# scenario host_logits_paths
bind_gpu_to_thread()
create_stream() -> 1
create_event() -> 2
adapter_id_for(-1)
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)
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, s1@0:len=5, stream=0)
argmax_on_device(row=0, stream=0) -> 10
alloc_sequence() -> slot 1
adapter_id_for(-1)
acquire_adapter_slot(-1)
ep_broadcast_cmd_for_seq(seq=1, 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@1:len=5, stream=0)
copy_logits_to_host(row=0, off=0, bytes=128)
alloc_sequence() -> slot 2
adapter_id_for(-1)
acquire_adapter_slot(-1)
ep_broadcast_cmd_for_seq(seq=2, 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, s3@2:len=5, stream=0)
copy_logits_to_host(row=0, off=0, bytes=128)
fx launch Decode{n=3, ctx=Unified, readback=HostLogits}
decode_batch(tokens=[10, 20, 30], seqs=["s1@0:len=5", "s2@1:len=5", "s3@2:len=5"], stream=0)
commit_ctx(s1@0:len=6, n=1, base=5, row=0)
commit_ctx(s2@1:len=6, n=1, base=5, row=1)
commit_ctx(s3@2:len=6, n=1, base=5, row=2)
copy_logits_to_host(row=0, off=0, bytes=384)
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s1@0:len=6)
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s2@1:len=6)
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s3@2:len=6)
num_free_blocks() -> 997
-- tick 2
fx launch Decode{n=3, ctx=Unified, readback=HostLogits}
decode_batch(tokens=[11, 21, 62], seqs=["s1@0:len=6", "s2@1:len=6", "s3@2:len=6"], stream=0)
commit_ctx(s1@0:len=7, n=1, base=6, row=0)
commit_ctx(s2@1:len=7, n=1, base=6, row=1)
commit_ctx(s3@2:len=7, n=1, base=6, row=2)
copy_logits_to_host(row=0, off=0, bytes=384)
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s1@0:len=7)
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s2@1:len=7)
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s3@2:len=7)
num_free_blocks() -> 997
-- tick 3
fx launch Decode{n=3, ctx=Unified, readback=HostLogits}
decode_batch(tokens=[12, 22, 31], seqs=["s1@0:len=7", "s2@1:len=7", "s3@2:len=7"], stream=0)
commit_ctx(s1@0:len=8, n=1, base=7, row=0)
commit_ctx(s2@1:len=8, n=1, base=7, row=1)
commit_ctx(s3@2:len=8, n=1, base=7, row=2)
copy_logits_to_host(row=0, off=0, bytes=384)
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s1@0:len=8)
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s2@1:len=8)
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s3@2:len=8)
num_free_blocks() -> 997
-- tick 4
fx launch Decode{n=3, ctx=Unified, readback=HostLogits}
decode_batch(tokens=[13, 23, 32], seqs=["s1@0:len=8", "s2@1:len=8", "s3@2:len=8"], stream=0)
commit_ctx(s1@0:len=9, n=1, base=8, row=0)
commit_ctx(s2@1:len=9, n=1, base=8, row=1)
commit_ctx(s3@2:len=9, n=1, base=8, row=2)
copy_logits_to_host(row=0, off=0, bytes=384)
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s1@0:len=9)
free_sequence(s1@0:len=9, blocks=1)
ep_broadcast_cmd_for_seq(seq=0, cmd=0xfffffff1)
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s2@1:len=9)
free_sequence(s2@1:len=9, blocks=1)
ep_broadcast_cmd_for_seq(seq=1, cmd=0xfffffff1)
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s3@2:len=9)
free_sequence(s3@2:len=9, blocks=1)
ep_broadcast_cmd_for_seq(seq=2, cmd=0xfffffff1)
num_free_blocks() -> 1000
-- tick 5
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: T10 TL11:2 TL12:2 TL13:2 Done(stop, n=5, think=0, cached=0, acc=0, guard=None)
out s2: T20 T21 T22 T23 Done(stop, n=5, think=0, cached=0, acc=0, guard=None)
out s3: T30 T62 T31 T32 Done(stop, n=5, think=0, cached=0, acc=0, guard=None)