[pull] main from NVIDIA:main

.pre-commit-config.yaml

@@ -1447,7 +1447,7 @@ repos:
         additional_dependencies:
         - tomli
         # add ignore words list
-        args: ["-L", "Mor,ans,thirdparty,subtiles,PARD,pard,therefrom", "--skip", "ATTRIBUTIONS-*.md,*.svg", "--skip", "security_scanning/*", "--skip", "tensorrt_llm/_torch/visual_gen/jit_kernels/*"]
+        args: ["-L", "Mor,ans,thirdparty,subtiles,PARD,pard,indx,therefrom", "--skip", "ATTRIBUTIONS-*.md,*.svg", "--skip", "security_scanning/*", "--skip", "tensorrt_llm/_torch/visual_gen/jit_kernels/*"]
         exclude: 'scripts/attribution/data/cas/.*$'
 -   repo: https://github.com/astral-sh/ruff-pre-commit
     rev: v0.9.4

ATTRIBUTIONS-Python.md

@@ -62375,7 +62375,7 @@ Copyright 2018- The Hugging Face team. All rights reserved.
   - `Homepage`: https://github.com/huggingface/transformers
 
 
-## triton (3.5.1)
+## triton (3.6.0)
 
 ### Licenses
 License: `MIT License`
@@ -62413,7 +62413,7 @@ License: `MIT License`
   - `Homepage`: https://github.com/triton-lang/triton/
 
 
-## triton-kernels (3.5.1)
+## triton-kernels (3.6.0)
 
 ### Licenses
 License: `MIT License`
@@ -62444,7 +62444,7 @@ SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
 ```
 
 ### URLs
-  - `Source`: https://github.com/triton-lang/triton/tree/v3.5.1/python/triton_kernels
+  - `Source`: https://github.com/triton-lang/triton/tree/v3.6.0/python/triton_kernels
 
 
 ## tritonclient (2.63.0)

README.md

@@ -9,8 +9,8 @@ state-of-the-art optimizations to perform inference efficiently on NVIDIA GPUs.<
 [![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-13.1.0-green)](https://developer.nvidia.com/cuda-downloads)
-[![torch](https://img.shields.io/badge/torch-2.9.1-green)](https://pytorch.org)
+[![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)
 [![version](https://img.shields.io/badge/release-1.3.0rc9-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)
 

cpp/tensorrt_llm/thop/fp8Op.cpp

@@ -112,7 +112,7 @@ std::tuple<Tensor, Tensor> e4m3_quantize_helper(Tensor input, at::optional<Tenso
     if (scales.has_value())
     {
         // static quantization will use float scales by default.
-        scales_ = scales.value();
+        scales_ = scales.value().clone();
         CHECK_TH_CUDA(scales_);
         CHECK_TYPE(scales_, torch::kFloat32);
         e4m3_static_quantize(input, quantized_input, scales_, stream, quantize_mode);

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=25.12-py3
-ARG TRITON_BASE_TAG=25.12-py3
+ARG BASE_TAG=26.02-py3
+ARG TRITON_BASE_TAG=26.02-py3
 ARG DEVEL_IMAGE=devel
 
 FROM ${BASE_IMAGE}:${BASE_TAG} AS base
@@ -52,7 +52,7 @@ RUN --mount=type=bind,source=docker/common,target=/opt/docker/common \
 # Install constraints after install.sh so cleanup() doesn't delete the file mid-RUN
 COPY constraints.txt /tmp/constraints.txt
 RUN --mount=type=cache,target=/root/.cache/pip \
-    pip3 install --no-cache-dir -r /tmp/constraints.txt && \
+    pip3 install --ignore-installed --no-cache-dir -r /tmp/constraints.txt && \
     rm /tmp/constraints.txt && \
     pip3 uninstall -y nbconvert || true
 
@@ -67,7 +67,7 @@ RUN --mount=type=bind,source=docker/common,target=/opt/docker/common \
