Repository navigation
Conversation
A fatal CUDA error (e.g. an out-of-bounds kernel write that leaves the context in cudaErrorIllegalAddress) makes an instance unable to serve, yet it kept reporting READY, so orchestrators never restarted it. Detect sticky/unrecoverable CUDA errors after a failed run, latch the instance unhealthy, and report NOT READY via TRITONBACKEND_ModelInstanceReady so the instance can be restarted.
 |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
whoisj
requested changes
Sep 18, 2026
whoisj
left a comment
whoisj
left a comment
There was a problem hiding this comment.
overall, I think the change makes sense, but would like cudaErrorAssert addressed.
mc-nv
reviewed
Sep 21, 2026
A device-side assert leaves the CUDA context unusable like the other sticky errors, so the instance must latch NOT READY on it too.
whoisj
approved these changes
Sep 29, 2026
mc-nv
approved these changes
Sep 29, 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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 — After a fatal, unrecoverable CUDA error during inference, the backend marks the instance unhealthy and reports it NOT READY instead of continuing to report READY.
Why — Sticky CUDA errors (illegal address, launch failure, ECC…) corrupt the context so every later request fails, but the backend returns success from Execute and the model keeps reporting READY — orchestrators never restart it, so a fatal GPU error becomes a silent, indefinite outage.
Change (src/onnxruntime.cc) — IsFatalCudaError() classifier; probe the CUDA context after a failed OrtRun and latch an atomic device_healthy_; implement TRITONBACKEND_ModelInstanceReady to return UNAVAILABLE once unhealthy (mirrors python_backend). Healthy path and CPU-only builds untouched.
Verification — Built and run in tritonserver:26.07-py3 on an A10: a fatal CUDA error now flips readiness to NOT READY (was READY before). IsFatalCudaError unit-tested; clang-format clean.