# scenario beam
bind_gpu_to_thread()
create_stream() -> 1
create_event() -> 2
adapter_id_for(-1)
adapter_id_for(-1)
adapter_id_for(-1)
num_free_blocks() -> 1000
-- tick 1
alloc_sequence() -> slot 0
adapter_id_for(-1)
acquire_adapter_slot(-1)
generate_beam_batch(n=1) -> [[10, 11, 12, 63]]
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s1@0:len=0)
free_sequence(s1@0:len=0, blocks=0)
ep_broadcast_cmd_for_seq(seq=0, cmd=0xfffffff1)
alloc_sequence() -> slot 0
adapter_id_for(-1)
acquire_adapter_slot(-1)
ep_broadcast_cmd_for_seq(seq=0, cmd=0xfffffff0)
ep_broadcast_cmd(0x5)
ep_broadcast_cmd(0x0)
ep_broadcast_cmd(0x5)
ep_broadcast_tokens(n=5)
ep_sync_vision_embeds(n=5)
prefill(n=5, s2@0:len=5, stream=0)
argmax_on_device(row=0, stream=0) -> 20
fx launch Decode{n=1, ctx=Unified, readback=Argmax}
decode_batch(tokens=[20], seqs=["s2@0:len=5"], stream=0)
commit_ctx(s2@0:len=6, n=1, base=5, row=0)
argmax_batch(row=0, n=1, stream=0) -> [21]
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s2@0:len=6)
num_free_blocks() -> 999
-- tick 2
fx launch Decode{n=1, ctx=Unified, readback=Argmax}
decode_batch(tokens=[21], seqs=["s2@0:len=6"], stream=0)
commit_ctx(s2@0:len=7, n=1, base=6, row=0)
argmax_batch(row=0, n=1, stream=0) -> [63]
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s2@0:len=7)
free_sequence(s2@0:len=7, blocks=1)
ep_broadcast_cmd_for_seq(seq=0, cmd=0xfffffff1)
num_free_blocks() -> 1000
-- tick 3
fx apply EpShutdown
ep_broadcast_cmd_for_seq(seq=0, cmd=0xffffffff)
fx apply Quiesce{stream=0}
synchronize(stream=0)
fx apply Quiesce{stream=1}
synchronize(stream=1)
fx teardown
teardown()
out s1: Response(tokens=[10, 11, 12, 63], stop, think=0, cached=0, acc=0, logprobs=0)
out s2: T20 T21 Done(stop, n=3, think=0, cached=0, acc=0, guard=None)