mc-nv · Oct 5, 2026 · Sep 24, 2026 · Oct 2, 2026 · Oct 2, 2026 · Oct 2, 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:26.09-py3-min
+ARG BASE_IMAGE=nvcr.io/nvidia/cuda-dl-base:26.09-cuda13.4-inference-runtime-ubuntu24.04
 
 ARG TRITON_CLIENT_REPO_SUBDIR=clientrepo
 ARG TRITON_REPO_ORGANIZATION=http://github.com/triton-inference-server

diff --git a/build.py b/build.py
@@ -59,6 +59,8 @@
 #   triton version ->
 #     (triton container version,
 #      upstream container version,
+#      cuda-dl-base version (container train and its CUDA version, which
+#        cuda-dl-base publishes as a matched pair and must be bumped together),
 #      ORT version,
 #      ORT OpenVINO version (use None to disable OpenVINO in ORT),
 #      Standalone OpenVINO version,
@@ -75,6 +77,7 @@
     "release_version": "2.74.0dev",
     "triton_container_version": "26.10dev",
     "upstream_container_version": "26.09",
+    "cuda_dl_base_version": "26.09-cuda13.4",
     "ort_version": "1.30.0",
     "ort_openvino_version": "2026.3.1",
     "standalone_openvino_version": "2026.3.1",
@@ -188,6 +191,18 @@ def target_machine():
     return platform.machine().lower()
 
 
+def default_build_container_image():
+    """Image the compile toolchain runs in.
+
+    Always CUDA-capable: the ONNX Runtime and OpenVINO backends build in this
+    container and need the CUDA toolchain even when Triton itself is being
+    built CPU-only.
+    """
+    return "nvcr.io/nvidia/cuda-dl-base:{}-devel-ubuntu24.04".format(
+        FLAGS.cuda_dl_base_version
+    )
+
+
 def container_versions(version, container_version, upstream_container_version):
     if container_version is None:
         container_version = FLAGS.triton_container_version
@@ -661,21 +676,14 @@ def onnxruntime_cmake_args(images, library_paths):
                 )
             )
 
-    if "base" in images:
-        cargs.append(
-            cmake_backend_arg(
-                "onnxruntime", "TRITON_BUILD_CONTAINER", None, images["base"]
-            )
-        )
-    else:
-        cargs.append(
-            cmake_backend_arg(
-                "onnxruntime",
-                "TRITON_BUILD_CONTAINER_VERSION",
-                None,
-                FLAGS.upstream_container_version,
-            )
+    cargs.append(
+        cmake_backend_arg(
+            "onnxruntime",
+            "TRITON_BUILD_CONTAINER",
+            None,
+            images.get("base", default_build_container_image()),
         )
+    )
 
     # TODO: TPRD-333 OpenVino extension is not currently supported by our manylinux build
     if (
@@ -719,21 +727,14 @@ def openvino_cmake_args():
             FLAGS.standalone_openvino_version,
         )
     ]
-    if "base" in images:
-        cargs.append(
-            cmake_backend_arg(
-                "openvino", "TRITON_BUILD_CONTAINER", None, images["base"]
-            )
-        )
-    else:
-        cargs.append(
-            cmake_backend_arg(
-                "openvino",
-                "TRITON_BUILD_CONTAINER_VERSION",
-                None,
-                FLAGS.upstream_container_version,
-            )
+    cargs.append(
+        cmake_backend_arg(
+            "openvino",
+            "TRITON_BUILD_CONTAINER",
+            None,
+            images.get("base", default_build_container_image()),
         )
+    )
     return cargs
 
 
@@ -1693,14 +1694,24 @@ def create_build_dockerfiles(
     elif target_platform() == "rhel":
         raise KeyError("A base image must be specified when targeting RHEL")
     elif FLAGS.enable_gpu:
-        base_image = "nvcr.io/nvidia/tritonserver:{}-py3-min".format(
-            FLAGS.upstream_container_version
-        )
+        base_image = default_build_container_image()
     else:
         base_image = "ubuntu:24.04"
 
     if "inference" in images:
         inference_image = images["inference"]
+    elif "base" in images:
+        # An explicit --image=base override has always supplied the runtime
+        # image as well. Leave it doing so, rather than silently swapping the
+        # final image for the default and discarding whatever runtime
+        # dependencies the override was chosen for.
+        inference_image = None
+    elif FLAGS.enable_gpu and target_platform() != "rhel" and "vllm" not in backends:
+        inference_image = (
+            "nvcr.io/nvidia/cuda-dl-base:{}-inference-runtime-ubuntu24.04".format(
+                FLAGS.cuda_dl_base_version
+            )
+        )
     else:
         inference_image = None
 
