fix: remove excessive newlines in system prompt(s) for alpaca (#936)

* new evals_per_epoch and saves_per_epoch to make things cleaner

* update per PR feedback

.github/workflows/tests.yml

@@ -71,8 +71,9 @@ jobs:
 
       - name: Install dependencies
         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

@@ -25,8 +25,10 @@ Features:
 - [Installation](#installation)
   - [Docker](#docker)
   - [Conda/Pip venv](#condapip-venv)
+  - [Runpod](#runpod)
   - [LambdaLabs](#lambdalabs)
   - [Windows](#windows)
+  - [Launching on public clouds via SkyPilot](#launching-on-public-clouds-via-skypilot)
 - [Dataset](#dataset)
   - [How to Add Custom Prompts](#how-to-add-custom-prompts)
   - [How to Use Custom Pretokenized Dataset](#how-to-use-your-custom-pretokenized-dataset)
@@ -63,18 +65,21 @@ 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     | ✅         | ❓    | ❓     | ❓             | ❓                 | ❓          | ❓            |
+|             | 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 ⚡
@@ -83,14 +88,19 @@ Get started with Axolotl in just a few steps! This quickstart guide will walk yo
 
 **Requirements**: Python >=3.9 and Pytorch >=2.0.
 
+`pip3 install "axolotl[flash-attn,deepspeed] @ git+https://github.com/OpenAccess-AI-Collective/axolotl"`
+
+### For developers
 ```bash
 git clone https://github.com/OpenAccess-AI-Collective/axolotl
 cd axolotl
 
 pip3 install packaging
 pip3 install -e '.[flash-attn,deepspeed]'
-pip3 install -U git+https://github.com/huggingface/peft.git
+```
 
+### Usage
+```bash
 # finetune lora
 accelerate launch -m axolotl.cli.train examples/openllama-3b/lora.yml
 
@@ -111,7 +121,6 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
   ```bash
   docker run --gpus '"all"' --rm -it winglian/axolotl:main-py3.10-cu118-2.0.1
   ```
-  - `winglian/axolotl-runpod:main-latest`: for runpod or use this [direct link](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
 
   Or run on the current files for development:
 
@@ -126,13 +135,15 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
   A more powerful Docker command to run would be this:
 
   ```bash
-  docker run --gpus '"all"' --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=volume,src=axolotl,target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-py3.10-cu118-2.0.1
+  docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=volume,src=axolotl,target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-py3.10-cu118-2.0.1
   ```
 
   It additionally:
   * Prevents memory issues when running e.g. deepspeed (e.g. you could hit SIGBUS/signal 7 error) through `--ipc` and `--ulimit` args.
   * Persists the downloaded HF data (models etc.) and your modifications to axolotl code through `--mount`/`-v` args.
   * The `--name` argument simply makes it easier to refer to the container in vscode (`Dev Containers: Attach to Running Container...`) or in your terminal.
+  * The `--privileged` flag gives all capabilities to the container.
+  * The `--shm-size 10g` argument increases the shared memory size. Use this if you see `exitcode: -7` errors using deepspeed.
 
   [More information on nvidia website](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html#setincshmem)
 
@@ -154,6 +165,10 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
         ```
         Get the token at huggingface.co/settings/tokens
 
+#### Runpod
+
+Use `winglian/axolotl-runpod:main-latest` or use this [direct link](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
+
 #### LambdaLabs
   <details>
 
@@ -201,6 +216,28 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
 #### Windows
 Please use WSL or Docker!
 
+
+#### Launching on public clouds via SkyPilot
+To launch on GPU instances (both on-demand and spot instances) on 7+ clouds (GCP, AWS, Azure, OCI, and more), you can use [SkyPilot](https://skypilot.readthedocs.io/en/latest/index.html):
+```bash
+pip install "skypilot-nightly[gcp,aws,azure,oci,lambda,kubernetes,ibm,scp]"  # choose your clouds
+sky check
+```
+Get the [example YAMLs](https://github.com/skypilot-org/skypilot/tree/master/llm/axolotl) of using Axolotl to finetune `mistralai/Mistral-7B-v0.1`:
+```
+git clone https://github.com/skypilot-org/skypilot.git
+cd skypilot/llm/axolotl
+```
+Use one command to launch:
+```bash
+# On-demand
+HF_TOKEN=xx sky launch axolotl.yaml --env HF_TOKEN
+
+# Managed spot (auto-recovery on preemption)
+HF_TOKEN=xx BUCKET=<unique-name> sky spot launch axolotl-spot.yaml --env HF_TOKEN --env BUCKET
+```
+
+
 ### Dataset
 
 Axolotl supports a variety of dataset formats. Below are some of the formats you can use.
@@ -210,7 +247,7 @@ 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": "..."}]}
   ```
@@ -397,6 +434,12 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
     - path: knowrohit07/know_sql
       type: context_qa.load_v2
       train_on_split: validation
+
+  # loading from s3 or gcs
+  # s3 creds will be loaded from the system default and gcs only supports public access
+  dataset:
+    - path: s3://path_to_ds # Accepts folder with arrow/parquet or file path like above. Supports s3, gcs.
+      ...
   ```
 
 - loading
@@ -459,6 +502,15 @@ is_falcon_derived_model:
 is_llama_derived_model:
 # Please note that if you set this to true, `padding_side` will be set to "left" by default
 is_mistral_derived_model:
+is_qwen_derived_model:
+
+# optional overrides to the base model configuration
+model_config:
+  # RoPE Scaling https://github.com/huggingface/transformers/pull/24653
+  rope_scaling:
+    type: # linear | dynamic
+    factor: # float
+
 
 # Whether you are training a 4-bit GPTQ quantized model
 gptq: true
@@ -483,7 +535,7 @@ float16: true
 
 # A list of one or more datasets to finetune the model with
 datasets:
-  # HuggingFace dataset repo | "json" for local dataset, make sure to fill data_files
+  # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
   - path: vicgalle/alpaca-gpt4
   # The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection]
     type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
@@ -491,9 +543,12 @@ datasets:
     data_files: # Optional[str] path to source data files
     shards: # Optional[int] number of shards to split data into
     name: # Optional[str] name of dataset configuration to load
+    train_on_split: train # Optional[str] name of dataset split to load from
 
     # Optional[str] fastchat conversation type, only used with type: sharegpt
     conversation:  # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+    field_human: # Optional[str]. Human key to use for conversation.
+    field_model: # Optional[str]. Assistant key to use for conversation.
 
   # Custom user prompt
   - path: repo
@@ -559,6 +614,12 @@ 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.
@@ -606,7 +667,8 @@ wandb_mode: # "offline" to save run metadata locally and not sync to the server,
 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
@@ -624,13 +686,16 @@ gradient_accumulation_steps: 1
 micro_batch_size: 2
 eval_batch_size:
 num_epochs: 4
-warmup_steps: 100
+warmup_steps: 100  # cannot use with warmup_ratio
+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.
@@ -640,6 +705,9 @@ 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:
 
@@ -726,10 +794,6 @@ 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:
-# RoPE Scaling https://github.com/huggingface/transformers/pull/24653
-rope_scaling:
-  type: # linear | dynamic
-  factor: # float
 
 # Resume from a specific checkpoint dir
 resume_from_checkpoint:
@@ -902,7 +966,7 @@ wandb_mode:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 ```
 
@@ -958,6 +1022,10 @@ 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.

