diff --git a/README.md b/README.md
index 1f9e7fa262b7..6f26d9665339 100644
--- a/README.md
+++ b/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)
 
diff --git a/constraints.txt b/constraints.txt
index 519a9a29f9fc..09a933566229 100644
--- a/constraints.txt
+++ b/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
diff --git a/cpp/include/tensorrt_llm/runtime/virtualMemory.h b/cpp/include/tensorrt_llm/runtime/virtualMemory.h
index a7e95b42d707..d74673f37e01 100644
--- a/cpp/include/tensorrt_llm/runtime/virtualMemory.h
+++ b/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);
diff --git a/cpp/tensorrt_llm/deep_ep/CMakeLists.txt b/cpp/tensorrt_llm/deep_ep/CMakeLists.txt
index 562c9e7d694c..e00815eed13e 100644
--- a/cpp/tensorrt_llm/deep_ep/CMakeLists.txt
+++ b/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})
diff --git a/cpp/tensorrt_llm/flash_mla/CMakeLists.txt b/cpp/tensorrt_llm/flash_mla/CMakeLists.txt
index e87f12275f60..024cee521d10 100644
--- a/cpp/tensorrt_llm/flash_mla/CMakeLists.txt
+++ b/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
diff --git a/cpp/tensorrt_llm/runtime/virtualMemory.cpp b/cpp/tensorrt_llm/runtime/virtualMemory.cpp
index 9b23866d6281..0d08012a29d8 100644
--- a/cpp/tensorrt_llm/runtime/virtualMemory.cpp
+++ b/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));
 
diff --git a/cpp/tests/unit_tests/common/cudaDriverWrapperTest.cpp b/cpp/tests/unit_tests/common/cudaDriverWrapperTest.cpp
index 2ba48cd9fb81..0c99fb955cd7 100644
--- a/cpp/tests/unit_tests/common/cudaDriverWrapperTest.cpp
+++ b/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);
diff --git a/cpp/tests/unit_tests/runtime/virtualMemoryTest.cpp b/cpp/tests/unit_tests/runtime/virtualMemoryTest.cpp
index 85d49433d882..f2045e7659d1 100644
--- a/cpp/tests/unit_tests/runtime/virtualMemoryTest.cpp
+++ b/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();
 
diff --git a/docker/Dockerfile.multi b/docker/Dockerfile.multi
index d4bf170470ac..34a7db7b8b5d 100644
--- a/docker/Dockerfile.multi
+++ b/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
diff --git a/docker/Makefile b/docker/Makefile
index 9a9c44ce8b56..0ca78eef80a0 100644
--- a/docker/Makefile
+++ b/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
diff --git a/docker/common/install_cuda_toolkit.sh b/docker/common/install_cuda_toolkit.sh
index 512a1043fb8e..41494e617c04 100644
--- a/docker/common/install_cuda_toolkit.sh
+++ b/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)
diff --git a/docker/common/install_pytorch.sh b/docker/common/install_pytorch.sh
index 5419bcc6de9b..4ec997fdbeca 100644
--- a/docker/common/install_pytorch.sh
+++ b/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() {
diff --git a/docker/common/install_tensorrt.sh b/docker/common/install_tensorrt.sh
index 9ae8be04cb21..66bf24deb2b4 100644
--- a/docker/common/install_tensorrt.sh
+++ b/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
diff --git a/docs/source/legacy/reference/support-matrix.md b/docs/source/legacy/reference/support-matrix.md
index 4c7629843299..e63f439b9446 100644
--- a/docs/source/legacy/reference/support-matrix.md
+++ b/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
diff --git a/jenkins/Build.groovy b/jenkins/Build.groovy
index 405280496af8..66c62c1b1580 100644
--- a/jenkins/Build.groovy
+++ b/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
diff --git a/jenkins/L0_Test.groovy b/jenkins/L0_Test.groovy
index fbdbd4df3091..0bfb7418a8a1 100644
--- a/jenkins/L0_Test.groovy
+++ b/jenkins/L0_Test.groovy
@@ -1,4 +1,4 @@
-@Library(['bloom-jenkins-shared-lib@main', 'trtllm-jenkins-shared-lib@main']) _
+@Library(['bloom-jenkins-shared-lib@emma/update_nsc_login_node', 'trtllm-jenkins-shared-lib@main']) _
 
 import java.lang.InterruptedException
 import groovy.transform.Field
@@ -46,7 +46,7 @@ LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE = env.wheelDockerImagePy312
 LLM_WHEEL_DOCKER_IMAGE = env.wheelDockerImage
 
