Code

Algorithm Optimization

This code has been written in C# with cooperation of Chinese Academy of Science, Nanjing Biostratigraphy Research Team

• C# CONOP9

Deep Learning Framework Enablement on eGPU-macOS 10.13.6

As officially Pytorch doesn't support for macOS cuda, I used this repository to build pytorch on macOS cuda. This branch 1.12.0-fixed branch is the current stable branch with MPI+CUDA enabled.

As officially Tensorflow doesn't support for macOS cuda, I used this repository to build tensorflow 2.8+ on macOS cuda. This branch v2.9.1-fixed branch is the current investigation branch.

As officially JAX doesn't support macOS(at least testing on my macOS 10.13.6, cuda 10.1), trying to fix building issue and enabling CUDA on macOS becomes a task that help me reuse Nx 1080 cards for ML acceleration.

Optimized primitives for collective multi-GPU communication migrated to Mac OS X (10.13 - 10.13.6). In order to make library and nccl-test compatible, each nccl library version will be unique mapped to nccl-test on macOS version.

As official bazel requires libtool provided by Xcode to accept params file as arguments, such feature breaks the building of libraries on macOS 10.13.6 (Xcode 10.1) that is the basis of tensorflow on macOS with GPU supports. The building issue of bazel 5.2.0 mentioned below will unavoidly happened on the legacies system that have been on the side line of bazel's release plan.O ne patch on top of 5.2.0 has been applied for bazel, which is the main purpose of this repository.

Miscellaneous

Implementation based on torch, lightning as well as language models.

Implementation based on torch, lightning as well as language models.

Course assignement given by Berkely Fall 2018, refer to Deep Reinforcement Learning CS285.