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",

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
 

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/llama-2/tiny-llama.yml

@@ -1,4 +1,4 @@
-base_model: PY007/TinyLlama-1.1B-step-50K-105b
+base_model: PY007/TinyLlama-1.1B-intermediate-step-715k-1.5T
 
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
@@ -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
@@ -54,9 +54,9 @@ 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/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

@@ -21,7 +21,7 @@ pad_to_sequence_len: true
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -46,10 +46,10 @@ 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 +1,79 @@
+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
+
+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

@@ -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

@@ -1,5 +1,5 @@
 base_model: microsoft/phi-1_5
-model_type: MixFormerSequentialForCausalLM
+model_type: PhiForCausalLM
 tokenizer_type: AutoTokenizer
 is_llama_derived_model: false
 trust_remote_code: true
@@ -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

@@ -0,0 +1,68 @@
+base_model: Qwen/Qwen-7B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+is_qwen_derived_model: true
+trust_remote_code: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./lora-out
+
+sequence_len: 2048  # supports up to 8192
+sample_packing: false
+pad_to_sequence_len:
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+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:

examples/qwen/qlora.yml

@@ -0,0 +1,68 @@
+base_model: Qwen/Qwen-7B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+is_qwen_derived_model: true
+trust_remote_code: 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: ./lora-out
+
+sequence_len: 2048  # supports up to 8192
+sample_packing: false
+pad_to_sequence_len:
+
+adapter: qlora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+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:

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/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

requirements.txt

@@ -1,22 +1,21 @@
---extra-index-url https://download.pytorch.org/whl/cu118
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
-torch==2.0.1
-auto-gptq==0.4.2
+auto-gptq==0.5.1
 packaging
 peft==0.6.0
-transformers @ git+https://github.com/huggingface/transformers.git@acc394c4f5e1283c19783581790b3dc3105a3697
+transformers @ git+https://github.com/huggingface/transformers.git@e5079b0b2abcef11ecbdae60ba4a6636c57b725d
+tokenizers==0.15.0
 bitsandbytes>=0.41.1
-accelerate @ git+https://github.com/huggingface/accelerate@80da9cfb09bb3cc9f1b385cb55d6b90d025a5fd9
+accelerate==0.24.1
 deepspeed
 addict
 fire
 PyYAML>=6.0
-datasets
-flash-attn>=2.3.0
+datasets>=2.15.0
+flash-attn==2.3.3
 sentencepiece
 wandb
 einops
-xformers>=0.0.22
+xformers==0.0.22
 optimum==1.13.2
 hf_transfer
 colorama
@@ -30,5 +29,11 @@ scipy
 scikit-learn==1.2.2
 pynvml
 art
-fschat==0.2.29
-gradio
+fschat==0.2.34
+gradio==3.50.2
+tensorboard
+
+# remote filesystems
+s3fs
+gcsfs
+# adlfs

setup.py

@@ -46,10 +46,13 @@ 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

@@ -29,6 +29,7 @@ from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process
 from axolotl.utils.models import load_tokenizer
 from axolotl.utils.tokenization import check_dataset_labels
+from axolotl.utils.trainer import prepare_optim_env
 from axolotl.utils.wandb_ import setup_wandb_env_vars
 
 project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
@@ -46,7 +47,7 @@ def print_axolotl_text_art(suffix=None):
     ascii_text = "  axolotl"
     if suffix:
         ascii_text += f"  x  {suffix}"
-    ascii_art = text2art(" axolotl", font=font)
+    ascii_art = text2art(ascii_text, font=font)
 
     if is_main_process():
         print(ascii_art)
@@ -71,7 +72,7 @@ def do_merge_lora(
 
     LOG.info("running merge of LoRA with base model")
     model = model.merge_and_unload()
-    model.to(dtype=torch.float16)
+    model.to(dtype=cfg.torch_dtype)
 
     if cfg.local_rank == 0:
         LOG.info(f"saving merged model to: {str(Path(cfg.output_dir) / 'merged')}")
@@ -296,6 +297,8 @@ def load_cfg(config: Path = Path("examples/"), **kwargs):
 
     validate_config(cfg)
 
+    prepare_optim_env(cfg)
+
     normalize_config(cfg)
 
     setup_wandb_env_vars(cfg)

src/axolotl/core/trainer_builder.py

@@ -11,7 +11,7 @@ from abc import abstractmethod
 from dataclasses import dataclass, field
 from functools import partial
 from pathlib import Path
-from typing import Optional, Union
+from typing import Optional
 
 import torch
 import transformers
@@ -25,13 +25,16 @@ 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.dataloader import MultipackDistributedDataloader
+from axolotl.utils.collators import (
+    BatchSamplerDataCollatorForSeq2Seq,
+    MambaDataCollator,
+)
 from axolotl.utils.samplers import MultipackBatchSampler
 from axolotl.utils.schedulers import get_cosine_schedule_with_quadratic_warmup
 
@@ -49,6 +52,9 @@ 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."},
@@ -215,9 +221,7 @@ class AxolotlTrainer(Trainer):
             )
         return super().get_train_dataloader()
 
-    def get_eval_dataloader(
-        self, eval_dataset: Optional[Dataset] = None
-    ) -> Union[DataLoader, MultipackDistributedDataloader]:
+    def get_eval_dataloader(self, eval_dataset: Optional[Dataset] = None) -> DataLoader:
         if self.args.sample_packing and self.args.eval_sample_packing is not False:
             eval_dataset = (
                 eval_dataset if eval_dataset is not None else self.eval_dataset
@@ -260,7 +264,7 @@ class AxolotlTrainer(Trainer):
     def get_bench_dataloader(
         self,
         bench_dataset: Dataset,
-    ) -> Union[DataLoader, MultipackDistributedDataloader]:
+    ) -> DataLoader:
         dataloader_params = {
             "batch_size": self.args.eval_batch_size,
             "collate_fn": self.bench_data_collator,
@@ -287,6 +291,32 @@ class AxolotlTrainer(Trainer):
         return super().compute_loss(model, inputs, return_outputs=return_outputs)
 
 
+class AxolotlMambaTrainer(AxolotlTrainer):
+    """
+    Mamba specific trainer to handle loss calculation
+    """
+
+    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
@@ -433,6 +463,9 @@ 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):
@@ -461,14 +494,19 @@ 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):
-        warmup_steps = (
-            self.cfg.warmup_steps
-            if self.cfg.warmup_steps is not None
-            else min(int(0.03 * total_num_steps), 100)
-        )
+        warmup_steps = None
+        if self.cfg.warmup_steps is not None:
+            warmup_steps = self.cfg.warmup_steps
+        elif self.cfg.warmup_ratio is not None:
+            warmup_steps = max(int(self.cfg.warmup_ratio * total_num_steps), 0)
+        else:
+            warmup_steps = min(int(0.03 * total_num_steps), 100)
+
         logging_steps = (
             self.cfg.logging_steps
             if self.cfg.logging_steps is not None
@@ -483,10 +521,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["fp16"] = (
             self.cfg.fp16 and not self.cfg.bf16
         ) or False
-        if self.cfg.fp8:
-            training_arguments_kwargs["fp16"] = False
-            training_arguments_kwargs["bf16"] = False
-
         training_arguments_kwargs["tf32"] = self.cfg.tf32
         training_arguments_kwargs["warmup_steps"] = warmup_steps
         training_arguments_kwargs["logging_steps"] = logging_steps
@@ -529,7 +563,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             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:
@@ -550,16 +584,16 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 "dataloader_prefetch_factor"
             ] = self.cfg.dataloader_prefetch_factor
 
-        if self.cfg.eval_steps:
+        if self.cfg.val_set_size == 0:
+            # no eval set, so don't eval
+            training_arguments_kwargs["evaluation_strategy"] = "no"
+        elif self.cfg.eval_steps:
             training_arguments_kwargs["evaluation_strategy"] = "steps"
             training_arguments_kwargs["eval_steps"] = self.cfg.eval_steps
         elif self.cfg.evaluation_strategy:
             training_arguments_kwargs[
                 "evaluation_strategy"
             ] = self.cfg.evaluation_strategy
-        elif self.cfg.val_set_size == 0:
-            # no eval set, so don't eval
-            training_arguments_kwargs["evaluation_strategy"] = "no"
         else:
             # we have an eval set, but no steps defined, default to use epoch
             training_arguments_kwargs["evaluation_strategy"] = "epoch"
@@ -647,7 +681,7 @@ 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"
@@ -665,7 +699,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             self.cfg.sample_packing if self.cfg.sample_packing else False
         )
         training_arguments_kwargs["eval_sample_packing"] = (
-            self.cfg.sample_packing if self.cfg.sample_packing else False
+            self.cfg.sample_packing
+            if self.cfg.eval_sample_packing is not False
+            else False
         )
         training_arguments_kwargs[
             "sample_packing_seq_len_multiplier"
@@ -675,6 +711,7 @@ 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_args = (
             AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
                 **training_arguments_kwargs,
@@ -729,11 +766,7 @@ 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(**data_collator_kwargs),
             bench_data_collator=transformers.DataCollatorForSeq2Seq(
                 self.tokenizer,
                 return_tensors="pt",
@@ -753,3 +786,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             ] = self.cfg.micro_batch_size
 
         return trainer
+
+    def build_collator(self, **kwargs):
+        if self.cfg.model_config_type == "mamba":
+            return MambaDataCollator(tokenizer=self.tokenizer)
+
+        return BatchSamplerDataCollatorForSeq2Seq(
+            self.tokenizer,
+            return_tensors="pt",
+            **kwargs,
+        )

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/models/phi/__init__.py

@@ -3,4 +3,6 @@ MixFormers model architecture used for phi models
 """
 
 from .configuration_mixformer_sequential import MixFormerSequentialConfig  # noqa
+from .configuration_phi import PhiConfig  # noqa
 from .modeling_mixformer_sequential import MixFormerSequentialForCausalLM  # noqa
+from .modeling_phi import PhiForCausalLM  # noqa

src/axolotl/models/phi/configuration_phi.py

@@ -0,0 +1,65 @@
+# pylint: skip-file
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import math
+from typing import Optional
+
+from transformers import PretrainedConfig
+
+
+class PhiConfig(PretrainedConfig):
+    """Phi configuration."""
