Unable to run `NGP` examples

[cecabert]

Hi,

I'm trying to run the train_ngp_nerf_{occ,prop}.py examples on a RTX3090Ti. However I'm always running into the following error when computing the gradient:

Traceback (most recent call last):
  File "nerfacc/examples/train_ngp_nerf_prop.py", line 247, in <module>
    grad_scaler.scale(loss).backward()
  File "site-packages/torch/_tensor.py", line 487, in backward
    torch.autograd.backward(
  File "site-packages/torch/autograd/__init__.py", line 200, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
  File "site-packages/torch/autograd/function.py", line 274, in apply
    return user_fn(self, *args)
  File "site-packages/tinycudann/modules.py", line 107, in backward
    input_grad, params_grad = _module_function_backward.apply(ctx, doutput, input, params, output)
  File "site-packages/torch/autograd/function.py", line 506, in apply
    return super().apply(*args, **kwargs)  # type: ignore[misc]
  File "site-packages/tinycudann/modules.py", line 118, in forward
    input_grad, params_grad = ctx_fwd.native_tcnn_module.bwd(ctx_fwd.native_ctx, input, params, output, scaled_grad)
RuntimeError: include/tiny-cuda-nn/gpu_memory.h:459 cuMemAddressReserve(&m_base_address, m_max_size, 0, 0, 0) failed with error CUDA_ERROR_OUT_OF_MEMORY

I've tried to play with the init_batch_size and target_sample_batch_size values but it didn't make any difference. On the other hand, the train_mlp_nerf.py runs fine. I don't know If it is related to the configuration of the example, or on the tiny-cuda-nn side, or the nerfacc side.

Any help/feedback would be much appreciated.

[yvdu]

I have the same problem here with

RuntimeError: D:/anaconda/envs/base2/Lib/site-packages/tiny-cuda-nn/include\tiny-cuda-nn/gpu_memory.h:592 cuMemSetAccess(m_base_address + m_size, n_bytes_to_allocate, &access_desc, 1) failed: CUDA_ERROR_OUT_OF_MEMORY

I've minimized the batch size but it's not working