SFT Llama 3.1 8B - full training vs LoRA vs QLoRA

sft_llama.py

# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
# Everything below was run on 8 x H100 GPUs

# Full training with packing (learns chat template)
accelerate launch --config_file=examples/accelerate_configs/deepspeed_zero3.yaml scratch/sft_llama.py \
    --model_name_or_path meta-llama/Llama-3.1-8B \
    --dataset_name trl-lib/Capybara \
    --report_to wandb \
    --learning_rate 2.0e-5 \
    --per_device_train_batch_size 8 \
    --gradient_accumulation_steps 1 \
    --gradient_checkpointing \
    --output_dir Llama-3.1-8B-SFT-full-packing \
    --logging_steps 10 \
    --num_train_epochs 1 \
    --push_to_hub \
    --packing \
    --bf16

# LoRA with packing and all-linear modules and lm_head and embed_tokens (learns chat template)
accelerate launch --config_file=examples/accelerate_configs/multi_gpu.yaml scratch/sft_llama.py \
    --model_name_or_path meta-llama/Llama-3.1-8B \
    --dataset_name trl-lib/Capybara \
    --report_to wandb \
    --learning_rate 2.0e-4 \
    --per_device_train_batch_size 8 \
    --gradient_accumulation_steps 1 \
    --gradient_checkpointing \
    --ddp_find_unused_parameters False \
    --output_dir Llama-3.1-8B-SFT-LoRA-packing \
    --logging_steps 10 \
    --num_train_epochs 1 \
    --push_to_hub \
    --use_peft \
    --lora_r 16 \
    --lora_alpha 32 \
    --lora_target_modules all-linear \
    --lora_modules_to_save lm_head embed_tokens \
    --packing

# LoRA with packing and all-linear modules (doesn't learn chat template!)
accelerate launch --config_file=examples/accelerate_configs/multi_gpu.yaml scratch/sft_llama.py \
    --model_name_or_path meta-llama/Llama-3.1-8B \
    --dataset_name trl-lib/Capybara \
    --report_to wandb \
    --learning_rate 2.0e-4 \
    --per_device_train_batch_size 8 \
    --gradient_accumulation_steps 1 \
    --gradient_checkpointing \
    --ddp_find_unused_parameters False \
    --output_dir Llama-3.1-8B-SFT-LoRA-packing-no-saved-modules \
    --logging_steps 10 \
    --num_train_epochs 1 \
    --push_to_hub \
    --use_peft \
    --lora_r 16 \
    --lora_alpha 32 \
    --lora_target_modules all-linear \
    --packing

# QLoRA with packing (learns chat template)
accelerate launch --config_file=examples/accelerate_configs/multi_gpu.yaml scratch/sft_llama.py \
    --model_name_or_path meta-llama/Llama-3.1-8B \
    --dataset_name trl-lib/Capybara \
    --report_to wandb \
    --learning_rate 2.0e-4 \
    --per_device_train_batch_size 8 \
    --gradient_accumulation_steps 1 \
    --gradient_checkpointing \
    --ddp_find_unused_parameters False \
    --output_dir Llama-3.1-8B-SFT-QLoRA-packing \
    --logging_steps 10 \
    --num_train_epochs 1 \
    --push_to_hub \
    --use_peft \
    --lora_r 16 \
    --lora_alpha 32 \
    --lora_target_modules all-linear \
    --lora_modules_to_save lm_head embed_tokens \
    --packing \
    --load_in_4bit
"""

from datasets import load_dataset
from transformers import AutoTokenizer

from trl import (
    ModelConfig,
    SFTConfig,
    SFTScriptArguments,
    SFTTrainer,
    TrlParser,
    get_kbit_device_map,
    get_peft_config,
    get_quantization_config,
)


if __name__ == "__main__":
    parser = TrlParser((SFTScriptArguments, SFTConfig, ModelConfig))
    script_args, training_args, model_config = parser.parse_args_and_config()
    print(f"Script args: {script_args}\n")
    print(f"Training args: {training_args}\n")
    print(f"Model config: {model_config}\n")

