[TRTLLM-11715][infra] Fix in upgrade dependencies for dlfw 26.04 stack

README.md

@@ -8,8 +8,8 @@ TensorRT LLM
 [![Ask DeepWiki](https://deepwiki.com/badge.svg)](https://deepwiki.com/NVIDIA/TensorRT-LLM)
 [![python](https://img.shields.io/badge/python-3.12-green)](https://www.python.org/downloads/release/python-3123/)
 [![python](https://img.shields.io/badge/python-3.10-green)](https://www.python.org/downloads/release/python-31012/)
-[![cuda](https://img.shields.io/badge/cuda-13.1.1-green)](https://developer.nvidia.com/cuda-downloads)
-[![torch](https://img.shields.io/badge/torch-2.10.0-green)](https://pytorch.org)
+[![cuda](https://img.shields.io/badge/cuda-13.2.1-green)](https://developer.nvidia.com/cuda-downloads)
+[![torch](https://img.shields.io/badge/torch-2.11.0-green)](https://pytorch.org)
 [![version](https://img.shields.io/badge/release-1.3.0rc18-green)](https://github.com/NVIDIA/TensorRT-LLM/blob/main/tensorrt_llm/version.py)
 [![license](https://img.shields.io/badge/license-Apache%202-blue)](https://github.com/NVIDIA/TensorRT-LLM/blob/main/LICENSE)
 

constraints.txt

@@ -1,11 +1,5 @@
 # These vulnerabilities were inherited from the base image (pytorch:25.12-py3) and should be removed when the base image
 # is updated.
-# WAR against https://github.com/advisories/GHSA-8rrh-rw8j-w5fx
-wheel>=0.46.2
-# WAR against https://github.com/advisories/GHSA-qjxf-f2mg-c6mc
-tornado>=6.5.5
-# WAR against https://github.com/advisories/GHSA-3936-cmfr-pm3m
-black>=26.3.1
 # Upgrade base image nvidia-cutlass-dsl 4.3.5 to 4.4.2
 nvidia-cutlass-dsl>=4.4.2
 # The `nvidia-cutlass-dsl` package does not pin numpy at all, which can be problematic in certain CI

cpp/include/tensorrt_llm/runtime/virtualMemory.h

@@ -505,10 +505,7 @@ class CudaVirtualMemoryAllocator
             {
                 std::size_t gpuAlignment = 1;
                 CUmemAllocationProp const prop{CU_MEM_ALLOCATION_TYPE_PINNED, CU_MEM_HANDLE_TYPE_NONE,
-                    {
-                        CU_MEM_LOCATION_TYPE_DEVICE,
-                        device,
-                    }};
+                    CUmemLocation{CU_MEM_LOCATION_TYPE_DEVICE, {device}}};
                 TLLM_CU_CHECK(
                     cuMemGetAllocationGranularity(&gpuAlignment, &prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
                 alignment = std::lcm(getpagesize(), gpuAlignment);

cpp/tensorrt_llm/deep_ep/CMakeLists.txt

@@ -120,6 +120,15 @@ if(NOT CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
   set(CMAKE_C_COMPILER gcc)
   set(CMAKE_CXX_COMPILER g++)
   set(CMAKE_CUDA_HOST_COMPILER g++)
+  # PyTorch's cmake/public/cuda.cmake (loaded transitively by
+  # find_package(Torch)) appends -Xcompiler=-fclang-abi-compat=17 to
+  # CMAKE_CUDA_FLAGS whenever the parent build is configured with Clang>=18 (see
+  # pytorch PR #175233). Since this subdirectory falls back to GCC for NVSHMEM
+  # compatibility, that Clang-only flag would be forwarded to g++ via `nvcc
+  # -ccbin=g++` and abort the build with: g++: error: unrecognized command-line
+  # option '-fclang-abi-compat=17'
+  string(REGEX REPLACE "-Xcompiler=-fclang-abi-compat=[0-9]+" ""
+                       CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS}")
 endif()
 
 # Add nvshmem external project
@@ -204,6 +213,12 @@ target_compile_options(
 target_compile_definitions(
   deep_ep_cpp_tllm PRIVATE DISABLE_AGGRESSIVE_PTX_INSTRS
                            TORCH_EXTENSION_NAME=deep_ep_cpp_tllm)
+# Newer CUDA containers provide NVSHMEM headers in the default CUDA include
+# directory. DeepEP must compile against the vendored NVSHMEM headers because it
+# links the vendored NVSHMEM static library below.
+target_include_directories(
+  deep_ep_cpp_tllm BEFORE
+  PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/nvshmem-build/src/include)
 target_link_libraries(
   deep_ep_cpp_tllm PRIVATE nvshmem_project::nvshmem ${TORCH_LIBRARIES}
                            ${TORCH_PYTHON_LIB})

cpp/tensorrt_llm/flash_mla/CMakeLists.txt

@@ -44,6 +44,15 @@ if(CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|arm64" AND CMAKE_CXX_COMPILER_ID
   set(CMAKE_CUDA_HOST_COMPILER ${GCC_EXECUTABLE})
   message(
     STATUS "FlashMLA: Using GCC at ${GCC_EXECUTABLE} for CUDA compilation")
+  # PyTorch's cmake/public/cuda.cmake (loaded transitively by
+  # find_package(Torch)) appends -Xcompiler=-fclang-abi-compat=17 to
+  # CMAKE_CUDA_FLAGS whenever the parent build is configured with Clang>=18 (see
+  # pytorch PR #175233). Since CUDA host compilation here falls back to GCC,
+  # that Clang-only flag would be forwarded to g++ via `nvcc -ccbin=g++` and
+  # abort the build with: g++: error: unrecognized command-line option
+  # '-fclang-abi-compat=17'
+  string(REGEX REPLACE "-Xcompiler=-fclang-abi-compat=[0-9]+" ""
+                       CMAKE_CUDA_FLAGS "${CMAKE_CUDA_FLAGS}")
 endif()
 
 # Check CUDA version and architecture support

cpp/tensorrt_llm/runtime/virtualMemory.cpp

@@ -344,11 +344,7 @@ void CudaVirtualMemoryAllocator::allocate(Pointer* ptr, std::size_t n, int devic
 
     CUDAVirtualMemoryChunk::Configurators configurators;
     configurators.push_back(std::make_unique<UnicastConfigurator>(address, alignedSize,
-        CUmemAccessDesc{{
-                            CU_MEM_LOCATION_TYPE_DEVICE,
-                            device,
-                        },
-            CU_MEM_ACCESS_FLAGS_PROT_READWRITE}));
+        CUmemAccessDesc{CUmemLocation{CU_MEM_LOCATION_TYPE_DEVICE, {device}}, CU_MEM_ACCESS_FLAGS_PROT_READWRITE}));
 
     switch (mConfig->mMode)
     {
@@ -368,10 +364,7 @@ void CudaVirtualMemoryAllocator::allocate(Pointer* ptr, std::size_t n, int devic
 
     mConfig->mManager.add(address, mConfig->mTag,
         std::make_unique<LocalCreator<>>(CUmemAllocationProp{CU_MEM_ALLOCATION_TYPE_PINNED, CU_MEM_HANDLE_TYPE_NONE,
-                                             {
-                                                 CU_MEM_LOCATION_TYPE_DEVICE,
-                                                 device,
-                                             }},
+                                             CUmemLocation{CU_MEM_LOCATION_TYPE_DEVICE, {device}}},
             alignedSize),
         std::move(configurators));
 

cpp/tests/unit_tests/common/cudaDriverWrapperTest.cpp

@@ -31,7 +31,7 @@ TEST(TestCudaDriverWrapper, TllmCuCheckFailingWithValidParametersDoesNotThrow)
         CUmemAllocationHandleType::CU_MEM_HANDLE_TYPE_NONE,
         CUmemLocation{
             CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
-            0,
+            {0},
         },
         nullptr};
     auto const granularity = tensorrt_llm::common::getAllocationGranularity();
@@ -51,7 +51,7 @@ TEST(TestCudaDriverWrapper, TllmCuCheckFailingWithInvalidParametersThrows)
         CUmemAllocationHandleType::CU_MEM_HANDLE_TYPE_NONE,
         CUmemLocation{
             CUmemLocationType::CU_MEM_LOCATION_TYPE_DEVICE,
-            0,
+            {0},
         },
         nullptr};
     ASSERT_THROW(TLLM_CU_CHECK(cuMemCreate(&handle, -1, &prop, 0ULL)), tensorrt_llm::common::TllmException);

cpp/tests/unit_tests/runtime/virtualMemoryTest.cpp

@@ -142,19 +142,12 @@ TEST_F(VirtualMemoryTest, TestBasic)
 