 // DLFW torch image
-DLFW_IMAGE = "urm.nvidia.com/docker/nvidia/pytorch:26.02-py3"
+DLFW_IMAGE = "urm.nvidia.com/docker/nvidia/pytorch:26.04-py3"
 
 //Ubuntu base image
 UBUNTU_22_04_IMAGE = "urm.nvidia.com/docker/ubuntu:22.04"
@@ -1293,7 +1293,7 @@ def runLLMTestlistWithSbatch(pipeline, platform, testList, config=VANILLA_CONFIG
                     """
                 } else {
                     if(nodeCount > 1) {
-                        srunArgs.add("--mpi=pmi2")
+                        srunArgs.add("--mpi=pmix")
                     }
 
                     def scriptContent = """
@@ -2293,7 +2293,7 @@ def launchTestListCheck(pipeline)
             def llmPath = sh (script: "realpath .", returnStdout: true).trim()
             def llmSrc = "${llmPath}/TensorRT-LLM/src"
             trtllm_utils.llmExecStepWithRetry(pipeline, script: "pip3 install -r ${llmSrc}/requirements-dev.txt")
-            sh "NVIDIA_TRITON_SERVER_VERSION=26.02 LLM_ROOT=${llmSrc} LLM_BACKEND_ROOT=${llmSrc}/triton_backend python3 ${llmSrc}/scripts/check_test_list.py --l0 --qa --waive"
+            sh "NVIDIA_TRITON_SERVER_VERSION=26.04 LLM_ROOT=${llmSrc} LLM_BACKEND_ROOT=${llmSrc}/triton_backend python3 ${llmSrc}/scripts/check_test_list.py --l0 --qa --waive"
         } catch (InterruptedException e) {
             throw e
         } catch (Exception e) {
@@ -4119,7 +4119,9 @@ def launchTestJobs(pipeline, testFilter)
         "DGX_B200-8_GPUs-PyTorch-2": ["auto:dgx-b200-flex", "l0_dgx_b200", 2, 3, 8, 1, true],
         "DGX_B200-8_GPUs-PyTorch-3": ["auto:dgx-b200-flex", "l0_dgx_b200", 3, 3, 8, 1, true],
         "DGX_B200-8_GPUs-AutoDeploy-Post-Merge-1": ["auto:dgx-b200-flex", "l0_dgx_b200", 1, 1, 8, 1, true],
-        "DGX_B200-4_GPUs-Verl-Post-Merge-1": ["auto:dgx-b200-flex", "l0_verl", 1, 1, 4, 1, true],
+        // Disable Verl stage due to https://nvbugs/6236818.
+        // Please re-enable it after the bug is fixed.
+        // "DGX_B200-4_GPUs-Verl-Post-Merge-1": ["auto:dgx-b200-flex", "l0_verl", 1, 1, 4, 1, true],
         "B300-PyTorch-1": ["auto:dgx-b300-flex", "l0_b300", 1, 2, 1, 1, true],
         "B300-PyTorch-2": ["auto:dgx-b300-flex", "l0_b300", 2, 2, 1, 1, true],
         "DGX_B300-4_GPUs-PyTorch-1": ["auto:dgx-b300-flex", "l0_dgx_b300", 1, 1, 4, 1, true],
@@ -4590,14 +4592,14 @@ def launchTestJobs(pipeline, testFilter)
                             def platform = cpu_arch == X86_64_TRIPLE ? "x86_64" : "sbsa"
                             trtllm_utils.llmExecStepWithRetry(pipeline, script: "wget https://developer.download.nvidia.com/compute/cuda/repos/${ubuntu_version}/${platform}/cuda-keyring_1.1-1_all.deb")
                             trtllm_utils.llmExecStepWithRetry(pipeline, script: "dpkg -i cuda-keyring_1.1-1_all.deb")
-                            trtllm_utils.llmExecStepWithRetry(pipeline, script: "apt-get update && apt-get install -y cuda-toolkit-13-1")
+                            trtllm_utils.llmExecStepWithRetry(pipeline, script: "apt-get update && apt-get install -y cuda-toolkit-13-2")
                         }
-                        // Extra PyTorch CUDA 13.0 install for all bare-metal environments (Default PyTorch is for CUDA 12.8)
+                        // Extra PyTorch CUDA 13.2 install for all bare-metal environments (Default PyTorch is for CUDA 12.8)
                         if (values[6]) {
-                            echo "###### Extra PyTorch CUDA 13.0 install Start ######"
+                            echo "###### Extra PyTorch CUDA 13.2 install Start ######"
                             // Use internal mirror instead of https://download.pytorch.org/whl/cu130 for better network stability.
                             // PyTorch CUDA 13.0 package and torchvision package can be installed as expected.
-                            trtllm_utils.llmExecStepWithRetry(pipeline, script: "pip3 install torch==2.10.0+cu130 torchvision==0.25.0+cu130 --extra-index-url https://urm.nvidia.com/artifactory/api/pypi/pytorch-cu128-remote/simple --extra-index-url https://download.pytorch.org/whl/cu130")
+                            trtllm_utils.llmExecStepWithRetry(pipeline, script: "pip3 install torch==2.11.0+cu130 torchvision==0.26.0+cu130 --extra-index-url https://urm.nvidia.com/artifactory/api/pypi/pytorch-cu128-remote/simple --extra-index-url https://download.pytorch.org/whl/cu130")
                         }
 
