thunderfx: pass bare torch.compile to _splitter

[crcrpar]

What does this PR do?

Since we use torch.compile w its default values, it doesn't look necessary to register it as an attribute of ThunderCompiler.

[t-vi]

Thank you @crcrpar
Anyone from the ThunderFX side who want's to take a look before merging?

[shino16]

Thank you for pointing this out, I've been wondering why this exists too. This seems to be left off in #1535

[shino16]

Why don't we pass this torch_compile_kwargs to torch.compile for submodules? I think it makes more sense to propagate pytorch options for submodule compilation too.

lightning-thunder/thunder/dynamo/compiler.py

Line 326 in 5ece7ef

compiled = torch.compile(fn, backend=backend, **torch_compile_kwargs)

[crcrpar]

Why don't we pass this torch_compile_kwargs to torch.compile for submodules? I think it makes more sense to propagate pytorch options for submodule compilation too.

lightning-thunder/thunder/dynamo/compiler.py

Line 326 in 5ece7ef

compiled = torch.compile(fn, backend=backend, **torch_compile_kwargs)

I've #2196 and #2198. The PR is a zombie now. Feel free to take it over if you want.

[shino16]

Thank you for the reference. If this was an intended design choice, I won't argue against it.