Input arrays of 2 MiB or larger produce incorrect results with CUDA shared memory

triton-inference-server / fil_backend

FIL backend for the Triton Inference Server

Apache License 2.0

72 stars 36 forks source link

Closed wphicks closed 3 years ago

wphicks commented 3 years ago

First 13330 results are random floats rather than 0s or 1s with predict_proba off