                         def libEnv = []
diff --git a/jenkins/current_image_tags.properties b/jenkins/current_image_tags.properties
index e220c19ae3bd..b2a4cd876b2b 100644
--- a/jenkins/current_image_tags.properties
+++ b/jenkins/current_image_tags.properties
@@ -13,8 +13,8 @@
 #     images are adopted from PostMerge pipelines, the abbreviated commit hash is used instead.
 IMAGE_NAME=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm
 
-LLM_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-26.02-py3-x86_64-ubuntu24.04-trt10.15.1.29-skip-tritondevel-202606051544-14972
-LLM_SBSA_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-26.02-py3-sbsa-ubuntu24.04-trt10.15.1.29-skip-tritondevel-202606051544-14972
-LLM_ROCKYLINUX8_PY310_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.1.0-devel-rocky8-x86_64-rocky8-py310-trt10.15.1.29-skip-tritondevel-202606051544-14972
-LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.1.0-devel-rocky8-x86_64-rocky8-py312-trt10.15.1.29-skip-tritondevel-202606051544-14972
-LLM_SBSA_WHEEL_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.1.0-devel-ubuntu24.04-sbsa-ubuntu24.04-py312-trt10.15.1.29-skip-tritondevel-202606051544-14972
+LLM_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-26.04-py3-x86_64-ubuntu24.04-trt10.16.1.11-skip-tritondevel-202606091350-12643
+LLM_SBSA_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:pytorch-26.04-py3-aarch64-ubuntu24.04-trt10.16.1.11-skip-tritondevel-202606091350-12643
+LLM_ROCKYLINUX8_PY310_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.2.1-devel-rocky8-x86_64-rocky8-py310-trt10.16.1.11-skip-tritondevel-202606091350-12643
+LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.2.1-devel-rocky8-x86_64-rocky8-py312-trt10.16.1.11-skip-tritondevel-202606091350-12643
+LLM_SBSA_WHEEL_DOCKER_IMAGE=urm.nvidia.com/sw-tensorrt-docker/tensorrt-llm:cuda-13.2.1-devel-ubuntu24.04-sbsa-ubuntu24.04-py312-trt10.16.1.11-skip-tritondevel-202606091350-12643
diff --git a/jenkins/scripts/perf/README.md b/jenkins/scripts/perf/README.md
index 5e68643ec06d..8c2f89679615 100644
--- a/jenkins/scripts/perf/README.md
+++ b/jenkins/scripts/perf/README.md
@@ -270,9 +270,16 @@ and benchmark srun steps never see MPI flags.
 **Where MPI is configured:**
 - `jenkins/scripts/perf/disaggregated/slurm_launch_draft.sh` — `--mpi=pmix` on
   ctx/gen srun commands only
-- `jenkins/scripts/perf/local/submit.py` — `--mpi=pmi2` for aggregated mode only,
+- `jenkins/scripts/perf/local/submit.py` — `--mpi=pmix` for aggregated mode only,
   no MPI flag for disaggregated mode (handled by the draft template)
-- `jenkins/L0_Test.groovy` — `--mpi=pmi2` for non-disagg multi-node only
+- `jenkins/L0_Test.groovy` — `--mpi=pmix` for non-disagg multi-node only
+
+> The DLFW base image (PyTorch 26.03+) ships an Open MPI build with PMIx support
+> only and **no classic PMI-2**. Under `--mpi=pmi2` srun exposes a PMI-2 server
+> but no PMIx server, so OMPI's `pmix3x_client` fails with
+> `OPAL ERROR: Unreachable in pmix3x_client.c at line 111` and aborts in
+> `MPI_Init_thread`. `--mpi=pmix` matches what the disagg path has always used
+> and works against the new OMPI build.
 
