grad) == 0:
         LOG.warning("there are no parameters that require gradient updates")
-    model.config.use_cache = False
+    if hasattr(model, "config"):
+        model.config.use_cache = False
 
     if cfg.flash_optimum:
         model = BetterTransformer.transform(model)
@@ -551,10 +711,15 @@ def load_lora(model, cfg, inference=False):
 
     if cfg.lora_model_dir:
         LOG.debug("Loading pretained PEFT - LoRA")
+        model_kwargs: Any = {}
+        if cfg.lora_on_cpu:
+            model_kwargs["max_memory"] = {"cpu": "256GiB"}
+            model_kwargs["device_map"] = {"": "cpu"}
         model = PeftModel.from_pretrained(
             model,
             cfg.lora_model_dir,
             is_trainable=(not inference),
+            **model_kwargs,
         )
     else:
         model = get_peft_model(model, lora_config)

src/axolotl/utils/tokenization.py

@@ -31,8 +31,8 @@ def check_example_labels(example, tokenizer, text_only=False):
         )
         colored_tokens.append(colored_token)
 
-    delimiter = "" if text_only else " "
-    LOG.info(delimiter.join(colored_tokens))
+    output = " ".join(colored_tokens)
+    LOG.info(output)
     LOG.info("\n\n\n")
 
-    return " ".join(colored_tokens)
+    return output

src/axolotl/utils/trainer.py

@@ -12,7 +12,7 @@ from accelerate.logging import get_logger
 from datasets import set_caching_enabled
 from torch.utils.data import DataLoader, RandomSampler
 
-from axolotl.core.trainer_builder import HFCausalTrainerBuilder
+from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFDPOTrainerBuilder
 from axolotl.utils.distributed import is_main_process, reduce_and_broadcast, zero_first
 from axolotl.utils.samplers import MultipackBatchSampler
 
@@ -131,8 +131,10 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
                     )
 
         # Phi doesn't want the attention_mask feature when training
-        if "CodeGenTokenizer" in tokenizer.__class__.__name__ or (
-            cfg.is_mistral_derived_model and cfg.flash_attention
+        if (
+            "CodeGenTokenizer" in tokenizer.__class__.__name__
+            or (cfg.is_mistral_derived_model and cfg.flash_attention)
+            or cfg.model_config_type == "mamba"
         ):
             train_dataset = train_dataset.remove_columns("attention_mask")
             if eval_dataset:
@@ -141,6 +143,16 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
     return train_dataset, eval_dataset
 
 
+def process_pretraining_datasets_for_packing(train_dataset, sequence_len):
+    drop_long = partial(drop_long_seq, sequence_len=sequence_len)
+
+    train_dataset = train_dataset.filter(drop_long)
+    train_dataset = train_dataset.map(
+        add_position_ids,
+    )
+    return train_dataset
+
+
 def calculate_total_num_steps(cfg, train_dataset, update=True):
     if not cfg.total_num_tokens:
         total_num_tokens = np.sum(
@@ -153,7 +165,9 @@ def calculate_total_num_steps(cfg, train_dataset, update=True):
         if update:
             cfg.total_num_tokens = total_num_tokens
 
-    if not cfg.total_supervised_tokens:
+    skip_estimates = cfg.model_config_type == "mamba"
+
+    if not skip_estimates and not cfg.total_supervised_tokens:
         total_supervised_tokens = (
             train_dataset.data.column("labels")
             .to_pandas()
@@ -167,7 +181,7 @@ def calculate_total_num_steps(cfg, train_dataset, update=True):
         if update:
             cfg.total_supervised_tokens = total_supervised_tokens
 
-    if cfg.sample_packing:
+    if not skip_estimates and cfg.sample_packing:
         # we have to drop anything longer then sequence len otherwise
         # flash attention with position ids fails
 