    ################
    # Model init kwargs & Tokenizer
    ################
    quantization_config = get_quantization_config(model_config)
    model_kwargs = dict(
        revision=model_config.model_revision,
        trust_remote_code=model_config.trust_remote_code,
        attn_implementation=model_config.attn_implementation,
        torch_dtype=model_config.torch_dtype,
        use_cache=False if training_args.gradient_checkpointing else True,
        device_map=get_kbit_device_map() if quantization_config is not None else None,
        quantization_config=quantization_config,
    )
    training_args.model_init_kwargs = model_kwargs
    tokenizer = AutoTokenizer.from_pretrained(
        model_config.model_name_or_path, trust_remote_code=model_config.trust_remote_code, use_fast=True
    )
    tokenizer.pad_token = tokenizer.eos_token
    print(f"Using pad token: {tokenizer.pad_token}")
    # Set Llama chat template
    tokenizer.chat_template = "{{- bos_token }}\n{%- if custom_tools is defined %}\n    {%- set tools = custom_tools %}\n{%- endif %}\n{%- if not tools_in_user_message is defined %}\n    {%- set tools_in_user_message = true %}\n{%- endif %}\n{%- if not date_string is defined %}\n    {%- set date_string = \"26 Jul 2024\" %}\n{%- endif %}\n{%- if not tools is defined %}\n    {%- set tools = none %}\n{%- endif %}\n\n{#- This block extracts the system message, so we can slot it into the right place. #}\n{%- if messages[0]['role'] == 'system' %}\n    {%- set system_message = messages[0]['content']|trim %}\n    {%- set messages = messages[1:] %}\n{%- else %}\n    {%- set system_message = \"\" %}\n{%- endif %}\n\n{#- System message + builtin tools #}\n{{- \"<|start_header_id|>system<|end_header_id|>\\n\\n\" }}\n{%- if builtin_tools is defined or tools is not none %}\n    {{- \"Environment: ipython\\n\" }}\n{%- endif %}\n{%- if builtin_tools is defined %}\n    {{- \"Tools: \" + builtin_tools | reject('equalto', 'code_interpreter') | join(\", \") + \"\\n\\n\"}}\n{%- endif %}\n{{- \"Cutting Knowledge Date: December 2023\\n\" }}\n{{- \"Today Date: \" + date_string + \"\\n\\n\" }}\n{%- if tools is not none and not tools_in_user_message %}\n    {{- \"You have access to the following functions. To call a function, please respond with JSON for a function call.\" }}\n    {{- 'Respond in the format {\"name\": function name, \"parameters\": dictionary of argument name and its value}.' }}\n    {{- \"Do not use variables.\\n\\n\" }}\n    {%- for t in tools %}\n        {{- t | tojson(indent=4) }}\n        {{- \"\\n\\n\" }}\n    {%- endfor %}\n{%- endif %}\n{{- system_message }}\n{{- \"<|eot_id|>\" }}\n\n{#- Custom tools are passed in a user message with some extra guidance #}\n{%- if tools_in_user_message and not tools is none %}\n    {#- Extract the first user message so we can plug it in here #}\n    {%- if messages | length != 0 %}\n        {%- set first_user_message = messages[0]['content']|trim %}\n        {%- set messages = messages[1:] %}\n    {%- else %}\n        {{- raise_exception(\"Cannot put tools in the first user message when there's no first user message!\") }}\n{%- endif %}\n    {{- '<|start_header_id|>user<|end_header_id|>\\n\\n' -}}\n    {{- \"Given the following functions, please respond with a JSON for a function call \" }}\n    {{- \"with its proper arguments that best answers the given prompt.\\n\\n\" }}\n    {{- 'Respond in the format {\"name\": function name, \"parameters\": dictionary of argument name and its value}.' }}\n    {{- \"Do not use variables.\\n\\n\" }}\n    {%- for t in tools %}\n        {{- t | tojson(indent=4) }}\n        {{- \"\\n\\n\" }}\n    {%- endfor %}\n    {{- first_user_message + \"<|eot_id|>\"}}\n{%- endif %}\n\n{%- for message in messages %}\n    {%- if not (message.role == 'ipython' or message.role == 'tool' or 'tool_calls' in message) %}\n        {{- '<|start_header_id|>' + message['role'] + '<|end_header_id|>\\n\\n'+ message['content'] | trim + '<|eot_id|>' }}\n    {%- elif 'tool_calls' in message %}\n        {%- if not message.tool_calls|length == 1 %}\n            {{- raise_exception(\"This model only supports single tool-calls at once!\") }}\n        {%- endif %}\n        {%- set tool_call = message.tool_calls[0].function %}\n        {%- if builtin_tools is defined and tool_call.name in builtin_tools %}\n            {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' -}}\n            {{- \"<|python_tag|>\" + tool_call.name + \".call(\" }}\n            {%- for arg_name, arg_val in tool_call.arguments | items %}\n                {{- arg_name + '=\"' + arg_val + '\"' }}\n                {%- if not loop.last %}\n                    {{- \", \" }}\n                {%- endif %}\n                {%- endfor %}\n            {{- \")\" }}\n        {%- else  %}\n            {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' -}}\n            {{- '{\"name\": \"' + tool_call.name + '\", ' }}\n            {{- '\"parameters\": ' }}\n            {{- tool_call.arguments | tojson }}\n            {{- \"}\" }}\n        {%- endif %}\n        {%- if builtin_tools is defined %}\n            {#- This means we're in ipython mode #}\n            {{- \"<|eom_id|>\" }}\n        {%- else %}\n            {{- \"<|eot_id|>\" }}\n        {%- endif %}\n    {%- elif message.role == \"tool\" or message.role == \"ipython\" %}\n        {{- \"<|start_header_id|>ipython<|end_header_id|>\\n\\n\" }}\n        {%- if message.content is mapping or message.content is iterable %}\n            {{- message.content | tojson }}\n        {%- else %}\n            {{- message.content }}\n        {%- endif %}\n        {{- \"<|eot_id|>\" }}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|start_header_id|>assistant<|end_header_id|>\\n\\n' }}\n{%- endif %}\n"

