Initial support - AOTAutograd - Test accuracy for TorchBench models

[anijain2305]

List of bugs
Eager

• torch.ops.aten.add_ is slow - Calling torch.ops.aten.add_ is ludicrously slow pytorch#74943

Torchscript bugs

• fastNLP_BERT - https://github.com/facebookresearch/torchdynamo/issues/93
• moco - Debug TorchScript error from moco #94
• slomo - Debug TorchScript error for Slomo #95
• -inf error - Handle -inf for Torchscript of FX graphs #124
• vision maskrcnn - Debug error for vision maskrcnn #170

TorchDyanamo bugs

• hf_GPT2/speech transformer/hf_t5 - Debug issue with AOTAutograd for speech_transformer/hf_GPT2/hf_T5 #85
• tacotron2 - Debug issue with tacotron2 #82

Torchbench issues

• Most of them tracked at Enable train tests when CI infra supports better GPU benchmark#652

AOTAutograd issues

• Functionalization tracker - AOT Autograd - Functionalization pass with torch.dispatch #88
• (inplace) pytorch_struct - AOT Autograd - Default partitioner fails for gather op (model pytorch_struct) functorch#591
• (inplace) pytorch_struct second error - AOT Autograd fails to get correct grads for view and Inplace Relu functorch#514
• tts_angular - AOT Autograd - LSTM - grads not generated (model tts_angular) functorch#586
• Contiguous tensors - AOT Autograd - Contiguous tensors functorch#537

NVFuser issues

• Assertion Error 1 - codegen uses undefined identifier csarofeen/pytorch#1502
• Jit trace fails for training
• Replay error - Codegen error on transform_replay csarofeen/pytorch#1538
• index error

[ciciewang]

need to work on the last 20% of the model.

[ciciewang]

will also polish up the 80%.

[ciciewang]

resolved one failed issue, one more to go.

[anijain2305]

Stack has changed over time. Closing this. Will open new one for errors.