[RFC] Extend Turbine tensor.py to handle torch_dispatches, nn.Module, and torch computational functions

[raikonenfnu]

No description provided.

[raikonenfnu]

@stellaraccident Any thoughts on defaulting to using async-exec mode? So that we won't need to have these guards

https://github.com/nod-ai/SHARK-Turbine/pull/114/files#diff-26bd2a602db8753e912ba73e3cebffe2fd5924c04a2dd090e460e4119f2b8e56R171

[stellaraccident]

@stellaraccident Any thoughts on defaulting to using async-exec mode? So that we won't need to have these guards

https://github.com/nod-ai/SHARK-Turbine/pull/114/files#diff-26bd2a602db8753e912ba73e3cebffe2fd5924c04a2dd090e460e4119f2b8e56R171

No time like the present. The worst it can do is reveal bugs and this is the time to hunt them down.

[raikonenfnu]

I am thinking of unifiying the SpecializedExecutable and the newly created EagerSpecializedExecutable before trying to land this. I think the main problem lies in if we want to introduce execution-model=async-* dynamo/cpu.py backend.

The main issue with this is IIRC a dynamo backend of torch.compile needs to output torch.Tensor. Which means we'd need to sync up all the HalFences in the executable's _returns_to_user method before we convert it to torch.Tensor. This AFAICT would result in it not really being asynchronous per se. 😅

[stellaraccident]

Very neat and a good start! Some general code comments and then this looks good for a first step.