chore!: Downgrade default TensorRT and cuDNN issue to address cuDNN

incompatablity with PyTorch 1.5.0.

TensorRT 7.1 and cuDNN 8 are still supported provided user is using a
version of PyTorch compiled with cuDNN 8 / cuBLAS 11

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>

Author: narendasan

README.md

@@ -75,8 +75,8 @@ torch.jit.save(trt_ts_module, "trt_torchscript_module.ts")
 - Bazel 3.3.1
 - Libtorch 1.5.1
 - CUDA 10.2
-- cuDNN 8.0.1
-- TensorRT 7.1.3
+- cuDNN 7.6.5 (by default, cuDNN 8 supported with compatable PyTorch build)
+- TensorRT 7.0.0 (by default, TensorRT 7.1 supported with compatable PyTorch build)
 
 ## Prebuilt Binaries and Wheel files
 

WORKSPACE

@@ -54,7 +54,7 @@ http_archive(
     build_file = "@//third_party/libtorch:BUILD",
     strip_prefix = "libtorch",
     urls = ["https://download.pytorch.org/libtorch/cu102/libtorch-cxx11-abi-shared-with-deps-1.5.1.zip"],
-    sha256 = "0efdd4e709ab11088fa75f0501c19b0e294404231442bab1d1fb953924feb6b5"
+    sha256 = "cf0691493d05062fe3239cf76773bae4c5124f4b039050dbdd291c652af3ab2a"
 )
 
 http_archive(
@@ -71,18 +71,18 @@ http_archive(
 
 http_archive(
     name = "cudnn",
-    urls = ["https://developer.nvidia.com/compute/machine-learning/cudnn/secure/8.0.1.13/10.2_20200626/cudnn-10.2-linux-x64-v8.0.1.13.tgz",],
+    urls = ["https://developer.nvidia.com/compute/machine-learning/cudnn/secure/7.6.5.32/Production/10.2_20191118/cudnn-10.2-linux-x64-v7.6.5.32.tgz",],
     build_file = "@//third_party/cudnn/archive:BUILD",
-    sha256 = "0c106ec84f199a0fbcf1199010166986da732f9b0907768c9ac5ea5b120772db",
+    sha256 = "600267f2caaed2fd58eb214ba669d8ea35f396a7d19b94822e6b36f9f7088c20",
     strip_prefix = "cuda"
 )
 
 http_archive(
     name = "tensorrt",
-    urls = ["https://developer.nvidia.com/compute/machine-learning/tensorrt/secure/7.1/tars/TensorRT-7.1.3.4.Ubuntu-18.04.x86_64-gnu.cuda-10.2.cudnn8.0.tar.gz",],
+    urls = ["https://developer.nvidia.com/compute/machine-learning/tensorrt/secure/7.0/7.0.0.11/tars/TensorRT-7.0.0.11.Ubuntu-18.04.x86_64-gnu.cuda-10.2.cudnn7.6.tar.gz",],
     build_file = "@//third_party/tensorrt/archive:BUILD",
-    sha256 = "9205bed204e2ae7aafd2e01cce0f21309e281e18d5bfd7172ef8541771539d41",
-    strip_prefix = "TensorRT-7.1.3.4"
+    sha256 = "c7d73b2585b18aae68b740249efa8c8ba5ae852abe9a023720595432a8eb4efd",
+    strip_prefix = "TensorRT-7.0.0.11"
 )
 
 ####################################################################################

docsrc/tutorials/installation.rst

@@ -26,13 +26,13 @@ You can install the python package using
 .. code-block:: sh
 
     # Python 3.5
-    pip3 install  https://github.com/NVIDIA/TRTorch/releases/download/v0.0.2/trtorch-0.0.2-cp35-cp35m-linux_x86_64.whl
+    pip3 install  https://github.com/NVIDIA/TRTorch/releases/download/v0.0.3/trtorch-0.0.3-cp35-cp35m-linux_x86_64.whl
     # Python 3.6
-    pip3 install  https://github.com/NVIDIA/TRTorch/releases/download/v0.0.2/trtorch-0.0.2-cp36-cp36m-linux_x86_64.whl
+    pip3 install  https://github.com/NVIDIA/TRTorch/releases/download/v0.0.3/trtorch-0.0.3-cp36-cp36m-linux_x86_64.whl
     # Python 3.7
-    pip3 install  https://github.com/NVIDIA/TRTorch/releases/download/v0.0.2/trtorch-0.0.2-cp37-cp37m-linux_x86_64.whl
+    pip3 install  https://github.com/NVIDIA/TRTorch/releases/download/v0.0.3/trtorch-0.0.3-cp37-cp37m-linux_x86_64.whl
     # Python 3.8
-    pip3 install  https://github.com/NVIDIA/TRTorch/releases/download/v0.0.2/trtorch-0.0.2-cp38-cp38-linux_x86_64.whl
+    pip3 install  https://github.com/NVIDIA/TRTorch/releases/download/v0.0.3/trtorch-0.0.3-cp38-cp38-linux_x86_64.whl
 
 .. _bin-dist:
 
@@ -73,6 +73,10 @@ the CUDA driver installed and the container must have CUDA)
 
 The correct LibTorch version will be pulled down for you by bazel.
 
+    NOTE: For best compatability with official PyTorch, use TensorRT 7.0 and cuDNN 7.6 however TRTorch itself supports
+    TensorRT 7.1 and cuDNN 8.0 for usecases such as using NVIDIA compiled distributions of PyTorch that use cuDNN 8
+    e.g. aarch64 or custom compiled version of PyTorch that use cuDNN 8.
+
 You then have two compilation options:
 
 .. _build-from-archive: