Painless Inference Acceleration (PIA)

A toolkit for LLM inference without 😭 . Currently it contains our work LOOKAHEAD, a framework which accelerates LLM inference without loss of accuracy, other works will release soon.

News or Update 🔥

• [2024/01] We support all models of baichuan family (Baichuan-7b & 13b, Baichuan2-7b & 13b).

• [2024/01] We fully support repetition_penalty parameter.

• [2024/01] We support Mistral & Mixtral. example

• [2023/12] We released our Lookahead paper on arXiv!

• [2023/12] PIA released 💪 !!! Fast, Faster, Fastest 🐆 !!!

Models we support

• GLM
• Baichuan & Baichuan 2
• BLOOM
• ChatGLM 2 & 3
• GPT-2
• GPT-J
• InterLM
• LLaMA & LLaMA-2
• Mistral
• Mixtral
• OPT
• Qwen

Known issuss & TODO

ISSUE 1. Repetition_penalty is not fully supported, we will fix it in the future.

ISSUE 2. Lookahead may generate responses different from original ones due to low-precise data type (i.e., fp16 or bf16), the responses would be the same with fp32.

ISSUE 3. Baichuan tokenizer cannot be initialized with the lastest version transformers (4.30.2 can work).

ISSUE 4. Qwen model may generate slightly different responses with lookahead when the repetition_penalty parameter is set.

TODO1: Support the latest version 🤗 transformers ]. Currently it's based on 4.30.2.

TODO2: Integrate our work FastCoT

TODO3: Optimize batch inference implementation with flash-attention.

Performance Comparison

Performance is measured by token/s(tokens per second) of generation tokens.

Public datasets and models

We use the test set for evaluation and the train set for trie-tree cache construction. The hyper-parameters are tuned by grid searching. The tag fused indicates operators are fused with triton, the implementation can be found in modeling_llama_batch.py.

We test 5 examples with Llama2-7b-chat and dolly dataset, inference time without lookahead (the left figure) is 15.7s (48.2token/s), while inference time with lookahead is 6.4s (112.9token/s), speedup is 2.34.

Private datasets and models

We use the first 1000 samples for evaluation and the rest for trie-tree cache construction. The hyper-parameters are decoding_length=128 and branch_lenght=32.

Our method could obtain significant acceleration in RAG (Retrieval Augmented Generation) scenarios. However, there is no real-life datasets available currently. Therefore, we only evaluate on our private datasets and models. AntGLM-10B is a LLM developed by Ant Group with GLM architecture.

We test 5 examples with AntGLM-10B and AntRag dataset, inference time without lookahead (the left figure) is 16.9s (33.8token/s), while inference time with lookahead is 3.9s (147.6token/s), speedup is 4.37.

Introduction

Our repo PIA (short for Painless Inference Acceleration) is used for LLM inference, it is based on 🤗 transformers library.

• It uses an on-the-fly trie-tree cache to prepare hierarchical multi-branch drafts, without the demand for assist models (e.g., speculative decoding) or additional head training (e.g., block decoding). With the efficient hierarchical structure, we can lookahead tens fo branches, therefore significantly improve generated tokens in a forward pass.

• You can also benefit from our optimized fuesed operation kernels.

Note that our work is different from the other method named lookahead decoding.

Lookahead workflow

Lookahead mask

Trie construction

Trie retrieve

Lincense

CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)

Installation

1. Clone this repository and navigate to PainlessInferenceAcceleration

   git clone https://github.com/alipay/PainlessInferenceAcceleration.git
   cd PainlessInferenceAcceleration

2. Install Package

   python setup.py install

Quick Start

Below is an example for the simplest use of lookahead to inference:

import torch
from transformers import AutoTokenizer

from pia.lookahead.common.lookahead_cache import LookaheadCache
from pia.lookahead.models.llama.modeling_llama import LlamaForCausalLM

model_dir = 'meta-llama/Llama-2-7b-chat-hf'
model = LlamaForCausalLM.from_pretrained(model_dir
                                         , cache_dir='./'
                                         , torch_dtype=torch.float16
                                         , low_cpu_mem_usage=True
                                         , device_map='auto'
                                         )
tokenizer = AutoTokenizer.from_pretrained(model_dir)

prompt = "Hello, I'm am conscious and"
inputs = tokenizer(prompt, return_tensors="pt")

output_ids = model.generate(input_ids=inputs.input_ids.cuda(),
                            attention_mask=inputs.attention_mask.cuda(),
                            max_new_tokens=256,
                            decoding_kwargs={'use_lookahead': True}
                            )
response = tokenizer.decode(output_ids[0].tolist())
print(f'{response=}')

To use lookahead with other models, we can run the scripts in the path examples/. Each supported models are included and can be used for correctness evaluation.

python [model name]_example.py

Benchmarks

To evaluation speedup of lookahead, we can run the scripts in the path benchmarks/, the preprocess of datasets can be found in benchmarks/preprocess_sample.py.

To inspect running details of lookahead, we can turn on return_dict_in_generate, i.e.,

outputs = model.generate(...,
                        return_dict_in_generate=True
                        )
output_ids = outputs.sequences
kwargs = outputs.kwargs
# edls: short for effective decoding lengths, i.e., generate token count in a forward, therefore edls always >=1 ( even without lookahead, we will generate one token in a forward, so edls=1)
edls = kwargs['edls']
# dls: short of decoding lengths, i.e., token count in a forward, always >= 1. Note that it is set to 1 intead of prompt length in the prefill stage.
dls = kwargs['dls']
# fts: short for forward time(s), the first is the prefill time and others are decoding times.
fts = kwargs['fts']
#qts: short of query time(s), i.e., the time for retrieving a sub trie tree.
qts = kwargs['qts']

Customize Model

Tests

Tests can be run with:

cd pia/lookahead
pytest tests/ -s

Citations

@misc{zhao2023lookahead, title={Lookahead: An Inference Acceleration Framework for Large Language Model with Lossless Generation Accuracy}, author={Yao Zhao and Zhitian Xie and Chenyi Zhuang and Jinjie Gu}, year={2023}, eprint={2312.12728}, archivePrefix={arXiv}, primaryClass={cs.IR} }