{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "## Multiple GPUS" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "\n", "import torch\n", "import torch.nn.functional as F\n", "from torch.utils.data import Dataset, DataLoader\n", "from datautils import MyTrainDataset\n", "import torch.multiprocessing as mp\n", "from torch.utils.data.distributed import DistributedSampler\n", "from torch.nn.parallel import DistributedDataParallel as DDP\n", "from torch.distributed import init_process_group, destroy_process_group\n", "import os\n", "import argparse\n", "from datasets import load_dataset\n", "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig\n", "from peft import LoraConfig\n", "from trl import SFTTrainer\n", "from transformers import TrainingArguments\n", "\n", "def ddp_setup(rank, world_size):\n", " os.environ[\"MASTER_ADDR\"] = \"localhost\"\n", " os.environ[\"MASTER_PORT\"] = \"12355\"\n", " init_process_group(backend=\"nccl\", rank=rank, world_size=world_size)\n", " torch.cuda.set_device(rank)\n", "\n", "class Trainer:\n", " def __init__(self, model, train_data, optimizer, gpu_id, save_every):\n", " self.gpu_id = gpu_id\n", " self.model = model.to(gpu_id)\n", " self.train_data = train_data\n", " self.optimizer = optimizer\n", " self.save_every = save_every\n", " f.model = DDP(model, device_ids=[gpu_id])\n", "\n", " def _run_batch(self, source, targets):\n", " self.optimizer.zero_grad()\n", " output = self.model(source)\n", " loss = F.cross_entropy(output, targets)\n", " loss.backward()\n", " self.optimizer.step()\n", "\n", " def _run_epoch(self, epoch):\n", " b_sz = len(next(iter(self.train_data))[0])\n", " print(f\"[GPU{self.gpu_id}] Epoch {epoch} | Batchsize: {b_sz} | Steps: {len(self.train_data)}\")\n", " self.train_data.sampler.set_epoch(epoch)\n", " for source, targets in self.train_data:\n", " source = source.to(self.gpu_id)\n", " targets = targets.to(self.gpu_id)\n", " self._run_batch(source, targets)\n", "\n", " def _save_checkpoint(self, epoch):\n", " ckp = self.model.module.state_dict()\n", " PATH = \"checkpoint.pt\"\n", " torch.save(ckp, PATH)\n", " print(f\"Epoch {epoch} | Training checkpoint saved at {PATH}\")\n", "\n", " def train(self, max_epochs):\n", " for epoch in range(max_epochs):\n", " self._run_epoch(epoch)\n", " if self.gpu_id == 0 and epoch % self.save_every == 0:\n", " self._save_checkpoint(epoch)\n", "\n", "def load_train_objs():\n", " dataset_name = \"ruslanmv/ai-medical-dataset\"\n", " dataset = load_dataset(dataset_name, split=\"train\")\n", " dataset = dataset.select(range(100))\n", "\n", " model_name = \"meta-llama/Meta-Llama-3-8B-Instruct\"\n", " tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)\n", " tokenizer.pad_token = tokenizer.eos_token\n", "\n", " bnb_config = BitsAndBytesConfig(\n", " load_in_4bit=True,\n", " bnb_4bit_quant_type=\"nf4\",\n", " bnb_4bit_compute_dtype=torch.float16,\n", " )\n", " model = AutoModelForCausalLM.from_pretrained(\n", " model_name,\n", " quantization_config=bnb_config,\n", " trust_remote_code=True,\n", " use_cache=False,\n", " device_map=\"auto\",\n", " )\n", "\n", " lora_alpha = 16\n", " lora_dropout = 0.1\n", " lora_r = 32\n", " peft_config = LoraConfig(\n", " lora_alpha=lora_alpha,\n", " lora_dropout=lora_dropout,\n", " r=lora_r,\n", " bias=\"none\",\n", " task_type=\"CAUSAL_LM\",\n", " target_modules=[\"k_proj\", \"q_proj\", \"v_proj\", \"up_proj\", \"down_proj\", \"gate_proj\"],\n", " modules_to_save=[\"embed_tokens\", \"input_layernorm\", \"post_attention_layernorm\", \"norm\"],\n", " )\n", "\n", " max_seq_length = 512\n", " output_dir = \"./results\"\n", " per_device_train_batch_size = 2\n", " gradient_accumulation_steps = 2\n", " optim = \"adamw_torch\"\n", " save_steps = 10\n", " logging_steps = 1\n", " learning_rate = 2e-4\n", " max_grad_norm = 0.3\n", " max_steps = 1\n", " warmup_ratio = 0.1\n", " lr_scheduler_type = \"cosine\"\n", "\n", " training_arguments = TrainingArguments(\n", " output_dir=output_dir,\n", " per_device_train_batch_size=per_device_train_batch_size,\n", " gradient_accumulation_steps=gradient_accumulation_steps,\n", " optim=optim,\n", " save_steps=save_steps,\n", " logging_steps=logging_steps,\n", " learning_rate=learning_rate,\n", " fp16=True,\n", " max_grad_norm=max_grad_norm,\n", " max_steps=max_steps,\n", " warmup_ratio=warmup_ratio,\n", " group_by_length=True,\n", " lr_scheduler_type=lr_scheduler_type,\n", " gradient_checkpointing=True,\n", " )\n", "\n", " return dataset, model, peft_config, tokenizer, training_arguments\n", "\n", "def prepare_dataloader(dataset, batch_size):\n", " return DataLoader(\n", " dataset,\n", " batch_size=batch_size,\n", " pin_memory=True,\n", " shuffle=False,\n", " sampler=DistributedSampler(dataset),\n", " )\n", "\n", "def main(rank, world_size, save_every, total_epochs, batch_size):\n", " ddp_setup(rank, world_size)\n", " dataset, model, peft_config, tokenizer, training_arguments = load_train_objs()\n", " train_data = prepare_dataloader(dataset, batch_size)\n", " trainer = SFTTrainer(\n", " model=model,\n", " train_dataset=dataset,\n", " peft_config=peft_config,\n", " dataset_text_field=\"context\",\n", " max_seq_length=max_seq_length,\n", " tokenizer=tokenizer,\n", " args=training_arguments,\n", " )\n", " trainer = Trainer(model, train_data, optimizer=trainer.optimizer, gpu_id=rank, save_every=save_every)\n", " trainer.train(total_epochs)\n", " destroy_process_group()\n", "\n", "TOTAL_EPOCHS = 10\n", "SAVE_EVERY = 2\n", "BATCH_SIZE = 32\n", "\n", "if __name__ == \"__main__\":\n", " world_size = torch.cuda.device_count()\n", " mp.set_start_method(\"spawn\", force=True) # Add this line\n", " mp.spawn(main, args=(world_size, SAVE_EVERY, TOTAL_EPOCHS, BATCH_SIZE), nprocs=world_size)\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import torch\n", "import torch.nn.functional as F\n", "from torch.utils.data import Dataset, DataLoader\n", "import torch.multiprocessing as mp\n", "from torch.utils.data.distributed import DistributedSampler\n", "from torch.nn.parallel import DistributedDataParallel as DDP\n", "from torch.distributed import init_process_group, destroy_process_group\n", "import os\n", "import argparse\n", "from datasets import load_dataset\n", "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoTokenizer\n", "from peft import LoraConfig\n", "from trl import SFTTrainer\n", "from transformers import TrainingArguments\n", "def ddp_setup(rank, world_size):\n", " \"\"\"\n", " Args:\n", " rank: Unique identifier of each process\n", " world_size: Total number of processes\n", " \"\"\"\n", " os.environ[\"MASTER_ADDR\"] = \"localhost\"\n", " os.environ[\"MASTER_PORT\"] = \"12355\"\n", " init_process_group(backend=\"nccl\", rank=rank, world_size=world_size)\n", " torch.cuda.set_device(rank)\n", "\n", "class