@@ -2541,7 +2552,7 @@ def enable_all():
         "--image",
         action="append",
         required=False,
-        help='Use specified Docker image in build as <image-name>,<full-image-name>. <image-name> can be "base", "gpu-base", or "pytorch".',
+        help='Use specified Docker image in build as <image-name>,<full-image-name>. <image-name> can be "base", "gpu-base", "pytorch", or "inference".',
     )
 
     parser.add_argument(
@@ -2708,6 +2719,12 @@ def enable_all():
         default=DEFAULT_TRITON_VERSION_MAP["upstream_container_version"],
         help="This flag sets the upstream container version for Triton Inference Server to be built. Default: the latest released version.",
     )
+    parser.add_argument(
+        "--cuda-dl-base-version",
+        required=False,
+        default=DEFAULT_TRITON_VERSION_MAP["cuda_dl_base_version"],
+        help="This flag sets the cuda-dl-base container train and its CUDA version, as a matched pair (e.g. 26.09-cuda13.4), used for the default build and runtime base images. Default: the latest supported version.",
+    )
     parser.add_argument(
         "--ort-version",
         required=False,

diff --git a/compose.py b/compose.py
@@ -90,7 +90,7 @@
        images["min"]
    )

    import build

    df += build.dockerfile_prepare_container_linux(
        argmap, backends, FLAGS.enable_gpu, platform.machine().lower()
@@ -176,7 +176,7 @@
        # Read from TRITON_VERSION file in server repo to determine version
        with open("TRITON_VERSION", "r") as vfile:
            version = vfile.readline().strip()
        import build

        _, FLAGS.container_version = build.container_versions(
            version, None, FLAGS.container_version
@@ -379,6 +379,16 @@
         "the container version will be chosen automatically based on the "
         "repository branch.",
     )
+    parser.add_argument(
+        "--cuda-dl-base-version",
+        type=str,
+        required=False,
+        help="The cuda-dl-base container train and its CUDA version, as a matched "
+        "pair (e.g. 26.09-cuda13.4), used for the 'min' container. If not "
+        "specified the value from build.py's version map is used, which "
+        "corresponds to the branch compose.py is on rather than to "
+        "--container-version.",
+    )
     parser.add_argument(
         "--image",
         action="append",
@@ -465,14 +475,32 @@
             log('image "{}": "{}"'.format(parts[0], parts[1]))
             images[parts[0]] = parts[1]
     else:
+        container_version_specified = FLAGS.container_version is not None
         get_container_version_if_not_specified()
         if FLAGS.enable_gpu:
+            import build
+
+            # cuda-dl-base couples its train to a CUDA version, so it cannot be
+            # derived from --container-version. Warn rather than silently pairing
+            # a requested `full` with a `min` from a different release.
+            if FLAGS.cuda_dl_base_version is None:
+                FLAGS.cuda_dl_base_version = build.DEFAULT_TRITON_VERSION_MAP[
+                    "cuda_dl_base_version"
+                ]
+                if container_version_specified:
+                    log(
+                        "warning: --container-version was specified but "
+                        "--cuda-dl-base-version was not, so the 'min' container "
+                        "will use {} and may not match the 'full' container".format(
+                            FLAGS.cuda_dl_base_version
+                        )
+                    )
             images = {
                 "full": "nvcr.io/nvidia/tritonserver:{}-py3".format(
                     FLAGS.container_version
                 ),
-                "min": "nvcr.io/nvidia/tritonserver:{}-py3-min".format(
-                    FLAGS.container_version
+                "min": "nvcr.io/nvidia/cuda-dl-base:{}-inference-runtime-ubuntu24.04".format(
+                    FLAGS.cuda_dl_base_version
                 ),
             }
         else:

diff --git a/deploy/gke-marketplace-app/README.md b/deploy/gke-marketplace-app/README.md
@@ -1,5 +1,5 @@
 <!--
-# Copyright (c) 2021-2024, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# Copyright (c) 2021-2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 #
 # Redistribution and use in source and binary forms, with or without
 # modification, are permitted provided that the following conditions
@@ -141,13 +141,13 @@ Please note that A100 MIG in GKE does not support GPU metrics yet, also Triton G
 
 Second, go to this [GKE Marketplace link](https://console.cloud.google.com/marketplace/details/nvidia-ngc-public/triton-inference-server) to deploy Triton application.
 