+
+    model_type = "phi"
+    attribute_map = {
+        "max_position_embeddings": "n_positions",
+        "hidden_size": "n_embd",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+    }
+
+    def __init__(
+        self,
+        vocab_size: int = 50304,
+        n_positions: int = 2048,
+        n_embd: int = 1024,
+        n_layer: int = 20,
+        n_inner: Optional[int] = None,
+        n_head: int = 16,
+        n_head_kv: Optional[int] = None,
+        rotary_dim: Optional[int] = 32,
+        activation_function: Optional[str] = "gelu_new",
+        flash_attn: bool = False,
+        flash_rotary: bool = False,
+        fused_dense: bool = False,
+        attn_pdrop: float = 0.0,
+        embd_pdrop: float = 0.0,
+        resid_pdrop: float = 0.0,
+        layer_norm_epsilon: float = 1e-5,
+        initializer_range: float = 0.02,
+        tie_word_embeddings: bool = False,
+        pad_vocab_size_multiple: int = 64,
+        **kwargs
+    ) -> None:
+        self.vocab_size = int(
+            math.ceil(vocab_size / pad_vocab_size_multiple) * pad_vocab_size_multiple
+        )
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_inner = n_inner
+        self.n_head = n_head
+        self.n_head_kv = n_head_kv
+        self.rotary_dim = min(rotary_dim, n_embd // n_head)
+        self.activation_function = activation_function
+        self.flash_attn = flash_attn
+        self.flash_rotary = flash_rotary
+        self.fused_dense = fused_dense
+        self.attn_pdrop = attn_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.resid_pdrop = resid_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -321,6 +321,8 @@ def flashattn_forward(
         # only on first autoregressive step q,k,v have same seqlen
         is_causal = key_states.shape == query_states.shape
 
+    dropout_rate = 0.0 if not self.training else getattr(self, "attention_dropout", 0.0)
+
     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(
@@ -330,7 +332,12 @@ def flashattn_forward(
         qkv = rearrange(qkv, "b s ... -> (b s) ...")
 
         output = flash_attn_varlen_qkvpacked_func(
-            qkv, cu_seqlens, max_seqlen, 0.0, softmax_scale=None, causal=True
+            qkv,
+            cu_seqlens,
+            max_seqlen,
+            dropout_p=dropout_rate,
+            softmax_scale=None,
+            causal=True,
         )
         output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
     elif query_states.shape == key_states.shape:
@@ -353,7 +360,7 @@ def flashattn_forward(
             qkv_unpad,
             cu_seqlens_q,
             max_seqlen_q,
-            0.0,
+            dropout_p=dropout_rate,
             softmax_scale=None,
             causal=is_causal,
         )
@@ -366,6 +373,7 @@ def flashattn_forward(
             output = flash_attn_kvpacked_func(
                 query_states,
                 torch.stack([key_states, value_states], 2),
+                dropout_p=dropout_rate,
                 causal=is_causal,
             )
         else:
@@ -398,7 +406,7 @@ def flashattn_forward(
                 cu_seqlens_k,
                 max_seqlen_q,
                 max_seqlen_k,
-                0.0,
+                dropout_p=dropout_rate,
                 softmax_scale=None,
                 causal=is_causal,
             )

src/axolotl/monkeypatch/llama_attn_hijack_sdp.py

@@ -25,6 +25,8 @@ def sdp_attention_forward(
     past_key_value: Optional[Tuple[torch.Tensor]] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
+    padding_mask: Optional[torch.LongTensor] = None,  # pylint: disable=unused-argument
+    **kwargs,  # pylint: disable=unused-argument
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
     # pylint: disable=duplicate-code
     bsz, q_len, _ = hidden_states.size()

src/axolotl/monkeypatch/llama_attn_hijack_xformers.py

@@ -29,6 +29,8 @@ def xformers_forward(
     past_key_value: Optional[Tuple[torch.Tensor]] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
+    padding_mask: Optional[torch.LongTensor] = None,  # pylint: disable=unused-argument
+    **kwargs,  # pylint: disable=unused-argument
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
     # pylint: disable=duplicate-code
     bsz, q_len, _ = hidden_states.size()

src/axolotl/monkeypatch/mistral_attn_hijack_flash.py

@@ -201,6 +201,8 @@ def flashattn_forward(
         # only on first autoregressive step q,k,v have same seqlen
         is_causal = key_states.shape == query_states.shape
 
+    dropout_rate = 0.0 if not self.training else getattr(self, "attention_dropout", 0.0)
+
     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(
@@ -213,7 +215,7 @@ def flashattn_forward(
             qkv,
             cu_seqlens,
             max_seqlen,
-            0.0,
+            dropout_p=dropout_rate,
             softmax_scale=None,
             causal=True,
             window_size=window_size,
@@ -239,7 +241,7 @@ def flashattn_forward(
             qkv_unpad,
             cu_seqlens_q,
             max_seqlen_q,
-            0.0,
+            dropout_p=dropout_rate,
             softmax_scale=None,
             causal=is_causal,
             window_size=window_size,
@@ -253,6 +255,7 @@ def flashattn_forward(
             output = flash_attn_kvpacked_func(
                 query_states,
                 torch.stack([key_states, value_states], 2),
+                dropout_p=dropout_rate,
                 causal=is_causal,
                 window_size=window_size,
             )
@@ -286,7 +289,7 @@ def flashattn_forward(
                 cu_seqlens_k,
                 max_seqlen_q,
                 max_seqlen_k,
-                0.0,
+                dropout_p=dropout_rate,
                 softmax_scale=None,
                 causal=is_causal,
                 window_size=window_size,

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.MISTRAL_ATTENTION_CLASSES[
+        "flash_attention_2"
+    ] = MixtralMultipackFlashAttention2

src/axolotl/monkeypatch/mixtral/modeling_mixtral.py

@@ -0,0 +1,379 @@
+"""
+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._use_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._use_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/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|>",
     )
 )
 

src/axolotl/prompters.py

@@ -22,13 +22,19 @@ class PromptStyle(Enum):
     CHATML = "chatml"
 
 
-class AlpacaPrompter:
+class Prompter:
+    """
+    Base prompter class for all prompters
+    """
+
+
+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
@@ -69,7 +75,7 @@ class AlpacaPrompter:
         else:
             res = (
                 self.system_format.format(system=self.system_no_input_prompt)
-                if self.system_prompt
+                if self.system_no_input_prompt
                 else ""
             ) + self.turn_no_input_format.format(instruction=instruction)
         if output:
@@ -159,7 +165,7 @@ class NomicGPT4AllPrompter(AlpacaPrompter):
     """
 
 
-class ReflectAlpacaPrompter:
+class ReflectAlpacaPrompter(Prompter):
     """
     Prompter for ReflectAlpaca
     """
@@ -254,7 +260,7 @@ SHAREGPT_ASSERTION_FAILED_ROLE = (
 )
 
 
-class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
+class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
     """
     A prompter that generates prompts for the ShareGPT
     """
@@ -349,7 +355,7 @@ class ShareGPTPrompterV2(ShareGPTPrompter):
         )
 
 
-class UnsupportedPrompter:
+class UnsupportedPrompter(Prompter):
     """
     A dummy class for custom prompters
     """

src/axolotl/train.py

@@ -82,7 +82,8 @@ def train(
         cfg, train_dataset, eval_dataset, model, 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 +93,8 @@ 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:

src/axolotl/utils/callbacks.py

@@ -124,6 +124,36 @@ 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 = [

