xFasterTransformer
xFasterTransformer copied to clipboard
[Distribute] Add distribute support for continuous batching api.
Master and slaves should both run according to the following workflow:
while True:
model.set_input_cb()
model.forward_cb()
model.free_seqs()
Need to fix following Error after complete all requests in vllm:
Abort(68817167) on node 1 (rank 1 in comm 0): Fatal error in PMPI_Test: Other MPI error, error stack:
PMPI_Test(186)..................: MPI_Test(request=0x74a0f28, flag=0x7f3d67ffdd3c, status=0x1) failed
MPIR_Test(79)...................:
MPIDU_Sched_progress_state(1060): Invalid communicator
@pujiang2018 Ready.