Adding support for hpu device

[arathi-hlab]

Adding a new device support named 'hpu'

[bsochack]

all my comments were resolved. It still needs to be reviewed by a maintainer.

[jeromean]

can you paste a snapshot of the results?

[arathi-hlab]

@jeromean here is the result for both train and eval mode

test_BERT_pytorch_train_hpu (main.TestBenchmark) ... ok
test_Background_Matting_train_hpu (main.TestBenchmark) ... ERROR
test_LearningToPaint_train_hpu (main.TestBenchmark) ... ok
test_Super_SloMo_train_hpu (main.TestBenchmark) ... ok
test_alexnet_train_hpu (main.TestBenchmark) ... ok
test_basic_gnn_edgecnn_train_hpu (main.TestBenchmark) ... ERROR
test_basic_gnn_gcn_train_hpu (main.TestBenchmark) ... ok
test_basic_gnn_gin_train_hpu (main.TestBenchmark) ... ok
test_basic_gnn_sage_train_hpu (main.TestBenchmark) ... ok
test_cm3leon_generate_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_dcgan_train_hpu (main.TestBenchmark) ... ok
test_demucs_train_hpu (main.TestBenchmark) ... ok
test_densenet121_train_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_101_c4_train_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_101_dc5_train_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_101_fpn_train_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_50_c4_train_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_50_dc5_train_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_50_fpn_train_hpu (main.TestBenchmark) ... ok
test_detectron2_fcos_r_50_fpn_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "FCOS train is not supported by upstream detectron2. See GH Issue: facebookresearch/detectron2#4369.", skipping...'
test_detectron2_maskrcnn_r_101_c4_train_hpu (main.TestBenchmark) ... ok
test_detectron2_maskrcnn_r_101_fpn_train_hpu (main.TestBenchmark) ... ok
test_detectron2_maskrcnn_r_50_c4_train_hpu (main.TestBenchmark) ... ok
test_detectron2_maskrcnn_r_50_fpn_train_hpu (main.TestBenchmark) ... ok
test_detectron2_maskrcnn_train_hpu (main.TestBenchmark) ... ok
test_dlrm_train_hpu (main.TestBenchmark) ... ok
test_doctr_det_predictor_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_doctr_reco_predictor_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_drq_train_hpu (main.TestBenchmark) ... ok
test_fastNLP_Bert_train_hpu (main.TestBenchmark) ... ERROR
test_functorch_dp_cifar10_train_hpu (main.TestBenchmark) ... ok
test_functorch_maml_omniglot_train_hpu (main.TestBenchmark) ... ok
test_hf_Albert_train_hpu (main.TestBenchmark) ... ok
test_hf_Bart_train_hpu (main.TestBenchmark) ... ok
test_hf_Bert_large_train_hpu (main.TestBenchmark) ... ok
test_hf_Bert_train_hpu (main.TestBenchmark) ... ok
test_hf_BigBird_train_hpu (main.TestBenchmark) ... ok
test_hf_DistilBert_train_hpu (main.TestBenchmark) ... ok
test_hf_GPT2_large_train_hpu (main.TestBenchmark) ... ok
test_hf_GPT2_train_hpu (main.TestBenchmark) ... ok
test_hf_Longformer_train_hpu (main.TestBenchmark) ... ok
test_hf_Reformer_train_hpu (main.TestBenchmark) ... ok
test_hf_Roberta_base_train_hpu (main.TestBenchmark) ... ok
test_hf_T5_base_train_hpu (main.TestBenchmark) ... ok
test_hf_T5_generate_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_hf_T5_large_train_hpu (main.TestBenchmark) ... ok
test_hf_T5_train_hpu (main.TestBenchmark) ... ok
test_hf_Whisper_train_hpu (main.TestBenchmark) ... skipped 'This test is skipped by its metadata'
test_hf_clip_train_hpu (main.TestBenchmark) ... ok
test_hf_distil_whisper_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Training is not implemented.", skipping...'
test_lennard_jones_train_hpu (main.TestBenchmark) ... ok
test_llama_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_llama_v2_7b_16h_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Make sure to set HUGGING_FACE_HUB_TOKEN so you can download weights", skipping...'
test_llava_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_maml_omniglot_train_hpu (main.TestBenchmark) ... ok
test_maml_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "MAML model doesn't support train.", skipping...'
test_microbench_unbacked_tolist_sum_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Train test is not implemented.", skipping...'
test_mnasnet1_0_train_hpu (main.TestBenchmark) ... ok
test_mobilenet_v2_train_hpu (main.TestBenchmark) ... ok
test_mobilenet_v3_large_train_hpu (main.TestBenchmark) ... ok
test_moco_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "hpu not supported", skipping...'
test_moondream_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_nanogpt_train_hpu (main.TestBenchmark) ... ok
test_nvidia_deeprecommender_train_hpu (main.TestBenchmark) ... ERROR
test_opacus_cifar10_train_hpu (main.TestBenchmark) ... ok
test_phlippe_densenet_train_hpu (main.TestBenchmark) ... ok
test_phlippe_resnet_train_hpu (main.TestBenchmark) ... ok
test_pyhpc_equation_of_state_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_pyhpc_isoneutral_mixing_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_pyhpc_turbulent_kinetic_energy_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_pytorch_CycleGAN_and_pix2pix_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "DataParallel is currently not supported on HPU. Please use torch.nn.DistributedDataParallel instead.", skipping...'
test_pytorch_stargan_train_hpu (main.TestBenchmark) ... ok
test_pytorch_unet_train_hpu (main.TestBenchmark) ... ok
test_resnet152_train_hpu (main.TestBenchmark) ... ok
test_resnet18_train_hpu (main.TestBenchmark) ... ok
test_resnet50_train_hpu (main.TestBenchmark) ... ok
test_resnext50_32x4d_train_hpu (main.TestBenchmark) ... ok
test_sam_fast_train_hpu (main.TestBenchmark) ... ERROR
test_sam_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_shufflenet_v2_x1_0_train_hpu (main.TestBenchmark) ... ok
test_simple_gpt_tp_manual_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_simple_gpt_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Model's DEFAULT_TRAIN_BSIZE is not implemented.", skipping...'
test_soft_actor_critic_train_hpu (main.TestBenchmark) ... ok
test_speech_transformer_train_hpu (main.TestBenchmark) ... ok
test_squeezenet1_1_train_hpu (main.TestBenchmark) ... ok
test_stable_diffusion_text_encoder_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Make sure to set HUGGING_FACE_HUB_TOKEN so you can download weights", skipping...'
test_stable_diffusion_unet_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "Make sure to set HUGGING_FACE_HUB_TOKEN so you can download weights", skipping...'
test_tacotron2_train_hpu (main.TestBenchmark) ... ok
test_timm_efficientdet_train_hpu (main.TestBenchmark) ... skipped 'Method train on hpu is not implemented because "The original model code forces the use of CUDA.", skipping...'
test_timm_efficientnet_train_hpu (main.TestBenchmark) ... ok
test_timm_nfnet_train_hpu (main.TestBenchmark) ... ok
test_timm_regnet_train_hpu (main.TestBenchmark) ... ok
test_timm_resnest_train_hpu (main.TestBenchmark) ... ok
test_timm_vision_transformer_large_train_hpu (main.TestBenchmark) ... ok
test_timm_vision_transformer_train_hpu (main.TestBenchmark) ... ok
test_timm_vovnet_train_hpu (main.TestBenchmark) ... ok
test_torch_multimodal_clip_train_hpu (main.TestBenchmark) ... ok
test_tts_angular_train_hpu (main.TestBenchmark) ... ok
test_vgg16_train_hpu (main.TestBenchmark) ... ok
test_vision_maskrcnn_train_hpu (main.TestBenchmark) ... ok
test_yolov3_train_hpu (main.TestBenchmark) ... ok

