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
91
92
93
94
95
96
97
98
99
100

# scenario cancel_mid_stream
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)
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)
argmax_on_device(row=0, stream=0) -> 20
mtp_slot_draft_capacity(0)
mtp_slot_draft_capacity(1)
sync_secondary()
can_batch_verify([2, 2]) -> true
decode_batch(tokens=[10, 20], seqs=["s1@0:len=5", "s2@1:len=5"], stream=0)
argmax_batch(row=0, n=2, stream=0) -> [11, 21]
stash_verify_hidden_rows([0, 1], stream=0)
run_mtp_propose_batched(tokens=[11, 21], positions=[6, 6], stash=[0, 1], n=1, stream=0) -> [[12], [22]]
start_checkpoint_async(s1@0:len=6)
start_checkpoint_async(s2@1:len=6)
num_free_blocks() -> 998
-- tick 2
mtp_slot_draft_capacity(0)
mtp_slot_draft_capacity(1)
can_batch_verify([2, 2]) -> true
sync_secondary()
decode_verify_batched(tokens=[11, 12, 21, 22], ks=[2, 2], seqs=["s1@0:len=8", "s2@1:len=8"], stream=0, woa=true) -> [12, 13, 22, 63]
stash_verify_hidden_rows([1, 3], stream=0)
stash_verify_catchup_rows([(0, 0), (4, 2)])
gdn_fold_accepted(slots=[0, 1], accepted=[2, 2], k=2) -> true
commit_accepted_prefix(s1@0:len=8, na=2, k=2)
trim_proposer_state(s1@0:len=8, na=1, stream=0)
run_mtp_catchup_batched(tokens=[[12]], slots=[0], pos=[7]) -> 0
save_hidden_for_mtp_from_stash(0, stream=0)
run_mtp_propose_multi(token=13, pos=8, n=1, s1@0:len=8, stream=0, mask=false) -> [14]
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s2@1:len=8)
free_sequence(s2@1:len=8, blocks=1)
ep_broadcast_cmd_for_seq(seq=1, cmd=0xfffffff1)
num_free_blocks() -> 999
-- tick 3
mtp_slot_draft_capacity(0)
sync_secondary()
ep_broadcast_cmd_for_seq(seq=0, cmd=0xfffffff2)
ep_broadcast_cmd(0xd)
ep_broadcast_cmd(0xe)
decode_verify_graphed(tokens=[13, 14], s1@0:len=10) -> [14, 15]
ep_broadcast_cmd(0x1)
commit_accepted_prefix(s1@0:len=10, na=2, k=2)
dflash_eagle_accept_append(s1@0:len=10)
save_hidden_for_mtp(1, stream=0)
trim_proposer_state(s1@0:len=10, na=1, stream=0)
run_mtp_propose_multi(token=15, pos=10, n=1, s1@0:len=10, stream=0, mask=false) -> [16]
decode_marconi_checkpoint(s1@0:len=10)
num_free_blocks() -> 999
-- tick 4
mtp_slot_draft_capacity(0)
sync_secondary()
ep_broadcast_cmd_for_seq(seq=0, cmd=0xfffffff2)
ep_broadcast_cmd(0xf)
ep_broadcast_cmd(0x10)
decode_verify_graphed(tokens=[15, 16], s1@0:len=12) -> [16, 17]
ep_broadcast_cmd(0x1)
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s1@0:len=12)
free_sequence(s1@0:len=12, blocks=1)
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 T14 T15 Done(stop, n=6, think=0, cached=0, acc=3, guard=None)
out s2: T20 T21 T22 Done(stop, n=4, think=0, cached=0, acc=1, guard=None)