Optimizing Memory for Large Language Model Inference and Fine-Tuning

Giant language fashions (LLMs) like GPT-4, Bloom, and LLaMA have achieved exceptional capabilities by scaling as much as billions of parameters. Nonetheless, deploying these huge fashions for inference or fine-tuning is difficult as a consequence of their immense reminiscence necessities. On this technical weblog, we are going to discover methods for estimating and optimizing reminiscence consumption throughout LLM inference and fine-tuning throughout numerous {hardware} setups.

Understanding Reminiscence Necessities

The reminiscence required to load an LLM is primarily decided by the variety of parameters and the numerical precision used to retailer the parameters. A easy rule of thumb is:

Related articles

Drasi by Microsoft: A New Approach to Tracking Rapid Data Changes

2024-11-21

Even Nvidia’s CEO is obsessed with Google’s NotebookLM AI tool

2024-11-21

• Loading a mannequin with X billion parameters requires roughly 4X GB of VRAM in 32-bit float precision
• Loading a mannequin with X billion parameters requires roughly 2X GB of VRAM in 16-bit bfloat16/float16 precision

For instance, loading the 175B parameter GPT-3 mannequin would require roughly 350GB of VRAM in bfloat16 precision. As of right this moment, the biggest commercially obtainable GPUs just like the NVIDIA A100 and H100 supply solely 80GB of VRAM, necessitating tensor parallelism and mannequin parallelism methods.

Throughout inference, the reminiscence footprint is dominated by the mannequin parameters and the non permanent activation tensors produced. A high-level estimate for the height reminiscence utilization throughout inference is the sum of the reminiscence required to load the mannequin parameters and the reminiscence for activations.

Quantifying Inference Reminiscence

Let’s quantify the reminiscence necessities for inference utilizing the OctoCode mannequin, which has round 15 billion parameters in bfloat16 format (~ 31GB). We’ll use the Transformers library to load the mannequin and generate textual content:

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import torch
mannequin = AutoModelForCausalLM.from_pretrained("bigcode/octocoder",
torch_dtype=torch.bfloat16,
device_map="auto",
pad_token_id=0)
tokenizer = AutoTokenizer.from_pretrained("bigcode/octocoder")
pipe = pipeline("text-generation", mannequin=mannequin, tokenizer=tokenizer)
immediate = "Query: Please write a Python perform to transform bytes to gigabytes.nnAnswer:"
outcome = pipe(immediate, max_new_tokens=60)[0]["generated_text"][len(prompt):]
def bytes_to_gigabytes(bytes):
return bytes / 1024 / 1024 / 1024
bytes_to_gigabytes(torch.cuda.max_memory_allocated())

Output:

The height GPU reminiscence utilization is round 29GB, which aligns with our estimate of 31GB for loading the mannequin parameters in bfloat16 format.

Optimizing Inference Reminiscence with Quantization

Whereas bfloat16 is the widespread precision used for coaching LLMs, researchers have discovered that quantizing the mannequin weights to decrease precision information varieties like 8-bit integers (int8) or 4-bit integers can considerably scale back reminiscence utilization with minimal accuracy loss for inference duties like textual content technology.

Let’s have a look at the reminiscence financial savings from 8-bit and 4-bit quantization of the OctoCode mannequin:

</div>
# 8-bit quantization
mannequin = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_8bit=True, 
pad_token_id=0)
pipe = pipeline("text-generation", mannequin=mannequin, tokenizer=tokenizer)
outcome = pipe(immediate, max_new_tokens=60)[0]["generated_text"][len(prompt):]
bytes_to_gigabytes(torch.cuda.max_memory_allocated())</pre>

Output:

# 4-bit quantization
mannequin = AutoModelForCausalLM.from_pretrained("bigcode/octocoder", load_in_4bit=True,
low_cpu_mem_usage=True, pad_token_id=0)
pipe = pipeline("text-generation", mannequin=mannequin, tokenizer=tokenizer)
outcome = pipe(immediate, max_new_tokens=60)[0]["generated_text"][len(prompt):]
bytes_to_gigabytes(torch.cuda.max_memory_allocated())

Output:

With 8-bit quantization, the reminiscence requirement drops from 31GB to 15GB, whereas 4-bit quantization additional reduces it to simply 9.5GB! This permits working the 15B parameter OctoCode mannequin on shopper GPUs just like the RTX 3090 (24GB VRAM).

