mc-nv · Sep 2, 2026 · Aug 12, 2026 · Sep 1, 2026 · Aug 20, 2026
diff --git a/Dockerfile.sdk b/Dockerfile.sdk
@@ -29,7 +29,7 @@
 #
 
 # Base image on the minimum Triton container
-ARG BASE_IMAGE=nvcr.io/nvidia/tritonserver:26.07-py3-min
+ARG BASE_IMAGE=nvcr.io/nvidia/tritonserver:26.08-py3-min
 
 ARG TRITON_CLIENT_REPO_SUBDIR=clientrepo
 ARG TRITON_REPO_ORGANIZATION=http://github.com/triton-inference-server
@@ -41,7 +41,7 @@ ARG TRITON_ENABLE_GPU=ON
 ARG JAVA_BINDINGS_JAVACPP_PRESETS_TAG=1.5.8
 ARG JAVA_BINDINGS_MAVEN_URL=https://archive.apache.org/dist/maven/maven-3/3.9.16/binaries/apache-maven-3.9.16-bin.tar.gz
 # DCGM version to install for Model Analyzer
-ARG DCGM_VERSION=4.5.3-1
+ARG DCGM_VERSION=4.6.1-1
 
 ARG NVIDIA_TRITON_SERVER_SDK_VERSION=unknown
 ARG NVIDIA_BUILD_ID=unknown

diff --git a/README.md b/README.md
@@ -29,8 +29,8 @@
 
 >[!WARNING]
 >You are currently on the `main` branch which tracks under-development progress
->towards the next release. The current release is version [2.71.0](https://github.com/triton-inference-server/server/releases/latest)
->and corresponds to the 26.07 container release on NVIDIA GPU Cloud (NGC).
+>towards the next release. The current release is version [2.72.0](https://github.com/triton-inference-server/server/releases/latest)
+>and corresponds to the 26.08 container release on NVIDIA GPU Cloud (NGC).
 
