Repository navigation
fix: Require TRITON_BUILD_CONTAINER instead of synthesizing a py3-min image - #366
Merged
Merged
Conversation
… image
When only TRITON_BUILD_CONTAINER_VERSION was given, the build container was
invented as nvcr.io/nvidia/tritonserver:${VERSION}-py3-min. That image is
being retired, so the fallback is a latent failure: configure succeeds and the
build dies much later on a docker pull 404 for an image nobody asked for.
Stop inventing one. Building ONNX Runtime from source now requires
TRITON_BUILD_CONTAINER, which turns a missing image into a configure-time
error naming exactly what to supply. TRITON_BUILD_CONTAINER_VERSION had no
other use in this repo -- it only fed the synthesis -- so it goes with it and
the two conditions collapse into one guard.
Unlike the OpenVINO backend, this one needs the CUDA toolchain in its build
container when GPU support is enabled, so both README examples name a
CUDA-capable image and the header comment says so.
Breaking for standalone cmake builds: configuring with only
-DTRITON_BUILD_CONTAINER_VERSION now fails with
"TRITON_BUILD_ONNXRUNTIME_VERSION requires TRITON_BUILD_CONTAINER". The README
examples are updated in the same commit so the documented path does not point
at a removed variable. Builds driven by server/build.py are unaffected -- it
always passes a resolved image.
 |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
The examples pinned ONNX Runtime 1.14.1 and OpenVINO 2021.2.200 -- versions the surrounding text ties to Triton 23.04 -- while naming a 26.09 CUDA 13.4 build container, leaving no example of a matched configuration. Move the versions and the worked reference to 26.09: ORT 1.30.0 and OpenVINO 2026.3.1, matching that release's TRITON_VERSION_MAP entry. Reported by Greptile on #366.
whoisj
approved these changes
Oct 5, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does the PR do?
TRITON_BUILD_CONTAINER_VERSIONinventedtritonserver:${VERSION}-py3-min, an image being retired — configuresucceeded and the build died later on a
docker pull404.TRITON_BUILD_CONTAINER.TRITON_BUILD_CONTAINER_VERSIONonly fed that synthesis, so it goes too.Standalone
cmakebuilds passing only-DTRITON_BUILD_CONTAINER_VERSIONnowfail with
TRITON_BUILD_ONNXRUNTIME_VERSION requires TRITON_BUILD_CONTAINER.This is the intent — loud and immediate, rather than an opaque 404 once
py3-min is withdrawn. Both README examples are updated in the same commit so
the documented path never references a removed variable.
Builds driven by
server/build.pyare unaffected: it always passes aresolved image. The pre-built-ONNX-Runtime path is untouched — the guard sits
inside the build-from-source branch.
Note vs. the sibling OpenVINO change
Unlike
openvino_backend, this backend does need the CUDA toolchain in itsbuild container when GPU support is enabled (
TRITON_ENABLE_GPUdefaults ON,plus
TRITON_BUILD_CUDA_VERSION/HOME). Both README examples therefore name aCUDA-capable image, and the header comment says so.
Where should the reviewer start?
CMakeLists.txt— the removed synthesis and the singleFATAL_ERRORguard.Test plan
Verified with
cmake4.4.3:TRITON_BUILD_CONTAINER-DTRITON_BUILD_CONTAINER_VERSION=23.04only-DTRITON_BUILD_CONTAINER=<image>Also
./build.py --dryrun --enable-gpu --backend onnxruntimeinserver: thegenerated
cmake_buildcarriesTRITON_BUILD_CONTAINER=nvcr.io/nvidia/cuda-dl-base:26.09-cuda13.4-devel-ubuntu24.04,confirming
build.pynever relied on the removed fallback.Related PRs:
Related Issues: