support for batched sharegpt tokenization to skip bad data

.github/workflows/base.yml

@@ -25,11 +25,6 @@ jobs:
             python_version: "3.10"
             pytorch: 2.0.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
-          - cuda: "118"
-            cuda_version: 11.8.0
-            python_version: "3.10"
-            pytorch: 2.1.0
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
     steps:
       - name: Checkout
         uses: actions/checkout@v3

.github/workflows/main.yml

@@ -23,12 +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"
-            pytorch: 2.1.0
-            axolotl_extras:
     runs-on: [self-hosted, gpu, docker]
     steps:
       - name: Checkout
@@ -52,12 +46,9 @@ jobs:
           build-args: |
             BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
             CUDA=${{ matrix.cuda }}
-            PYTORCH_VERSION=${{ matrix.pytorch }}
           file: ./docker/Dockerfile
           push: ${{ github.event_name != 'pull_request' }}
-          tags: |
-            ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
-            ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
+          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
@@ -77,11 +68,6 @@ jobs:
             pytorch: 2.0.1
             axolotl_extras:
             is_latest: true
-          - cuda: 118
-            cuda_version: 11.8.0
-            python_version: "3.10"
-            pytorch: 2.1.0
-            axolotl_extras:
     runs-on: [self-hosted, gpu, docker]
     steps:
       - name: Checkout

.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-nested-blocks,

README.md

@@ -23,10 +23,9 @@ Features:
 - [Supported Features](#axolotl-supports)
 - [Quickstart](#quickstart-)
 - [Installation](#installation)
-  - [Docker](#docker)
-  - [Conda/Pip venv](#condapip-venv)
-  - [LambdaLabs](#lambdalabs)
-  - [Windows](#windows)
+  - [Docker Installation](#environment)
+  - [Conda/Pip venv Installation](#condapip-venv)
+  - [LambdaLabs Installation](#lambdalabs)
 - [Dataset](#dataset)
   - [How to Add Custom Prompts](#how-to-add-custom-prompts)
   - [How to Use Custom Pretokenized Dataset](#how-to-use-your-custom-pretokenized-dataset)
@@ -51,7 +50,7 @@ Features:
       <b>Axolotl provides a unified repository for fine-tuning <br />a variety of AI models with ease</b>
     </p>
     <p>
-      Go ahead and Axolotl questions!!
+      Go ahead and axolotl questions!!
     </p>
     <img src="https://github.com/OpenAccess-AI-Collective/axolotl/actions/workflows/pre-commit.yml/badge.svg?branch=main" alt="pre-commit">
     <img alt="PyTest Status" src="https://github.com/OpenAccess-AI-Collective/axolotl/actions/workflows/tests.yml/badge.svg?branch=main">
@@ -103,7 +102,7 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
 
 ### Environment
 
-#### Docker
+- Docker
   ```bash
   docker run --gpus '"all"' --rm -it winglian/axolotl:main-py3.10-cu118-2.0.1
   ```
@@ -115,12 +114,12 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
   docker compose up -d
   ```
 
-#### Conda/Pip venv
+- Conda/Pip venv
   1. Install python >=**3.9**
 
   2. Install pytorch stable https://pytorch.org/get-started/locally/
 
-  3. Install Axolotl along with python dependencies
+  3. Install axolotl along with python dependencies
         ```bash
         pip3 install packaging
         pip3 install -e '.[flash-attn,deepspeed]'
@@ -131,7 +130,7 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
         ```
         Get the token at huggingface.co/settings/tokens
 
-#### LambdaLabs
+- LambdaLabs
   <details>
 
   <summary>Click to Expand</summary>
@@ -175,8 +174,7 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
   ```
   </details>
 
-#### Windows
-Please use WSL or Docker!
+- Windows: Please use WSL or Docker!
 
 ### Dataset
 
@@ -297,24 +295,25 @@ Have dataset(s) in one of the following format (JSONL recommended):
 
 #### How to add custom prompts
 
-For a dataset that is preprocessed for instruction purposes:
-
-```json
-{"instruction": "...", "output": "..."}
-```
-
-You can use this example in your YAML config:
-
+Using yaml. Example:
 ```yaml
 datasets:
   - path: repo
     type:
       system_prompt: ""
-      field_system: system
-      format: "[INST] {instruction} [/INST]"
-      no_input_format: "[INST] {instruction} [/INST]"
+      no_input_format: |-
+        User: {instruction}<|end_of_turn|>
+        Assistant:
+      format: |-
+        User: {instruction}
+        {input}<|end_of_turn|>
+        Assistant:
 ```
 
+Using file:
+1. Add your method to a file in [prompt_strategies](src/axolotl/prompt_strategies). Please see other files as example.
+2. Use your custom file name as the dataset type `<prompt_strategies_file>.load_<load_fn>`.
+
 #### How to use your custom pretokenized dataset
 