 # Triton Inference Server
 
@@ -90,16 +90,16 @@ Inference Server with the
 
 ```bash
 # Step 1: Create the example model repository
-git clone -b r26.07 https://github.com/triton-inference-server/server.git
+git clone -b r26.08 https://github.com/triton-inference-server/server.git
 cd server/docs/examples
 ./fetch_models.sh
 
 # Step 2: Launch triton from the NGC Triton container
-docker run --gpus=1 --rm --net=host -v ${PWD}/model_repository:/models nvcr.io/nvidia/tritonserver:26.07-py3 tritonserver --model-repository=/models --model-control-mode explicit --load-model densenet_onnx
+docker run --gpus=1 --rm --net=host -v ${PWD}/model_repository:/models nvcr.io/nvidia/tritonserver:26.08-py3 tritonserver --model-repository=/models --model-control-mode explicit --load-model densenet_onnx
 
 # Step 3: Sending an Inference Request
 # In a separate console, launch the image_client example from the NGC Triton SDK container
-docker run -it --rm --net=host nvcr.io/nvidia/tritonserver:26.07-py3-sdk /workspace/install/bin/image_client -m densenet_onnx -c 3 -s INCEPTION /workspace/images/mug.jpg
+docker run -it --rm --net=host nvcr.io/nvidia/tritonserver:26.08-py3-sdk /workspace/install/bin/image_client -m densenet_onnx -c 3 -s INCEPTION /workspace/images/mug.jpg
 
 # Inference should return the following
 Image '/workspace/images/mug.jpg':

diff --git a/build.py b/build.py
@@ -74,12 +74,11 @@
 DEFAULT_TRITON_VERSION_MAP = {
     "release_version": "2.73.0dev",
     "triton_container_version": "26.09dev",
-    "upstream_container_version": "26.07",
-    "ort_version": "1.27.0",
-    "ort_openvino_version": "2026.2.0",
-    "standalone_openvino_version": "2026.2.0",
-    "dcgm_version": "4.5.3-1",
-    "rhel_py_version": "3.12.3",
+    "upstream_container_version": "26.08",
+    "ort_version": "1.28.0",
+    "ort_openvino_version": "2026.3.0",
+    "standalone_openvino_version": "2026.3.0",
+    "dcgm_version": "4.6.1-1",
 }
 
 CORE_BACKENDS = ["ensemble"]
@@ -554,8 +553,6 @@ def backend_cmake_args(images, components, be, install_dir, library_paths):
         args = onnxruntime_cmake_args(images, library_paths)
     elif be == "openvino":
         args = openvino_cmake_args()
-    elif be == "python":
-        args = python_cmake_args()
     elif be == "dali":
         args = dali_cmake_args()
     elif be == "pytorch":
@@ -614,18 +611,6 @@ def backend_cmake_args(images, components, be, install_dir, library_paths):
     return cargs
 
 
-def python_cmake_args():
-    cargs = []
-    if target_platform() == "rhel":
-        cargs.append(
-            cmake_backend_arg(
-                "python", "PYBIND11_PYTHON_VERSION", "STRING", FLAGS.rhel_py_version
-            )
-        )
-
-    return cargs
-
-
 def pytorch_cmake_args(images):
     if "pytorch" in images:
         image = images["pytorch"]
@@ -1041,9 +1026,6 @@ def create_dockerfile_buildbase_rhel(ddir, dockerfile_name, argmap):
             numactl-devel \\
             openssl-devel \\
             pkg-config \\
-            python3-pip \\
-            python3-scons \\
-            python3-setuptools \\
             rapidjson-devel \\
             re2-devel \\
             readline-devel \\
@@ -1077,20 +1059,21 @@ def create_dockerfile_buildbase_rhel(ddir, dockerfile_name, argmap):
 """.format(
             os.getenv("CCACHE_REMOTE_STORAGE")
         )
-    # Requires openssl-devel to be installed first for pyenv build to be successful
-    df += change_default_python_version_rhel(FLAGS.rhel_py_version)
+    df += set_python_location_to_manylinux_for_rhel()
+    df += restore_embeddable_python_lib_rhel()
     df += """
 
 RUN pip3 install --upgrade pip \\
       && pip3 install --upgrade \\
           auditwheel \\
           build \\
-          wheel \\
-          setuptools \\
+          cmake==4.0.3 \\
           docker \\
-          virtualenv \\
           patchelf==0.17.2 \\
-          cmake==4.0.3
+          scons \\
+          setuptools \\
+          virtualenv \\
+          wheel
 """
     df += f"""
 # Install boost version >= 1.78 for boost::span
@@ -1455,9 +1438,10 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach
         libb64-devel \\
         gperftools-devel \\
         wget \\
-        python3.12-pip \\
         numactl-devel
-
+"""
+        df += set_python_location_to_manylinux_for_rhel()
+        df += """
 RUN pip3 install patchelf==0.17.2
 
 """
@@ -1532,8 +1516,7 @@ def dockerfile_prepare_container_linux(argmap, backends, enable_gpu, target_mach
         openssl-devel \\
         readline-devel
 """
-            # Requires openssl-devel to be installed first for pyenv build to be successful
-            df += change_default_python_version_rhel(FLAGS.rhel_py_version)
+            df += set_python_location_to_manylinux_for_rhel()
             df += """
 RUN pip3 install --upgrade pip \\
     && pip3 install --upgrade \\
@@ -1674,30 +1657,30 @@ def add_cpu_libs_to_linux_dockerfile(backends, target_machine):
     return df
 
 
-def change_default_python_version_rhel(version):
