Added new protocol to improve performance on small operations.

Improved topology detection and tree/ring creation (#179, #262).

Improved multi-node tree performance by sending/receiving from different GPUs.

Added model-based tuning to switch between the different algorithms and protocols.

Reworked P2P/SHM detection in containers (#155, #248).

Added detection for duplicate CUDA devices and return an error (#231).

Added tuning for Google Cloud’s gVNIC platform.

Deep learning framework containers. Refer to theSupport Matrix for the supported container version.

This NCCL release supports CUDA 9.0, CUDA 10.0, and CUDA 10.2.

Sporadic NCCL error "ring 0 does not loop back to start” (#179).

NCCL doesn't form proper rings on GCP V100s (#262).