FMS HF Tuning

• Installation
• Data format
• Supported Models
• Training
  • Single GPU
  • Multiple GPUs with FSDP
• Tuning Techniques
  • LoRA Tuning Example
  • Prompt Tuning
  • Fine Tuning
  • FMS Acceleration
• Inference
  • Running a single example
  • Running multiple examples
  • Inference Results Format
  • Changing the Base Model for Inference
• Validation
• Trainer Controller Framework
• More Examples

This repo provides basic tuning scripts with support for specific models. The repo relies on Hugging Face SFTTrainer and PyTorch FSDP. Our approach to tuning is:

1. Models are loaded from Hugging Face transformers or the foundation-model-stack -- models are either optimized to use Flash Attention v2 directly or through SDPA
2. Hugging Face SFTTrainer for the training loop
3. FSDP as the backend for training

Installation

pip install fms-hf-tuning

Note: After installing, if you wish to use FlashAttention, then you need to install these requirements:

pip install fms-hf-tuning[dev]
pip install fms-hf-tuning[flash-attn]

FlashAttention requires the CUDA Toolit to be pre-installed.

If you wish to use aim, then you need to install it:

pip install fms-hf-tuning[aim]

If you wish to use fms-acceleration, you need to install it.

pip install fms-hf-tuning[fms-accel]

fms-acceleration is a collection of plugins that packages that accelerate fine-tuning / training of large models, as part of the fms-hf-tuning suite. For more details on see this section below.

Data format

We support the following data formats:

1. JSON formats with a single sequence and a specified response_template to use for masking on completion.

1.1 Pre-process the JSON/JSONL dataset

Pre-process the JSON/JSONL dataset to contain a single sequence of each data instance containing input + Response. The trainer is configured to expect a response template as a string. For example, if one wants to prepare the alpaca format data to feed into this trainer, it is quite easy and can be done with the following code.

PROMPT_DICT = {
    "prompt_input": (
        "Below is an instruction that describes a task, paired with an input that provides further context. "
        "Write a response that appropriately completes the request.\n\n"
        "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:"
    ),
    "prompt_no_input": (
        "Below is an instruction that describes a task. "
        "Write a response that appropriately completes the request.\n\n"
        "### Instruction:\n{instruction}\n\n### Response:"
    ),
}

def format_alpaca_fn(example):
    prompt_input, prompt_no_input = PROMPT_DICT['prompt_input'], PROMPT_DICT['prompt_no_input']
    output = prompt_input.format_map(example) if example.get("input", "") != "" else prompt_no_input.format_map(example)
    output = f"{output} {example['output']}"
    return {"output": output}

ds = datasets.load_dataset('json', data_files='./stanford_alpaca/alpaca_data.json')

alpaca_ds = ds['train'].map(format_alpaca_fn, remove_columns=['instruction', 'input'])
alpaca_ds.to_json("sft_alpaca_data.json")

The response template corresponding to the above dataset and the Llama tokenizer is: \n### Response:".

The same way can be applied to any dataset, with more info can be found here.

Once the JSON is converted using the formatting function, pass the dataset_text_field containing the single sequence to the trainer.

1.2 Format JSON/JSONL on the fly

Pass a JSON/JSONL and a data_formatter_template to use the formatting function on the fly while tuning. The template should specify fields of JSON with {{field}}. While tuning, the data will be converted to a single sequence using the template.
JSON fields can contain alpha-numeric characters, spaces and the following special symbols - "." , "_", "-".

Example: Train.json [{ "input" : <text>, "output" : <text>, }, ... ]
data_formatter_template: ### Input: {{input}} \n\n##Label: {{output}}

Formatting will happen on the fly while tuning. The keys in template should match fields in JSON file. The response template corresponding to the above template will need to be supplied. in this case, response template = \n## Label:.

In conclusion, if using the reponse_template and single sequence, either the data_formatter_template argument or dataset_text_field needs to be supplied to the trainer.

2. JSONL with input and output fields (no response template)

Pass a JSONL containing fields "input" with source text and "output" with class labels. Pre-format the input as you see fit. The output field will simply be concatenated to the end of input to create single sequence, and input will be masked.

The "input" and "output" field names are mandatory and cannot be changed.

Example: Train.jsonl

{"input": "### Input: Colorado is a state in USA ### Output:", "output": "USA : Location"} 
{"input": "### Input: Arizona is also a state in USA ### Output:", "output": "USA : Location"}

Supported Models

Current supported and tested models are Llama2 (7 and 13B configurations have been tested) and GPTBigCode.

