Implementing Single-Dim Strategies With Sharding Validator
summary
Prototype separates semantic correctness from pruning by requiring each operator to declare Single Mesh Dim Strategies for one mesh dimension while infra composes them across N dims and picks the lowest‑cost plan. A landed Truth Table-style sharding validator enumerates 1D placements, runs DTensor simulations in local tensor mode, and compares against eager outputs to find missing or incorrect rules. Example run surfaces three incorrect strategies that get fixed and shows zero missing cases after validation, cutting unexpected redistributions.
AI-written summary · about
related talks
01h 03mScaling up machine learning without tears (and what do programming languages ha...
17mScaling DSpark Training Using vLLM, Speculators and Mooncake
24mValidity and the future of blockchains
37mSharding: Growing Systems from Node-scale to Planet-scale
13mSharding the State Machine: Automated Modular Reasoning for Complex Concurrent System
17mState of distributed validators in 2025
14mImproving Censorship Resistance Using DVT
20mScheduler and Sharding Considerations for Network Efficiency