Add e2e test for ia3 ft

Co-authored-by: NanoCode012 <kevinvong@rocketmail.com>

.github/workflows/main.yml

@@ -23,7 +23,6 @@ jobs:
             python_version: "3.10"
             pytorch: 2.0.1
             axolotl_extras:
-            is_latest: true
           - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.10"
@@ -55,9 +54,7 @@ jobs:
             PYTORCH_VERSION=${{ matrix.pytorch }}
           file: ./docker/Dockerfile
           push: ${{ github.event_name != 'pull_request' }}
-          tags: |
-            ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
-            ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
+          tags: ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
           labels: ${{ steps.metadata.outputs.labels }}
   build-axolotl-runpod:
     needs: build-axolotl

.pylintrc

@@ -12,3 +12,4 @@ generated-members=numpy.*, torch.*
 disable=missing-function-docstring, line-too-long, import-error,
     too-many-arguments, too-many-locals, too-many-statements, too-many-branches, too-few-public-methods,
     too-many-instance-attributes, fixme, import-outside-toplevel, logging-fstring-interpolation,
+    too-many-boolean-expressions,

README.md

@@ -32,6 +32,7 @@ Features:
   - [How to Use Custom Pretokenized Dataset](#how-to-use-your-custom-pretokenized-dataset)
 - [Config](#config)
   - [Train](#train)
+  - [Training w/ Deepspeed](#training-with-deepspeed)
   - [Inference](#inference)
   - [Merge LORA to Base](#merge-lora-to-base)
 - [Common Errors](#common-errors-)
@@ -95,7 +96,7 @@ accelerate launch -m axolotl.cli.train examples/openllama-3b/lora.yml
 
 # inference
 accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
-    --lora_model_dir="./lora-out"
+    --peft_model_dir="./lora-out"
 ```
 
 ## Installation
@@ -114,25 +115,6 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
   docker compose up -d
   ```
 
-  <details>
-
-  <summary>Docker advanced</summary>
-
-  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
-  ```
-
-  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.
-
-  [More information on nvidia website](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html#setincshmem)
-
-  </details>
-
 #### Conda/Pip venv
   1. Install python >=**3.9**
 
@@ -374,13 +356,6 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
         - typescript
       type: ... # unimplemented custom format
 
-  # fastchat conversation
-  # See 'conversation' options: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
-  datasets:
-    - path: ...
-      type: sharegpt
-      conversation: chatml
-
   # local
   datasets:
     - path: data.jsonl # or json
@@ -409,17 +384,17 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
 - lora
   ```yaml
   adapter: lora # qlora or leave blank for full finetune
-  lora_r: 8
-  lora_alpha: 16
-  lora_dropout: 0.05
-  lora_target_modules:
+  peft_r: 8
+  peft_alpha: 16
+  peft_dropout: 0.05
+  peft_target_modules:
     - q_proj
     - v_proj
   ```
 
 <details>
 
-<summary>All yaml options (click me)</summary>
+<summary>All yaml options</summary>
 
 ```yaml
 # This is the huggingface model that contains *.pt, *.safetensors, or *.bin files
@@ -486,9 +461,7 @@ 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
-
-    # 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
+    conversation:  # Optional[str] fastchat conversation type, only used with type: sharegpt
 
   # Custom user prompt
   - path: repo
@@ -558,15 +531,15 @@ total_num_tokens:
 adapter: lora
 # If you already have a lora model trained that you want to load, put that here.
 # This means after training, if you want to test the model, you should set this to the value of `lora_out_dir`.
-lora_model_dir:
+peft_model_dir:
 
 # LoRA hyperparameters
 # For more details about the following options, see:
 # https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2
-lora_r: 8
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules:
+peft_r: 8
+peft_alpha: 16
+peft_dropout: 0.05
+peft_target_modules:
   - q_proj
   - v_proj
 #  - k_proj
@@ -574,13 +547,13 @@ lora_target_modules:
 #  - gate_proj
 #  - down_proj
 #  - up_proj
-lora_target_linear: # If true, will target all linear layers
+peft_target_linear: # if true, will target all linear layers
 
 # If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.
 # For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.
 # `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.
 # https://github.com/huggingface/peft/issues/334#issuecomment-1561727994
-lora_modules_to_save:
+peft_modules_to_save:
 #  - embed_tokens
 #  - lm_head
 
@@ -588,7 +561,8 @@ lora_modules_to_save:
 # If you merge the adapter to the base model, a subdirectory `merged` will be created under this directory.
 # Make sure `lora_model_dir` points to this directory if you want to use the trained model.
 lora_out_dir:
-lora_fan_in_fan_out: false
+peft_fan_in_fan_out: false
+peft_feedforward_modules:  # ffn modules for IA3, for llama down projection
 
 # ReLoRA configuration
 # Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
@@ -618,14 +592,14 @@ gradient_accumulation_steps: 1
 # The number of samples to include in each batch. This is the number of samples sent to each GPU.
 micro_batch_size: 2
 eval_batch_size:
-num_epochs: 4
+num_epochs: 3
 warmup_steps: 100
 learning_rate: 0.00003
 lr_quadratic_warmup:
 logging_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
+eval_steps: # Leave empty to eval at each epoch
 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.
@@ -711,8 +685,6 @@ xformers_attention:
 flash_attention:
 flash_attn_cross_entropy:  # Whether to use flash-attention cross entropy implementation - advanced use only
 flash_attn_rms_norm:  # Whether to use flash-attention rms norm implementation - advanced use only
-flash_attn_fuse_qkv: # Whether to fuse QKV into a single operation
-flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
 # Whether to use scaled-dot-product attention
 # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
 sdp_attention:
@@ -842,41 +814,14 @@ Run
 accelerate launch -m axolotl.cli.train your_config.yml
 ```
 
-#### Preprocess dataset
-
-You can optionally pre-tokenize dataset with the following before finetuning.
-This is recommended for large datasets.
-
-- Set `push_dataset_to_hub: hf_user/repo` to push it to Huggingface.
-- Use `--debug` to see preprocessed examples.
-
-```bash
-python -m axolotl.cli.preprocess your_config.yml
-```
-
 #### Multi-GPU
 
-Below are the options available in axolotl for training with multiple GPUs. Note that DeepSpeed
-is the recommended multi-GPU option currently because FSDP may experience
-[loss instability](https://github.com/huggingface/transformers/issues/26498).
-
-##### DeepSpeed
-
-Deepspeed is an optimization suite for multi-gpu systems allowing you to train much larger models than you
-might typically be able to fit into your GPU's VRAM. More information about the various optimization types
-for deepspeed is available at https://huggingface.co/docs/accelerate/main/en/usage_guides/deepspeed#what-is-integrated
-
-We provide several default deepspeed JSON configurations for ZeRO stage 1, 2, and 3.
-
-```yaml
-deepspeed: deepspeed/zero1.json
+You can optionally pre-tokenize dataset with the following before finetuning:
+```bash
+CUDA_VISIBLE_DEVICES="" accelerate launch -m axolotl.cli.train your_config.yml --prepare_ds_only
 ```
 
-```shell
-accelerate launch -m axolotl.cli.train examples/llama-2/config.py --deepspeed deepspeed/zero1.json
-```
-
-##### FSDP
+##### Config
 
 - llama FSDP
 ```yaml
@@ -901,13 +846,31 @@ wandb_run_id:
 wandb_log_model:
 ```
 
+### Training with Deepspeed
+
+Deepspeed is an optimization suite for multi-gpu systems allowing you to train much larger models than you
+might typically be able to fit into your GPU's VRAM. More information about the various optimization types
+for deepspeed is available at https://huggingface.co/docs/accelerate/main/en/usage_guides/deepspeed#what-is-integrated
+
+We provide several default deepspeed JSON configurations for ZeRO stage 1, 2, and 3.
+
+```shell
+accelerate launch -m axolotl.cli.train examples/llama-2/config.py --deepspeed deepspeed/zero1.json
+```
+
+or
+
+```yaml
+deepspeed: deepspeed/zero1.json
+```
+
 ### Inference
 
 Pass the appropriate flag to the train command:
 
 - Pretrained LORA:
   ```bash
-  python -m axolotl.cli.inference examples/your_config.yml --lora_model_dir="./lora-output-dir"
+  python -m axolotl.cli.inference examples/your_config.yml --peft_model_dir="./lora-output-dir"
   ```
 - Full weights finetune:
   ```bash
@@ -928,7 +891,7 @@ Please use `--sample_packing False` if you have it on and receive the error simi
 Add below flag to train command above
 
 ```bash
-python3 -m axolotl.cli.merge_lora examples/your_config.yml --lora_model_dir="./completed-model" --load_in_8bit=False --load_in_4bit=False
+python3 -m axolotl.cli.merge_lora examples/your_config.yml --peft_model_dir="./completed-model" --load_in_8bit=False --load_in_4bit=False
 ```
 
 If you run out of CUDA memory, you can try to merge in system RAM with
@@ -939,8 +902,6 @@ CUDA_VISIBLE_DEVICES="" python3 -m axolotl.cli.merge_lora ...
 
 ## Common Errors 🧰
 
-See also the [FAQ's](./docs/faq.md).
-
 > If you encounter a 'Cuda out of memory' error, it means your GPU ran out of memory during the training process. Here's how to resolve it:
 
 Please reduce any below

deepspeed/zero3.json

@@ -1,6 +1,14 @@
 {
   "zero_optimization": {
     "stage": 3,
+    "offload_optimizer": {
+      "device": "cpu",
+      "pin_memory": true
+    },
+    "offload_param": {
+      "device": "cpu",
+      "pin_memory": true
+    },
     "overlap_comm": true,
     "contiguous_gradients": true,
     "sub_group_size": 0,
@@ -33,13 +41,12 @@
     }
   },
   "scheduler": {
-    "type": "WarmupDecayLR",
+    "type": "WarmupLR",
     "params": {
       "warmup_min_lr": "auto",
       "warmup_max_lr": "auto",
       "warmup_num_steps": "auto",
-      "warmup_type": "linear",
-      "total_num_steps": "auto"
+      "warmup_type": "linear"
     }
   },
   "gradient_accumulation_steps": "auto",

docs/faq.md

@@ -1,18 +0,0 @@
-# Axolotl FAQ's
-
-
-> The trainer stopped and hasn't progressed in several minutes.
-
-Usually an issue with the GPU's communicating with each other. See the [NCCL doc](../docs/nccl.md)
-
-> Exitcode -9
-
-This usually happens when you run out of system RAM.
-
-> Exitcode -7 while using deepspeed
-
-Try upgrading deepspeed w: `pip install -U deepspeed`
-
-> AttributeError: 'DummyOptim' object has no attribute 'step'
-
-You may be using deepspeed with single gpu. Please don't set `deepspeed:` in yaml or cli.

examples/cerebras/btlm-ft.yml

@@ -1,4 +1,5 @@
 base_model: cerebras/btlm-3b-8k-base
+base_model_config: cerebras/btlm-3b-8k-base
 model_type: AutoModelForCausalLM
 tokenizer_type: GPT2Tokenizer
 trust_remote_code: true
@@ -17,7 +18,7 @@ dataset_prepared_path: last_prepared_run
 val_set_size: 0.01
 
 adapter:
-lora_model_dir:
+peft_model_dir:
 sequence_len: 2048
 max_packed_sequence_len:
 sample_packing: false

examples/cerebras/qlora.yml

@@ -1,4 +1,5 @@
 base_model: cerebras/Cerebras-GPT-1.3B
+base_model_config: cerebras/Cerebras-GPT-1.3B
 load_in_8bit: false
 load_in_4bit: true
 strict: false
@@ -9,7 +10,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.01
 adapter: qlora
-lora_model_dir:
+peft_model_dir:
 sequence_len: 2048
 max_packed_sequence_len: 2048
 lora_r: 16
@@ -49,7 +50,7 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 save_steps:
 debug:
 deepspeed:

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

@@ -1,4 +1,5 @@
 base_model: codellama/CodeLlama-13b-hf
+base_model_config: codellama/CodeLlama-13b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -19,7 +20,7 @@ sample_packing: true
 pad_to_sequence_len: true
 
 adapter: lora
-lora_model_dir:
+peft_model_dir:
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
@@ -34,7 +35,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 4
+num_epochs: 3
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,7 +55,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 save_steps:
 debug:
 deepspeed:

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

@@ -1,4 +1,5 @@
 base_model: codellama/CodeLlama-13b-hf
+base_model_config: codellama/CodeLlama-13b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -15,7 +16,7 @@ val_set_size: 0.01
 output_dir: ./qlora-out
 
 adapter: qlora
-lora_model_dir:
+peft_model_dir:
 
 sequence_len: 4096
 sample_packing: true
@@ -36,7 +37,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 4
+num_epochs: 3
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,7 +57,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 save_steps:
 debug:
 deepspeed:

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

@@ -1,4 +1,5 @@
 base_model: codellama/CodeLlama-34b-hf
+base_model_config: codellama/CodeLlama-34b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -19,7 +20,7 @@ sample_packing: true
 pad_to_sequence_len: true
 
 adapter: lora
-lora_model_dir:
+peft_model_dir:
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
@@ -34,7 +35,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 4
+num_epochs: 3
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,7 +55,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 save_steps:
 debug:
 deepspeed:

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

@@ -1,4 +1,5 @@
 base_model: codellama/CodeLlama-34b-hf
+base_model_config: codellama/CodeLlama-34b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -15,7 +16,7 @@ val_set_size: 0.01
 output_dir: ./qlora-out
 
 adapter: qlora
-lora_model_dir:
+peft_model_dir:
 
 sequence_len: 4096
 sample_packing: true
@@ -36,7 +37,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 4
+num_epochs: 3
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,7 +57,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 save_steps:
 debug:
 deepspeed:

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

@@ -1,4 +1,5 @@
 base_model: codellama/CodeLlama-7b-hf
+base_model_config: codellama/CodeLlama-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -19,7 +20,7 @@ sample_packing: true
 pad_to_sequence_len: true
 
 adapter: lora
-lora_model_dir:
+peft_model_dir:
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
@@ -34,7 +35,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 4
+num_epochs: 3
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,7 +55,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 save_steps:
 debug:
 deepspeed:

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

@@ -1,4 +1,5 @@
 base_model: codellama/CodeLlama-7b-hf
+base_model_config: codellama/CodeLlama-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -15,7 +16,7 @@ val_set_size: 0.01
 output_dir: ./qlora-out
 
 adapter: qlora
-lora_model_dir:
+peft_model_dir:
 
 sequence_len: 4096
 sample_packing: true
@@ -36,7 +37,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 4
+num_epochs: 3
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,7 +57,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 save_steps:
 debug:
 deepspeed:

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

@@ -1,4 +1,5 @@
 base_model: tiiuae/falcon-7b
+base_model_config: tiiuae/falcon-7b
 trust_remote_code: true
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
@@ -14,7 +15,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.01
 adapter: lora
-lora_model_dir:
+peft_model_dir:
 sequence_len: 2048
 max_packed_sequence_len:
 lora_r: 16

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

@@ -1,6 +1,7 @@
 # 1b: tiiuae/falcon-rw-1b
 # 40b: tiiuae/falcon-40b
 base_model: tiiuae/falcon-7b
+base_model_config: tiiuae/falcon-7b
 # required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
 trust_remote_code: true
 model_type: AutoModelForCausalLM
@@ -21,7 +22,7 @@ dataset_prepared_path:
 val_set_size: 0.01
 # enable QLoRA
 adapter: qlora
-lora_model_dir:
+peft_model_dir:
 sequence_len: 2048
 max_packed_sequence_len:
 
@@ -53,7 +54,7 @@ output_dir: ./qlora-out
 # decrease if OOM, increase for max VRAM utilization
 micro_batch_size: 1
 gradient_accumulation_steps: 2
-num_epochs: 4
+num_epochs: 3
 # Optimizer for QLoRA
 optimizer: paged_adamw_32bit
 torchdistx_path:

examples/falcon/config-7b.yml

@@ -1,4 +1,5 @@
 base_model: tiiuae/falcon-7b
+base_model_config: tiiuae/falcon-7b
 trust_remote_code: true
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
@@ -14,7 +15,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.01
 adapter:
-lora_model_dir:
+peft_model_dir:
 sequence_len: 2048
 max_packed_sequence_len:
 lora_r: 64

examples/gptj/qlora.yml

@@ -1,4 +1,5 @@
 base_model: EleutherAI/gpt-j-6b
+base_model_config: EleutherAI/gpt-j-6b
 load_in_8bit: false
 load_in_4bit: true
 strict: false
@@ -9,7 +10,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.01
 adapter: qlora
-lora_model_dir:
+peft_model_dir:
 sequence_len: 2048
 max_packed_sequence_len:
 lora_r: 8
@@ -46,7 +47,7 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 save_steps:
 debug:
 deepspeed:

examples/jeopardy-bot/config.yml

@@ -1,4 +1,5 @@
 base_model: huggyllama/llama-7b
+base_model_config: huggyllama/llama-7b
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 load_in_8bit: false
@@ -8,7 +9,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.02
 adapter:
-lora_model_dir:
+peft_model_dir:
 sequence_len: 512
 max_packed_sequence_len:
 lora_r:
@@ -24,7 +25,7 @@ wandb_log_model:
 output_dir: ./jeopardy-bot-7b
 gradient_accumulation_steps: 1
 micro_batch_size: 1
-num_epochs: 4
+num_epochs: 3
 optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine

examples/llama-2/README.md

@@ -9,16 +9,12 @@ gradient_accumulation_steps: 2
 micro_batch_size: 1
 
 ```shell
-accelerate launch -m axolotl.cli.train examples/llama-2/qlora.yml
+accelerate launch scripts/finetune.py examples/llama-2/qlora.yml
+
 ```
 or
 
 ```shell
-accelerate launch -m axolotl.cli.train examples/llama-2/lora.yml
-```
+accelerate launch scripts/finetune.py examples/llama-2/lora.yml
 
