support QLoRA

README.md

@@ -9,12 +9,14 @@
 
 ## Changelog
 
+[23/06/03] Now we support quantized training and inference (aka QLoRA). Try `--quantization_bit 4/8` argument to work with quantized model. (experimental feature)
+
 [23/05/31] Now we support training the BLOOM & BLOOMZ models in this repo. Try `--model_name_or_path bigscience/bloomz-7b1-mt` argument to use the BLOOMZ model.
 
 ## Supported Models
 
-- [LLaMA](https://github.com/facebookresearch/llama) (7B, 13B, 33B, 65B)
-- [BLOOM](https://huggingface.co/bigscience/bloom) & [BLOOMZ](https://huggingface.co/bigscience/bloomz) (560M, 1.1B, 1.7B, 3B, 7.1B, 176B)
+- [LLaMA](https://github.com/facebookresearch/llama) (7B/13B/33B/65B)
+- [BLOOM](https://huggingface.co/bigscience/bloom) & [BLOOMZ](https://huggingface.co/bigscience/bloomz) (560M/1.1B/1.7B/3B/7.1B/176B)
 
 ## Supported Training Approaches
 
@@ -22,12 +24,15 @@
   - Full-parameter training
   - Partial-parameter training
   - [LoRA](https://arxiv.org/abs/2106.09685)
+  - [QLoRA](https://arxiv.org/abs/2305.14314)
 - [Supervised fine-tuning](https://arxiv.org/abs/2109.01652)
   - Full-parameter training
   - Partial-parameter training
   - [LoRA](https://arxiv.org/abs/2106.09685)
+  - [QLoRA](https://arxiv.org/abs/2305.14314)
 - [RLHF](https://arxiv.org/abs/2203.02155)
   - [LoRA](https://arxiv.org/abs/2106.09685)
+  - [QLoRA](https://arxiv.org/abs/2305.14314)
 
 ## Provided Datasets
 
@@ -209,6 +214,8 @@ CUDA_VISIBLE_DEVICES=0 python src/train_sft.py \
     --predict_with_generate
 ```
 
+We recommend using `--per_device_eval_batch_size=1` and `--max_target_length 128` in INT8 evaluation.
+
 ### CLI Demo
 
 ```bash

src/utils/common.py

@@ -11,7 +11,8 @@ from transformers import (
     AutoModelForCausalLM,
     AutoTokenizer,
     HfArgumentParser,
-    Seq2SeqTrainingArguments
+    Seq2SeqTrainingArguments,
+    BitsAndBytesConfig
 )
 from transformers.utils import check_min_version
 from transformers.utils.versions import require_version
@@ -167,12 +168,27 @@ def load_pretrained(
 
     # Quantization configurations (using bitsandbytes library).
     if model_args.quantization_bit is not None:
-        assert model_args.quantization_bit == 8, "We only accept 8-bit quantization."
-        require_version("bitsandbytes>=0.39.0", "To fix: pip install bitsandbytes>=0.39.0")
-        #require_version("transformers>=4.30.0.dev0", "To fix: pip install git+https://github.com/huggingface/transformers.git")
-        #require_version("peft>=0.4.0.dev0", "To fix: pip install git+https://github.com/huggingface/peft.git")
-        #require_version("accelerate>=0.20.0.dev0", "To fix: pip install git+https://github.com/huggingface/accelerate.git")
-        config_kwargs["load_in_8bit"] = True
+        if model_args.quantization_bit == 8:
+            require_version("bitsandbytes>=0.37.0", "To fix: pip install bitsandbytes>=0.37.0")
+            config_kwargs["load_in_8bit"] = True
+            config_kwargs["quantization_config"] = BitsAndBytesConfig(
+                load_in_8bit=True,
+                llm_int8_threshold=6.0
+            )
+        elif model_args.quantization_bit == 4:
+            require_version("bitsandbytes>=0.39.0", "To fix: pip install bitsandbytes>=0.39.0")
+            require_version("transformers>=4.30.0.dev0", "To fix: pip install git+https://github.com/huggingface/transformers.git")
+            require_version("peft>=0.4.0.dev0", "To fix: pip install git+https://github.com/huggingface/peft.git")
+            require_version("accelerate>=0.20.0.dev0", "To fix: pip install git+https://github.com/huggingface/accelerate.git")
+            config_kwargs["load_in_4bit"] = True
+            config_kwargs["quantization_config"] = BitsAndBytesConfig(
+                load_in_4bit=True,
+                bnb_4bit_compute_dtype=finetuning_args.compute_dtype,
+                bnb_4bit_use_double_quant=model_args.double_quantization,
+                bnb_4bit_quant_type=model_args.quantization_type
+            )
+        else:
+            raise NotImplementedError
         is_mergeable = False
         logger.info("Quantizing model to {} bit.".format(model_args.quantization_bit))
 
@@ -183,7 +199,7 @@ def load_pretrained(
     model = AutoModelForCausalLM.from_pretrained(
         model_args.model_name_or_path,
         config=config,
-        torch_dtype=torch.float16, # the model weights are float16 type
+        torch_dtype=torch.bfloat16 if finetuning_args.compute_dtype == torch.bfloat16 else torch.float16,
         low_cpu_mem_usage=True,
         **config_kwargs
     )
@@ -237,13 +253,13 @@ def prepare_args(
 
     # Check arguments (do not check finetuning_args since it may be loaded from checkpoints)
     if stage != "sft" and training_args.predict_with_generate:
-        raise ValueError("`predict_with_generate` cannot be set as True in PT, RM and PPO stages.")
+        raise ValueError("`predict_with_generate` cannot be set as True at PT, RM and PPO stages.")
 
     if training_args.do_train and training_args.predict_with_generate:
         raise ValueError("`predict_with_generate` cannot be set as True while training.")
 
     if training_args.do_predict and (not training_args.predict_with_generate):
-        raise ValueError("Please enable `predict_with_generate` for saving model predictions.")
+        raise ValueError("Please enable `predict_with_generate` to save model predictions.")
 
     if model_args.quantization_bit is not None and (not training_args.do_train):
         logger.warning("Evaluating model in 4/8-bit mode may cause lower scores.")
@@ -257,6 +273,14 @@ def prepare_args(
 
     training_args.optim = "adamw_torch" if training_args.optim == "adamw_hf" else training_args.optim # suppress warning
 
+    if model_args.quantization_bit is not None:
+        if training_args.fp16:
+            finetuning_args.compute_dtype = torch.float16
+        elif training_args.bf16:
+            finetuning_args.compute_dtype = torch.bfloat16
+        else:
+            finetuning_args.compute_dtype = torch.float32
+
     # Log on each process the small summary:
     logger.info(
         f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}\n"

src/utils/config.py

@@ -1,5 +1,6 @@
 import os
 import json
+import torch
 from typing import List, Literal, Optional
 from dataclasses import asdict, dataclass, field
 
@@ -207,6 +208,10 @@ class FinetuningArguments:
                   LLaMA choices: [\"q_proj\", \"k_proj\", \"v_proj\", \"o_proj\", \"up_proj\", \"down_proj\"], \
                   BLOOM choices: [\"query_key_value\", \"dense\", \"dense_\"]"}
     )
+    compute_dtype: Optional[torch.dtype] = field(
+        default=None,
+        metadata={"help": "Used in quantization configs. Do not specify this argument manually."}
+    )
 
     def __post_init__(self):
         if isinstance(self.lora_target, str):