README.md

FMS HF Tuning

• Installation
• Data format support
• Supported Models
• Training
  • Single GPU
  • Multiple GPUs with FSDP
  • Tips on Parameters to Set
• Tuning Techniques
  • LoRA Tuning Example
  • GPTQ-LoRA with AutoGPTQ 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
• Experiment Tracking
• 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

Basic Installation

pip install fms-hf-tuning

Using FlashAttention

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.

Using FMS-Acceleration

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 see this section below.

Using Experiment Trackers

To use experiment tracking with popular tools like Aim, note that some trackers are considered optional dependencies and can be installed with the following command:

pip install fms-hf-tuning[aim]

For more details on how to enable and use the trackers, Please see, the experiment tracking section below.

Data Support

Users can pass training data in a single file using the --training_data_path argument along with other arguments required for various use cases (see details below) and the file can be in any of the supported formats. Alternatively, you can use our powerful data preprocessing backend to preprocess datasets on the fly.

Below, we mention the list of supported data usecases via --training_data_path argument. For details of our advanced data preprocessing see more details in Advanced Data Preprocessing.

Supported Data Formats

We support the following data formats via --training_data_path argument

Data Format	Tested Support
JSON	✅
JSONL	✅
PARQUET	✅
ARROW	✅

Use cases supported with training_data_path argument

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

1.1 Pre-process the dataset

Pre-process the 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 data is converted using the formatting function, pass the dataset_text_field containing the single sequence to the trainer.

1.2 Format the dataset on the fly

Pass a dataset and a data_formatter_template to use the formatting function on the fly while tuning. The template should specify fields of the dataset with {{field}}. While tuning, the data will be converted to a single sequence using the template. Data 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 the dataset 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. Dataset with input and output fields (no response template)

Pass a supported dataset 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: For a JSON dataset like, 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"}

3. Chat Style Single/Multi turn datasets

Pass a dataset containing single/multi turn chat dataset. Your dataset could follow this format:

$ head -n 1 train.jsonl
{"messages": [{"content": "You are an AI language model developed by IBM Research. You are a cautious assistant. You carefully follow instructions. You are helpful and harmless and you follow ethical guidelines and promote positive behavior.", "role": "system"}, {"content": "Look up a word that rhymes with exist", "role": "user"}, {"content": "I found a word that rhymes with \"exist\":\n1\\. Mist", "role": "assistant"}], "group": "lab_extension", "dataset": "base/full-extension", "metadata": "{\"num_turns\": 1}"}

This format supports both single and multi-turn chat scenarios.

The chat template used to render the dataset will default to tokenizer.chat_template from the model's tokenizer configuration. This can be overridden using the --chat_template <chat-template-string> argument. For example, models like ibm-granite/granite-3.0-8b-instruct, which include a chat template in their tokenizer_config.json, do not require users to provide a chat template to process the data.

Users do need to pass --response_template and --instruction_template which are pieces of text representing start of assistant and human response inside the formatted chat template. For the granite model above for example, the values shall be.

--instruction_template "<|start_of_role|>user<|end_of_role|>"
--response_template "<|start_of_role|>assistant<|end_of_role|>"

The code internally uses DataCollatorForCompletionOnlyLM to perform masking of text ensuring model learns only on the assistant responses for both single and multi turn chat.

3. Pre tokenized datasets.

Users can also pass a pretokenized dataset (containing input_ids and labels columns) as --training_data_path argument e.g.

python tuning/sft_trainer.py ... --training_data_path twitter_complaints_tokenized_with_maykeye_tinyllama_v0.arrow

4. Advanced data preprocessing.

For advanced data preprocessing support including mixing and custom preprocessing of datasets please see this document.

Supported Models

• For each tuning technique, we run testing on a single large model of each architecture type and claim support for the smaller models. For example, with QLoRA technique, we tested on granite-34b GPTBigCode and claim support for granite-20b-multilingual.

