met 1.0.0-beta-preview

Metrale Engine: pure Rust LLM inference server (HTTP API, scheduler, TUI, CLI).
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64

# scenario chunked_prefill
bind_gpu_to_thread()
create_stream() -> 1
create_event() -> 2
adapter_id_for(-1)
adapter_id_for(-1)
num_free_blocks() -> 1000
-- tick 1
alloc_sequence_for(budget=25) -> slot 0
adapter_id_for(-1)
acquire_adapter_slot(-1)
ep_broadcast_cmd_for_seq(seq=0, cmd=0xfffffff0)
ep_broadcast_cmd(0x14)
ep_broadcast_cmd(0x0)
ep_broadcast_cmd(0x14)
ep_broadcast_tokens(n=20)
ep_sync_vision_embeds(n=20)
prefill_chunk(start=0, len=20, last=true, s1@0:len=20, stream=1)
record_event(event=2, stream=1)
stream_wait_event(stream=0, event=2)
argmax_on_device(row=0, stream=0) -> 10
fx launch Decode{n=1, ctx=Unified, readback=Argmax}
decode_batch(tokens=[10], seqs=["s1@0:len=20"], stream=0)
commit_ctx(s1@0:len=21, n=1, base=20, row=0)
argmax_batch(row=0, n=1, stream=0) -> [11]
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s1@0:len=21)
num_free_blocks() -> 998
-- tick 2
fx launch Decode{n=1, ctx=Unified, readback=Argmax}
decode_batch(tokens=[11], seqs=["s1@0:len=21"], stream=0)
commit_ctx(s1@0:len=22, n=1, base=21, row=0)
argmax_batch(row=0, n=1, stream=0) -> [12]
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s1@0:len=22)
num_free_blocks() -> 998
-- tick 3
fx launch Decode{n=1, ctx=Unified, readback=Argmax}
decode_batch(tokens=[12], seqs=["s1@0:len=22"], stream=0)
commit_ctx(s1@0:len=23, n=1, base=22, row=0)
argmax_batch(row=0, n=1, stream=0) -> [13]
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s1@0:len=23)
num_free_blocks() -> 998
-- tick 4
fx launch Decode{n=1, ctx=Unified, readback=Argmax}
decode_batch(tokens=[13], seqs=["s1@0:len=23"], stream=0)
commit_ctx(s1@0:len=24, n=1, base=23, row=0)
argmax_batch(row=0, n=1, stream=0) -> [63]
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s1@0:len=24)
free_sequence(s1@0:len=24, blocks=2)
ep_broadcast_cmd_for_seq(seq=0, 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 T11 T12 T13 Done(stop, n=5, think=0, cached=0, acc=0, guard=None)