 - Do not pass a `type:`
@@ -397,15 +396,15 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
 <summary>All yaml options</summary>
 
 ```yaml
-# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files
-# This can also be a relative path to a model on disk
+# this is the huggingface model that contains *.pt, *.safetensors, or *.bin files
+# this can also be a relative path to a model on disk
 base_model: ./llama-7b-hf
-# You can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)
+# you can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)
 base_model_ignore_patterns:
-# If the base_model repo on hf hub doesn't include configuration .json files,
-# You can set that here, or leave this empty to default to base_model
+# if the base_model repo on hf hub doesn't include configuration .json files,
+# you can set that here, or leave this empty to default to base_model
 base_model_config: ./llama-7b-hf
-# You can specify to choose a specific model revision from huggingface hub
+# you can specify to choose a specific model revision from huggingface hub
 model_revision:
 # Optional tokenizer configuration override in case you want to use a different tokenizer
 # than the one defined in the base model
@@ -420,24 +419,23 @@ trust_remote_code:
 tokenizer_use_fast:
 # Whether to use the legacy tokenizer setting, defaults to True
 tokenizer_legacy:
-# Resize the model embeddings when new tokens are added to multiples of 32
-# This is reported to improve training speed on some models
+# resize the model embeddings when new tokens are added to multiples of 32
+# this is reported to improve training speed on some models
 resize_token_embeddings_to_32x:
 
-# Used to identify which the model is based on
+# used to identify which the model is based on
 is_falcon_derived_model:
 is_llama_derived_model:
-# Please note that if you set this to true, `padding_side` will be set to "left" by default
 is_mistral_derived_model:
 
-# Whether you are training a 4-bit GPTQ quantized model
+# whether you are training a 4-bit GPTQ quantized model
 gptq: true
 gptq_groupsize: 128 # group size
 gptq_model_v1: false # v1 or v2
 
-# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer
+# this will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer
 load_in_8bit: true
-# Use bitsandbytes 4 bit
+# use bitsandbytes 4 bit
 load_in_4bit:
 
 # Use CUDA bf16
@@ -451,9 +449,9 @@ tf32: true # require >=ampere
 bfloat16: true # require >=ampere
 float16: true
 
-# A list of one or more datasets to finetune the model with
+# a list of one or more datasets to finetune the model with
 datasets:
-  # HuggingFace dataset repo | "json" for local dataset, make sure to fill data_files
+  # hf dataset repo | "json" for local dataset, make sure to fill data_files
   - path: vicgalle/alpaca-gpt4
   # The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection]
     type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
@@ -463,18 +461,17 @@ datasets:
     name: # Optional[str] name of dataset configuration to load
     conversation:  # Optional[str] fastchat conversation type, only used with type: sharegpt
 
-  # Custom user prompt
+  # custom user prompt
   - path: repo
     type:
-      # The below are defaults. only set what's needed.
+      # the below are defaults. only set what's needed.
       system_prompt: ""
-      system_format: "{system}"
       field_system: system
       field_instruction: instruction
-      field_input: input
-      field_output: output
+      field_output: input
 
-      # Customizable to be single line or multi-line
+      # customizable to be single line or multi-line
+      system_format: "{system}"
       # 'format' can include {input}
       format: |-
         User: {instruction} {input}
@@ -482,13 +479,13 @@ datasets:
       # 'no_input_format' cannot include {input}
       no_input_format: "{instruction} "
 
-      # For `completion` datsets only, uses the provided field instead of `text` column
+      # for completions datsets, uses the provided field if not `text`
       field:
 
-# Axolotl attempts to save the dataset as an arrow after packing the data together so
+# axolotl attempts to save the dataset as an arrow after packing the data together so
 # subsequent training attempts load faster, relative path
 dataset_prepared_path: data/last_run_prepared
-# Push prepared dataset to hub
+# push prepared dataset to hub
 push_dataset_to_hub: # repo path
 # The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`
 # if not set.
@@ -498,8 +495,8 @@ hub_model_id: # repo path to push finetuned model
 # how to push checkpoints to hub
 # https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy
 hub_strategy:
-# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets
-# Required to be true when used in combination with `push_dataset_to_hub`
+# whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets
+# required to be true when used in combination with `push_dataset_to_hub`
 hf_use_auth_token: # boolean
 # How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.
 val_set_size: 0.04
@@ -508,34 +505,30 @@ dataset_shard_num:
 # Index of shard to use for whole dataset
 dataset_shard_idx:
 
-# The maximum length of an input to train with, this should typically be less than 2048
+# the maximum length of an input to train with, this should typically be less than 2048
 # as most models have a token/context limit of 2048
 sequence_len: 2048
-# Pad inputs so each step uses constant sized buffers
-# This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently
+# pad inputs so each step uses constant sized buffers
+# this will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently
 pad_to_sequence_len:
-# Max sequence length to concatenate training samples together up to
-# Inspired by StackLLaMA. see https://huggingface.co/blog/stackllama#supervised-fine-tuning
+# max sequence length to concatenate training samples together up to
+# inspired by StackLLaMA. see https://huggingface.co/blog/stackllama#supervised-fine-tuning
 # FutureWarning: This will soon be DEPRECATED
 max_packed_sequence_len: 1024
-# Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'
+# use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'
 sample_packing:
-# Set to 'false' if getting errors during eval with sample_packing on.
+# set to 'false' if getting errors during eval with sample_packing on.
 eval_sample_packing:
-# You can set these packing optimizations AFTER starting a training at least once.
+# you can set these packing optimizations AFTER starting a training at least once.
 # The trainer will provide recommended values for these values.
 sample_packing_eff_est:
 total_num_tokens:
 
-# If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model
+# if you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model
 adapter: lora
-# If you already have a lora model trained that you want to load, put that here.
-# This means after training, if you want to test the model, you should set this to the value of `lora_out_dir`.
+# if you already have a lora model trained that you want to load, put that here
+# lora hyperparameters
 lora_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
@@ -547,48 +540,36 @@ lora_target_modules:
 #  - gate_proj
 #  - down_proj
 #  - up_proj
-lora_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_target_linear: # if true, will target all linear layers
 lora_modules_to_save:
 #  - embed_tokens
 #  - lm_head
-
-# Once you complete training, the model will be saved to the following directory.
-# If you merge the adapter to the base model, a subdirectory `merged` will be created under this directory.
-# Make sure `lora_model_dir` points to this directory if you want to use the trained model.
 lora_out_dir:
 lora_fan_in_fan_out: false
 
 # ReLoRA configuration
-# Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
-relora_steps: # Number of steps per ReLoRA restart
-relora_warmup_steps: # Number of per-restart warmup steps
-relora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings
+# must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
+relora_steps: # number of steps per ReLoRA restart
+relora_warmup_steps: # number of per-restart warmup steps
+relora_cpu_offload: # true to perform lora weight merges on cpu during restarts, for modest gpu memory savings
 
 # wandb configuration if you're using it
 wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
-wandb_project: # Your wandb project name
-wandb_entity: # A wandb Team name if using a Team
+wandb_project: # your wandb project name
+wandb_entity: # a wandb Team name if using a Team
 wandb_watch:
-wandb_run_id: # Set the name of your wandb run
+wandb_run_id: # set the name of your wandb run
 wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only at the end of training
 
-# Where to save the full-finetuned model to
+# where to save the finished model to
 output_dir: ./completed-model
 
-# Whether to use torch.compile and which backend to use
+# whether to use torch.compile and which backend to use
 torch_compile:  # bool
 torch_compile_backend:  # Optional[str]
 
-# Training hyperparameters
-
-# If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.
+# training hyperparameters
 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: 3
@@ -596,47 +577,44 @@ 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
-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.
-# e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 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
+save_total_limit: # checkpoints saved at a time
 max_steps:
 
-eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
-eval_table_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
+eval_table_size: # approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
+eval_table_max_new_tokens: # total number of tokens generated for predictions sent to wandb. Default is 128
 
-# Save model as safetensors (require safetensors package)
+# save model as safetensors (require safetensors package)
 save_safetensors:
 
-# Whether to mask out or include the human's prompt from the training labels
+# whether to mask out or include the human's prompt from the training labels
 train_on_inputs: false
-# Group similarly sized data to minimize padding.
-# May be slower to start, as it must download and sort the entire dataset.
-# Note that training loss may have an oscillating pattern with this enabled.
+# group similarly sized data to minimize padding
+# may be slower to start, as it must download and sort the entire dataset
+# note that training loss may have an oscillating pattern with this enabled
 group_by_length: false
 
 # Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 gradient_checkpointing: false
 
-# Stop training after this many evaluation losses have increased in a row
+# stop training after this many evaluation losses have increased in a row
 # https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback
 early_stopping_patience: 3
 
-# Specify a scheduler and kwargs to use with the optimizer
+# specify a scheduler and kwargs to use with the optimizer
 lr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine
 lr_scheduler_kwargs:
 
-# For one_cycle optim
-lr_div_factor: # Learning rate div factor
+# for one_cycle optim
+lr_div_factor: # learning rate div factor
 
-# For log_sweep optim
+# for log_sweep optim
 log_sweep_min_lr:
 log_sweep_max_lr:
 
-# Specify optimizer
+# specify optimizer
 # Valid values are driven by the Transformers OptimizerNames class, see:
 # https://github.com/huggingface/transformers/blob/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134
 #
@@ -662,7 +640,7 @@ log_sweep_max_lr:
 # - paged_lion_32bit
 # - paged_lion_8bit
 optimizer:
-# Specify weight decay
+# specify weight decay
 weight_decay:
 # adamw hyperparams
 adam_beta1:
@@ -671,58 +649,49 @@ adam_epsilon:
 # Gradient clipping max norm
 max_grad_norm:
 
-# Augmentation techniques
-# NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings
-# currently only supported on Llama and Mistral
-noisy_embedding_alpha:
-
-# Whether to bettertransformers
+# whether to bettertransformers
 flash_optimum:
-# Whether to use xformers attention patch https://github.com/facebookresearch/xformers:
+# whether to use xformers attention patch https://github.com/facebookresearch/xformers:
 xformers_attention:
-# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
+# whether to use flash attention patch https://github.com/Dao-AILab/flash-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
+# whether to use scaled-dot-product attention
 # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
 sdp_attention:
 # Landmark attention (only llama)
 landmark_attention:
 # xpos RoPE see https://github.com/kaiokendev/cutoff-len-is-context-len/blob/main/util/xpos_rope_llama_monkey_patch.py
-# LLaMA only
+# llama only
 xpos_rope:
 # RoPE Scaling https://github.com/huggingface/transformers/pull/24653
 rope_scaling:
   type: # linear | dynamic
   factor: # float
 
-# Resume from a specific checkpoint dir
+# resume from a specific checkpoint dir
 resume_from_checkpoint:
-# If resume_from_checkpoint isn't set and you simply want it to start where it left off.
-# Be careful with this being turned on between different models.
+# if resume_from_checkpoint isn't set and you simply want it to start where it left off
+# be careful with this being turned on between different models
 auto_resume_from_checkpoints: false
 
-# Don't mess with this, it's here for accelerate and torchrun
+# don't mess with this, it's here for accelerate and torchrun
 local_rank:
 
-# Add or change special tokens.
-# If you add tokens here, you don't need to add them to the `tokens` list.
+# add or change special tokens
 special_tokens:
   # bos_token: "<s>"
   # eos_token: "</s>"
   # unk_token: "<unk>"
-
-# Add extra tokens.
+# add extra tokens
 tokens:
 
 # FSDP
 fsdp:
 fsdp_config:
 
-# Deepspeed config path. e.g., deepspeed/zero3.json
+# Deepspeed config path
 deepspeed:
 
 # Advanced DDP Arguments
@@ -748,66 +717,6 @@ strict:
 
 </details>
 
-<details>
-<summary> Understanding of batch size and gradient accumulation steps </summary>
-<br/>
-Gradient accumulation means accumulating gradients over several mini-batches and updating the model weights afterward. When the samples in each batch are diverse, this technique doesn't significantly impact learning.
-
-This method allows for effective training with larger effective batch sizes without needing proportionally larger memory. Here's why:
-
-1. **Memory Consumption with Batch Size**: The primary reason increasing the batch size impacts memory is due to the storage requirements for intermediate activations. When you forward propagate a batch through a network, you have to store the activations at each layer for each sample in the batch, because these activations are used during backpropagation to compute gradients. Therefore, larger batches mean more activations, leading to greater GPU memory consumption.
-
-2. **Gradient Accumulation**: With gradient accumulation, you're effectively simulating a larger batch size by accumulating gradients over several smaller batches (or micro-batches). However, at any given time, you're only forward and backward propagating a micro-batch. This means you only store activations for the micro-batch, not the full accumulated batch. As a result, you can simulate the effect of a larger batch size without the memory cost of storing activations for a large batch.
-
-**Example 1:**
-Micro batch size: 3
-Gradient accumulation steps: 2
-Number of GPUs: 3
-Total batch size = 3 * 2 * 3 = 18
-
-```
-| GPU 1          | GPU 2          | GPU 3          |
-|----------------|----------------|----------------|
-| S1, S2, S3     | S4, S5, S6     | S7, S8, S9     |
-| e1, e2, e3     | e4, e5, e6     | e7, e8, e9     |
-|----------------|----------------|----------------|
-| → (accumulate) | → (accumulate) | → (accumulate) |
-|----------------|----------------|----------------|
-| S10, S11, S12  | S13, S14, S15  | S16, S17, S18  |
-| e10, e11, e12  | e13, e14, e15  | e16, e17, e18  |
-|----------------|----------------|----------------|
-| → (apply)      | → (apply)      | → (apply)      |
-
-Accumulated gradient for the weight w1 after the second iteration (considering all GPUs):
-Total gradient for w1 = e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8 + e9 + e10 + e11 + e12 + e13 + e14 + e15 + e16 + e17 + e18
-
-Weight update for w1:
-w1_new = w1_old - learning rate x (Total gradient for w1 / 18)
-```
-
-**Example 2:**
-Micro batch size: 2
-Gradient accumulation steps: 1
-Number of GPUs: 3
-Total batch size = 2 * 1 * 3 = 6
-
-```
-| GPU 1     | GPU 2     | GPU 3     |
-|-----------|-----------|-----------|
-| S1, S2    | S3, S4    | S5, S6    |
-| e1, e2    | e3, e4    | e5, e6    |
-|-----------|-----------|-----------|
-| → (apply) | → (apply) | → (apply) |
-
-Accumulated gradient for the weight w1 (considering all GPUs):
-Total gradient for w1 = e1 + e2 + e3 + e4 + e5 + e6
-
-Weight update for w1:
-w1_new = w1_old - learning rate × (Total gradient for w1 / 6)
-```
-
-</details>
-
 ### Train
 
 Run
@@ -819,7 +728,7 @@ accelerate launch -m axolotl.cli.train your_config.yml
 
 You can optionally pre-tokenize dataset with the following before finetuning:
 ```bash
-CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train your_config.yml --prepare_ds_only
+CUDA_VISIBLE_DEVICES="" accelerate launch -m axolotl.cli.train your_config.yml --prepare_ds_only
 ```
 
 ##### Config
@@ -883,10 +792,6 @@ Pass the appropriate flag to the train command:
     --base_model="./completed-model" --prompter=None --load_in_8bit=True
   ```
 
