Remove attention_is_all_you_need and pytorch_struct (#1833)

Summary:
This is a follow-up of #1831

Pull Request resolved: #1833

Reviewed By: janeyx99

Differential Revision: D48432770

Pulled By: xuzhao9

fbshipit-source-id: 2ba8dff02d2ab0df17703427abad126708252762

torchbenchmark/models/ADDING_MODELS.md

@@ -62,7 +62,7 @@ the build steps from install.py.  Avoid getting too fancy trying to be cross-pla
 compatible (e.g., Windows/Mac/etc., or using package managers like yum/dnf) - if it's
 not easy to build, there may be easier models to target.
 
-[Example install.py](attention_is_all_you_need_pytorch/install.py)
+[Example install.py](BERT_pytorch/install.py)
 
 ### Mini-dataset
 By the time install.py script runs, a miniature version of the dataset is expected to be 
@@ -98,7 +98,7 @@ Important: be deliberate about support for cpu/gpu and jit/no-jit.  In the case
 your model is instantiated in an unsupported configuration, the convention is to return
 a model object from \_\_init\_\_ but raise NotImplementedError() from all its methods.
 
-See the [BenchmarkModel API](https://github.com/pytorch/benchmark/blob/master/torchbenchmark/util/model.py) to get started. The [attention is all you need](attention_is_all_you_need_pytorch/__init__.py) benchmark can serve as a good example.
+See the [BenchmarkModel API](https://github.com/pytorch/benchmark/blob/master/torchbenchmark/util/model.py) to get started. The [BERT_pytorch](BERT_pytorch/__init__.py) benchmark can serve as a good example.
 
 ### JIT
 As an optional step, make whatever modifications necessary to the model code to enable it to script or trace.  If doing this,

torchbenchmark/models/nvidia_deeprecommender/__init__.py

@@ -9,7 +9,6 @@
 
 import torch
 
-from torchbenchmark.models.attention_is_all_you_need_pytorch.train import train
 from ...util.model import BenchmarkModel
 from torchbenchmark.tasks import RECOMMENDATION
 from typing import Tuple

torchbenchmark/util/env_check.py

@@ -43,7 +43,6 @@
 # Need lower tolerance on GPU. GPU kernels have non deterministic kernels for these models.
 REQUIRE_HIGHER_TOLERANCE = {
     "alexnet",
-    "attention_is_all_you_need_pytorch",
     "densenet121",
     "hf_Albert",
     "vgg16",

userbenchmark/optim/run.py

@@ -142,7 +142,6 @@ def get_unstable_models() -> Set[str]:
     'LearningToPaint',
     'Super_SloMo',
     'alexnet',
-    'attention_is_all_you_need_pytorch',
     'basic_gnn_edgecnn',
     'basic_gnn_gcn',
     'basic_gnn_gin',
@@ -205,7 +204,6 @@ def get_unstable_models() -> Set[str]:
     'phlippe_resnet',
     'pytorch_CycleGAN_and_pix2pix',
     'pytorch_stargan',
-    'pytorch_struct',
     'pytorch_unet',
     'resnet152',
     'resnet18',
@@ -285,7 +283,7 @@ def get_unstable_models() -> Set[str]:
     # torch.compile()'d optimizer.step() has too many arguments in C++
     # See GH issue: https://github.com/pytorch/pytorch/issues/97361
     {'model': m, 'device': 'cpu', 'func_str': 'pt2_', 'defaults': []} for m in [
-        'BERT_pytorch', 'Background_Matting', 'Super_SloMo', 'attention_is_all_you_need_pytorch',
+        'BERT_pytorch', 'Background_Matting', 'Super_SloMo',
         'densenet121', 'detectron2_fasterrcnn_r_101_c4', 'detectron2_fasterrcnn_r_101_dc5',
         'detectron2_fasterrcnn_r_101_fpn', 'detectron2_fasterrcnn_r_50_fpn', 'detectron2_maskrcnn',
         'detectron2_maskrcnn_r_101_c4', 'detectron2_maskrcnn_r_101_fpn',