Nonetheless, notice that extra aggressive quantization like 4-bit can generally result in accuracy degradation in comparison with 8-bit or bfloat16 precision. There is a trade-off between reminiscence financial savings and accuracy that customers ought to consider for his or her use case.

Quantization is a strong method that may allow LLM deployment on resource-constrained environments like cloud cases, edge gadgets, and even cell phones by drastically lowering the reminiscence footprint.

Estimating Reminiscence for Tremendous-Tuning

Whereas quantization is primarily used for environment friendly inference, methods like tensor parallelism and mannequin parallelism are essential for managing reminiscence necessities in the course of the coaching or fine-tuning of huge language fashions.

The height reminiscence consumption throughout fine-tuning is often 3-4 occasions larger than inference as a consequence of extra reminiscence necessities for:

• Gradients
• Optimizer states
• Activations from the ahead move saved for backpropagation

A conservative estimate is that fine-tuning an LLM with X billion parameters requires round 4 * (2X) = 8X GB of VRAM in bfloat16 precision.

For instance, fine-tuning the 7B parameter LLaMA mannequin would require roughly 7 * 8 = 56GB of VRAM per GPU in bfloat16 precision. This exceeds the reminiscence capability of present GPUs, necessitating distributed fine-tuning methods.

Distributed Tremendous-Tuning Methods

A number of distributed fine-tuning strategies have been proposed to beat GPU reminiscence constraints for big fashions:

1. Information Parallelism: The basic information parallelism method replicates the complete mannequin throughout a number of GPUs whereas splitting and distributing the coaching information batches. This reduces coaching time linearly with the variety of GPUs however doesn’t scale back the height reminiscence requirement on every GPU.
2. ZeRO Stage 3: A complicated type of information parallelism that partitions the mannequin parameters, gradients, and optimizer states throughout GPUs. It reduces reminiscence in comparison with basic information parallelism by conserving solely the required partitioned information on every GPU throughout totally different phases of coaching.
3. Tensor Parallelism: As a substitute of replicating the mannequin, tensor parallelism divides the mannequin parameters into rows or columns and distributes them throughout GPUs. Every GPU operates on a partitioned set of parameters, gradients, and optimizer states, resulting in substantial reminiscence financial savings.
4. Pipeline Parallelism: This method partitions the mannequin layers throughout totally different GPUs/staff, with every system executing a subset of the layers. Activations are handed between staff, lowering peak reminiscence however rising communication overhead.

Estimating reminiscence utilization for these distributed strategies is non-trivial because the distribution of parameters, gradients, activations, and optimizer states varies throughout methods. Furthermore, totally different parts just like the transformer physique and language modeling head might exhibit totally different reminiscence allocation behaviors.

The LLMem Answer

Researchers just lately proposed LLMem, an answer that precisely estimates GPU reminiscence consumption when making use of distributed fine-tuning strategies to LLMs throughout a number of GPUs.

LLMem considers components like recombining parameters earlier than computation (ZeRO Stage 3), output gathering within the backward move (tensor parallelism), and the totally different reminiscence allocation methods for the transformer physique and language modeling head.

Experimental outcomes present that LLMem can estimate peak GPU reminiscence utilization for fine-tuning LLMs on a single GPU with error charges of as much as 1.6%, outperforming the state-of-the-art DNNMem’s common error fee of 42.6%. When making use of distributed fine-tuning strategies to LLMs with over a billion parameters on a number of GPUs, LLMem achieves a formidable common error fee of 3.0%.