• LoRA Layers supported : All the linear layers of a model + output lm_head layer. Users can specify layers as a list or use all-linear as a shortcut. Layers are specific to a model architecture and can be specified as noted here

• Legend:

  ✅ Ready and available

  ✔️ Ready and available - compatible architecture (*see first bullet point above)

  🚫 Not supported

  ? May be supported, but not tested

Model Name & Size	Model Architecture	Full Finetuning	Low Rank Adaptation (i.e. LoRA)	qLoRA(quantized LoRA)
Granite PowerLM 3B	GraniteForCausalLM	✅*	✅*	✅*
Granite 3.0 2B	GraniteForCausalLM	✔️*	✔️*	✔️*
Granite 3.0 8B	GraniteForCausalLM	✅*	✅*	✔️
GraniteMoE 1B	GraniteMoeForCausalLM	✅	✅**	?
GraniteMoE 3B	GraniteMoeForCausalLM	✅	✅**	?
Granite 3B	LlamawithCausalLM	✅	✔️	✔️
Granite 8B	LlamawithCausalLM	✅	✅	✅
Granite 13B	GPTBigCodeForCausalLM	✅	✅	✔️
Granite 20B	GPTBigCodeForCausalLM	✅	✔️	✔️
Granite 34B	GPTBigCodeForCausalLM	🚫	✅	✅
Llama3.1-8B	LLaMA 3.1	✅***	✔️	✔️
Llama3.1-70B(same architecture as llama3)	LLaMA 3.1	🚫 - same as Llama3-70B	✔️	✔️
Llama3.1-405B	LLaMA 3.1	🚫	🚫	✅
Llama3-8B	LLaMA 3	✅	✅	✔️
Llama3-70B	LLaMA 3	🚫	✅	✅
aLLaM-13b	LlamaForCausalLM	✅	✅	✅
Mixtral 8x7B	Mixtral	✅	✅	✅
Mistral-7b	Mistral	✅	✅	✅
Mistral large	Mistral	🚫	🚫	🚫

(*) - Supported with fms-hf-tuning v2.0.1 or later

(**) - Supported for q,k,v,o layers . all-linear target modules does not infer on vLLM yet.

(***) - Supported from platform up to 8k context length - same architecture as llama3-8b

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:

Tips on Parameters to Set

Saving checkpoints while training

By default, save_strategy is set to "epoch" in the TrainingArguments. This means that checkpoints will be saved on each epoch. This can also be set to "steps" to save on every "save_steps" or "no" to not save any checkpoints.

Checkpoints are saved to the given output_dir, which is a required field. If save_strategy="no", the output_dir will only contain the training logs with loss details.

A useful flag to set to limit the number of checkpoints saved is save_total_limit. Older checkpoints are deleted from the output_dir to limit the number of checkpoints, for example, if save_total_limit=1, this will only save the last checkpoint. However, while tuning, two checkpoints will exist in output_dir for a short time as the new checkpoint is created and then the older one will be deleted. If the user sets a validation dataset and load_best_model_at_end, then the best checkpoint will be saved.

Saving model after training

save_model_dir can optionally be set to save the tuned model using SFTTrainer.save_model(). This can be used in tandem with save_strategy="no" to only save the designated checkpoint and not any intermediate checkpoints, which can help to save space.

save_model_dir can be set to a different directory than output_dir. If set to the same directory, the designated checkpoint, training logs, and any intermediate checkpoints will all be saved to the same directory as seen below.