-    df = f"""
-# The python library version available for install via 'yum install python3.X-devel' does not
-# match the version of python inside the RHEL base container. This means that python packages
-# installed within the container will not be picked up by the python backend stub process pybind
-# bindings. It must instead must be installed via pyenv.
-ENV PYENV_ROOT=/opt/pyenv_build
-RUN curl https://pyenv.run | bash
-ENV PATH="${{PYENV_ROOT}}/bin:$PATH"
-RUN eval "$(pyenv init -)"
-RUN CONFIGURE_OPTS=\"--with-openssl=/usr/lib64\" && pyenv install {version} \\
-    && cp ${{PYENV_ROOT}}/versions/{version}/lib/libpython3* /usr/lib64/
-
-# RHEL image has several python versions. It's important
-# to set the correct version, otherwise, packages that are
-# pip installed will not be found during testing.
-ENV PYVER={version} PYTHONPATH=/opt/python/v
-RUN ln -sf ${{PYENV_ROOT}}/versions/${{PYVER}}* ${{PYTHONPATH}}
-ENV PYBIN=${{PYTHONPATH}}/bin
-ENV PYTHON_BIN_PATH=${{PYBIN}}/python${{PYVER}} PATH=${{PYBIN}}:${{PATH}}
+def set_python_location_to_manylinux_for_rhel():
+    df = """
+ENV PATH="/opt/_internal/pipx/shared/bin:$PATH"
 """
     return df
 
 
+def restore_embeddable_python_lib_rhel():
+    # The manylinux base container configures CPython with --disable-shared,
+    # so no libpython*.so is ever built, and its finalize.sh then compresses
+    # every libpython*.a into static-libs-for-embedding-only.tar.xz and drops
+    # the originals. triton_python_backend_stub links pybind11::embed and so
+    # needs that archive. Restore it in place -- the path it unpacks to is
+    # exactly the LIBDIR the interpreter's sysconfig reports, which is where
+    # pybind11's FindPythonLibsNew looks.
+    #
+    # This deliberately fails the build if the archive is absent: without it
+    # the stub link fails much later with an unresolvable -lpython<ver>.
+    return """
+RUN tar -xvf /opt/_internal/static-libs-for-embedding-only.tar.xz \\
+        -C /opt/_internal
+"""
+
+
 def create_build_dockerfiles(
     container_build_dir, images, backends, repoagents, caches, endpoints
 ):
@@ -2131,17 +2114,6 @@ def backend_build(
     cmake_script.mkdir(os.path.join(install_dir, "backends"))
     cmake_script.rmdir(os.path.join(install_dir, "backends", be))
 
-    # The python library version available for install via 'yum install python3.X-devel' does not
-    # match the version of python inside the RHEL base container. This means that python packages
-    # installed within the container will not be picked up by the python backend stub process pybind
-    # bindings. It must instead must be installed via pyenv. We package it here for better usability.
-    if target_platform() == "rhel" and be == "python":
-        major_minor_version = ".".join((FLAGS.rhel_py_version).split(".")[:2])
-        version_matched_files = "/usr/lib64/libpython" + major_minor_version + "*"
-        cmake_script.cp(
-            version_matched_files, os.path.join(repo_install_dir, "backends", be)
-        )
-
     cmake_script.cpdir(
         os.path.join(repo_install_dir, "backends", be),
         os.path.join(install_dir, "backends"),
@@ -2760,12 +2732,6 @@ def enable_all():
         default=DEFAULT_TRITON_VERSION_MAP["dcgm_version"],
         help="This flag sets the DCGM version for Triton Inference Server to be built. Default: the latest supported version.",
     )
-    parser.add_argument(
-        "--rhel-py-version",
-        required=False,
-        default=DEFAULT_TRITON_VERSION_MAP["rhel_py_version"],
-        help="This flag sets the Python version for RHEL platform of Triton Inference Server to be built. Default: the latest supported version.",
-    )
     parser.add_argument(
         "--docker-build-secret",
         action="append",

diff --git a/compose.py b/compose.py
@@ -33,7 +33,7 @@
 FLAGS = None
 
 
-#### helper functions
+# helper functions
 def log(msg, force=False):
     if force or not FLAGS.quiet:
         try:
@@ -255,35 +255,35 @@ def create_argmap(images, skip_pull):
     vars = p_path.stdout
     log_verbose("inspect args: {}".format(vars))
 
-    e0 = re.search("TRITON_SERVER_GPU_ENABLED=([\S]{1,}) ", vars)
+    e0 = re.search("TRITON_SERVER_GPU_ENABLED=([\\S]{1,}) ", vars)
     e1 = re.search("CUDA_VERSION", vars)
     gpu_enabled = False
-    if e0 != None:
+    if e0 is not None:
         gpu_enabled = e0.group(1) == "1"
-    elif e1 != None:
+    elif e1 is not None:
         gpu_enabled = True
     fail_if(
         gpu_enabled != enable_gpu,
         "Error: full container provided was build with "
         "'TRITON_SERVER_GPU_ENABLED' as {} and you are composing container"
         "with 'TRITON_SERVER_GPU_ENABLED' as {}".format(gpu_enabled, enable_gpu),
     )
-    e = re.search("TRITON_SERVER_VERSION=([\S]{6,}) ", vars)
+    e = re.search("TRITON_SERVER_VERSION=([\\S]{6,}) ", vars)
     version = "" if e is None else e.group(1)
     fail_if(
         len(version) == 0,
         "docker inspect to find triton server version failed, {}".format(p_path.stderr),
     )
-    e = re.search("NVIDIA_TRITON_SERVER_VERSION=([\S]{5,}) ", vars)
+    e = re.search("NVIDIA_TRITON_SERVER_VERSION=([\\S]{5,}) ", vars)
     container_version = "" if e is None else e.group(1)
     fail_if(
         len(container_version) == 0,
         "docker inspect to find triton container version failed, {}".format(vars),
     )
-    dcgm_ver = re.search("DCGM_VERSION=([\S]{4,}) ", vars)
+    dcgm_ver = re.search("DCGM_VERSION=([\\S]{4,}) ", vars)
     dcgm_version = ""
     if dcgm_ver is None:
-        dcgm_version = "4.5.3-1"
+        dcgm_version = "4.6.1-1"
         log(
             "WARNING: DCGM version not found from image, installing the earlierst version {}".format(
                 dcgm_version

diff --git a/deploy/aws/values.yaml b/deploy/aws/values.yaml
@@ -27,7 +27,7 @@
 replicaCount: 1
 
 image:
-  imageName: nvcr.io/nvidia/tritonserver:26.07-py3
+  imageName: nvcr.io/nvidia/tritonserver:26.08-py3
   pullPolicy: IfNotPresent
   modelRepositoryPath: s3://triton-inference-server-repository/model_repository
   numGpus: 1

diff --git a/deploy/fleetcommand/Chart.yaml b/deploy/fleetcommand/Chart.yaml
@@ -26,7 +26,7 @@
 
 apiVersion: v1
 # appVersion is the Triton version; update when changing release
-appVersion: 2.71.0
+appVersion: 2.72.0
 description: Triton Inference Server (Fleet Command)
 name: triton-inference-server
 # version is the Chart version; update when changing anything in the chart

diff --git a/deploy/fleetcommand/values.yaml b/deploy/fleetcommand/values.yaml
@@ -27,7 +27,7 @@
 replicaCount: 1
 
 image:
-  imageName: nvcr.io/nvidia/tritonserver:26.07-py3
+  imageName: nvcr.io/nvidia/tritonserver:26.08-py3
   pullPolicy: IfNotPresent
   numGpus: 1
   serverCommand: tritonserver
@@ -47,13 +47,13 @@ image:
     #
     # To set model control mode, uncomment and configure below
     # TODO: Fix the following url, it is invalid
-    # See https://github.com/triton-inference-server/server/blob/r26.07/docs/user_guide/model_management.md
+    # See https://github.com/triton-inference-server/server/blob/r26.08/docs/user_guide/model_management.md
     #  for more details
     #- --model-control-mode=explicit|poll|none
     #
     # Additional server args
     #
-    # see https://github.com/triton-inference-server/server/blob/r26.07/README.md
+    # see https://github.com/triton-inference-server/server/blob/r26.08/README.md
     #  for more details
 
 service:

diff --git a/deploy/gcp/values.yaml b/deploy/gcp/values.yaml
@@ -27,7 +27,7 @@
 replicaCount: 1
 
 image:
-  imageName: nvcr.io/nvidia/tritonserver:26.07-py3
+  imageName: nvcr.io/nvidia/tritonserver:26.08-py3
   pullPolicy: IfNotPresent
   modelRepositoryPath: gs://triton-inference-server-repository/model_repository
   numGpus: 1

diff --git a/deploy/gke-marketplace-app/benchmark/perf-analyzer-script/triton_client.yaml b/deploy/gke-marketplace-app/benchmark/perf-analyzer-script/triton_client.yaml
@@ -33,7 +33,7 @@ metadata:
   namespace: default
 spec:
   containers:
-  - image: nvcr.io/nvidia/tritonserver:26.07-py3-sdk
+  - image: nvcr.io/nvidia/tritonserver:26.08-py3-sdk
     imagePullPolicy: Always
     name: nv-triton-client
     securityContext:

diff --git a/deploy/gke-marketplace-app/server-deployer/build_and_push.sh b/deploy/gke-marketplace-app/server-deployer/build_and_push.sh
@@ -28,8 +28,8 @@
 export REGISTRY=gcr.io/$(gcloud config get-value project | tr ':' '/')
 export APP_NAME=tritonserver
 export MAJOR_VERSION=2.67
-export MINOR_VERSION=2.71.0
-export NGC_VERSION=26.07-py3
+export MINOR_VERSION=2.72.0
+export NGC_VERSION=26.08-py3
 
 docker pull nvcr.io/nvidia/$APP_NAME:$NGC_VERSION
 

diff --git a/deploy/gke-marketplace-app/server-deployer/chart/triton/Chart.yaml b/deploy/gke-marketplace-app/server-deployer/chart/triton/Chart.yaml
@@ -28,4 +28,4 @@ apiVersion: v1
 appVersion: "2.68"
 description: Triton Inference Server
 name: triton-inference-server
-version: 2.71.0
+version: 2.72.0
diff --git a/deploy/gke-marketplace-app/server-deployer/chart/triton/values.yaml b/deploy/gke-marketplace-app/server-deployer/chart/triton/values.yaml
@@ -31,14 +31,14 @@ maxReplicaCount: 3
 tritonProtocol: HTTP
 # HPA GPU utilization autoscaling target
 HPATargetAverageValue: 85
-modelRepositoryPath: gs://triton_sample_models/26.07
-publishedVersion: '2.71.0'
+modelRepositoryPath: gs://triton_sample_models/26.08
+publishedVersion: '2.72.0'
 gcpMarketplace: true
 
 image:
   registry: gcr.io
   repository: nvidia-ngc-public/tritonserver
-  tag: 26.07-py3
+  tag: 26.08-py3
   pullPolicy: IfNotPresent
   # modify the model repository here to match your GCP storage bucket
   numGpus: 1

diff --git a/deploy/gke-marketplace-app/server-deployer/data-test/schema.yaml b/deploy/gke-marketplace-app/server-deployer/data-test/schema.yaml
@@ -27,7 +27,7 @@
 x-google-marketplace:
   schemaVersion: v2
   applicationApiVersion: v1beta1
-  publishedVersion: '2.71.0'
+  publishedVersion: '2.72.0'
   publishedVersionMetadata:
     releaseNote: >-
       Initial release.