@@ -272,10 +286,16 @@ def prepare_optim_env(cfg):
         setup_fsdp_envs(cfg)
     elif cfg.deepspeed:
         os.environ["ACCELERATE_USE_DEEPSPEED"] = "true"
+        os.environ["ACCELERATE_DEEPSPEED_CONFIG_FILE"] = cfg.deepspeed
 
 
 def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps):
-    trainer_builder = HFCausalTrainerBuilder(cfg, model, tokenizer)
+    if cfg.rl:
+        trainer_builder = HFDPOTrainerBuilder(cfg, model[0], tokenizer)
+        trainer_builder.model_ref = model[1]
+    else:
+        trainer_builder = HFCausalTrainerBuilder(cfg, model[0], tokenizer)
+
     trainer_builder.train_dataset = train_dataset
     trainer_builder.eval_dataset = eval_dataset
 

src/axolotl/utils/wandb_.py

@@ -2,20 +2,20 @@
 
 import os
 
+from axolotl.utils.dict import DictDefault
 
-def setup_wandb_env_vars(cfg):
-    if cfg.wandb_mode and cfg.wandb_mode == "offline":
-        os.environ["WANDB_MODE"] = cfg.wandb_mode
-    elif cfg.wandb_project and len(cfg.wandb_project) > 0:
-        os.environ["WANDB_PROJECT"] = cfg.wandb_project
+
+def setup_wandb_env_vars(cfg: DictDefault):
+    for key in cfg.keys():
+        if key.startswith("wandb_"):
+            value = cfg.get(key, "")
+
+            if value and isinstance(value, str) and len(value) > 0:
+                os.environ[key.upper()] = value
+
+    # Enable wandb if project name is present
+    if cfg.wandb_project and len(cfg.wandb_project) > 0:
         cfg.use_wandb = True
-        if cfg.wandb_entity and len(cfg.wandb_entity) > 0:
-            os.environ["WANDB_ENTITY"] = cfg.wandb_entity
-        if cfg.wandb_watch and len(cfg.wandb_watch) > 0:
-            os.environ["WANDB_WATCH"] = cfg.wandb_watch
-        if cfg.wandb_log_model and len(cfg.wandb_log_model) > 0:
-            os.environ["WANDB_LOG_MODEL"] = cfg.wandb_log_model
-        if cfg.wandb_run_id and len(cfg.wandb_run_id) > 0:
-            os.environ["WANDB_RUN_ID"] = cfg.wandb_run_id
+        os.environ.pop("WANDB_DISABLED", None)  # Remove if present
     else:
         os.environ["WANDB_DISABLED"] = "true"

tests/core/test_trainer_builder.py

@@ -0,0 +1,59 @@
+"""
+unit tests for axolotl.core.trainer_builder
+"""
+import pytest
+
+from axolotl.core.trainer_builder import HFDPOTrainerBuilder
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.models import load_model, load_tokenizer
+
+
+@pytest.fixture(name="cfg")
+def fixture_cfg():
+    return DictDefault(
+        {
+            "base_model": "TinyLlama/TinyLlama-1.1B-Chat-v0.6",
+            "model_type": "AutoModelForCausalLM",
+            "tokenizer_type": "LlamaTokenizer",
+            "micro_batch_size": 1,
+            "gradient_accumulation_steps": 1,
+            "learning_rate": 0.00005,
+            "save_steps": 100,
+            "output_dir": "./model-out",
+            "warmup_steps": 10,
+            "gradient_checkpointing": False,
+            "optimizer": "adamw_torch",
+            "sequence_len": 2048,
+            "rl": True,
+            "adam_beta1": 0.998,
+            "adam_beta2": 0.9,
+            "adam_epsilon": 0.00001,
+            "dataloader_num_workers": 1,
+            "dataloader_pin_memory": True,
+        }
+    )
+
+
+@pytest.fixture(name="tokenizer")
+def fixture_tokenizer(cfg):
+    return load_tokenizer(cfg)
+
+
+@pytest.fixture(name="model")
+def fixture_model(cfg, tokenizer):
+    return load_model(cfg, tokenizer)
+
+
+class TestHFDPOTrainerBuilder:
+    """
+    TestCase class for DPO trainer builder
+    """
+
+    def test_build_training_arguments(self, cfg, model, tokenizer):
+        builder = HFDPOTrainerBuilder(cfg, model, tokenizer)
+        training_arguments = builder.build_training_arguments(100)
+        assert training_arguments.adam_beta1 == 0.998
+        assert training_arguments.adam_beta2 == 0.9
+        assert training_arguments.adam_epsilon == 0.00001
+        assert training_arguments.dataloader_num_workers == 1
+        assert training_arguments.dataloader_pin_memory is True