Training

Single GPU

Below example runs fine tuning with the given datasets and model:

1. Using pre-processed dataset for training.

# if you want to use one GPU on multi-gpu machine
export CUDA_VISIBLE_DEVICES=0

# MODEL_PATH=meta-llama/Llama-2-7b-hf # Huggingface model id or path to a checkpoint
# TRAIN_DATA_PATH=twitter_complaints.json # Path to the dataset
                  # contains data in single sequence {"output": "### Input: text \n\n### Response: text"}
# OUTPUT_PATH=out # Path to the output folder where the checkpoints are saved

python tuning/sft_trainer.py  \
--model_name_or_path $MODEL_PATH  \
--tokenizer_name_or_path $MODEL_PATH \ # This field is optional and if not specified, tokenizer from model_name_or_path will be used
--training_data_path $TRAIN_DATA_PATH  \
--output_dir $OUTPUT_PATH  \
--num_train_epochs 5  \
--per_device_train_batch_size 4  \
--gradient_accumulation_steps 4  \
--learning_rate 1e-5  \
--response_template "\n### Response:"  \
--dataset_text_field "output"

2. Using formatter with JSON/JSONL files

# if you want to use one GPU on multi-gpu machine
export CUDA_VISIBLE_DEVICES=0

# MODEL_PATH=meta-llama/Llama-2-7b-hf # Huggingface model id or path to a checkpoint
# TRAIN_DATA_PATH=twitter_complaints.json # Path to the dataset
                  # contains data in form of [{"input": text , "output": text}]
# OUTPUT_PATH=out # Path to the output folder where the checkpoints are saved

python tuning/sft_trainer.py  \
--model_name_or_path $MODEL_PATH  \
--tokenizer_name_or_path $MODEL_PATH \ # This field is optional and if not specified, tokenizer from model_name_or_path will be used
--training_data_path $TRAIN_DATA_PATH  \
--output_dir $OUTPUT_PATH  \
--num_train_epochs 5  \
--per_device_train_batch_size 4  \
--gradient_accumulation_steps 4  \
--learning_rate 1e-5  \
--response_template "\n## Label:"  \
--data_formatter_template: "### Input: {{input}} \n\n##Label: {{output}}"

Multiple GPUs with FSDP

The recommendation is to use huggingface accelerate to launch multi-gpu jobs, in particular when using FSDP:

• accelerate is written on top of torch.distributed.run.
• accelerate launch CLI highly similar to torchrun, spawns multiple jobs (one for each gpu).
• tightly integrated with huggingface Trainer.

accelerate launch CLI to be run with specific command line arguments, see example below. Default arguments handled by passing in a --config_file argument; see reference docs and fixtures/accelerate_fsdp_defaults.yaml for sample defaults.

Below example runs multi-GPU fine tuning on 8 GPUs with FSDP:

# Please set the environment variables:
# MASTER_PORT=1234 # The port at which the process with rank 0 listens to and should be set to an unused port
# MODEL_PATH=meta-llama/Llama-2-7b-hf # Huggingface model id or path to a checkpoint
# TRAIN_DATA_PATH=twitter_complaints.json # Path to the training dataset
# OUTPUT_PATH=out # Path to the output folder where the checkpoints are saved

accelerate launch \
--main_process_port $MASTER_PORT \
--config_file fixtures/accelerate_fsdp_defaults.yaml \
--num_processes=8 \ 
--main_process_port=$MASTER_PORT \
tuning/sft_trainer.py \
--model_name_or_path $MODEL_PATH \
--training_data_path $TRAIN_DATA_PATH \
--torch_dtype bfloat16 \
--output_dir $OUTPUT_PATH \
--num_train_epochs 5 \
--per_device_train_batch_size 4 \
--gradient_accumulation_steps 4 \
--learning_rate 1e-5 \
--response_template "\n### Response:" \
--dataset_text_field "output" \
--tokenizer_name_or_path $MODEL_PATH  # This field is optional and if not specified, tokenizer from model_name_or_path will be used

To summarize you can pick either python for single-GPU jobs or use accelerate launch for multi-GPU jobs. The following tuning techniques can be applied:

Tuning Techniques:

LoRA Tuning Example

Set peft_method to "lora". You can additionally pass any arguments from LoraConfig.