-Users can leave everything as default if their models have already been tested/validated with Triton. They can provide a GCS path pointing to the model repository containing their models. By default, we provide a BERT large model optimized by TensorRT in a public demo GCS bucket that is compatible with the `xx.yy` release of Triton Server in `gs://triton_sample_models/xx_yy`. However, please take note of the following about this demo bucket:
+Users can leave everything as default if their models have already been tested/validated with Triton. They can provide a GCS path pointing to the model repository containing their models. By default, we provide a BERT large model optimized by TensorRT in a public demo GCS bucket that is compatible with the `YY.MM` release of Triton Server in `gs://triton_sample_models/YY.MM`. However, please take note of the following about this demo bucket:
 - The TensorRT engine provided in the demo bucket is only compatible with Tesla T4 GPUs.
 - This bucket is located in `us-central1`, so loading from this bucket into Triton in other regions may be affected.
 - The first deployment of this Triton GKE application will be slower than consecutive runs because the image needs to be pulled into the GKE cluster.
 - You can find an example of how this model is generated and uploaded [here](trt-engine/README.md).
 
-Where <xx.yy> is the version of NGC Triton container needed.
+Where <YY.MM> is the version of NGC Triton container needed.
 
 ![GKE Marketplace Application UI](ui.png)
 

diff --git a/docs/README.md b/docs/README.md
@@ -46,10 +46,10 @@ Containers](http://docs.nvidia.com/deeplearning/dgx/preparing-containers/index.h
 Pull the image using the following command.
 
 ```
-$ docker pull nvcr.io/nvidia/tritonserver:<yy.mm>-py3
+$ docker pull nvcr.io/nvidia/tritonserver:<YY.MM>-py3
 ```
 
-Where \<yy.mm\> is the version of Triton that you want to pull. For a complete list of all the variants and versions of the Triton Inference Server Container,  visit the [NGC Page](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/tritonserver). More information about customizing the Triton Container can be found in [this section](customization_guide/compose.md) of the User Guide.
+Where \<YY.MM\> is the version of Triton that you want to pull. For a complete list of all the variants and versions of the Triton Inference Server Container,  visit the [NGC Page](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/tritonserver). More information about customizing the Triton Container can be found in [this section](customization_guide/compose.md) of the User Guide.
 
 ## **Getting Started**
 

diff --git a/docs/customization_guide/build.md b/docs/customization_guide/build.md
@@ -103,11 +103,11 @@ building with Docker.
     *tritonserver_buildbase* image is based on a minimal/base
     image. When building with GPU support (--enable-gpu), the *min*
     image is the
-    [\<xx.yy\>-py3-min](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/tritonserver)
+    [cuda-dl-base \<YY.MM\>-cuda\<X.Y\>-devel](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/cuda-dl-base)
     image pulled from [NGC](https://ngc.nvidia.com) that contains the
     CUDA, cuDNN, TensorRT and other dependencies that are required to
     build Triton. When building without GPU support, the *min* image
-    is the standard ubuntu:22.04 image.
+    is the standard ubuntu:24.04 image.
 
   * Run the cmake_build script within the *tritonserver_buildbase*
     image to actually build Triton. The cmake_build script performs
@@ -339,7 +339,7 @@ available for a non-GPU / CPU-only build: `identity`, `repeat`, `ensemble`,
 CPU-only builds of the PyTorch backends require some CUDA stubs
 and runtime dependencies that are not present in the CPU-only base container.
 These are retrieved from a GPU base container, which can be changed with the
-`--image=gpu-base,nvcr.io/nvidia/tritonserver:<xx.yy>-py3-min` flag.
+`--image=gpu-base,nvcr.io/nvidia/tritonserver:<YY.MM>-py3-min` flag.
 
 ### Building Without Docker
 
@@ -362,11 +362,9 @@ $ ./build.py -v --enable-all
 From Dockerfile.buildbase you can see what dependencies you need to
 install on your host system. Note that when building with --enable-gpu
 (or --enable-all), Dockerfile.buildbase depends on the
-[\<xx.yy\>-py3-min](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/tritonserver)
-image pulled from [NGC](https://ngc.nvidia.com). Unfortunately, a
-Dockerfile is not currently available for the
-[\<xx.yy\>-py3-min](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/tritonserver)
-image. Instead, you must manually install [CUDA and
+[cuda-dl-base \<YY.MM\>-cuda\<X.Y\>-devel](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/cuda-dl-base)
+image pulled from [NGC](https://ngc.nvidia.com). Rather than reproducing
+that image, you must manually install [CUDA and
 cuDNN](#cuda-cublas-cudnn) and [TensorRT](#tensorrt) dependencies as
 described below.
 

diff --git a/docs/customization_guide/compose.md b/docs/customization_guide/compose.md
@@ -72,14 +72,15 @@ This may result in different GPU statistic reporting behavior.
 
