Fix ParallelTaskAssigner's use of the HloPassPipeline interface
We were creating the ParallelTaskAssignment contained in ParallelTaskAssigner with an unoptimized module and then trying to ParallelTaskAssigning::Run on an optimized module. This meant that the flop counts in HloCostAnalysis were cached using bogus HloInstruction* pointers, which meant our parallel task assignment was not effective. PiperOrigin-RevId: 174909618
Loading
Please sign in to comment