    ################
    # Dataset
    ################
    dataset = load_dataset(script_args.dataset_name)


    ################
    # Training
    ################
    trainer = SFTTrainer(
        model=model_config.model_name_or_path,
        args=training_args,
        train_dataset=dataset[script_args.dataset_train_split],
        eval_dataset=dataset[script_args.dataset_test_split],
        tokenizer=tokenizer,
        peft_config=get_peft_config(model_config),
    )
    trainer.train()

    # Save and push to hub
    trainer.save_model(training_args.output_dir)
    if training_args.push_to_hub:
        trainer.push_to_hub(dataset_name=script_args.dataset_name)

Run inference tests with

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# model_name = "lewtun/Llama-3.1-8B-SFT-full-packing"
# model_name = "lewtun/Llama-3.1-8B-SFT-LoRA-packing"
# model_name = "lewtun/Llama-3.1-8B-SFT-LoRA-packing-no-saved-modules"
model_name = "lewtun/Llama-3.1-8B-SFT-QLoRA-packing"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto")

print(f"EOS token: {tokenizer.eos_token}")
print(f"PAD token: {tokenizer.pad_token}")
print(f"Generation config: {model.generation_config}")

inputs = tokenizer.apply_chat_template([{"role": "user", "content": "What is the capital of France?"}], return_tensors="pt", add_generation_prompt=True).to("cuda")

eot_token_id = tokenizer.encode("<|eot_id|>", add_special_tokens=False)
print(f"EOT token: {tokenizer.decode(eot_token_id)}")

outputs = model.generate(inputs, max_new_tokens=128, do_sample=False, eos_token_id=eot_token_id, pad_token_id=tokenizer.eos_token_id)

print(tokenizer.decode(outputs[0], skip_special_tokens=False))