1.Tensorrt-llm安裝

os: ubuntu 22.04

1.1搭建docker 環境

切換到 root 用戶

sodu passwd root

更新apt

sudo apt-get update --fix-missing 

更新docker

sudo apt-get upgrade docker-ce

安裝nvidia 容器運行時，避免如下錯誤

Error response from daemon: could not select device driver "" with capabilities: [[gpu]].

• 安裝方法如?Installing the NVIDIA Container Toolkit?

安裝nvidia image

docker run  --runtime=nvidia --gpus all --name tllm --entrypoint /bin/bash -it nvidia/cuda:12.3.0-devel-ubuntu22.04 

注意： cuda 版本不必和宿主機cuda 版本一致。

1.2.?安裝tensorrt-llm

安裝python 環境

apt-get update && apt-get -y install python3.10 python3-pip openmpi-bin libopenmpi-dev git git-lfs

安裝tensorrt-llm

pip3 install tensorrt_llm==0.10.0 -U  --extra-index-url https://pypi.nvidia.com

注意：截至2024年6月建議安裝0.10.0

2. 量化

weight only 量化

轉換huggince face 模型為checkpoint

CUDA_VISIBLE_DEVICES=0 python convert_checkpoint.py \
--model_version v2_7b --model_dir ./baichuan7B \
--dtype float16 --output_dir ./ckpt --use_weight_only

編譯模型

trtllm-build --checkpoint_dir /ckpt \
--output_dir /engine --gemm_plugin float16 \--max_batch_size=1 --max_input_len=2048 \
--max_output_len=128

逐層量化

在轉換時期修改

/usr/local/lib/python3.10/dist-packages/tensorrt_llm/models/convert_utils.py(65)weight_only_quantize_dict()

其中exclusive 為不想量化的decode block

def weight_only_quantize_dict(weights: Dict[str, torch.Tensor],quant_algo: str,quant_weights=['qkv.weight', 'dense.weight', 'fc.weight','proj.weight', 'gate.weight'],plugin: bool = True):exclusive = ['24','25', '26', '27', '28','29','30','31'] # <- Hereif quant_algo not in [QuantAlgo.W4A16, QuantAlgo.W8A16]:return weightsfor name in list(weights):"""exclu """conti = False for exlu in exclusive:if exlu in name:conti = Trueprint("exclu: ", exlu) if conti:continue"""ori code """if any([_name in name for _name in quant_weights]) and weights[name].dtype != torch.int8:quant_weight, quant_scale = weight_only_quantize(weight=weights[name], quant_algo=quant_algo, plugin=plugin)weights[name] = quant_weightweights[name.replace('.weight', '.per_channel_scale')] = quant_scalereturn weights

在之前trtllm-build前修改

修改/usr/local/lib/python3.10/dist-packages/tensorrt_llm/quantization/quantize.py

同樣exclusive_modules 為不想量化的層。

def weight_only_quantize(model,quant_config: QuantConfig,current_key_name=None):assert quant_config.quant_mode.is_weight_only()exclude_modules = quant_config.exclude_modules or ['lm_head', 'router', '20', '21', '22', '23', '24','25','26','27','28','29','30','31']for name, module in model.named_children():if current_key_name is None:current_key_name = []current_key_name.append(name)print(current_key_name)if len(list(module.children())) > 0:weight_only_quantize(module, quant_config, current_key_name)if isinstance(module, ColumnLinear) and name not in exclude_modules and current_key_name[2] not in exclude_modules:

精度比較

量化層數
全部32層	98.78%
量化0-27層	98.87%
量化0-23層	99.06%
fp16	99.26

可以看出放棄最后幾層的量化是對模型精度有略微提升的。

smoothquant 量化

python convert_checkpoint.py --model_version v2_7b --model_dir ./baichuan7b --dtype float16 --output_dir /ckpt -sq 0.5 --per_token --per_channel 

CUDA_VISIBLE_DEVICES=0 trtllm-build --checkpoint_dir /ckpt --output_dir /engine --gemm_plugin float16 --max_batch_size=1 --max_input_len=1024 --max_output_len=256