feat(triton): Update Triton Base Image to 26.08

container/README.md

@@ -253,9 +253,9 @@ docker build -t dynamo:latest-sglang-xpu-runtime -f container/sglang-runtime-xpu
 # Build Triton runtime image (prebuilt Dynamo wheels from PyPI on the upstream
 # Triton release image). --network=host lets the build reach PyPI.
 # The Triton release is selected with --build-arg RUNTIME_IMAGE_TAG=<tag>
-# (defaults to 26.07-py3); pick any nvcr.io/nvidia/tritonserver:<tag>.
+# (defaults to 26.08-py3); pick any nvcr.io/nvidia/tritonserver:<tag>.
 container/render.py --framework=triton --target=runtime --output-short-filename
-docker build --network=host --build-arg RUNTIME_IMAGE_TAG=26.07-py3 -t dynamo:latest-triton-runtime -f container/rendered.Dockerfile .
+docker build --network=host --build-arg RUNTIME_IMAGE_TAG=26.08-py3 -t dynamo:latest-triton-runtime -f container/rendered.Dockerfile .
 ```
 
 The `--device` flag selects the accelerator backend. It defaults to `cuda`; pass `--device=xpu`

container/context.yaml

@@ -182,8 +182,14 @@ triton:
     # Build toolchain only, pinned to CUDA 13.1: the Rust cudarc crate rejects 13.2 at build time.
     # The wheels load CUDA dynamically, so they run on the 13.2 runtime image regardless.
     base_image_tag: 26.02-cuda13.1-devel-ubuntu24.04
-    # This value will be updated monthly by the Triton Team after each release.
     runtime_image_tag: 26.07-py3
+  cuda13.4:
+    base_image: nvcr.io/nvidia/cuda-dl-base
+    runtime_image: nvcr.io/nvidia/tritonserver
+    # Build toolchain only, pinned to CUDA 13.1: the Rust cudarc crate rejects 13.4 at build time.
+    # The wheels load CUDA dynamically, so they run on the 13.4 runtime image regardless.
+    base_image_tag: 26.02-cuda13.1-devel-ubuntu24.04
+    runtime_image_tag: 26.08-py3
   enable_media_ffmpeg: "false"
   enable_gpu_memory_service: "true"
   # KVBM is disabled: it is an LLM KV-cache feature whose Python NIXL path

container/render.py

@@ -100,11 +100,11 @@ def parse_args():
         "--cuda-version",
         type=str,
         default="13.0",
-        choices=["13.0", "13.1"],
+        choices=["13.0", "13.1", "13.2", "13.4"],
         help=(
             "CUDA version to use. [13.0 for vllm and sglang, 13.1 for trtllm].\n"
             "Not required for non-cuda devices.\n"
-            "Not supported by Triton - CUDA version is predefined by its release image."
+            "Only Triton supports versions greater than 13.1."
         ),
     )
     parser.add_argument("--make-efa", action="store_true", help="Enable AWS EFA")
@@ -165,7 +165,7 @@ def validate_args(args):
             "target": [
                 "runtime",
             ],
-            "cuda_version": ["13.2"],
+            "cuda_version": ["13.2", "13.4"],
         },
         "dynamo": {
             "device": ["cuda"],
@@ -182,19 +182,6 @@ def validate_args(args):
         },
     }
 
-    # Triton's CUDA family is fixed by its release image, so it cannot be chosen
-    # by the user: reject an explicitly-passed --cuda-version (detected from argv
-    # since the arg has a default) and pin it to Triton's single valid value.
-    if args.framework == "triton":
-        if any(
-            a == "--cuda-version" or a.startswith("--cuda-version=") for a in sys.argv
-        ):
-            raise ValueError(
-                "--cuda-version cannot be specified for triton: its CUDA family is "
-                "fixed by the Triton release image."
-            )
-        args.cuda_version = valid_inputs["triton"]["cuda_version"][0]
-
     if args.framework in valid_inputs:
         cuda_version_valid = (
             args.device != "cuda"

docs/fern/pages/developer-guide/knowledge-base/modular-components/backends/triton/overview.md

@@ -45,9 +45,10 @@ The Triton and KServe features below are not yet supported end-to-end through th
 
 | Triton release tag                      |  CUDA  | NVIDIA driver |
 | :-------------------------------------- | :----: | :-----------: |
-| `nvcr.io/nvidia/tritonserver:26.07-py3` | `13.2` |  `610.43.02`  | <!-- Update version with each Triton release. -->
+| `nvcr.io/nvidia/tritonserver:26.07-py3` | `13.2` |  `610.43`  |
+| `nvcr.io/nvidia/tritonserver:26.08-py3` | `13.4` |  `615.61`  |
 
-The Triton release is pinned in [`container/context.yaml`](https://github.com/ai-dynamo/dynamo/tree/main/container/context.yaml) under `triton.cuda13.2` (`runtime_image_tag`/`base_image_tag`), the same way the other framework runtimes pin their image.
+The Triton release is pinned in [`container/context.yaml`](https://github.com/ai-dynamo/dynamo/tree/main/container/context.yaml) under `triton.cuda13.2` and `triton.cuda13.4` (`runtime_image_tag`/`base_image_tag`), the same way the other framework runtimes pin their image.
 
 ## Prerequisites
 
@@ -170,21 +171,21 @@ Common flags:
 
 ## Configuring the Triton version
 
-The Triton release is pinned by `triton.cuda13.2.runtime_image_tag` in [`container/context.yaml`](https://github.com/ai-dynamo/dynamo/tree/main/container/context.yaml) (default `26.07-py3`), mirroring how the other framework runtimes pin their image.
+The Triton release is pinned by `triton.cuda13.4.runtime_image_tag` in [`container/context.yaml`](https://github.com/ai-dynamo/dynamo/tree/main/container/context.yaml) (default `26.08-py3`), mirroring how the other framework runtimes pin their image.
 The CUDA family is fixed by the Triton release, so `--cuda-version` is auto-derived; passing it explicitly is rejected.
 
 To build into a different Triton release, override `RUNTIME_IMAGE_TAG` at build time (no `context.yaml` edit needed). Keep `BASE_IMAGE_TAG` on CUDA 13.1 (default `26.02-cuda13.1-devel-ubuntu24.04`):
-the Rust `cudarc` crate rejects CUDA 13.2 at build time, and the wheels load CUDA dynamically so they run on the 13.2 runtime image regardless:
+the Rust `cudarc` crate rejects other CUDA versions at build time, and the wheels load CUDA dynamically so they run on the 13.4 runtime image regardless:
 
 ```bash
 python container/render.py --framework=triton --target=runtime --output-short-filename
 docker buildx build --network=host \
-  --build-arg RUNTIME_IMAGE_TAG=26.07-py3 \
+  --build-arg RUNTIME_IMAGE_TAG=26.08-py3 \
   --build-arg BASE_IMAGE_TAG=26.02-cuda13.1-devel-ubuntu24.04 \
-  -f container/rendered.Dockerfile -t dynamo:triton-26.07 .
+  -f container/rendered.Dockerfile -t dynamo:triton-26.08 .
 ```
 
-To make a release the default, edit `triton.cuda13.2` in `container/context.yaml`.
+To make a release the default, edit `triton.cuda13.4` in `container/context.yaml`.
 
 ## Metrics