[inductor] TIMM training failures tracker

[desertfire]

[Editied]

benchmarks/timm_models.py -d cuda --inductor --training --float32 --use-eval-mode

Snapshot of Aug 22:

• [x] pytorch/torchdynamo#862
  • jx_nest_base
• [x] pytorch/torchdynamo#965
  • nasnetalarge
  • pnasnet5large
• [ ] OOM Error, pytorch/torchdynamo#822 :
  • adv_inception_v3
  • cait_m36_384
  • beit_base_patch16_224
  • convit_base
  • convmixer_768_32
  • convnext_base
  • deit_base_distilled_patch16_224
  • densenet121
  • dpn107
  • ecaresnet101d
  • gluon_senet154
  • gmixer_24_224
  • gmlp_s16_224
  • legacy_senet154
  • mixer_b16_224
  • mixnet_l
  • mobilevit_s
  • res2net101_26w_4s
  • resnest101e
  • sebotnet33ts_256
  • swin_base_patch4_window7_224
  • xcit_large_24_p8_224
  • volo_d1_224
  • vit_base_patch16_224
  • tf_mixnet_l
  • tnt_s_patch16_224
  • twins_pcpvt_base
• [ ] Accuracy Error :
  • coat_lite_mini
  • dla102
  • ese_vovnet19b_dw
  • fbnetc_100
  • ghostnet_100
  • gluon_inception_v3
  • gluon_xception65
  • hrnet_w18
  • inception_v3
  • mnasnet_100
  • mobilenetv2_100
  • rexnet_100
  • selecsls42b
  • spnasnet_100

The following ones also fail with aot_nvfuser:

• [ ] levit_128
• [ ] convnext_base
• [ ] legacy_senet154

[eellison]

I got errors for both convnext_base and xcit_large_24_p8_224 when running with --accuracy-aot-nop.

legacy_senet154 also fails in eager - Variation in Eager runs itself

[desertfire]

I got errors for both convnext_base and xcit_large_24_p8_224 when running with --accuracy-aot-nop.

legacy_senet154 also fails in eager - Variation in Eager runs itself

Thanks for checking. Do you mind to file separate issues to track those for closer attention to folks working on AOTAutograd?

[desertfire]

Combine with https://github.com/pytorch/pytorch/issues/93777