📖 [Story] Coverage for Core ATen Ops

[gs-olive]

TL;DR

Prioritze coverage for the core ATen opset.

Goal(s)

• Determine priority order, based on criteria such as key model requirements, which operators from the core ATen opset should be implemented first
• Track progress of operator implementations and visualize tasks

Tasks

Tasks

Beta Give feedback

1. Add support for aten.copy #2435
   +0
   
2. Add support for logicial aten operators #2439
   6 of 6
   component: converters feature request +2
   
3. Add support for padding aten operators #2461
   9 of 9
   +0
   
4. Support aten._scaled_dot_product_flash_attention #2427
   +0
   
5. aten.remainder.Scalar #2564
   component: converters +1
   
6. aten.arange.start_step #2492
   +0
   
7. aten.sort #2508
   +0
   
8. aten.reflection_pad1d
9. aten.remainder.Tensor #2565
   component: converters +1
   
10. aten.roll #2567
    component: converters +1
    
11. aten._pdist_forward #2568
    component: converters +1
    
12. aten.any #2545
    +0
    
13. aten.topk #2499
    +0
    
14. aten.replication_pad3d
15. aten.reflection_pad2d
16. aten.clamp.Tensor #2517
    +0
    
17. aten.pixel_shuffle #2594
    component: converters +1
    
18. aten.scalar_tensor #2593
    component: converters +1
    
19. aten.upsample_nearest2d.vec
20. aten.replication_pad2d
21. aten.constant_pad_nd
22. aten.adaptive_avg_pool1d #2603
    component: converters +1
    
23. aten.reflection_pad3d
24. aten.amin #2493
    +0
    
25. aten.argmin #2498
    +0
    
26. aten.trunc #2536
    component: converters +1
    
27. aten.flip #2535
    component: converters +1
    
28. aten.log10 #2602
    +0
    
29. aten.log2 #2601
    +0
    
30. aten.any.dim
31. aten.isnan #2712
    component: converters +1
    
32. aten.gather #2534
    component: converters +1
    
33. aten.as_strided #2734
    component: converters +1
    
34. aten._cdist_forward #2725
    component: converters +1
    
35. aten._local_scalar_dense #2743
    Blocked [TensorRT] component: converters +2
    
36. aten.sym_storage_offset #2757
    component: converters +1
    
37. aten.index_put #2544
    +0
    
38. Add support for aten.select_scatter #2436
    +0
    
39. aten.scatter_add #2737
    +0
    
40. aten.scatter.value #2705
    +0
    
41. aten.select_scatter
42. aten.scatter_reduce.two #2739
    +0
    
43. Add support for aten.slice_scatter #2434
    +0
    
44. aten.native_dropout #2494
    component: converters +1
    
45. aten.sym_size.int #2496
    +0
    
46. aten.sym_stride.int #2497
    +0
    
47. aten.scatter.src
48. aten.empty_strided #2758
    component: converters +1
    
49. aten.randn #2571
    +0
    
50. aten.rand #2572
    +0
    
51. aten.randperm #2573
    +0
    
52. aten.empty.memory_format #2738
    +0
    
53. aten.atan2 #2586
    component: converters +1
    
54. aten.expm1 #2713
    component: converters +1
    
55. aten.adaptive_avg_pool2d #2611
    component: converters +1
    
56. aten.adaptive_avg_pool3d #2612
    component: converters +1
    
57. aten.index_select #2708
    component: converters +1
    
58. aten.embedding_bag #2533
    component: converters +1
    
59. aten.nonzero #2516
    Blocked [TensorRT] +1
    
60. aten._adaptive_avg_pool2d_backward
61. aten.log1p #2760
    component: converters +1
    
62. aten.atan2.out #2828
    component: converters +1
    
63. aten.resize_ #2872
    component: converters +1
    
64. aten.diagonal #2873
    component: converters +1
    

[zewenli98]

Based on my common sense, the following ops should be prioritized:

Element-wise ops (I'm writing):
torch.ops.aten.ne.Scalar
torch.ops.aten.ne.Tensor
torch.ops.aten.ge.Scalar
torch.ops.aten.ge.Tensor
torch.ops.aten.le.Scalar
torch.ops.aten.le.Tensor
torch.ops.aten.bitwise_and.Scalar
torch.ops.aten.bitwise_and.Tensor
torch.ops.aten.bitwise_and.Scalar_Tensor
torch.ops.aten.bitwise_or.Scalar
torch.ops.aten.bitwise_or.Tensor
torch.ops.aten.bitwise_or.Scalar_Tensor
torch.ops.aten.bitwise_xor.Scalar
torch.ops.aten.bitwise_xor.Tensor
torch.ops.aten.bitwise_xor.Scalar_Tensor
torch.ops.aten.bitwise_not

Padding-related ops:
torch.ops.aten.pad.default
torch.ops.aten.constant_pad_nd.default
torch.ops.aten.reflection_pad1d.default
torch.ops.aten.reflection_pad2d.default
torch.ops.aten.reflection_pad3d.default
torch.ops.aten.replication_pad1d.default
torch.ops.aten.replication_pad2d.default
torch.ops.aten.replication_pad3d.default

Others:
torch.ops.aten.amin
torch.ops.aten.argmin
torch.ops.aten.arange.start_step
torch.ops.aten.native_dropout
torch.ops.aten.rand
torch.ops.aten.randn
torch.ops.aten.sort
torch.ops.aten.copy
torch.ops.aten.topk
torch.ops.aten.clamp
torch.ops.aten.isnan
torch.ops.aten.nonzero
torch.ops.aten.index_select
torch.ops.aten.flip
torch.ops.aten.trunc

[narendasan]

Additional context #1809