Update sample_finetune.py

sample_finetune.py

@@ -1,28 +1,68 @@
-import torch
+import sys
+import logging
+
+import datasets
 from datasets import load_dataset
+from peft import LoraConfig
+import torch
+import transformers
 from trl import SFTTrainer
-from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
+from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, BitsAndBytesConfig
 
 """
 A simple example on using SFTTrainer and Accelerate to finetune Phi-3 models. For
-a more advanced example, please follow HF alignment-handbook/scripts/run_sft.py
-
-1. Install accelerate: 
+a more advanced example, please follow HF alignment-handbook/scripts/run_sft.py.
+This example has utilized DeepSpeed ZeRO3 offload to reduce the memory usage. The
+script can be run on V100 or later generation GPUs. Here are some suggestions on 
+futher reducing memory consumption:
+    - reduce batch size
+    - decrease lora dimension
+    - restrict lora target modules
+Please follow these steps to run the script:
+1. Install dependencies: 
     conda install -c conda-forge accelerate
-2. Setup accelerate config:
+    pip3 install -i https://pypi.org/simple/ bitsandbytes
+    pip3 install peft
+    pip3 install deepspeed
+2. Setup accelerate and deepspeed config based on the machine used:
     accelerate config
-to simply use all the GPUs available:
-    python -c "from accelerate.utils import write_basic_config; write_basic_config(mixed_precision='bf16')"
-check accelerate config:
+Here is a sample config for deepspeed zero3:
+    compute_environment: LOCAL_MACHINE
+    debug: false
+    deepspeed_config:
+    gradient_accumulation_steps: 1
+    offload_optimizer_device: none
+    offload_param_device: none
+    zero3_init_flag: true
+    zero3_save_16bit_model: true
+    zero_stage: 3
+    distributed_type: DEEPSPEED
+    downcast_bf16: 'no'
+    enable_cpu_affinity: false
+    machine_rank: 0
+    main_training_function: main
+    mixed_precision: bf16
+    num_machines: 1
+    num_processes: 4
+    rdzv_backend: static
+    same_network: true
+    tpu_env: []
+    tpu_use_cluster: false
+    tpu_use_sudo: false
+    use_cpu: false
+3. check accelerate config:
     accelerate env
-3. Run the code:
+4. Run the code:
     accelerate launch sample_finetune.py
 """
 
+logger = logging.getLogger(__name__)
+
+
 ###################
 # Hyper-parameters
 ###################
-args = {
+training_config = {
     "bf16": True,
     "do_eval": False,
     "learning_rate": 5.0e-06,
@@ -35,7 +75,7 @@ args = {
     "output_dir": "./checkpoint_dir",
     "overwrite_output_dir": True,
     "per_device_eval_batch_size": 4,
-    "per_device_train_batch_size": 8,
+    "per_device_train_batch_size": 4,
     "remove_unused_columns": True,
     "save_steps": 100,
     "save_total_limit": 1,
@@ -45,8 +85,43 @@ args = {
     "gradient_accumulation_steps": 1,
     "warmup_ratio": 0.2,
     }
-    
-training_args = TrainingArguments(**args)
+
+peft_config = {
+    "r": 16,
+    "lora_alpha": 32,
+    "lora_dropout": 0.05,
+    "bias": "none",
+    "task_type": "CAUSAL_LM",
+    "target_modules": "all-linear",
+    "modules_to_save": None,
+}
+train_conf = TrainingArguments(**training_config)
+peft_conf = LoraConfig(**peft_config)
+
+
+###############
+# Setup logging
+###############
+logging.basicConfig(
+    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+    handlers=[logging.StreamHandler(sys.stdout)],
+)
+log_level = train_conf.get_process_log_level()
+logger.setLevel(log_level)
+datasets.utils.logging.set_verbosity(log_level)
+transformers.utils.logging.set_verbosity(log_level)
+transformers.utils.logging.enable_default_handler()
+transformers.utils.logging.enable_explicit_format()
+
+# Log on each process a small summary
+logger.warning(
+    f"Process rank: {train_conf.local_rank}, device: {train_conf.device}, n_gpu: {train_conf.n_gpu}"
+    + f" distributed training: {bool(train_conf.local_rank != -1)}, 16-bits training: {train_conf.fp16}"
+)
+logger.info(f"Training/evaluation parameters {train_conf}")
+logger.info(f"PEFT parameters {peft_conf}")
+
 
 ################
 # Modle Loading
@@ -58,14 +133,16 @@ model_kwargs = dict(
     trust_remote_code=True,
     attn_implementation="flash_attention_2",  # loading the model with flash-attenstion support
     torch_dtype=torch.bfloat16,
-    device_map="cuda",
+    device_map=None
 )
 model = AutoModelForCausalLM.from_pretrained(checkpoint_path, **model_kwargs)
 tokenizer = AutoTokenizer.from_pretrained(checkpoint_path)
+tokenizer.model_max_length = 2048
 tokenizer.pad_token = tokenizer.unk_token  # use unk rather than eos token to prevent endless generation
 tokenizer.pad_token_id = tokenizer.convert_tokens_to_ids(tokenizer.pad_token)
 tokenizer.padding_side = 'right'
 
+
 ##################
 # Data Processing
 ##################
@@ -82,26 +159,36 @@ def apply_chat_template(
     return example
 
 raw_dataset = load_dataset("HuggingFaceH4/ultrachat_200k")
-column_names = list(raw_dataset["train_sft"].features)
+train_dataset = raw_dataset["train_sft"]
+test_dataset = raw_dataset["test_sft"]
+column_names = list(train_dataset.features)
 
-processed_dataset = raw_dataset.map(
+processed_train_dataset = train_dataset.map(
     apply_chat_template,
     fn_kwargs={"tokenizer": tokenizer},
-    num_proc=12,
+    num_proc=10,
     remove_columns=column_names,
-    desc="Applying chat template",
+    desc="Applying chat template to train_sft",
 )
-train_dataset = processed_dataset["train_sft"]
-eval_dataset = processed_dataset["test_sft"]
+
+processed_test_dataset = test_dataset.map(
+    apply_chat_template,
+    fn_kwargs={"tokenizer": tokenizer},
+    num_proc=10,
+    remove_columns=column_names,
+    desc="Applying chat template to test_sft",
+)
+
 
 ###########
 # Training
 ###########
 trainer = SFTTrainer(
     model=model,
-    args=training_args,
-    train_dataset=train_dataset,
-    eval_dataset=eval_dataset,
+    args=train_conf,
+    peft_config=peft_conf,
+    train_dataset=processed_train_dataset,
+    eval_dataset=processed_test_dataset,
     max_seq_length=2048,
     dataset_text_field="text",
     tokenizer=tokenizer,
@@ -113,16 +200,18 @@ trainer.log_metrics("train", metrics)
 trainer.save_metrics("train", metrics)
 trainer.save_state()
 
+
 #############
 # Evaluation
 #############
 tokenizer.padding_side = 'left'
 metrics = trainer.evaluate()
-metrics["eval_samples"] = len(eval_dataset)
+metrics["eval_samples"] = len(processed_test_dataset)
 trainer.log_metrics("eval", metrics)
 trainer.save_metrics("eval", metrics)
 
-############
-# Save model
-############
-trainer.save_model(training_args.output_dir)
+
+# ############
+# # Save model
+# ############
+trainer.save_model(train_conf.output_dir)