# Args you can pass
r: int =8 
lora_alpha: int = 32
target_modules: List[str] = field(
  default_factory=lambda: ["q_proj", "v_proj"],
      metadata={
            "help": "The names of the modules to apply LORA to. LORA selects modules which either \
            completely match or "
            'end with one of the strings. If the value is ["all-linear"], \
            then LORA selects all linear and Conv1D '
            "modules except for the output layer."
        },
    )
bias = "none"
lora_dropout: float = 0.05

Example command to run:

python tuning/sft_trainer.py \
--model_name_or_path $MODEL_PATH \
--tokenizer_name_or_path $MODEL_PATH \ # This field is optional and if not specified, tokenizer from model_name_or_path will be used
--training_data_path $TRAIN_DATA_PATH \
--output_dir $OUTPUT_PATH \
--num_train_epochs 40 \
--per_device_train_batch_size 4 \
---learning_rate 1e-4 \
--response_template "\n### Label:" \
--dataset_text_field "output" \
--peft_method "lora" \
--r 8 \
--lora_dropout 0.05 \
--lora_alpha 16 \
--target_modules c_attn c_proj

Equally you can pass in a JSON configuration for running tuning. See build doc for more details. The above can also be passed in as JSON:

{
    "model_name_or_path": $MODEL_PATH,
    "training_data_path": $TRAIN_DATA_PATH,
    "output_dir": $OUTPUT_PATH,
    "num_train_epochs": 40.0,
    "per_device_train_batch_size": 4,
    "learning_rate": 1e-4,
    "response_template": "\n### Label:",
    "dataset_text_field": "output",
    "peft_method": "lora",
    "r": 8,
    "lora_dropout": 0.05,
    "lora_alpha": 16,
    "target_modules": ["c_attn", "c_proj"]
}

Notice the target_modules that are set are the default values. target_modules are the names of the modules to apply the adapter to. If this is specified, only the modules with the specified names will be replaced. When passing a list of strings, either an exact match will be performed or it is checked if the name of the module ends with any of the passed strings. If this is specified as all-linear, then all linear/Conv1D modules are chosen, excluding the output layer. If this is not specified, modules will be chosen according to the model architecture. If the architecture is not known, an error will be raised — in this case, you should specify the target modules manually. See HuggingFace docs for more details.

For each model, the target_modules will depend on the type of model architecture. You can specify linear or attention layers to target_modules. To obtain list of target_modules for a model:

from transformers import AutoModelForCausalLM
# load the model
model = AutoModelForCausalLM.from_pretrained(MODEL_PATH)
# see the module list
model.modules

# to get just linear layers
import re
model_modules = str(model.modules)
pattern = r'\((\w+)\): Linear'
linear_layer_names = re.findall(pattern, model_modules)

names = []
for name in linear_layer_names:
    names.append(name)
target_modules = list(set(names))

For example for LLaMA model the modules look like:

<bound method Module.modules of LlamaForCausalLM(
  (model): LlamaModel(
    (embed_tokens): Embedding(32000, 4096, padding_idx=0)
    (layers): ModuleList(
      (0-31): 32 x LlamaDecoderLayer(
        (self_attn): LlamaSdpaAttention(
          (q_proj): Linear(in_features=4096, out_features=4096, bias=False)
          (k_proj): Linear(in_features=4096, out_features=4096, bias=False)
          (v_proj): Linear(in_features=4096, out_features=4096, bias=False)
          (o_proj): Linear(in_features=4096, out_features=4096, bias=False)
          (rotary_emb): LlamaRotaryEmbedding()
        )
        (mlp): LlamaMLP(
          (gate_proj): Linear(in_features=4096, out_features=11008, bias=False)
          (up_proj): Linear(in_features=4096, out_features=11008, bias=False)
          (down_proj): Linear(in_features=11008, out_features=4096, bias=False)
          (act_fn): SiLU()
        )
        (input_layernorm): LlamaRMSNorm()
        (post_attention_layernorm): LlamaRMSNorm()
      )
    )
    (norm): LlamaRMSNorm()
  )
  (lm_head): Linear(in_features=4096, out_features=32000, bias=False)
)>

You can specify attention or linear layers. With the CLI, you can specify layers with --target_modules "q_proj" "v_proj" "k_proj" "o_proj" or --target_modules "all-linear".

Recommended target modules per model architecture

As per LoRA paper, section 4.2 , by using the query and value projection matrices, we can achieve reasonable quality with efficient GPU utilization. Hence, while thinking about what LoRA adapters to specify, we recommend starting with query and value matrices. You could also refer to the defaults specified by PEFT library for popular model architectures in section TRANSFORMERS_MODELS_TO_LORA_TARGET_MODULES_MAPPING as a good starting point.

---

Prompt Tuning:

Specify peft_method to 'pt' . You can additionally pass any arguments from PromptTuningConfig.