     # WAR against https://github.com/advisories/GHSA-58pv-8j8x-9vj2
     rm -rf /usr/local/lib/python3.12/dist-packages/setuptools/_vendor/jaraco.context-5.3.0.dist-info && \
     # WAR against https://github.com/advisories/GHSA-8rrh-rw8j-w5fx
-    rm -rf /usr/local/lib/python3.12/dist-packages/setuptools/_vendor/wheel-0.45.1.dist-info
+    rm -rf /usr/local/lib/python3.12/dist-packages/setuptools/_vendor/wheel-*.dist-info
 
 # Generate OSS attribution file for devel image
 ARG TRT_LLM_VER

docker/Makefile

@@ -202,16 +202,16 @@ 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 = 13.1.0-devel-rockylinux8
+jenkins-rockylinux8_%: BASE_TAG = 13.1.1-devel-rockylinux8
 
 rockylinux8_%: STAGE = tritondevel
 rockylinux8_%: BASE_IMAGE = nvcr.io/nvidia/cuda
-rockylinux8_%: BASE_TAG = 13.1.0-devel-rockylinux8
+rockylinux8_%: BASE_TAG = 13.1.1-devel-rockylinux8
 
 # For x86_64 and aarch64
 ubuntu22_%: STAGE = tritondevel
 ubuntu22_%: BASE_IMAGE = nvcr.io/nvidia/cuda
-ubuntu22_%: BASE_TAG = 13.1.0-devel-ubuntu22.04
+ubuntu22_%: BASE_TAG = 13.1.1-devel-ubuntu22.04
 
 trtllm_%: STAGE = release
 trtllm_%: PUSH_TO_STAGING := 0

docker/common/install_base.sh

@@ -38,9 +38,14 @@ set_bash_env() {
 cleanup() {
   # Clean up apt/dnf cache
   if [ -f /etc/debian_version ]; then
+    echo "Removing python3-pygments from Ubuntu..."
+    apt-get remove -y python3-pygments || true
+    apt-get autoremove -y || true
     apt-get clean
     rm -rf /var/lib/apt/lists/*
   elif [ -f /etc/redhat-release ]; then
+    echo "Removing python3-pygments from Rocky Linux..."
+    dnf remove -y python3-pygments || true
     dnf clean all
     rm -rf /var/cache/dnf
   fi

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="13.1.0_590.44.01"
+CUDA_VER="13.1.1_590.48.01"
 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-25-12.html#rel-25-12
-TORCH_VERSION="2.9.1"
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-02.html#rel-26-02
+TORCH_VERSION="2.10.0"
 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="10.14.1.48"
+TRT_VER="10.15.1.29"
 # Align with the pre-installed cuDNN / cuBLAS / NCCL versions from
-# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-25-12.html#rel-25-12
-CUDA_VER="13.1" # 13.1.0
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-02.html#rel-26-02
+CUDA_VER="13.1" # 13.1.1
 # Keep the installation for cuDNN if users want to install PyTorch with source codes.
 # PyTorch 2.x can compile with cuDNN v9.
-CUDNN_VER="9.17.0.29-1"
-NCCL_VER="2.28.9-1+cuda13.0"
-CUBLAS_VER="13.2.0.9-1"
+CUDNN_VER="9.19.0.56-1"
+NCCL_VER="2.29.2-1+cuda13.1"
+CUBLAS_VER="13.2.1.1-1"
 # Align with the pre-installed CUDA / NVCC / NVRTC versions from
 # https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
-NVRTC_VER="13.1.80-1"
+NVRTC_VER="13.1.115-1"
 CUDA_RUNTIME="13.1.80-1"
-CUDA_DRIVER_VERSION="590.44.01-1.el8"
+CUDA_DRIVER_VERSION="590.48.01-1.el8"
 
 for i in "$@"; do
     case $i in
@@ -120,10 +120,6 @@ install_tensorrt() {
     PARSED_PY_VERSION=$(echo "${PY_VERSION//./}")
 
     TRT_CUDA_VERSION=${CUDA_VER}
-    # No CUDA 13.1 version for TensorRT yet. Use CUDA 13.0 package instead.