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 +150,31 @@ 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,
+        }

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()}
@@ -70,15 +70,22 @@ def normalize_config(cfg):
     else:
         torch.backends.cuda.matmul.allow_tf32 = cfg.tf32 or False
 
-    if cfg.fp8:
-        cfg.torch_dtype = torch.bfloat16
-    elif cfg.bf16 or cfg.bfloat16:
+    if cfg.bf16 or cfg.bfloat16:
         cfg.torch_dtype = torch.bfloat16
     elif cfg.load_in_8bit or cfg.fp16 or cfg.float16:
         cfg.torch_dtype = torch.float16
     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:
@@ -124,6 +131,19 @@ def normalize_config(cfg):
         or (cfg.model_type and "mistral" in cfg.model_type.lower())
     )
 
+    cfg.is_qwen_derived_model = (
+        (
+            hasattr(model_config, "model_type")
+            and model_config.model_type
+            in [
+                "qwen",
+            ]
+        )
+        or cfg.is_qwen_derived_model
+        or "qwen" in cfg.base_model.lower()
+        or (cfg.model_type and "qwen" in cfg.model_type.lower())
+    )
+
     if isinstance(cfg.learning_rate, str):
         cfg.learning_rate = float(cfg.learning_rate)
 
@@ -167,7 +187,11 @@ def validate_config(cfg):
             "batch_size is not recommended. Please use gradient_accumulation_steps instead.",
             "To calculate the equivalent gradient_accumulation_steps, divide batch_size / micro_batch_size / number of gpus.",
         )
-    if cfg.eval_batch_size != cfg.micro_batch_size:
+    if (
+        cfg.eval_batch_size
+        and cfg.micro_batch_size
+        and cfg.eval_batch_size != cfg.micro_batch_size
+    ):
         LOG.warning(
             "eval_batch_size != micro_batch_size. This can lead to VRAM instability."
         )
@@ -337,6 +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."
@@ -371,6 +416,24 @@ def validate_config(cfg):
             "If you want to full finetune, please turn off load_in_8bit and load_in_4bit."
         )
 
+    if cfg.rope_scaling:
+        LOG.warning("`rope_scaling` should now be be a key under `model_config`")
+
+    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:
+        LOG.warning(
+            "Gradient checkpointing is broken for Qwen models for transformers>=4.35.0, except main branch."
+        )
+
+    if cfg.wandb_run_id and not cfg.wandb_name:
+        cfg.wandb_name = cfg.wandb_run_id
+
+        LOG.warning(
+            "wandb_run_id sets the ID of the run. If you would like to set the name, please use wandb_name instead."
+        )
+
     # TODO
     # MPT 7b
     # https://github.com/facebookresearch/bitsandbytes/issues/25

src/axolotl/utils/data.py

@@ -3,7 +3,7 @@ import functools
 import hashlib
 import logging
 from pathlib import Path
-from typing import Any, Dict, List, Tuple, Union
+from typing import Dict, List, Tuple, Union
 
 import torch
 from datasets import (
@@ -34,6 +34,7 @@ from axolotl.prompters import (
     JeopardyPrompter,
     MultipleChoiceConcisePrompter,
     MultipleChoiceExplainPrompter,
+    Prompter,
     ReflectAlpacaPrompter,
     SummarizeTLDRPrompter,
     UnsupportedPrompter,
@@ -78,6 +79,14 @@ def prepare_dataset(cfg, tokenizer):
         train_dataset, eval_dataset = process_datasets_for_packing(
             cfg, train_dataset, eval_dataset, tokenizer
         )
+
+    if eval_dataset and cfg.sample_packing and cfg.eval_sample_packing is not False:
+        total_eval_steps = calculate_total_num_steps(cfg, eval_dataset, update=False)
+        if total_eval_steps == 0:
+            raise ValueError(
+                "eval dataset split is too small for sample_packing. You should set `eval_sample_packing: False`. "
+            )
+
     if cfg.max_steps:
         total_num_steps = min(
             calculate_total_num_steps(cfg, train_dataset), cfg.max_steps
@@ -90,7 +99,7 @@ def prepare_dataset(cfg, tokenizer):
 
 def load_tokenized_prepared_datasets(
     tokenizer, cfg, default_dataset_prepared_path
-) -> DatasetDict:
+) -> Tuple[DatasetDict, List[Prompter]]:
     tokenizer_name = tokenizer.__class__.__name__
     ds_hash = str(
         md5(
@@ -98,7 +107,12 @@ def load_tokenized_prepared_datasets(
                 str(cfg.sequence_len)
                 + "@"
                 + "|".join(
-                    sorted([f"{d.path}:{d.type}:{d.shards}" for d in cfg.datasets])
+                    sorted(
+                        [
+                            f"{d.path}:{d.type}:{d.shards}:{d.conversation}"
+                            for d in cfg.datasets
+                        ]
+                    )
                 )
                 + "|"
                 + tokenizer_name
@@ -164,6 +178,66 @@ def load_tokenized_prepared_datasets(
             except (FileNotFoundError, ConnectionError):
                 pass
 
+            ds_from_cloud = False
+            storage_options = {}
+            remote_file_system = None
+            if config_dataset.path.startswith("s3://"):
+                try:
+                    import aiobotocore.session  # type: ignore
+                    import s3fs  # type: ignore
+                except ImportError as exc:
+                    raise ImportError(
+                        "s3:// paths require aiobotocore and s3fs to be installed"
+                    ) from exc
+
+                # Takes credentials from ~/.aws/credentials for default profile
+                s3_session = aiobotocore.session.AioSession(profile="default")
+                storage_options = {"session": s3_session}
+                remote_file_system = s3fs.S3FileSystem(**storage_options)
+            elif config_dataset.path.startswith(
+                "gs://"
+            ) or config_dataset.path.startswith("gcs://"):
+                try:
+                    import gcsfs  # type: ignore
+                except ImportError as exc:
+                    raise ImportError(
+                        "gs:// or gcs:// paths require gcsfs to be installed"
+                    ) from exc
+
+                # gcsfs will use default credentials from the environment else anon
+                # https://gcsfs.readthedocs.io/en/latest/#credentials
+                storage_options = {"token": None}
+                remote_file_system = gcsfs.GCSFileSystem(**storage_options)
+            # TODO: Figure out how to get auth creds passed
+            # elif config_dataset.path.startswith("adl://") or config_dataset.path.startswith("abfs://"):
+            #     try:
+            #         import adlfs
+            #     except ImportError as exc:
+            #        raise ImportError(
+            #            "adl:// or abfs:// paths require adlfs to be installed"
+            #        ) from exc
+
+            #     # Gen 1
+            #     storage_options = {
+            #         "tenant_id": TENANT_ID,
+            #         "client_id": CLIENT_ID,
+            #         "client_secret": CLIENT_SECRET,
+            #     }
+            #     # Gen 2
+            #     storage_options = {
+            #         "account_name": ACCOUNT_NAME,
+            #         "account_key": ACCOUNT_KEY,
+            #     }
+
+            #     remote_file_system = adlfs.AzureBlobFileSystem(**storage_options)
+            try:
+                if remote_file_system and remote_file_system.exists(
+                    config_dataset.path
+                ):
+                    ds_from_cloud = True
+            except (FileNotFoundError, ConnectionError):
+                pass
+
             # prefer local dataset, even if hub exists
             local_path = Path(config_dataset.path)
             if local_path.exists():
@@ -177,17 +251,8 @@ def load_tokenized_prepared_datasets(
                         split=None,
                     )
                 elif local_path.is_file():
