Skip to content

Fix Mistral GPTQ and DPO #3

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
u66u wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Fix Mistral GPTQ and DPO #3

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
89047b5
use_reentrant=False in Mistral_GPTQ
u66u Jan 29, 2024
70ad823
put quantization_config_loading outside the function so its accessibl…
u66u Jan 29, 2024
6837d24
fix Mistral DPO
u66u Jan 29, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 12 additions & 10 deletions Mistral_7B_Instruct_GPTQ_finetune.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@
"metadata": {},
"outputs": [],
"source": [
"# !pip install --upgrade trl peft accelerate bitsandbytes datasets auto-gptq optimum -q"
"!pip install --upgrade trl peft accelerate bitsandbytes datasets auto-gptq optimum -q"
]
},
{
Expand Down Expand Up @@ -156,6 +156,15 @@
"metadata": {},
"outputs": [],
"source": [
"tokenizer = AutoTokenizer.from_pretrained(\n",
" pretrained_model_name_or_path,\n",
" padding_side=\"left\",\n",
" add_eos_token=True,\n",
" add_bos_token=True,\n",
")\n",
"\n",
"quantization_config_loading = GPTQConfig(bits=4, use_exllama=False, tokenizer=tokenizer)\n",
"\n",
"def build_qlora_model(\n",
" pretrained_model_name_or_path: str = \"TheBloke/Mistral-7B-Instruct-v0.1-GPTQ\",\n",
" gradient_checkpointing: bool = True,\n",
Expand All @@ -182,17 +191,9 @@
"\n",
" # In below as well, when using any GPTQ model\n",
" # comment-out the quantization_config param\n",
"\n",
" tokenizer = AutoTokenizer.from_pretrained(\n",
" pretrained_model_name_or_path,\n",
" padding_side=\"left\",\n",
" add_eos_token=True,\n",
" add_bos_token=True,\n",
" )\n",
" \n",
" tokenizer.pad_token = tokenizer.eos_token\n",
"\n",
" quantization_config_loading = GPTQConfig(bits=4, use_exllama=False, tokenizer=tokenizer)\n",
"\n",
" model = AutoModelForCausalLM.from_pretrained(\n",
" pretrained_model_name_or_path,\n",
" # quantization_config=bnb_config,\n",
Expand Down Expand Up @@ -672,6 +673,7 @@
" warmup_steps=5,\n",
" per_device_train_batch_size=1,\n",
" gradient_checkpointing=True,\n",
" gradient_checkpointing_kwargs={\"use_reentrant\": False},\n",
" gradient_accumulation_steps=4,\n",
" max_steps=1000,\n",
" learning_rate=2.5e-5,\n",
Expand Down
162 changes: 73 additions & 89 deletions Mistral_7b_FineTuning_with_DPO_Direct_Preference_Optimization.ipynb
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,13 +40,12 @@
"from dataclasses import dataclass, field\n",
"from typing import Any, Dict, List, NewType, Optional, Tuple\n",
"import transformers\n",
"from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, BitsAndBytesConfig\n",
"from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, BitsAndBytesConfig, pipeline\n",
"from datasets import load_dataset\n",
"from peft import LoraConfig, PeftModel, get_peft_model, prepare_model_for_kbit_training\n",
"from trl import DPOTrainer\n",
"import bitsandbytes as bnb\n",
"from google.colab import\n",
"\n",
"import google.colab\n",
"\n",
"model_name = \"teknium/OpenHermes-2.5-Mistral-7B\"\n",
"\n",
Expand Down Expand Up @@ -209,94 +208,79 @@
"metadata": {},
"outputs": [],
"source": [
"def train(model_name,\n",
" dataset,\n",
" tokenizer,\n",
" new_model,\n",
" #wandb_project: str = \"\",\n",
" #wandb_run_name: str = \"\",\n",
" #wandb_watch: str = \"\", # options: false | gradients | all\n",
" #wandb_log_model: str = \"\", # options: false | true\n",
" ):\n",
" peft_config = LoraConfig(\n",
" r=16,\n",
" lora_alpha=16,\n",
" lora_dropout=0.05,\n",
" bias=\"none\",\n",
" task_type=\"CAUSAL_LM\",\n",
" target_modules: List[str] =['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']\n",
" )\n",
" assert (\n",
" model_name\n",
" ), \"Please specify a --base_model, e.g. --base_model='huggyllama/llama-7b'\"\n",
"\n",
" # Check if parameter passed or if set within environ\n",
" '''\n",
