[Windows Desktop 26.814.5167.0] launching CUDA llama-cli through exec crashes the Codex task
What version of the Codex App are you using?
26.814.5167.0 (Windows x64 MSIX)
What platform is your computer?
Windows x64, NVIDIA RTX 5070 Ti
What issue are you seeing?
Launching a CUDA-enabled llama-cli.exe through a Codex Desktop code-mode exec call crashes/restarts the ChatGPT/Codex task instead of containing the child process. This reproduced twice in the same task. After the app recovered, the tool result was only aborted.
The executable was llama.cpp build 10375 (ba360efe1), built with Clang 20.1.8 for Windows x86_64. The GPU had about 15.3 GiB free before the second attempt, and no llama-cli, llama-perplexity, or llama-server process remained after recovery.
I found no matching Windows Application Error or WER event in the last 12 hours, so I cannot yet identify whether the failing process is codex-code-mode-host.exe, ChatGPT.exe, or another Electron helper.
Steps to reproduce
From a Codex Desktop task on Windows, run a CUDA llama.cpp CLI through the exec tool. The command used was equivalent to:
& '...\llama-cli.exe' `
-m '...\nanbeige4.2-3b-Q8_0.gguf' `
-ngl all -c 512 -b 512 -ub 512 `
--override-kv 'nanbeige.num_loops=int:1' `
-p 'Write a valid JSON object with keys name and value.' `
-n 32 --temp 0 --no-display-prompt -no-cnv --no-warmup
A first attempt with -p 'Hello' -n 1 caused the same app/task crash. Ordinary short PowerShell, Git, ripgrep, and nvidia-smi exec calls in the same task work normally.
Expected behavior
Codex should either return the child process output/exit status or report that the command runner failed. A native child executable should not crash or restart the Codex/ChatGPT task.
Additional information
This may be related to #38203 (Windows codex-code-mode-host.exe fail-fast and stuck exec calls), but this report has a deterministic external CUDA executable trigger on a newer desktop build. I am not claiming the same native root cause without a crash dump or matching WER record.
1 Comment
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action