-Please use `--sample_packing False` if you have it on and receive the error similar to below:
-
-> RuntimeError: stack expects each tensor to be equal size, but got [1, 32, 1, 128] at entry 0 and [1, 32, 8, 128] at entry 1
-
 ### Merge LORA to base
 
 Add below flag to train command above
@@ -903,8 +808,6 @@ 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

docker/Dockerfile

@@ -5,9 +5,6 @@ ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ARG AXOLOTL_EXTRAS=""
 ARG CUDA="118"
 ENV BNB_CUDA_VERSION=$CUDA
-ARG PYTORCH_VERSION="2.0.1"
-
-ENV PYTORCH_VERSION=$PYTORCH_VERSION
 
 RUN apt-get update && \
     apt-get install -y vim curl
@@ -19,7 +16,6 @@ RUN git clone --depth=1 https://github.com/OpenAccess-AI-Collective/axolotl.git
 WORKDIR /workspace/axolotl
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
-RUN sed -i "s/torch==.*/torch==$PYTORCH_VERSION/" requirements.txt
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install -e .[flash-attn,$AXOLOTL_EXTRAS]; \
     else \

docker/Dockerfile-base

@@ -14,7 +14,7 @@ ARG CUDA="118"
 ENV PYTHON_VERSION=$PYTHON_VERSION
 
 RUN apt-get update \
-    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev && rm -rf /var/lib/apt/lists/* \
+    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev && rm -rf /var/lib/apt/lists/*
     && wget \
     https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
     && mkdir /root/.conda \
@@ -57,6 +57,11 @@ FROM base-builder
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
 ENV TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST
 
+# recompile apex
+RUN python3 -m pip uninstall -y apex
+RUN git clone https://github.com/NVIDIA/apex
+RUN cd apex && python3 -m pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" ./
+
 RUN mkdir -p /workspace/builds
 COPY --from=bnb-builder /workspace/bitsandbytes /workspace/builds/bitsandbytes
 

docs/faq.md

@@ -1,14 +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`

docs/llava.md

@@ -1,36 +0,0 @@
-# LLaVA
-
-### Installing dependencies
-
-```shell
-git clone https://github.com/haotian-liu/LLaVA.git
-cd LLaVA
-pip install --no-deps -e .
-```
-
-### Downloading assets
-
-LLaVA doesn't support remote datasets, so both the JSON and image assets need to be downloaded locally
-
-```shell
-mkdir llava
-mkdir data
-cd llava
-curl -L -O https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain/resolve/main/images.zip
-unzip images.zip
-
-cd ../data
-curl -L -O https://huggingface.co/datasets/liuhaotian/LLaVA-Pretrain/resolve/main/blip_laion_cc_sbu_558k.json
-```
-
-### Pretraining
-
-Pretraining aligns the vision model with the language model.
-
-```shell
-accelerate launch -m axolotl.cli.train_mm examples/multimodal/pretrain-llava-llama.yml
-```
-
-### Finetuning
-
-TBD

docs/multipack.md

@@ -1,51 +0,0 @@
-# Multipack
-
-4k context, bsz =4,
-each character represents 256 tokens
-X represents a padding token
-
-```
-   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5
-[[ A A A A A A A A A A A ]
-   B B B B B B ]
-   C C C C C C C ]
-   D D D D ]]
-
-[[ E E E E E E E E ]
- [ F F F F ]
- [ G G G ]
- [ H H H H ]]
-
-[[ I I I ]
- [ J J J ]
- [ K K K K K]
- [ L L L ]]
-```
-
-after padding to longest input in each step
-```
-   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5
-[[ A A A A A A A A A A A ]
-   B B B B B B X X X X X X ]
-   C C C C C C C X X X X ]
-   D D D D X X X X X X X ]]
-
-[[ E E E E E E E E ]
- [ F F F F X X X X ]
- [ G G G X X X X X ]
- [ H H H H X X X X ]]
-
-[[ I I I X X ]
- [ J J J X X ]
- [ K K K K K ]
- [ L L L X X ]]
-```
-
-w packing ( note it's the same effective number of tokens per step, but a true bsz of 1)
-```
-   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5
-[[ A A A A A A A A A A A B B B B B
-   B C C C C C C C D D D D E E E E
-   E E E E F F F F F G G G H H H H
-   I I I J J J J K K K K K L L L X ]]
-```

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

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

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

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

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

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

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

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

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

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

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

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

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

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/fft_optimized.yml

@@ -1,72 +0,0 @@
-base_model: NousResearch/Llama-2-7b-hf
-model_type: LlamaForCausalLM
-tokenizer_type: LlamaTokenizer
-is_llama_derived_model: true
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: mhenrichsen/alpaca_2k_test
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
-output_dir: ./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:
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_run_id:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: true
-fp16: false
-tf32: false
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-flash_attn_cross_entropy: false
-flash_attn_rms_norm: true
-flash_attn_fuse_qkv: false
-flash_attn_fuse_mlp: true
-
-warmup_steps: 100
-eval_steps: 0.05
-eval_table_size:
-save_steps:
-debug:
-deepspeed: #deepspeed/zero2.json # multi-gpu only
-weight_decay: 0.1
-fsdp:
-fsdp_config:
-special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

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

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

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

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

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

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
@@ -15,8 +16,8 @@ val_set_size: 0.01
 output_dir: ./out
 
 sequence_len: 8192
-sample_packing: true
-pad_to_sequence_len: true
+sample_packing:
+pad_to_sequence_len:
 
 wandb_project:
 wandb_entity:
@@ -29,7 +30,7 @@ micro_batch_size: 2
 num_epochs: 3
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
-learning_rate: 0.000005
+learning_rate: 0.0002
 
 train_on_inputs: false
 group_by_length: false
@@ -47,7 +48,7 @@ flash_attention: true
 
 warmup_steps: 10
 eval_steps: 20
-eval_table_size:
+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
@@ -18,8 +19,8 @@ adapter: qlora
 lora_model_dir:
 
 sequence_len: 8192
-sample_packing: true
-pad_to_sequence_len: true
+sample_packing: True
+pad_to_sequence_len: True
 
 lora_r: 32
 lora_alpha: 16
@@ -42,7 +43,7 @@ wandb_run_id:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
-micro_batch_size: 2
+micro_batch_size: 4
 num_epochs: 1
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
@@ -64,7 +65,7 @@ flash_attention: true
 
 warmup_steps: 10
 eval_steps: 20
-eval_table_size:
+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

examples/multimodal/pretrain-llava-llama.yml

@@ -1,66 +0,0 @@
-base_model: NousResearch/Llama-2-7b-hf
-model_type: LlamaForCausalLM
-tokenizer_type: LlamaTokenizer
-is_llama_derived_model: true
-
-# multimodal pretrain
-multimodal: true
-mm_vision_tower: openai/clip-vit-large-patch14
-tune_mm_mlp_adapter: true
-mm_freeze_backbone: true
-mm_vision_select_layer: -2
-mm_projector_type: mlp2x_gelu
-mm_image_folder: ./llava/
-mm_use_im_patch_token: false
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: ./data/blip_laion_cc_sbu_558k.json
-dataset_prepared_path:
-val_set_size: 0.0
-output_dir: ./out
-
-sequence_len: 2048
-sample_packing: false
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_run_id:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_torch
-lr_scheduler: cosine
-learning_rate: 0.002
-
-train_on_inputs: false
-group_by_length: false
-bf16: true
-fp16: false
-tf32: false
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 10
-eval_steps:
-save_steps: 0.1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-  pad_token: "<unk>"

examples/multimodal/pretrain-llava-mistral.yml

@@ -1,66 +0,0 @@
-base_model: mistralai/Mistral-7B-v0.1
-model_type: MistralForCausalLM
-tokenizer_type: LlamaTokenizer
-is_mistral_derived_model: true
-
-# multimodal pretrain
-multimodal: true
-mm_vision_tower: openai/clip-vit-large-patch14
-tune_mm_mlp_adapter: true
-mm_freeze_backbone: true
-mm_vision_select_layer: -2
-mm_projector_type: mlp2x_gelu
-mm_image_folder: ./llava/
-mm_use_im_patch_token: false
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: ./data/blip_laion_cc_sbu_558k.json
-dataset_prepared_path:
-val_set_size: 0.0
-output_dir: ./out
-
-sequence_len: 2048
-sample_packing: false
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_run_id:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_torch
-lr_scheduler: cosine
-learning_rate: 0.002
-
-train_on_inputs: false
-group_by_length: false
-bf16: true
-fp16: false
-tf32: false
-
-gradient_checkpointing: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 10
-eval_steps:
-save_steps:
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-  pad_token: "<unk>"

