flashinfer.comm.all_gather_matmul

flashinfer.comm.all_gather_matmul(inp: Tensor, w: Tensor, group: ProcessGroup, *, backend: str = 'auto', verbose: bool = False)

Run push-wait all-gather matmul.

backend="auto" preserves the existing cuTile-on-Blackwell and Triton-otherwise routing. backend="cake" selects the exact source-built Blackwell implementation and rejects unsupported inputs.