-    if [ "$CUDA_VER" = "13.1" ]; then
-        TRT_CUDA_VERSION="13.0"
-    fi
     TRT_VER_SHORT=$(echo $TRT_VER | cut -d. -f1-3)
 
     if [ -z "$RELEASE_URL_TRT" ];then

docs/source/installation/build-from-source-linux.md

@@ -2,7 +2,7 @@
 
 # Building from Source Code on Linux
 
-This document provides instructions for building TensorRT LLM from source code on Linux. Building from source is recommended for achieving optimal performance, enabling debugging capabilities, or when you need a different [GNU CXX11 ABI](https://gcc.gnu.org/onlinedocs/libstdc++/manual/using_dual_abi.html) configuration than what is available in the pre-built TensorRT LLM wheel on PyPI. Note that the current pre-built TensorRT LLM wheel on PyPI is linked against PyTorch 2.9.1, which uses the new CXX11 ABI.
+This document provides instructions for building TensorRT LLM from source code on Linux. Building from source is recommended for achieving optimal performance, enabling debugging capabilities, or when you need a different [GNU CXX11 ABI](https://gcc.gnu.org/onlinedocs/libstdc++/manual/using_dual_abi.html) configuration than what is available in the pre-built TensorRT LLM wheel on PyPI. Note that the current pre-built TensorRT LLM wheel on PyPI is linked against PyTorch 2.10.0, which uses the new CXX11 ABI.
 
 
 ## Prerequisites

docs/source/installation/linux.md

@@ -17,7 +17,7 @@
 
    ```bash
    # By default, PyTorch CUDA 12.8 package is installed. Install PyTorch CUDA 13.0 package to align with the CUDA version used for building TensorRT LLM wheels.
-   pip3 install torch==2.9.1 torchvision --index-url https://download.pytorch.org/whl/cu130
+   pip3 install torch==2.10.0 torchvision --index-url https://download.pytorch.org/whl/cu130
 
    sudo apt-get -y install libopenmpi-dev
 
@@ -40,7 +40,7 @@
    pip3 install --ignore-installed pip setuptools wheel && pip3 install tensorrt_llm
    ```
 
-   > **Note:** The TensorRT LLM wheel on PyPI is built with PyTorch 2.9.1. This version may be incompatible with the NVIDIA NGC PyTorch 25.12 container, which uses a more recent PyTorch build from the main branch. If you are using this container or a similar environment, please install the pre-built wheel located at `/app/tensorrt_llm` inside the TensorRT LLM NGC Release container instead.
+   > **Note:** The TensorRT LLM wheel on PyPI is built with PyTorch 2.10.0. This version may be incompatible with the NVIDIA NGC PyTorch 25.12 container, which uses a more recent PyTorch build from the main branch. If you are using this container or a similar environment, please install the pre-built wheel located at `/app/tensorrt_llm` inside the TensorRT LLM NGC Release container instead.
 
    **This project will download and install additional third-party open source software projects. Review the license terms of these open source projects before use.**
 

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

@@ -154,7 +154,7 @@ The following table shows the supported software for TensorRT-LLM.
