Skip to content

PyTorch2.0's jit is still not good enough to support jit in tc #139

@refraction-ray

Description

@refraction-ray
Contributor

Issue Description

Simply replacing torch.jit.script or torch.jit.trace with backend.jit still fail for tc functions

Example scripts:

@torch.jit.script
def f(param):
    c = tc.Circuit(6)
    for i in range(5):
        for j in range(5):
            c.rzz(i, i+1, theta=param[i, j])
    return c.expectation_ps(z=[1])

f(torch.ones([5, 5]))

or

@partial(torch.jit.trace, example_inputs=torch.ones([5, 5]))
def f(param):
    c = tc.Circuit(6)
    for i in range(5):
        for j in range(5):
            c.rzz(i, i+1, theta=param[i, j])
    return c.expectation_ps(z=[1])

f(torch.ones([5, 5]))

actually the latter somehow works, but very fragile, for example, if the jit transformation is nested with grad or vmap operation, torch mostly fails

Proposed Solution

  1. Wait for further development of torch or 2. use tf/jax backend with torch interface instead or 3. actually maybe slightly fix in the exsisting tc codebase may work but currently have no time to try 4. or try torch.compile later.

Additional References

Activity

refraction-ray

refraction-ray commented on Apr 10, 2023

@refraction-ray
ContributorAuthor

On the other hand, torch.vmap seems to work fine at least at syntax level, detailed performance is not benchmarked

refraction-ray

refraction-ray commented on Apr 10, 2023

@refraction-ray
ContributorAuthor
refraction-ray

refraction-ray commented on Apr 26, 2024

@refraction-ray
ContributorAuthor

torch2.3 is okay for functional transformation nesting, but this version doesn't include support for macOS x86...

https://dev-discuss.pytorch.org/t/pytorch-macos-x86-builds-deprecation-starting-january-2024/1690

pytorch/pytorch#114602

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

      Development

      No branches or pull requests

        Participants

        @refraction-ray

        Issue actions

          PyTorch2.0's jit is still not good enough to support jit in tc · Issue #139 · tencent-quantum-lab/tensorcircuit