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

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

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

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

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

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

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
@@ -27,8 +28,8 @@ num_epochs: 3
 learning_rate: 0.00001
 train_on_inputs: false
 group_by_length: false
-bf16: true
-tf32: true
+bf16: True
+tf32: True
 early_stopping_patience:
 resume_from_checkpoint:
 local_rank:

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:

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:

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

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
@@ -16,7 +16,7 @@ flash-attn>=2.3.0
 sentencepiece
 wandb
 einops
-xformers>=0.0.22
+xformers
 optimum
 hf_transfer
 colorama

setup.py

@@ -21,14 +21,6 @@ def parse_requirements():
             ):
                 # Handle standard packages
                 _install_requires.append(line)
-
-    # TODO(wing) remove once xformers release supports torch 2.1.0
-    if "torch==2.1.0" in _install_requires:
-        _install_requires.pop(_install_requires.index("xformers>=0.0.22"))
-        _install_requires.append(
-            "xformers @ git+https://github.com/facebookresearch/xformers.git@main"
-        )
-
     return _install_requires, _dependency_links
 
 
@@ -46,7 +38,7 @@ setup(
     dependency_links=dependency_links,
     extras_require={
         "flash-attn": [
-            "flash-attn>=2.3.0",
+            "flash-attn>=2.2.1",
         ],
         "deepspeed": [
             "deepspeed",

src/axolotl/cli/__init__.py

@@ -2,7 +2,6 @@
 
 import importlib
 import logging
-import math
 import os
 import random
 import sys
@@ -195,7 +194,6 @@ def load_cfg(config: Path = Path("examples/"), **kwargs):
     # load the config from the yaml file
     with open(config, encoding="utf-8") as file:
         cfg: DictDefault = DictDefault(yaml.safe_load(file))
-    cfg.axolotl_config_path = config
     # if there are any options passed in the cli, if it is something that seems valid from the yaml,
     # then overwrite the value
     cfg_keys = cfg.keys()
@@ -216,46 +214,6 @@ def load_cfg(config: Path = Path("examples/"), **kwargs):
     return cfg
 
 
-def load_mm_dataset(
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,  # pylint: disable=unused-argument
-    model,
-):
-    # pylint: disable=duplicate-code
-    from llava.train.train import DataArguments, LazySupervisedDataset
-
-    vision_tower = model.get_vision_tower()
-    data_args = DataArguments(
-        data_path=cfg.datasets[0]["path"],
-        lazy_preprocess=cfg.mm_lazy_preprocess
-        if cfg.mm_lazy_preprocess is not None
-        else True,
-        is_multimodal=True,
-        image_folder=cfg.mm_image_folder or None,
-        image_aspect_ratio=cfg.mm_image_aspect_ratio or "square",
-        image_grid_pinpoints=cfg.mm_image_grid_pinpoints or None,
-    )
-    data_args.image_processor = vision_tower.image_processor
-    data_args.mm_use_im_start_end = cfg.mm_use_im_start_end or False
-    tokenizer = load_tokenizer(cfg)
-    train_dataset = LazySupervisedDataset(
-        tokenizer=tokenizer,
-        data_path=data_args.data_path,
-        data_args=data_args,
-    )
-
-    total_num_steps = int(
-        math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
-    )
-
-    return TrainDatasetMeta(
-        train_dataset=train_dataset,
-        eval_dataset=None,
-        total_num_steps=total_num_steps,
-    )
-
-
 def load_datasets(
     *,
     cfg: DictDefault,

src/axolotl/cli/inference.py

@@ -14,7 +14,6 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
     parsed_cfg = load_cfg(config, **kwargs)
-    parsed_cfg.sample_packing = False
     parser = transformers.HfArgumentParser((TrainerCliArgs))
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True

src/axolotl/cli/train.py

@@ -1,12 +1,10 @@
 """
 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,
@@ -16,11 +14,8 @@ 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")
-
 
 def do_cli(config: Path = Path("examples/"), **kwargs):
     # pylint: disable=duplicate-code
@@ -32,14 +27,6 @@ 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:

src/axolotl/cli/train_mm.py

@@ -1,59 +0,0 @@
-"""
-CLI to run training on a model
-"""
-import logging
-from pathlib import Path
-
-import fire
-import torch
-import transformers
-from colorama import Fore
-
-from axolotl.cli import (
-    check_accelerate_default_config,
-    check_user_token,
-    load_cfg,
-    load_mm_dataset,
-    print_axolotl_text_art,
-)
-from axolotl.common.cli import TrainerCliArgs
-from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
-from axolotl.train import train
-from axolotl.utils.models import load_model, load_tokenizer
-
-LOG = logging.getLogger("axolotl.cli.train")
-
-
-def do_cli(config: Path = Path("examples/"), **kwargs):
-    # pylint: disable=duplicate-code
-    print_axolotl_text_art()
-    parsed_cfg = load_cfg(config, **kwargs)
-    check_accelerate_default_config()
-    check_user_token()
-    parser = transformers.HfArgumentParser((TrainerCliArgs))
-    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
-
-    tokenizer = load_tokenizer(parsed_cfg)
-    model, _ = load_model(parsed_cfg, tokenizer)
-    dataset_meta = load_mm_dataset(
-        cfg=parsed_cfg, cli_args=parsed_cli_args, model=model
-    )
-    del model
-    torch.cuda.empty_cache()
-    if parsed_cli_args.prepare_ds_only:
-        return
-    train(cfg=parsed_cfg, cli_args=parsed_cli_args, dataset_meta=dataset_meta)
-
-
-if __name__ == "__main__":
-    fire.Fire(do_cli)

src/axolotl/common/const.py

@@ -1,5 +0,0 @@
-"""
-Various shared constants
-"""
-
-DEFAULT_DATASET_PREPARED_PATH = "last_run_prepared"

src/axolotl/core/trainer_builder.py

@@ -1,771 +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 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.schedulers import get_cosine_schedule_with_quadratic_warmup
-
-try:
-    import torch._dynamo  # pylint: disable=ungrouped-imports
-except ImportError:
-    pass
-
-try:
-    from llava.train.llava_trainer import get_mm_adapter_state_maybe_zero_3
-except ImportError:
-
-    def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
-        raise ImportError("missing LLaVA package")
-
-
-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."}
-    )
-    tune_mm_mlp_adapter: bool = field(
-        default=False,
-        metadata={"help": "Whether to train the multimodal projector adapter"},
-    )
-    freeze_mm_mlp_adapter: bool = field(
-        default=False,
-        metadata={"help": "Whether to freeze the multimodal projector adapter"},
-    )
-    mm_projector_lr: Optional[float] = field(
-        default=None, metadata={"help": "Learning rate for the multimodal projector"}
-    )
-
-
-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)
-
-    def _save_checkpoint(self, model, trial, metrics=None):
-        if getattr(self.args, "tune_mm_mlp_adapter", False):
-            from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
-
-            checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
-
-            run_dir = self._get_output_dir(trial=trial)
-            output_dir = os.path.join(run_dir, checkpoint_folder)
-
-            weights_to_save = self._get_mm_mlp_adapter_weights()
-
-            if self.args.local_rank in (0, -1):
-                self.model.config.save_pretrained(output_dir)
-                torch.save(
-                    weights_to_save, os.path.join(output_dir, "mm_projector.bin")
-                )
-        else:
-            super()._save_checkpoint(model, trial, metrics)
-
-    def _get_mm_mlp_adapter_weights(self):
-        # Only save Adapter
-        keys_to_match = ["mm_projector", "vision_resampler"]
-        if getattr(self.args, "use_im_start_end", False):
-            keys_to_match.extend(["embed_tokens", "embed_in"])
-
-        return get_mm_adapter_state_maybe_zero_3(
-            self.model.named_parameters(), keys_to_match
-        )
-
-    def _save(self, output_dir: Optional[str] = None, state_dict=None):
-        if getattr(self.args, "tune_mm_mlp_adapter", False):
-            pass
-        else:
-            super()._save(output_dir, state_dict)
-
-
-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):
-        # TODO
-        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
-
-        # multimodal: llava
-        training_arguments_kwargs["tune_mm_mlp_adapter"] = self.cfg.tune_mm_mlp_adapter
-        training_arguments_kwargs[
-            "freeze_mm_mlp_adapter"
-        ] = self.cfg.freeze_mm_mlp_adapter
-        training_arguments_kwargs["mm_projector_lr"] = self.cfg.mm_projector_lr
-
-        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")
-
-        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_collator_kwargs = self.build_data_collator()
-
-        trainer = trainer_cls(
-            model=self.model,
-            train_dataset=self.train_dataset,
-            eval_dataset=self.eval_dataset,
-            args=training_args,
-            callbacks=self.get_callbacks(),
-            **trainer_collator_kwargs,
-            **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
-
-    def build_data_collator(self):
-        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
-
-        collator_kwargs = {}
-        if self.cfg.multimodal:
-            from llava.train.train import DataCollatorForSupervisedDataset
-
-            collator_kwargs["data_collator"] = DataCollatorForSupervisedDataset(
-                tokenizer=self.tokenizer,
-            )
-        else:
-            collator_kwargs["data_collator"] = DataCollatorForSeq2Seq(
-                self.tokenizer,
-                return_tensors="pt",
-                **data_collator_kwargs,
-            )
-
-        if self.cfg.do_bench_eval:
-            collator_kwargs[
-                "bench_data_collator"
-            ] = transformers.DataCollatorForSeq2Seq(
-                self.tokenizer,
-                return_tensors="pt",
-                **data_collator_kwargs,
-            )
-
-        return collator_kwargs

src/axolotl/datasets.py

@@ -5,7 +5,7 @@ import os
 from typing import List
 
 import torch
-from datasets import Dataset, IterableDataset
+from datasets import Dataset, IterableDataset, Sequence, Value
 
 from .prompt_tokenizers import PromptTokenizingStrategy
 
@@ -42,11 +42,15 @@ class TokenizedPromptDataset(Dataset):
         if self.prompt_tokenizer.supports_batched:
             map_kwargs["batched"] = True
             map_kwargs["batch_size"] = 100
-        return dataset.map(
-            self.prompt_tokenizer.tokenize_prompt,
-            num_proc=num_proc,
-            remove_columns=features,
-            **map_kwargs,
+        return (
+            dataset.map(
+                self.prompt_tokenizer.tokenize_prompt,
+                num_proc=num_proc,
+                remove_columns=features,
+                **map_kwargs,
+            )
+            .cast_column("input_ids", Sequence(feature=Value(dtype="int32", id=None)))
+            .cast_column("labels", Sequence(feature=Value(dtype="int32", id=None)))
         )
 