-To launch a full finetuning with 16-bit precision:
-
-```shell
-accelerate launch -m axolotl.cli.train examples/llama-2/fft_optimized.yml
 ```

examples/llama-2/gptq-lora.yml

@@ -1,4 +1,5 @@
 base_model: TheBloke/Llama-2-7B-GPTQ
+base_model_config: TheBloke/Llama-2-7B-GPTQ
 is_llama_derived_model: false
 gptq: true
 gptq_disable_exllama: true
@@ -17,7 +18,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.01
 adapter: lora
-lora_model_dir:
+peft_model_dir:
 sequence_len: 4096
 sample_packing:
 lora_r: 8
@@ -37,7 +38,7 @@ wandb_log_model:
 output_dir: ./model-out
 gradient_accumulation_steps: 1
 micro_batch_size: 1
-num_epochs: 4
+num_epochs: 3
 optimizer: adamw_torch
 adam_beta2: 0.95
 adam_eps: 0.00001

examples/llama-2/ia3.yml

@@ -1,9 +1,10 @@
-base_model: NousResearch/Llama-2-7b-hf
+base_model: meta-llama/Llama-2-7b-hf
+base_model_config: meta-llama/Llama-2-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 is_llama_derived_model: true
 
-load_in_8bit: false
+load_in_8bit: true
 load_in_4bit: false
 strict: false
 
@@ -12,19 +13,21 @@ datasets:
     type: alpaca
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.01
-output_dir: ./out
+output_dir: ./ia3-out
 
 sequence_len: 4096
 sample_packing: true
 pad_to_sequence_len: true
 
-adapter:
-lora_model_dir:
-lora_r:
-lora_alpha:
-lora_dropout:
-lora_target_linear:
-lora_fan_in_fan_out:
+adapter: ia3
+peft_model_dir:
+peft_target_modules:
+  - k_proj
+  - v_proj
+  - down_proj
+peft_feedforward_modules:
+  - down_proj
+peft_fan_in_fan_out: false
 
 wandb_project:
 wandb_entity:
@@ -33,8 +36,8 @@ wandb_run_id:
 wandb_log_model:
 
 gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 1
+micro_batch_size: 2
+num_epochs: 5
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -52,17 +55,14 @@ local_rank:
 logging_steps: 1
 xformers_attention:
 flash_attention: true
-flash_attn_cross_entropy: false
-flash_attn_rms_norm: true
-flash_attn_fuse_qkv: false
-flash_attn_fuse_mlp: true
 
-warmup_steps: 100
+warmup_steps: 10
 eval_steps: 0.05
 eval_table_size:
+eval_table_max_new_tokens:
 save_steps:
 debug:
-deepspeed: #deepspeed/zero2.json # multi-gpu only
+deepspeed:
 weight_decay: 0.1
 fsdp:
 fsdp_config:

examples/llama-2/lora.yml

@@ -1,4 +1,5 @@
 base_model: NousResearch/Llama-2-7b-hf
+base_model_config: NousResearch/Llama-2-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 is_llama_derived_model: true
@@ -19,7 +20,7 @@ sample_packing: true
 pad_to_sequence_len: true
 
 adapter: lora
-lora_model_dir:
+peft_model_dir:
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
@@ -34,7 +35,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 4
+num_epochs: 3
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,7 +55,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 eval_table_size:
 eval_table_max_new_tokens: 128
 save_steps:

examples/llama-2/qlora.yml

@@ -1,4 +1,5 @@
 base_model: NousResearch/Llama-2-7b-hf
+base_model_config: NousResearch/Llama-2-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 is_llama_derived_model: true
@@ -15,7 +16,7 @@ val_set_size: 0.01
 output_dir: ./qlora-out
 
 adapter: qlora
-lora_model_dir:
+peft_model_dir:
 
 sequence_len: 4096
 sample_packing: true
@@ -36,7 +37,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 4
+num_epochs: 3
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,7 +57,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 eval_table_size:
 save_steps:
 debug:

examples/llama-2/relora.yml

@@ -1,4 +1,5 @@
 base_model: NousResearch/Llama-2-7b-hf
+base_model_config: NousResearch/Llama-2-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 is_llama_derived_model: true
@@ -15,7 +16,7 @@ val_set_size: 0.01
 output_dir: ./relora-out
 
 adapter: qlora
-lora_model_dir:
+peft_model_dir:
 
 sequence_len: 4096
 sample_packing: true
@@ -40,7 +41,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 4
-num_epochs: 4
+num_epochs: 3
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -60,7 +61,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 save_steps: 50
 debug:
 deepspeed:

examples/llama-2/tiny-llama.yml

@@ -1,4 +1,5 @@
 base_model: PY007/TinyLlama-1.1B-step-50K-105b
+base_model_config: PY007/TinyLlama-1.1B-step-50K-105b
 
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
@@ -19,7 +20,7 @@ sequence_len: 4096
 sample_packing: true
 
 adapter: lora
-lora_model_dir:
+peft_model_dir:
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
@@ -34,7 +35,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 4
+num_epochs: 3
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,7 +55,7 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
+eval_steps: 20
 eval_table_size:
 save_steps:
 debug:

examples/mistral/config.yml

@@ -1,4 +1,5 @@
 base_model: mistralai/Mistral-7B-v0.1
+base_model_config: mistralai/Mistral-7B-v0.1
 model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
 is_mistral_derived_model: true
@@ -26,7 +27,7 @@ wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 4
+num_epochs: 3
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.000005
@@ -46,8 +47,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-eval_table_size:
+eval_steps: 20
+eval_table_size: 5
 eval_table_max_new_tokens: 128
 save_steps:
 debug:

examples/mistral/qlora.yml

@@ -1,4 +1,5 @@
 base_model: mistralai/Mistral-7B-v0.1
+base_model_config: mistralai/Mistral-7B-v0.1
 model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
 is_mistral_derived_model: true
@@ -63,8 +64,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-eval_table_size:
+eval_steps: 20
+eval_table_size: 5
 eval_table_max_new_tokens: 128
 save_steps:
 debug:

examples/mpt-7b/config.yml

@@ -1,4 +1,5 @@
 base_model: mosaicml/mpt-7b
+base_model_config: mosaicml/mpt-7b
 tokenizer_type: AutoTokenizer
 trust_remote_code: true  # required for mpt as their model class is not merged into transformers yet
 load_in_8bit: false
@@ -8,7 +9,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.02
 adapter:
-lora_model_dir:
+peft_model_dir:
 sequence_len: 2048
 max_packed_sequence_len:
 lora_r: 8
@@ -26,7 +27,7 @@ wandb_log_model:
 output_dir: ./mpt-alpaca-7b
 gradient_accumulation_steps: 1
 micro_batch_size: 1
-num_epochs: 4
+num_epochs: 3
 optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine

examples/openllama-3b/config.yml

@@ -1,4 +1,5 @@
 base_model: openlm-research/open_llama_3b_v2
+base_model_config: openlm-research/open_llama_3b_v2
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 load_in_8bit: false
@@ -11,7 +12,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.02
 adapter:
-lora_model_dir:
+peft_model_dir:
 sequence_len: 1024
 sample_packing: true
 lora_r:

examples/openllama-3b/lora.yml

@@ -1,4 +1,5 @@
 base_model: openlm-research/open_llama_3b_v2
+base_model_config: openlm-research/open_llama_3b_v2
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 load_in_8bit: true
@@ -11,7 +12,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.02
 adapter: lora
-lora_model_dir:
+peft_model_dir:
 sequence_len: 1024
 sample_packing: true
 lora_r: 8

examples/openllama-3b/qlora.yml

@@ -1,4 +1,5 @@
 base_model: openlm-research/open_llama_3b_v2
+base_model_config: openlm-research/open_llama_3b_v2
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 load_in_8bit: false
@@ -11,7 +12,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.01
 adapter: qlora
-lora_model_dir:
+peft_model_dir:
 sequence_len: 1024
 sample_packing: true
 lora_r: 8

examples/phi/phi-ft.yml

@@ -1,4 +1,5 @@
 base_model: microsoft/phi-1_5
+base_model_config: microsoft/phi-1_5
 model_type: MixFormerSequentialForCausalLM
 tokenizer_type: AutoTokenizer
 is_llama_derived_model: false
@@ -21,7 +22,7 @@ sample_packing: true
 pad_to_sequence_len:
 
 adapter:
-lora_model_dir:
+peft_model_dir:
 lora_r:
 lora_alpha:
 lora_dropout:

examples/phi/phi-qlora.yml

@@ -1,4 +1,5 @@
 base_model: microsoft/phi-1_5
+base_model_config: microsoft/phi-1_5
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
 is_llama_derived_model: false
@@ -21,7 +22,7 @@ sample_packing: false  # not CURRENTLY compatible with LoRAs
 pad_to_sequence_len:
 
 adapter: qlora
-lora_model_dir:
+peft_model_dir:
 lora_r: 64
 lora_alpha: 32
 lora_dropout: 0.05

examples/pythia-12b/config.yml

@@ -1,4 +1,5 @@
 base_model: EleutherAI/pythia-12b-deduped
+base_model_config: EleutherAI/pythia-12b-deduped
 base_model_ignore_patterns: pytorch*  # prefer safetensors
 model_type: GPTNeoXForCausalLM
 tokenizer_type: AutoTokenizer
@@ -12,7 +13,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.05
 adapter:
-lora_model_dir:
+peft_model_dir:
 sequence_len: 2048
 max_packed_sequence_len: 2048
 lora_r: 64

examples/pythia/lora.yml

@@ -1,4 +1,5 @@
 base_model: EleutherAI/pythia-1.4b-deduped
+base_model_config: EleutherAI/pythia-1.4b-deduped
 load_in_8bit: true
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
@@ -6,7 +7,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.05
 adapter: lora
-lora_model_dir:
+peft_model_dir:
 sequence_len: 512
 lora_r: 16
 lora_alpha: 32
@@ -23,7 +24,7 @@ wandb_log_model:
 output_dir: ./lora-alpaca-pythia
 gradient_accumulation_steps: 1
 micro_batch_size: 4
-num_epochs: 4
+num_epochs: 3
 learning_rate: 0.00001
 train_on_inputs: false
 group_by_length: false
@@ -33,5 +34,5 @@ early_stopping_patience:
 resume_from_checkpoint:
 local_rank:
 weight_decay: 0.1
-eval_steps: 0.05
+eval_steps: 20
 logging_steps: 1

examples/redpajama/config-3b.yml

@@ -1,4 +1,5 @@
 base_model: togethercomputer/RedPajama-INCITE-Chat-3B-v1
+base_model_config: togethercomputer/RedPajama-INCITE-Chat-3B-v1
 model_type: GPTNeoXForCausalLM
 tokenizer_type: AutoTokenizer
 trust_remote_code:
@@ -9,7 +10,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.02
 adapter:
-lora_model_dir:
+peft_model_dir:
 sequence_len: 2048
 max_packed_sequence_len:
 lora_r: 8
@@ -27,7 +28,7 @@ wandb_log_model:
 output_dir: ./redpajama-alpaca-3b
 batch_size: 4
 micro_batch_size: 1
-num_epochs: 4
+num_epochs: 3
 optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine

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

@@ -1,4 +1,5 @@
 base_model: replit/replit-code-v1-3b
+base_model_config: replit/replit-code-v1-3b
 trust_remote_code: true
 load_in_8bit: false
 datasets:
@@ -7,7 +8,7 @@ datasets:
 dataset_prepared_path:
 val_set_size: 0.05
 adapter: lora
-lora_model_dir:
+peft_model_dir:
 sequence_len: 2048
 max_packed_sequence_len:
 lora_r: 8
@@ -26,7 +27,7 @@ wandb_log_model:
 output_dir: ./lora-replit
 batch_size: 8
 micro_batch_size: 1
-num_epochs: 4
+num_epochs: 3
 optimizer:
 torchdistx_path:
 lr_scheduler:

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

@@ -1,6 +1,7 @@
 # An example finetuning Saleforce's XGen-7b model with 8k context using qlora
 # on Tim Dettmer's Guanaco dataset.
 base_model: Salesforce/xgen-7b-8k-base
+base_model_config: Salesforce/xgen-7b-8k-base
 trust_remote_code: true
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
@@ -19,7 +20,7 @@ dataset_prepared_path:
 val_set_size: 0.01
 # enable QLoRA
 adapter: qlora
-lora_model_dir:
+peft_model_dir:
 sequence_len: 8192
 max_packed_sequence_len:
 
@@ -51,7 +52,7 @@ output_dir: ./qlora-out
 # decrease if OOM, increase for max VRAM utilization
 micro_batch_size: 1
 gradient_accumulation_steps: 1
-num_epochs: 4
+num_epochs: 3
 # Optimizer for QLoRA
 optimizer: paged_adamw_32bit
 torchdistx_path:

gitbook/README.md

@@ -1 +0,0 @@
-# Page

gitbook/SUMMARY.md

@@ -1,4 +0,0 @@
-# Table of contents
-
-* [Page](README.md)
-* [Small dev details](small-dev-details.md)

gitbook/small-dev-details.md

@@ -1,3 +0,0 @@
-# Small dev details
-
-/

requirements.txt

@@ -4,7 +4,7 @@ torch==2.0.1
 auto-gptq
 packaging
 peft @ git+https://github.com/huggingface/peft.git
-transformers @ git+https://github.com/huggingface/transformers.git@acc394c4f5e1283c19783581790b3dc3105a3697
+transformers @ git+https://github.com/huggingface/transformers.git@bd6205919aad4d3a2300a39a98a642f1cc3a5348
 bitsandbytes>=0.41.1
 accelerate @ git+https://github.com/huggingface/accelerate@80da9cfb09bb3cc9f1b385cb55d6b90d025a5fd9
 deepspeed
@@ -31,4 +31,3 @@ scikit-learn==1.2.2
 pynvml
 art
 fschat==0.2.29
-tensor_parallel

scripts/finetune.py

@@ -45,6 +45,8 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
         shard(cfg=parsed_cfg, cli_args=parsed_cli_args)
     else:
         dataset_meta = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+        if parsed_cli_args.prepare_ds_only:
+            return
         train(cfg=parsed_cfg, cli_args=parsed_cli_args, dataset_meta=dataset_meta)
 
 

src/axolotl/cli/__init__.py

@@ -222,9 +222,7 @@ def load_datasets(
 ) -> TrainDatasetMeta:
     tokenizer = load_tokenizer(cfg)
 
-    train_dataset, eval_dataset, total_num_steps, prompters = prepare_dataset(
-        cfg, tokenizer
-    )
+    train_dataset, eval_dataset, total_num_steps = prepare_dataset(cfg, tokenizer)
 
     if cli_args.debug or cfg.debug:
         LOG.info("check_dataset_labels...")
@@ -240,10 +238,6 @@ def load_datasets(
             text_only=cli_args.debug_text_only,
         )
 
-        LOG.info("printing prompters...")
-        for prompter in prompters:
-            LOG.info(prompter)
-
     return TrainDatasetMeta(
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,

src/axolotl/cli/preprocess.py

@@ -1,53 +0,0 @@
-"""
-CLI to run training on a model
-"""
-import logging
-from pathlib import Path
-
-import fire
-import transformers
-from colorama import Fore
-
-from axolotl.cli import (
-    check_accelerate_default_config,
-    check_user_token,
-    load_cfg,
-    load_datasets,
-    print_axolotl_text_art,
-)
-from axolotl.common.cli import PreprocessCliArgs
-from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
-
-LOG = logging.getLogger("axolotl.cli.preprocess")
-
-
-def do_cli(config: Path = Path("examples/"), **kwargs):
-    # pylint: disable=duplicate-code
-    print_axolotl_text_art()
-    parsed_cfg = load_cfg(config, **kwargs)
-    check_accelerate_default_config()
-    check_user_token()
-    parser = transformers.HfArgumentParser((PreprocessCliArgs))
-    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
-        return_remaining_strings=True
-    )
-    if not parsed_cfg.dataset_prepared_path:
-        msg = (
-            Fore.RED
-            + "preprocess CLI called without dataset_prepared_path set, "
-            + f"using default path: {DEFAULT_DATASET_PREPARED_PATH}"
-            + Fore.RESET
-        )
-        LOG.warning(msg)
-        parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
-
-    _ = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
-    LOG.info(
-        Fore.GREEN
-        + f"Success! Preprocessed data path: `dataset_prepared_path: {parsed_cfg.dataset_prepared_path}`"
-        + Fore.RESET
-    )
-
-
-if __name__ == "__main__":
-    fire.Fire(do_cli)

src/axolotl/cli/train.py

@@ -6,6 +6,7 @@ from pathlib import Path
 
 import fire
 import transformers
+from colorama import Fore
 
 from axolotl.cli import (
     check_accelerate_default_config,
@@ -15,6 +16,7 @@ from axolotl.cli import (
     print_axolotl_text_art,
 )
 from axolotl.common.cli import TrainerCliArgs
+from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.train import train
 
 LOG = logging.getLogger("axolotl.cli.train")
@@ -30,7 +32,18 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True
     )
+    if parsed_cli_args.prepare_ds_only and not parsed_cfg.dataset_prepared_path:
+        msg = (
+            Fore.RED
+            + "--prepare_ds_only called without dataset_prepared_path set."
+            + Fore.RESET
+        )
+        LOG.warning(msg)
+        parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
+
     dataset_meta = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    if parsed_cli_args.prepare_ds_only:
+        return
     train(cfg=parsed_cfg, cli_args=parsed_cli_args, dataset_meta=dataset_meta)
 
 

src/axolotl/common/cli.py

@@ -25,22 +25,11 @@ class TrainerCliArgs:
     debug_num_examples: int = field(default=5)
     inference: bool = field(default=False)
     merge_lora: bool = field(default=False)
+    prepare_ds_only: bool = field(default=False)
     prompter: Optional[str] = field(default=None)
     shard: bool = field(default=False)
 
 
-@dataclass
-class PreprocessCliArgs:
-    """
-    dataclass representing arguments for preprocessing only
-    """
-
-    debug: bool = field(default=False)
-    debug_text_only: bool = field(default=False)
-    debug_num_examples: int = field(default=1)
-    prompter: Optional[str] = field(default=None)
-
-
 def load_model_and_tokenizer(
     *,
     cfg: DictDefault,

src/axolotl/core/trainer_builder.py

@@ -1,711 +0,0 @@
-"""
-Builder for the training args and trainer
-"""
-
-import abc
-import importlib
-import logging
-import math
-import os
-import sys
-from abc import abstractmethod
-from dataclasses import dataclass, field
-from functools import partial
-from pathlib import Path
-from typing import Optional, Union
-
-import tensor_parallel as tp
-import torch
-import transformers
-from datasets import Dataset
-from torch.optim.lr_scheduler import OneCycleLR
-from torch.utils.data import DataLoader, DistributedSampler, SequentialSampler
-from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
-from transformers.trainer_pt_utils import SequentialDistributedSampler
-
-from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
-from axolotl.utils.callbacks import (
-    EvalFirstStepCallback,
-    GPUStatsCallback,
-    SaveAxolotlConfigtoWandBCallback,
-    SaveBetterTransformerModelCallback,
-    bench_eval_callback_factory,
-    log_prediction_callback_factory,
-)
-from axolotl.utils.collators import DataCollatorForSeq2Seq
-from axolotl.utils.dataloader import MultipackDistributedDataloader
-from axolotl.utils.distributed import is_distributed
-from axolotl.utils.schedulers import get_cosine_schedule_with_quadratic_warmup
-
-try:
-    import torch._dynamo  # pylint: disable=ungrouped-imports
-except ImportError:
-    pass
-
-LOG = logging.getLogger("axolotl.core.trainer_builder")
-
-
-@dataclass
-class AxolotlTrainingArguments(TrainingArguments):
-    """
-    Extend the base TrainingArguments for axolotl helpers
-    """
-
-    lr_quadratic_warmup: bool = field(
-        default=False,
-        metadata={"help": "Use quadratic warmup for cosine scheduling."},
-    )
-    sample_packing: bool = field(
-        default=False,
-        metadata={"help": "Use sample packing for efficient training."},
-    )
-    eval_sample_packing: Optional[bool] = field(
-        default=None,
-        metadata={"help": "Use sample packing for efficient evals."},
-    )
-    sample_packing_efficiency: float = field(
-        default=1.0,
-        metadata={"help": "Sample packing efficiency for calculating batch length."},
-    )
-    max_seq_length: int = field(
-        default=2048,
-        metadata={"help": "The maximum sequence length the model can handle"},
-    )
-    sample_packing_seq_len_multiplier: int = field(
-        default=1,
-        metadata={"help": "the multiplier for the max len for packed sequences"},
-    )
-    relora_steps: Optional[int] = field(
-        default=None,
-        metadata={"help": "how often to reset for ReLoRA"},
-    )
-    relora_warmup_steps: Optional[int] = field(
-        default=None,
-        metadata={"help": "how many warmup steps to take after reset for ReLoRA"},
-    )
-    bench_split: Optional[str] = field(
-        default="eval", metadata={"help": "The benchmark split to run on"}
-    )
-    bench_dataset: Optional[str] = field(
-        default="pharaouk/dharma-1/dharma_1_mini.json",
-        metadata={
-            "help": "Benchmark dataset to use: options are `mmlu-zs`, `mmlu-fs`, or the full path to the dataset file"
-        },
-    )
-    do_bench_eval: Optional[bool] = field(
-        default=False, metadata={"help": "Whether to run the Benchmark evaluation."}
-    )
-    max_bench_samples: Optional[int] = field(
-        default=None,
-        metadata={
-            "help": "If set, only evaluates on `max_bench_samples` of the benchmark dataset."
-        },
-    )
-    bench_source_max_len: int = field(
-        default=2048, metadata={"help": "Maximum source sequence length for bench."}
-    )
-    tensor_parallel: bool = field(
-        default=False, metadata={"help": "Use tensor parallelism to train"}
-    )
-
-
-class AxolotlTrainer(Trainer):
-    """
-    Extend the base Trainer for axolotl helpers
-    """
-
-    args = None  # type: AxolotlTrainingArguments
-
-    def __init__(self, *args, num_epochs=1, bench_data_collator=None, **kwargs):
-        self.num_epochs = num_epochs
-        self.bench_data_collator = bench_data_collator
-        super().__init__(*args, **kwargs)
-
-    def create_scheduler(
-        self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
-    ):
-        """
-        Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or
-        passed as an argument.
-
-        Args:
-            num_training_steps (int): The number of training steps to do.