tests/e2e/test_mamba.py

@@ -0,0 +1,65 @@
+"""
+E2E tests for lora llama
+"""
+
+import logging
+import os
+import unittest
+from pathlib import Path
+
+from axolotl.cli import load_datasets
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.train import train
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+
+from .utils import with_temp_dir
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestMistral(unittest.TestCase):
+    """
+    Test case for Llama models using LoRA
+    """
+
+    @with_temp_dir
+    def test_fft(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "state-spaces/mamba-130m",
+                "model_type": "MambaLMHeadModel",
+                "tokenizer_type": "AutoTokenizer",
+                "tokenizer_config": "EleutherAI/gpt-neox-20b",
+                "flash_attention": False,
+                "sequence_len": 1024,
+                "load_in_8bit": False,
+                "val_set_size": 0.0,
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "gradient_checkpointing": False,
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": None,
+                "save_safetensors": False,
+            }
+        )
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "pytorch_model.bin").exists()

tests/e2e/test_mixtral.py

@@ -0,0 +1,109 @@
+"""
+E2E tests for mixtral
+"""
+
+import logging
+import os
+import unittest
+from pathlib import Path
+
+from transformers.utils import is_torch_bf16_gpu_available
+
+from axolotl.cli import load_datasets
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.train import train
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+
+from .utils import with_temp_dir
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestMixtral(unittest.TestCase):
+    """
+    Test case for Llama models using LoRA
+    """
+
+    @with_temp_dir
+    def test_qlora(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "tokenizer_config": "mistralai/Mixtral-8x7B-v0.1",
+                "flash_attention": True,
+                "sequence_len": 1024,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 16,
+                "lora_alpha": 32,
+                "lora_dropout": 0.1,
+                "lora_target_linear": True,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": 10,
+            }
+        )
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "adapter_model.bin").exists()
+
+    @with_temp_dir
+    def test_ft(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "tokenizer_config": "mistralai/Mixtral-8x7B-v0.1",
+                "flash_attention": True,
+                "sequence_len": 1024,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": 10,
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "pytorch_model.bin").exists()

tests/e2e/test_mixtral_samplepack.py

@@ -0,0 +1,123 @@
+"""
+E2E tests for mixtral
+"""
+
+import logging
+import os
+import unittest
+from pathlib import Path
+
+from transformers.utils import is_torch_bf16_gpu_available
+
+from axolotl.cli import load_datasets
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.train import train
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+
+from .utils import with_temp_dir
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestMixtral(unittest.TestCase):
+    """
+    Test case for Llama models using LoRA
+    """
+
+    @with_temp_dir
+    def test_qlora(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "tokenizer_config": "mistralai/Mixtral-8x7B-v0.1",
+                "flash_attention": True,
+                "sequence_len": 2048,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 16,
+                "lora_alpha": 32,
+                "lora_dropout": 0.1,
+                "lora_target_linear": True,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": 10,
+                "sample_packing": True,
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "adapter_model.bin").exists()
+
+    @with_temp_dir
+    def test_ft(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "tokenizer_config": "mistralai/Mixtral-8x7B-v0.1",
+                "flash_attention": True,
+                "sequence_len": 2048,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": 10,
+                "sample_packing": True,
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        model, _ = train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (
+            "axolotl.monkeypatch.mixtral.modeling_mixtral"
+            in model.model.layers[0].self_attn.__class__.__module__
+        )
+        assert (
+            "MixtralMultipackFlashAttention2"
+            in model.model.layers[0].self_attn.__class__.__name__
+        )
+        assert (Path(temp_dir) / "pytorch_model.bin").exists()

