Parallel execution of operators

Hi All,

I am wondering is there a way to schedule graph nodes so that they can run in parallel on parallel hardware?

In other words, how we can express nodes (operators) in the json file so that the runtime runs them in parallel?

Thanks, S.

Hi, TVM already takes advantage of intra-op parallelism, e.g., parallelizing the execution of a convolution itself. However, I have the same question if inter-op parallelism is already supported or if there any plans to support that.

Inter-op parallelism is going to be very relevant for other TVM features like the new partitioning algorithms as it is discussed in the following RFC: [Relay] Improved graph partitioning algorithm

@comaniac, @FrozenGene, @srkreddy1238 could you please comment about this?