2025-08-11 nightly release (40f1a12)

Author: pytorchbot

.github/workflows/build-test-linux-aarch64-jetpack.yml

@@ -1,17 +1,16 @@
 name: Build and test Linux aarch64 wheels for Jetpack
 
 on:
-  # TODO: Uncomment this when we have a stable release
-  # pull_request:
-  # push:
-  #   branches:
-  #     - main
-  #     - nightly
-  #     - release/*
-  #   tags:
-  #     # NOTE: Binary build pipelines should only get triggered on release candidate builds
-  #     # Release candidate tags look like: v1.11.0-rc1
-  #     - v[0-9]+.[0-9]+.[0-9]+-rc[0-9]+
+  pull_request:
+  push:
+    branches:
+      - main
+      - nightly
+      - release/*
+    tags:
+      # NOTE: Binary build pipelines should only get triggered on release candidate builds
+      # Release candidate tags look like: v1.11.0-rc1
+      - v[0-9]+.[0-9]+.[0-9]+-rc[0-9]+
   workflow_dispatch:
 
 jobs:

.github/workflows/build_wheels_linux_aarch64.yml

@@ -264,7 +264,7 @@ jobs:
           if [[ ${{ inputs.is-jetpack }} == false ]]; then
             ${CONDA_RUN} python setup.py bdist_wheel
           else
-            ${CONDA_RUN} python setup.py bdist_wheel --jetpack --plat-name=linux_tegra_aarch64
+            ${CONDA_RUN} python setup.py bdist_wheel --jetpack
           fi
       - name: Repair Manylinux_2_28 Wheel
         shell: bash -l {0}
@@ -337,8 +337,8 @@ jobs:
     needs: build
     name: upload-wheel-${{ matrix.python_version }}-${{ matrix.desired_cuda }}-${{ matrix.gpu_arch_type }}-${{ inputs.is-jetpack }}
     uses: pytorch/test-infra/.github/workflows/_binary_upload.yml@main
-    # for jetpack builds, only upload to pytorch index for nightly builds
-    if: ${{ inputs.is-jetpack == false || (github.event_name == 'push' && startsWith(github.event.ref, 'refs/heads/nightly')) }}
+    # for jetpack builds, do not upload to pytorch nightly index, only upload to https://pypi.jetson-ai-lab.io/ manually for each release
+    if: ${{ inputs.is-jetpack == false }}
     with:
       repository: ${{ inputs.repository }}
       ref: ${{ inputs.ref }}

MODULE.bazel

@@ -90,10 +90,9 @@ http_archive(
 http_archive(
     name = "torch_l4t",
     build_file = "@//third_party/libtorch:BUILD",
-    sha256 = "6eff643c0a7acda92734cc798338f733ff35c7df1a4434576f5ff7c66fc97319",
     strip_prefix = "torch",
     type = "zip",
-    urls = ["https://pypi.jetson-ai-lab.dev/jp6/cu126/+f/6ef/f643c0a7acda9/torch-2.7.0-cp310-cp310-linux_aarch64.whl"],
+    urls = ["https://pypi.jetson-ai-lab.io/jp6/cu126/+f/62a/1beee9f2f1470/torch-2.8.0-cp310-cp310-linux_aarch64.whl"],
 )
 
 # Download these tarballs manually from the NVIDIA website

core/conversion/conversion.cpp

@@ -202,13 +202,7 @@ void AddInputs(ConversionCtx* ctx, c10::ArrayRef<const torch::jit::Value*> input
   TORCHTRT_CHECK(
       profile->isValid(),
       "Optimization profile is invalid, please check the input range provided (conversion.AddInputs)");
-
   ctx->cfg->addOptimizationProfile(profile);
-#if NV_TENSORRT_MAJOR > 7 || (NV_TENSORRT_MAJOR == 7 && NV_TENSORRT_MINOR >= 1)
-  if (ctx->enabled_precisions.find(nvinfer1::DataType::kINT8) != ctx->enabled_precisions.end()) {
-    ctx->cfg->setCalibrationProfile(profile);
-  }
-#endif
 }
 
 void MarkOutputs(ConversionCtx* ctx, at::ArrayRef<const torch::jit::Value*> outputs) {

core/conversion/conversionctx/ConversionCtx.cpp

@@ -31,8 +31,7 @@ std::ostream& operator<<(std::ostream& os, const BuilderSettings& s) {
     if (s.device.device_type == nvinfer1::DeviceType::kDLA) {
     os << "\n    DLACore: " << s.device.dla_core;
     }
-    os << "\n    Engine Capability: " << s.capability                                      \
-       << "\n    Calibrator Created: " << (s.calibrator != nullptr);
+    os << "\n    Engine Capability: " << s.capability;
     return os;
 }
 // clang-format on
@@ -64,15 +63,7 @@ ConversionCtx::ConversionCtx(BuilderSettings build_settings)
         cfg->setFlag(nvinfer1::BuilderFlag::kFP16);
         break;
       case nvinfer1::DataType::kINT8:
-        TORCHTRT_CHECK(
-            builder->platformHasFastInt8(), "Requested inference in INT8 but platform does not support INT8");
-        cfg->setFlag(nvinfer1::BuilderFlag::kINT8);
-        if (!settings.calibrator) {
-          LOG_INFO(
-              "Int8 precision has been enabled but no calibrator provided. This assumes the network has Q/DQ nodes obtained from Quantization aware training. For more details, refer to https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#work-with-qat-networks");
-        } else {
-          cfg->setInt8Calibrator(settings.calibrator);
-        }
+        LOG_DEBUG("INT8 precision has been enabled, we assume the network has Q/DQ nodes obtained from modelopt");
         break;
       case nvinfer1::DataType::kFLOAT:
         break;