 ### Key Rules
 
diff --git a/jenkins/scripts/perf/aggregated/slurm_launch_draft.sh b/jenkins/scripts/perf/aggregated/slurm_launch_draft.sh
index 0421439d8236..e8daf893f87b 100644
--- a/jenkins/scripts/perf/aggregated/slurm_launch_draft.sh
+++ b/jenkins/scripts/perf/aggregated/slurm_launch_draft.sh
@@ -12,7 +12,7 @@ chmod +x $runScript
 # Run aggregated test
 echo "Starting aggregated test..."
 world_size=${world_size:-$((totalNodes * gpusPerNodePerServer))}
-if ! srun "${srunArgs[@]}" --mpi=pmi2 --kill-on-bad-exit=1 \
+if ! srun "${srunArgs[@]}" --mpi=pmix --kill-on-bad-exit=1 \
     -N $totalNodes \
     --ntasks=$world_size \
     --ntasks-per-node=$gpusPerNodePerServer \
diff --git a/jenkins/scripts/perf/local/slurm_run.sh b/jenkins/scripts/perf/local/slurm_run.sh
index d10dc6fc23a8..a874b4fb965d 100755
--- a/jenkins/scripts/perf/local/slurm_run.sh
+++ b/jenkins/scripts/perf/local/slurm_run.sh
@@ -15,6 +15,15 @@ fi
 
 cd $llmSrcNode/tests/integration/defs
 
+# Force PMIx to use the in-memory hash GDS instead of ds12/ds21 shared-memory.
+# Under `srun --mpi=pmix` with the DLFW 26.04 OpenMPI build, the shared-memory
+# GDS modes can fail to publish UCX worker addresses across nodes, producing:
+#   pml_ucx.c:178  Error: Failed to receive UCX worker address: Not found (-13)
+#   pml_ucx.c:482  Error: Failed to resolve UCX endpoint for rank N
+# See https://github.com/open-mpi/ompi/issues/6981. Setting this is a no-op
+# when PMIx isn't used.
+export PMIX_MCA_gds=hash
+
 # Turn off "exit on error" so the following lines always run
 set +e
 
diff --git a/jenkins/scripts/perf/local/submit.py b/jenkins/scripts/perf/local/submit.py
index 97527eb9a8c4..fbb57182e810 100755
--- a/jenkins/scripts/perf/local/submit.py
+++ b/jenkins/scripts/perf/local/submit.py
@@ -368,7 +368,7 @@ def generate_srun_args(args, runtime_mode, timestamp, llm_src="", hardware_confi
 
     lines.append("--container-env=NVIDIA_IMEX_CHANNELS")
 
