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
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90

# scenario batched_prefill_waves
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_for(budget=24) -> 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
alloc_sequence_for(budget=20) -> slot 1
adapter_id_for(-1)
acquire_adapter_slot(-1)
ep_broadcast_cmd_for_seq(seq=1, cmd=0xfffffff0)
ep_broadcast_cmd(0x8)
ep_broadcast_cmd(0x0)
ep_broadcast_cmd(0x11)
ep_broadcast_tokens(n=17)
ep_sync_vision_embeds(n=17)
prefill_chunk(start=0, len=8, last=false, s2@1:len=8, stream=1)
record_event(event=2, stream=1)
stream_wait_event(stream=0, event=2)
poll_innerq()
mixed_forward(decode=[10], prefill=s2@1:len=16 start=8 len=8 last=false, stream=1)
record_event(event=2, stream=1)
stream_wait_event(stream=0, event=2)
argmax_batch(row=0, n=1, stream=0) -> [11]
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s1@0:len=21)
num_free_blocks() -> 996
-- tick 2
poll_innerq()
mixed_forward(decode=[11], prefill=s2@1:len=17 start=16 len=1 last=true, stream=1)
record_event(event=2, stream=1)
stream_wait_event(stream=0, event=2)
argmax_on_device(row=1, stream=0) -> 20
record_event(event=2, stream=1)
stream_wait_event(stream=0, event=2)
argmax_batch(row=0, n=1, stream=0) -> [12]
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s1@0:len=22)
num_free_blocks() -> 996
-- tick 3
fx launch Decode{n=2, ctx=Unified, readback=Argmax}
decode_batch(tokens=[12, 20], seqs=["s1@0:len=22", "s2@1:len=17"], stream=0)
commit_ctx(s1@0:len=23, n=1, base=22, row=0)
commit_ctx(s2@1:len=18, n=1, base=17, row=1)
argmax_batch(row=0, n=2, stream=0) -> [63, 21]
fx apply MarconiCheckpoint
decode_marconi_checkpoint(s2@1:len=18)
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s1@0:len=23)
free_sequence(s1@0:len=23, blocks=2)
ep_broadcast_cmd_for_seq(seq=0, cmd=0xfffffff1)
fx apply CompactSlot{target=0}
compact_sequence(s2@1:len=18, new_slot=0)
num_free_blocks() -> 998
-- tick 4
fx launch Decode{n=1, ctx=Unified, readback=Argmax}
decode_batch(tokens=[21], seqs=["s2@0:len=18"], stream=0)
commit_ctx(s2@0:len=19, n=1, base=18, row=0)
argmax_batch(row=0, n=1, stream=0) -> [63]
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s2@0:len=19)
free_sequence(s2@0:len=19, 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 Done(stop, n=4, think=0, cached=0, acc=0, guard=None)
out s2: T20 T21 Done(stop, n=3, think=0, cached=0, acc=0, guard=None)