MOGD: move tensors between devices

Angryrou / UDAO2022

Apache License 2.0

0 stars 0 forks source link

MOGD: move tensors between devices #82

Closed julhien closed 11 months ago

julhien commented 11 months ago

Adapt MOGD use of tensors to support the use of one or several gpus.

Adapt tests with different values for GPU deterministic case vs. CPU deterministic case.

N.B. currently the deterministic behavior does not allow some tests to run unless the following variable is set. export CUBLAS_WORKSPACE_CONFIG=:16:8 (see https://docs.nvidia.com/cuda/cublas/index.html#cublasApi_reproducibility)

julhien commented 11 months ago

Current dependencies on/for this PR:

trace-parsing-and-modeling-spark
- PR #82 👈

This stack of pull requests is managed by Graphite.

julhien commented 11 months ago

PF algorithm should work with GPU too. I'm just setting it as CPU in the tests to keep the same results. The GPU version of MOGD is tested in the MOGD tests.