Can magpy support graph break?

[hanwen-sun]

Hello,I was wondering if MagPy supports graph breaking. After reviewing the code, I noticed that each frame can have multiple cached_graphs. Additionally, in the dynamic control flow section of the paper, it is mentioned that "MAGPY will cut the user program before the dynamic jump and run this jump eagerly."
However, when I tested MagPy with a simple example involving a print statement and a tensor condition, Dynamo breaks the graph and compiles multiple graphs, but MagPy raises an error. Here is the test case:

import torch
from frontend.compile import compile

def builtin_function(x):
    if x.sum() > 10:
        x = x * 2
    else:
        x = x / 2
    
    return x

x = torch.randn(1024)
compiled = compile(builtin_function)
y = compiled(x)

And another example:

class Example(torch.nn.Module):
    def __init__(self):
        super(Example, self).__init__()
        self.conv = torch.nn.Conv2d(3, 3, 3)
        self.relu = torch.nn.ReLU()

    def forward(self, x):
        x = self.conv(x)
        print(x)
        x = self.relu(x)
        return x

Are there any examples where MagPy performs graph breaking and compiles multiple graphs, caching them? Thank you!

[Jamesswang]

@hanwen-sun Hi, I'm experiencing the same trouble, have you solved the problem now?

[hanwen-sun]

@Jamesswang I think magpy can do graph break and fall back to python, but now it cannot understand when to do graph break, you can refer to the code in guard_tracker.py, when restart() and commit() called, there may be a subgraph. Actually I think the inplementation of magpy is not so complete now.

[Jamesswang]

@hanwen-sun Got it. Thanks for the answer.