-                    ds_type = "json"
-                    if config_dataset.ds_type:
-                        ds_type = config_dataset.ds_type
-                    elif ".parquet" in config_dataset.path:
-                        ds_type = "parquet"
-                    elif ".arrow" in config_dataset.path:
-                        ds_type = "arrow"
-                    elif ".csv" in config_dataset.path:
-                        ds_type = "csv"
-                    elif ".txt" in config_dataset.path:
-                        ds_type = "text"
+                    ds_type = get_ds_type(config_dataset)
+
                     ds = load_dataset(
                         ds_type,
                         name=config_dataset.name,
@@ -207,6 +272,22 @@ def load_tokenized_prepared_datasets(
                     data_files=config_dataset.data_files,
                     token=use_auth_token,
                 )
+            elif ds_from_cloud and remote_file_system:
+                if remote_file_system.isdir(config_dataset.path):
+                    ds = load_from_disk(
+                        config_dataset.path,
+                        storage_options=storage_options,
+                    )
+                elif remote_file_system.isfile(config_dataset.path):
+                    ds_type = get_ds_type(config_dataset)
+                    ds = load_dataset(
+                        ds_type,
+                        name=config_dataset.name,
+                        data_files=config_dataset.path,
+                        streaming=False,
+                        split=None,
+                        storage_options=storage_options,
+                    )
             else:
                 if isinstance(config_dataset.data_files, str):
                     fp = hf_hub_download(
@@ -298,11 +379,29 @@ def load_tokenized_prepared_datasets(
     return dataset, prompters
 
 
+def get_ds_type(config_dataset: DictDefault):
+    """
+    Get the dataset type from the path if it's not specified
+    """
+    ds_type = "json"
+    if config_dataset.ds_type:
+        ds_type = config_dataset.ds_type
+    elif ".parquet" in config_dataset.path:
+        ds_type = "parquet"
+    elif ".arrow" in config_dataset.path:
+        ds_type = "arrow"
+    elif ".csv" in config_dataset.path:
+        ds_type = "csv"
+    elif ".txt" in config_dataset.path:
+        ds_type = "text"
+    return ds_type
+
+
 def load_prepare_datasets(
     tokenizer: PreTrainedTokenizerBase,
     cfg,
     default_dataset_prepared_path,
-) -> Tuple[Dataset, Dataset, List[Any]]:
+) -> Tuple[Dataset, Dataset, List[Prompter]]:
     max_packed_sequence_len = (
         cfg.max_packed_sequence_len if cfg.max_packed_sequence_len else cfg.sequence_len
     )
@@ -311,7 +410,7 @@ def load_prepare_datasets(
     )  # make sure we don't accidentally set it larger than sequence_len
 
     tokenizer_name = tokenizer.__class__.__name__
-    prompters = []
+    prompters: List[Prompter] = []
     if cfg.max_packed_sequence_len is not None:
         # see if we can go ahead and load the stacked dataset
         seed = f"@{str(cfg.seed)}" if cfg.seed else ""
@@ -445,14 +544,13 @@ def load_prepare_datasets(
         train_fingerprint = md5(to_hash_train)
         test_fingerprint = md5(to_hash_test)
 
-        with zero_first(is_main_process()):
-            dataset = dataset.train_test_split(
-                test_size=cfg.val_set_size,
-                shuffle=False,
-                seed=cfg.seed or 42,
-                train_new_fingerprint=train_fingerprint,
-                test_new_fingerprint=test_fingerprint,
-            )
+        dataset = dataset.train_test_split(
+            test_size=cfg.val_set_size,
+            shuffle=False,
+            seed=cfg.seed or 42,
+            train_new_fingerprint=train_fingerprint,
+            test_new_fingerprint=test_fingerprint,
+        )
 
         train_dataset = dataset["train"]
         eval_dataset = dataset["test"]

src/axolotl/utils/dataloader.py

@@ -1,342 +0,0 @@
-# pylint: skip-file
-import hashlib
-import itertools
-import logging
-import math
-import time
-from queue import Queue
-from threading import Thread
-from typing import Any, Callable, List, Union
-
-import numba
-import numpy as np
-from torch.utils.data import DistributedSampler, Sampler
-
-LOG = logging.getLogger("axolotl.utils.dataloader")
-
-
-@numba.njit
-def ffd_check(a: np.ndarray, c: int, n: int):
-    # First-fit-decreasing bin packing
-    # Check if a[] could fit in n bins with capacity c
-    # https://en.wikipedia.org/wiki/First-fit-decreasing_bin_packing
-
-    a = np.sort(a)[::-1]
-    bins = np.full((n,), c, dtype=a.dtype)
-    for size in a:
-        not_found = True
-        for idx in range(n):
-            if bins[idx] >= size:
-                bins[idx] -= size
-                not_found = False
-                break
-
-        if not_found:
-            return False
-
-    return True
-
-
-@numba.njit
-def ffd_with_result(a: np.ndarray, c: int, start_index: int):
-    # First-fit-decreasing bin packing (with result return)
-
-    indices = np.argsort(a)[::-1]
-    a = a[indices]
-
-    bins: List[Any] = []
-    bins_result: List[Any] = []
-    for a_id, size in enumerate(a):
-        add_new = True
-        for idx in range(len(bins)):
-            if bins[idx] >= size:
-                bins[idx] -= size
-                bins_result[idx].append(indices[a_id] + start_index)
-                add_new = False
-                break
-
-        if add_new:
-            bins.append(c - size)
-            bins_result.append([indices[a_id] + start_index])
-
-    return bins_result, len(a)
-
-
-@numba.njit
-def allocate(
-    lengths: np.ndarray, lengths_cumsum: np.ndarray, rank: int, c: int, n: int
-):
-    """
-    :param lengths: array of lengths of each sample
-    :param lengths_cumsum: cumulative sum of consecutive lengths
-    :param rank: rank for this process
-    :param c: length of tokens per batch
-    :param n: number of ranks
-    :return:
-    """
-    # Dynamic batch allocator, similar to Multifit
-    # https://en.wikipedia.org/wiki/Multifit_algorithm
-    # ~99.5% efficiency on OpenChat training set (12 * 2048 ctx len)
-
-    s = 0
-    start_index = 0
-    result = []
-    result_totseqs = []
-
-    while True:
-        # binary search [left, right)
-        left = 1
-        right = 1 + np.searchsorted(lengths_cumsum[start_index:], s + c * n, "right")
-
-        while right - left > 1:
-            mid = (left + right) // 2
-            if ffd_check(lengths[start_index : start_index + mid], c, n):
-                left = mid
-            else:
-                right = mid
-
-        # use length left
-        batch, tot_seqs = ffd_with_result(
-            lengths[start_index : start_index + left], c, start_index
-        )
-        if len(batch) < n:
-            break
-
-        start_index += left
-        s = lengths_cumsum[start_index - 1]
-
-        # add local rank
-        result.append(batch[rank])
-        # add total seqs for all ranks
-        result_totseqs.append(tot_seqs)
-        # yield batch[rank], tot_seqs, s, len(result) * c * n
-    return result, result_totseqs, s, len(result) * c * n
-
-
-def chunk(iterable, n):
-    """
-    Chunk data into tuples of length n
-    """
-    # batched('ABCDEFG', 3) --> ABC DEF G
-    if n < 1:
-        raise ValueError("n must be at least one")
-    it = iter(iterable)
-    while batch := tuple(itertools.islice(it, n)):
-        yield batch
-
-
-def hash_indices(lst: List[int]) -> str:
-    # Convert the list of integers to a string representation
-    concatenated = ",".join(map(str, lst))
-
-    # Generate the hash
-    sha256 = hashlib.sha256()
-    sha256.update(concatenated.encode())
-
-    return sha256.hexdigest()
-
-
-class MultipackDistributedDataloader:
-    """Unpadded data loading using Multipack.