**Ran 101 tests in 860.358s

FAILED (errors=5, skipped=24)**

========================================================================================================== EVAL MODE
test_BERT_pytorch_eval_hpu (main.TestBenchmark) ... ok
test_Background_Matting_eval_hpu (main.TestBenchmark) ... skipped 'Method eval on hpu is not implemented because "", skipping...'
test_LearningToPaint_eval_hpu (main.TestBenchmark) ... ok
test_Super_SloMo_eval_hpu (main.TestBenchmark) ... ok
test_alexnet_eval_hpu (main.TestBenchmark) ... ok
test_basic_gnn_edgecnn_eval_hpu (main.TestBenchmark) ... ok
test_basic_gnn_gcn_eval_hpu (main.TestBenchmark) ... ok
test_basic_gnn_gin_eval_hpu (main.TestBenchmark) ... ok
test_basic_gnn_sage_eval_hpu (main.TestBenchmark) ... ok
test_cm3leon_generate_eval_hpu (main.TestBenchmark) ... ok
test_dcgan_eval_hpu (main.TestBenchmark) ... ok
test_demucs_eval_hpu (main.TestBenchmark) ... ok
test_densenet121_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_101_c4_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_101_dc5_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_101_fpn_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_50_c4_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_50_dc5_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_fasterrcnn_r_50_fpn_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_fcos_r_50_fpn_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_maskrcnn_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_maskrcnn_r_101_c4_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_maskrcnn_r_101_fpn_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_maskrcnn_r_50_c4_eval_hpu (main.TestBenchmark) ... ok
test_detectron2_maskrcnn_r_50_fpn_eval_hpu (main.TestBenchmark) ... ok
test_dlrm_eval_hpu (main.TestBenchmark) ... ok
test_doctr_det_predictor_eval_hpu (main.TestBenchmark) ... ERROR
test_doctr_reco_predictor_eval_hpu (main.TestBenchmark) ... ERROR
test_drq_eval_hpu (main.TestBenchmark) ... ok
test_fastNLP_Bert_eval_hpu (main.TestBenchmark) ... ERROR
test_functorch_dp_cifar10_eval_hpu (main.TestBenchmark) ... ok
test_functorch_maml_omniglot_eval_hpu (main.TestBenchmark) ... ok
test_hf_Albert_eval_hpu (main.TestBenchmark) ... ok
test_hf_Bart_eval_hpu (main.TestBenchmark) ... ok
test_hf_Bert_eval_hpu (main.TestBenchmark) ... ok
test_hf_Bert_large_eval_hpu (main.TestBenchmark) ... ok
test_hf_BigBird_eval_hpu (main.TestBenchmark) ... ok
test_hf_DistilBert_eval_hpu (main.TestBenchmark) ... ok
test_hf_GPT2_eval_hpu (main.TestBenchmark) ... ok
test_hf_GPT2_large_eval_hpu (main.TestBenchmark) ... ok
test_hf_Longformer_eval_hpu (main.TestBenchmark) ... ok
test_hf_Reformer_eval_hpu (main.TestBenchmark) ... ok
test_hf_Roberta_base_eval_hpu (main.TestBenchmark) ... ok
test_hf_T5_base_eval_hpu (main.TestBenchmark) ... ok
test_hf_T5_eval_hpu (main.TestBenchmark) ... ok
test_hf_T5_generate_eval_hpu (main.TestBenchmark) ... ok
test_hf_T5_large_eval_hpu (main.TestBenchmark) ... ok
test_hf_Whisper_eval_hpu (main.TestBenchmark) ... ok
test_hf_clip_eval_hpu (main.TestBenchmark) ... ok
test_hf_distil_whisper_eval_hpu (main.TestBenchmark) ... ok
test_lennard_jones_eval_hpu (main.TestBenchmark) ... ok
test_llama_eval_hpu (main.TestBenchmark) ... ERROR
test_llama_v2_7b_16h_eval_hpu (main.TestBenchmark) ... skipped 'Method eval on hpu is not implemented because "Make sure to set HUGGING_FACE_HUB_TOKEN so you can download weights", skipping...'
test_llava_eval_hpu (main.TestBenchmark) ... ok
test_maml_eval_hpu (main.TestBenchmark) ... ok
test_maml_omniglot_eval_hpu (main.TestBenchmark) ... ok