-    # Single-GPU aggregated jobs run one process without MPI -- drop --mpi=pmi2
+    # Single-GPU aggregated jobs run one process without MPI -- drop --mpi=pmix
     is_single_gpu_aggr = (
         is_aggr and hardware_config is not None and hardware_config.get("total_gpus") == 1
     )
@@ -376,7 +376,7 @@ def generate_srun_args(args, runtime_mode, timestamp, llm_src="", hardware_confi
     if args.mpi_type:
         lines.append(f"--mpi={args.mpi_type}")
     elif is_aggr and not is_single_gpu_aggr:
-        lines.append("--mpi=pmi2")
+        lines.append("--mpi=pmix")
 
     return lines
 
@@ -561,7 +561,7 @@ def main():
         "--mpi-type",
         default="",
         help="MPI type for srun (e.g. pmix, pmi2). If not set, aggregated runs default to"
-        " --mpi=pmi2; non-aggregated runs omit --mpi entirely.",
+        " --mpi=pmix; non-aggregated runs omit --mpi entirely.",
     )
     parser.add_argument(
         "--disagg-server-port",
diff --git a/jenkins/scripts/slurm_run.sh b/jenkins/scripts/slurm_run.sh
index eea9733d2d2d..87b932285229 100755
--- a/jenkins/scripts/slurm_run.sh
+++ b/jenkins/scripts/slurm_run.sh
@@ -88,6 +88,15 @@ if [ "${UCX_TLS:-}" = "tcp" ]; then
     unset UCX_TLS
     echo "Unset UCX_TLS (cluster injected UCX_TLS=tcp)"
 fi
+
+# Force PMIx to use the in-memory hash GDS instead of ds12/ds21 shared-memory.
+# Under `srun --mpi=pmix` with the DLFW 26.04 OpenMPI build, the shared-memory
+# GDS modes can fail to publish UCX worker addresses across nodes, producing:
+#   pml_ucx.c:178  Error: Failed to receive UCX worker address: Not found (-13)
+#   pml_ucx.c:482  Error: Failed to resolve UCX endpoint for rank N
+# See https://github.com/open-mpi/ompi/issues/6981. Setting this is a no-op
+# when PMIx isn't used.
+export PMIX_MCA_gds=hash
 echo "Library Path:"
 echo "$LD_LIBRARY_PATH"
 env | sort
diff --git a/requirements.txt b/requirements.txt
index 9fe0fbf59814..e8d7cab33435 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -21,14 +21,14 @@ pandas
 h5py==3.12.1
 StrEnum
 sentencepiece>=0.1.99
-tensorrt~=10.15.1
-# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-02.html#rel-26-02 uses 2.11.0a0.
-torch>=2.10.0,<=2.11.0a0
+tensorrt~=10.16.1
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-04.html#rel-26-04 uses 2.12.0a0.
+torch>=2.11.0,<=2.12.0a0
 torchvision
 nvidia-modelopt[torch]~=0.37.0
-# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-02.html#rel-26-02 uses 2.29.2
-# torch 2.10.0+cu130 depends on nvidia-nccl-cu13==2.28.9
-nvidia-nccl-cu13>=2.28.9,<=2.29.2
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-04.html#rel-26-04 uses 2.29.7
+# torch 2.11.0+cu130 depends on nvidia-nccl-cu13==2.28.9
+nvidia-nccl-cu13>=2.28.9,<=2.29.7
 nvidia-cuda-nvrtc
 transformers==5.5.4
 prometheus_client
@@ -89,3 +89,5 @@ etcd-sdk-python==0.0.7
 python-multipart
 smg-grpc-proto>=0.4.2
 cache-dit>=1.3.5
+librosa
+msgpack
diff --git a/scripts/rename_docker_images.py b/scripts/rename_docker_images.py
old mode 100755
new mode 100644
index dacbcb350ea3..a38983398ec4
--- a/scripts/rename_docker_images.py
+++ b/scripts/rename_docker_images.py
@@ -23,10 +23,6 @@
 def parse_arguments() -> _ap.Namespace:
     parser = _ap.ArgumentParser(
         description="Rename Docker images based on the given instructions.")
-    parser.add_argument(
-        'src_branch',
-        type=str,
-        help="The name of the source branch releasing the Docker image.")
     parser.add_argument(
         'src_build_id',
         type=int,
@@ -52,6 +48,22 @@ def parse_arguments() -> _ap.Namespace:
         help=
         "The new stage part of the destination image name (default: tritondevel)."
     )
+    source_group = parser.add_mutually_exclusive_group()
+    source_group.add_argument(
+        "--commit",
+        type=str,
+        default=None,
+        help=
+        "GitHub mode: short commit SHA embedded in the source image tag (default: `git rev-parse --short HEAD`)."
+    )
+    source_group.add_argument(
+        "--src-branch",
+        dest="src_branch",
+        type=str,
+        default=None,
+        help="GitLab mode: source branch name embedded in the source image tag. "
+        "When set, the tag is built as `<prefix>-<src_branch>-<src_build_id>` instead of the GitHub-PR layout."
+    )
     return parser.parse_args()
 
