The GitHub Actions job "Java CI" on opennlp.git/OPENNLP-1982-sentence-vectors-inference-options has failed. Run started by GitHub user krickert (triggered by krickert).
Head commit for run: 576206d29415eb777a253b6125ab03f7566c7dcb / Kristian Rickert <[email protected]> GH-1389, GH-1390: Let SentenceVectorsDL run on a GPU and fix the first CUDA request in a fresh JVM SentenceVectorsDL always created default session options, so it ran on the CPU even with opennlp-dl-gpu. A new constructor takes InferenceOptions, like NameFinderDL and DocumentCategorizerDL, and uses the GPU settings and the lower case setting from them; a lower case setting in the options takes precedence over the lowerCase argument. The existing constructors keep their behavior. AbstractDL.sessionOptions added the CUDA provider before any OrtEnvironment existed. Loading the provider needs ONNX Runtime's default logger, which only exists once an environment does, so the first GPU request in a fresh JVM failed with "Attempt to use DefaultLogger but none has been registered". The environment is now created first. This affects NameFinderDL and DocumentCategorizerDL as well. A test runs the first CUDA request in a child JVM. Other tests request a device id no machine has and check that ONNX Runtime rejects it for SentenceVectorsDL, NameFinderDL and DocumentCategorizerDL, and compare vectors on CUDA device 0 with the CPU session (skipped without a CUDA provider or device). The opennlp-dl-gpu test classpath no longer has the CPU onnxruntime next to onnxruntime_gpu. The machine learning chapter documents the GPU setup. Fixes #1389 Fixes #1390 Co-authored-by: Richard Zowalla <[email protected]> Co-authored-by: Martin Wiesner <[email protected]> Signed-off-by: Kristian Rickert <[email protected]> Report URL: https://github.com/apache/opennlp/actions/runs/38104092851 With regards, GitHub Actions via GitBox