 

src/axolotl/models/llava/llava_mistral.py

@@ -1,167 +0,0 @@
-"""
-LLaVA Mistral classes
-"""
-
-from typing import List, Optional, Tuple, Union
-
-import torch
-from llava.model.llava_arch import LlavaMetaForCausalLM, LlavaMetaModel
-from torch import nn
-from torch.nn import CrossEntropyLoss
-from transformers import (
-    AutoConfig,
-    AutoModelForCausalLM,
-    MistralConfig,
-    MistralForCausalLM,
-    MistralModel,
-)
-from transformers.modeling_outputs import CausalLMOutputWithPast
-
-
-class LlavaMistralConfig(MistralConfig):
-    """
-    HF Transformers Config for Mistral w LLaVA
-    """
-
-    model_type = "llava_mistral"
-
-
-class LlavaMistralModel(LlavaMetaModel, MistralModel):
-    """
-    HF Transformers Model for Mistral w LLaVA
-    """
-
-    config_class = LlavaMistralConfig
-
-    def __init__(
-        self, config: LlavaMistralConfig
-    ):  # pylint: disable=useless-parent-delegation
-        super().__init__(config)
-
-
-class LlavaMistralForCausalLM(MistralForCausalLM, LlavaMetaForCausalLM):
-    """
-    HF Transformers Causal Model for Mistral w LLaVA
-    """
-
-    config_class = LlavaMistralConfig
-
-    def __init__(self, config: LlavaMistralConfig):
-        super().__init__(config)
-        self.model = LlavaMistralModel(config)
-
-        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
-
-        # Initialize weights and apply final processing
-        self.post_init()
-
-    def get_model(self):
-        return self.model
-
-    def forward(
-        self,
-        input_ids: torch.LongTensor = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        past_key_values: Optional[List[torch.FloatTensor]] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        labels: Optional[torch.LongTensor] = None,
-        use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        images: Optional[torch.FloatTensor] = None,
-        return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, CausalLMOutputWithPast]:
-        output_attentions = (
-            output_attentions
-            if output_attentions is not None
-            else self.config.output_attentions
-        )
-        output_hidden_states = (
-            output_hidden_states
-            if output_hidden_states is not None
-            else self.config.output_hidden_states
-        )
-        return_dict = (
-            return_dict if return_dict is not None else self.config.use_return_dict
-        )
-
-        (
-            input_ids,
-            attention_mask,
-            past_key_values,
-            inputs_embeds,
-            labels,
-        ) = self.prepare_inputs_labels_for_multimodal(
-            input_ids, attention_mask, past_key_values, labels, images
-        )
-
-        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
-        outputs = self.model(
-            input_ids=input_ids,
-            attention_mask=attention_mask,
-            past_key_values=past_key_values,
-            inputs_embeds=inputs_embeds,
-            use_cache=use_cache,
-            output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
-            return_dict=return_dict,
-        )
-
-        hidden_states = outputs[0]
-        logits = self.lm_head(hidden_states)
-
-        loss = None
-        if labels is not None:
-            # Shift so that tokens < n predict n
-            shift_logits = logits[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
-            # Flatten the tokens
-            loss_fct = CrossEntropyLoss()
-            shift_logits = shift_logits.view(-1, self.config.vocab_size)
-            shift_labels = shift_labels.view(-1)
-            # Enable model/pipeline parallelism
-            shift_labels = shift_labels.to(shift_logits.device)
-            loss = loss_fct(shift_logits, shift_labels)
-
-        if not return_dict:
-            output = (logits,) + outputs[1:]
-            return (loss,) + output if loss is not None else output
-
-        return CausalLMOutputWithPast(
-            loss=loss,
-            logits=logits,
-            past_key_values=outputs.past_key_values,
-            hidden_states=outputs.hidden_states,
-            attentions=outputs.attentions,
-        )
-
-    def prepare_inputs_for_generation(
-        self,
-        input_ids,
-        past_key_values=None,
-        attention_mask=None,
-        inputs_embeds=None,
-        **kwargs
-    ):
-        if past_key_values:
-            input_ids = input_ids[:, -1:]
-
-        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
-        if inputs_embeds is not None and past_key_values is None:
-            model_inputs = {"inputs_embeds": inputs_embeds}
-        else:
-            model_inputs = {"input_ids": input_ids}
-
-        model_inputs.update(
-            {
-                "past_key_values": past_key_values,
-                "use_cache": kwargs.get("use_cache"),
-                "attention_mask": attention_mask,
-                "images": kwargs.get("images", None),
-            }
-        )
-        return model_inputs
-
-
-AutoConfig.register("llava_mistral", LlavaMistralConfig)
-AutoModelForCausalLM.register(LlavaMistralConfig, LlavaMistralForCausalLM)

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(
@@ -261,14 +147,9 @@ 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)
 
     query_states = query_states.view(
         bsz, q_len, self.num_heads, self.head_dim

src/axolotl/monkeypatch/llama_embeddings_hijack.py

@@ -1,40 +0,0 @@
-"""
-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_attn_hijack_flash.py

@@ -14,9 +14,6 @@ from flash_attn.flash_attn_interface import (  # pylint: disable=ungrouped-impor
     flash_attn_varlen_qkvpacked_func,
 )
 from transformers.modeling_outputs import BaseModelOutputWithPast
-from transformers.models.mistral.modeling_mistral import (
-    MistralAttention as OriginalMistralAttention,
-)
 from transformers.models.mistral.modeling_mistral import (
     MistralDecoderLayer as OriginalMistralDecoderLayer,
 )
@@ -45,44 +42,6 @@ def replace_mistral_attn_with_flash_attn(
         )
 
 
-@torch.jit.script
-def _make_sliding_window_causal_mask(
-    bsz: int,
-    tgt_len: int,
-    dtype: torch.dtype,
-    device: torch.device,
-    past_key_values_length: int = 0,
-    sliding_window: int = 4096,
-):
-    """
-    Make causal mask used for sliding window attention
-    """
-    tensor = torch.full(
-        (tgt_len, tgt_len),
-        fill_value=1,
-        device=device,
-    )
-    mask = torch.tril(tensor, diagonal=0)
-    # make the mask banded to account for sliding window
-    # NOTE: HF implementation is wrong as of 14-10-2023 for torch.triu, needs +1
-    mask = torch.triu(mask, diagonal=-sliding_window + 1)
-    mask = torch.log(mask).to(dtype)
-
-    if past_key_values_length > 0:
-        mask = torch.cat(
-            [
-                torch.zeros(
-                    tgt_len, past_key_values_length, dtype=dtype, device=device
-                ),
-                mask,
-            ],
-            dim=-1,
-        )
-    return mask[None, None, :, :].expand(
-        bsz, 1, tgt_len, tgt_len + past_key_values_length
-    )
-
-
 # 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(
@@ -94,29 +53,11 @@ def _prepare_decoder_attention_mask(
     sliding_window,
 ):  # pylint: disable=unused-argument
     # [bsz, seq_len]
-    if attention_mask is None:
-        return attention_mask
-
-    # NOTE: attention mask and sliding masks are only broadcastable in certain scenarios.
-    # Without attention_mask.shape[0] == 1, error will trigger after eval loss but only when wandb is enabled.
-    if input_shape[-1] > 1 and attention_mask.shape[0] == 1:
-        sliding_window_mask = _make_sliding_window_causal_mask(
-            bsz=input_shape[0],
-            tgt_len=input_shape[1],
-            dtype=inputs_embeds.dtype,
-            device=inputs_embeds.device,
-            past_key_values_length=past_key_values_length,
-            sliding_window=sliding_window,
-        )
-        attention_mask = attention_mask + sliding_window_mask
-    else:
-        LOG.info("skipping sliding window mask, not broadcastable with attention mask")
-
     return attention_mask
 
 
 def flashattn_forward(
-    self: OriginalMistralAttention,
+    self,
     hidden_states: torch.Tensor,
     attention_mask: Optional[torch.Tensor] = None,
     position_ids: Optional[torch.LongTensor] = None,
@@ -150,41 +91,10 @@ def flashattn_forward(
         query_states, key_states, cos, sin, position_ids
     )
 