core/conversion/conversionctx/ConversionCtx.h

@@ -26,7 +26,6 @@ struct BuilderSettings {
   bool allow_shape_tensors = false;
   ir::Device device;
   nvinfer1::EngineCapability capability = TRT_ENGINE_CAPABILITY_STANDARD;
-  nvinfer1::IInt8Calibrator* calibrator = nullptr;
   uint64_t num_avg_timing_iters = 1;
   uint64_t workspace_size = 0;
   uint64_t dla_sram_size = DLA_SRAM_SIZE;

cpp/BUILD

@@ -7,14 +7,12 @@ cc_library(
     srcs = [
         "src/compile_spec.cpp",
         "src/logging.cpp",
-        "src/ptq.cpp",
         "src/torch_tensorrt.cpp",
         "src/types.cpp",
     ],
     hdrs = [
         "include/torch_tensorrt/logging.h",
         "include/torch_tensorrt/macros.h",
-        "include/torch_tensorrt/ptq.h",
         "include/torch_tensorrt/torch_tensorrt.h",
     ],
     linkstatic = True,

cpp/CMakeLists.txt

@@ -4,15 +4,13 @@ add_library(${lib_name} OBJECT)
 set(CXX_SRCS
     "${CMAKE_CURRENT_SOURCE_DIR}/src/compile_spec.cpp"
     "${CMAKE_CURRENT_SOURCE_DIR}/src/logging.cpp"
-    "${CMAKE_CURRENT_SOURCE_DIR}/src/ptq.cpp"
     "${CMAKE_CURRENT_SOURCE_DIR}/src/torch_tensorrt.cpp"
     "${CMAKE_CURRENT_SOURCE_DIR}/src/types.cpp"
 )
 
 set(HEADER_FILES
     "${CMAKE_CURRENT_SOURCE_DIR}/include/torch_tensorrt/logging.h"
     "${CMAKE_CURRENT_SOURCE_DIR}/include/torch_tensorrt/macros.h"
-    "${CMAKE_CURRENT_SOURCE_DIR}/include/torch_tensorrt/ptq.h"
     "${CMAKE_CURRENT_SOURCE_DIR}/include/torch_tensorrt/torch_tensorrt.h"
 )
 

cpp/bin/torchtrtc/fileio.h

@@ -23,7 +23,6 @@
 #include "torch/torch.h"
 
 #include "torch_tensorrt/logging.h"
-#include "torch_tensorrt/ptq.h"
 #include "torch_tensorrt/torch_tensorrt.h"
 
 namespace torchtrtc {

cpp/bin/torchtrtc/main.cpp

@@ -7,7 +7,6 @@
 #include "torch/script.h"
 
 #include "torch_tensorrt/logging.h"
-#include "torch_tensorrt/ptq.h"
 #include "torch_tensorrt/torch_tensorrt.h"
 
 #include "accuracy.h"
@@ -335,8 +334,6 @@ int main(int argc, char** argv) {
     calibration_cache_file_path = torchtrtc::fileio::resolve_path(args::get(calibration_cache_file));
   }
 
-  auto calibrator = torchtrt::ptq::make_int8_cache_calibrator(calibration_cache_file_path);
-
   compile_settings.require_full_compilation = require_full_compilation;
 
   if (torch_executed_ops || torch_executed_mods) {
@@ -367,13 +364,9 @@ int main(int argc, char** argv) {
         compile_settings.enabled_precisions.insert(torch::kF16);
       } else if (dtype == torchtrt::DataType::kChar) {
         compile_settings.enabled_precisions.insert(torch::kI8);
-        if (calibration_cache_file) {
-          compile_settings.ptq_calibrator = calibrator;
-        } else {
-          torchtrt::logging::log(
-              torchtrt::logging::Level::kINFO,
-              "Int8 precision has been enabled but no calibrator provided. This assumes the network has Q/DQ nodes obtained from Quantization aware training. For more details, refer to https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/index.html#work-with-qat-networks");
-        }
+        torchtrt::logging::log(
+            torchtrt::logging::Level::kDEBUG,
+            "Int8 precision has been enabled which assumes the network has Q/DQ nodes obtained");
       } else {
         std::stringstream ss;
         ss << "Invalid precision given for enabled kernel precision, options are [ float | float32 | f32 | fp32 | half | float16 | f16 | fp16 | char | int8 | i8 ], found: ";