diff --git a/docker/Dockerfile.multi b/docker/Dockerfile.multi
index a6f9fd57185f..5ee9a46eba5f 100644
--- a/docker/Dockerfile.multi
+++ b/docker/Dockerfile.multi
@@ -15,8 +15,8 @@
 # Multi-stage Dockerfile
 ARG BASE_IMAGE=nvcr.io/nvidia/pytorch
 ARG TRITON_IMAGE=nvcr.io/nvidia/tritonserver
-ARG BASE_TAG=26.08-py3
-ARG TRITON_BASE_TAG=26.08-py3
+ARG BASE_TAG=26.09-py3
+ARG TRITON_BASE_TAG=26.09-py3
 ARG DEVEL_IMAGE=devel
 
 FROM ${BASE_IMAGE}:${BASE_TAG} AS base
diff --git a/docker/Makefile b/docker/Makefile
index 4c1042168b92..3278733b3b40 100644
--- a/docker/Makefile
+++ b/docker/Makefile
@@ -208,21 +208,21 @@ 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 = 13.3.1-devel-rockylinux8
+jenkins-rockylinux8_%: BASE_TAG = 13.4.1-devel-rockylinux8
 
 rockylinux8_%: STAGE = tritondevel
 rockylinux8_%: BASE_IMAGE = nvcr.io/nvidia/cuda
-rockylinux8_%: BASE_TAG = 13.3.1-devel-rockylinux8
+rockylinux8_%: BASE_TAG = 13.4.1-devel-rockylinux8
 
 # For x86_64
 ubuntu22_%: STAGE = tritondevel
 ubuntu22_%: BASE_IMAGE = nvcr.io/nvidia/cuda
-ubuntu22_%: BASE_TAG = 13.3.1-devel-ubuntu22.04
+ubuntu22_%: BASE_TAG = 13.4.1-devel-ubuntu22.04
 
 # For x86_64 and aarch64
 ubuntu24_%: STAGE = tritondevel
 ubuntu24_%: BASE_IMAGE = nvcr.io/nvidia/cuda
-ubuntu24_%: BASE_TAG = 13.3.1-devel-ubuntu24.04
+ubuntu24_%: BASE_TAG = 13.4.1-devel-ubuntu24.04
 
 trtllm_%: STAGE = release
 trtllm_%: PUSH_TO_STAGING := 0
diff --git a/docker/common/install_cuda_libs.sh b/docker/common/install_cuda_libs.sh
index d076d6e583cb..8b0231f3d85e 100644
--- a/docker/common/install_cuda_libs.sh
+++ b/docker/common/install_cuda_libs.sh
@@ -12,9 +12,9 @@ CUDA_VER="13.4" # the NGC PyTorch image's CUDA_VERSION, major.minor
 # included, then ends up on a cuDNN anyone can install, and the tests validate that combination
 # rather than one only NGC can reproduce. The version guard below will not match on NGC PyTorch,
 # so its cuDNN is deliberately purged and reinstalled at this version.
-CUDNN_VER="9.25.1.1-1"
-NCCL_VER="2.30.7-1+cuda13.3"
-CUBLAS_VER="13.7.0.27-1"
+CUDNN_VER="9.26.0.51-1"
+NCCL_VER="2.31.2-1+cuda13.4"
+CUBLAS_VER="13.8.0.4-1"
 # Align with the pre-installed CUDA / NVCC / NVRTC versions from
 # https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
 NVRTC_VER="13.4.59-1"
diff --git a/docker/common/install_mooncake.sh b/docker/common/install_mooncake.sh
index 0935c7f31a2f..5bc4686e1ad9 100644
--- a/docker/common/install_mooncake.sh
+++ b/docker/common/install_mooncake.sh
@@ -30,7 +30,7 @@ apt-get install -y --no-install-recommends \
 
 mkdir -p /third-party-source
 
-git clone --depth 1 https://github.com/alibaba/yalantinglibs.git
+git clone --depth 1 -b 0.5.5 https://github.com/alibaba/yalantinglibs.git
 tar -czf /third-party-source/yalantinglibs.tar.gz yalantinglibs
 cd yalantinglibs
 mkdir build && cd build
@@ -53,3 +53,14 @@ cd ../..
 rm -rf Mooncake
 
 echo "export LD_LIBRARY_PATH=${MOONCAKE_INSTALL_PATH}/lib:\$LD_LIBRARY_PATH" >> "${ENV}"
+
+# `make install` also emits a `mooncake` package that omits
+# libmooncake_store.so, so importing mooncake.store from it fails. It has to go
+# before the wheel is installed: CMake writes
+# store.cpython-312-x86_64-linux-gnu.so where the wheel writes store.so, and
+# importlib prefers the interpreter-tagged suffix, so the broken extension
+# would win even after pip reports success. The directory is the one
+# mooncake-integration/CMakeLists.txt chose, which this repeats.
+MOONCAKE_CMAKE_PACKAGE="$(python3 -c "import sys; print([s for s in sys.path if 'packages' in s][0])")/mooncake"
+echo "removing CMake-generated mooncake package: ${MOONCAKE_CMAKE_PACKAGE}"
+rm -rf "${MOONCAKE_CMAKE_PACKAGE}"
diff --git a/docker/common/install_pytorch.sh b/docker/common/install_pytorch.sh
index 87e95de810d6..184d0c94cba2 100644
--- a/docker/common/install_pytorch.sh
+++ b/docker/common/install_pytorch.sh
@@ -12,7 +12,7 @@ fi
 
 # 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-26-08.html#rel-26-08
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-09.html#rel-26-09
 TORCH_VERSION="2.14.0"
 SYSTEM_ID=$(grep -oP '(?<=^ID=).+' /etc/os-release | tr -d '"')
 
diff --git a/jenkins/L0_Test.groovy b/jenkins/L0_Test.groovy
index edeabb07295c..eb35d3083ded 100644
--- a/jenkins/L0_Test.groovy
+++ b/jenkins/L0_Test.groovy
@@ -71,7 +71,7 @@ ARTIFACTORY_DOCKER_HOST = "artifactory.nvidia.com"
 ARTIFACTORY_CREDENTIALS_ID = "trtllm-artifactory-credentials"
 
 // DLFW torch image
-DLFW_IMAGE = "urm.nvidia.com/docker/nvidia/pytorch:26.08-py3"
+DLFW_IMAGE = "urm.nvidia.com/docker/nvidia/pytorch:26.09-py3"
 
 MODEL_EXPRESS_VERSION = "0.5.1"
 MODEL_EXPRESS_SERVER_IMAGE = "urm.nvidia.com/docker/nvidia/ai-dynamo/modelexpress-server:${MODEL_EXPRESS_VERSION}"
@@ -4145,7 +4145,7 @@ def launchTestListCheck(pipeline)
             trtllm_utils.llmExecStepWithRetry(pipeline, script: "pip3 install -r ${llmSrc}/requirements-dev.txt")
             // --validate --parity: after --l0/--qa generate the collectable lists, assert every
             // statically-verified parametrize ID is actually collectable (validate<->collection parity).
-            sh "NVIDIA_TRITON_SERVER_VERSION=26.08 LLM_ROOT=${llmSrc} LLM_BACKEND_ROOT=${llmSrc}/triton_backend python3 ${llmSrc}/scripts/check_test_list.py --l0 --qa --waive --validate --parity"
+            sh "NVIDIA_TRITON_SERVER_VERSION=26.09 LLM_ROOT=${llmSrc} LLM_BACKEND_ROOT=${llmSrc}/triton_backend python3 ${llmSrc}/scripts/check_test_list.py --l0 --qa --waive --validate --parity"
         } catch (InterruptedException e) {
             throw e
         } catch (Exception e) {
diff --git a/jenkins/current_image_tags.properties b/jenkins/current_image_tags.properties
index b615eb04d28a..b43224170147 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=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm
 