tests/e2e/test_model_patches.py

@@ -0,0 +1,99 @@
+"""
+E2E smoke tests to check that the monkeypatches are in place for certain configurations
+"""
+
+import unittest
+
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.models import load_model, load_tokenizer
+
+from .utils import with_temp_dir
+
+
+class TestModelPatches(unittest.TestCase):
+    """
+    TestCases for the multipack monkey patches
+    """
+
+    @with_temp_dir
+    def test_mixtral_multipack(self, temp_dir):
+        cfg = DictDefault(
+            {
+                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "tokenizer_config": "mistralai/Mixtral-8x7B-v0.1",
+                "flash_attention": True,
+                "sample_packing": True,
+                "sequence_len": 2048,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": 10,
+            }
+        )
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        tokenizer = load_tokenizer(cfg)
+        model, _ = load_model(cfg, tokenizer, inference=cli_args.inference)
+
+        assert (
+            "axolotl.monkeypatch.mixtral.modeling_mixtral"
+            in model.model.layers[0].self_attn.__class__.__module__
+        )
+        assert (
+            "MixtralMultipackFlashAttention2"
+            in model.model.layers[0].self_attn.__class__.__name__
+        )
+
+    @with_temp_dir
+    def test_mistral_multipack(self, temp_dir):
+        cfg = DictDefault(
+            {
+                "base_model": "openaccess-ai-collective/tiny-mistral",
+                "flash_attention": True,
+                "sample_packing": True,
+                "sequence_len": 2048,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": 10,
+            }
+        )
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        tokenizer = load_tokenizer(cfg)
+        model, _ = load_model(cfg, tokenizer, inference=cli_args.inference)
+
+        assert (
+            "axolotl.monkeypatch.mistral_attn_hijack_flash"
+            in model.model.layers[0].self_attn.forward.__module__
+        )

tests/test_packed_pretraining.py

@@ -0,0 +1,82 @@
+"""Module for testing streaming dataset sequence packing"""
+import unittest
+from functools import partial
+
+import torch
+from datasets import load_dataset
+from torch.utils.data import DataLoader
+from transformers import AutoTokenizer
+
+from axolotl.utils.collators import PretrainingBatchSamplerDataCollatorForSeq2Seq
+from axolotl.utils.data import encode_packed_pretraining
+
+
+class TestPacking(unittest.TestCase):
+    """
+    Test class for packing streaming dataset sequences
+    """
+
+    def setUp(self) -> None:
+        # pylint: disable=duplicate-code
+        self.tokenizer = AutoTokenizer.from_pretrained("huggyllama/llama-7b")
+        self.tokenizer.pad_token = "</s>"
+        self.max_seq_length = 2048
+        self.batch_size = 2
+
+    def test_packing_stream_dataset(self):
+        # pylint: disable=duplicate-code
+        dataset = load_dataset(
+            "c4",
+            "en",
+            streaming=True,
+        )["train"]
+
+        collate_fn = PretrainingBatchSamplerDataCollatorForSeq2Seq(
+            self.tokenizer,
+            return_tensors="pt",
+            padding=True,
+            pad_to_multiple_of=self.max_seq_length,
+        )
+
+        encode = partial(
+            encode_packed_pretraining,
+            self.tokenizer,
+            collate_fn,
+            max_seq_length=self.max_seq_length,
+            batch_size=self.batch_size,
+        )
+
+        dataset = dataset.map(
+            encode,
+            batched=True,
+            input_columns="text",
+            remove_columns=dataset.features.keys(),
+        )
+
+        trainer_loader = DataLoader(
+            dataset,
+            batch_size=1,
+            collate_fn=None,
+            drop_last=True,
+        )
+        idx = 0
+        for data in trainer_loader:
+            if idx > 10:
+                break
+            assert data["input_ids"].shape == torch.Size(
+                [1, self.batch_size * self.max_seq_length]
+            )
+            assert data["position_ids"].shape == torch.Size(
+                [1, self.batch_size * self.max_seq_length]
+            )
+            assert data["labels"].shape == torch.Size(
+                [1, self.batch_size * self.max_seq_length]
+            )
+            assert data["attention_mask"].shape == torch.Size(
+                [1, self.batch_size * self.max_seq_length]
+            )
+            idx += 1
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_prompt_tokenizers.py