-    use_sliding_windows = (
-        hasattr(self.config, "sliding_window") is not None
-        and kv_seq_len > self.config.sliding_window
-    )
-
-    if use_sliding_windows:
-        window_size = (self.config.sliding_window, self.config.sliding_window)
-    else:
-        window_size = (-1, -1)
-
     if past_key_value is not None:
-        # Activate slicing cache only if the config has a value `sliding_windows` attribute
-        if (
-            hasattr(self.config, "sliding_window")
-            and kv_seq_len > self.config.sliding_window
-        ):
-            slicing_tokens = kv_seq_len - self.config.sliding_window
-
-            past_key = past_key_value[0]
-            past_value = past_key_value[1]
-
-            past_key = past_key[:, :, slicing_tokens:, :].contiguous()
-            past_value = past_value[:, :, slicing_tokens:, :].contiguous()
-
-            if past_key.shape[-2] != self.config.sliding_window - 1:
-                raise ValueError(
-                    f"past key much have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
-                    f" {past_key.shape}"
-                )
-
-            past_key_value = (past_key, past_value) if use_cache else None
-
-        if past_key_value is not None:
-            key_states = torch.cat([past_key_value[0], key_states], dim=2)
-            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+        # reuse k, v, self_attention
+        key_states = torch.cat([past_key_value[0], key_states], dim=2)
+        value_states = torch.cat([past_key_value[1], value_states], dim=2)
 
     past_key_value = (key_states, value_states) if use_cache else None
 
@@ -210,13 +120,7 @@ def flashattn_forward(
         qkv = rearrange(qkv, "b s ... -> (b s) ...")
 
         output = flash_attn_varlen_qkvpacked_func(
-            qkv,
-            cu_seqlens,
-            max_seqlen,
-            0.0,
-            softmax_scale=None,
-            causal=True,
-            window_size=window_size,
+            qkv, cu_seqlens, max_seqlen, 0.0, softmax_scale=None, causal=True
         )
         output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
     elif query_states.shape == key_states.shape:
@@ -242,7 +146,6 @@ def flashattn_forward(
             0.0,
             softmax_scale=None,
             causal=is_causal,
-            window_size=window_size,
         )
         output = output_pad_fn(output_unpad)
     else:
@@ -254,7 +157,6 @@ def flashattn_forward(
                 query_states,
                 torch.stack([key_states, value_states], 2),
                 causal=is_causal,
-                window_size=window_size,
             )
         else:
             (  # pylint: disable=unbalanced-tuple-unpacking
@@ -289,7 +191,6 @@ def flashattn_forward(
                 0.0,
                 softmax_scale=None,
                 causal=is_causal,
-                window_size=window_size,
             )
             output = output_pad_fn(output_unpad)
 

src/axolotl/monkeypatch/mistral_embeddings_hijack.py

@@ -1,40 +0,0 @@
-"""
-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/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/alpaca_chat.py

@@ -1,6 +1,6 @@
-"""Module for Alpaca prompt strategy classes"""
+"""Module containing the AlpacaQAPromptTokenizingStrategy class"""
 
-from typing import Any, Dict, Optional, Tuple
+from typing import Tuple
 
 from axolotl.prompt_tokenizers import (
     AlpacaPromptTokenizingStrategy,
@@ -9,13 +9,9 @@ from axolotl.prompt_tokenizers import (
 from axolotl.prompters import AlpacaPrompter, PromptStyle, UnpromptedPrompter
 
 
-def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
-    prompt_style = PromptStyle.CHAT.value
-    if ds_cfg and "conversation" in ds_cfg:
-        prompt_style = ds_cfg["conversation"]
-
+def load(tokenizer, cfg):
     return AlpacaPromptTokenizingStrategy(
-        AlpacaPrompter(prompt_style),
+        AlpacaPrompter(PromptStyle.CHAT.value),
         tokenizer,
         cfg.train_on_inputs,
         cfg.sequence_len,

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
-    return SimpleShareGPTPromptTokenizingStrategy(
+    strat = ShareGPTPromptTokenizingStrategy(
         ShareGPTPrompterV2(
             conversation=conversation,
             role_key_model=field_model,
@@ -34,6 +34,9 @@ def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
         cfg.train_on_inputs,
         cfg.sequence_len,
     )
+    if ds_cfg and ds_cfg["skip"]:
+        strat.skip_invalid = True
+    return strat
 
 
 def load_role(tokenizer, cfg):
@@ -54,13 +57,38 @@ def load_guanaco(tokenizer, cfg):
     )
 
 
-class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
+def load_nous(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
+    conversation = (
+        ds_cfg["conversation"] if ds_cfg and "conversation" in ds_cfg else 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
+    return NousShareGPTPromptTokenizingStrategy(
+        ShareGPTPrompterV2(
+            conversation=conversation,
+            role_key_model=field_model,
+            role_key_human=field_human,
+        ),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+
+
+class NousShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
     """
-    basic sharegpt strategy to grab conversations from the sample row
+    basic sharegpt strategy used by nous/ldj for input/output keyed data
     """
 
-    def get_conversation_thread(self, prompt):
-        return prompt["conversations"]
+    def get_conversation_thread(self):
+        return "conversation"
+
+    def map_conversation_thread(self, conversation):
+        turns = []
+        for turn in conversation:
+            turns.append({"from": "human", "value": turn["input"]})
+            turns.append({"from": "gpt", "value": turn["output"]})
+        return turns
 
 
 class SimpleRoleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
@@ -68,10 +96,11 @@ class SimpleRoleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrateg
     basic sharegpt strategy to grab conversations from the sample row, but uses role instead of from
     """
 
-    def get_conversation_thread(self, prompt):
-        conversations = prompt["conversations"]
+    def map_conversation_thread(self, conversation):
         # remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
-        turns = [{"from": t["role"], "value": t["value"]} for t in conversations]
+        turns = [
+            {"from": turn["role"], "value": turn["value"]} for turn in conversation
+        ]
         return turns
 
 
@@ -80,11 +109,11 @@ class GuanacoShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
     sharegpt strategy that remaps oasst data to sharegpt format
     """
 
-    def get_conversation_thread(self, prompt):
-        conversations = prompt["conversations"]
+    def map_conversation_thread(self, conversation):
         # remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
         role_map = {"prompter": "human", "assistant": "gpt"}
         turns = [
-            {"from": role_map[t["role"]], "value": t["text"]} for t in conversations
+            {"from": role_map[turn["role"]], "value": turn["text"]}
+            for turn in conversation
         ]
         return turns

src/axolotl/prompt_tokenizers.py

@@ -2,7 +2,9 @@
 
 import abc
 import copy
+import functools
 import logging
+from collections import defaultdict
 from typing import Dict, List, Tuple, Union
 
 from fastchat.conversation import Conversation
@@ -45,8 +47,6 @@ 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
 
@@ -58,34 +58,57 @@ class PromptTokenizingStrategy(abc.ABC):
     def supports_batched(self):
         return False
 
+    @functools.lru_cache(maxsize=128)
+    def _get_user_token(self):
+        try:
+            id_or_ids = self.tokenizer.convert_tokens_to_ids("<|USER|>")
+            if isinstance(id_or_ids, (int,)):
+                return id_or_ids
+        except KeyError:
+            pass
+        return False
+
+    @functools.lru_cache(maxsize=128)
+    def _get_assistant_token(self):
+        try:
+            id_or_ids = self.tokenizer.convert_tokens_to_ids("<|ASSISTANT|>")
+            if isinstance(id_or_ids, (int,)):
+                return id_or_ids
+        except KeyError:
+            pass
+        return False
+
     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 +144,7 @@ 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
         )
@@ -269,7 +292,7 @@ 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
@@ -329,105 +352,141 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
     Tokenizing strategy for ShareGPT prompts.
     """
 
-    def get_conversation_thread(self, prompt):
-        return prompt["conversations"]
+    _skip_invalid = False
+
+    @property
+    def supports_batched(self):
+        return True
+
+    @property
+    def skip_invalid(self):
+        return self._skip_invalid
+
+    @skip_invalid.setter
+    def skip_invalid(self, value):
+        self._skip_invalid = value
+
+    def get_conversation_thread(self):
+        return "conversations"
+
+    def map_conversation_thread(self, conversation):
+        return conversation
 
     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
-        )
+        tokenized_res = defaultdict(lambda: [])
+        conv_field = self.get_conversation_thread()
+        for prmpt in prompt[conv_field]:
+            result, current_len = tokenize_prompt_default()
+            user_token = self._get_user_token()
+            assistant_token = self._get_assistant_token()
+            conversation: Conversation = (
+                self.prompter._conversation  # pylint: disable=protected-access
+            )
+            try:
+                for _, part in enumerate(
+                    self.prompter.build_prompt(self.map_conversation_thread(prmpt))
+                ):
+                    if isinstance(part, tuple):
+                        if conversation.roles[0] in part[0]:
+                            turn = part[0] + part[1] if not user_token else part[1]
+                            # this is still the user query, we should
+                            if not part[1].strip():
+                                err_msg = f"user turn has empty text: {prmpt}"
+                                if self.skip_invalid:
+                                    raise ValueError(err_msg)
+                                LOG.warning(err_msg)
+                            res = self._tokenize(
+                                turn,
+                                add_eos_token=False,
+                                strip_bos_token=True,
+                            )
+                            if user_token:
+                                res["input_ids"] = [user_token, *res["input_ids"]]
+                            # 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
+                            turn = part[0] + part[1] if not assistant_token else part[1]
+                            # this should be the assistant response, should end with an eos token
+                            if not part[1].strip():
+                                err_msg = f"assistant turn has empty text: {prmpt}"
+                                if self.skip_invalid:
+                                    raise ValueError(err_msg)
+                                LOG.warning(err_msg)
+                            res = self._tokenize(
+                                turn,
+                                add_eos_token=True,
+                                strip_bos_token=True,
+                            )
+                            if assistant_token:
+                                res["input_ids"] = [
+                                    assistant_token,
+                                    *res["input_ids"],
+                                ]
+                            # not masked out from labels
+                            labels = copy.deepcopy(res["input_ids"])
+                        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:
+                            err_msg = f"unhandled role: {part[0]}"
+                            if self.skip_invalid:
+                                raise ValueError(err_msg)
+                            LOG.warning(err_msg)
+                            continue
 
