Do not use cache to avoid lock overlap with conda

triton-inference-server / fil_backend

FIL backend for the Triton Inference Server

Apache License 2.0

68 stars 35 forks source link

Do not use cache to avoid lock overlap with conda #217

Closed wphicks closed 2 years ago

wphicks commented 2 years ago

Currently, BuildKit parallelism causes conda installs to overlap from different layers, intermittently preventing one conda process from getting a lock. This PR eliminates the shared cache between layers. The performance impact of this is minimal and offset by the new use of mamba for environment installs.