Trainer:\n", " def __init__(self, model, train_data, optimizer, gpu_id, save_every):\n", " self.gpu_id = gpu_id\n", " self.model = model.to(gpu_id)\n", " self.train_data = train_data\n", " self.optimizer = optimizer\n", " self.save_every = save_every\n", " self.model = DDP(model, device_ids=[gpu_id])\n", "\n", " def _run_batch(self, source, targets):\n", " self.optimizer.zero_grad()\n", " output = self.model(source)\n", " loss = F.cross_entropy(output, targets)\n", " loss.backward()\n", " self.optimizer.step()\n", "\n", " def _run_epoch(self, epoch):\n", " b_sz = len(next(iter(self.train_data))[0])\n", " print(f\"[GPU{self.gpu_id}] Epoch {epoch} | Batchsize: {b_sz} | Steps: {len(self.train_data)}\")\n", " self.train_data.sampler.set_epoch(epoch)\n", " for source, targets in self.train_data:\n", " source = source.to(self.gpu_id)\n", " targets = targets.to(self.gpu_id)\n", " self._run_batch(source, targets)\n", "\n", " def _save_checkpoint(self, epoch):\n", " ckp = self.model.module.statt()\n", " PATH = \"checkpoint.pt\"\n", " torch.save(ckp, PATH)\n", " print(f\"Epoch {epoch} | Training checkpoint saved at {PATH}\")\n", "\n", " def train(self, max_epochs):\n", " for epoch in range(max_epochs):\n", " self._run_epoch(epoch)\n", " if self.gpu_id == 0 and epoch % self.save_every == 0:\n", " self._save_checkpoint(epoch)\n", "\n", "def load_train_objs():\n", " dataset_name = \"ruslanmv/ai-medical-dataset\"\n", " dataset = load_dataset(dataset_name, split=\"train\")\n", " dataset = dataset.select(range(100))\n", "\n", " model_name = \"meta-llama/Meta-Llama-3-8B-Instruct\"\n", " tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)\n", " tokenizer.pad_token = tokenizer.eos_token\n", "\n", " bnb_config = BitsAndBytesConfig(\n", " load_in_4bit=True,\n", " bnb_4bit_quant_type=\"nf4\",\n", " bnb_4bit_compute_dtype=torch.float16,\n", " )\n", " model = AutoModelForCausalLM.from_pretrained(\n", " model_name,\n", " quantization_config=bnb_config,\n", " trust_remote_code=True,\n", " use_cache=False,\n", " device_map=\"auto\",\n", " )\n", "\n", " lora_alpha = 16\n", " lora_dropout = 0.1\n", " lora_r = 32\n", " peft_config = LoraConfig(\n", " lora_alpha=lora_alpha,\n", " lora_dropout=lora_dropout,\n", " r=lora_r,\n", " bias=\"none\",\n", " task_type=\"CAUSAL_LM\",\n", " target_modules=[\"k_proj\", \"q_proj\", \"v_proj\", \"up_proj\", \"down_proj\", \"gate_proj\"],\n", " modules_to_save=[\"embed_tokens\", \"input_layernorm\", \"post_attention_layernorm\", \"norm\"],\n", " )\n", "\n", " max_seq_length = 512\n", " output_dir = \"./results\"\n", " per_device_train_batch_size = 2\n", " gradient_accumulation_steps = 2\n", " optim = \"adamw_torch\"\n", " save_steps = 10\n", " logging_steps = 1\n", " learning_rate = 2e-4\n", " max_grad_norm = 0.3\n", " max_steps = 1\n", " warmup_ratio = 0.1\n", " lr_scheduler_type = \"cosine\"\n", "\n", " training_arguments = TrainingArguments(\n", " output_dir=output_dir,\n", " per_device_train_batch_size=per_device_train_batch_size,\n", " gradient_accumulation_steps=gradient_accumulation_steps,\n", " optim=optim,\n", " save_steps=save_steps,\n", " logging_steps=logging_steps,\n", " learning_rate=learning_rate,\n", " fp16=True,\n", " max_grad_norm=max_grad_norm,\n", " max_steps=max_steps,\n", " warmup_ratio=warmup_ratio,\n", " group_by_length=True,\n", " lr_scheduler_type=lr_scheduler_type,\n", " gradient_checkpointing=True,\n", " )\n", "\n", " return