-    Adapted from https://github.com/imoneoi/openchat/blob/v3_fix_mle_loss/ochat/training_deepspeed/multipack_dataloader.py
-    Approximate (at most ~1.22x) the optimal solution of the identical-machines scheduling problem, which is NP-hard.
-    """
-
-    def __init__(
-        self,
-        dataset: Any,
-        collate_fn: Callable,
-        seq_max_length: int = 2048,
-        batch_size: int = 1,
-        sampler: Union[Sampler, DistributedSampler] = None,
-        packing_efficiency_estimate: float = 1.0,
-        sample_packing_seq_len_multiplier: int = 1,
-        device_count: int = 1,
-        prefetch_max: int = 1000,
-        num_epochs: int = 1,
-    ):
-        # Dataset
-        self.dataset = dataset
-        self.lengths = (
-            dataset.data.column("position_ids")
-            .to_pandas()
-            .apply(lambda x: x[-1] + 1)
-            .values
-        )
-        assert isinstance(self.lengths, np.ndarray)
-        assert batch_size % sample_packing_seq_len_multiplier == 0
-        assert batch_size >= sample_packing_seq_len_multiplier
-        self.sampler = sampler
-        self.batch_size = batch_size
-        self.sample_packing_seq_len_multiplier = sample_packing_seq_len_multiplier
-        self.seq_max_length = seq_max_length
-        self.batch_max_length = batch_size * seq_max_length
-        self.collate_fn = collate_fn
-        self.num_epochs = num_epochs
-
-        self.num_replicas = 1
-        self.rank = 0
-
-        # statistics
-        self.eff_total_used = 0
-        self.eff_total_slots = 0
-        self.packing_efficiency_estimate = packing_efficiency_estimate or 1.0
-        self.device_count = device_count
-
-        # maxsize is maximum number of samples in queue
-        self.prefetch_max = prefetch_max
-        self.queue: Queue = Queue(maxsize=prefetch_max)
-        self.thread = None
-
-    def _worker(self):
-        LOG.info(
-            f"[WORKER] Epochs: {self.num_epochs}, Samples: {self.len_w_stats()*self.batch_size}"
-        )
-        for epoch in range(self.num_epochs):
-            for sample in self._internal_batch_generator():
-                while True:
-                    if self.queue.full():
-                        time.sleep(1)
-                    else:
-                        break
-                self.queue.put(sample)
-
-            # stop the queue when epoch is done
-            self.queue.put(None)
-
-    def __iter__(self):
-        if hasattr(self.sampler, "set_epoch"):
-            new_epoch = self.sampler.epoch + 1
-            self.sampler.set_epoch(new_epoch)
-            LOG.info(f"calling sampler.set_epoch({new_epoch})")
-
-        if self.thread is None:
-            self.thread = Thread(target=self._worker, daemon=True)
-            self.thread.start()
-
-        while True:
-            item = self.queue.get()
-
-            if item is None:
-                break
-            yield item
-
-    def generate_batches(self, set_stats=False):
-        LOG.info("generating packed batches")
-        if self.sampler:
-            indices = [idx for idx in self.sampler]
-        else:
-            indices = range(0, len(self.dataset))
-
-        LOG.info(hash_indices(indices))
-        lengths = self.lengths[indices]
-        lengths_cumsum = np.cumsum(lengths)
-
-        batches, totseqs, total_used, total_slots = allocate(
-            lengths=lengths,
-            lengths_cumsum=lengths_cumsum,
-            rank=self.rank,
-            # c=self.batch_max_length,
-            c=self.seq_max_length * self.sample_packing_seq_len_multiplier,
-            n=self.num_replicas,
-        )
-
-        batches = [[indices[b_idx] for b_idx in batch] for batch in batches]
-
-        # statistics
-        if set_stats:
-            self.eff_total_used += total_used
-            self.eff_total_slots += total_slots
-
-        return batches, totseqs
-
-    def _internal_batch_generator(self):
-        all_batches, _ = self.generate_batches(set_stats=True)
-        features = self.dataset.features.keys()
-        len_remaining = self._len_est()
-        for batches in chunk(
-            all_batches, self.batch_size // self.sample_packing_seq_len_multiplier
-        ):
-            chunked_data = []
-            attn_mask_cum_idx = 0
-            for batch in batches:
-                concatenated = {}
-                batched_data = [self.dataset[batch_idx] for batch_idx in batch]
-                for feature in features:
-                    if feature == "length":
-                        continue
-                    if feature == "attention_mask":
-                        arrays = [
-                            (attn_mask_cum_idx + idx + 1) * np.array(item[feature])
-                            for idx, item in enumerate(batched_data)
-                            if feature in item
-                        ]
-                        attn_mask_cum_idx += len(batched_data)
-                        concatenated[feature] = np.concatenate(arrays)
-                    else:
-                        arrays = [
-                            np.array(item[feature])
-                            for item in batched_data
-                            if feature in item
-                        ]
-                        concatenated[feature] = np.concatenate(arrays)
-                chunked_data.append(concatenated)
-            yield self.collate_fn(chunked_data)
-            len_remaining -= 1
-            if not len_remaining:
-                return
-        # yield a no-op for cases where we don't have any data left to pack
-        for i in range(0, len_remaining):
-            yield self.collate_fn(
-                [
-                    {
-                        "input_ids": [0],
-                        "labels": [-100],
-                        "attention_mask": [True],
-                        "position_ids": [0],
-                    }
-                ]
-            )
-
-    def _len_est(self):
-        lengths_sum = np.sum(self.lengths)
-        lengths_sum_per_device = lengths_sum // self.device_count
-        LOG.info(
-            f"packing_efficiency_estimate: {self.packing_efficiency_estimate} "
-            f"total_num_tokens per device: {lengths_sum_per_device}"
-        )
-
-        # shave off 1% + 1 for dealing with variance in packing from random sampler to sampler
-        return (
-            math.floor(
-                0.99
-                * lengths_sum_per_device
-                / self.packing_efficiency_estimate
-                // self.seq_max_length
-                // self.batch_size
-            )
-            - 1
-        )
-
-    def __len__(self):
-        # this doesn't return the actual length b/c with distributed samplers, not all dataloaders get
-        # the same share of total tokens
-        # if not self.eff_total_used:
-        #     batches, _ = self.generate_batches(set_stats=True)
-        # LOG.info(
-        #     f"packing_efficiency_estimate: {self.packing_efficiency_estimate} "
-        #     f"actual packing efficiency: {self.efficiency()}"
-        # )
-        return max(1, self._len_est())
-
-    def len_w_stats(self):
-        if not self.eff_total_used:
-            batches, _ = self.generate_batches(set_stats=True)
-        LOG.info(
-            f"packing_efficiency_estimate: {self.packing_efficiency_estimate} "
-            f"actual packing efficiency: {self.efficiency()}"
-        )
-        return max(1, self._len_est())
-
-    def efficiency(self):
-        return self.eff_total_used / self.eff_total_slots

src/axolotl/utils/models.py

@@ -4,6 +4,7 @@ import math
 import os
 from typing import Optional, Tuple  # noqa: F401
 
+import addict
 import bitsandbytes as bnb
 import torch
 import transformers
@@ -17,11 +18,11 @@ from transformers import (  # noqa: F401
     AutoTokenizer,
     BitsAndBytesConfig,
     GPTQConfig,
-    LlamaConfig,
     PreTrainedModel,
     PreTrainedTokenizerBase,
 )
 
+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.dict import DictDefault
@@ -29,12 +30,51 @@ 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
-    return 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
 
 
 def load_tokenizer(cfg):
@@ -51,7 +91,7 @@ def load_tokenizer(cfg):
     if cfg.tokenizer_type:
         tokenizer_cls = getattr(transformers, cfg.tokenizer_type)
 
-    tokenizer_config = cfg.tokenizer_config or cfg.base_model_config
+    tokenizer_config = cfg.tokenizer_config or cfg.base_model_config or cfg.base_model
     tokenizer = tokenizer_cls.from_pretrained(
         tokenizer_config,
         trust_remote_code=cfg.trust_remote_code or False,
@@ -65,6 +105,7 @@ def load_tokenizer(cfg):
             "LlamaTokenizer",
             "LlamaTokenizerFast",
             "CodeLlamaTokenizer",
+            "CodeLlamaTokenizerFast",
         ]
         and hasattr(tokenizer, "pad_token")
         and not tokenizer.pad_token
@@ -80,11 +121,40 @@ def load_tokenizer(cfg):
     if cfg.is_mistral_derived_model and cfg.flash_attention and not cfg.sample_packing:
         tokenizer.padding_side = "left"
 
+    # Qwen base only has single token, so we need to set the special tokens
+    if cfg.is_qwen_derived_model:
+        token_ids = ["bos_token_id", "eos_token_id", "pad_token_id", "unk_token_id"]
+        for attr_name in token_ids:
+            if getattr(tokenizer, attr_name) is None:
+                setattr(tokenizer, attr_name, tokenizer.eod_id)
+
+        token_names = ["bos_token", "eos_token", "pad_token", "unk_token"]
+        for attr_name in token_names:
+            if getattr(tokenizer, attr_name) is None:
+                setattr(tokenizer, attr_name, "<|endoftext|>")
+
     if cfg.special_tokens:
         for k, val in cfg.special_tokens.items():
             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(
             [
@@ -110,7 +180,6 @@ def load_model(
     Load a model for a given configuration and tokenizer.
