TensorRT 8.6

[johnnynunez]

When will tensor 8.6 be added? On new GPU's using CUDA 12 does not work with version 8.5
https://github.com/NVIDIA/TensorRT/releases/tag/v8.6.0

[gs-olive]

We are working to switch our main branch over to TRT 8.6 now (PR #1852), and our upcoming release 1.4 will be based on TRT 8.6. Switching to CUDA 12 is not yet planned, since our current Torch distribution dependency does not have a CUDA 12-compatible build, but you can build Torch-TensorRT from scratch using your own Torch/CUDA versions by replacing the libtorch, libtorch_pre_cxx11_abi, and cuda sections in the WORKSPACE file, and rebuilding Torch-TensorRT from scratch:

TensorRT/WORKSPACE

Lines 40 to 70 in 1d78f43

	# CUDA should be installed on the system locally
	new_local_repository(
	name = "cuda",
	build_file = "@//third_party/cuda:BUILD",
	path = "/usr/local/cuda-11.7/",
	)
	new_local_repository(
	name = "cublas",
	build_file = "@//third_party/cublas:BUILD",
	path = "/usr",
	)
	#############################################################################################################
	# Tarballs and fetched dependencies (default - use in cases when building from precompiled bin and tarballs)
	#############################################################################################################
	http_archive(
	name = "libtorch",
	build_file = "@//third_party/libtorch:BUILD",
	sha256 = "7c4b8754830fef23ec19c5eaf414794cee9597b435df055f5c1d0471d3e81568",
	strip_prefix = "libtorch",
	urls = ["https://download.pytorch.org/libtorch/nightly/cu117/libtorch-cxx11-abi-shared-with-deps-2.1.0.dev20230314%2Bcu117.zip"],
	)
	http_archive(
	name = "libtorch_pre_cxx11_abi",
	build_file = "@//third_party/libtorch:BUILD",
	sha256 = "f1e64a75dd12d0ba4c8c1f61947299e0a9c50684dff64f0cfbf355aa7a13e8cf",
	strip_prefix = "libtorch",
	urls = ["https://download.pytorch.org/libtorch/nightly/cu117/libtorch-shared-with-deps-2.1.0.dev20230314%2Bcu117.zip"],
	)

[johnnynunez]

We are working to switch our main branch over to TRT 8.6 now (PR #1852), and our upcoming release 1.4 will be based on TRT 8.6. Switching to CUDA 12 is not yet planned, since our current Torch distribution dependency does not have a CUDA 12-compatible build, but you can build Torch-TensorRT from scratch using your own Torch/CUDA versions by replacing the libtorch, libtorch_pre_cxx11_abi, and cuda sections in the WORKSPACE file, and rebuilding Torch-TensorRT from scratch:

TensorRT/WORKSPACE

Lines 40 to 70 in 1d78f43

	# CUDA should be installed on the system locally
	new_local_repository(
	name = "cuda",
	build_file = "@//third_party/cuda:BUILD",
	path = "/usr/local/cuda-11.7/",
	)
	new_local_repository(
	name = "cublas",
	build_file = "@//third_party/cublas:BUILD",
	path = "/usr",
	)
	#############################################################################################################
	# Tarballs and fetched dependencies (default - use in cases when building from precompiled bin and tarballs)
	#############################################################################################################
	http_archive(
	name = "libtorch",
	build_file = "@//third_party/libtorch:BUILD",
	sha256 = "7c4b8754830fef23ec19c5eaf414794cee9597b435df055f5c1d0471d3e81568",
	strip_prefix = "libtorch",
	urls = ["https://download.pytorch.org/libtorch/nightly/cu117/libtorch-cxx11-abi-shared-with-deps-2.1.0.dev20230314%2Bcu117.zip"],
	)
	http_archive(
	name = "libtorch_pre_cxx11_abi",
	build_file = "@//third_party/libtorch:BUILD",
	sha256 = "f1e64a75dd12d0ba4c8c1f61947299e0a9c50684dff64f0cfbf355aa7a13e8cf",
	strip_prefix = "libtorch",
	urls = ["https://download.pytorch.org/libtorch/nightly/cu117/libtorch-shared-with-deps-2.1.0.dev20230314%2Bcu117.zip"],
	)

Yes I have seen that they are working on CUDA 12 on the main branch. In fact they will remove cuda 11.7 to add CUDA 12.1.
pytorch/pytorch#98832
pytorch/pytorch#98986
pytorch/pytorch#98398

[narendasan]

We support TRT 8.6 so closing this issue