FlagPerf是一款面向AI异构芯片的通用基准测试平台。我们希望探索开源、开放、灵活、公正、客观的AI芯片评测体系,提供行业价值,促进AI产业生态发展。 更多模型及框架支持持续开发中,欢迎加入共同建设,助力AI产业生态发展。
under review表示对应case的支持已开发完毕,在review中;Incoming表示正在添加或计划添加中;N/A表示不支持或尚无计划添加
你可以点击模型或训练框架来跳转到对应case的训练脚本,✅来跳转到对应厂商的运行配置。
模型 | 框架 | 英伟达 | 昆仑芯 | 天数智芯 | 华为昇腾 |
BERT | PyTorch | ✅ | ✅ | ✅ | N/A |
Paddle | ✅ | ✅ | ✅ | N/A | |
MindSpore | N/A | N/A | N/A | Incoming | |
CPM | PyTorch | ✅ | ✅ | ✅ | N/A |
GLM(medium) | PyTorch | ✅ | ✅ | ✅ | N/A |
ResNet50 | PyTorch | ✅ | ✅ | ✅ | N/A |
TensorFlow2 | ✅ | N/A | N/A | N/A | |
MindSpore | N/A | N/A | N/A | under review | |
MobileNetV2 | PyTorch | ✅ | ✅ | ✅ | N/A |
MindSpore | N/A | N/A | N/A | Incoming | |
VisionTransformer | PyTorch | ✅ | ✅ | ✅ | N/A |
FasterRCNN | PyTorch | ✅ | ✅ | ✅ | N/A |
MindSpore | N/A | N/A | N/A | Incoming | |
BigTransfer | PyTorch | ✅ | ✅ | ✅ | N/A |
Wav2Vec2 | PyTorch | ✅ | ✅ | N/A | N/A |
EfficientNet | PyTorch | ✅ | ✅ | ✅ | N/A |
MaskRCNN | PyTorch | ✅ | N/A | ✅ | N/A |
RetinaNet | PyTorch | ✅ | N/A | ✅ | N/A |
Tacotron2 | PyTorch | ✅ | ✅ | ✅ | N/A |
WaveGlow | PyTorch | ✅ | ✅ | N/A | N/A |
Transformer | PyTorch | ✅ | ✅ | ✅ | N/A |
SwinTransformer | PyTorch | ✅ | N/A | ✅ | N/A |
DLRM | PyTorch | Incoming | N/A | N/A | N/A |
RNNT | PyTorch | Incoming | N/A | N/A | N/A |
Llama(7B) | MindSpore | N/A | N/A | N/A | Incoming |
Paddle | under review | N/A | N/A | N/A | |
Llama(13B) | MindSpore | N/A | N/A | N/A | Incoming |
Paddle | Incoming | N/A | N/A | N/A | |
GLM(6B) | MindSpore | N/A | N/A | N/A | Incoming |
GLM(13B) | MindSpore | N/A | N/A | N/A | Incoming |
GPT3(6.7B) | Paddle | Incoming | N/A | N/A | N/A |
GPT3(13B) | Paddle | Incoming | N/A | N/A | N/A |
yolov5_large | PyTorch | Incoming | Incoming | N/A | N/A |
T5_small | PyTorch | ✅ | Incoming | N/A | N/A |
GPT2 | PyTorch | under review | Incoming | N/A | N/A |
TransformerXL | PyTorch | Incoming | Incoming | N/A | N/A |
whisper | PyTorch | Incoming | Incoming | N/A | N/A |
DistilBERT | PyTorch | Incoming | Incoming | N/A | N/A |
Roberta | PyTorch | Incoming | Incoming | N/A | N/A |
DETR | PyTorch | Incoming | Incoming | N/A | N/A |
Longformer | PyTorch | Incoming | Incoming | N/A | N/A |
你可以点击模型来跳转到对应case的推理脚本及结果。
模型 | 英伟达+tensorrt/inductor | 昆仑芯+xtcl | 天数智芯+ixrt |
resnet50 | f32/f16 | f32 | f16 |
BertLarge | f32/f16 | under review | Incoming |
VisionTransformer | f32/f16 | under review | N/A |
Yolov5_large | f32/f16 | Incoming | Incoming |
Stable Diffusion v1.4 | f32/f16 | Incoming | N/A |
SwinTransformer | f32/f16 | Incoming | N/A |
Llama2-7B-mmlu | f32/f16 | N/A | N/A |
Aquila-7B-mmlu | fp16 | N/A | N/A |
DLRM | Incoming | N/A | N/A |
RNNT | Incoming | N/A | N/A |
本项目目前由北京智源人工智能研究院、天数智芯、百度PaddlePaddle、昆仑芯、华为昇腾、华为昇思MindSpore共同建设中。 诚邀各框架、芯片、编译器团队与个人参与!
本项目基于Apache 2.0 license。
本项目的代码来源于不同的代码仓库,关于各模型测试Case的情况,请参考各模型测试Case目录的文档。