 `compose.py` requires two containers: a `min` container which is the
 base the compose container is built from and a `full` container from which the
-script will extract components. The version of the `min` and `full` container
-is determined by the branch of Triton `compose.py` is on.
+script will extract components. The version of the `full` container is
+determined by the branch of Triton `compose.py` is on, while the `min`
+container is the CUDA DL base image that branch builds against.
 For example, running
 ```
 python3 compose.py --backend pytorch --repoagent checksum
 ```
 on branch [r26.09](https://github.com/triton-inference-server/server/tree/r26.09) pulls:
-- `min` container `nvcr.io/nvidia/tritonserver:26.09-py3-min`
+- `min` container `nvcr.io/nvidia/cuda-dl-base:26.09-cuda13.4-inference-runtime-ubuntu24.04`
 - `full` container `nvcr.io/nvidia/tritonserver:26.09-py3`
 
 Alternatively, users can specify the version of Triton container to pull from
@@ -91,7 +92,7 @@ python3 compose.py --backend pytorch --repoagent checksum --container-version 26
 2. Specifying `--image min,<min container image name> --image full,<full container image name>`.
    The user is responsible for specifying compatible `min` and `full` containers.
 ```
-python3 compose.py --backend pytorch --repoagent checksum --image min,nvcr.io/nvidia/tritonserver:26.09-py3-min --image full,nvcr.io/nvidia/tritonserver:26.09-py3
+python3 compose.py --backend pytorch --repoagent checksum --image min,nvcr.io/nvidia/cuda-dl-base:26.09-cuda13.4-inference-runtime-ubuntu24.04 --image full,nvcr.io/nvidia/tritonserver:26.09-py3
 ```
 Method 1 and 2 will result in the same composed container. Furthermore,
 `--image` flag overrides the `--container-version` flag when both are specified.

diff --git a/docs/getting_started/quickstart.md b/docs/getting_started/quickstart.md
@@ -1,5 +1,5 @@
 <!--
-# Copyright (c) 2018-2024, NVIDIA CORPORATION. All rights reserved.
+# Copyright (c) 2018-2026, NVIDIA CORPORATION. All rights reserved.
 #
 # Redistribution and use in source and binary forms, with or without
 # modification, are permitted provided that the following conditions
@@ -73,10 +73,10 @@ for Docker to recognize the GPU(s). The --gpus=1 flag indicates that 1
 system GPU should be made available to Triton for inferencing.
 
 ```
-$ docker run --gpus=1 --rm -p8000:8000 -p8001:8001 -p8002:8002 -v/full/path/to/docs/examples/model_repository:/models nvcr.io/nvidia/tritonserver:<xx.yy>-py3 tritonserver --model-repository=/models
+$ docker run --gpus=1 --rm -p8000:8000 -p8001:8001 -p8002:8002 -v/full/path/to/docs/examples/model_repository:/models nvcr.io/nvidia/tritonserver:<YY.MM>-py3 tritonserver --model-repository=/models
 ```
 
-Where \<xx.yy\> is the version of Triton that you want to use (and
+Where \<YY.MM\> is the version of Triton that you want to use (and
 pulled above). After you start Triton you will see output on the
 console showing the server starting up and loading the model. When you
 see output like the following, Triton is ready to accept inference
@@ -106,7 +106,7 @@ On a system without GPUs, Triton should be run without using the
 above.
 
 ```
-$ docker run --rm -p8000:8000 -p8001:8001 -p8002:8002 -v/full/path/to/docs/examples/model_repository:/models nvcr.io/nvidia/tritonserver:<xx.yy>-py3 tritonserver --model-repository=/models
+$ docker run --rm -p8000:8000 -p8001:8001 -p8002:8002 -v/full/path/to/docs/examples/model_repository:/models nvcr.io/nvidia/tritonserver:<YY.MM>-py3 tritonserver --model-repository=/models
 ```
 
 Because the --gpus flag is not used, a GPU is not available and Triton
@@ -136,17 +136,17 @@ Use docker pull to get the client libraries and examples image
 from NGC.
 
 ```
-$ docker pull nvcr.io/nvidia/tritonserver:<xx.yy>-py3-sdk
+$ docker pull nvcr.io/nvidia/tritonserver:<YY.MM>-py3-sdk
 ```
 
-Where \<xx.yy\> is the version that you want to pull. Run the client
+Where \<YY.MM\> is the version that you want to pull. Run the client
 image.
 
 ```
-$ docker run -it --rm --net=host nvcr.io/nvidia/tritonserver:<xx.yy>-py3-sdk
+$ docker run -it --rm --net=host nvcr.io/nvidia/tritonserver:<YY.MM>-py3-sdk
 ```
 
-From within the nvcr.io/nvidia/tritonserver:<xx.yy>-py3-sdk
+From within the nvcr.io/nvidia/tritonserver:<YY.MM>-py3-sdk
 image, run the example image-client application to perform image
 classification using the example densenet_onnx model.