Top Level Torch Compile Issue Tracker

[drisspg]

Summary

This issue is used to track multiple issues encountered with torch.compile for Float8Tensor

Float8Experimental

Sub-Issues/Pull Requests

• PR to add subclass at graph_boundary: Add tests for Float8Tensor at graph boundaries #196
• ISSUE: Use Activation Hooks failing with AotAutograd for dynamic linear #223
• PR: The min cut partitioner does not want to recmpute weight cast: Enable fp8_weight recomputation during backwards pass #185

PyTorch Core

Sub-Issues/Pull Requests

• ISSUE: request: torch.compile support for tensor subclass subgraph boundaries pytorch/pytorch#117115
  • PR: fix multiple-fake-modes bug with compile + subclasses pytorch/pytorch#118191
• ISSUE: Min Cut Partitioner Issue with float8_experimental pytorch/pytorch#117901
• ISSUE: dynamo <> autograd.Function assumes that grad_outputs have the same subclass-ness as forward outputs pytorch/pytorch#117662

• ISSUE: torch.compile + __torch_dispatch__ subclasses: support custom attribute accesses on intermediates pytorch/pytorch#117596
• ISSUE: dynamo + autograd.Function: dynamo doesn't model multiple ctx.save_for_backward calls. pytorch/pytorch#117652 (not blocking, I forgot to push my changes to the PR)
• PR: dynamo: support attribute access on tensor subclasses without sources pytorch/pytorch#117666
• PR: dynamo: respect autograd.Function + multiple save_for_backward calls pytorch/pytorch#117667
• ISSUE: Dynamo codegen for mapping outputs back to user variables should contain safety checks (no fake tensors, no variable trackers) pytorch/pytorch#118211

[vkuzo]

closing this since this seems done, but please feel free to reopen if needed