 * -
   - Software Compatibility
 * - Container
-  - [25.12](https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html)
+  - [26.02](https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html)
 * - TensorRT
   - [10.14](https://docs.nvidia.com/deeplearning/tensorrt/release-notes/index.html)
 * - Precision

jenkins/Build.groovy

@@ -438,7 +438,7 @@ 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 = "r25.12"
+    def tritonShortTag = "r26.02"
     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

jenkins/L0_Test.groovy

@@ -40,7 +40,7 @@ LLM_ROCKYLINUX8_PY310_DOCKER_IMAGE = env.wheelDockerImagePy310
 LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE = env.wheelDockerImagePy312
 
 // DLFW torch image
-DLFW_IMAGE = "urm.nvidia.com/docker/nvidia/pytorch:25.12-py3"
+DLFW_IMAGE = "urm.nvidia.com/docker/nvidia/pytorch:26.02-py3"
 
 //Ubuntu base image
 UBUNTU_22_04_IMAGE = "urm.nvidia.com/docker/ubuntu:22.04"
@@ -1146,6 +1146,7 @@ def runLLMTestlistWithSbatch(pipeline, platform, testList, config=VANILLA_CONFIG
 
                     // Output is the corresponding scriptLaunchPathLocal script under the disaggMode
                     sh """
+                        pip3 install pyyaml && \\
                         python3 ${scriptSubmitLocalPath} \\
                         --run-ci \\
                         --llm-src ${llmSrcLocal} \\
@@ -1957,7 +1958,7 @@ def launchTestListCheck(pipeline)
             def llmPath = sh (script: "realpath .", returnStdout: true).trim()
             def llmSrc = "${llmPath}/TensorRT-LLM/src"
             trtllm_utils.llmExecStepWithRetry(pipeline, script: "pip3 install -r ${llmSrc}/requirements-dev.txt")
-            sh "NVIDIA_TRITON_SERVER_VERSION=25.12 LLM_ROOT=${llmSrc} LLM_BACKEND_ROOT=${llmSrc}/triton_backend python3 ${llmSrc}/scripts/check_test_list.py --l0 --qa --waive"
+            sh "NVIDIA_TRITON_SERVER_VERSION=26.02 LLM_ROOT=${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) {
@@ -2967,17 +2968,19 @@ def runLLMBuild(pipeline, cpu_arch, reinstall_dependencies=false, wheel_path="",
     echo "uploading ${wheelName} to ${cpu_arch}/${wheel_path}"
     trtllm_utils.uploadArtifacts("tensorrt_llm/build/${wheelName}",  "${UPLOAD_PATH}/${cpu_arch}/${wheel_path}")
 
-    if (reinstall_dependencies == true) {
+    if (reinstall_dependencies) {
         // Test installation in the new environment
-        def pip_keep = "-e 'pip'"
+        // Reserve CUDA 13.0 torch and torchvision packages
+        def pip_keep = "^pip==|^torch==|^torchvision=="
         def remove_trt = "rm -rf /usr/local/tensorrt"
         if (env.alternativeTRT) {
-            pip_keep += " -e tensorrt"
+            pip_keep += "|^tensorrt=="
             remove_trt = "echo keep /usr/local/tensorrt"
         }
-        sh "#!/bin/bash \n" + "pip3 list --format=freeze | egrep -v ${pip_keep} | xargs pip3 uninstall -y"
-        sh "#!/bin/bash \n" + "yum remove -y libcudnn* libnccl* libcublas* && ${remove_trt}"
+        sh "bash -c 'pip3 list --format=freeze | grep -Ev \"${pip_keep}\" | xargs -r pip3 uninstall -y'"
+        sh "bash -c 'yum remove -y libcudnn* libnccl* libcublas* && ${remove_trt}'"
     }
+
     // Test preview installation
     trtllm_utils.llmExecStepWithRetry(pipeline, script: "#!/bin/bash \n" + "cd tensorrt_llm/ && pip3 install pytest build/tensorrt_llm-*.whl")
     if (env.alternativeTRT) {
@@ -3016,17 +3019,22 @@ def runPackageSanityCheck(pipeline, wheel_path, reinstall_dependencies=false, cp
         trtllm_utils.replaceWithAlternativeTRT(env.alternativeTRT, cpver)
         sh "bash -c 'pip3 show tensorrt || true'"
     }
+
     if (reinstall_dependencies) {
         // Test installation in the new environment