-        # support for custom roles from the dataset, only useful for vicuna style prompts/roles
-        role_remap = []
+                    # pylint: disable=duplicate-code
+                    result, current_len = parse_tokenized_to_result(
+                        result,
+                        current_len,
+                        res,
+                        labels,
+                        pad_token_id=self.tokenizer.pad_token_id,
+                    )
+                for key, val in sorted(result.items(), key=lambda x: x[0]):
+                    tokenized_res[key].append(val)
+            except (KeyError, AssertionError, IndexError) as err:
+                raise InvalidDataException(str(err)) from err
+            except ValueError as err:
+                LOG.warning("skipping prompt: %s", str(err))
+        return tokenized_res
+
+    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 (
-            conversation.name == "vicuna_v1.1"
-            and "roles" in prompt
-            and len(prompt["roles"]) >= 2
+            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
         ):
-            role_remap = [
-                {"from": conversation.roles[0], "to": prompt["roles"][0]},
-                {"from": conversation.roles[1], "to": prompt["roles"][1]},
-            ]
+            result["input_ids"].append(self.tokenizer.eos_token_id)
+            result["attention_mask"].append(1)
 
-        try:
-            for _, part in enumerate(
-                self.prompter.build_prompt(self.get_conversation_thread(prompt))
-            ):
-                if not isinstance(part, tuple):
-                    LOG.warning(f"expected tuple, got {part}")
-                    continue
+        if (
+            len(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:]
 
-                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
-
-                # pylint: disable=duplicate-code
-                result, current_len = parse_tokenized_to_result(
-                    result,
-                    current_len,
-                    res,
-                    labels,
-                    pad_token_id=self.tokenizer.pad_token_id,
-                )
-            return result
-        except (KeyError, AssertionError, IndexError) as err:
-            raise InvalidDataException(str(err)) from err
+        result["labels"] = result["input_ids"].copy()
+        return result
 
 
 def tokenize_prompt_default() -> Tuple[Dict[str, List[int]], int]:

src/axolotl/prompters.py

@@ -274,11 +274,9 @@ class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
             raise err
 
         conv.messages = []
-        for _, sentence in enumerate(source):
+        for j, sentence in enumerate(source):
             role = roles[sentence["from"]]
-            if len(conv.messages) > 0 and (
-                (role == conv.messages[-1][0]) or (role not in conv.roles)
-            ):
+            if role != conv.roles[j % 2]:
                 LOG.warning(f"{SHAREGPT_ASSERTION_FAILED_ROLE}: {sentence}")
             conv.append_message(role, sentence["value"])
 

src/axolotl/train.py

@@ -12,7 +12,6 @@ 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
@@ -20,14 +19,6 @@ from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_model, load_tokenizer
 from axolotl.utils.trainer import setup_trainer
 
-try:
-    from llava.train.train import safe_save_model_for_hf_trainer
-except ImportError:
-
-    def safe_save_model_for_hf_trainer(trainer: transformers.Trainer, output_dir: str):
-        raise ImportError("missing LLaVA package")
-
-
 project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
 src_dir = os.path.join(project_root, "src")
 sys.path.insert(0, src_dir)
@@ -48,7 +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}")
@@ -64,7 +58,9 @@ def train(
 
     safe_serialization = cfg.save_safetensors is True
 
-    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
+    if (
+        cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints
+    ) or cfg.resume_from_checkpoint is True:
         possible_checkpoints = [
             str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
         ]
@@ -77,7 +73,9 @@ def train(
             LOG.info(
                 f"Using Auto-resume functionality to start with checkpoint at {cfg.resume_from_checkpoint}"
             )
-    resume_from_checkpoint = cfg.resume_from_checkpoint
+    resume_from_checkpoint = (
+        cfg.resume_from_checkpoint if cfg.resume_from_checkpoint is not True else None
+    )
 
     trainer = setup_trainer(
         cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps
@@ -125,11 +123,6 @@ def train(
 
     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.")
@@ -145,28 +138,10 @@ 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.multimodal:
-        safe_save_model_for_hf_trainer(trainer=trainer, output_dir=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.args.should_save,
-            save_function=trainer.accelerator.save,
-            state_dict=trainer.accelerator.get_state_dict(trainer.model_wrapped),
-        )
-    elif trainer.args.should_save:
+    elif cfg.local_rank == 0:
         if cfg.flash_optimum:
             model = BetterTransformer.reverse(model)
-        # TODO figure out if `trainer.save_model(cfg.output_dir)` is sufficient here
+
         model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
 
     if not cfg.hub_model_id:

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
@@ -514,27 +514,3 @@ def log_prediction_callback_factory(trainer: Trainer, tokenizer):
             return control
 
     return LogPredictionCallback
-
-
-class SaveAxolotlConfigtoWandBCallback(TrainerCallback):
-    """Callback to save axolotl config to wandb"""
-
-    def __init__(self, axolotl_config_path):
-        self.axolotl_config_path = axolotl_config_path
-
-    def on_train_begin(
-        self,
-        args: AxolotlTrainingArguments,  # pylint: disable=unused-argument
-        state: TrainerState,  # pylint: disable=unused-argument
-        control: TrainerControl,
-        **kwargs,  # pylint: disable=unused-argument
-    ):
-        if is_main_process():
-            try:
-                artifact = wandb.Artifact(name="axolotl-config", type="config")
-                artifact.add_file(local_path=self.axolotl_config_path)
-                wandb.run.log_artifact(artifact)
-                LOG.info("Axolotl config has been saved to WandB as an artifact.")
-            except (FileNotFoundError, ConnectionError) as err:
-                LOG.warning(f"Error while saving Axolotl config to WandB: {err}")
-        return control

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
 
@@ -122,9 +119,6 @@ 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)
 
 
@@ -195,15 +189,9 @@ 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")
 
-    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 cfg.relora_steps:
         if cfg.adapter not in ("lora", "qlora"):
             raise ValueError("cfg.adapter must be lora or qlora to use ReLoRA")
@@ -217,9 +205,6 @@ 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."
@@ -354,30 +339,6 @@ 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.multimodal:
-        try:
-            import llava  # noqa: F401 # pylint:disable=unused-import
-        except ImportError as exc:
-            LOG.warning(
-                "LLaVA package required for multimodal training. See docs/llava.md for more information."
-            )
-            raise exc
-
     # TODO
     # MPT 7b
     # https://github.com/facebookresearch/bitsandbytes/issues/25

src/axolotl/utils/data.py

@@ -16,7 +16,6 @@ from datasets import (
 from huggingface_hub import hf_hub_download
 from transformers import PreTrainedTokenizerBase
 
-from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.datasets import ConstantLengthDataset, TokenizedPromptDataset
 from axolotl.prompt_strategies import load
 from axolotl.prompt_tokenizers import (
@@ -45,6 +44,7 @@ from axolotl.utils.trainer import (
 )
 
 LOG = logging.getLogger("axolotl")
+DEFAULT_DATASET_PREPARED_PATH = "last_run_prepared"
 
 
 def md5(to_hash: str, encoding: str = "utf-8") -> str:
@@ -54,19 +54,8 @@ def md5(to_hash: str, encoding: str = "utf-8") -> str:
         return hashlib.md5(to_hash.encode(encoding)).hexdigest()  # nosec
 
 
-def prepare_dataset(cfg, tokenizer, model=None):
-    if cfg.multimodal:
-        if not model:
-            raise ValueError("missing model argument")
-        from llava.train.train import LazySupervisedDataset
-
-        with zero_first(is_main_process()):
-            eval_dataset = None
-            train_dataset = LazySupervisedDataset(
-                tokenizer=tokenizer,
-            )
-
-    elif not cfg.pretraining_dataset:
+def prepare_dataset(cfg, tokenizer):
+    if not cfg.pretraining_dataset:
         with zero_first(is_main_process()):
             train_dataset, eval_dataset = load_prepare_datasets(
                 tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
@@ -169,21 +158,23 @@ def load_tokenized_prepared_datasets(
                     token=use_auth_token,
                 )
                 ds_from_hub = True
-            except (FileNotFoundError, ConnectionError):
+            except (FileNotFoundError, ValueError):
                 pass
 