@@ -2,6 +2,7 @@
 import json
 import logging
 import unittest
+from copy import deepcopy
 from pathlib import Path
 from typing import Optional
 
@@ -25,6 +26,50 @@ from axolotl.prompters import AlpacaPrompter, PromptStyle, ShareGPTPrompterV2
 
 LOG = logging.getLogger("axolotl")
 
+test_data = {
+    "multi_turn_sys": {
+        "conversations": [
+            {"from": "system", "value": "lorem"},
+            {"from": "human", "value": "abc"},
+            {"from": "gpt", "value": "ipsum"},
+            {"from": "human", "value": "123"},
+            {"from": "gpt", "value": "sit"},
+        ]
+    },
+    "single_turn_sys": {
+        "conversations": [
+            {"from": "system", "value": "lorem"},
+            {"from": "human", "value": "abc"},
+            {"from": "gpt", "value": "ipsum"},
+        ]
+    },
+    "single_turn_no_sys": {
+        "conversations": [
+            {"from": "human", "value": "abc"},
+            {"from": "gpt", "value": "ipsum"},
+        ]
+    },
+    "multi_turn_no_sys": {
+        "conversations": [
+            {"from": "human", "value": "abc"},
+            {"from": "gpt", "value": "ipsum"},
+            {"from": "human", "value": "123"},
+            {"from": "gpt", "value": "sit"},
+        ]
+    },
+}
+
+
+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):
     """
@@ -114,6 +159,70 @@ class TestPromptTokenizationStrategies(unittest.TestCase):
                 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, 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, 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"],

tests/test_tokenizers.py

@@ -3,6 +3,8 @@ Test cases for the tokenizer loading
 """
 import unittest
 
+import pytest
+
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_tokenizer
 
@@ -31,6 +33,40 @@ class TestTokenizers(unittest.TestCase):
         tokenizer = load_tokenizer(cfg)
         assert "Fast" not in tokenizer.__class__.__name__
 
+    def test_special_tokens_modules_to_save(self):
+        # setting special_tokens to new token
+        cfg = DictDefault(
+            {
+                "tokenizer_config": "huggyllama/llama-7b",
+                "adapter": "lora",
+                "special_tokens": {"bos_token": "[INST]"},
+            }
+        )
+        with pytest.raises(
+            ValueError,
+            match=r".*Please set lora_modules_to_save*",
+        ):
+            load_tokenizer(cfg)
+
+        # setting special_tokens but not changing from default
+        cfg = DictDefault(
+            {
+                "tokenizer_config": "huggyllama/llama-7b",
+                "adapter": "lora",
+                "special_tokens": {"bos_token": "<s>"},
+            }
+        )
+        load_tokenizer(cfg)
+
+        # non-adapter setting special_tokens
+        cfg = DictDefault(
+            {
+                "tokenizer_config": "huggyllama/llama-7b",
+                "special_tokens": {"bos_token": "[INST]"},
+            }
+        )
+        load_tokenizer(cfg)
+
 
 if __name__ == "__main__":
     unittest.main()

tests/test_validation.py

@@ -1,6 +1,7 @@
 """Module for testing the validation module"""
 
