Neural Speed is designed to provide the efficient inference of large language models (LLMs) on Intel platforms through the state-of-the-art (SOTA) low-bit quantization and sparsity. The work is highly inspired from llama.cpp and provides the below features:
- Modular design to support new models
- Highly optimized low precision kernels
- Utilize AMX, VNNI, AVX512F and AVX2 instruction set
- Support CPU (x86 platforms only) and Intel GPU (WIP)
- Support 4bits and 8bits quantization
Neural Speed is under active development so APIs are subject to change.
| Hardware | Optimization |
|---|---|
| Intel Xeon Scalable Processors | ✔ |
| Intel Xeon CPU Max Series | ✔ |
| Intel Core Processors | ✔ |
| Intel Arc GPU Series | WIP |
| Intel Data Center GPU Max Series | WIP |
| Intel Gaudi2 | Not yet |
Neural Speed supports the following models:
| Model Name | INT8 | INT4 | Transformer Version | ||
|---|---|---|---|---|---|
| RTN | GPTQ | RTN | GPTQ | ||
| LLaMA2-7B, LLaMA2-13B, LLaMA2-70B | ✅ | ✅ | ✅ | ✅ | Latest |
| LLaMA-7B, LLaMA-13B | ✅ | ✅ | ✅ | ✅ | Latest |
| GPT-J-6B | ✅ | ✅ | Latest | ||
| GPT-NeoX-20B | ✅ | ✅ | Latest | ||
| Dolly-v2-3B | ✅ | ✅ | 4.28.1 or newer | ||
| MPT-7B, MPT-30B | ✅ | ✅ | Latest | ||
| Falcon-7B, Falcon-40B | ✅ | ✅ | Latest | ||
| BLOOM-7B | ✅ | ✅ | Latest | ||
| OPT-125m, OPT-1.3B, OPT-13B | ✅ | ✅ | Latest | ||
| ChatGLM-6B, ChatGLM2-6B | ✅ | ✅ | 4.33.1 | ||
| Baichuan-13B-Chat, Baichuan2-13B-Chat | ✅ | ✅ | 4.33.1 | ||
| Mistral-7B | ✅ | ✅ | ✅ | ✅ | 4.34.0 or newer |
| Qwen-7B, Qwen-14B | ✅ | ✅ | Latest | ||
| Model Name | INT8 | INT4 | Transformer Version | ||
|---|---|---|---|---|---|
| RTN | GPTQ | RTN | GPTQ | ||
| Code-LLaMA-7B, Code-LLaMA-13B | ✅ | ✅ | ✅ | ✅ | Latest |
| StarCoder-1B, StarCoder-3B, StarCoder-15.5B | ✅ | ✅ | Latest | ||
pip install .# Linux and WSL
git submodule update --init --recursive
mkdir build
cd build
cmake .. -G Ninja
ninja# Windows
# Install VisualStudio 2022 and open 'Developer PowerShell for VS 2022'
mkdir build
cd build
cmake ..
cmake --build . -j --config ReleaseThere are two methods for utilizing the Neural Speed:
Please refer to intel extension for transformers for detailed usage.
You can use Python API to run Hugging Face model simply. Here is the sample code:
from transformers import AutoTokenizer, TextStreamer
from intel_extension_for_transformers.transformers import AutoModelForCausalLM
model_name = "Intel/neural-chat-7b-v1-1" # Hugging Face model_id or local model
prompt = "Once upon a time, there existed a little girl,"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
inputs = tokenizer(prompt, return_tensors="pt").input_ids
streamer = TextStreamer(tokenizer)
model = AutoModelForCausalLM.from_pretrained(model_name, load_in_4bit=True)
outputs = model.generate(inputs, streamer=streamer, max_new_tokens=300)warning: If you want to use
from_pretrainAPI: please follow Transformer-based API
You can run LLM with one-click python script including conversion, quantization and inference.
python scripts/run.py model-path --weight_dtype int4 -p "She opened the door and see"
Argument description of run.py (supported MatMul combinations):
| Argument | Description |
|---|---|
| model | Directory containing model file or model id: String |
| --weight_dtype | Data type of quantized weight: int4/int8/fp8(=fp8_e4m3)/fp8_e5m2/fp4(=fp4e2m1)/nf4 (default int4) |
| --alg | Quantization algorithm: sym/asym (default sym) |
| --group_size | Group size: Int, 32/128/-1 (per channel) (default: 32) |
| --scale_dtype | Data type of scales: fp32/bf16/fp8 (dafault fp32) |
| --compute_dtype | Data type of Gemm computation: int8/bf16/fp16/fp32 (default: fp32) |
| --use_ggml | Enable ggml for quantization and inference |
| -p / --prompt | Prompt to start generation with: String (default: empty) |
| -n / --n_predict | Number of tokens to predict: Int (default: -1, -1 = infinity) |
| -t / --threads | Number of threads to use during computation: Int (default: 56) |
| -b / --batch_size_truncate | Batch size for prompt processing: Int (default: 512) |
| -c / --ctx_size | Size of the prompt context: Int (default: 512, can not be larger than specific model's context window length) |
| -s / --seed | NG seed: Int (default: -1, use random seed for < 0) |
| --repeat_penalty | Penalize repeat sequence of tokens: Float (default: 1.1, 1.0 = disabled) |
| --color | Colorise output to distinguish prompt and user input from generations |
| --keep | Number of tokens to keep from the initial prompt: Int (default: 0, -1 = all) |
| --shift-roped-k | Use ring-buffer and thus do not re-computing after reaching ctx_size (default: False) |
Besides the one-click script, Neural Speed also offers the detailed script: 1) convert and quantize, and 2) inference.
Neural Speed assumes the compatible model format as llama.cpp and ggml. You can also convert the model by following the below steps:
# convert the model directly use model id in Hugging Face. (recommended)
python scripts/convert.py --outtype f32 --outfile ne-f32.bin EleutherAI/gpt-j-6b
# or you can download fp32 model (e.g., LLAMA2) from Hugging Face at first, then convert the pytorch model to ggml format.
git clone https://huggingface.co/meta-llama/Llama-2-7b-chat-hf
python scripts/convert.py --outtype f32 --outfile ne-f32.bin model_path
# To convert model with PEFT(Parameter-Efficient Fine-Tuning) adapter, you need to merge the PEFT adapter into the model first, use below command to merge the PEFT adapter and save the merged model, afterwards you can use 'scripts/convert.py' just like above mentioned.
python scripts/load_peft_and_merge.py --model_name_or_path meta-llama/Llama-2-7b-hf --peft_name_or_path dfurman/llama-2-7b-instruct-peft --save_path ./Llama-2-7b-hf-instruct-peft
# quantize weights of fp32 ggml bin
# model_name: llama, llama2, mpt, falcon, gptj, starcoder, dolly
# optimized INT4 model with group size 128 (recommended)
python scripts/quantize.py --model_name llama2 --model_file ne-f32.bin --out_file ne-q4_j.bin --weight_dtype int4 --group_size 128 --compute_dtype int8
# Alternativly you could run ggml q4_0 format like following
python scripts/quantize.py --model_name llama2 --model_file ne-f32.bin --out_file ne-q4_0.bin --weight_dtype int4 --use_ggml
# optimized INT4 model with group size 32
python scripts/quantize.py --model_name llama2 --model_file ne-f32.bin --out_file ne-q4_j.bin --weight_dtype int4 --group_size 32 --compute_dtype int8
Argument description of quantize.py (supported MatMul combinations):
| Argument | Description |
|---|---|
| --model_file | Path to the fp32 model: String |
| --out_file | Path to the quantized model: String |
| --build_dir | Path to the build file: String |
| --config | Path to the configuration file: String (default: "") |
| --nthread | Number of threads to use: Int (default: 1) |
| --weight_dtype | Data type of quantized weight: int4/int8/fp8(=fp8_e4m3)/fp8_e5m2/fp4(=fp4_e2m1)/nf4 (default: int4) |
| --alg | Quantization algorithm to use: sym/asym (default: sym) |
| --group_size | Group size: Int 32/128/-1 (per channel) (default: 32) |
| --scale_dtype | Data type of scales: bf16/fp32/fp8 (default: fp32) |
| --compute_dtype | Data type of Gemm computation: int8/bf16/fp16/fp32 (default: fp32) |
| --use_ggml | Enable ggml for quantization and inference |
Our Neural Speed supports INT4 / INT8 / FP8 (E4M3, E5M2) / FP4 (E2M1) / NF4 weight-only quantization and FP32 / FP16 / BF16 / INT8 computation forward matmul on the Intel platforms. Here are the all supported data types combinations for matmul operations (quantization and forward).
This table will be updated frequently due to active development
| Weight dtype | Compute dtype (default value) | Scale dtype (default value) | Quantization scheme (default value) |
|---|---|---|---|
| FP32 | FP32 | NA | NA |
| INT8 | INT8 / BF16 / FP16 / FP32 (FP32) | BF16 / FP32 (FP32) | sym / asym (sym) |
| INT4 | INT8 / BF16 / FP16 / FP32 (FP32) | BF16 / FP32 (FP32) | sym / asym (sym) |
| FP8 (E4M3, E5M2) | BF16 / FP16 / FP32 (FP32) | FP8 (FP8) | sym (sym) |
| FP4 (E2M1) | BF16 / FP16 / FP32 (FP32) | BF16 / FP32 (FP32) | sym (sym) |
| NF4 | BF16 / FP16 / FP32 (FP32) | BF16 / FP32 (FP32) | sym (sym) |
We provide LLM inference script to run the quantized model. Please reach us if you want to run using C++ API directly.