# prompt_tuning_init can be either "TEXT" or "RANDOM"
prompt_tuning_init: str = "TEXT"
num_virtual_tokens: int = 8
# prompt_tuning_init_text only applicable if prompt_tuning_init= "TEXT"
prompt_tuning_init_text: str = "Classify if the tweet is a complaint or not:"
tokenizer_name_or_path: str = "llama-7b-hf"

Example command you can run:

python tuning/sft_trainer.py  \
--model_name_or_path $MODEL_PATH  \
--training_data_path $TRAIN_DATA_PATH  \
--output_dir $OUTPUT_PATH  \
--num_train_epochs 5  \
--per_device_train_batch_size 1  \
--learning_rate 0.03  \
--response_template "\n### Label:"  \
--dataset_text_field "output" \
--peft_method pt \
--tokenizer_name_or_path $MODEL_PATH \ # This field is optional and if not specified, tokenizer from model_name_or_path will be used
--prompt_tuning_init "RANDOM" \
--prompt_tuning_init_text "From the following input, identify target sentiment of following types: neutral, negative, positive"

Equally you can pass in a JSON configuration for running tuning. See build doc for more details. The above can also be passed in as JSON:

{
    "model_name_or_path": $MODEL_PATH,
    "training_data_path": $TRAIN_DATA_PATH,
    "output_dir": $OUTPUT_PATH,
    "num_train_epochs": 5.0,
    "per_device_train_batch_size": 1,
    "learning_rate": 0.03,
    "response_template": "\n### Label:",
    "dataset_text_field": "output",
    "peft_method": "pt",
    "tokenizer_name_or_path": $MODEL_PATH,
    "prompt_tuning_init": "RANDOM",
    "prompt_tuning_init_text": "From the following input, identify target sentiment of following types: neutral, negative, positive"
}

Fine Tuning:

Set peft_method to 'None' or do not provide peft_method flag.

Full fine tuning needs more compute resources, so it is advised to use the MultiGPU method. Example command:

accelerate launch \
--num_processes=4
--config_file fixtures/accelerate_fsdp_defaults.yaml \
tuning/sft_trainer.py  \
--model_name_or_path $MODEL_PATH  \
--tokenizer_name_or_path $MODEL_PATH \ # This field is optional and if not specified, tokenizer from model_name_or_path will be used
--training_data_path $TRAIN_DATA_PATH  \
--output_dir $OUTPUT_PATH  \
--num_train_epochs 5  \
--per_device_train_batch_size 4  \
--learning_rate 1e-5  \
--response_template "\n### Label:"  \
--dataset_text_field "output" \
--peft_method "None"

Equally you can pass in a JSON configuration for running tuning. See build doc for more details. The above can also be passed in as JSON:

{
    "model_name_or_path": $MODEL_PATH,
    "training_data_path": $TRAIN_DATA_PATH,
    "output_dir": $OUTPUT_PATH,
    "num_train_epochs": 5.0,
    "per_device_train_batch_size": 4,
    "learning_rate": 1e-5,
    "response_template": "\n### Label:",
    "dataset_text_field": "output",
    "peft_method": "None"
}

FMS Acceleration

fms-acceleration is fuss-free approach to access a curated collection of acceleration plugins that acclerate your tuning/sft-trainer.py experience. Accelerations that apply to a variety of use-cases, e.g., PeFT / full-finetuning, are being planned for. As such, the accelerations are grouped into plugins; only install the plugins needed for the acceleration of interest. The plugins are housed in the seperate repository found here.

To access fms-acceleration features the [fms-accel] dependency must first be installed:

$ pip install fms-hf-tuning[fms-accel]

Furthermore, the required fms-acceleration plugin must be installed. This is done via the command line utility fms_acceleration.cli. To show available plugins:

$ python -m fms_acceleration.cli plugins

as well as to install the fms_acceleration_peft:

$ python -m fms_acceleration.cli install fms_acceleration_peft

If you do not know what plugin to install (or forget), the framework will remind

An acceleration feature is requested by specifying the '--auto_gptq' argument, but the this requires acceleration packages to be installed. Please do:
- python -m fms_acceleration.cli install fms_acceleration_peft

The list of configurations for various fms_acceleration plugins:

• quantized_lora_config: For quantized 4bit LoRA training
  • --auto_gptq: 4bit GPTQ-LoRA with AutoGPTQ
  • --bnb_qlora: 4bit QLoRA with bitsandbytes
• fused_ops_and_kernels (experimental):
  • --fused_lora: fused lora for more efficient LoRA training.
  • --fast_kernels: fast cross-entropy, rope, rms loss kernels.