     CUDAVirtualMemoryChunk::CreatorPtr creator
         = std::make_unique<LocalCreator<>>(CUmemAllocationProp{CU_MEM_ALLOCATION_TYPE_PINNED, CU_MEM_HANDLE_TYPE_NONE,
-                                               {
-                                                   CU_MEM_LOCATION_TYPE_DEVICE,
-                                                   0,
-                                               }},
+                                               CUmemLocation{CU_MEM_LOCATION_TYPE_DEVICE, {0}}},
             size);
 
     CUDAVirtualMemoryChunk::Configurators configurators;
     configurators.push_back(std::make_unique<UnicastConfigurator>(address, size,
-        CUmemAccessDesc{{
-                            CU_MEM_LOCATION_TYPE_DEVICE,
-                            0,
-                        },
-            CU_MEM_ACCESS_FLAGS_PROT_READWRITE}));
+        CUmemAccessDesc{CUmemLocation{CU_MEM_LOCATION_TYPE_DEVICE, {0}}, CU_MEM_ACCESS_FLAGS_PROT_READWRITE}));
 
     CUDAVirtualMemoryChunk vm(std::move(creator), std::move(configurators));
     ASSERT_EQ(vm.status(), CUDAVirtualMemoryChunk::RELEASED);
@@ -212,19 +205,12 @@ TEST_P(VirtualMemoryOffloadConfigurator, Test)
 
     CUDAVirtualMemoryChunk::CreatorPtr creator
         = std::make_unique<LocalCreator<>>(CUmemAllocationProp{CU_MEM_ALLOCATION_TYPE_PINNED, CU_MEM_HANDLE_TYPE_NONE,
-                                               {
-                                                   CU_MEM_LOCATION_TYPE_DEVICE,
-                                                   0,
-                                               }},
+                                               CUmemLocation{CU_MEM_LOCATION_TYPE_DEVICE, {0}}},
             size);
 
     CUDAVirtualMemoryChunk::Configurators configurators;
     configurators.push_back(std::make_unique<UnicastConfigurator>(address, size,
-        CUmemAccessDesc{{
-                            CU_MEM_LOCATION_TYPE_DEVICE,
-                            0,
-                        },
-            CU_MEM_ACCESS_FLAGS_PROT_READWRITE}));
+        CUmemAccessDesc{CUmemLocation{CU_MEM_LOCATION_TYPE_DEVICE, {0}}, CU_MEM_ACCESS_FLAGS_PROT_READWRITE}));
     configurators.push_back(std::make_unique<OffloadConfigurator>(address, size, backType, stream.get(), false));
 
     CUDAVirtualMemoryChunk vm(std::move(creator), std::move(configurators));
@@ -611,14 +597,14 @@ TEST_F(VirtualMemoryTest, TestFacilities)
     {
 
         // Create original CUDAVirtualMemoryChunk
-        CUDAVirtualMemoryChunk::CreatorPtr creator
-            = std::make_unique<LocalCreator<>>(CUmemAllocationProp{CU_MEM_ALLOCATION_TYPE_PINNED,
-                                                   CU_MEM_HANDLE_TYPE_NONE, {CU_MEM_LOCATION_TYPE_DEVICE, 0}},
-                size);
+        CUDAVirtualMemoryChunk::CreatorPtr creator = std::make_unique<LocalCreator<>>(
+            CUmemAllocationProp{CU_MEM_ALLOCATION_TYPE_PINNED, CU_MEM_HANDLE_TYPE_NONE,
+                CUmemLocation{CU_MEM_LOCATION_TYPE_DEVICE, {0}}},
+            size);
 