# recommed to use numactl to bind cores in Intel cpus for better performance
# if you use different core numbers, please also change -t arg value
# please type prompt about codes when run `StarCoder`, for example, -p "def fibonnaci(".
#Linux and WSL
OMP_NUM_THREADS=<physic_cores> numactl -m 0 -C 0-<physic_cores-1> python scripts/inference.py --model_name llama -m ne-q4_j.bin -c 512 -b 1024 -n 256 -t <physic_cores> --color -p "She opened the door and see"
# if you want to generate fixed outputs, please set --seed arg, for example:
OMP_NUM_THREADS=<physic_cores> numactl -m 0 -C 0-<physic_cores-1> python scripts/inference.py --model_name llama -m ne-q4_j.bin -c 512 -b 1024 -n 256 -t <physic_cores> --color -p "She opened the door and see" --seed 12
# if you want to reduce repeated generated texts, please set --repeat_penalty (value > 1.0, default = 1.0), for example:
OMP_NUM_THREADS=<physic_cores> numactl -m 0 -C 0-<physic_cores-1> python scripts/inference.py --model_name llama -m ne-q4_j.bin -c 512 -b 1024 -n 256 -t <physic_cores> --color -p "She opened the door and see" --repeat_penalty 1.2
#Windows
#Recommend to build and run our project in WSL to get a better and stable performance
python scripts/inference.py --model_name llama -m ne-q4_j.bin -c 512 -b 1024 -n 256 -t <physic_cores|P-cores> --color -p "She opened the door and see"Argument description of inference.py:
| Argument | Description |
|---|---|
| --model_name | Model name: String |
| -m / --model | Path to the executed model: String |
| --build_dir | Path to the build file: String |
| -p / --prompt | Prompt to start generation with: String (default: empty) |
| -n / --n_predict | Number of tokens to predict: Int (default: -1, -1 = infinity) |
| -t / --threads | Number of threads to use during computation: Int (default: 56) |
| -b / --batch_size | Batch size for prompt processing: Int (default: 512) |
| -c / --ctx_size | Size of the prompt context: Int (default: 512, can not be larger than specific model's context window length) |
| -s / --seed | NG seed: Int (default: -1, use random seed for < 0) |
| --repeat_penalty | Penalize repeat sequence of tokens: Float (default: 1.1, 1.0 = disabled) |
| --color | Colorise output to distinguish prompt and user input from generations |
| --keep | Number of tokens to keep from the initial prompt: Int (default: 0, -1 = all) |
| --shift-roped-k | Use ring-buffer and thus do not re-computing after reaching ctx_size (default: False) |
| --glm_tokenizer | The path of the chatglm tokenizer: String (default: THUDM/chatglm-6b) |
| --memory-f32 --memory-f16 --memory-auto |
Data type of kv memory (default to auto); If set to auto, the runtime will try with bestla flash attn managed format (currently requires GCC11+ & AMX) and fall back to fp16 if failed |
We support tensor parallelism strategy for distributed inference/training on multi-node and multi-socket. You can refer to tensor_parallelism.md to enable this feature.
You can consider adding your own models via graph developer document.
You can customize the stopping criteria according to your own needs by processing the input_ids to determine if text generation needs to be stopped.
Here is a simple example, which requires a minimum generation length of 80 tokens. Once the min_length is met, encountering a terminator eos_token_id will end the generation.
import torch
from typing import List
from transformers import StoppingCriteria, StoppingCriteriaList
class StopOnTokens(StoppingCriteria):
def __init__(self, min_length: int, start_length: int, stop_token_id: List[int]):
self.min_length = min_length
self.start_length = start_length
self.stop_token_id = stop_token_id
def __call__(
self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs
) -> bool:
if input_ids.shape[-1] - self.start_length > self.min_length:
for stop_id in self.stop_token_id:
if input_ids[0][input_ids.shape[-1] - 1] == stop_id:
return True
return False
stopping_criteria = StoppingCriteriaList(
[
StopOnTokens(
min_length=80,
start_length=inputs.shape[1],
stop_token_id=[tokenizer.eos_token_id],
)
]
)
outputs = model.generate(inputs, streamer=streamer, stopping_criteria=stopping_criteria)Enable verbose mode and control tracing information using the NEURAL_SPEED_VERBOSE environment variable.
Available modes:
- 0: Print all tracing information. Comprehensive output, including: evaluation time and operator profiling.
- 1: Print evaluation time. Time taken for each evaluation.
- 2: Profile individual operator. Identify performance bottleneck within the model.