             # prefer local dataset, even if hub exists
             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(
-                        d.path,
-                        name=d.name,
-                        data_files=d.data_files,
-                        streaming=False,
-                        split=None,
-                    )
+                    if not d.type:
+                        ds = load_from_disk(d.path)
+                    else:
+                        ds = load_dataset(
+                            d.path,
+                            name=d.name,
+                            data_files=d.data_files,
+                            streaming=False,
+                            split=None,
+                        )
                 elif local_path.is_file():
                     ds_type = "json"
                     if d.ds_type:
@@ -368,7 +359,7 @@ def load_tokenized_prepared_datasets(
         if len(datasets) > 1:
             LOG.info("shuffle merged datasets")
             dataset = dataset.shuffle(seed=seed)
-        if cfg.local_rank == 0:
+        if cfg.local_rank == 0 and cfg.dataset_prepared_path:
             LOG.info(f"Saving merged prepared dataset to disk... {prepared_ds_path}")
             dataset.save_to_disk(prepared_ds_path)
             if cfg.push_dataset_to_hub:

src/axolotl/utils/models.py

@@ -180,26 +180,6 @@ 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,
@@ -255,102 +235,7 @@ def load_model(
             model_kwargs["use_flash_attention_2"] = True
 
     try:
-        if cfg.multimodal:
-            from llava.train.train import DataArguments, ModelArguments
-
-            if cfg.is_llama_derived_model:
-                from llava.model.language_model.llava_llama import LlavaLlamaForCausalLM
-
-                model = LlavaLlamaForCausalLM.from_pretrained(
-                    cfg.base_model,
-                )
-            elif cfg.is_mistral_derived_model:
-                from axolotl.models.llava.llava_mistral import LlavaMistralForCausalLM
-
-                model = LlavaMistralForCausalLM.from_pretrained(
-                    cfg.base_model,
-                )
-            else:
-                raise NotImplementedError(
-                    "unhandled model architecture for multimodal training"
-                )
-
-            if cfg.mm_freeze_backbone:
-                model.model.requires_grad_(False)
-
-            if cfg.gradient_checkpointing:
-                if hasattr(model, "enable_input_require_grads"):
-                    model.enable_input_require_grads()
-                else:
-
-                    def make_inputs_require_grad(
-                        module,
-                        input,
-                        output,
-                    ):  # pylint: disable=redefined-builtin,unused-argument
-                        output.requires_grad_(True)
-
-                    model.get_input_embeddings().register_forward_hook(
-                        make_inputs_require_grad
-                    )
-
-            model_args = ModelArguments(
-                model_name_or_path=cfg.base_model,
-                version="v0",
-                freeze_backbone=cfg.mm_freeze_backbone or False,
-                tune_mm_mlp_adapter=cfg.tune_mm_mlp_adapter or False,
-                vision_tower=cfg.mm_vision_tower,
-                mm_vision_select_layer=cfg.mm_vision_select_layer or -1,
-                pretrain_mm_mlp_adapter=cfg.pretrain_mm_mlp_adapter,
-                mm_projector_type=cfg.mm_projector_type or "linear",
-                mm_use_im_start_end=cfg.mm_use_im_start_end or False,
-                mm_use_im_patch_token=cfg.mm_use_im_patch_token,
-                mm_vision_select_feature=cfg.mm_vision_select_feature or "patch",
-            )
-
-            if cfg.mm_vision_tower is not None:
-                model.get_model().initialize_vision_modules(
-                    model_args=model_args, fsdp=cfg.fsdp
-                )
-
-            vision_tower = model.get_vision_tower()
-            vision_tower.to(dtype=cfg.torch_dtype, device=cfg.device)
-
-            # pylint: disable=duplicate-code
-            data_args = DataArguments(
-                data_path=cfg.datasets[0]["path"],
-                lazy_preprocess=cfg.mm_lazy_preprocess
-                if cfg.mm_lazy_preprocess is not None
-                else True,
-                is_multimodal=True,
-                image_folder=cfg.mm_image_folder or None,
-                image_aspect_ratio=cfg.mm_image_aspect_ratio or "square",
-                image_grid_pinpoints=cfg.mm_image_grid_pinpoints or None,
-            )
-            data_args.image_processor = vision_tower.image_processor
-            model.config.image_aspect_ratio = data_args.image_aspect_ratio
-            model.config.image_grid_pinpoints = data_args.image_grid_pinpoints
-            model.config.tune_mm_mlp_adapter = cfg.tune_mm_mlp_adapter
-            if cfg.tune_mm_mlp_adapter:
-                model.requires_grad_(False)
-                for (
-                    p  # pylint: disable=invalid-name
-                ) in model.get_model().mm_projector.parameters():
-                    p.requires_grad = True
-
-            model.config.freeze_mm_mlp_adapter = cfg.freeze_mm_mlp_adapter
-            if cfg.freeze_mm_mlp_adapter:
-                for (
-                    p  # pylint: disable=invalid-name
-                ) in model.get_model().mm_projector.parameters():
-                    p.requires_grad = False
-
-            model.config.mm_use_im_start_end = (
-                data_args.mm_use_im_start_end
-            ) = cfg.mm_use_im_start_end
-            model.config.mm_use_im_patch_token = cfg.mm_use_im_patch_token
-            model.initialize_vision_tokenizer(model_args, tokenizer=tokenizer)
-        elif cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
+        if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
             from transformers import LlamaForCausalLM
 
             config_kwargs = {}
@@ -367,20 +252,6 @@ 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
@@ -511,7 +382,7 @@ def load_model(
         if model_config.model_type == "btlm":
             # don't upcast lm_head for btlm
             continue
-        if "lm_head" in name or "embed_tokens" in name:
+        if any(x in name for x in ["lm_head", "embed_tokens", "wte", "wpe"]):
             if hasattr(module, "weight"):
                 module.to(torch.float32)
 
@@ -615,19 +486,8 @@ 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()
-    multimodal_keywords = [
-        "mm_projector",
-        "vision_tower",
-        "vision_resampler",
-    ]  # for LLaVA
     for name, module in model.named_modules():
-        if any(mm_keyword in name for mm_keyword in multimodal_keywords):
-            continue
-        if (
-            isinstance(module, cls)
-            or "Linear" in module.__class__.__name__
-            and module.__class__.__name__ not in ("LlamaLinearScalingRotaryEmbedding",)
-        ):
+        if isinstance(module, cls) or "Linear" in module.__class__.__name__:
             names = name.split(".")
             lora_module_names.add(names[0] if len(names) == 1 else names[-1])
 

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 +1,39 @@
 """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 AxolotlTrainer, HFCausalTrainerBuilder
+from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
+from axolotl.utils.callbacks import (
+    EvalFirstStepCallback,
+    GPUStatsCallback,
+    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 +42,7 @@ 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 +109,269 @@ 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"]))
@@ -138,9 +422,7 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset, tokenizer):
                     )
 
         # Phi doesn't want the attention_mask feature when training
-        if "CodeGenTokenizer" in tokenizer.__class__.__name__ or (
-            cfg.is_mistral_derived_model and cfg.flash_attention
-        ):
+        if "CodeGenTokenizer" in tokenizer.__class__.__name__:
             train_dataset = train_dataset.remove_columns("attention_mask")
             if eval_dataset:
                 eval_dataset = eval_dataset.remove_columns("attention_mask")
@@ -259,16 +541,248 @@ def setup_fsdp_envs(cfg):
         ] = cfg.fsdp_config.fsdp_transformer_layer_cls_to_wrap
 
 
-def setup_trainer(
-    cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps
-) -> AxolotlTrainer:
+def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps):
     if cfg.fsdp:
         setup_fsdp_envs(cfg)
     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.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

@@ -29,6 +29,7 @@ class TestLoraLlama(unittest.TestCase):
         cfg = DictDefault(
             {
                 "base_model": "JackFram/llama-68m",
+                "base_model_config": "JackFram/llama-68m",
                 "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
@@ -71,6 +72,7 @@ class TestLoraLlama(unittest.TestCase):
         cfg = DictDefault(
             {
                 "base_model": "JackFram/llama-68m",
+                "base_model_config": "JackFram/llama-68m",
                 "tokenizer_type": "LlamaTokenizer",
                 "sequence_len": 1024,
                 "sample_packing": True,
@@ -115,6 +117,7 @@ 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_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_prompt_tokenizers.py

@@ -90,73 +90,6 @@ class TestPromptTokenizationStrategies(unittest.TestCase):
             strat.tokenize_prompt(conversation)
             assert "assistant turn has empty text" in self._caplog.records[1].message
 
-    def test_sharegpt_warnings_turns(self):
-        conversation = {
-            "conversations": [
-                {"from": "system", "value": "lorem"},
-                {"from": "gpt", "value": "ipsum"},
-                {"from": "human", "value": "dolor"},
-                {"from": "human", "value": "dolor"},
-                {"from": "gpt", "value": "sit"},
-            ]
-        }
-        prompter = ShareGPTPrompterV2()
-        strat = ShareGPTPromptTokenizingStrategy(
-            prompter,
-            self.tokenizer,
-            False,
-            2048,
-        )
-        with self._caplog.at_level(logging.WARNING):
-            strat.tokenize_prompt(conversation)
-            assert (
-                "Role did not alternate between turns (gpt and human)"
-                in self._caplog.records[0].message
-            )
-
-    def test_sharegpt_changes_roles(self):
-        conversation = {
-            "roles": ["USER", "CHARACTER"],
-            "conversations": [
-                {"from": "system", "value": "lorem"},
-                {"from": "gpt", "value": "ipsum"},
-                {"from": "human", "value": "dolor"},
-                {"from": "gpt", "value": "sit"},
-            ],
-        }
-        prompter = ShareGPTPrompterV2()
-        strat = ShareGPTPromptTokenizingStrategy(
-            prompter,
-            self.tokenizer,
-            False,
-            2048,
-        )
-        with self._caplog.at_level(logging.WARNING):
-            res = strat.tokenize_prompt(conversation)
-            assert "CHARACTER" in self.tokenizer.decode(res["input_ids"])
-
-    def test_sharegpt_assistant_label_ignore(self):
-        conversation = {
-            "roles": ["user", "assistant"],
-            "conversations": [
-                {"from": "system", "value": "lorem"},
-                {"from": "gpt", "value": "ipsum"},
-                {"from": "human", "value": "dolor"},
-                {"from": "gpt", "value": "sit"},
-            ],
-        }
-        prompter = ShareGPTPrompterV2()
-        strat = ShareGPTPromptTokenizingStrategy(
-            prompter,
-            self.tokenizer,
-            False,
-            2048,
-        )
-        with self._caplog.at_level(logging.WARNING):
-            res = strat.tokenize_prompt(conversation)
-            idx = res["input_ids"].index(20255)  # assistant token
-            assert res["labels"][idx] == -100
-
     def test_no_sys_prompt(self):
         """
         tests the interface between the user and assistant parts

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)