Ways you can use `save_model_dir` and more tips:

For example, if save_model_dir is set to a sub-directory of output_dirand save_total_limit=1 with LoRA tuning, the directory would look like:

$ ls /tmp/output_dir/
checkpoint-35  save_model_dir  training_logs.jsonl

$ ls /tmp/output_dir/save_model_dir/
README.md	     adapter_model.safetensors	special_tokens_map.json  tokenizer.model	training_args.bin
adapter_config.json  added_tokens.json		tokenizer.json		 tokenizer_config.json

Here is an fine tuning example of how the directory would look if output_dir is set to the same value as save_model_dir and save_total_limit=2. Note the checkpoint directories as well as the training_logs.jsonl:

$ ls /tmp/same_dir

added_tokens.json	model-00001-of-00006.safetensors  model-00006-of-00006.safetensors  tokenizer_config.json
checkpoint-16		model-00002-of-00006.safetensors  model.safetensors.index.json	    training_args.bin
checkpoint-20		model-00003-of-00006.safetensors  special_tokens_map.json	    training_logs.jsonl
config.json		model-00004-of-00006.safetensors  tokenizer.json
generation_config.json	model-00005-of-00006.safetensors  tokenizer.model

Optimizing writing checkpoints

Writing models to Cloud Object Storage (COS) is an expensive operation. Saving model checkpoints to a local directory causes much faster training times than writing to COS. You can use output_dir and save_model_dir to control which type of storage you write your checkpoints and final model to.

You can set output_dir to a local directory and set save_model_dir to COS to save time on write operations while ensuring checkpoints are saved.

In order to achieve the fastest train time, set save_strategy="no", as saving no checkpoints except for the final model will remove intermediate write operations all together.

Resuming tuning from checkpoints

If the output directory already contains checkpoints, tuning will automatically resume from the latest checkpoint in the directory specified by the output_dir flag. To start tuning from scratch and ignore existing checkpoints, set the resume_from_checkpoint flag to False.

You can also use the resume_from_checkpoint flag to resume tuning from a specific checkpoint by providing the full path to the desired checkpoint as a string. This flag is passed as an argument to the trainer.train() function of the SFTTrainer.

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=None,
  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 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 specified as lm_head which is an output layer, the lm_head layer will be chosen. See the Note of this section on recommended target modules by model architecture.
• 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.

How to get list of LoRA target_modules of a model

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.

How to specify lm_head as a target module

Since lm_head is an output layer, it will not be included as a target module if you specify all-linear. You can, however, specify to apply the LoRA adapter to the lm_head layer by explicitly naming it in the target_modules arg.

NOTE: Specifying ["lm_head", "all-linear"] will not tune the lm_head layer, but will run the equivalent of ["all-linear"]. To include lm_head, you must explicitly specify all of the layers to tune on. Using the example of the Llama model above, you would need to list "q_proj" "v_proj" "k_proj" "o_proj" "lm_head" to tune the all linear layers including lm_head. These 5 layers will be produced in the LoRA adapter.

Example 1:

{
    "target_modules": ["lm_head"] // this produces lm_head layer only
}

Example 2:

{
    "target_modules": ["lm_head", "c_proj", "c_attn", "c_fc"] // this produces lm_head, c_proj, c_attn and c_fc layers 
}

Example 3:

{
    "target_modules": ["lm_head", "all-linear"] // this produces the equivalent of all-linear only, no lm_head
}

Post-processing needed for inference on VLLM

In order to run inference of LoRA adapters on vLLM, any new token embeddings added while tuning needs to be moved out of 'adapters.safetensors' to a new file 'new_embeddings.safetensors'. The 'adapters.safetensors' should only have LoRA weights and should not have modified embedding vectors. This is a requirement to support vLLM's paradigm that one base model can serve multiple adapters. New token embedding vectors are appended to the embedding matrix read from the base model by vLLM.