         CUDAVirtualMemoryChunk::Configurators configurators;
-        configurators.push_back(std::make_unique<UnicastConfigurator>(
-            address, size, CUmemAccessDesc{{CU_MEM_LOCATION_TYPE_DEVICE, 0}, CU_MEM_ACCESS_FLAGS_PROT_READWRITE}));
+        configurators.push_back(std::make_unique<UnicastConfigurator>(address, size,
+            CUmemAccessDesc{CUmemLocation{CU_MEM_LOCATION_TYPE_DEVICE, {0}}, CU_MEM_ACCESS_FLAGS_PROT_READWRITE}));
 
         CUDAVirtualMemoryChunk original(std::move(creator), std::move(configurators));
         original.materialize();
@@ -977,19 +963,12 @@ TEST_F(VirtualMemoryManagerTest, TestBasic)
 
     CUDAVirtualMemoryChunk::CreatorPtr creator
         = std::make_unique<LocalCreator<>>(CUmemAllocationProp{CU_MEM_ALLOCATION_TYPE_PINNED, CU_MEM_HANDLE_TYPE_NONE,
-                                               {
-                                                   CU_MEM_LOCATION_TYPE_DEVICE,
-                                                   0,
-                                               }},
+                                               CUmemLocation{CU_MEM_LOCATION_TYPE_DEVICE, {0}}},
             size);
 
     CUDAVirtualMemoryChunk::Configurators configurators;
     configurators.push_back(std::make_unique<UnicastConfigurator>(address, size,
-        CUmemAccessDesc{{
-                            CU_MEM_LOCATION_TYPE_DEVICE,
-                            0,
-                        },
-            CU_MEM_ACCESS_FLAGS_PROT_READWRITE}));
+        CUmemAccessDesc{CUmemLocation{CU_MEM_LOCATION_TYPE_DEVICE, {0}}, CU_MEM_ACCESS_FLAGS_PROT_READWRITE}));
 
     auto memoryBegin = getCurrentProcessMemoryInfo();
 

docker/Dockerfile.multi

@@ -1,8 +1,8 @@
 # Multi-stage Dockerfile
 ARG BASE_IMAGE=nvcr.io/nvidia/pytorch
 ARG TRITON_IMAGE=nvcr.io/nvidia/tritonserver
-ARG BASE_TAG=26.02-py3
-ARG TRITON_BASE_TAG=26.02-py3
+ARG BASE_TAG=26.04-py3
+ARG TRITON_BASE_TAG=26.04-py3
 ARG DEVEL_IMAGE=devel
 
 FROM ${BASE_IMAGE}:${BASE_TAG} AS base

docker/Makefile

@@ -202,21 +202,21 @@ jenkins-rockylinux8_%: PYTHON_VERSION_TAG_ID = $(if $(findstring 3.12,${PYTHON_V
 jenkins-rockylinux8_%: IMAGE_WITH_TAG = $(shell . ../jenkins/current_image_tags.properties && echo $$LLM_ROCKYLINUX8_${PYTHON_VERSION_TAG_ID}_DOCKER_IMAGE)
 jenkins-rockylinux8_%: STAGE = tritondevel
 jenkins-rockylinux8_%: BASE_IMAGE = nvcr.io/nvidia/cuda
-jenkins-rockylinux8_%: BASE_TAG = 13.1.1-devel-rockylinux8
+jenkins-rockylinux8_%: BASE_TAG = 13.2.1-devel-rockylinux8
 
 rockylinux8_%: STAGE = tritondevel
 rockylinux8_%: BASE_IMAGE = nvcr.io/nvidia/cuda
-rockylinux8_%: BASE_TAG = 13.1.1-devel-rockylinux8
+rockylinux8_%: BASE_TAG = 13.2.1-devel-rockylinux8
 
 # For x86_64
 ubuntu22_%: STAGE = tritondevel
 ubuntu22_%: BASE_IMAGE = nvcr.io/nvidia/cuda
-ubuntu22_%: BASE_TAG = 13.1.1-devel-ubuntu22.04
+ubuntu22_%: BASE_TAG = 13.2.1-devel-ubuntu22.04
 
