List view
By v0.9 we *must* figure out: * built-in gating function averaging; * built-in tensor compression & quantization; * faster beam search with tunable accuracy/latency trade-off; * measure & improve tests coverage;
No due date•6/6 issues closedGoal: actually support training on 1000s of nodes Features: * server: switch from pythonic connection_handler to asyncio + gRPC * client: implement parallel fault-tolerant backward for moe.py * dht: implement bulk store/get operations with caching
No due date•12/12 issues closed