 
@@ -60,6 +72,14 @@ def get_current_timestamp() -> str:
     return _dt.datetime.now(_dt.timezone.utc).strftime("%Y%m%d%H%M")
 
 
+def get_current_commit() -> str:
+    """Return the 7-char short SHA of the current git HEAD."""
+    return _sp.run(["git", "rev-parse", "--short=7", "HEAD"],
+                   check=True,
+                   text=True,
+                   capture_output=True).stdout.strip()[:7]
+
+
 def run_shell_command(command: str, dry_run: bool) -> None:
     """Run a shell command and display its output.
 
@@ -178,29 +198,37 @@ def find_and_replace_in_files(directory, file_extension: str,
 
 
 def rename_images(*,
-                  src_branch: str,
                   src_build_id: int,
                   dst_mr: int,
                   stage: str,
+                  commit: str | None = None,
+                  src_branch: str | None = None,
                   timestamp: str | None = None,
                   dry_run: bool = False) -> None:
-    print(
-        f"Renaming images for branch {src_branch} and build id {src_build_id} to {dst_mr}"
-    )
+    if src_branch is not None:
+        src_branch_sanitized = src_branch.replace("/", "_")
+        src_suffix = f"{src_branch_sanitized}-{src_build_id}"
+        print(
+            f"Renaming images (GitLab mode) from branch {src_branch} build id {src_build_id} to MR {dst_mr}"
+        )
+    else:
+        commit = (commit or get_current_commit())[:7]
+        src_suffix = f"{commit}-github-pr-{dst_mr}-{src_build_id}"
+        print(
+            f"Renaming images (GitHub mode) at commit {commit} build id {src_build_id} to MR {dst_mr}"
+        )
     if dry_run:
         print("Dry-run mode enabled. No actual changes will be made.")
     else:
         print("Renaming images...")
 
     timestamp = timestamp or get_current_timestamp()
-    src_branch_sanitized = src_branch.replace("/", "_")
     base_dir = find_script_directory().parent
     current_tags_path = base_dir / "jenkins" / CURRENT_TAG_FILE
 
     for dst_key, src_pattern in IMAGE_MAPPING.items():
         print(f"Processing {dst_key} ...")
-        src_image = f"{src_pattern}-{src_branch_sanitized}-{src_build_id}".replace(
-            "__stage__", stage)
+        src_image = f"{src_pattern}-{src_suffix}".replace("__stage__", stage)
         dst_image_old = extract_line_after_prefix(current_tags_path,
                                                   dst_key + "=").strip('"')
         dst_image = replace_text_between_dashes(
diff --git a/tensorrt_llm/_torch/compilation/backend.py b/tensorrt_llm/_torch/compilation/backend.py
index b952236e3916..43319ac9425c 100644
--- a/tensorrt_llm/_torch/compilation/backend.py
+++ b/tensorrt_llm/_torch/compilation/backend.py
@@ -5,7 +5,7 @@
 import torch
 import torch._inductor.config as inductor_config
 from torch._functorch.aot_autograd import aot_module_simplified
-from torch._inductor.compile_fx import compile_fx, select_decomp_table
+from torch._inductor.compile_fx import compile_fx_inner, select_decomp_table
 from torch._inductor.pattern_matcher import PatternMatcherPass
 from torch._subclasses import FakeTensor
 from torch.fx import GraphModule
@@ -153,7 +153,7 @@ def optimize(
             self.generate_events(num_events)
             return gm
         elif self.enable_inductor:
-            return compile_fx(gm, example_inputs)
+            return compile_fx_inner(gm, example_inputs)
         else:
             return gm
 
