Simon Fan

icon indicating an email [email protected]

icon location San Francisco

Results 11 comments of Simon Fan

DISABLED test_unused_output (__main__.TestAutogradWithCompiledAutograd)

https://github.com/pytorch/pytorch/pull/126146

DISABLED test_type_conversions (__main__.TestAutogradWithCompiledAutograd)

https://github.com/pytorch/pytorch/pull/126146

DISABLED test_too_many_grads (__main__.TestAutogradWithCompiledAutograd)

https://github.com/pytorch/pytorch/pull/126146

[compiled autograd][cudagraphs] Inputs runtime wrapper to move cpu scalars to cuda

@pytorchbot merge

DISABLED test_power_function (__main__.TestAutogradWithCompiledAutograd)

https://github.com/pytorch/pytorch/pull/126146

[compiled autograd] torch.compile API

@albanD Could you elaborate more? If some tensor's autograd graph contains a node created by torch.compile (e.g. CompiledFunctionBackward), it should run compiled autograd. The context manager API that would let...

[compiled autograd] torch.compile API

Yes, CA tries to compile the entire backward, even the nodes originating from fw graph break regions. Without this, we would have basically the same graphs as AOTAutograd, and can't...

[compiled autograd] torch.compile API

That makes a lot of sense to me, I'll explore the implementation

[compiled autograd] torch.compile API

@jansel could you take another look

[compiled autograd] torch.compile API

@pytorchbot merge