To do this postprocessing, the tuning script sft_trainer.py will generate a file 'added_tokens_info.json' with model artifacts. After tuning, you can run script 'post_process_adapters_vLLM.py' :

# model_path: Path to saved model artifacts which has file 'added_tokens_info.json'
# output_model_path: Optional. If you want to store modified \
#    artifacts in a different directory rather than modify in-place.
python scripts/post_process_adapters_vLLM.py \
--model_path "/testing/tuning/output/post-process-LoRA-saved" \
--output_model_path "/testing/tuning/output/post-process-LoRA-modified"

Alternatively, if using SDK :

# function in tuning/utils/merge_model_utils.py
post_process_vLLM_adapters_new_tokens(
    path_to_checkpoint="/testing/tuning/output/post-process-LoRA-saved",
    modified_checkpoint_path=None,
    num_added_tokens=1,
)
# where num_added_tokens is returned by sft_trainer.train()

---

GPTQ-LoRA with AutoGPTQ Tuning Example

This method is similar to LoRA Tuning, but the base model is a quantized model. We currently only support GPTQ-LoRA model that has been quantized with 4-bit AutoGPTQ technique. Bits-and-Bytes (BNB) quantized LoRA is not yet enabled. The qLoRA tuning technique is enabled via the fms-acceleration package. You can see details on a sample configuration of Accelerated GPTQ-LoRA here.

To use GPTQ-LoRA technique, you can set the quantized_lora_config defined here. See the Notes section of FMS Acceleration doc below for usage. The only kernel we are supporting currently is triton_v2.

In addition, LoRA tuning technique is required to be used, set peft_method to "lora" and pass any arguments from LoraConfig.

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 \
--auto_gptq triton_v2 \ # setting quantized_lora_config 
--torch_dtype float16 \ # need this for triton_v2
--fp16 \ # need this for triton_v2

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"],
    "auto_gptq": ["triton_v2"], // setting quantized_lora_config
    "torch_dtype": "float16", // need this for triton_v2
    "fp16": true // need this for triton_v2
}

Similarly to LoRA, the target_modules are the names of the modules to apply the adapter to. See the LoRA section on target_modules for more info.

Note that with LoRA tuning technique, setting all-linear on target_modules returns linear modules. And with qLoRA tuning technique, all-linear returns all quant linear modules, excluding lm_head.

---

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:
  • --fused_lora: fused lora for more efficient LoRA training.
  • --fast_kernels: fast cross-entropy, rope, rms loss kernels.
• attention_and_distributed_packing:
  • --padding_free: technique to process multiple examples in single batch without adding padding tokens that waste compute.
  • --multipack: technique for multi-gpu training to balance out number of tokens processed in each device, to minimize waiting time.

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.
• 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.
• fused_ops_and_kernels works for full-finetuning, LoRA, QLoRA and GPTQ-LORA,
  • pass --fast_kernels True True True for full finetuning/LoRA
  • pass --fast_kernels True True True --auto_gptq triton_v2 --fused_lora auto_gptq True for GPTQ-LoRA
  • pass --fast_kernels True True True --bitsandbytes nf4 --fused_lora bitsandbytes True for QLoRA
  • Note the list of supported models here.
• Notes on Padding Free
  • works for both single and multi-gpu.
  • works on both pretokenized and untokenized datasets
  • verified against the version found in HF main, merged in via PR huggingface/transformers#31629.
• Notes on Multipack
  • works only for multi-gpu.
  • currently only includes the version of multipack optimized for linear attention implementations like flash-attn.

Note: To pass the above flags via a JSON config, each of the flags expects the value to be a mixed type list, so the values must be a list. For example:

{
  "fast_kernels": [true, true, true],
  "padding_free": ["huggingface"],
  "multipack": [16],
  "auto_gptq": ["triton_v2"]
}

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

Experiment Tracking

Experiment tracking in fms-hf-tuning allows users to track their experiments with known trackers like Aimstack, MLflow Tracking or custom trackers built into the code like FileLoggingTracker

The code supports currently two trackers out of the box,

• FileLoggingTracker : A built in tracker which supports logging training loss to a file.
• Aimstack : A popular opensource tracker which can be used to track any metrics or metadata from the experiments.
• MLflow Tracking : Another popular opensource tracker which stores metrics, metadata or even artifacts from experiments.

Further details on enabling and using the trackers mentioned above can be found here.

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.