-            optimizer (torch.optim.Optimizer): The training optimizer
-        """
-
-        # fmt: off
-        if self.lr_scheduler is None:  # type: ignore  # pylint: disable=access-member-before-definition
-            # fmt: on
-            if (
-                self.args.lr_scheduler_type == "cosine"
-                and self.args.lr_quadratic_warmup is True
-            ):
-                self.lr_scheduler = get_cosine_schedule_with_quadratic_warmup(  # pylint: disable=attribute-defined-outside-init
-                    optimizer,
-                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
-                    num_training_steps=num_training_steps,
-                )
-            else:
-                return super().create_scheduler(num_training_steps, optimizer)
-        return self.lr_scheduler
-
-    def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
-        if self.args.world_size > 1 and self.args.sample_packing:
-            return DistributedSampler(
-                self.train_dataset,
-                num_replicas=self.args.world_size,
-                rank=self.args.process_index,
-                seed=self.args.seed,
-            )
-        return super()._get_train_sampler()
-
-    def _get_eval_sampler(
-        self, eval_dataset: Dataset
-    ) -> Optional[torch.utils.data.Sampler]:
-        if (
-            self.args.world_size > 1
-            and self.args.sample_packing
-            and self.args.eval_sample_packing is not False
-        ):
-            return SequentialDistributedSampler(
-                eval_dataset,
-                num_replicas=self.args.world_size,
-                rank=self.args.process_index,
-                batch_size=self.args.per_device_eval_batch_size,
-            )
-        return super()._get_eval_sampler(eval_dataset)
-
-    def get_train_dataloader(self) -> Union[DataLoader, MultipackDistributedDataloader]:
-        if self.args.sample_packing:
-            train_sampler = self._get_train_sampler()
-            return self.accelerator.prepare(
-                MultipackDistributedDataloader(
-                    self.train_dataset,
-                    batch_size=self._train_batch_size,
-                    seq_max_length=self.args.max_seq_length,
-                    collate_fn=self.data_collator,
-                    sampler=train_sampler,
-                    packing_efficiency_estimate=self.args.sample_packing_efficiency,
-                    sample_packing_seq_len_multiplier=self.args.sample_packing_seq_len_multiplier,
-                    device_count=int(os.environ.get("WORLD_SIZE", 1)),
-                    num_epochs=self.num_epochs,
-                )
-            )
-        return super().get_train_dataloader()
-
-    def get_eval_dataloader(
-        self, eval_dataset: Optional[Dataset] = None
-    ) -> Union[DataLoader, MultipackDistributedDataloader]:
-        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
-            )
-
-            eval_sampler = self._get_eval_sampler(eval_dataset)
-            return self.accelerator.prepare(
-                MultipackDistributedDataloader(
-                    eval_dataset,
-                    batch_size=self.args.eval_batch_size,
-                    seq_max_length=self.args.max_seq_length,
-                    collate_fn=self.data_collator,
-                    sampler=eval_sampler,
-                    packing_efficiency_estimate=self.args.sample_packing_efficiency,
-                    sample_packing_seq_len_multiplier=self.args.eval_batch_size,
-                    device_count=int(os.environ.get("WORLD_SIZE", 1)),
-                    num_epochs=self.num_epochs,
-                )
-            )
-        return super().get_eval_dataloader(eval_dataset)
-
-    def _get_bench_sampler(
-        self, bench_dataset: Dataset
-    ) -> Optional[torch.utils.data.Sampler]:
-        if self.args.world_size <= 1:
-            return SequentialSampler(bench_dataset)
-        return None
-
-    def get_bench_dataloader(
-        self,
-        bench_dataset: Dataset,
-    ) -> Union[DataLoader, MultipackDistributedDataloader]:
-        dataloader_params = {
-            "batch_size": self.args.eval_batch_size,
-            "collate_fn": self.bench_data_collator,
-            "num_workers": self.args.dataloader_num_workers,
-            "pin_memory": self.args.dataloader_pin_memory,
-        }
-
-        if not isinstance(bench_dataset, torch.utils.data.IterableDataset):
-            dataloader_params["sampler"] = self._get_bench_sampler(bench_dataset)
-            dataloader_params["drop_last"] = self.args.dataloader_drop_last
-
-        return DataLoader(bench_dataset, **dataloader_params)
-        # return self.accelerator.prepare(DataLoader(bench_dataset, **dataloader_params))
-
-    def compute_loss(self, model, inputs, return_outputs=False):
-        # use one's weighted cross entropy loss calc
-        # if self.args.sample_packing:
-        #     labels = inputs.pop("labels")
-        #     outputs = model(**inputs)
-        #     loss = trainer_weighted_loss(outputs, labels, shift_labels=True)
-        #     return (loss, outputs) if return_outputs else loss
-        return super().compute_loss(model, inputs, return_outputs=return_outputs)
-
-    def _wrap_model(self, model, training=True, dataloader=None):
-        if self.args.tensor_parallel:
-            model = tp.tensor_parallel(model, distributed=is_distributed())
-            model.hf_device_map = tp.infer_sharded_device_map(model)
-        else:
-            model = super()._wrap_model(model, training=training, dataloader=dataloader)
-        return model
-
-
-class OneCycleLRSchedulerTrainer(AxolotlTrainer):
-    """
-    Trainer subclass that uses the OneCycleLR scheduler
-    """
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.lr_scheduler = None
-
-    def create_scheduler(
-        self,
-        num_training_steps: int,
-        optimizer: Optional[torch.optim.Optimizer] = None,
-    ):
-        optimizer = self.optimizer if optimizer is None else optimizer
-        num_warmup_steps = self.args.get_warmup_steps(num_training_steps)
-        pct_start = num_warmup_steps / num_training_steps
-
-        self.lr_scheduler = OneCycleLR(
-            optimizer,
-            max_lr=self.args.learning_rate,
-            total_steps=num_training_steps,
-            pct_start=pct_start,
-            div_factor=6,
-        )
-
-        return self.lr_scheduler
-
-
-class ReLoRATrainer(AxolotlTrainer):
-    """
-    Trainer subclass that uses the OneCycleLR scheduler
-    """
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.lr_scheduler = None
-
-    def create_scheduler(
-        self,
-        num_training_steps: int,
-        optimizer: Optional[torch.optim.Optimizer] = None,
-    ):
-        optimizer = self.optimizer if optimizer is None else optimizer
-        lr_scheduler = super().create_scheduler(num_training_steps, optimizer)
-
-        if self.args.relora_steps:
-            warmup_steps = (
-                self.args.relora_warmup_steps if self.args.relora_warmup_steps else 10
-            )
-            self.lr_scheduler = ReLoRAScheduler(
-                optimizer,
-                lr_scheduler,
-                self.args.relora_steps,
-                warmup_steps,
-            )
-        else:
-            self.lr_scheduler = lr_scheduler
-
-        return self.lr_scheduler
-
-
-class TrainerBuilderBase(abc.ABC):
-    """
-    Base class for trainer builder
-    """
-
-    _train_dataset = None
-    _eval_dataset = None
-
-    def __init__(self, cfg, model, tokenizer):
-        self.cfg = cfg
-        self.model = model
-        self.tokenizer = tokenizer
-
-    @property
-    def train_dataset(self):
-        return self._train_dataset
-
-    @train_dataset.setter
-    def train_dataset(self, dataset):
-        self._train_dataset = dataset
-
-    @property
-    def eval_dataset(self):
-        return self._eval_dataset
-
-    @eval_dataset.setter
-    def eval_dataset(self, dataset):
-        self._eval_dataset = dataset
-
-    @abstractmethod
-    def build(self, total_num_steps):
-        pass
-
-    @abstractmethod
-    def get_callbacks(self):
-        pass
-
-    @abstractmethod
-    def get_post_trainer_create_callbacks(self, trainer):
-        """
-        Callbacks added after the trainer is created, usually b/c these need access to the trainer
-        """
-
-
-class HFCausalTrainerBuilder(TrainerBuilderBase):
-    """
-    Build the HuggingFace training args/trainer for Causal models
-    """
-
-    def hook_pre_create_training_args(self, training_arguments_kwargs):
-        # TODO
-        return training_arguments_kwargs
-
-    def hook_post_create_training_args(self, training_arguments):
-        # TODO
-        return training_arguments
-
-    def hook_pre_create_trainer(self, trainer_kwargs, trainer_cls):
-        # TODO
-        return trainer_kwargs, trainer_cls
-
-    def hook_post_create_trainer(self, trainer):
-        if self.cfg.tensor_parallel:
-            trainer.model = trainer.accelerator.prepare_model(
-                trainer.model, device_placement=True
-            )
-        return trainer
-
-    def get_callbacks(self):
-        callbacks = []
-        callbacks.append(GPUStatsCallback(self.cfg))
-        callbacks.append(EvalFirstStepCallback)
-
-        if self.cfg.relora_steps:
-            callbacks.append(ReLoRACallback(self.cfg))
-
-        if (
-            hasattr(self.model, "use_bettertransformer")
-            and self.model.use_bettertransformer is True
-        ):
-            callbacks.append(SaveBetterTransformerModelCallback)
-
-        if self.cfg.use_wandb:
-            callbacks.append(
-                SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
-            )
-
-        return callbacks
-
-    def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = []
-        if self.cfg.use_wandb and self.cfg.eval_table_size > 0:
-            LogPredictionCallback = log_prediction_callback_factory(
-                trainer, self.tokenizer
-            )
-            callbacks.append(LogPredictionCallback(self.cfg))
-
-        if self.cfg.do_bench_eval:
-            callbacks.append(bench_eval_callback_factory(trainer, self.tokenizer))
-
-        if self.cfg.early_stopping_patience:
-            early_stop_cb = EarlyStoppingCallback(
-                self.cfg.early_stopping_patience,
-            )
-            callbacks.append(early_stop_cb)
-
-        return callbacks
-
-    def _get_trainer_cls(self):
-        if self.cfg.lr_scheduler == "one_cycle" and (
-            self.cfg.fsdp or self.cfg.adapter == "qlora"
-        ):
-            return OneCycleLRSchedulerTrainer
-        if self.cfg.relora_steps:
-            return ReLoRATrainer
-        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)
-        )
-        logging_steps = (
-            self.cfg.logging_steps
-            if self.cfg.logging_steps is not None
-            else max(min(int(0.005 * total_num_steps), 10), 1)
-        )
-
-        training_arguments_kwargs = {}
-        if self.cfg.bf16 == "full":
-            training_arguments_kwargs["bf16_full_eval"] = True
-        else:
-            training_arguments_kwargs["bf16"] = self.cfg.bf16
-        training_arguments_kwargs["fp16"] = (
-            self.cfg.fp16 and not self.cfg.bf16
-        ) or False
-        training_arguments_kwargs["tf32"] = self.cfg.tf32
-        training_arguments_kwargs["warmup_steps"] = warmup_steps
-        training_arguments_kwargs["logging_steps"] = logging_steps
-
-        if self.cfg.seed:
-            training_arguments_kwargs["seed"] = self.cfg.seed
-
-        if self.cfg.gradient_checkpointing:
-            training_arguments_kwargs[
-                "gradient_checkpointing"
-            ] = self.cfg.gradient_checkpointing
-        if self.cfg.fsdp:
-            training_arguments_kwargs["fsdp"] = self.cfg.fsdp
-            if self.cfg.fsdp_config:
-                training_arguments_kwargs["fsdp_config"] = dict(self.cfg.fsdp_config)
-
-        # deepspeed
-        if self.cfg.deepspeed:
-            training_arguments_kwargs["deepspeed"] = self.cfg.deepspeed
-
-        if self.cfg.lr_quadratic_warmup is not None:
-            training_arguments_kwargs[
-                "lr_quadratic_warmup"
-            ] = self.cfg.lr_quadratic_warmup
-
-        if self.cfg.adam_beta1:
-            training_arguments_kwargs["adam_beta1"] = self.cfg.adam_beta1
-        if self.cfg.adam_beta2:
-            training_arguments_kwargs["adam_beta2"] = self.cfg.adam_beta2
-        if self.cfg.adam_epsilon:
-            training_arguments_kwargs["adam_epsilon"] = self.cfg.adam_epsilon
-        if self.cfg.max_grad_norm:
-            training_arguments_kwargs["max_grad_norm"] = self.cfg.max_grad_norm
-
-        if self.cfg.hub_model_id:
-            training_arguments_kwargs["hub_model_id"] = self.cfg.hub_model_id
-            training_arguments_kwargs["push_to_hub"] = True
-            training_arguments_kwargs["hub_private_repo"] = True
-
-            if self.cfg.hub_strategy:
-                training_arguments_kwargs["hub_strategy"] = self.cfg.hub_strategy
-
-        if self.cfg.save_safetensors:
-            training_arguments_kwargs["save_safetensors"] = self.cfg.save_safetensors
-
-        if self.cfg.sample_packing_eff_est:
-            training_arguments_kwargs[
-                "sample_packing_efficiency"
-            ] = self.cfg.sample_packing_eff_est
-
-        if self.cfg.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"
-
-        if self.cfg.save_steps:
-            training_arguments_kwargs["save_strategy"] = "steps"
-            training_arguments_kwargs["save_steps"] = self.cfg.save_steps
-        elif self.cfg.save_strategy:
-            training_arguments_kwargs["save_strategy"] = self.cfg.save_strategy
-        else:
-            # default to saving each epoch if not defined
-            training_arguments_kwargs["save_strategy"] = "epoch"
-
-        if self.cfg.do_bench_eval:
-            training_arguments_kwargs["do_bench_eval"] = self.cfg.do_bench_eval
-            if self.cfg.bench_dataset:
-                training_arguments_kwargs["bench_dataset"] = self.cfg.bench_dataset
-        if self.cfg.metric_for_best_model:
-            training_arguments_kwargs[
-                "metric_for_best_model"
-            ] = self.cfg.metric_for_best_model
-        if self.cfg.greater_is_better:
-            training_arguments_kwargs["greater_is_better"] = self.cfg.greater_is_better
-
-        if self.cfg.torch_compile:
-            if torch.__version__ < "2.1.0":  # pylint: disable=protected-access
-                LOG.warning("torch>=2.1.0 required for torch_compile to work properly")
-            elif torch._dynamo:  # pylint: disable=protected-access
-                torch._dynamo.config.suppress_errors = (  # pylint: disable=protected-access
-                    True
-                )
-                training_arguments_kwargs["torch_compile"] = self.cfg.torch_compile
-                if self.cfg.torch_compile_backend:
-                    training_arguments_kwargs[
-                        "torch_compile_backend"
-                    ] = self.cfg.torch_compile_backend
-
-        # DDP Config
-        if self.cfg.ddp_timeout:
-            training_arguments_kwargs["ddp_timeout"] = self.cfg.ddp_timeout
-        # see https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html
-        if self.cfg.ddp_bucket_cap_mb:
-            training_arguments_kwargs["ddp_bucket_cap_mb"] = self.cfg.ddp_bucket_cap_mb
-        if self.cfg.ddp_broadcast_buffers is not None:
-            training_arguments_kwargs[
-                "ddp_broadcast_buffers"
-            ] = self.cfg.ddp_broadcast_buffers
-
-        # these are all the "standard" kwargs that are def used
-        training_arguments_kwargs["max_steps"] = (
-            total_num_steps if self.cfg.max_steps else -1
-        )
-        training_arguments_kwargs["max_seq_length"] = self.cfg.sequence_len
-        training_arguments_kwargs[
-            "per_device_train_batch_size"
-        ] = self.cfg.micro_batch_size
-        training_arguments_kwargs[
-            "per_device_eval_batch_size"
-        ] = self.cfg.eval_batch_size
-        training_arguments_kwargs[
-            "gradient_accumulation_steps"
-        ] = self.cfg.gradient_accumulation_steps
-        training_arguments_kwargs[
-            "eval_accumulation_steps"
-        ] = self.cfg.gradient_accumulation_steps
-        training_arguments_kwargs["num_train_epochs"] = self.cfg.num_epochs
-        training_arguments_kwargs["learning_rate"] = self.cfg.learning_rate
-        training_arguments_kwargs["output_dir"] = self.cfg.output_dir
-        training_arguments_kwargs["save_total_limit"] = (
-            self.cfg.save_total_limit if self.cfg.save_total_limit else 4
-        )
-        training_arguments_kwargs["load_best_model_at_end"] = (
-            (
-                self.cfg.load_best_model_at_end is not False
-                or self.cfg.early_stopping_patience
-            )
-            and self.cfg.val_set_size > 0
-            and self.cfg.save_steps
-            and self.cfg.eval_steps
-            and self.cfg.save_steps % self.cfg.eval_steps == 0
-        ) or False
-        training_arguments_kwargs["ddp_find_unused_parameters"] = (
-            False if self.cfg.ddp else None
-        )
-        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
-        )
-        training_arguments_kwargs["optim"] = (
-            self.cfg.optimizer if self.cfg.optimizer else "adamw_hf"
-        )
-        training_arguments_kwargs["lr_scheduler_type"] = (
-            self.cfg.lr_scheduler
-            if self.cfg.lr_scheduler
-            and self.cfg.lr_scheduler not in ("one_cycle", "log_sweep")
-            else "cosine"
-        )
-        training_arguments_kwargs["weight_decay"] = (
-            self.cfg.weight_decay if self.cfg.weight_decay is not None else 0.0
-        )
-        training_arguments_kwargs["sample_packing"] = (
-            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
-        )
-        training_arguments_kwargs[
-            "sample_packing_seq_len_multiplier"
-        ] = self.cfg.micro_batch_size
-        training_arguments_kwargs["relora_steps"] = self.cfg.relora_steps
-        training_arguments_kwargs["relora_warmup_steps"] = self.cfg.relora_warmup_steps
-        training_arguments_kwargs["tensor_parallel"] = self.cfg.tensor_parallel is True
-
-        training_arguments_kwargs = self.hook_pre_create_training_args(
-            training_arguments_kwargs
-        )
-        training_args = (
-            AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
-                **training_arguments_kwargs,
-            )
-        )
-        training_args = self.hook_post_create_training_args(training_args)
-        trainer_kwargs = {}
-
-        if self.cfg.optimizer == "adamw_anyprecision":
-            if Path(self.cfg.torchdistx_path).exists():
-                sys.path.append(self.cfg.torchdistx_path)
-                importlib.import_module("torchdistx")
-
-        data_collator_kwargs = {
-            "padding": True,  # True/"longest" is the default
-        }
-        if self.cfg.pad_to_sequence_len:
-            data_collator_kwargs["pad_to_multiple_of"] = 64 * math.ceil(
-                self.cfg.sequence_len / 64
-            )
-        else:
-            # A100 is best at 64, while others at 8. Let's use the larger so we don't have to check
-            # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
-            data_collator_kwargs["pad_to_multiple_of"] = 64
-
-        if self.cfg.is_llama_derived_model and self.cfg.landmark_attention:
-            from axolotl.monkeypatch.llama_landmark_attn import (
-                add_mem_tokens,
-                get_mem_id,
-                set_model_mem_id,
-            )
-
-            set_model_mem_id(self.model, self.tokenizer)
-
-            LOG.info("Adding landmark attention tokens to dataset")
-
-            for dataset in [self.train_dataset, self.eval_dataset]:
-                dataset = dataset.map(
-                    partial(
-                        add_mem_tokens, mem_freq=50, mem_id=get_mem_id(self.tokenizer)
-                    ),
-                    batched=False,
-                    num_proc=32,
-                )
-
-        trainer_cls = self._get_trainer_cls()
-        trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
-            trainer_kwargs, trainer_cls
-        )
-        trainer = trainer_cls(
-            model=self.model,
-            train_dataset=self.train_dataset,
-            eval_dataset=self.eval_dataset,
-            args=training_args,
-            data_collator=DataCollatorForSeq2Seq(
-                self.tokenizer,
-                return_tensors="pt",
-                **data_collator_kwargs,
-            ),
-            bench_data_collator=transformers.DataCollatorForSeq2Seq(
