+- Selur's Little Message Board (https://forum.selur.net)
+-- Forum: Talk, Talk, Talk (https://forum.selur.net/forum-5.html)
+--- Forum: Small Talk (https://forum.selur.net/forum-7.html)
+--- Thread: torch_tensorrt and TensorRT 10.0 compatibility (/thread-3760.html)
torch_tensorrt and TensorRT 10.0 compatibility - Dan64 - 05.07.2024
Hello Selur,
I'm trying compiling a filter with torch_tensorrt but I found a compatibility problem regarding torch_tensorrt and TensorRT 10.0.
Looking to NVIDIA migration guide
The TensorRT 8.x function build_engine() is substitute by the function build_serialized_network()
Also the TensorRT 8.x parameters INetworkDefinition.has_explicit_precision and INetworkDefinition.has_implicit_batch_dimension() were removed.
But all these parameters are used in the torch_tensorrt 2.4.0.dev installed in Hybrid.
See for example:
torch_tensorrt\fx\fx2trt.py, rows: 211, 213, 257
This version is targeting TensorRT 8.x and not the installed version TensorRT 10.0
I'm wondering if the current Hybrid environment need to be changed to get torch_tensorrt working.
Dan
RE: torch_tensorrt and TensorRT 10.0 compatibility - Dan64 - 05.07.2024
I managed to don't use tensorrt.fx and now I'm able to compile (till now only fp32).
Dan