-        def pip_keep = "-e 'pip'"
+        // Reserve CUDA 13.0 torch and torchvision packages
+        def pip_keep = "^pip==|^torch==|^torchvision=="
         def remove_trt = "rm -rf /usr/local/tensorrt"
         if (env.alternativeTRT) {
-            pip_keep += " -e tensorrt"
+            pip_keep += "|^tensorrt=="
             remove_trt = "echo keep /usr/local/tensorrt"
         }
-        sh "bash -c 'pip3 list --format=freeze | egrep -v ${pip_keep} | xargs pip3 uninstall -y'"
+        sh "bash -c 'pip3 list --format=freeze | grep -Ev \"${pip_keep}\" | xargs -r pip3 uninstall -y'"
         sh "bash -c 'yum remove -y libcudnn* libnccl* libcublas* && ${remove_trt}'"
     }
+    //WAR: remove python3-pygments first since it is installed in NGC PyTorch image
+    trtllm_utils.llmExecStepWithRetry(pipeline, script: "apt-get remove -y python3-pygments")
+
     // Test preview installation
     trtllm_utils.llmExecStepWithRetry(pipeline, script: "bash -c 'pip3 install pytest tensorrt_llm-*.whl'")
     if (env.alternativeTRT) {
@@ -3477,7 +3485,7 @@ def launchTestJobs(pipeline, testFilter)
     // Python version and OS for sanity check
     x86SanityCheckConfigs = [
         "PY312-DLFW": [
-            LLM_DOCKER_IMAGE,  // Workaround ABI incompatibilities between PyTorch 2.9.1 and 2.10.0a0
+            LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE,  // Workaround ABI incompatibilities between PyTorch 2.9.1 and 2.10.0a0
             "B200_PCIe",
             X86_64_TRIPLE,
             false,
@@ -3515,8 +3523,8 @@ def launchTestJobs(pipeline, testFilter)
             AARCH64_TRIPLE,
             false,
             "",
-            DLFW_IMAGE,
-            false, // Extra PyTorch CUDA 13.0 install
+            UBUNTU_24_04_IMAGE,
+            true, // Extra PyTorch CUDA 13.0 install
         ],
         "PY312-DLFW": [
             LLM_DOCKER_IMAGE,
@@ -3604,6 +3612,8 @@ def launchTestJobs(pipeline, testFilter)
                         // Clean up the pip constraint file from the base NGC PyTorch image.
                         if (values[5] == DLFW_IMAGE) {
                             trtllm_utils.llmExecStepWithRetry(pipeline, script: "[ -f /etc/pip/constraint.txt ] && : > /etc/pip/constraint.txt || true")
+                            // Remove the python3-pygments pip package because the dlfw image already includes a Debian pygments package, which conflicts with the pip-installed version.
+                            trtllm_utils.llmExecStepWithRetry(pipeline, script: "apt-get remove -y python3-pygments")
                         }
                         trtllm_utils.llmExecStepWithRetry(pipeline, script: "apt-get update && apt-get install -y python3-pip git rsync curl wget")
                         trtllm_utils.checkoutSource(LLM_REPO, env.gitlabCommit, LLM_ROOT, false, true)
@@ -3622,11 +3632,7 @@ def launchTestJobs(pipeline, testFilter)
                             echo "###### Extra PyTorch CUDA 13.0 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.
-                            if (k8s_arch == "amd64") {
-                                trtllm_utils.llmExecStepWithRetry(pipeline, script: "pip3 install torch==2.9.1+cu130 torchvision==0.24.1+cu130 --extra-index-url https://urm.nvidia.com/artifactory/api/pypi/pytorch-cu128-remote/simple --extra-index-url https://download.pytorch.org/whl/cu130")
-                            } else {
-                                trtllm_utils.llmExecStepWithRetry(pipeline, script: "pip3 install torch==2.9.1+cu130 torchvision==0.24.1 --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==2.10.0+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")
                         }
 
                         def libEnv = []