test_microbench_unbacked_tolist_sum_eval_hpu (main.TestBenchmark) ... ok
test_mnasnet1_0_eval_hpu (main.TestBenchmark) ... ok
test_mobilenet_v2_eval_hpu (main.TestBenchmark) ... ok
test_mobilenet_v3_large_eval_hpu (main.TestBenchmark) ... ok
test_moco_eval_hpu (main.TestBenchmark) ... skipped 'Method eval on hpu is not implemented because "hpu not supported", skipping...'
test_moondream_eval_hpu (main.TestBenchmark) ... ok
test_nanogpt_eval_hpu (main.TestBenchmark) ... ok
test_nvidia_deeprecommender_eval_hpu (main.TestBenchmark) ... ERROR
test_opacus_cifar10_eval_hpu (main.TestBenchmark) ... ok
test_phlippe_densenet_eval_hpu (main.TestBenchmark) ... ok
test_phlippe_resnet_eval_hpu (main.TestBenchmark) ... ok
test_pyhpc_equation_of_state_eval_hpu (main.TestBenchmark) ... ok
test_pyhpc_isoneutral_mixing_eval_hpu (main.TestBenchmark) ... ok
test_pyhpc_turbulent_kinetic_energy_eval_hpu (main.TestBenchmark) ... ok
test_pytorch_CycleGAN_and_pix2pix_eval_hpu (main.TestBenchmark) ... skipped 'Method eval on hpu is not implemented because "DataParallel is currently not supported on HPU. Please use torch.nn.DistributedDataParallel instead.", skipping...'
test_pytorch_stargan_eval_hpu (main.TestBenchmark) ... ok
test_pytorch_unet_eval_hpu (main.TestBenchmark) ... ok
test_resnet152_eval_hpu (main.TestBenchmark) ... ok
test_resnet18_eval_hpu (main.TestBenchmark) ... ok
test_resnet50_eval_hpu (main.TestBenchmark) ... ok
test_resnext50_32x4d_eval_hpu (main.TestBenchmark) ... ok
test_sam_eval_hpu (main.TestBenchmark) ... ERROR
test_sam_fast_eval_hpu (main.TestBenchmark) ... ERROR
test_shufflenet_v2_x1_0_eval_hpu (main.TestBenchmark) ... ok
test_simple_gpt_eval_hpu (main.TestBenchmark) ... skipped 'Method eval on hpu is not implemented because "Model requires CUDA", skipping...'
test_simple_gpt_tp_manual_eval_hpu (main.TestBenchmark) ... skipped 'Method eval on hpu is not implemented because "Model requires CUDA", skipping...'
test_soft_actor_critic_eval_hpu (main.TestBenchmark) ... ok
test_speech_transformer_eval_hpu (main.TestBenchmark) ... ok
test_squeezenet1_1_eval_hpu (main.TestBenchmark) ... ok
test_stable_diffusion_text_encoder_eval_hpu (main.TestBenchmark) ... skipped 'Method eval on hpu is not implemented because "Make sure to set HUGGING_FACE_HUB_TOKEN so you can download weights", skipping...'
test_stable_diffusion_unet_eval_hpu (main.TestBenchmark) ... skipped 'Method eval on hpu is not implemented because "Make sure to set HUGGING_FACE_HUB_TOKEN so you can download weights", skipping...'
test_tacotron2_eval_hpu (main.TestBenchmark) ... ok
test_timm_efficientdet_eval_hpu (main.TestBenchmark) ... skipped 'Method eval on hpu is not implemented because "The original model code forces the use of CUDA.", skipping...'
test_timm_efficientnet_eval_hpu (main.TestBenchmark) ... ok
test_timm_nfnet_eval_hpu (main.TestBenchmark) ... ok
test_timm_regnet_eval_hpu (main.TestBenchmark) ... ok
test_timm_resnest_eval_hpu (main.TestBenchmark) ... ok
test_timm_vision_transformer_eval_hpu (main.TestBenchmark) ... ok
test_timm_vision_transformer_large_eval_hpu (main.TestBenchmark) ... ok
test_timm_vovnet_eval_hpu (main.TestBenchmark) ... ok
test_torch_multimodal_clip_eval_hpu (main.TestBenchmark) ... ok
test_tts_angular_eval_hpu (main.TestBenchmark) ... ok
test_vgg16_eval_hpu (main.TestBenchmark) ... ok
test_vision_maskrcnn_eval_hpu (main.TestBenchmark) ... ok
test_yolov3_eval_hpu (main.TestBenchmark) ... ok
**Ran 101 tests in 974.877s

FAILED (errors=7, skipped=9)**

[arathi-hlab]

@jeromean can you review and approve the change

[EikanWang]

@xuzhao9 , I wonder if you could help review this PR?