" use_wandb = len(wandb_project) > 0 or (\n",
" \"WANDB_PROJECT\" in os.environ and len(os.environ[\"WANDB_PROJECT\"]) > 0\n",
" )\n",
" # Only overwrite environ if wandb param passed\n",
" if len(wandb_project) > 0:\n",
" os.environ[\"WANDB_PROJECT\"] = wandb_project\n",
" if len(wandb_watch) > 0:\n",
" os.environ[\"WANDB_WATCH\"] = wandb_watch\n",
" if len(wandb_log_model) > 0:\n",
" os.environ[\"WANDB_LOG_MODEL\"] = wandb_log_model\n",
" '''\n",
"\n",
" # Base Model\n",
" model = AutoModelForCausalLM.from_pretrained(\n",
" model_name,\n",
" torch_dtype=torch.float16,\n",
" load_in_4bit=True\n",
" )\n",
" model.config.use_cache = False\n",
"# Train the model\n",
"\n",
"peft_config = LoraConfig(\n",
" r=16,\n",
" lora_alpha=16,\n",
" lora_dropout=0.05,\n",
" bias=\"none\",\n",
" task_type=\"CAUSAL_LM\",\n",
" target_modules =['k_proj', 'gate_proj', 'v_proj', 'up_proj', 'q_proj', 'o_proj', 'down_proj']\n",
")\n",
"\n",
" # Reference model\n",
" ref_model = AutoModelForCausalLM.from_pretrained(\n",
" model_name,\n",
" torch_dtype=torch.float16,\n",
" load_in_4bit=True\n",
" )\n",
"assert (\n",
" model_name\n",
"), \"Please specify a model in the first cell\"\n",
"\n",
" # Training arguments\n",
" training_args = DPOConfig(\n",
" num_train_epochs=3,\n",
" per_device_train_batch_size=1,\n",
" gradient_accumulation_steps=4,\n",
" gradient_checkpointing=True,\n",
" learning_rate=5e-5,\n",
" lr_scheduler_type=\"linear\",\n",
" max_steps=200,\n",
" save_strategy=\"no\",\n",
" logging_steps=1,\n",
" output_dir=new_model,\n",
" optim=\"paged_adamw_32bit\",\n",
" warmup_steps=100,\n",
" fp16=True,\n",
" # report_to=\"wandb\",\n",
" )\n",
"# Check if parameter passed or if set within environ\n",
"'''\n",
"use_wandb = len(wandb_project) > 0 or (\n",
" \"WANDB_PROJECT\" in os.environ and len(os.environ[\"WANDB_PROJECT\"]) > 0\n",
")\n",
"# Only overwrite environ if wandb param passed\n",
"if len(wandb_project) > 0:\n",
" os.environ[\"WANDB_PROJECT\"] = wandb_project\n",
"if len(wandb_watch) > 0:\n",
" os.environ[\"WANDB_WATCH\"] = wandb_watch\n",
"if len(wandb_log_model) > 0:\n",
" os.environ[\"WANDB_LOG_MODEL\"] = wandb_log_model\n",
"'''\n",
"\n",
"# Base Model\n",
"model = AutoModelForCausalLM.from_pretrained(\n",
" model_name,\n",
" torch_dtype=torch.float16,\n",
" load_in_4bit=True\n",
")\n",
"model.config.use_cache = False\n",
"\n",
" dpo_trainer = DPOTrainer(\n",
" model,\n",
" ref_model,\n",
" args=training_args,\n",
" train_dataset=dataset,\n",
" tokenizer=tokenizer,\n",
" peft_config=peft_config,\n",
" beta=0.1,\n",
" max_prompt_length=1024,\n",
" max_length=1536,\n",
" )"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"dpo_trainer.train(model_name, dataset, tokenizer, new_model)"
"# Reference model\n",
"ref_model = AutoModelForCausalLM.from_pretrained(\n",
" model_name,\n",
" torch_dtype=torch.float16,\n",
" load_in_4bit=True\n",
")\n",
"\n",
"# Training arguments\n",
"training_args = DPOConfig(\n",
" num_train_epochs=3,\n",
" per_device_train_batch_size=1,\n",
" gradient_accumulation_steps=4,\n",
" gradient_checkpointing=True,\n",
" learning_rate=5e-5,\n",
" lr_scheduler_type=\"linear\",\n",
" max_steps=200,\n",
" save_strategy=\"no\",\n",
" logging_steps=1,\n",
" output_dir=new_model,\n",
" optim=\"paged_adamw_32bit\",\n",
" warmup_steps=100,\n",
" fp16=True,\n",
" # report_to=\"wandb\",\n",
")\n",
"\n",
"dpo_trainer = DPOTrainer(\n",
" model,\n",
" ref_model,\n",
" args=training_args,\n",
" train_dataset=dataset,\n",
" tokenizer=tokenizer,\n",
" peft_config=peft_config,\n",
" beta=0.1,\n",
" max_prompt_length=1024,\n",
" max_length=1536,\n",
")"
]
},
{
Expand Down Expand Up @@ -366,7 +350,7 @@
" prompt = tokenizer.apply_chat_template(message, add_generation_prompt=True, tokenize=False)\n",
"\n",
" chat_pipeline = pipeline(\n",
" \"text-generation\",\n",
" task=\"text-generation\",\n",
" model=new_model,\n",
" tokenizer=tokenizer\n",
" )\n",
Expand Down