Difference between revisions of "TensorRT"
Jump to navigation
Jump to search
(Created page with "TensorRT is a timing-based kernel schedule. It is non-deterministic so it cannot be used for Grassland nodes. "Each time an inference engine is generated, it could be slightly...") |
|||
| Line 1: | Line 1: | ||
TensorRT is a timing-based kernel schedule. It is non-deterministic so it cannot be used for Grassland nodes. "Each time an inference engine is generated, it could be slightly different, particularly if there is varying load on the machine used to run TensorRT. There is a solution planned for this"<ref>https://github.com/NVIDIA/tensorflow-determinism/tree/8d4580609c2ac5ae9d5e33956bec802152220449#sources-of-non-determinism-beyond-tensorflow</ref>. | TensorRT is a timing-based kernel schedule. It is non-deterministic (see [[Determinism]]) so it cannot be used for Grassland nodes until that is fixed. "Each time an inference engine is generated, it could be slightly different, particularly if there is varying load on the machine used to run TensorRT. There is a solution planned for this"<ref>https://github.com/NVIDIA/tensorflow-determinism/tree/8d4580609c2ac5ae9d5e33956bec802152220449#sources-of-non-determinism-beyond-tensorflow</ref>. | ||
Latest revision as of 20:29, 21 February 2020
TensorRT is a timing-based kernel schedule. It is non-deterministic (see Determinism) so it cannot be used for Grassland nodes until that is fixed. "Each time an inference engine is generated, it could be slightly different, particularly if there is varying load on the machine used to run TensorRT. There is a solution planned for this"[1].