-                self.tokenizer,
-                return_tensors="pt",
-                **data_collator_kwargs,
-            ),
-            callbacks=self.get_callbacks(),
-            num_epochs=self.cfg.num_epochs,
-            **trainer_kwargs,
-        )
-        trainer = self.hook_post_create_trainer(trainer)
-        for callback in self.get_post_trainer_create_callbacks(trainer):
-            trainer.add_callback(callback)
-
-        return trainer

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -13,18 +13,12 @@ import transformers
 from einops import rearrange
 from flash_attn.bert_padding import pad_input, unpad_input
 from transformers.modeling_outputs import BaseModelOutputWithPast
-from transformers.models.llama.modeling_llama import LlamaAttention
 from transformers.models.llama.modeling_llama import (
     LlamaDecoderLayer as OriginalLlamaDecoderLayer,
 )
-from transformers.models.llama.modeling_llama import (
-    LlamaMLP,
-    apply_rotary_pos_emb,
-    repeat_kv,
-)
-from xformers.ops import SwiGLU
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
 
-from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids, set_module_name
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
 
 try:
     from flash_attn.flash_attn_interface import (  # pylint: disable=ungrouped-imports
@@ -44,28 +38,6 @@ except ImportError:
 LOG = logging.getLogger("axolotl")
 
 
-def replace_llama_mlp_with_swiglu(model):
-    for name, module in model.named_modules():
-        if isinstance(module, LlamaMLP):
-            mlp = FusedMLP(
-                module.config, module.gate_proj, module.up_proj, module.down_proj
-            )
-            set_module_name(model, name, mlp)
-
-
-def replace_llama_qkv_with_fused(model):
-    for name, module in model.named_modules():
-        if isinstance(module, LlamaAttention):
-            qkv = FusedAttention(
-                module.config,
-                module.q_proj,
-                module.k_proj,
-                module.v_proj,
-                module.o_proj,
-            )
-            set_module_name(model, name, qkv)
-
-
 def replace_llama_attn_with_flash_attn(
     packed: Optional[bool] = False,
     cross_entropy: Optional[bool] = False,
@@ -114,92 +86,6 @@ def replace_llama_attn_with_flash_attn(
             )
 
 
-class FusedAttention(LlamaAttention):
-    """
-    Fused QKV Attention layer for incrementally improved training efficiency
-    """
-
-    def __init__(
-        self,
-        config,
-        q: torch.nn.Linear,  # pylint: disable=invalid-name
-        k: torch.nn.Linear,  # pylint: disable=invalid-name
-        v: torch.nn.Linear,  # pylint: disable=invalid-name
-        o: torch.nn.Linear,  # pylint: disable=invalid-name
-    ):
-        super().__init__(config)
-        self.config = config
-        self.init_device = next(iter(q.state_dict().values())).device
-
-        # define equivalent fused qkv projection
-        self.out_features: List[int] = [q.out_features, k.out_features, v.out_features]
-        self.qkv_proj = torch.nn.Linear(
-            q.in_features, sum(self.out_features), device=self.init_device, bias=False
-        )
-        self.o_proj = o
-
-        # overwrite initialized weights with pretrained weights
-        self.qkv_proj.weight.data = torch.cat(
-            (q.weight.data, k.weight.data, v.weight.data), dim=0
-        )
-
-    def _post_training(self, model, name):
-        q_proj, k_proj, v_proj = torch.split(
-            self.qkv_proj.weight.data, self.out_features, dim=0
-        )
-
-        new_attn = LlamaAttention(self.config)
-        new_attn.q_proj.weight.data = q_proj
-        new_attn.k_proj.weight.data = k_proj
-        new_attn.v_proj.weight.data = v_proj
-        new_attn.o_proj.weight.data = self.o_proj.weight.data
-
-        set_module_name(model, name, new_attn)
-
-
-class FusedMLP(torch.nn.Module):
-    """
-    Fused MLP layer for incrementally improved training efficiency
-    """
-
-    def __init__(
-        self,
-        config,
-        gate_proj: torch.nn.Linear,
-        up_proj: torch.nn.Linear,
-        down_proj: torch.nn.Linear,
-    ):
-        super().__init__()
-        self.config = config
-        self.swiglu = SwiGLU(
-            in_features=config.hidden_size,
-            hidden_features=config.intermediate_size,
-            bias=False,
-            _pack_weights=True,
-        )
-        # overwrite initialized weights with pretrained weights
-        self.swiglu.w12.weight.data = torch.cat(
-            (gate_proj.weight.data, up_proj.weight.data), dim=0
-        )
-        self.swiglu.w3.weight.data = down_proj.weight.data
-
-    def _post_training(self, model, name):
-        w1, w2 = torch.split(  # pylint: disable=invalid-name
-            self.swiglu.w12.weight.data, self.config.intermediate_size, dim=0
-        )
-
-        # Assign the split weights back to the original layers
-        new_mlp = LlamaMLP(self.config)
-        new_mlp.gate_proj.weight.data = w1
-        new_mlp.up_proj.weight.data = w2
-        new_mlp.down_proj.weight.data = self.swiglu.w3.weight.data
-
-        set_module_name(model, name, new_mlp)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:  # pylint: disable=invalid-name
-        return self.swiglu(x)
-
-
 # Disable the transformation of the attention mask in LlamaModel as the flash attention
 # requires the attention mask to be the same as the key_padding_mask
 def _prepare_decoder_attention_mask(
@@ -230,6 +116,8 @@ def flashattn_forward(
     attention_mask: [bsz, q_len]
     """
     # pylint: disable=duplicate-code
+    original_dtype = hidden_states.dtype
+
     bsz, q_len, _ = hidden_states.size()
 
     if not hasattr(self, "pretraining_tp"):
@@ -261,14 +149,16 @@ def flashattn_forward(
         value_states = torch.cat(value_states, dim=-1)
 
     else:
-        if isinstance(self, FusedAttention):
-            query_states, key_states, value_states = self.qkv_proj(hidden_states).split(
-                self.out_features, dim=-1
-            )
-        else:
-            query_states = self.q_proj(hidden_states)
-            key_states = self.k_proj(hidden_states)
-            value_states = self.v_proj(hidden_states)
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+    if query_states.dtype == torch.float32:
+        query_states = query_states.to(dtype=original_dtype)
+    if key_states.dtype == torch.float32:
+        key_states = key_states.to(dtype=original_dtype)
+    if value_states.dtype == torch.float32:
+        value_states = value_states.to(dtype=original_dtype)
 
     query_states = query_states.view(
         bsz, q_len, self.num_heads, self.head_dim
@@ -428,6 +318,10 @@ def flashattn_forward(
     else:
         attn_output = self.o_proj(attn_output)
 
+    # handle conversion back for IA3
+    if attn_output.dtype == torch.float32:
+        attn_output = attn_output.to(dtype=original_dtype)
+
     return attn_output, None, past_key_value
 
 
@@ -621,6 +515,7 @@ def llama_model_forward(
     )
 
     hidden_states = inputs_embeds
+    original_dtype = hidden_states.dtype
 
     if self.gradient_checkpointing and self.training:
         if use_cache:
@@ -678,6 +573,10 @@ def llama_model_forward(
 
         hidden_states = layer_outputs[0]
 
+        # handle conversion back for IA3
+        if hidden_states.dtype == torch.float32:
+            hidden_states = hidden_states.to(dtype=original_dtype)
+
         if use_cache:
             next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)
 

src/axolotl/monkeypatch/llama_embeddings_hijack.py

@@ -0,0 +1,40 @@
+"""
+patch to add noisy embeddings per https://arxiv.org/abs/2310.05914
+"""
+
+import torch
+import transformers.models.llama.modeling_llama
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+def replace_llama_embeddings_with_uniform_distribution(noise_alpha=5):
+    # pylint: disable=duplicate-code
+    def noised_embed(orig_embed, noise_alpha, model):
+        def new_func(input_ids):
+            # during training, we add noise to the embedding
+            # during generation, we don't add noise to the embedding
+            if model.training:
+                embed_init = orig_embed(input_ids)
+                dims = torch.tensor(embed_init.size(1) * embed_init.size(2))
+                mag_norm = noise_alpha / torch.sqrt(dims)
+                return embed_init + torch.zeros_like(embed_init).uniform_(
+                    -mag_norm, mag_norm
+                )
+            return orig_embed(input_ids)
+
+        return new_func
+
+    def post_init(orig_post_init):
+        def new_func(self):
+            orig_post_init(self)
+            self.embed_tokens.forward = noised_embed(
+                self.embed_tokens.forward, noise_alpha, self
+            )
+
+        return new_func
+
+    transformers.models.llama.modeling_llama.LlamaModel.post_init = post_init(
+        transformers.models.llama.modeling_llama.LlamaModel.post_init
+    )

src/axolotl/monkeypatch/mistral_embeddings_hijack.py

@@ -0,0 +1,40 @@
+"""
+patch to add noisy embeddings per https://arxiv.org/abs/2310.05914
+"""
+
+import torch
+import transformers.models.mistral.modeling_mistral
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+def replace_mistral_embeddings_with_uniform_distribution(noise_alpha=5):
+    # pylint: disable=duplicate-code
+    def noised_embed(orig_embed, noise_alpha, model):
+        def new_func(input_ids):
+            # during training, we add noise to the embedding
+            # during generation, we don't add noise to the embedding
+            if model.training:
+                embed_init = orig_embed(input_ids)
+                dims = torch.tensor(embed_init.size(1) * embed_init.size(2))
+                mag_norm = noise_alpha / torch.sqrt(dims)
+                return embed_init + torch.zeros_like(embed_init).uniform_(
+                    -mag_norm, mag_norm
+                )
+            return orig_embed(input_ids)
+
+        return new_func
+
+    def post_init(orig_post_init):
+        def new_func(self):
+            orig_post_init(self)
+            self.embed_tokens.forward = noised_embed(
+                self.embed_tokens.forward, noise_alpha, self
+            )
+
+        return new_func
+
+    transformers.models.mistral.modeling_mistral.MistralModel.post_init = post_init(
+        transformers.models.mistral.modeling_mistral.MistralModel.post_init
+    )

src/axolotl/monkeypatch/neft_embeddings.py

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

src/axolotl/monkeypatch/utils.py

@@ -101,16 +101,3 @@ def get_cu_seqlens_from_pos_ids(position_ids):
         max_seq_lens.append(max_seq_len)
 
     return torch.stack(results).to(dtype=torch.int32), torch.stack(max_seq_lens)
-
-
-def set_module_name(model, name, value):
-    if "." in name:
-        parent_name = name.rsplit(".", 1)[0]
-        child_name = name[len(parent_name) + 1 :]
-        parent = model.get_submodule(parent_name)
-    else:
-        parent_name = ""
-        parent = model
-        child_name = name
-
-    setattr(parent, child_name, value)

src/axolotl/prompt_strategies/sharegpt.py

@@ -24,7 +24,7 @@ def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
     )
     field_human = ds_cfg["field_human"] if ds_cfg and "field_human" in ds_cfg else None
     field_model = ds_cfg["field_model"] if ds_cfg and "field_model" in ds_cfg else None
-    strategy = SimpleShareGPTPromptTokenizingStrategy(
+    return SimpleShareGPTPromptTokenizingStrategy(
         ShareGPTPrompterV2(
             conversation=conversation,
             role_key_model=field_model,
@@ -34,9 +34,6 @@ def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
         cfg.train_on_inputs,
         cfg.sequence_len,
     )
-    if ds_cfg and "strict" in ds_cfg:
-        strategy.strict = ds_cfg["strict"]
-    return strategy
 
 
 def load_role(tokenizer, cfg):
@@ -62,26 +59,8 @@ class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
     basic sharegpt strategy to grab conversations from the sample row
     """
 
-    _strict = True
-
-    @property
-    def strict(self):
-        return self._strict
-
-    @strict.setter
-    def strict(self, strict):
-        self._strict = strict
-
     def get_conversation_thread(self, prompt):
-        conversations = prompt["conversations"]
-        if self.strict:
-            return conversations
-        # remap roles - allow for assistant turn
-        role_map = {"human": "human", "assistant": "gpt", "gpt": "gpt"}
-        turns = [
-            {"from": role_map[t["from"]], "value": t["value"]} for t in conversations
-        ]
-        return turns
+        return prompt["conversations"]
 
 
 class SimpleRoleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):

src/axolotl/prompt_tokenizers.py

@@ -45,8 +45,6 @@ class PromptTokenizingStrategy(abc.ABC):
         self.prompter = prompter
         self.tokenizer: PreTrainedTokenizer = tokenizer
         self.train_on_inputs = train_on_inputs
-        # sequence_len and max_length can be different for CompletionPromptTokenizingStrategy.
-        # TODO: Document how they are different.
         self.sequence_len = sequence_len
         self.max_length = sequence_len
 
@@ -61,31 +59,34 @@ class PromptTokenizingStrategy(abc.ABC):
     def _tokenize(
         self, prompt: str, add_eos_token: bool = True, strip_bos_token: bool = False
     ) -> BatchEncoding:
-        empty = BatchEncoding(data={"input_ids": [], "attention_mask": []})
+        result: BatchEncoding
         if not prompt:
             LOG.warning("Empty text requested for tokenization.")
-            return empty
-
-        result = self.tokenizer(
-            prompt,
-            truncation=True,
-            max_length=self.max_length,
-            padding=False,
-            return_tensors=None,
-        )
+            result = BatchEncoding(data={"input_ids": [], "attention_mask": []})
+        else:
+            result = self.tokenizer(
+                prompt,
+                truncation=True,
+                max_length=self.max_length,
+                padding=False,
+                return_tensors=None,
+            )
         if len(result["input_ids"]) == 0:
             LOG.warning("Tokenizer result is empty. You may want to audit your dataset")
-            return empty
-
         if (
-            result["input_ids"][-1] != self.tokenizer.eos_token_id
+            len(result["input_ids"]) > 0
+            and result["input_ids"][-1] != self.tokenizer.eos_token_id
             and len(result["input_ids"]) < self.max_length
             and add_eos_token
         ):
             result["input_ids"].append(self.tokenizer.eos_token_id)
             result["attention_mask"].append(1)
 
-        if result["input_ids"][0] == self.tokenizer.bos_token_id and strip_bos_token:
+        if (
+            len(result["input_ids"]) > 0
+            and result["input_ids"][0] == self.tokenizer.bos_token_id
+            and strip_bos_token
+        ):
             result["input_ids"] = result["input_ids"][1:]
             result["attention_mask"] = result["attention_mask"][1:]
 
@@ -121,7 +122,7 @@ class InstructionPromptTokenizingStrategy(PromptTokenizingStrategy):
         if not self.train_on_inputs:
             user_prompt_len = len(tokenized_prompt["input_ids"])
             # TODO this could be sped up using numpy array slicing
-            tokenized_prompt["labels"] = [IGNORE_INDEX] * user_prompt_len
+            tokenized_prompt["labels"] = [-100] * user_prompt_len
         tokenized_res_prompt = self._tokenize(
             response, strip_bos_token=True, add_eos_token=True
         )
@@ -245,7 +246,6 @@ class ReflectionPromptTokenizingStrategy(PromptTokenizingStrategy):
         raise NotImplementedError
 
     def tokenize_prompt(self, prompt):
-        # pylint: disable=duplicate-code
         (
             instruction,
             input,  # pylint: disable=redefined-builtin
@@ -270,7 +270,7 @@ class ReflectionPromptTokenizingStrategy(PromptTokenizingStrategy):
             user_prompt_len = len(tokenized_user_prompt["input_ids"])
             # TODO this could be sped up using numpy array slicing
             tokenized_full_prompt["labels"] = [
-                IGNORE_INDEX
+                -100
             ] * user_prompt_len + tokenized_full_prompt["labels"][user_prompt_len:]
 
         return tokenized_full_prompt
@@ -334,7 +334,6 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
         return prompt["conversations"]
 
     def tokenize_prompt(self, prompt):
-        # Initial values. We will append to these as we go through the conversation.
         result, current_len = tokenize_prompt_default()
         conversation: Conversation = (
             self.prompter._conversation.copy()  # pylint: disable=protected-access
@@ -356,67 +355,62 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
             for _, part in enumerate(
                 self.prompter.build_prompt(self.get_conversation_thread(prompt))
             ):
-                if not isinstance(part, tuple):
-                    LOG.warning(f"expected tuple, got {part}")
-                    continue
-
-                user, assistant = conversation.roles
-                role, content = part
-
-                # Uses "in" because role contains extra characters
-                if user in role:
-                    role = (
-                        role.replace(role_remap[0]["from"], role_remap[0]["to"])
-                        if role_remap
-                        else role
-                    )
-                    turn = role + content
-                    # this is still the user query, we should
-                    if not content.strip():
-                        LOG.warning(f"user turn has empty text: {prompt}")
-                    res = self._tokenize(
-                        turn,
-                        add_eos_token=False,
-                        strip_bos_token=True,
-                    )
-                    # everything from this is masked out from the labels
-                    labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
-                elif assistant in role:
-                    # TODO label assistant token/tokens w/ IGNORE_TOKEN_ID
-                    role = (
-                        role.replace(role_remap[1]["from"], role_remap[1]["to"])
-                        if role_remap
-                        else role
-                    )
-                    turn = role + content
-                    # this should be the assistant response, should end with an eos token
-                    if not content.strip():
-                        LOG.warning(f"assistant turn has empty text: {prompt}")
-                    res = self._tokenize(
-                        turn,
-                        add_eos_token=True,
-                        strip_bos_token=True,
-                    )
-                    role_res = self._tokenize(
-                        role.rstrip(),
-                        add_eos_token=False,
-                        strip_bos_token=True,
-                    )
-                    # not masked out from labels
-                    labels = copy.deepcopy(res["input_ids"])
-                    len_role = len(role_res["input_ids"])
-                    labels[:len_role] = [IGNORE_TOKEN_ID] * min(len_role, len(labels))
-                elif role == "":
-                    turn = content
-                    # this is only ever the first part, should include the bos token and the user query
-                    res = self._tokenize(
-                        turn, add_eos_token=False, strip_bos_token=False
-                    )
-                    # everything from this is masked out from the labels
-                    labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
-                else:
-                    LOG.warning(f"unhandled role: {role}")
-                    continue
+                if isinstance(part, tuple):
+                    if conversation.roles[0] in part[0]:
+                        role = (
+                            part[0].replace(role_remap[0]["from"], role_remap[0]["to"])
+                            if role_remap
+                            else part[0]
+                        )
+                        turn = role + part[1]
+                        # this is still the user query, we should
+                        if not part[1].strip():
+                            LOG.warning(f"user turn has empty text: {prompt}")
+                        res = self._tokenize(
+                            turn,
+                            add_eos_token=False,
+                            strip_bos_token=True,
+                        )
+                        # everything from this is masked out from the labels
+                        labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
+                    elif conversation.roles[1] in part[0]:
+                        # TODO label assistant token/tokens w/ IGNORE_TOKEN_ID
+                        role = (
+                            part[0].replace(role_remap[1]["from"], role_remap[1]["to"])
+                            if role_remap
+                            else part[0]
+                        )
+                        turn = role + part[1]
+                        # this should be the assistant response, should end with an eos token
+                        if not part[1].strip():
+                            LOG.warning(f"assistant turn has empty text: {prompt}")
+                        res = self._tokenize(
+                            turn,
+                            add_eos_token=True,
+                            strip_bos_token=True,
+                        )
+                        role_res = self._tokenize(
+                            role.rstrip(),
+                            add_eos_token=False,
+                            strip_bos_token=True,
+                        )
+                        # not masked out from labels
+                        labels = copy.deepcopy(res["input_ids"])
+                        len_role = len(role_res["input_ids"])
+                        labels[:len_role] = [IGNORE_TOKEN_ID] * min(
+                            len_role, len(labels)
+                        )
+                    elif part[0] == "":
+                        turn = part[1]
+                        # this is only ever the first part, should include the bos token and the user query
+                        res = self._tokenize(
+                            turn, add_eos_token=False, strip_bos_token=False
+                        )
+                        # everything from this is masked out from the labels
+                        labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
+                    else:
+                        LOG.warning(f"unhandled role: {part[0]}")
+                        continue
 
                 # pylint: disable=duplicate-code
                 result, current_len = parse_tokenized_to_result(
@@ -430,6 +424,38 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
         except (KeyError, AssertionError, IndexError) as err:
             raise InvalidDataException(str(err)) from err
 
+    def _tokenize(self, prompt, add_eos_token=True, strip_bos_token=False):
+        if not prompt.strip():
+            LOG.warning("Empty text requested for tokenization.")
+            result = BatchEncoding(data={"input_ids": [], "attention_mask": []})
+        else:
+            result = self.tokenizer(
+                prompt,
+                truncation=True,
+                max_length=self.sequence_len,
+                padding=False,
+                return_tensors=None,
+            )
+        if (
+            len(result["input_ids"]) > 0
+            and result["input_ids"][-1] != self.tokenizer.eos_token_id
+            and len(result["input_ids"]) < self.sequence_len
+            and add_eos_token
+        ):
+            result["input_ids"].append(self.tokenizer.eos_token_id)
+            result["attention_mask"].append(1)
+
+        if (
+            len(result["input_ids"]) > 0
+            and result["input_ids"][0] == self.tokenizer.bos_token_id
+            and strip_bos_token
+        ):
+            result["input_ids"] = result["input_ids"][1:]
+            result["attention_mask"] = result["attention_mask"][1:]
+
+        result["labels"] = result["input_ids"].copy()
+        return result
+
 
 def tokenize_prompt_default() -> Tuple[Dict[str, List[int]], int]:
     """

src/axolotl/prompters.py

@@ -4,12 +4,10 @@ import logging
 from enum import Enum
 from typing import Generator, Optional, Union
 
-from colorama import Fore
 from fastchat.conversation import Conversation, get_conv_template
 
 LOG = logging.getLogger("axolotl")
 IGNORE_TOKEN_ID = -100
-REPR_TEMPLATE = "\n<start>\n" + Fore.CYAN + "{full_prompt}" + Fore.RESET + "\n<end>\n"
 
 
 class PromptStyle(Enum):
@@ -57,15 +55,20 @@ class AlpacaPrompter:
             )
             self.system_format = "<|im_start|>system\n{system}<|im_end|>\n"
 
-    def _build_result(self, instruction, input_text, output):
+    def build_prompt(
+        self,
+        instruction: str,
+        input: Union[None, str] = None,  # pylint: disable=redefined-builtin
+        output: Union[None, str] = None,
+    ) -> Generator[str, None, None]:
         # returns the full prompt from instruction and optional input
         # if a label (=response, =output) is provided, it's also appended.
-        if input_text:
+        if input:
             res = (
                 self.system_format.format(system=self.system_prompt)
                 if self.system_prompt
                 else ""
-            ) + self.turn_format.format(instruction=instruction, input=input_text)
+            ) + self.turn_format.format(instruction=instruction, input=input)
         else:
             res = (
                 self.system_format.format(system=self.system_no_input_prompt)
@@ -74,21 +77,7 @@ class AlpacaPrompter:
             ) + self.turn_no_input_format.format(instruction=instruction)
         if output:
             res = f"{res}{output}"
-
-        return res
-
-    def build_prompt(
-        self,
-        instruction: str,
-        input: Union[None, str] = None,  # pylint: disable=redefined-builtin
-        output: Union[None, str] = None,
-    ) -> Generator[str, None, None]:
-        yield self._build_result(instruction, input, output)
-
-    def __repr__(self) -> str:
-        return REPR_TEMPLATE.format(
-            full_prompt=self._build_result("{instruction}", "{input}", "{output}")
-        )
+        yield res
 
 
 class UnpromptedPrompter(AlpacaPrompter):
@@ -202,14 +191,14 @@ class ReflectAlpacaPrompter:
             )
             self.response_split = "ASSISTANT:"
 
-    def _build_result(
+    def build_prompt(
         self,
         instruction: str,
         input: Union[None, str] = None,  # pylint: disable=redefined-builtin
         output: Union[None, str] = None,
         reflection: Union[None, str] = None,
         corrected: Union[None, str] = None,
-    ):
+    ) -> Generator[str, None, None]:
         # returns the full prompt from instruction and optional input
         # if a label (=response, =output) is provided, it's also appended.
         if input:
@@ -223,30 +212,7 @@ class ReflectAlpacaPrompter:
                 corrected=corrected,
             )
             res = f"{res}{label}"
-
-        return res
-
-    def build_prompt(
-        self,
-        instruction: str,
-        input: Union[None, str] = None,  # pylint: disable=redefined-builtin
-        output: Union[None, str] = None,
-        reflection: Union[None, str] = None,
-        corrected: Union[None, str] = None,
-    ) -> Generator[str, None, None]:
-        # pylint: disable=duplicate-code
-        yield self._build_result(
-            instruction,
-            input,
-            output,
-            reflection,
-            corrected,
-        )
-
-    def __repr__(self) -> str:
-        return REPR_TEMPLATE.format(
-            full_prompt=self._build_result("{instruction}", "{input}", "{output}")
-        )
+        yield res
 
 
 SHAREGPT_ASSERTION_FAILED_ROLE = (
@@ -281,7 +247,7 @@ class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
         if role_key_model:
             self.role_key_model = role_key_model
 
-    def _build_result(self, source):
+    def build_prompt(self, source) -> Generator[str, None, None]:
         if len(source) < 2:
             # If there isn't a back and forth conversation, ignore it
             # also happens on the data splitting leaving empty conversations
@@ -316,20 +282,11 @@ class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
                 LOG.warning(f"{SHAREGPT_ASSERTION_FAILED_ROLE}: {sentence}")
             conv.append_message(role, sentence["value"])
 
-        return conv.get_turns()
-
-    def build_prompt(self, source) -> Generator[str, None, None]:
-        turns = self._build_result(source)
-
-        for part in turns:
+        for part in conv.get_turns():
             if part[0] and not part[1]:
                 LOG.warning(f"role with empty message: {part[0]}")
             yield part
 
-    def __repr__(self) -> str:
-        turns = self._build_result([{"from": "{from}", "value": "{value}"}])
-        return "\n".join([REPR_TEMPLATE.format(full_prompt=part) for part in turns])
-
 
 class ShareGPTPrompterV2(ShareGPTPrompter):
     """
@@ -347,15 +304,3 @@ class ShareGPTPrompterV2(ShareGPTPrompter):
             role_key_human=role_key_human,
             role_key_model=role_key_model,
         )
-
-
-class UnsupportedPrompter:
-    """
-    A dummy class for custom prompters
-    """
-
-    def __init__(self) -> None:
-        pass
-
-    def __repr__(self):
-        return "Pre-tokenized or custom dataset types are unsupported for logging"

src/axolotl/train.py

@@ -12,11 +12,9 @@ import torch
 import transformers.modelcard
 from datasets import Dataset
 from optimum.bettertransformer import BetterTransformer
-from transformers.deepspeed import is_deepspeed_zero3_enabled
 
 from axolotl.common.cli import TrainerCliArgs
 from axolotl.logging_config import configure_logging
-from axolotl.monkeypatch import neft_embeddings
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_model, load_tokenizer
 from axolotl.utils.trainer import setup_trainer
@@ -41,7 +39,10 @@ class TrainDatasetMeta:
 
 
 def train(
-    *, cfg: DictDefault, cli_args: TrainerCliArgs, dataset_meta: TrainDatasetMeta
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,
+    dataset_meta: TrainDatasetMeta,
 ):
     # load the tokenizer first
     LOG.info(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
@@ -108,7 +109,6 @@ def train(
     if cfg.group_by_length:
         LOG.info("hang tight... sorting dataset for group_by_length")
 
-    pretrain_hooks(cfg, trainer)
     if cfg.flash_optimum:
         with torch.backends.cuda.sdp_kernel(
             enable_flash=True, enable_math=True, enable_mem_efficient=True
@@ -116,15 +116,9 @@ def train(
             trainer.train(resume_from_checkpoint=resume_from_checkpoint)
     else:
         trainer.train(resume_from_checkpoint=resume_from_checkpoint)
-    post_train_hooks(cfg, trainer)
 
     LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
 
-    # post training
-    for name, module in model.named_modules():
-        if hasattr(module, "_post_training"):
-            module._post_training(model, name)  # pylint: disable=protected-access
-
     if trainer.is_fsdp_enabled:
         trainer.accelerator.state.fsdp_plugin.set_state_dict_type("FULL_STATE_DICT")
         LOG.info("Set FSDP state dict type to FULL_STATE_DICT for saving.")
@@ -140,22 +134,6 @@ def train(
     # only save on rank 0, otherwise it corrupts output on multi-GPU when multiple processes attempt to write the same file
     if cfg.fsdp:
         trainer.save_model(cfg.output_dir)
-    elif cfg.deepspeed and is_deepspeed_zero3_enabled():
-        # Copied over from: https://github.com/huggingface/accelerate/blob/5ae611118057232f441055f7ef9ba0b0f2b8d533/docs/source/usage_guides/deepspeed.md#saving-and-loading
-        trainer.accelerator.wait_for_everyone()
-        unwrapped_model = trainer.accelerator.unwrap_model(trainer.model_wrapped)
-
-        # Saves the whole/unpartitioned fp16 model when in ZeRO Stage-3 to the output directory if
-        # `stage3_gather_16bit_weights_on_model_save` is True in DeepSpeed Config file or
-        # `zero3_save_16bit_model` is True in DeepSpeed Plugin.
-        # For Zero Stages 1 and 2, models are saved as usual in the output directory.
-        # The model name saved is `pytorch_model.bin`
-        unwrapped_model.save_pretrained(
-            cfg.output_dir,
-            is_main_process=trainer.accelerator.is_main_process,
-            save_function=trainer.accelerator.save,
-            state_dict=trainer.accelerator.get_state_dict(trainer.model_wrapped),
-        )
     elif cfg.local_rank == 0:
         if cfg.flash_optimum:
             model = BetterTransformer.reverse(model)
@@ -166,23 +144,3 @@ def train(
         trainer.create_model_card(model_name=cfg.output_dir.lstrip("./"))
 
     return model, tokenizer
-
-
-def pretrain_hooks(cfg, trainer):
-    """
-    Run hooks right before kicking off the training
-    :param cfg:
-    :param trainer:
-    :return:
-    """
-    neft_embeddings.pretrain_hook(cfg, trainer)
-
-
-def post_train_hooks(cfg, trainer):
-    """
-    Run hooks right after training completes
-    :param cfg:
-    :param trainer:
-    :return:
-    """
-    neft_embeddings.post_train_hook(cfg, trainer)

src/axolotl/utils/bench.py

@@ -1,13 +1,10 @@
 """Benchmarking and measurement utilities"""
 import functools
-import logging
 
 import pynvml
 import torch
 from pynvml.nvml import NVMLError
 
-LOG = logging.getLogger("axolotl.utils.bench")
-
 
 def check_cuda_device(default_value):
     """
@@ -65,14 +62,7 @@ def gpu_memory_usage_smi(device=0):
 
 
 def log_gpu_memory_usage(log, msg, device):
-    if not torch.cuda.is_available():
-        return (0, 0, 0)
-
-    try:
-        usage, cache, misc = gpu_memory_usage_all(device)
-    except ValueError as exc:
-        LOG.exception(exc)
-        return (0, 0, 0)
+    usage, cache, misc = gpu_memory_usage_all(device)
     extras = []
     if cache > 0:
         extras.append(f"+{cache:.03f}GB cache")

src/axolotl/utils/callbacks.py

@@ -37,7 +37,7 @@ from axolotl.utils.distributed import (
 )
 
 if TYPE_CHECKING:
-    from axolotl.core.trainer_builder import AxolotlTrainingArguments
+    from axolotl.utils.trainer import AxolotlTrainingArguments
 
 LOG = logging.getLogger("axolotl.callbacks")
 IGNORE_INDEX = -100

src/axolotl/utils/config.py

@@ -79,9 +79,6 @@ def normalize_config(cfg):
 
     cfg.dataset_processes = cfg.dataset_processes or os.cpu_count()
 
-    if not cfg.base_model_config:
-        cfg.base_model_config = cfg.base_model
-
     model_config = load_model_config(cfg)
     cfg.model_config_type = model_config.model_type
 
@@ -122,11 +119,20 @@ def normalize_config(cfg):
         or (cfg.model_type and "mistral" in cfg.model_type.lower())
     )
 
-    if isinstance(cfg.learning_rate, str):
-        cfg.learning_rate = float(cfg.learning_rate)
-
     log_gpu_memory_usage(LOG, "baseline", cfg.device)
 
+    if cfg.adapter is not None:
+        for key in list(cfg.keys()):
+            if key.startswith("lora_"):
+                new_key = key.replace("lora_", "peft_")
+                LOG.warning(
+                    PendingDeprecationWarning(
+                        f"{key} soon to be deprecated. please use {new_key}"
+                    )
+                )
+                cfg[new_key] = cfg[key]
+                del cfg[key]
+
 
 def validate_config(cfg):
     if is_torch_bf16_gpu_available():
@@ -195,14 +201,11 @@ def validate_config(cfg):
             if not cfg.load_in_4bit:
                 raise ValueError("Require cfg.load_in_4bit to be True for qlora")
 
-        if cfg.flash_attn_fuse_qkv or cfg.flash_attn_fuse_mlp:
-            raise ValueError("Fused modules are not supported with QLoRA")
-
     if not cfg.load_in_8bit and cfg.adapter == "lora":
-        LOG.warning("We recommend setting `load_in_8bit: true` for LORA finetuning")
+        LOG.warning("We recommend setting `load_in_8bit: true` for LoRA finetuning")
 
-    if cfg.adapter == "lora" and (cfg.flash_attn_fuse_qkv or cfg.flash_attn_fuse_mlp):
-        raise ValueError("Fused modules are not supported with LoRA")
+    if not cfg.load_in_8bit and cfg.adapter == "ia3":
+        LOG.warning("We recommend setting `load_in_8bit: true` for IA3 finetuning")
 
     if cfg.relora_steps:
         if cfg.adapter not in ("lora", "qlora"):
@@ -217,9 +220,6 @@ def validate_config(cfg):
         if cfg.lr_scheduler == "one_cycle":
             raise ValueError("ReLoRA is not compatible with the one_cycle scheduler")
 
-        if cfg.flash_attn_fuse_qkv or cfg.flash_attn_fuse_mlp:
-            raise ValueError("Fused modules are not supported with ReLoRA")
-
     if cfg.trust_remote_code:
         LOG.warning(
             "`trust_remote_code` is set to true. Please make sure that you reviewed the remote code/model."
@@ -354,25 +354,6 @@ def validate_config(cfg):
             "eval_steps and evaluation_strategy are not supported with val_set_size == 0"
         )
 
-    if (
-        cfg.sample_packing
-        and cfg.eval_table_size
-        and cfg.eval_sample_packing is not False
-    ):
-        raise ValueError(
-            "eval_table_size and eval_sample_packing are not supported together with sample_packing. Please set 'eval_sample_packing' to false."
-        )
-
-    if not cfg.adapter and (cfg.load_in_8bit or cfg.load_in_4bit):
-        raise ValueError(
-            "load_in_8bit and load_in_4bit are not supported without setting an adapter."
-            "If you want to full finetune, please turn off load_in_8bit and load_in_4bit."
-        )
-
-    if cfg.tensor_parallel and cfg.gradient_checkpointing:
-        raise ValueError(
-            "TensorParallelPreTrainedModel does not support gradient checkpointing"
-        )
     # 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 (
@@ -36,7 +36,6 @@ from axolotl.prompters import (
     MultipleChoiceExplainPrompter,
     ReflectAlpacaPrompter,
     SummarizeTLDRPrompter,
-    UnsupportedPrompter,
 )
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process, zero_first
@@ -56,10 +55,9 @@ def md5(to_hash: str, encoding: str = "utf-8") -> str:
 
 
 def prepare_dataset(cfg, tokenizer):
-    prompters = []
     if not cfg.pretraining_dataset:
         with zero_first(is_main_process()):
-            train_dataset, eval_dataset, prompters = load_prepare_datasets(
+            train_dataset, eval_dataset = load_prepare_datasets(
                 tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
             )
     else:
@@ -72,7 +70,7 @@ def prepare_dataset(cfg, tokenizer):
         # https://discuss.huggingface.co/t/how-to-use-huggingface-trainer-streaming-datasets-without-wrapping-it-with-torchdatas-iterablewrapper/25230
         train_dataset = train_dataset.with_format("torch")
         eval_dataset = None
-        return train_dataset, eval_dataset, cfg.max_steps, prompters
+        return train_dataset, eval_dataset, cfg.max_steps
 
     with zero_first(is_main_process()):
         train_dataset, eval_dataset = process_datasets_for_packing(
@@ -85,7 +83,7 @@ def prepare_dataset(cfg, tokenizer):
         LOG.info(f"Maximum number of steps set at {total_num_steps}")
     else:
         total_num_steps = calculate_total_num_steps(cfg, train_dataset, tokenizer)
-    return train_dataset, eval_dataset, total_num_steps, prompters
+    return train_dataset, eval_dataset, total_num_steps
 
 
 def load_tokenized_prepared_datasets(
@@ -111,7 +109,6 @@ def load_tokenized_prepared_datasets(
         else Path(default_dataset_prepared_path) / ds_hash
     )
     dataset = None
-    prompters = []
     use_auth_token = cfg.hf_use_auth_token
     try:
         if cfg.push_dataset_to_hub:
@@ -150,13 +147,13 @@ def load_tokenized_prepared_datasets(
                     yield dataset
 
         # pylint: disable=invalid-name
-        for config_dataset in for_d_in_datasets(cfg.datasets):
+        for d in for_d_in_datasets(cfg.datasets):
             ds: Union[Dataset, DatasetDict] = None
             ds_from_hub = False
             try:
                 load_dataset(
-                    config_dataset.path,
-                    name=config_dataset.name,
+                    d.path,
+                    name=d.name,
                     streaming=True,
                     token=use_auth_token,
                 )
@@ -165,33 +162,33 @@ def load_tokenized_prepared_datasets(
                 pass
 
             # prefer local dataset, even if hub exists
-            local_path = Path(config_dataset.path)
+            local_path = Path(d.path)
             if local_path.exists():
                 if local_path.is_dir():
                     # TODO dirs with arrow or parquet files could be loaded with `load_from_disk`
                     ds = load_dataset(
-                        config_dataset.path,
-                        name=config_dataset.name,
-                        data_files=config_dataset.data_files,
+                        d.path,
+                        name=d.name,
+                        data_files=d.data_files,
                         streaming=False,
                         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:
+                    if d.ds_type:
+                        ds_type = d.ds_type
+                    elif ".parquet" in d.path:
                         ds_type = "parquet"
-                    elif ".arrow" in config_dataset.path:
+                    elif ".arrow" in d.path:
                         ds_type = "arrow"
-                    elif ".csv" in config_dataset.path:
+                    elif ".csv" in d.path:
                         ds_type = "csv"
-                    elif ".txt" in config_dataset.path:
+                    elif ".txt" in d.path:
                         ds_type = "text"
                     ds = load_dataset(
                         ds_type,
-                        name=config_dataset.name,
-                        data_files=config_dataset.path,
+                        name=d.name,
+                        data_files=d.path,
                         streaming=False,
                         split=None,
                     )
@@ -201,25 +198,25 @@ def load_tokenized_prepared_datasets(
                     )
             elif ds_from_hub:
                 ds = load_dataset(
-                    config_dataset.path,
-                    name=config_dataset.name,
+                    d.path,
+                    name=d.name,
                     streaming=False,
-                    data_files=config_dataset.data_files,
+                    data_files=d.data_files,
                     token=use_auth_token,
                 )
             else:
-                if isinstance(config_dataset.data_files, str):
+                if isinstance(d.data_files, str):
                     fp = hf_hub_download(
-                        repo_id=config_dataset.path,
+                        repo_id=d.path,
                         repo_type="dataset",
-                        filename=config_dataset.data_files,
+                        filename=d.data_files,
                     )
-                elif isinstance(config_dataset.data_files, list):
+                elif isinstance(d.data_files, list):
                     fp = []
-                    for file in config_dataset.data_files:
+                    for file in d.data_files:
                         fp.append(
                             hf_hub_download(
-                                repo_id=config_dataset.path,
+                                repo_id=d.path,
                                 repo_type="dataset",
                                 filename=file,
                             )
@@ -229,27 +226,21 @@ def load_tokenized_prepared_datasets(
                         "data_files must be either a string or list of strings"
                     )
                 ds = load_dataset(
-                    "json",
-                    name=config_dataset.name,
-                    data_files=fp,
-                    streaming=False,
-                    split=None,
+                    "json", name=d.name, data_files=fp, streaming=False, split=None
                 )
             if not ds:
                 raise ValueError("unhandled dataset load")
             # support for using a subset of the data
-            if config_dataset.shards:
+            if d.shards:
                 if "train" in ds:
                     ds = ds.shuffle(seed=seed)["train"].shard(
-                        num_shards=config_dataset.shards, index=0
+                        num_shards=d.shards, index=0
                     )
                 else:
-                    ds = ds.shuffle(seed=seed).shard(
-                        num_shards=config_dataset.shards, index=0
-                    )
+                    ds = ds.shuffle(seed=seed).shard(num_shards=d.shards, index=0)
 
             d_base_type = d_prompt_style = None
-            d_type = config_dataset.type
+            d_type = d.type
             if isinstance(d_type, str):
                 d_type_split = d_type.split(":")
                 d_base_type = d_type_split[0]
@@ -258,26 +249,108 @@ def load_tokenized_prepared_datasets(
                 ds = ds["train"]
             elif (
                 isinstance(ds, DatasetDict)
-                and config_dataset.train_on_split
-                and config_dataset.train_on_split in ds
+                and d.train_on_split
+                and d.train_on_split in ds
             ):
-                ds = ds[config_dataset.train_on_split]
+                ds = ds[d.train_on_split]
             elif isinstance(ds, DatasetDict):
                 raise ValueError(
-                    f"no train split found for dataset {config_dataset.path}, you may specify a split with 'train_on_split: `"
+                    f"no train split found for dataset {d.path}, you may specify a split with 'train_on_split: `"
+                )
+            if (
+                "input_ids" in ds.features
+                and "attention_mask" in ds.features
+                and "labels" in ds.features
+            ):
+                # dataset is already tokenized, just drop it straight in
+                datasets.append(ds)
+            elif isinstance(d.type, DictDefault):
+                ds_strategy = load("user_defined", tokenizer, cfg, d.type.to_dict())
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            elif ds_strategy := load(d.type, tokenizer, cfg, d):
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            elif d_base_type == "alpaca":
+                ds_strategy = AlpacaPromptTokenizingStrategy(
+                    AlpacaPrompter(d_prompt_style),
+                    tokenizer,
+                    cfg.train_on_inputs,
+                    cfg.sequence_len,
+                )
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            elif d_base_type == "explainchoice":
+                ds_strategy = AlpacaMultipleChoicePromptTokenizingStrategy(
+                    MultipleChoiceExplainPrompter(d_prompt_style),
+                    tokenizer,
+                    cfg.train_on_inputs,
+                    cfg.sequence_len,
+                )
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            elif d_base_type == "concisechoice":
+                ds_strategy = AlpacaMultipleChoicePromptTokenizingStrategy(
+                    MultipleChoiceConcisePrompter(d_prompt_style),
+                    tokenizer,
+                    cfg.train_on_inputs,
+                    cfg.sequence_len,
+                )
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            elif d_base_type == "summarizetldr":
+                ds_strategy = SummarizeTLDRPromptTokenizingStrategy(
+                    SummarizeTLDRPrompter(d_prompt_style),
+                    tokenizer,
+                    cfg.train_on_inputs,
+                    cfg.sequence_len,
+                )
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            elif d_base_type == "jeopardy":
+                ds_strategy = JeopardyPromptTokenizingStrategy(
+                    JeopardyPrompter(d_prompt_style),
+                    tokenizer,
+                    cfg.train_on_inputs,
+                    cfg.sequence_len,
+                )
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            elif d_base_type == "oasst":
+                ds_strategy = OpenAssistantPromptTokenizingStrategy(
+                    AlpacaPrompter(d_prompt_style),
+                    tokenizer,
+                    cfg.train_on_inputs,
+                    cfg.sequence_len,
+                )
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            elif d_base_type == "gpteacher":
+                ds_strategy = GPTeacherPromptTokenizingStrategy(
+                    GPTeacherPrompter(d_prompt_style),
+                    tokenizer,
+                    cfg.train_on_inputs,
+                    cfg.sequence_len,
+                )
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            elif d_base_type == "reflection":
+                ds_strategy = AlpacaReflectionPTStrategy(
+                    ReflectAlpacaPrompter(d_prompt_style),
+                    tokenizer,
+                    cfg.train_on_inputs,
+                    cfg.sequence_len,
+                )
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            else:
+                suffix = ""
+                if ":load_" in d.type:
+                    suffix = f" Did you mean {d.type.replace(':load_', '.load_')}?"
+                LOG.error(f"unhandled prompt tokenization strategy: {d.type}. {suffix}")
+                raise ValueError(
+                    f"unhandled prompt tokenization strategy: {d.type} {suffix}"
                 )
-
-            dataset_wrapper, dataset_prompter = get_dataset_wrapper(
-                config_dataset=config_dataset,
-                dataset=ds,
-                tokenizer=tokenizer,
-                cfg=cfg,
-                d_base_type=d_base_type,
-                d_prompt_style=d_prompt_style,
-            )
-            datasets.append(dataset_wrapper)
-            prompters.append(dataset_prompter)
-
         LOG.info("merging datasets")
         dataset = concatenate_datasets(datasets)
 
@@ -295,14 +368,14 @@ def load_tokenized_prepared_datasets(
                     f"{cfg.push_dataset_to_hub}/{ds_hash}", private=True
                 )
 
-    return dataset, prompters
+    return dataset
 
 
 def load_prepare_datasets(
     tokenizer: PreTrainedTokenizerBase,
     cfg,
     default_dataset_prepared_path,
-) -> Tuple[Dataset, Dataset, List[Any]]:
+) -> Tuple[Dataset, Dataset]:
     max_packed_sequence_len = (
         cfg.max_packed_sequence_len if cfg.max_packed_sequence_len else cfg.sequence_len
     )
@@ -311,7 +384,6 @@ def load_prepare_datasets(
     )  # make sure we don't accidentally set it larger than sequence_len
 
     tokenizer_name = tokenizer.__class__.__name__
-    prompters = []
     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 ""
@@ -367,7 +439,7 @@ def load_prepare_datasets(
                     f"{cfg.push_dataset_to_hub}/{ds_hash}", private=True
                 )
         else:
-            dataset, prompters = load_tokenized_prepared_datasets(
+            dataset = load_tokenized_prepared_datasets(
                 tokenizer, cfg, default_dataset_prepared_path
             )
 
@@ -409,7 +481,7 @@ def load_prepare_datasets(
                         private=True,
                     )
     else:
-        dataset, prompters = load_tokenized_prepared_datasets(
+        dataset = load_tokenized_prepared_datasets(
             tokenizer, cfg, default_dataset_prepared_path
         )
 
@@ -460,124 +532,7 @@ def load_prepare_datasets(
         train_dataset = dataset
         eval_dataset = None
 
-    return train_dataset, eval_dataset, prompters
-
-
-def get_dataset_wrapper(
-    config_dataset, dataset, tokenizer, cfg, d_base_type, d_prompt_style
-):
-    dataset_wrapper = None
-    dataset_prompter = None
-
-    if (
-        "input_ids" in dataset.features
-        and "attention_mask" in dataset.features
-        and "labels" in dataset.features
-    ):
-        # dataset is already tokenized, just drop it straight in
-        dataset_prompter = UnsupportedPrompter()
-        dataset_wrapper = dataset
-    elif isinstance(config_dataset.type, DictDefault):
-        ds_strategy = load(
-            "user_defined", tokenizer, cfg, config_dataset.type.to_dict()
-        )
-        dataset_prompter = UnsupportedPrompter()
-        dataset_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
-    elif ds_strategy := load(config_dataset.type, tokenizer, cfg, config_dataset):
-        dataset_prompter = UnsupportedPrompter()
-        dataset_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
-    elif d_base_type == "alpaca":
-        dataset_prompter = AlpacaPrompter(d_prompt_style)
-        ds_strategy = AlpacaPromptTokenizingStrategy(
-            dataset_prompter,
-            tokenizer,
-            cfg.train_on_inputs,
-            cfg.sequence_len,
-        )
-        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
-        dataset_wrapper = ds_wrapper
-    elif d_base_type == "explainchoice":
-        dataset_prompter = MultipleChoiceExplainPrompter(d_prompt_style)
-        ds_strategy = AlpacaMultipleChoicePromptTokenizingStrategy(
-            dataset_prompter,
-            tokenizer,
-            cfg.train_on_inputs,
-            cfg.sequence_len,
-        )
-        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
-        dataset_wrapper = ds_wrapper
-    elif d_base_type == "concisechoice":
-        dataset_prompter = MultipleChoiceConcisePrompter(d_prompt_style)
-        ds_strategy = AlpacaMultipleChoicePromptTokenizingStrategy(
-            dataset_prompter,
-            tokenizer,
-            cfg.train_on_inputs,
-            cfg.sequence_len,
-        )
-        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
-        dataset_wrapper = ds_wrapper
-    elif d_base_type == "summarizetldr":
-        dataset_prompter = SummarizeTLDRPrompter(d_prompt_style)
-        ds_strategy = SummarizeTLDRPromptTokenizingStrategy(
-            dataset_prompter,
-            tokenizer,
-            cfg.train_on_inputs,
-            cfg.sequence_len,
-        )
-        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
-        dataset_wrapper = ds_wrapper
-    elif d_base_type == "jeopardy":
-        dataset_prompter = JeopardyPrompter(d_prompt_style)
-        ds_strategy = JeopardyPromptTokenizingStrategy(
-            dataset_prompter,
-            tokenizer,
-            cfg.train_on_inputs,
-            cfg.sequence_len,
-        )
-        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
-        dataset_wrapper = ds_wrapper
-    elif d_base_type == "oasst":
-        dataset_prompter = AlpacaPrompter(d_prompt_style)
-        ds_strategy = OpenAssistantPromptTokenizingStrategy(
-            dataset_prompter,
-            tokenizer,
-            cfg.train_on_inputs,
-            cfg.sequence_len,
-        )
-        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
-        dataset_wrapper = ds_wrapper
-    elif d_base_type == "gpteacher":
-        dataset_prompter = GPTeacherPrompter(d_prompt_style)
-        ds_strategy = GPTeacherPromptTokenizingStrategy(
-            dataset_prompter,
-            tokenizer,
-            cfg.train_on_inputs,
-            cfg.sequence_len,
-        )
-        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
-        dataset_wrapper = ds_wrapper
-    elif d_base_type == "reflection":
-        dataset_prompter = ReflectAlpacaPrompter(d_prompt_style)
-        ds_strategy = AlpacaReflectionPTStrategy(
-            dataset_prompter,
-            tokenizer,
-            cfg.train_on_inputs,
-            cfg.sequence_len,
-        )
-        ds_wrapper = TokenizedPromptDataset(ds_strategy, dataset)
-        dataset_wrapper = ds_wrapper
-    else:
-        suffix = ""
-        if ":load_" in config_dataset.type:
-            suffix = f" Did you mean {config_dataset.type.replace(':load_', '.load_')}?"
-        LOG.error(
-            f"unhandled prompt tokenization strategy: {config_dataset.type}. {suffix}"
-        )
-        raise ValueError(
-            f"unhandled prompt tokenization strategy: {config_dataset.type} {suffix}"
-        )
-
-    return dataset_wrapper, dataset_prompter
+    return train_dataset, eval_dataset
 
 
 def encode_pretraining(

src/axolotl/utils/dataloader.py

@@ -3,9 +3,6 @@ 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
@@ -152,8 +149,6 @@ class MultipackDistributedDataloader:
         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
@@ -172,7 +167,6 @@ class MultipackDistributedDataloader:
         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
@@ -183,44 +177,6 @@ class MultipackDistributedDataloader:
         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:
@@ -250,7 +206,11 @@ class MultipackDistributedDataloader:
 
         return batches, totseqs
 
-    def _internal_batch_generator(self):
+    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})")
         all_batches, _ = self.generate_batches(set_stats=True)
         features = self.dataset.features.keys()
         len_remaining = self._len_est()

src/axolotl/utils/models.py

@@ -7,7 +7,6 @@ from typing import Optional, Tuple  # noqa: F401
 import bitsandbytes as bnb
 import torch
 import transformers
-import transformers.utils.bitsandbytes
 from optimum.bettertransformer import BetterTransformer
 from peft import PeftConfig, prepare_model_for_kbit_training
 from peft.tuners.lora import QuantLinear
@@ -32,7 +31,7 @@ LOG = logging.getLogger("axolotl")
 
 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
+    trust_remote_code: bool = False or cfg.trust_remote_code
     return AutoConfig.from_pretrained(
         model_config_name, trust_remote_code=trust_remote_code
     )
@@ -73,6 +72,11 @@ def load_tokenizer(cfg):
         # set a pad_token, but use eos_token so we don't add a new token
         tokenizer.pad_token = LLAMA_DEFAULT_EOS_TOKEN
 
+    LOG.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
+    LOG.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
+    LOG.debug(f"PAD: {tokenizer.pad_token_id} / {tokenizer.pad_token}")
+    LOG.debug(f"UNK: {tokenizer.unk_token_id} / {tokenizer.unk_token}")
+
     if tokenizer.__class__.__name__ == "GPTNeoXTokenizerFast":
         tokenizer.add_special_tokens({"pad_token": "[PAD]"})
         os.environ["TOKENIZERS_PARALLELISM"] = "false"
@@ -94,11 +98,6 @@ def load_tokenizer(cfg):
             ]
         )
 
-    LOG.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
-    LOG.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
-    LOG.debug(f"PAD: {tokenizer.pad_token_id} / {tokenizer.pad_token}")
-    LOG.debug(f"UNK: {tokenizer.unk_token_id} / {tokenizer.unk_token}")
-
     return tokenizer
 
 
@@ -181,6 +180,26 @@ def load_model(
         LOG.info("patching with flash attention")
         replace_mistral_attn_with_flash_attn(packed=cfg.sample_packing)
 
+    if cfg.is_llama_derived_model and cfg.noisy_embedding_alpha:
+        from axolotl.monkeypatch.llama_embeddings_hijack import (
+            replace_llama_embeddings_with_uniform_distribution,
+        )
+
+        LOG.info("patching with noisy embeddings")
+        replace_llama_embeddings_with_uniform_distribution(
+            noise_alpha=cfg.noisy_embedding_alpha
+        )
+
+    if cfg.is_mistral_derived_model and cfg.noisy_embedding_alpha:
+        from axolotl.monkeypatch.mistral_embeddings_hijack import (
+            replace_mistral_embeddings_with_uniform_distribution,
+        )
+
+        LOG.info("patching with noisy embeddings")
+        replace_mistral_embeddings_with_uniform_distribution(
+            noise_alpha=cfg.noisy_embedding_alpha
+        )
+
     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,
@@ -222,7 +241,7 @@ def load_model(
             load_in_4bit=True,
             llm_int8_threshold=6.0,
             llm_int8_has_fp16_weight=False,
-            bnb_4bit_compute_dtype=torch.float16,
+            bnb_4bit_compute_dtype=cfg.torch_dtype,
             bnb_4bit_use_double_quant=True,
             bnb_4bit_quant_type="nf4",
         )
@@ -236,12 +255,7 @@ def load_model(
             model_kwargs["use_flash_attention_2"] = True
 
     try:
-        if (
-            cfg.is_llama_derived_model
-            and not cfg.trust_remote_code
-            and not cfg.gptq
-            and not cfg.tensor_parallel
-        ):
+        if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
             from transformers import LlamaForCausalLM
 
             config_kwargs = {}
@@ -258,20 +272,6 @@ def load_model(
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
                 **model_kwargs,
             )
-
-            if cfg.flash_attention and not inference:
-                from axolotl.monkeypatch.llama_attn_hijack_flash import (
-                    replace_llama_mlp_with_swiglu,
-                    replace_llama_qkv_with_fused,
-                )
-
-                if cfg.flash_attn_fuse_mlp:
-                    LOG.info("patching with SwiGLU")
-                    replace_llama_mlp_with_swiglu(model)
-
-                if cfg.flash_attn_fuse_qkv:
-                    LOG.info("patching with fused QKV")
-                    replace_llama_qkv_with_fused(model)
         # elif model_type == "GPTNeoXForCausalLM" and cfg.flash_attention:
         #     This is a WIP, still an issue with the backward pass
         #     RuntimeError: grad can be implicitly created only for scalar outputs
@@ -307,7 +307,7 @@ def load_model(
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
                 **model_kwargs,
             )
-        elif model_type and not cfg.trust_remote_code and not cfg.tensor_parallel:
+        elif model_type and not cfg.trust_remote_code:
             if cfg.gptq:
                 model = AutoModelForCausalLM.from_pretrained(
                     base_model,
@@ -322,17 +322,6 @@ def load_model(
                     trust_remote_code=cfg.trust_remote_code or False,
                     **model_kwargs,
                 )
-        elif cfg.tensor_parallel:
-            model_kwargs.pop("device_map")
-            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,
-                low_cpu_mem_usage=True,
-                offload_state_dict=True,
-                trust_remote_code=cfg.trust_remote_code or False,
-                **model_kwargs,
-            )
         else:
             config = AutoConfig.from_pretrained(
                 base_model,
@@ -383,18 +372,15 @@ def load_model(
             **model_kwargs,
         )
 
-    try:
-        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:
-            model.resize_token_embeddings(embeddings_len)
-        else:
-            model.tie_weights()
-    except NotImplementedError:
-        LOG.warning("`resize_token_embeddings` not implemented on model")
+    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:
+        model.resize_token_embeddings(embeddings_len)
+    else:
+        model.tie_weights()
 
     if (
         hasattr(model.config, "max_position_embeddings")
@@ -406,20 +392,6 @@ def load_model(
         )
         model.config.max_position_embeddings = cfg.sequence_len
 
-    if (
-        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")
-        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":
         log_gpu_memory_usage(LOG, "after model load", model.device)
 
@@ -434,21 +406,21 @@ def load_model(
             if hasattr(module, "weight"):
                 module.to(torch.float32)
 
-    needs_fa2_dtype = cfg.adapter or cfg.fsdp
-    if (cfg.adapter == "lora" and load_in_8bit) or (
-        cfg.adapter == "qlora" and cfg.load_in_4bit
-    ):
+    require_peft: bool = False
+    if cfg.adapter in ["lora", "qlora", "ia3"]:
+        require_peft = True
+
+    if require_peft:
         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
         )
-        needs_fa2_dtype = True
 
     # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
     # convert them back to fp16/bf16 for flash-attn compatibility.
-    if needs_fa2_dtype or (cfg.flash_attention and cfg.is_llama_derived_model):
+    if require_peft or cfg.fsdp or (cfg.flash_attention and cfg.is_llama_derived_model):
         LOG.info("converting modules to %s for flash attention", cfg.torch_dtype)
         for name, module in model.named_modules():
             if "norm" in name:
@@ -457,7 +429,7 @@ def load_model(
                 if hasattr(module, "weight"):
                     module.to(cfg.torch_dtype)
 
-    model, lora_config = load_adapter(model, cfg, cfg.adapter)
+    model, peft_config = load_adapter(model, cfg, cfg.adapter)
 
     if cfg.ddp and not load_in_8bit:
         model.to(f"cuda:{cfg.local_rank}")
@@ -488,7 +460,7 @@ def load_model(
         log_gpu_memory_usage(LOG, "after adapters", model.device)
 
     # TODO resume_from_checkpoint handling
-    return model, lora_config
+    return model, peft_config
 
 
 def load_adapter(model, cfg, adapter, inference=False):
@@ -497,12 +469,9 @@ def load_adapter(model, cfg, adapter, inference=False):
     if adapter is None:
         return model, None
     if hasattr(model, "enable_input_require_grads"):
-        try:
-            model.enable_input_require_grads()
-        except NotImplementedError:
-            LOG.warning("enable_input_require_grads not implemented on model")
-    if adapter == "qlora" and cfg.tensor_parallel:
-        model, _ = load_tp_qlora(model)
+        model.enable_input_require_grads()
+    if adapter == "ia3":
+        return load_ia3(model, cfg, inference=inference)
     if adapter in ["lora", "qlora"]:
         return load_lora(model, cfg, inference=inference)
     if adapter == "llama-adapter":
@@ -521,11 +490,11 @@ def load_llama_adapter(model, cfg):
         task_type="CAUSAL_LM",
     )
 
-    if cfg.lora_model_dir:
+    if cfg.peft_model_dir:
         LOG.debug("Loading pretained PEFT - llama_adapter")
         model = PeftModel.from_pretrained(
             model,
-            cfg.lora_model_dir,
+            cfg.peft_model_dir,
             torch_dtype=torch.float16,
         )
     else:
@@ -538,7 +507,7 @@ def load_llama_adapter(model, cfg):
 
 def find_all_linear_names(model):
     cls = (bnb.nn.Linear4bit, bnb.nn.Linear8bitLt, torch.nn.Linear, QuantLinear)
-    lora_module_names = set()
+    peft_module_names = set()
     for name, module in model.named_modules():
         if (
             isinstance(module, cls)
@@ -546,31 +515,12 @@ def find_all_linear_names(model):
             and module.__class__.__name__ not in ("LlamaLinearScalingRotaryEmbedding",)
         ):
             names = name.split(".")
-            lora_module_names.add(names[0] if len(names) == 1 else names[-1])
+            peft_module_names.add(names[0] if len(names) == 1 else names[-1])
 
-    if "lm_head" in lora_module_names:  # needed for 16-bit
-        lora_module_names.remove("lm_head")
+    if "lm_head" in peft_module_names:  # needed for 16-bit
+        peft_module_names.remove("lm_head")
 
-    return list(lora_module_names)
-
-
-def load_tp_qlora(model):
-    from transformers.utils.bitsandbytes import replace_with_bnb_linear
-
-    model = replace_with_bnb_linear(
-        model,
-        quantization_config=BitsAndBytesConfig(
-            load_in_4bit=True,
-            llm_int8_threshold=6.0,
-            llm_int8_has_fp16_weight=False,
-            bnb_4bit_compute_dtype=torch.float16,
-            bnb_4bit_use_double_quant=True,
-            bnb_4bit_quant_type="nf4",
-        ),
-    )
-    model.is_loaded_in_4bit = True
-
-    return model, None
+    return list(peft_module_names)
 
 
 def load_lora(model, cfg, inference=False):
@@ -578,34 +528,68 @@ def load_lora(model, cfg, inference=False):
 
     from peft import LoraConfig, PeftModel, get_peft_model
 
-    lora_target_modules = list(cfg.lora_target_modules or [])
+    peft_target_modules = list(cfg.peft_target_modules or [])
 
-    if cfg.lora_target_linear:
+    if cfg.peft_target_linear:
         linear_names = find_all_linear_names(model)
         LOG.info(f"found linear modules: {repr(linear_names)}")
-        lora_target_modules = list(set(lora_target_modules + linear_names))
+        peft_target_modules = list(set(peft_target_modules + linear_names))
 
-    lora_config = LoraConfig(
-        r=cfg.lora_r,
-        lora_alpha=cfg.lora_alpha,
-        target_modules=lora_target_modules,
-        lora_dropout=cfg.lora_dropout,
-        fan_in_fan_out=cfg.lora_fan_in_fan_out,
-        modules_to_save=cfg.lora_modules_to_save if cfg.lora_modules_to_save else None,
+    peft_config = LoraConfig(
+        r=cfg.peft_r,
+        lora_alpha=cfg.peft_alpha,
+        target_modules=peft_target_modules,
+        lora_dropout=cfg.peft_dropout,
+        fan_in_fan_out=cfg.peft_fan_in_fan_out,
+        modules_to_save=cfg.peft_modules_to_save if cfg.peft_modules_to_save else None,
         bias="none",
         task_type="CAUSAL_LM",
     )
 
-    if cfg.lora_model_dir:
+    if cfg.peft_model_dir:
         LOG.debug("Loading pretained PEFT - LoRA")
         model = PeftModel.from_pretrained(
             model,
-            cfg.lora_model_dir,
+            cfg.peft_model_dir,
             is_trainable=(not inference),
         )
     else:
-        model = get_peft_model(model, lora_config)
+        model = get_peft_model(model, peft_config)
 
     model.print_trainable_parameters()
 
-    return model, lora_config
+    return model, peft_config
+
+
+def load_ia3(model, cfg, inference=False):
+    # type: (PreTrainedModel, DictDefault, bool) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
+
+    from peft import IA3Config, PeftModel, get_peft_model
+
+    peft_config_kwargs = {}
+    if cfg.peft_init_ia3_weights is not None:
+        peft_config_kwargs["init_ia3_weights"] = cfg.peft_init_ia3_weights
+    if cfg.peft_fan_in_fan_out is not None:
+        peft_config_kwargs["fan_in_fan_out"] = cfg.peft_fan_in_fan_out
+
+    peft_config = IA3Config(
+        target_modules=cfg.peft_target_modules,
+        feedforward_modules=cfg.peft_feedforward_modules,
+        modules_to_save=cfg.peft_modules_to_save,
+        task_type="CAUSAL_LM",
+        **peft_config_kwargs,
+    )
+
+    if cfg.peft_model_dir:
+        LOG.debug("Loading pretained PEFT - IA3")
+        model = PeftModel.from_pretrained(
+            model,
+            cfg.peft_model_dir,
+            is_trainable=(not inference),
+        )
+    else:
+        model = get_peft_model(model, peft_config)
+
+    model.print_trainable_parameters()
+
+    return model, peft_config

src/axolotl/utils/tokenization.py

@@ -34,5 +34,6 @@ def check_example_labels(example, tokenizer, text_only=False):
     delimiter = "" if text_only else " "
     LOG.info(delimiter.join(colored_tokens))
     LOG.info("\n\n\n")
+    print(" ".join(colored_tokens))
 
     return " ".join(colored_tokens)

src/axolotl/utils/trainer.py

@@ -1,19 +1,40 @@
 """Module containing the Trainer class and related functions"""
+import importlib
 import logging
 import math
 import os
+import sys
 from contextlib import contextmanager
+from dataclasses import dataclass, field
 from functools import partial
-from typing import List
+from pathlib import Path
+from typing import List, Optional, Union
 
 import numpy as np
 import torch
 import torch.cuda
 import torch.distributed as dist
-from datasets import set_caching_enabled
-from torch.utils.data import DistributedSampler, RandomSampler
+import transformers
+from datasets import Dataset, set_caching_enabled
+from torch.optim.lr_scheduler import OneCycleLR
+from torch.utils.data import (
+    DataLoader,
+    DistributedSampler,
+    RandomSampler,
+    SequentialSampler,
+)
+from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
+from transformers.trainer_pt_utils import SequentialDistributedSampler
 
-from axolotl.core.trainer_builder import HFCausalTrainerBuilder
+from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
+from axolotl.utils.callbacks import (
+    EvalFirstStepCallback,
+    GPUStatsCallback,
+    SaveAxolotlConfigtoWandBCallback,
+    SaveBetterTransformerModelCallback,
+    bench_eval_callback_factory,
+    log_prediction_callback_factory,
+)
 from axolotl.utils.collators import DataCollatorForSeq2Seq
 from axolotl.utils.dataloader import MultipackDistributedDataloader
 from axolotl.utils.distributed import (
@@ -22,6 +43,7 @@ from axolotl.utils.distributed import (
     reduce_and_broadcast,
     zero_first,
 )
+from axolotl.utils.schedulers import get_cosine_schedule_with_quadratic_warmup
 
 LOG = logging.getLogger("axolotl")
 
@@ -88,6 +110,269 @@ def trainer_weighted_loss(model_output, labels, shift_labels=True):
     return weighted_cross_entropy(logits, labels, weights)
 
 
+@dataclass
+class AxolotlTrainingArguments(TrainingArguments):
+    """
+    Extend the base TrainingArguments for axolotl helpers
+    """
+
+    lr_quadratic_warmup: bool = field(
+        default=False,
+        metadata={"help": "Use quadratic warmup for cosine scheduling."},
+    )
+    sample_packing: bool = field(
+        default=False,
+        metadata={"help": "Use sample packing for efficient training."},
+    )
+    eval_sample_packing: Optional[bool] = field(
+        default=None,
+        metadata={"help": "Use sample packing for efficient evals."},
+    )
+    sample_packing_efficiency: float = field(
+        default=1.0,
+        metadata={"help": "Sample packing efficiency for calculating batch length."},
+    )
+    max_seq_length: int = field(
+        default=2048,
+        metadata={"help": "The maximum sequence length the model can handle"},
+    )
+    sample_packing_seq_len_multiplier: int = field(
+        default=1,
+        metadata={"help": "the multiplier for the max len for packed sequences"},
+    )
+    relora_steps: Optional[int] = field(
+        default=None,
+        metadata={"help": "how often to reset for ReLoRA"},
+    )
+    relora_warmup_steps: Optional[int] = field(
+        default=None,
+        metadata={"help": "how many warmup steps to take after reset for ReLoRA"},
+    )
+    bench_split: Optional[str] = field(
+        default="eval", metadata={"help": "The benchmark split to run on"}
+    )
+    bench_dataset: Optional[str] = field(
+        default="pharaouk/dharma-1/dharma_1_mini.json",
+        metadata={
+            "help": "Benchmark dataset to use: options are `mmlu-zs`, `mmlu-fs`, or the full path to the dataset file"
+        },
+    )
+    do_bench_eval: Optional[bool] = field(
+        default=False, metadata={"help": "Whether to run the Benchmark evaluation."}
+    )
+    max_bench_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "If set, only evaluates on `max_bench_samples` of the benchmark dataset."
+        },
+    )
+    bench_source_max_len: int = field(
+        default=2048, metadata={"help": "Maximum source sequence length for bench."}
+    )
+
+
+class AxolotlTrainer(Trainer):
+    """
+    Extend the base Trainer for axolotl helpers
+    """
+
+    args = None  # type: AxolotlTrainingArguments
+
+    def __init__(self, *args, bench_data_collator=None, **kwargs):
+        self.bench_data_collator = bench_data_collator
+        super().__init__(*args, **kwargs)
+
+    def create_scheduler(
+        self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
+    ):
+        """
+        Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or
+        passed as an argument.
+
+        Args:
+            num_training_steps (int): The number of training steps to do.
+            optimizer (torch.optim.Optimizer): The training optimizer
+        """
+
+        # fmt: off
+        if self.lr_scheduler is None:  # type: ignore  # pylint: disable=access-member-before-definition
+            # fmt: on
+            if (
+                self.args.lr_scheduler_type == "cosine"
+                and self.args.lr_quadratic_warmup is True
+            ):
+                self.lr_scheduler = get_cosine_schedule_with_quadratic_warmup(  # pylint: disable=attribute-defined-outside-init
+                    optimizer,
+                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
+                    num_training_steps=num_training_steps,
+                )
+            else:
+                return super().create_scheduler(num_training_steps, optimizer)
+        return self.lr_scheduler
+
+    def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
+        if self.args.world_size > 1 and self.args.sample_packing:
+            return DistributedSampler(
+                self.train_dataset,
+                num_replicas=self.args.world_size,
+                rank=self.args.process_index,
+                seed=self.args.seed,
+            )
+        return super()._get_train_sampler()
+
+    def _get_eval_sampler(
+        self, eval_dataset: Dataset
+    ) -> Optional[torch.utils.data.Sampler]:
+        if (
+            self.args.world_size > 1
+            and self.args.sample_packing
+            and self.args.eval_sample_packing is not False
+        ):
+            return SequentialDistributedSampler(
+                eval_dataset,
+                num_replicas=self.args.world_size,
+                rank=self.args.process_index,
+                batch_size=self.args.per_device_eval_batch_size,
+            )
+        return super()._get_eval_sampler(eval_dataset)
+
+    def get_train_dataloader(self) -> Union[DataLoader, MultipackDistributedDataloader]:
+        if self.args.sample_packing:
+            train_sampler = self._get_train_sampler()
+            return self.accelerator.prepare(
+                MultipackDistributedDataloader(
+                    self.train_dataset,
+                    batch_size=self._train_batch_size,
+                    seq_max_length=self.args.max_seq_length,
+                    collate_fn=self.data_collator,
+                    sampler=train_sampler,
+                    packing_efficiency_estimate=self.args.sample_packing_efficiency,
+                    sample_packing_seq_len_multiplier=self.args.sample_packing_seq_len_multiplier,
+                    device_count=int(os.environ.get("WORLD_SIZE", 1)),
+                )
+            )
+        return super().get_train_dataloader()
+
+    def get_eval_dataloader(
+        self, eval_dataset: Optional[Dataset] = None
+    ) -> Union[DataLoader, MultipackDistributedDataloader]:
+        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
+            )
+
+            eval_sampler = self._get_eval_sampler(eval_dataset)
+            return self.accelerator.prepare(
+                MultipackDistributedDataloader(
+                    eval_dataset,
+                    batch_size=self.args.eval_batch_size,
+                    seq_max_length=self.args.max_seq_length,
+                    collate_fn=self.data_collator,
+                    sampler=eval_sampler,
+                    packing_efficiency_estimate=self.args.sample_packing_efficiency,
+                    sample_packing_seq_len_multiplier=self.args.eval_batch_size,
+                    device_count=int(os.environ.get("WORLD_SIZE", 1)),
+                )
+            )
+        return super().get_eval_dataloader(eval_dataset)
+
+    def _get_bench_sampler(
+        self, bench_dataset: Dataset
+    ) -> Optional[torch.utils.data.Sampler]:
+        if self.args.world_size <= 1:
+            return SequentialSampler(bench_dataset)
+        return None
+
+    def get_bench_dataloader(
+        self,
+        bench_dataset: Dataset,
+    ) -> Union[DataLoader, MultipackDistributedDataloader]:
+        dataloader_params = {
+            "batch_size": self.args.eval_batch_size,
+            "collate_fn": self.bench_data_collator,
+            "num_workers": self.args.dataloader_num_workers,
+            "pin_memory": self.args.dataloader_pin_memory,
+        }
+
+        if not isinstance(bench_dataset, torch.utils.data.IterableDataset):
+            dataloader_params["sampler"] = self._get_bench_sampler(bench_dataset)
+            dataloader_params["drop_last"] = self.args.dataloader_drop_last
+
+        return DataLoader(bench_dataset, **dataloader_params)
+        # return self.accelerator.prepare(DataLoader(bench_dataset, **dataloader_params))
+
+    def compute_loss(self, model, inputs, return_outputs=False):
+        # use one's weighted cross entropy loss calc
+        # if self.args.sample_packing:
+        #     labels = inputs.pop("labels")
+        #     outputs = model(**inputs)
+        #     loss = trainer_weighted_loss(outputs, labels, shift_labels=True)
+        #     return (loss, outputs) if return_outputs else loss
+        return super().compute_loss(model, inputs, return_outputs=return_outputs)
+
+
+class OneCycleLRSchedulerTrainer(AxolotlTrainer):
+    """
+    Trainer subclass that uses the OneCycleLR scheduler
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.lr_scheduler = None
+
+    def create_scheduler(
+        self,
+        num_training_steps: int,
+        optimizer: Optional[torch.optim.Optimizer] = None,
+    ):
+        optimizer = self.optimizer if optimizer is None else optimizer
+        num_warmup_steps = self.args.get_warmup_steps(num_training_steps)
+        pct_start = num_warmup_steps / num_training_steps
+
+        self.lr_scheduler = OneCycleLR(
+            optimizer,
+            max_lr=self.args.learning_rate,
+            total_steps=num_training_steps,
+            pct_start=pct_start,
+            div_factor=6,
+        )
+
+        return self.lr_scheduler
+
+
+class ReLoRATrainer(AxolotlTrainer):
+    """
+    Trainer subclass that uses the OneCycleLR scheduler
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.lr_scheduler = None
+
+    def create_scheduler(
+        self,
+        num_training_steps: int,
+        optimizer: Optional[torch.optim.Optimizer] = None,
+    ):
+        optimizer = self.optimizer if optimizer is None else optimizer
+        lr_scheduler = super().create_scheduler(num_training_steps, optimizer)
+
+        if self.args.relora_steps:
+            warmup_steps = (
+                self.args.relora_warmup_steps if self.args.relora_warmup_steps else 10
+            )
+            self.lr_scheduler = ReLoRAScheduler(
+                optimizer,
+                lr_scheduler,
+                self.args.relora_steps,
+                warmup_steps,
+            )
+        else:
+            self.lr_scheduler = lr_scheduler
+
+        return self.lr_scheduler
+
+
 def add_position_ids(sample):
     sample_len = len(sample["input_ids"])
     sample["position_ids"] = torch.arange(len(sample["input_ids"]))
@@ -216,7 +501,6 @@ def calculate_total_num_steps(cfg, train_dataset, tokenizer):
                 packing_efficiency_estimate=cfg.sample_packing_eff_est,
                 sample_packing_seq_len_multiplier=cfg.micro_batch_size,
                 device_count=int(os.environ.get("WORLD_SIZE", 1)),
-                num_epochs=cfg.num_epochs,
             )
             data_loader_len = data_loader.len_w_stats()
             actual_eff = data_loader.efficiency()
@@ -266,8 +550,245 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
     elif cfg.deepspeed:
         os.environ["ACCELERATE_USE_DEEPSPEED"] = "true"
 
-    trainer_builder = HFCausalTrainerBuilder(cfg, model, tokenizer)
-    trainer_builder.train_dataset = train_dataset