Notes:

• quantized_lora_config requires that it be used along with LoRA tuning technique. See LoRA tuning section on the LoRA parameters to pass.
• When setting --auto_gptq triton_v2 plus note to also pass --torch_dtype float16 and --fp16, or an exception will be raised. This is because these kernels only support this dtype.
• Currently, the fused_ops_and_kernels is to be used used together QLoRA or GPTQ-LORA via the quantized_lora_config. In the future it may be made more flexible such that fast_kernels can even be used with full-finetuning.
• When using fused_ops_and_kernels together with quantized_lora_config, make sure to appropriately set --fused_lora auto_gptq True or bitsandbytes True; the True sets fast_lora==True.
• Currently fused_ops_and_kernels only supports activating fast_loss,fast_rsm_layernorm,fast_rope_embeddings all to True, so pass --fast_kernels True True True.

Activate TRANSFORMERS_VERBOSITY=info to see the huggingface trainer printouts and verify that AccelerationFramework is activated!

# this printout will be seen in huggingface trainer logs if acceleration is activated
***** FMS AccelerationFramework *****
Active Plugin: AutoGPTQAccelerationPlugin. Python package: fms_acceleration_peft. Version: 0.0.1.
***** Running training *****
Num examples = 1,549
Num Epochs = 1
Instantaneous batch size per device = 4
Total train batch size (w. parallel, distributed & accumulation) = 4
Gradient Accumulation steps = 1
Total optimization steps = 200
Number of trainable parameters = 13,631,488

The fms_acceleration.cli can do more to search for all available configs, plugins and arguments, see the advanced flow.

Inference

Currently, we do not offer inference support as part of the library, but we provide a standalone script for running inference on tuned models for testing purposes. For a full list of options run python scripts/run_inference.py --help. Note that no data formatting / templating is applied at inference time.

Running a single example

If you want to run a single example through a model, you can pass it with the --text flag.

python scripts/run_inference.py \
--model my_checkpoint \
--text "This is a text the model will run inference on" \
--max_new_tokens 50 \
--out_file result.json

Running multiple examples

To run multiple examples, pass a path to a file containing each source text as its own line. Example:

Contents of source_texts.txt

This is the first text to be processed.
And this is the second text to be processed.

python scripts/run_inference.py \
--model my_checkpoint \
--text_file source_texts.txt \
--max_new_tokens 50 \
--out_file result.json

Inference Results Format

After running the inference script, the specified --out_file will be a JSON file, where each text has the original input string and the predicted output string, as follows. Note that due to the implementation of .generate() in Transformers, in general, the input string will be contained in the output string as well.

[
    {
        "input": "{{Your input string goes here}}",
        "output": "{{Generate result of processing your input string goes here}}"
    },
    ...
]

Changing the Base Model for Inference

If you tuned a model using a local base model, then a machine-specific path will be saved into your checkpoint by Peft, specifically the adapter_config.json. This can be problematic if you are running inference on a different machine than you used for tuning.

As a workaround, the CLI for inference provides an arg for --base_model_name_or_path, where a new base model may be passed to run inference with. This will patch the base_model_name_or_path in your checkpoint's adapter_config.json while loading the model, and restore it to its original value after completion. Alternatively, if you like, you can change the config's value yourself.

NOTE: This can also be an issue for tokenizers (with the tokenizer_name_or_path config entry). We currently do not allow tokenizer patching since the tokenizer can also be explicitly configured within the base model and checkpoint model, but may choose to expose an override for the tokenizer_name_or_path in the future.

Validation

We can use lm-evaluation-harness from EleutherAI for evaluating the generated model. For example, for the Llama-13B model, using the above command and the model at the end of Epoch 5, we evaluated MMLU score to be 53.9 compared to base model to be 52.8.

How to run the validation:

pip install -U transformers
pip install -U datasets
git clone https://github.com/EleutherAI/lm-evaluation-harness
cd lm-evaluation-harness
pip install -e .
python main.py \ 
--model hf-causal \
--model_args pretrained=$MODEL_PATH \ 
--output_path $OUTPUT_PATH/results.json \ 
--tasks boolq,piqa,hellaswag,winogrande,arc_easy,arc_challenge,hendrycksTest-*

The above runs several tasks with hendrycksTest-* being MMLU.

Trainer Controller Framework

Trainer controller is a framework for controlling the trainer loop using user-defined rules and metrics.

This framework helps users define rules to capture scenarios like criteria for stopping an ongoing training (E.g validation loss reaching a certain target, validation loss increasing with epoch, training loss values for last 100 steps increasing etc).

For details about how you can use set a custom stopping criteria and perform custom operations, see examples/trainercontroller_configs/Readme.md

More Examples

Prompt Tuning on Twitter Complaints

A good simple example can be found here which launches a Kubernetes-native PyTorchJob using the Kubeflow Training Operator with Kueue for the queue management of tuning jobs.