 # For x86_64 and aarch64
 ubuntu24_%: STAGE = tritondevel
 ubuntu24_%: BASE_IMAGE = nvcr.io/nvidia/cuda
-ubuntu24_%: BASE_TAG = 13.1.1-devel-ubuntu24.04
+ubuntu24_%: BASE_TAG = 13.2.1-devel-ubuntu24.04
 
 trtllm_%: STAGE = release
 trtllm_%: PUSH_TO_STAGING := 0

docker/common/install_cuda_toolkit.sh

@@ -5,7 +5,7 @@ set -ex
 # This script is used for reinstalling CUDA on Rocky Linux 8 with the run file.
 # CUDA version is usually aligned with the latest NGC CUDA image tag.
 # Only use when public CUDA image is not ready.
-CUDA_VER="13.1.1_590.48.01"
+CUDA_VER="13.2.1_595.58.03"
 CUDA_VER_SHORT="${CUDA_VER%_*}"
 
 NVCC_VERSION_OUTPUT=$(nvcc --version)

docker/common/install_pytorch.sh

@@ -4,8 +4,8 @@ set -ex
 
 # Use latest stable version from https://pypi.org/project/torch/#history
 # and closest to the version specified in
-# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-02.html#rel-26-02
-TORCH_VERSION="2.10.0"
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-04.html#rel-26-04
+TORCH_VERSION="2.11.0"
 SYSTEM_ID=$(grep -oP '(?<=^ID=).+' /etc/os-release | tr -d '"')
 
 prepare_environment() {

docker/common/install_tensorrt.sh

@@ -2,20 +2,20 @@
 
 set -ex
 
-TRT_VER="10.15.1.29"
+TRT_VER="10.16.1.11"
 # Align with the pre-installed cuDNN / cuBLAS / NCCL versions from
-# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-02.html#rel-26-02
-CUDA_VER="13.1" # 13.1.1
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-04.html#rel-26-04
+CUDA_VER="13.2" # 13.2.1
 # Keep the installation for cuDNN if users want to install PyTorch with source codes.
 # PyTorch 2.x can compile with cuDNN v9.
-CUDNN_VER="9.19.0.56-1"
-NCCL_VER="2.29.2-1+cuda13.1"
-CUBLAS_VER="13.2.1.1-1"
+CUDNN_VER="9.21.0.82-1"
+NCCL_VER="2.29.7-1+cuda13.2"
+CUBLAS_VER="13.4.0.1-1"
 # Align with the pre-installed CUDA / NVCC / NVRTC versions from
 # https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
-NVRTC_VER="13.1.115-1"
-CUDA_RUNTIME="13.1.80-1"
-CUDA_DRIVER_VERSION="590.48.01-1.el8"
+NVRTC_VER="13.2.78-1"
+CUDA_RUNTIME="13.2.75-1"
+CUDA_DRIVER_VERSION="595.58.03-1.el8"
 
 for i in "$@"; do
     case $i in

docs/source/legacy/reference/support-matrix.md

@@ -158,9 +158,9 @@ The following table shows the supported software for TensorRT-LLM.
 * -
   - Software Compatibility
 * - Container
-  - [26.02](https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html)
+  - [26.04](https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html)
 * - TensorRT
-  - [10.14](https://docs.nvidia.com/deeplearning/tensorrt/release-notes/index.html)
+  - [10.16](https://docs.nvidia.com/deeplearning/tensorrt/release-notes/index.html)
 * - Precision
   -
     - Blackwell (SM100/SM103/SM120) - FP32, FP16, BF16, FP8, FP4, INT8, INT4

jenkins/Build.groovy

@@ -408,7 +408,7 @@ def runLLMBuild(pipeline, buildFlags, tarName, is_linux_x86_64)
     def llmPath = sh (script: "realpath ${LLM_ROOT}",returnStdout: true).trim()
     // TODO: Remove after the cmake version is upgraded to 3.31.8
     // Get triton tag from docker/dockerfile.multi
-    def tritonShortTag = "r26.02"
+    def tritonShortTag = "r26.04"
     sh "cd ${LLM_ROOT}/triton_backend/inflight_batcher_llm && mkdir build && cd build && cmake .. -DTRTLLM_DIR=${llmPath} -DTRITON_COMMON_REPO_TAG=${tritonShortTag} -DTRITON_CORE_REPO_TAG=${tritonShortTag} -DTRITON_THIRD_PARTY_REPO_TAG=${tritonShortTag} -DTRITON_BACKEND_REPO_TAG=${tritonShortTag} -DUSE_CXX11_ABI=ON && make -j${buildJobs} install"
 
     // Step 3: packaging wheels into tarfile