-    trainer_builder.eval_dataset = eval_dataset
+    warmup_steps = (
+        cfg.warmup_steps
+        if cfg.warmup_steps is not None
+        else min(int(0.03 * total_num_steps), 100)
+    )
+    logging_steps = (
+        cfg.logging_steps
+        if cfg.logging_steps is not None
+        else max(min(int(0.005 * total_num_steps), 10), 1)
+    )
 
-    return trainer_builder.build(total_num_steps)
+    training_arguments_kwargs = {}
+    if cfg.bf16 == "full":
+        training_arguments_kwargs["bf16_full_eval"] = True
+    else:
+        training_arguments_kwargs["bf16"] = cfg.bf16
+    training_arguments_kwargs["fp16"] = (cfg.fp16 and not cfg.bf16) or False
+    training_arguments_kwargs["tf32"] = cfg.tf32
+    training_arguments_kwargs["warmup_steps"] = warmup_steps
+    training_arguments_kwargs["logging_steps"] = logging_steps
+
+    if cfg.seed:
+        training_arguments_kwargs["seed"] = cfg.seed
+
+    if cfg.gradient_checkpointing:
+        training_arguments_kwargs["gradient_checkpointing"] = cfg.gradient_checkpointing
+    if cfg.fsdp:
+        training_arguments_kwargs["fsdp"] = cfg.fsdp
+        if cfg.fsdp_config:
+            training_arguments_kwargs["fsdp_config"] = dict(cfg.fsdp_config)
+
+    # deepspeed
+    if cfg.deepspeed:
+        training_arguments_kwargs["deepspeed"] = cfg.deepspeed
+
+    if cfg.lr_quadratic_warmup is not None:
+        training_arguments_kwargs["lr_quadratic_warmup"] = cfg.lr_quadratic_warmup
+
+    if cfg.adam_beta1:
+        training_arguments_kwargs["adam_beta1"] = cfg.adam_beta1
+    if cfg.adam_beta2:
+        training_arguments_kwargs["adam_beta2"] = cfg.adam_beta2
+    if cfg.adam_epsilon:
+        training_arguments_kwargs["adam_epsilon"] = cfg.adam_epsilon
+    if cfg.max_grad_norm:
+        training_arguments_kwargs["max_grad_norm"] = cfg.max_grad_norm
+
+    if cfg.hub_model_id:
+        training_arguments_kwargs["hub_model_id"] = cfg.hub_model_id
+        training_arguments_kwargs["push_to_hub"] = True
+        training_arguments_kwargs["hub_private_repo"] = True
+
+        if cfg.hub_strategy:
+            training_arguments_kwargs["hub_strategy"] = cfg.hub_strategy
+
+    if cfg.save_safetensors:
+        training_arguments_kwargs["save_safetensors"] = cfg.save_safetensors
+
+    if cfg.sample_packing_eff_est:
+        training_arguments_kwargs[
+            "sample_packing_efficiency"
+        ] = cfg.sample_packing_eff_est
+
+    if cfg.eval_steps:
+        training_arguments_kwargs["evaluation_strategy"] = "steps"
+        training_arguments_kwargs["eval_steps"] = cfg.eval_steps
+    elif cfg.evaluation_strategy:
+        training_arguments_kwargs["evaluation_strategy"] = cfg.evaluation_strategy
+    elif 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"
+
+    if cfg.save_steps:
+        training_arguments_kwargs["save_strategy"] = "steps"
+        training_arguments_kwargs["save_steps"] = cfg.save_steps
+    elif cfg.save_strategy:
+        training_arguments_kwargs["save_strategy"] = cfg.save_strategy
+    else:
+        # default to saving each epoch if not defined
+        training_arguments_kwargs["save_strategy"] = "epoch"
+
+    if cfg.do_bench_eval:
+        training_arguments_kwargs["do_bench_eval"] = cfg.do_bench_eval
+        if cfg.bench_dataset:
+            training_arguments_kwargs["bench_dataset"] = cfg.bench_dataset
+    if cfg.metric_for_best_model:
+        training_arguments_kwargs["metric_for_best_model"] = cfg.metric_for_best_model
+    if cfg.greater_is_better:
+        training_arguments_kwargs["greater_is_better"] = cfg.greater_is_better
+
+    if cfg.torch_compile:
+        if torch.__version__ < "2.1.0":  # pylint: disable=protected-access
+            LOG.warning("torch>=2.1.0 required for torch_compile to work properly")
+        else:
+            import torch._dynamo  # pylint: disable=redefined-outer-name
+
+            torch._dynamo.config.suppress_errors = (  # pylint: disable=protected-access
+                True
+            )
+            training_arguments_kwargs["torch_compile"] = cfg.torch_compile
+            if cfg.torch_compile_backend:
+                training_arguments_kwargs[
+                    "torch_compile_backend"
+                ] = cfg.torch_compile_backend
+
+    # DDP Config
+    if cfg.ddp_timeout:
+        training_arguments_kwargs["ddp_timeout"] = cfg.ddp_timeout
+    # see https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html
+    if cfg.ddp_bucket_cap_mb:
+        training_arguments_kwargs["ddp_bucket_cap_mb"] = cfg.ddp_bucket_cap_mb
+    if cfg.ddp_broadcast_buffers is not None:
+        training_arguments_kwargs["ddp_broadcast_buffers"] = cfg.ddp_broadcast_buffers
+
+    training_args = AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
+        max_steps=total_num_steps if cfg.max_steps else -1,
+        max_seq_length=cfg.sequence_len,
+        per_device_train_batch_size=cfg.micro_batch_size,
+        per_device_eval_batch_size=cfg.eval_batch_size,
+        gradient_accumulation_steps=cfg.gradient_accumulation_steps,
+        eval_accumulation_steps=cfg.gradient_accumulation_steps,
+        num_train_epochs=cfg.num_epochs,
+        learning_rate=cfg.learning_rate,
+        output_dir=cfg.output_dir,
+        save_total_limit=cfg.save_total_limit if cfg.save_total_limit else 4,
+        load_best_model_at_end=(
+            (cfg.load_best_model_at_end is not False or cfg.early_stopping_patience)
+            and cfg.val_set_size > 0
+            and cfg.save_steps
+            and cfg.eval_steps
+            and cfg.save_steps % cfg.eval_steps == 0
+        )
+        or False,
+        ddp_find_unused_parameters=False if cfg.ddp else None,
+        group_by_length=cfg.group_by_length,
+        report_to="wandb" if cfg.use_wandb else None,
+        run_name=cfg.wandb_run_id if cfg.use_wandb else None,
+        optim=cfg.optimizer if cfg.optimizer else "adamw_hf",
+        lr_scheduler_type=cfg.lr_scheduler
+        if cfg.lr_scheduler and cfg.lr_scheduler not in ("one_cycle", "log_sweep")
+        else "cosine",
+        weight_decay=cfg.weight_decay if cfg.weight_decay is not None else 0.0,
+        sample_packing=cfg.sample_packing if cfg.sample_packing else False,
+        eval_sample_packing=cfg.eval_sample_packing,
+        sample_packing_seq_len_multiplier=cfg.micro_batch_size,
+        relora_steps=cfg.relora_steps,
+        relora_warmup_steps=cfg.relora_warmup_steps,
+        **training_arguments_kwargs,
+    )
+
+    trainer_kwargs = {}
+
+    if cfg.optimizer == "adamw_anyprecision":
+        if Path(cfg.torchdistx_path).exists():
+            sys.path.append(cfg.torchdistx_path)
+            importlib.import_module("torchdistx")
+
+    callbacks = []
+    callbacks.append(GPUStatsCallback(cfg))
+    callbacks.append(EvalFirstStepCallback)
+
+    if cfg.relora_steps:
+        callbacks.append(ReLoRACallback(cfg))
+
+    if hasattr(model, "use_bettertransformer") and model.use_bettertransformer is True:
+        callbacks.append(SaveBetterTransformerModelCallback)
+
+    data_collator_kwargs = {
+        "padding": True,  # True/"longest" is the default
+    }
+    if cfg.pad_to_sequence_len:
+        data_collator_kwargs["pad_to_multiple_of"] = 64 * math.ceil(
+            cfg.sequence_len / 64
+        )
+    else:
+        # A100 is best at 64, while others at 8. Let's use the larger so we don't have to check
+        # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
+        data_collator_kwargs["pad_to_multiple_of"] = 64
+
+    if cfg.is_llama_derived_model and cfg.landmark_attention:
+        from axolotl.monkeypatch.llama_landmark_attn import (
+            add_mem_tokens,
+            get_mem_id,
+            set_model_mem_id,
+        )
+
+        set_model_mem_id(model, tokenizer)
+
+        LOG.info("Adding landmark attention tokens to dataset")
+
+        for dataset in [train_dataset, eval_dataset]:
+            dataset = dataset.map(
+                partial(add_mem_tokens, mem_freq=50, mem_id=get_mem_id(tokenizer)),
+                batched=False,
+                num_proc=32,
+            )
+
+    trainer_cls = AxolotlTrainer
+    if cfg.lr_scheduler == "one_cycle" and (cfg.fsdp or cfg.adapter == "qlora"):
+        trainer_cls = OneCycleLRSchedulerTrainer
+    elif cfg.relora_steps:
+        trainer_cls = ReLoRATrainer
+    trainer = trainer_cls(
+        model=model,
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        args=training_args,
+        data_collator=DataCollatorForSeq2Seq(
+            tokenizer,
+            return_tensors="pt",
+            **data_collator_kwargs,
+        ),
+        bench_data_collator=transformers.DataCollatorForSeq2Seq(
+            tokenizer,
+            return_tensors="pt",
+            **data_collator_kwargs,
+        ),
+        callbacks=callbacks,
+        **trainer_kwargs,
+    )
+
+    if cfg.use_wandb and cfg.eval_table_size > 0:
+        LogPredictionCallback = log_prediction_callback_factory(trainer, tokenizer)
+        trainer.add_callback(LogPredictionCallback(cfg))
+
+    if cfg.use_wandb:
+        trainer.add_callback(SaveAxolotlConfigtoWandBCallback(cfg.axolotl_config_path))
+
+    if cfg.do_bench_eval:
+        trainer.add_callback(bench_eval_callback_factory(trainer, tokenizer))
+
+    # TODO on_save callback to sync checkpoints to GCP/AWS in background
+    if cfg.early_stopping_patience:
+        early_stop_cb = EarlyStoppingCallback(
+            cfg.early_stopping_patience,
+        )
+        trainer.add_callback(early_stop_cb)
+
+    return trainer

tests/e2e/test_fused_llama.py

@@ -1,72 +0,0 @@
-"""
-E2E tests for lora llama
-"""
-
-import logging
-import os
-import tempfile
-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
-
-LOG = logging.getLogger("axolotl.tests.e2e")
-os.environ["WANDB_DISABLED"] = "true"
-
-
-class TestFusedLlama(unittest.TestCase):
-    """
-    Test case for Llama models using Fused layers
-    """
-
-    def test_fft_packing(self):
-        # pylint: disable=duplicate-code
-        output_dir = tempfile.mkdtemp()
-        cfg = DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "flash_attention": True,
-                "flash_attn_fuse_qkv": True,
-                "flash_attn_fuse_mlp": True,
-                "sample_packing": True,
-                "sequence_len": 1024,
-                "val_set_size": 0.1,
-                "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
-                },
-                "datasets": [
-                    {
-                        "path": "mhenrichsen/alpaca_2k_test",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 2,
-                "micro_batch_size": 2,
-                "gradient_accumulation_steps": 1,
-                "output_dir": output_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_torch",
-                "lr_scheduler": "cosine",
-                "max_steps": 20,
-                "save_steps": 10,
-                "eval_steps": 10,
-            }
-        )
-        if is_torch_bf16_gpu_available():
-            cfg.bf16 = True
-        else:
-            cfg.fp16 = True
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(output_dir) / "pytorch_model.bin").exists()

tests/e2e/test_lora_llama.py

@@ -24,11 +24,16 @@ class TestLoraLlama(unittest.TestCase):
     """
 
     def test_lora(self):
+        """
+        support for legacy lora_ configs
+        :return:
+        """
         # pylint: disable=duplicate-code
         output_dir = tempfile.mkdtemp()
         cfg = DictDefault(
             {
                 "base_model": "JackFram/llama-68m",
+                "base_model_config": "JackFram/llama-68m",
                 "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
@@ -65,12 +70,108 @@ class TestLoraLlama(unittest.TestCase):
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(output_dir) / "adapter_model.bin").exists()
 
+    def test_lora_peft(self):
+        """
+        support for legacy lora_ configs
+        :return:
+        """
+        # pylint: disable=duplicate-code
+        output_dir = tempfile.mkdtemp()
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "base_model_config": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "peft_r": 32,
+                "peft_alpha": 64,
+                "peft_dropout": 0.05,
+                "peft_target_linear": True,
+                "val_set_size": 0.1,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": output_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+            }
+        )
+        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(output_dir) / "adapter_model.bin").exists()
+
+    def test_ia3_peft(self):
+        """
+        support for IA3 peft
+        :return:
+        """
+        # pylint: disable=duplicate-code
+        output_dir = tempfile.mkdtemp()
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "base_model_config": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "ia3",
+                "peft_r": 32,
+                "peft_alpha": 64,
+                "peft_dropout": 0.05,
+                "peft_target_modules": ["k_proj", "v_proj", "down_proj"],
+                "peft_feedforward_modules": ["down_proj"],
+                "val_set_size": 0.1,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": output_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+            }
+        )
+        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(output_dir) / "adapter_model.bin").exists()
+
     def test_lora_packing(self):
         # pylint: disable=duplicate-code
         output_dir = tempfile.mkdtemp()
         cfg = DictDefault(
             {
                 "base_model": "JackFram/llama-68m",
+                "base_model_config": "JackFram/llama-68m",
                 "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "sample_packing": True,
@@ -115,6 +216,7 @@ class TestLoraLlama(unittest.TestCase):
         cfg = DictDefault(
             {
                 "base_model": "TheBlokeAI/jackfram_llama-68m-GPTQ",
+                "base_model_config": "TheBlokeAI/jackfram_llama-68m-GPTQ",
                 "model_type": "AutoModelForCausalLM",
                 "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,

tests/e2e/test_mistral.py

@@ -31,6 +31,7 @@ class TestMistral(unittest.TestCase):
         cfg = DictDefault(
             {
                 "base_model": "openaccess-ai-collective/tiny-mistral",
+                "base_model_config": "openaccess-ai-collective/tiny-mistral",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "load_in_8bit": True,
@@ -76,6 +77,7 @@ class TestMistral(unittest.TestCase):
         cfg = DictDefault(
             {
                 "base_model": "openaccess-ai-collective/tiny-mistral",
+                "base_model_config": "openaccess-ai-collective/tiny-mistral",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "val_set_size": 0.1,

tests/e2e/test_mistral_samplepack.py

@@ -31,6 +31,7 @@ class TestMistral(unittest.TestCase):
         cfg = DictDefault(
             {
                 "base_model": "openaccess-ai-collective/tiny-mistral",
+                "base_model_config": "openaccess-ai-collective/tiny-mistral",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 1024,
@@ -77,6 +78,7 @@ class TestMistral(unittest.TestCase):
         cfg = DictDefault(
             {
                 "base_model": "openaccess-ai-collective/tiny-mistral",
+                "base_model_config": "openaccess-ai-collective/tiny-mistral",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 1024,

tests/e2e/test_phi.py

@@ -27,6 +27,7 @@ class TestPhi(unittest.TestCase):
         cfg = DictDefault(
             {
                 "base_model": "microsoft/phi-1_5",
+                "base_model_config": "microsoft/phi-1_5",
                 "trust_remote_code": True,
                 "model_type": "MixFormerSequentialForCausalLM",
                 "tokenizer_type": "AutoTokenizer",
@@ -70,6 +71,7 @@ class TestPhi(unittest.TestCase):
         cfg = DictDefault(
             {
                 "base_model": "microsoft/phi-1_5",
+                "base_model_config": "microsoft/phi-1_5",
                 "trust_remote_code": True,
                 "model_type": "MixFormerSequentialForCausalLM",
                 "tokenizer_type": "AutoTokenizer",

tests/test_cfg_normalization.py

@@ -0,0 +1,48 @@
+"""Module for testing the validation module"""
+
+import logging
+import unittest
+from typing import Optional
+
+import pytest
+
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+
+
+class NormalizationTest(unittest.TestCase):
+    """
+    Test the cfg normalization module
+    """
+
+    _caplog: Optional[pytest.LogCaptureFixture] = None
+
+    @pytest.fixture(autouse=True)
+    def inject_fixtures(self, caplog):
+        self._caplog = caplog
+
+    def test_lora_to_peft(self):
+        base_cfg = DictDefault(
+            {
+                "gradient_accumulation_steps": 1,
+                "micro_batch_size": 1,
+                "base_model": "NousResearch/Llama-2-7b-hf",
+                "base_model_config": "NousResearch/Llama-2-7b-hf",
+            }
+        )
+        cfg = base_cfg | DictDefault(
+            {
+                "adapter": "lora",
+                "lora_r": 128,
+                "lora_alpha": 64,
+            }
+        )
+        with self._caplog.at_level(logging.WARNING):
+            normalize_config(cfg)
+            assert any(
+                "soon to be deprecated. please use peft_" in record.message
+                for record in self._caplog.records
+            )
+
+            assert cfg.peft_r == 128
+            assert cfg.peft_alpha == 64

tests/test_normalize_config.py

@@ -1,46 +0,0 @@
-"""
-Test classes for checking functionality of the cfg normalization
-"""
-import unittest
-
-from axolotl.utils.config import normalize_config
-from axolotl.utils.dict import DictDefault
-
-
-class NormalizeConfigTestCase(unittest.TestCase):
-    """
-    test class for normalize_config checks
-    """
-
-    def _get_base_cfg(self):
-        return DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "base_model_config": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
-                "num_epochs": 1,
-                "micro_batch_size": 1,
-                "gradient_accumulation_steps": 1,
-            }
-        )
-
-    def test_lr_as_float(self):
-        cfg = (
-            self._get_base_cfg()
-            | DictDefault(  # pylint: disable=unsupported-binary-operation
-                {
-                    "learning_rate": "5e-5",
-                }
-            )
-        )
-
-        normalize_config(cfg)
-
-        assert cfg.learning_rate == 0.00005
-
-    def test_base_model_config_set_when_empty(self):
-        cfg = self._get_base_cfg()
-        del cfg.base_model_config
-        normalize_config(cfg)
-
-        assert cfg.base_model_config == cfg.base_model

tests/test_validation.py

@@ -565,87 +565,3 @@ class ValidationTest(unittest.TestCase):
         )
 
         validate_config(cfg)
-
-    def test_eval_table_size_conflict_eval_packing(self):
-        cfg = DictDefault(
-            {
-                "sample_packing": True,
-                "eval_table_size": 100,
-            }
-        )
-
-        with pytest.raises(
-            ValueError, match=r".*Please set 'eval_sample_packing' to false.*"
-        ):
-            validate_config(cfg)
-
-        cfg = DictDefault(
-            {
-                "sample_packing": True,
-                "eval_sample_packing": False,
-            }
-        )
-
-        validate_config(cfg)
-
-        cfg = DictDefault(
-            {
-                "sample_packing": False,
-                "eval_table_size": 100,
-            }
-        )
-
-        validate_config(cfg)
-
-        cfg = DictDefault(
-            {
-                "sample_packing": True,
-                "eval_table_size": 100,
-                "eval_sample_packing": False,
-            }
-        )
-
-        validate_config(cfg)
-
-    def test_load_in_x_bit_without_adapter(self):
-        cfg = DictDefault(
-            {
-                "load_in_4bit": True,
-            }
-        )
-
-        with pytest.raises(
-            ValueError,
-            match=r".*load_in_8bit and load_in_4bit are not supported without setting an adapter.*",
-        ):
-            validate_config(cfg)
-
-        cfg = DictDefault(
-            {
-                "load_in_8bit": True,
-            }
-        )
-
-        with pytest.raises(
-            ValueError,
-            match=r".*load_in_8bit and load_in_4bit are not supported without setting an adapter.*",
-        ):
-            validate_config(cfg)
-
-        cfg = DictDefault(
-            {
-                "load_in_4bit": True,
-                "adapter": "qlora",
-            }
-        )
-
-        validate_config(cfg)
-
-        cfg = DictDefault(
-            {
-                "load_in_8bit": True,
-                "adapter": "lora",
-            }
-        )
-
-        validate_config(cfg)