 import logging
+import os
 import unittest
 from typing import Optional
 
@@ -8,6 +9,7 @@ import pytest
 
 from axolotl.utils.config import validate_config
 from axolotl.utils.dict import DictDefault
+from axolotl.utils.wandb_ import setup_wandb_env_vars
 
 
 class ValidationTest(unittest.TestCase):
@@ -679,3 +681,120 @@ class ValidationTest(unittest.TestCase):
         )
 
         validate_config(cfg)
+
+    def test_add_tokens_adapter(self):
+        cfg = DictDefault(
+            {"adapter": "qlora", "load_in_4bit": True, "tokens": ["<|imstart|>"]}
+        )
+
+        with pytest.raises(
+            ValueError,
+            match=r".*lora_modules_to_save not properly set yet adding new tokens*",
+        ):
+            validate_config(cfg)
+
+        cfg = DictDefault(
+            {
+                "adapter": "qlora",
+                "load_in_4bit": True,
+                "tokens": ["<|imstart|>"],
+                "lora_modules_to_save": ["embed_tokens"],
+            }
+        )
+
+        with pytest.raises(
+            ValueError,
+            match=r".*lora_modules_to_save not properly set yet adding new tokens*",
+        ):
+            validate_config(cfg)
+
+        cfg = DictDefault(
+            {
+                "adapter": "qlora",
+                "load_in_4bit": True,
+                "tokens": ["<|imstart|>"],
+                "lora_modules_to_save": ["embed_tokens", "lm_head"],
+            }
+        )
+
+        validate_config(cfg)
+
+
+class ValidationWandbTest(ValidationTest):
+    """
+    Validation test for wandb
+    """
+
+    def test_wandb_set_run_id_to_name(self):
+        cfg = DictDefault(
+            {
+                "wandb_run_id": "foo",
+            }
+        )
+
+        with self._caplog.at_level(logging.WARNING):
+            validate_config(cfg)
+            assert any(
+                "wandb_run_id sets the ID of the run. If you would like to set the name, please use wandb_name instead."
+                in record.message
+                for record in self._caplog.records
+            )
+
+            assert cfg.wandb_name == "foo" and cfg.wandb_run_id == "foo"
+
+        cfg = DictDefault(
+            {
+                "wandb_name": "foo",
+            }
+        )
+
+        validate_config(cfg)
+
+        assert cfg.wandb_name == "foo" and cfg.wandb_run_id is None
+
+    def test_wandb_sets_env(self):
+        cfg = DictDefault(
+            {
+                "wandb_project": "foo",
+                "wandb_name": "bar",
+                "wandb_run_id": "bat",
+                "wandb_entity": "baz",
+                "wandb_mode": "online",
+                "wandb_watch": "false",
+                "wandb_log_model": "checkpoint",
+            }
+        )
+
+        validate_config(cfg)
+
+        setup_wandb_env_vars(cfg)
+
+        assert os.environ.get("WANDB_PROJECT", "") == "foo"
+        assert os.environ.get("WANDB_NAME", "") == "bar"
+        assert os.environ.get("WANDB_RUN_ID", "") == "bat"
+        assert os.environ.get("WANDB_ENTITY", "") == "baz"
+        assert os.environ.get("WANDB_MODE", "") == "online"
+        assert os.environ.get("WANDB_WATCH", "") == "false"
+        assert os.environ.get("WANDB_LOG_MODEL", "") == "checkpoint"
+        assert os.environ.get("WANDB_DISABLED", "") != "true"
+
+    def test_wandb_set_disabled(self):
+        cfg = DictDefault({})
+
+        validate_config(cfg)
+
+        setup_wandb_env_vars(cfg)
+
+        assert os.environ.get("WANDB_DISABLED", "") == "true"
+
+        cfg = DictDefault(
+            {
+                "wandb_project": "foo",
+            }
+        )
+
+        validate_config(cfg)
+
+        setup_wandb_env_vars(cfg)
+
+        assert os.environ.get("WANDB_DISABLED", "") != "true"