| # distributed-multi-gpu | |
| Two-GPU tasks. The agent moves the bytes itself (NVSHMEM, CUDA IPC, a Triton kernel over peer memory); torch's own collectives are blocked inside the graded call. | |
| **8 tasks.** | |
| | task | metric | roofline | what the kernel does | | |
| |---|---|---|---| | |
| | [`dist-allgather-gemm`](dist-allgather-gemm/) **2-GPU** | — | — | | | |
| | [`dist-allgather-gemm-overlap`](dist-allgather-gemm-overlap/) **2-GPU** | TFLOP/s | 1571 us | A sequence-parallel block starts with the mirror image of the reduce-scatter that ended the last one: each rank holds only its own slice of … | | |
| | [`dist-allreduce-rmsnorm`](dist-allreduce-rmsnorm/) **2-GPU** | GB/s | 14062 us | Every tensor-parallel transformer block ends the same way: the row-parallel projection leaves a PARTIAL sum on each rank, that partial has t… | | |
| | [`dist-moe-a2a-dispatch`](dist-moe-a2a-dispatch/) **2-GPU** | GB/s | 4688 us | Expert parallelism puts different experts on different GPUs, so after the router has spoken every token has to travel to whichever rank owns… | | |
| | [`dist-oneshot-allreduce`](dist-oneshot-allreduce/) **2-GPU** | GB/s | 3750 us | At the message sizes a tensor-parallel decoder actually produces — a few MB to a few tens of MB — NCCL's ring all-reduce is the wrong algori… | | |
| | [`dist-reduce-scatter-rmsnorm`](dist-reduce-scatter-rmsnorm/) **2-GPU** | GB/s | 8438 us | Sequence-parallel tensor parallelism replaces the block-ending all-reduce with a REDUCE-SCATTER: the row-parallel projection leaves a full-s… | | |
| | [`dist-tp-embedding-allreduce`](dist-tp-embedding-allreduce/) **2-GPU** | GB/s | 14062 us | The embedding table is the one weight too large to replicate, so it is sharded over the VOCABULARY: each rank holds a slice of the rows and … | | |
| | [`dist-tp-fused-kernels`](dist-tp-fused-kernels/) **2-GPU** | — | — | | | |