GPU vs NPU：Benchmark 模型数量对照

本文只回答一件事：图模式 / torch.compile 模型级 benchmark，GPU（NVIDIA 官方）和 NPU（Ascend）各自白名单里有多少模型。

侧	代码位置
GPU	ref_code/torch_nvidia/pytorch/benchmarks/dynamo/
NPU	ref_code/torch_ascend/pytorch/benchmarks/torchbench/

---

总览

	GPU（NVIDIA dynamo）	NPU（Ascend torchbench）
TorchBench	57	34
HuggingFace	39	5
TIMM	18	0（无独立套件）
合计	约 114	39

NPU 约为 GPU 的 1/3。
NPU 的 TorchBench 白名单里另含若干 hf_* / timm_*（torchbenchmark 仓内封装名），与 GPU 独立 HF/TIMM runner 的命名不是同一套。

---

GPU 支持的模型（约 114）

TorchBench（57）

alexnet, resnet18, resnet50, resnext50_32x4d, densenet121, vgg16, mobilenet_v2, mobilenet_v3_large, mobilenet_v2_quantized_qat, resnet50_quantized_qat, mnasnet1_0, shufflenet_v2_x1_0, squeezenet1_1, pytorch_unet, vision_maskrcnn, yolov3, detectron2_fasterrcnn_r_101_c4, detectron2_fasterrcnn_r_101_dc5, detectron2_fasterrcnn_r_101_fpn, detectron2_fasterrcnn_r_50_c4, detectron2_fasterrcnn_r_50_dc5, detectron2_fasterrcnn_r_50_fpn, detectron2_fcos_r_50_fpn, detectron2_maskrcnn, detectron2_maskrcnn_r_101_c4, detectron2_maskrcnn_r_101_fpn, detectron2_maskrcnn_r_50_c4, detectron2_maskrcnn_r_50_fpn, Background_Matting, Super_SloMo, dcgan, pytorch_stargan, pytorch_CycleGAN_and_pix2pix, DALLE2_pytorch, LearningToPaint, attention_is_all_you_need_pytorch, BERT_pytorch, fastNLP_Bert, speech_transformer, tacotron2, tts_angular, demucs, dlrm, fambench_dlrm, nvidia_deeprecommender, fambench_xlmr, maml, maml_omniglot, soft_actor_critic, drq, moco, opacus_cifar10, pytorch_struct, pplbench_beanmachine, pyhpc_equation_of_state, pyhpc_isoneutral_mixing, pyhpc_turbulent_kinetic_energy

HuggingFace（39）

AlbertForMaskedLM, AllenaiLongformerBase, BartForCausalLM, BertForMaskedLM, BlenderbotForCausalLM, BlenderbotForConditionalGeneration, DebertaV2ForMaskedLM, DistilBertForMaskedLM, DistillGPT2, ElectraForCausalLM, GPT2ForSequenceClassification, GPTJForCausalLM, GPTJForQuestionAnswering, GPTNeoForCausalLM, GPTNeoForSequenceClassification, GoogleFnet, LayoutLMForMaskedLM, M2M100ForConditionalGeneration, MBartForCausalLM, MT5ForConditionalGeneration, MegatronBertForCausalLM, MobileBertForMaskedLM, OPTForCausalLM, PLBartForCausalLM, PegasusForCausalLM, RobertaForCausalLM, T5ForConditionalGeneration, T5Small, TrOCRForCausalLM, XGLMForCausalLM, XLNetLMHeadModel, YituTechConvBert, meta-llama/Llama-3.2-1B, google/gemma-2-2b, google/gemma-3-4b-it, openai/whisper-tiny, Qwen/Qwen3-0.6B, mistralai/Mistral-7B-Instruct-v0.3, openai/gpt-oss-20b

TIMM（18）

adv_inception_v3, beit_base_patch16_224, convnextv2_nano.fcmae_ft_in22k_in1k, deit_base_distilled_patch16_224, deit_tiny_patch16_224.fb_in1k, dm_nfnet_f0, ghostnet_100, inception_v3, mobilenetv2_100, mobilenetv3_large_100, mobilevit_s, nfnet_l0, repvgg_a2, swin_base_patch4_window7_224, tf_efficientnet_b0, visformer_small, vit_base_patch14_dinov2.lvd142m, vit_base_patch16_siglip_256

---

NPU 支持的模型（39）

TorchBench（34）

alexnet, BERT_pytorch, dcgan, densenet121, hf_Albert, hf_Bart, hf_Bert, hf_DistilBert, hf_GPT2_large, hf_T5_base, hf_T5_large, LearningToPaint, lennard_jones, mobilenet_v2, nvidia_deeprecommender, phlippe_densenet, phlippe_resnet, pytorch_stargan, pytorch_unet, resnet152, resnet18, resnet50, resnext50_32x4d, shufflenet_v2_x1_0, soft_actor_critic, speech_transformer, squeezenet1_1, timm_regnet, timm_resnest, timm_vision_transformer, timm_vovnet, torch_multimodal_clip, tts_angular, vgg16

HuggingFace（5）

AlbertForMaskedLM, BartForCausalLM, BertForMaskedLM, DistilBertForMaskedLM, T5ForConditionalGeneration

TIMM

无独立白名单。仅通过 TorchBench 内的 timm_regnet / timm_resnest / timm_vision_transformer / timm_vovnet 覆盖少量家族。

---

一句话

• GPU：三大套件合计约 114 个模型。
• NPU：TorchBench + HF 合计 39 个模型，约为 GPU 的三分之一；缺独立 TIMM，HF 与检测类覆盖明显更少。