-LLM_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:pytorch-26.08-py3-x86_64-ubuntu24.04-skip-tritondevel-202609210610-19345
-LLM_SBSA_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:pytorch-26.08-py3-sbsa-ubuntu24.04-skip-tritondevel-202609210610-19345
-LLM_ROCKYLINUX8_PY310_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.4.1-devel-rocky8-x86_64-rocky8-py310-skip-tritondevel-202609210610-19345
-LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.4.1-devel-rocky8-x86_64-rocky8-py312-skip-tritondevel-202609210610-19345
-LLM_SBSA_WHEEL_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.4.1-devel-ubuntu24.04-sbsa-ubuntu24.04-py312-skip-tritondevel-202609210610-19345
+LLM_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:pytorch-26.09-py3-x86_64-ubuntu24.04-skip-tritondevel-202610050217-19679
+LLM_SBSA_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:pytorch-26.09-py3-sbsa-ubuntu24.04-skip-tritondevel-202610050217-19679
+LLM_ROCKYLINUX8_PY310_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.4.1-devel-rocky8-x86_64-rocky8-py310-skip-tritondevel-202610050217-19679
+LLM_ROCKYLINUX8_PY312_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.4.1-devel-rocky8-x86_64-rocky8-py312-skip-tritondevel-202610050217-19679
+LLM_SBSA_WHEEL_DOCKER_IMAGE=artifactory.nvidia.com/sw-tensorrt-llm-docker-local/tensorrt-llm:cuda-13.4.1-devel-ubuntu24.04-sbsa-ubuntu24.04-py312-skip-tritondevel-202610050217-19679
diff --git a/requirements.txt b/requirements.txt
index 63c1972c7c35..811de826fb74 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -23,17 +23,17 @@ pandas
 h5py==3.12.1
 StrEnum
 sentencepiece>=0.1.99
-# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-08.html#rel-26-08 uses 2.14.0a0.
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-09.html#rel-26-09 uses 2.14.0a0.
 # The floor is the NGC alpha, which PEP 440 orders below the public release of the same
 # series; the ceiling is that public release, which is what the resolver picks off the public
 # index. Ordered comparisons drop the local version, so the container's +nv build stays
 # admissible against the floor.
 torch>=2.14.0a0,<=2.14.0
 torchvision
-# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-08.html#rel-26-08 uses 2.30.7
-# The public torch above pins the same version, so an exact pin covers both it and the NGC
-# container's own NCCL; as with triton, carrying no local version keeps the +nv build admissible.
-nvidia-nccl-cu13==2.30.7
+# The floor is the exact pin of the public torch above; the ceiling is the NCCL that
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-09.html#rel-26-09
+# ships, so the resolver lands on the container's own NCCL version there instead of an older one.
+nvidia-nccl-cu13>=2.30.7,<=2.31.2
 # NCCL-EP uses nccl4py's nccl.ep package. The nccl-extensions wheel is built
 # from pinned source by CMake so it links against the container's NCCL.
 nccl4py>=0.4,<=0.5
@@ -79,7 +79,7 @@ meson
 ninja
 blake3
 soundfile
-# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-08.html#rel-26-08
+# https://docs.nvidia.com/deeplearning/frameworks/pytorch-release-notes/rel-26-09.html#rel-26-09
 # ships 3.8.0 and the torch pin above requires the same series, so floor and ceiling coincide.
 # The pin carries no local version, which keeps the container's own +nv build admissible -- it
 # must not be reinstalled underneath the rest of the stack that was built against it. This is
diff --git a/tests/unittest/_torch/thop/parallel/test_locality_domain_utils.py b/tests/unittest/_torch/thop/parallel/test_locality_domain_utils.py
index a5b7b5480104..1658ee7c198c 100644
--- a/tests/unittest/_torch/thop/parallel/test_locality_domain_utils.py
+++ b/tests/unittest/_torch/thop/parallel/test_locality_domain_utils.py
@@ -68,6 +68,17 @@ def check_locality_domain_support():
         pytest.skip("LOCALITY_DOMAIN localization is not supported on this system")
 
 
+@pytest.fixture(scope="module")
+def check_locality_domain_enabled(check_locality_domain_support):
+    """Skip unless locality_domain_device() actually switches the current domain.
+
+    Driver support alone is not enough: the feature is only enabled on SM107, and
+    elsewhere locality_domain_device() is a no-op.
+    """
+    if not is_locality_domain_enabled():
+        pytest.skip("LOCALITY_DOMAIN localization is not enabled on this system")
+
+
 class TestLocalityDomainSupport:
     """Tests for LOCALITY_DOMAIN support detection."""
 
@@ -666,7 +677,7 @@ def _pool_allocated_bytes(pool):
 class TestLocalityDomainNestedMemPool:
     """Nested optional_locality_domain_mem_pool scopes across locality domains."""
 
-    def test_nested_same_domain_does_not_reenter(self, check_locality_domain_support):
+    def test_nested_same_domain_does_not_reenter(self, check_locality_domain_enabled):
         """Re-entering the same pool is skipped; use_mem_pool would reject it."""
         manager = locality_domain_utils.get_locality_domain_resource_manager()
         device_id = torch.cuda.current_device()
@@ -684,7 +695,7 @@ def test_nested_same_domain_does_not_reenter(self, check_locality_domain_support
 
         assert getattr(manager.in_mem_pool_context, "active_pool_key", None) is None
 
-    def test_nested_other_domain_restores_outer_key(self, check_locality_domain_support):
+    def test_nested_other_domain_restores_outer_key(self, check_locality_domain_enabled):
         """A nested scope for a different domain enters it and unwinds to the outer one."""
         manager = locality_domain_utils.get_locality_domain_resource_manager()
         device_id = torch.cuda.current_device()
@@ -704,7 +715,7 @@ def test_nested_other_domain_restores_outer_key(self, check_locality_domain_supp
 
         assert getattr(manager.in_mem_pool_context, "active_pool_key", None) is None
 
-    def test_nested_other_domain_allocates_from_inner_pool(self, check_locality_domain_support):
+    def test_nested_other_domain_allocates_from_inner_pool(self, check_locality_domain_enabled):
         """A nested domain's tensor comes from its own pool, not the enclosing one."""
         try:
             pool0 = get_locality_domain_mempool(0)
@@ -728,7 +739,7 @@ def test_nested_other_domain_allocates_from_inner_pool(self, check_locality_doma
         assert after0 == before0, "nested domain 1 allocation leaked into domain 0's pool"
         del inner
 
-    def test_exception_restores_previous_key(self, check_locality_domain_support):
+    def test_exception_restores_previous_key(self, check_locality_domain_enabled):
         """An exception inside a nested scope must still unwind the key."""
         manager = locality_domain_utils.get_locality_domain_resource_manager()
         device_id = torch.cuda.current_device()