Meet MindSpore, the new open source AI framework!
summary
MindSpore claims to automate distributed training decisions by using an Auto Parallel component that measures device utilization and gradients, applies a cost model, and picks from a pool of parallelization strategies via dynamic programming to partition computation and run across devices. The talk contrasts that with hand-tuned TensorFlow/PyTorch setups, cites the scale of modern models (≈130 billion parameters on >1000 GPUs), and notes MindSpore currently targets Huawei chips with GPU/CPU support still needing community contributions.
AI-written summary · about
chapters (4)
related talks
24mMindSpore's Foundation Model Platform Promotes the Ecological Development of Foundat...
38mData Pre-processing in MindSpore
28mAutomatic Operator Generation for Open AI frameworks: MindSpore/AKG
15mCo-host Sponsor Session: Open, Innovate, Fail Fast: MindSpore Community Operation
11mSimAI: Unifying Architecture Design and Performance Tuning for Large-Scale Large...
20mOperating Frontier Intelligence at Scale
28mScaling AI Infrastructure
12mSmartMoE: Efficiently Training Sparsely-Activated Models through Combiningand...
