FindDefinition

tensorrt outputs are allocated once when engine created, so you don't need to assign a new ptr, just use outputs to create ```tv::Tensor``` by ```tv::from_blob``` you should check tensorrt plugin...

1. you should use ```get_indice_gen_tensors_from_workspace``` instead of manage workspace by your self. 2. If the spatial shape is too large (trigger use_int64_hash_k), the misaligned may occur. you can change ```hash_size...

if you use tensorrt, you can use matmul layer in tensorrt (skip sparse conv when construct tensorrt network), no need to implement a new one. By the way, torch use...

pytorch installed in conda will install a nvcc compiler (11.7 in your case), when you activate this environment, cuda 11.7 will override system cuda. so there are some ways to...

@callzhang can't reproduce your problem on pip installed torch 1.13+spconv-cu120+cumm-cu120+4090+driver525+cuda 11.8. Could you provide a mimimal reproduce script for this problem? My test command: '''python -m spconv.benchmark bench_basic f16'''

@Zhongdao could you share a script to help me reproduce this? I can't reproduce in my environment.

@Deephome could you provide a minimal reproduce script for this problem? I can't fix this without reproduce it in my development environment.

I don't have any mac os devices, so I can't provide prebuilts and support for mac os, you can try to install from source in mac os (cpu only).

should be fixed in cumm >= 0.4.10. you can install it and try again.

int32 is enough for indices, the possible problem is indices scalar overflow. that function already have code that switch to int64 if needed to avoid overflow. Please provide more information...