ColossalAI/colossalai/inference/pipeline

🐳 Pipeline Inference

Table of Contents

• 💡 Introduction
• 🔗 Design
• 🔨 Usage
  • Example
  • Quick start
• 📊 Performance

Introduction

Pipeline Inference is a module designed to make inference on a pipeline way. In inference systems, although there is no need to store intermediate information such as activations during forward propagation for backward propagation, the weights of some larger models still cannot fit on a single GPU for inference. This requires us to use model parallelism and other methods to reduce the memory occupation on a single GPU. Pipeline parallelism, as one of the traditional model parallelism approaches, has been widely used due to its reduced all-reduce communication requirements and simple layout. The main issue with pipeline parallelism, known as bubbles, can be almost eliminated in inference because the backward propagation that causes bubbles no longer exists in inference. This makes pipeline parallelism almost bubble-free in the ideal scenario where the sequence length is the same across the pipeline.

Design

Pipeline Inference is composed of three parts: PPInferEngine, MicroBatchManager and generate schedule.

1. PPInderEngine is the High-Level API for users to use. It is responsible for the following tasks:

   • Initialize the pipeline inference environment with PipelineStageManager and mdoel with ShardFormer.
   • Run the pipeline inference model.

2. MicroBatchManager is a structure to manage the micro-batch information. It is responsible for the following tasks:

   • Record each micro-batch information, like generated new tokens and kvcache.
   • Record each micro-batch inference state, like prefill, generate or done.
   • Update the micro-batch information.

3. generate schedule implements the simple pipeline inference layout. When pipeline size is 2, we use torch.distributed.P2Pop to implement the communication between stages, mainly to solve the race communication. When pipeline size is larger than 2, we use torch.distributed.broadcast which is faster than torch.distributed.P2Pop.

Usage

Example

from colossalai.pipeline import PPInferEngine
# Suppose the pipeline size is 2, and use fp16 to do infenrence. Use Llama as an example.
model = LlamaForCausalLM.from_pretrained('/path/to/model')
inputs = tokenizer("Hello, my dog is cute", "What a good day", return_tensors="pt")
engine = PPInferEngine(
    pp_size=2,
    dtype='fp16',
    micro_batch_size=1,
    new_length=10,
    model=model,
    model_policy=LlamaForCausalLMPipelinePolicy())

output = engine.inference([inputs])

Quick start

cd benchmark
sh run.sh

Performance

We conducted multiple benchmark tests to evaluate the performance. We compared the inference latency and throughputs between Pipeline Inference and hugging face pipeline. The test environment is 2*A10, 20G.

Llama Throughput(tokens/s)

7b, fp16

batch_size(micro_batch size)	2(1)	4(2)	8(4)	16(8)	32(8)	32(16)
Pipeline Inference(1024, 128)	33.31	59.98	98.92	143.47	152.61	OOM
Hugging Face(1024, 128)	41.43	65.30	91.93	114.62	OOM	OOM
Pipeline Inference(512, 512)	43.37	82.81	148.03	229.06	238.67	312.82
Hugging Face(512, 512)	49.13	84.91	132.87	178.30	OOM	OOM

7b, fp32

batch_size(micro_batch size)	2(1)	4(2)	8(4)	16(4)
Pipeline Inference(1024, 128)	20.61	31.23	45.20	47.46
Hugging Face(1024, 128)	19.80	29.37	OOM	OOM
Pipeline Inference(512, 512)	28.07	46.76	79.35	81.70
Hugging Face(512, 512)	25.67	43.97	60.67	OOM

13b, fp16

batch_size(micro_batch size)	2(1)	4(2)	8(4)	16(4)
Pipeline Inference(1024, 128)	21.73	38.06	61.02	64.30
Hugging Face(1024, 128)	23.48	37.59	53.44	OOM
Pipeline Inference(512, 512)	26.65	49.48	86.11	88.44
Hugging Face(512, 512)	27.45	47.74	74.46	OOM