dataset, model, peft_config, tokenizer, training_arguments\n", "\n", "def prepare_dataloader(dataset, batch_size):\n", " return DataLoader(\n", " dataset,\n", " batch_size=batch_size,\n", " pin_memory=True,\n", " shuffle=False,\n", " sampler=DistributedSampler(dataset),\n", " )" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import torch\n", "import torch.multiprocessing as mp\n", "\n", "def main(rank, world_size):\n", " # Define the parameters as constants\n", " TOTAL_EPOCHS = 10\n", " SAVE_EVERY = 2\n", " BATCH_SIZE = 32\n", " torch.cuda.init()\n", " ddp_setup(rank, world_size) \n", " dataset, model, peft_config, tokenizer, training_arguments = load_train_objs()\n", " train_data = prepare_dataloader(dataset, BATCH_SIZE) # Corrected batch_size variable\n", " trainer = SFTTrainer(\n", " model=model,\n", " train_dataset=dataset,\n", " peft_config=peft_config,\n", " dataset_text_field=\"context\",\n", " max_seq_length=max_seq_length,\n", " tokenizer=tokenizer,\n", " args=training_arguments,\n", " )\n", " trainer = Trainer(model, train_data, optimizer=trainer.optimizer, gpu_id=rank, save_every=SAVE_EVERY)\n", " trainer.train(TOTAL_EPOCHS)\n", " destroy_process_group()\n", "\n", "if __name__ == \"__main__\":\n", " mp.set_start_method('spawn') # Set start method to 'spawn'\n", " world_size = torch.cuda.device_count()\n", "\n", " # Workaround for Jupyter Notebook and interactive environments\n", " processes = []\n", " for rank in range(world_size):\n", " p = mp.Process(target=main, args=(rank, world_size))\n", " p.start()\n", " processes.append(p)\n", "\n", " for p in processes:\n", " p.join()" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import torch\n", "from transformers import AutoModelForCausalLM, AutoTokenizer\n", "from datasets import load_dataset\n", "from trl import SFTTrainer\n", "from transformers import TrainingArguments\n", "import os\n", "import socket\n", "\n", "# Distributed training setup (assuming all GPUs are available on a single machine)\n", "def init_distributed(rank, world_size):\n", " \"\"\"Initializes distributed training using `nccl` backend.\"\"\"\n", " if rank == 0:\n", " os.environ[\"MASTER_ADDR\"] = socket.gethostname() # Set MASTER_ADDR using rank 0's hostname\n", " else:\n", " # Wait a bit to ensure MASTER_ADDR is set before other ranks try to use it\n", " import time\n", " time.sleep(5)\n", " os.environ[\"MASTER_PORT\"] = \"12345\" # Set MASTER_PORT environment variable\n", " os.environ[\"RANK\"] = str(rank) # Set RANK environment variable\n", " os.environ[\"WORLD_SIZE\"] = str(world_size) # Set WORLD_SIZE environment variable\n", " torch.distributed.init_process_group(backend='nccl', init_method='env://')\n", "\n", "# Cleanup after training\n", "def cleanup_distributed():\n", " if torch.distributed.is_initialized():\n", " torch.distributed.destroy_process_group()\n", "\n", "# Model and tokenizer selection\n", "model_name = \"facebook/bart-base\" # Replace with your desired model\n", "tokenizer = AutoTokenizer.from_pretrained(model_name)\n", "model = AutoModelForCausalLM.from_pretrained(model_name)\n", "\n", "# Dataset loading (replace with your dataset and field names)\n", "dataset = load_dataset(\"glue\", \"mnli\", split=\"train\")\n", "text_field = \"premise\" # Assuming premise is the field containing text for prediction\n", "\n", "# Training arguments (adjust hyperparameters as