diff --git a/tensorrt_llm/_torch/compilation/piecewise_optimizer.py b/tensorrt_llm/_torch/compilation/piecewise_optimizer.py
index d4172201eb13..ceaa657c8fb3 100644
--- a/tensorrt_llm/_torch/compilation/piecewise_optimizer.py
+++ b/tensorrt_llm/_torch/compilation/piecewise_optimizer.py
@@ -4,7 +4,7 @@
 
 import torch
 from torch._guards import detect_fake_mode
-from torch._inductor.compile_fx import compile_fx
+from torch._inductor.compile_fx import compile_fx, compile_fx_inner
 from torch._subclasses import FakeTensor
 from torch.fx import GraphModule, Interpreter
 from torch.fx.passes.split_module import split_module
@@ -96,7 +96,8 @@ def call_module(self, target, args, kwargs):
                 runtime_num_tokens_idx,
                 self.capture_num_tokens,
                 self.graph_pool_handle,
-                compile_fx(submod, args) if self.enable_inductor else submod,
+                compile_fx_inner(submod, args)
+                if self.enable_inductor else submod,
                 self.enable_inductor,
                 self.piecewise_runner_idx == 0,
                 self.piecewise_runner_idx == self.piecewise_runner_num - 1,
diff --git a/tests/integration/defs/verl/verl_config.yml b/tests/integration/defs/verl/verl_config.yml
index db43920203de..286c59428670 100644
--- a/tests/integration/defs/verl/verl_config.yml
+++ b/tests/integration/defs/verl/verl_config.yml
@@ -15,8 +15,8 @@ verl_config:
     - "pip install nvidia-nvshmem-cu13==3.3.20"
     # Create nvshmem symlink (needed before DeepEP build)
     - >-
-      (cd /usr/local/lib/python3.12/dist-packages/nvidia/nvshmem/lib &&
-      ln -s libnvshmem_host.so.3 libnvshmem_host.so)
+      (cd /usr/local/cuda/targets/x86_64-linux/lib &&
+      [ -e libnvshmem_host.so ] || ln -s libnvshmem_host.so.3 libnvshmem_host.so)
     # Install DeepEP
     - >-
       git clone -b hybrid-ep https://github.com/deepseek-ai/DeepEP.git &&
@@ -36,7 +36,7 @@ verl_config:
 
   # The environment variables to expose in the container before setting up
   env_vars:
-    - "NVSHMEM_DIR=/usr/local/lib/python3.12/dist-packages/nvidia/nvshmem"
+    - "NVSHMEM_DIR=/usr/local/cuda/targets/x86_64-linux"
     - "LD_LIBRARY_PATH=\"${NVSHMEM_DIR}/lib:$LD_LIBRARY_PATH\""
     - "PATH=\"${NVSHMEM_DIR}/bin:$PATH\""
     - "TRTLLM_TEST_MODEL_PATH_ROOT=/tmp/verl-models"
diff --git a/tests/unittest/auto_deploy/singlegpu/custom_ops/moe/test_trtllm_moe.py b/tests/unittest/auto_deploy/singlegpu/custom_ops/moe/test_trtllm_moe.py
index bfb79a4d3e37..82651348ea0a 100644
--- a/tests/unittest/auto_deploy/singlegpu/custom_ops/moe/test_trtllm_moe.py
+++ b/tests/unittest/auto_deploy/singlegpu/custom_ops/moe/test_trtllm_moe.py
@@ -574,7 +574,7 @@ def break_fp4_bytes(a, dtype):
 
 # Scale the input and weights to prevent large absolute values.
 FP4_X_GEN_SCALE = 0.5
-FP4_W_GEN_SCALE = 0.1
+FP4_W_GEN_SCALE = 0.75
 
 
 @pytest.mark.parametrize("batch_size", BATCH_SIZES)