triton-inference-server · mc-nv · Aug 4, 2026 · Aug 4, 2026 · Aug 4, 2026 · Aug 5, 2026
diff --git a/Dockerfile.sdk b/Dockerfile.sdk
@@ -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/build.py b/build.py
@@ -75,11 +75,10 @@
     "release_version": "2.72.0dev",
     "triton_container_version": "26.08dev",
     "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",
+    "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"]
@@ -925,9 +910,6 @@ 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 \\
@@ -961,20 +943,21 @@ 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
@@ -1335,9 +1318,10 @@ 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
 
 """
@@ -1412,8 +1396,7 @@ 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 \\
@@ -1554,30 +1537,30 @@ 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
 ):
@@ -1987,17 +1970,6 @@ 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"),
@@ -2616,12 +2588,6 @@ 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(
         "--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/qa/L0_backend_python/env/test.sh b/qa/L0_backend_python/env/test.sh
@@ -41,10 +41,9 @@ rm -rf *.tar.gz
 install_build_deps
 install_conda
 
-# Test conda env without custom Python backend stub This environment should
-# always use the default Python version shipped in the container. For Ubuntu
-# 24.04 it is Python 3.12, for Ubuntu 22.04 is Python 3.10 and for Ubuntu 20.04
-# is 3.8.
+# Test conda env without custom Python backend stub. This environment should
+# always use the default Python version shipped in the container, which is
+# 3.12 -- nothing below that is supported.
 path_to_conda_pack='$$TRITON_MODEL_DIRECTORY/python_3_12_environment.tar.gz'
 create_conda_env "3.12" "python-3-12"
 conda install -c conda-forge libstdcxx-ng=14 -y

diff --git a/qa/L0_backend_python/examples/test.sh b/qa/L0_backend_python/examples/test.sh
@@ -42,33 +42,18 @@ pip3 uninstall -y numpy
 # NOTE: Using this subtest as a test case that involves using a python model with
 # numpy 2.X without changing the environments used in all the other test cases.
 if [ "$TEST_JETSON" == "0" ] && [[ ${TEST_WINDOWS} == 0 ]]; then
-    if [ ${PYTHON_ENV_VERSION} == "8" ]; then
-        # Python 3.8 does not support numpy 2.x, so installing numpy1.x
-        pip3 install "numpy<2"
-        pip3 install torch==2.0.0+cu117 -f https://download.pytorch.org/whl/torch_stable.html torchvision==0.15.0+cu117
-    else
-        # Python 3.9 >= supports numpy 2.x.
-        pip3 install "numpy>=2"
-        pip3 install torch==2.5.0 torchvision==0.20.0 --index-url https://download.pytorch.org/whl/cu124
-    fi
+    pip3 install "numpy>=2"
+    pip3 install torch==2.5.0 torchvision==0.20.0 --index-url https://download.pytorch.org/whl/cu124
 else
-    if [ ${PYTHON_ENV_VERSION} == "8" ]; then
-        # Python 3.8 does not support numpy 2.x, so installing numpy1.x
-        pip3 install "numpy<2"
-        pip3 install torch==2.0.0 -f https://download.pytorch.org/whl/torch_stable.html torchvision==0.15.0
-    else
-        # Python 3.9 >= supports numpy 2.x.
-        pip3 install "numpy>=2"
-        pip3 install torch==2.5.0 -f https://download.pytorch.org/whl/torch_stable.html torchvision==0.20.0
-    fi
+    pip3 install "numpy>=2"
+    pip3 install torch==2.5.0 -f https://download.pytorch.org/whl/torch_stable.html torchvision==0.20.0
 fi
 
 # Install `validators` for Model Instance Kind example
 pip3 install validators
 
 # Install JAX
-# Jax has dropped the support for Python 3.8. See https://jax.readthedocs.io/en/latest/changelog.html
-if [ "$TEST_JETSON" == "0" ] && [ ${PYTHON_ENV_VERSION} != "8" ]; then
+if [ "$TEST_JETSON" == "0" ]; then
     pip install -U "jax[cuda12]"
 fi
 
@@ -152,8 +137,7 @@ kill_server
 
 # JAX AddSub
 # JAX is not supported on Jetson
-# Jax has dropped the support for Python 3.8. See https://jax.readthedocs.io/en/latest/changelog.html
-if [ "$TEST_JETSON" == "0" ] && [ ${PYTHON_ENV_VERSION} != "8" ]; then
+if [ "$TEST_JETSON" == "0" ]; then
     CLIENT_LOG="../examples_jax_client.log"
     mkdir -p models/jax/1/
     cp examples/jax/model.py models/jax/1/model.py