needed)\n", "training_args = TrainingArguments(\n", " output_dir=\"./results\",\n", " per_device_train_batch_size=2, # Adjust based on GPU memory (might need to adjust)\n", " save_steps=500,\n", " save_total_limit=2,\n", " num_train_epochs=3, # Adjust training time as needed\n", ")\n", "\n", "world_size = torch.cuda.device_count()\n", "if world_size > 1:\n", " # Initialize distributed training\n", " init_distributed(rank=0, world_size=world_size) # Rank is assumed to be 0 here\n", "\n", " # Wrap model in DDP for distributed training\n", " model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[torch.cuda.current_device()])\n", "\n", " # Create SFTTrainer with distributed settings\n", " trainer = SFTTrainer(\n", " model=model,\n", " args=training_args, # Pass training_args as 'args' instead of 'training_args'\n", " train_dataset=dataset,\n", " dataset_text_field=text_field,\n", " compute_metrics=None, # You can define your custom metrics here\n", " )\n", " print(\"Trainer For distributed training loaded\")\n", "else:\n", " # For single-GPU training\n", " trainer = SFTTrainer(\n", " model=model,\n", " args=training_args, # Pass training_args as 'args' instead of 'training_args'\n", " train_dataset=dataset,\n", " dataset_text_field=text_field,\n", " compute_metrics=None, # You can define your custom metrics here\n", " )\n", " print(\"Trainer For single-GPU loaded\")\n", "\n", "# Start training\n", "trainer.train()\n", "\n", "# Cleanup after training\n", "cleanup_distributed()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import os\n", "import torch\n", "import torch.multiprocessing as mp\n", "from datasets import load_dataset\n", "import torch\n", "from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoTokenizer\n", "from peft import LoraConfig\n", "from trl import SFTTrainer\n", "from transformers import TrainingArguments\n", "from torch.nn.parallel import DistributedDataParallel as DDP\n", "\n", "\n", "# Distributed training setup\n", "def init_distributed():\n", " os.environ[\"MASTER_ADDR\"] = \"localhost\"\n", " os.environ[\"MASTER_PORT\"] = \"12345\"\n", " torch.distributed.init_process_group(backend='nccl', world_size=torch.cuda.device_count(), rank=rank)\n", "\n", "def cleanup_distributed():\n", " torch.distributed.destroy_process_group()\n", "\n", "def main_worker(rank, world_size):\n", " init_distributed()\n", "\n", " # Your model training and fine-tuning code goes here\n", " # Load the dataset\n", " dataset_name = \"ruslanmv/ai-medical-dataset\"\n", " dataset = load_dataset(dataset_name, split=\"train\")\n", " # Select the first 1M rows of the dataset\n", " dataset = dataset.select(range(100))\n", "\n", " # Load the model + tokenizer\n", " model_name = \"meta-llama/Meta-Llama-3-8B-Instruct\"\n", " tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)\n", " tokenizer.pad_token = tokenizer.eos_token\n", " bnb_config = BitsAndBytesConfig(\n", " load_in_4bit=True,\n", " bnb_4bit_quant_type=\"nf4\",\n", " bnb_4bit_compute_dtype=torch.float16,\n", " )\n", " model = AutoModelForCausalLM.from_pretrained(\n", " model_name,\n", " quantization_config=bnb_config,\n", " trust_remote_code=True,\n", " use_cache=False,\n", " )\n", "\n", " # Check for available GPUs\n", " device = torch.device(f\"cuda:{rank}\" if torch.cuda.is_available() else \"cpu\")\n", "\n", " # PEFT config\n", " lora_alpha = 1\n", " lora_dropout = 0.1\n", " lora_r = 32 # 64\n", " peft_config = LoraConfig(\n", " lora_alpha=lora_alpha,\n", " lora_dropout=lora_dropout,\n", " task_type=\"CAUSAL_LM\",\n", " target_modules=[\"k_proj\", \"q_proj\", \"v_proj\", \"up_proj\", \"down_proj\", \"gate_proj\"],\n", " modules_to_save=[\"embed_tokens\", \"input_layernorm\", \"post_attention_layernorm\", \"norm\"],\n", " )\n", "\n", " # Args\n", " max_seq_length = 512\n", " output_dir = \"./results\"\n", " per_device_train_batch_size = 2 # reduced batch size to avoid OOM\n", " gradient_accumulation_steps = 2\n", " optim = \"adamw_torch\"\n", " save_steps = 10\n", " logging_steps = 1\n", " learning_rate = 2e-4\n", " max_grad_norm = 0.3\n", " max_steps = 1 # 300 Approx the size of guanaco at bs 8, ga 2, 2 GPUs.\n", " warmup_ratio = 0.1\n", " lr_scheduler_type = \"cosine\"\n", " training_arguments = TrainingArguments(\n", " output_dir=output_dir,\n", " per_device_train_batch_size=per_device_train_batch_size,\n", " gradient_accumulation_steps=gradient_accumulation_steps,\n", " optim=optim,\n", " save_steps=save_steps,\n", " logging_steps=logging_steps,\n", " learning_rate=learning_rate,\n", " fp16=True,\n", " max_grad_norm=max_grad_norm,\n", " max_steps=max_steps,\n", " warmup_ratio=warmup_ratio,\n", " group_by_length=True,\n", " lr_scheduler_type=lr_scheduler_type,\n", " gradient_checkpointing=True, # gradient checkpointing\n", " #report_to=\"wandb\",\n", " )\n", "\n", " # Trainer\n", " trainer = SFTTrainer(\n", " model=model,\n", " train_dataset=dataset,\n", " peft_config=peft_config,\n", " dataset_text_field=\"context\",\n", " max_seq_length=max_seq_length,\n", " tokenizer=tokenizer,\n", " args=training_arguments,\n", " )\n", "\n", " # Train :)\n", " trainer.train()\n", " cleanup_distributed()\n", "\n", "\n", "if __name__ == \"__main__\":\n", " world_size = torch.cuda.device_count()\n", " mp.set_start_method('spawn') # Add this line to fix the error\n", " processes = []\n", " for rank in range(world_size):\n", " p = mp.Process(target=main_worker, args=(rank, world_size))\n", " p.start()\n", " processes.append(p)\n", " for p in processes:\n", " p.join()\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "def finetune():\n", " from datasets import load_dataset\n", " import torch\n", " from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, AutoTokenizer\n", " from peft import LoraConfig\n", " from trl import SFTTrainer\n", " from transformers import TrainingArguments\n", " from torch.nn.parallel import DistributedDataParallel as DDP\n", " # Load the dataset\n", " dataset_name = \"ruslanmv/ai-medical-dataset\"\n", " dataset = load_dataset(dataset_name, split=\"train\")\n", " # Select the first 1M rows of the dataset\n", " dataset = dataset.select(range(100))\n", " # Load the model + tokenizer\n", " model_name = \"meta-llama/Meta-Llama-3-8B-Instruct\"\n", " tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)\n", " tokenizer.pad_token = tokenizer.eos_token\n", " bnb_config = BitsAndBytesConfig(\n", " load_in_4bit=True,\n", " bnb_4bit_quant_type=\"nf4\",\n", " bnb_4bit_compute_dtype=torch.float16,\n", " )\n", " model = AutoModelForCausalLM.from_pretrained(\n", " model_name,\n", " quantization_config=bnb_config,\n", " trust_remote_code=True,\n", " use_cache=False,\n", " )\n", " # Check for available GPUs\n", " if torch.cuda.device_count() > 1:\n", " print(\"Multiple GPUs detected, enabling DataParallel...\")\n", " model = DDP(model) # Wrap the model with DDP\n", " else:\n", " print(\"Using single GPU...\")\n", " # PEFT config\n", " lora_alpha = 16\n", " lora_dropout = 0.1\n", " lora_r = 32 # 64\n", " peft_config = LoraConfig(\n", " lora_alpha=lora_alpha,\n", " lora_dropout=lora_dropout,\n", " r=lora_r,\n", " bias=\"none\",\n", " task_type=\"CAUSAL_LM\",\n", " target_modules=[\"k_proj\", \"q_proj\", \"v_proj\", \"up_proj\", \"down_proj\", \"gate_proj\"],\n", " modules_to_save=[\"embed_tokens\", \"input_layernorm\", \"post_attention_layernorm\", \"norm\"],\n", " )\n", " # Args\n", " max_seq_length = 512\n", " output_dir = \"./results\"\n", " per_device_train_batch_size = 2 # reduced batch size to avoid OOM\n", " gradient_accumulation_steps = 2\n", " optim = \"adamw_torch\"\n", " save_steps = 10\n", " logging_steps = 1\n", " learning_rate = 2e-4\n", " max_grad_norm = 0.3\n", " max_steps = 1 # 300 Approx the size of guanaco at bs 8, ga 2, 2 GPUs.\n", " warmup_ratio = 0.1\n", " lr_scheduler_type = \"cosine\"\n", "\n", " training_arguments = TrainingArguments(\n", " output_dir=output_dir,\n", " per_device_train_batch_size=per_device_train_batch_size,\n", " gradient_accumulation_steps=gradient_accumulation_steps,\n", " optim=optim,\n", " save_steps=save_steps,\n", " logging_steps=logging_steps,\n", " learning_rate=learning_rate,\n", " fp16=True,\n", " max_grad_norm=max_grad_norm,\n", " max_steps=max_steps,\n", " warmup_ratio=warmup_ratio,\n", " group_by_length=True,\n", " lr_scheduler_type=lr_scheduler_type,\n", " gradient_checkpointing=True, # gradient checkpointing\n", " #report_to=\"wandb\",\n", " )\n", " # Trainer\n", " trainer = SFTTrainer(\n", " model=model,\n", " train_dataset=dataset,\n", " peft_config=peft_config,\n", " dataset_text_field=\"context\",\n", " max_seq_length=max_seq_length,\n", " tokenizer=tokenizer,\n", " args=training_arguments,\n", " )\n", " # Train :)\n", " trainer.train()" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import os\n", "import torch\n", "import torch.multiprocessing as mp\n", "\n", "def init_distributed(rank, world_size, local_rank=0): # Add local_rank argument\n", " os.environ[\"MASTER_ADDR\"] = \"localhost\"\n", " os.environ[\"MASTER_PORT\"] = \"12345\" # Adjust port if needed\n", " if rank == 0:\n", " print(\"Initializing distributed process group...\")\n", " torch.distributed.init_process_group(backend='nccl', world_size=world_size, rank=rank)\n", " torch.cuda.set_device(local_rank) # Set unique GPU device for each process\n", "\n", "def cleanup_distributed():\n", " torch.distributed.destroy_process_group()\n", "\n", "def main_worker(rank, world_size):\n", " local_rank = rank % torch.cuda.device_count() # Assign unique local rank\n", " init_distributed(rank, world_size, local_rank)\n", " # Your model training and fine-tuning code goes here with model on local_rank GPU\n", " finetune() # Move model to assigned GPU\n", " cleanup_distributed()\n", "if __name__ == \"__main__\":\n", " world_size = torch.cuda.device_count()\n", "\n", " # Workaround for Jupyter Notebook and interactive environments\n", " processes = []\n", " for rank in range(world_size):\n", " p = mp.Process(target=main_worker, args=(rank, world_size))\n", " p.start()\n", " processes.append(p)\n", "\n", " for p in processes:\n", " p.join()\n" ] } ], "metadata": { "language_info": { "name": "python" } }, "nbformat": 4, "nbformat_minor": 2 }