search

mitsuba-renderer / drjit

Dr.Jit — A Just-In-Time-Compiler for Differentiable Rendering
BSD 3-Clause "New" or "Revised" License
593 stars 43 forks source link

Gradients get disabled in CUDA mode when using suspend_grad + .numpy() #257

Closed dvicini closed 2 months ago

dvicini commented 2 months ago

We ran into a weird issue that seems to only affect the CUDA mode.

The following code:

import drjit as dr
import numpy as np

a = dr.cuda.ad.TensorXf(np.random.rand(16,16))

dr.enable_grad(a)
print(dr.grad_enabled(a))
with dr.suspend_grad():
   a.numpy()

print(dr.grad_enabled(a))

prints

True
False

I.e., the gradients of a somehow get disabled. This does only happen in cuda, when using dr.llvm.ad.TensorXf, everything works as expected. Maybe some form of undefined behavior? Or some bug due to the migration of the tensor to CPU when copying to numpy?