     """
     base_model = cfg.base_model
-    base_model_config = cfg.base_model_config
     model_type = cfg.model_type
     model_config = load_model_config(cfg)
 
@@ -180,6 +249,18 @@ def load_model(
         LOG.info("patching with flash attention")
         replace_mistral_attn_with_flash_attn(packed=cfg.sample_packing)
 
+    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 flash attention")
+        replace_mixtral_attn_with_multipack_flash_attn()
+
     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,
@@ -201,6 +282,7 @@ def load_model(
     model_kwargs = {}
 
     model_kwargs["device_map"] = cfg.device_map
+    model_kwargs["max_memory"] = cfg.max_memory
     model_kwargs["torch_dtype"] = cfg.torch_dtype
 
     if cfg.model_revision:
@@ -226,28 +308,30 @@ def load_model(
             bnb_4bit_quant_type="nf4",
         )
     # 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_config._attn_implementation = (  # pylint: disable=protected-access
+                    "flash_attention_2"
+                )
+        else:
+            if model_config.model_type == "mixtral":
+                model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "flash_attention_2"
+                )
 
     try:
         if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
             from transformers import LlamaForCausalLM
 
-            config_kwargs = {}
-            if cfg.rope_scaling:
-                config_kwargs["rope_scaling"] = cfg.rope_scaling
-            config = LlamaConfig.from_pretrained(
-                base_model_config,
-                **config_kwargs,
-            )
             model = LlamaForCausalLM.from_pretrained(
                 base_model,
-                config=config,
+                config=model_config,
                 load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
                 **model_kwargs,
@@ -292,92 +376,99 @@ def load_model(
         #         device=cfg.device,
         #     )
         #     model.train() # sets to train instead of eval mode
-        elif model_type == "MixFormerSequentialForCausalLM":
-            from axolotl.models.phi import MixFormerSequentialForCausalLM
+        elif model_type == "PhiForCausalLM":
+            from axolotl.models.phi import PhiForCausalLM
 
-            model = MixFormerSequentialForCausalLM.from_pretrained(
+            model = PhiForCausalLM.from_pretrained(
                 base_model,
                 load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
                 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"]
+            del model_kwargs["max_memory"]
+
+            model = MambaLMHeadModel.from_pretrained(
+                base_model,
+                **model_kwargs,
+            )
         elif model_type and not cfg.trust_remote_code:
             if cfg.gptq:
                 model = AutoModelForCausalLM.from_pretrained(
                     base_model,
+                    config=model_config,
                     trust_remote_code=cfg.trust_remote_code or False,
                     **model_kwargs,
                 )
             else:
                 model = getattr(transformers, model_type).from_pretrained(
                     base_model,
+                    config=model_config,
                     load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
                     load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
                     trust_remote_code=cfg.trust_remote_code or False,
                     **model_kwargs,
                 )
         else:
-            config = AutoConfig.from_pretrained(
-                base_model,
-                trust_remote_code=cfg.trust_remote_code or False,
-            )
             # Shouldn't be a problem most of the time. will obviously error if the model doesn't support this
             # when training starts
             if (
-                hasattr(config, "max_seq_len")
-                and config.max_seq_len
-                and cfg.sequence_len > config.max_seq_len
+                hasattr(model_config, "max_seq_len")
+                and model_config.max_seq_len
+                and cfg.sequence_len > model_config.max_seq_len
             ):
-                config.max_seq_len = cfg.sequence_len
+                model_config.max_seq_len = cfg.sequence_len
                 LOG.warning(f"increasing context length to {cfg.sequence_len}")
             elif (
-                hasattr(config, "max_sequence_length")
-                and config.max_sequence_length
-                and cfg.sequence_len > config.max_sequence_length
+                hasattr(model_config, "max_sequence_length")
+                and model_config.max_sequence_length
+                and cfg.sequence_len > model_config.max_sequence_length
             ):
-                config.max_sequence_length = cfg.sequence_len
+                model_config.max_sequence_length = cfg.sequence_len
                 LOG.warning(f"increasing context length to {cfg.sequence_len}")
             if cfg.gptq:
                 model = AutoModelForCausalLM.from_pretrained(
                     base_model,
-                    config=config,
+                    config=model_config,
                     trust_remote_code=cfg.trust_remote_code or False,
                     **model_kwargs,
                 )
             else:
                 model = AutoModelForCausalLM.from_pretrained(
                     base_model,
-                    config=config,
+                    config=model_config,
                     load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
                     load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
                     trust_remote_code=cfg.trust_remote_code or False,
                     **model_kwargs,
                 )
     except Exception as err:  # pylint: disable=broad-exception-caught
-        LOG.error(
-            "Exception raised attempting to load model, retrying with AutoModelForCausalLM"
-        )
         LOG.exception(err)
-        model = AutoModelForCausalLM.from_pretrained(
-            base_model,
-            load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
-            load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-            trust_remote_code=cfg.trust_remote_code or False,
-            **model_kwargs,
-        )
+        raise err
 
     embeddings_len = (
         math.ceil(len(tokenizer) / 32) * 32
         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
     ):
@@ -387,20 +478,22 @@ 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)
@@ -415,15 +508,22 @@ def load_model(
                 module.to(torch.float32)
 
     needs_fa2_dtype = cfg.adapter or cfg.fsdp
+    skip_prepare_model_for_kbit_training = False
+
+    if cfg.model_config_type == "qwen" and cfg.adapter == "lora":
+        # Qwen doesn't play nicely with LoRA if this is enabled
+        skip_prepare_model_for_kbit_training = True
+
     if (cfg.adapter == "lora" and load_in_8bit) or (
         cfg.adapter == "qlora" and cfg.load_in_4bit
     ):
         LOG.info("converting PEFT model w/ prepare_model_for_kbit_training")
         if cfg.gradient_checkpointing:
             model.gradient_checkpointing_enable()
-        model = prepare_model_for_kbit_training(
-            model, use_gradient_checkpointing=cfg.gradient_checkpointing
-        )
+        if not skip_prepare_model_for_kbit_training:
+            model = prepare_model_for_kbit_training(
+                model, use_gradient_checkpointing=cfg.gradient_checkpointing
+            )
         needs_fa2_dtype = True
 
     # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
@@ -452,7 +552,8 @@ 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)

src/axolotl/utils/samplers/multipack.py

@@ -181,13 +181,16 @@ class MultipackBatchSampler(BatchSampler):
         )
 
         # shave off 1% + 1 for dealing with variance in packing from random sampler to sampler
-        return (
-            world_size
-            * math.floor(
-                0.99
-                * lengths_sum_per_device
-                / self.packing_efficiency_estimate
-                // self.batch_max_len
-            )
-            - 1
+        return max(
+            0,
+            (
+                world_size
+                * math.floor(
+                    0.99
+                    * lengths_sum_per_device
+                    / self.packing_efficiency_estimate
+                    // self.batch_max_len
+                )
+                - 1
+            ),
         )

src/axolotl/utils/trainer.py

@@ -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:
@@ -141,33 +143,38 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
     return train_dataset, eval_dataset
 
 
-def calculate_total_num_steps(cfg, train_dataset):
-    if cfg.sample_packing:
-        # we have to drop anything longer then sequence len otherwise
-        # flash attention with position ids fails
-        if not cfg.total_num_tokens:
-            total_num_tokens = np.sum(
-                train_dataset.data.column("input_ids")
-                .to_pandas()
-                .apply(lambda x: len(x))  # pylint: disable=unnecessary-lambda
-                .values
-            )
-            LOG.debug(f"total_num_tokens: {total_num_tokens}", main_process_only=True)
+def calculate_total_num_steps(cfg, train_dataset, update=True):
+    if not cfg.total_num_tokens:
+        total_num_tokens = np.sum(
+            train_dataset.data.column("input_ids")
+            .to_pandas()
+            .apply(lambda x: len(x))  # pylint: disable=unnecessary-lambda
+            .values
+        )
+        LOG.debug(f"total_num_tokens: {total_num_tokens}", main_process_only=True)
+        if update:
             cfg.total_num_tokens = total_num_tokens
 
-        if not cfg.total_supervised_tokens:
-            total_supervised_tokens = (
-                train_dataset.data.column("labels")
-                .to_pandas()
-                .apply(lambda x: np.sum(np.array(x) != -100))
-                .sum()
-            )
-            LOG.debug(
-                f"`total_supervised_tokens: {total_supervised_tokens}`",
-                main_process_only=True,
-            )
+    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()
+            .apply(lambda x: np.sum(np.array(x) != -100))
+            .sum()
+        )
+        LOG.debug(
+            f"`total_supervised_tokens: {total_supervised_tokens}`",
+            main_process_only=True,
+        )
+        if update:
             cfg.total_supervised_tokens = total_supervised_tokens
 
+    if not skip_estimates and cfg.sample_packing:
+        # we have to drop anything longer then sequence len otherwise
+        # flash attention with position ids fails
+
         if cfg.sample_packing_eff_est:
             total_num_steps = (
                 # match count to len est in dataloader
@@ -231,7 +238,8 @@ def calculate_total_num_steps(cfg, train_dataset):
             sample_packing_eff_est = (
                 math.ceil(sample_packing_actual_eff_all * 100.0) / 100.0
             )
-            cfg.sample_packing_eff_est = sample_packing_eff_est
+            if update:
+                cfg.sample_packing_eff_est = sample_packing_eff_est
             LOG.debug(
                 f"sample_packing_eff_est: {cfg.sample_packing_eff_est}",
                 main_process_only=True,
@@ -263,14 +271,14 @@ def setup_fsdp_envs(cfg):
         ] = cfg.fsdp_config.fsdp_transformer_layer_cls_to_wrap
 
 
-def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps):
+def prepare_optim_env(cfg):
     if cfg.fsdp:
         setup_fsdp_envs(cfg)
     elif cfg.deepspeed:
         os.environ["ACCELERATE_USE_DEEPSPEED"] = "true"
-    if cfg.fp8:
-        os.environ["ACCELERATE_MIXED_PRECISION"] = "fp8"
 
+
+def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps):
     trainer_builder = HFCausalTrainerBuilder(cfg, model, 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/e2e/test_lora_llama.py

@@ -101,6 +101,7 @@ class TestLoraLlama(unittest.TestCase):
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch",
                 "lr_scheduler": "cosine",
+                "bf16": True,
             }
         )
         normalize_config(cfg)

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_phi.py

@@ -31,7 +31,7 @@ class TestPhi(unittest.TestCase):
             {
                 "base_model": "microsoft/phi-1_5",
                 "trust_remote_code": True,
-                "model_type": "MixFormerSequentialForCausalLM",
+                "model_type": "PhiForCausalLM",
                 "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 512,
                 "sample_packing": False,
@@ -76,7 +76,7 @@ class TestPhi(unittest.TestCase):
             {
                 "base_model": "microsoft/phi-1_5",
                 "trust_remote_code": True,
-                "model_type": "MixFormerSequentialForCausalLM",
+                "model_type": "PhiForCausalLM",
                 "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 512,
                 "sample_packing": True,

tests/e2e/test_resume.py

@@ -0,0 +1,95 @@
+"""
+E2E tests for resuming training
+"""
+
+import logging
+import os
+import re
+import subprocess
+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 most_recent_subdir, with_temp_dir
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestResumeLlama(unittest.TestCase):
+    """
+    Test case for resuming training of llama models
+    """
+
+    @with_temp_dir
+    def test_resume_qlora(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "sample_packing": True,
+                "flash_attention": True,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 32,
+                "lora_alpha": 64,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "vicgalle/alpaca-gpt4",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "save_steps": 10,
+                "save_total_limit": 5,
+                "max_steps": 40,
+            }
+        )
+        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)
+
+        resume_cfg = cfg | DictDefault(
+            {
+                "resume_from_checkpoint": f"{temp_dir}/checkpoint-30/",
+            }
+        )
+        normalize_config(resume_cfg)
+        cli_args = TrainerCliArgs()
+
+        train(cfg=resume_cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "adapter_model.bin").exists()
+
+        tb_log_path_1 = most_recent_subdir(temp_dir + "/runs")
+        cmd = f"tensorboard --inspect  --logdir {tb_log_path_1}"
+        res = subprocess.run(
+            cmd, shell=True, text=True, capture_output=True, check=True
+        )
+        pattern = r"first_step\s+(\d+)"
+        first_steps = int(re.findall(pattern, res.stdout)[0])
+        assert first_steps == 31

tests/e2e/utils.py

@@ -1,10 +1,11 @@
 """
 helper utils for tests
 """
-
+import os
 import shutil
 import tempfile
 from functools import wraps
+from pathlib import Path
 
 
 def with_temp_dir(test_func):
@@ -20,3 +21,13 @@ def with_temp_dir(test_func):
             shutil.rmtree(temp_dir)
 
     return wrapper
+
+
+def most_recent_subdir(path):
+    base_path = Path(path)
+    subdirectories = [d for d in base_path.iterdir() if d.is_dir()]
+    if not subdirectories:
+        return None
+    subdir = max(subdirectories, key=os.path.getctime)
+
+    return subdir

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):
@@ -649,3 +651,113 @@ class ValidationTest(unittest.TestCase):
         )
 
         validate_config(cfg)
+
+    def test_warmup_step_no_conflict(self):
+        cfg = DictDefault(
+            {
+                "warmup_steps": 10,
+                "warmup_ratio": 0.1,
+            }
+        )
+
+        with pytest.raises(
+            ValueError,
+            match=r".*warmup_steps and warmup_ratio are mutually exclusive*",
+        ):
+            validate_config(cfg)
+
+        cfg = DictDefault(
+            {
+                "warmup_steps": 10,
+            }
+        )
+
+        validate_config(cfg)
+
+        cfg = DictDefault(
+            {
+                "warmup_ratio": 0.1,
+            }
+        )
+
+        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"