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
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149

# scenario batched_verify_k4
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(0x6)
ep_broadcast_cmd(0x0)
ep_broadcast_cmd(0x6)
ep_broadcast_tokens(n=6)
ep_sync_vision_embeds(n=6)
prefill(n=6, s2@1:len=6, stream=0)
argmax_on_device(row=0, stream=0) -> 20
alloc_sequence() -> slot 2
adapter_id_for(-1)
acquire_adapter_slot(-1)
ep_broadcast_cmd_for_seq(seq=2, cmd=0xfffffff0)
ep_broadcast_cmd(0x7)
ep_broadcast_cmd(0x0)
ep_broadcast_cmd(0x7)
ep_broadcast_tokens(n=7)
ep_sync_vision_embeds(n=7)
prefill(n=7, s3@2:len=7, stream=0)
argmax_on_device(row=0, stream=0) -> 30
mtp_slot_draft_capacity(0)
mtp_slot_draft_capacity(1)
mtp_slot_draft_capacity(2)
sync_secondary()
can_batch_verify([2, 2, 2]) -> true
decode_batch(tokens=[10, 20, 30], seqs=["s1@0:len=5", "s2@1:len=6", "s3@2:len=7"], stream=0)
argmax_batch(row=0, n=3, stream=0) -> [11, 21, 31]
stash_verify_hidden_rows([0, 1, 2], stream=0)
run_mtp_propose_batched(tokens=[11, 21, 31], positions=[6, 7, 8], stash=[0, 1, 2], n=3, stream=0) -> [[12, 13, 14], [22, 24, 24], [32, 33, 34]]
start_checkpoint_async(s1@0:len=6)
start_checkpoint_async(s2@1:len=7)
start_checkpoint_async(s3@2:len=8)
num_free_blocks() -> 997
-- tick 2
mtp_slot_draft_capacity(0)
mtp_slot_draft_capacity(1)
mtp_slot_draft_capacity(2)
can_batch_verify([4, 4, 3]) -> true
sync_secondary()
decode_verify_batched(tokens=[11, 12, 13, 14, 21, 22, 24, 24, 31, 32, 33], ks=[4, 4, 3], seqs=["s1@0:len=10", "s2@1:len=11", "s3@2:len=11"], stream=0, woa=true) -> [12, 13, 14, 15, 22, 23, 1, 1, 32, 33, 34]
stash_verify_hidden_rows([3, 5, 10], stream=0)
stash_verify_catchup_rows([(0, 0), (1, 1), (2, 2), (4, 4), (8, 8), (9, 9)])
gdn_fold_accepted(slots=[0, 1, 2], accepted=[4, 2, 3], k=4) -> true
commit_accepted_prefix(s1@0:len=10, na=4, k=4)
trim_proposer_state(s1@0:len=10, na=3, stream=0)
trim_proposer_state(s2@1:len=9, na=1, stream=0)
commit_accepted_prefix(s2@1:len=9, na=2, k=4)
commit_accepted_prefix(s3@2:len=11, na=3, k=3)
trim_proposer_state(s3@2:len=11, na=2, stream=0)
run_mtp_catchup_batched(tokens=[[12, 13, 14], [22], [32, 33]], slots=[0, 4, 8], pos=[7, 8, 9]) -> 0
run_mtp_propose_batched(tokens=[15, 23, 34], positions=[10, 9, 11], stash=[0, 1, 2], n=3, stream=0) -> [[16, 17, 18], [24, 25, 27], [35, 36, 37]]
num_free_blocks() -> 997
-- tick 3
mtp_slot_draft_capacity(0)
mtp_slot_draft_capacity(1)
mtp_slot_draft_capacity(2)
can_batch_verify([4, 4, 3]) -> true
sync_secondary()
decode_verify_batched(tokens=[15, 16, 17, 18, 23, 24, 25, 27, 34, 35, 36], ks=[4, 4, 3], seqs=["s1@0:len=14", "s2@1:len=13", "s3@2:len=14"], stream=0, woa=true) -> [16, 17, 18, 19, 24, 25, 26, 1, 35, 36, 37]
stash_verify_hidden_rows([3, 6, 10], stream=0)
stash_verify_catchup_rows([(0, 0), (1, 1), (2, 2), (4, 4), (5, 5), (8, 8), (9, 9)])
gdn_fold_accepted(slots=[0, 1, 2], accepted=[4, 3, 3], k=4) -> true
commit_accepted_prefix(s1@0:len=14, na=4, k=4)
trim_proposer_state(s1@0:len=14, na=3, stream=0)
trim_proposer_state(s2@1:len=12, na=2, stream=0)
commit_accepted_prefix(s2@1:len=12, na=3, k=4)
commit_accepted_prefix(s3@2:len=14, na=3, k=3)
trim_proposer_state(s3@2:len=14, na=2, stream=0)
run_mtp_catchup_batched(tokens=[[16, 17, 18], [24, 25], [35, 36]], slots=[0, 4, 8], pos=[11, 10, 12]) -> 0
run_mtp_propose_batched(tokens=[19, 26, 37], positions=[14, 12, 14], stash=[0, 1, 2], n=3, stream=0) -> [[20, 21, 63], [27, 28, 30], [38, 39, 40]]
num_free_blocks() -> 997
-- tick 4
mtp_slot_draft_capacity(0)
mtp_slot_draft_capacity(1)
mtp_slot_draft_capacity(2)
can_batch_verify([4, 4, 3]) -> true
sync_secondary()
decode_verify_batched(tokens=[19, 20, 21, 63, 26, 27, 28, 30, 37, 38, 39], ks=[4, 4, 3], seqs=["s1@0:len=18", "s2@1:len=16", "s3@2:len=17"], stream=0, woa=true) -> [20, 21, 63, 63, 27, 28, 29, 1, 38, 39, 40]
stash_verify_hidden_rows([3, 6, 10], stream=0)
stash_verify_catchup_rows([(0, 0), (1, 1), (2, 2), (4, 4), (5, 5), (8, 8), (9, 9)])
gdn_fold_accepted(slots=[0, 1, 2], accepted=[4, 3, 3], k=4) -> true
trim_proposer_state(s2@1:len=15, na=2, stream=0)
commit_accepted_prefix(s2@1:len=15, na=3, k=4)
commit_accepted_prefix(s3@2:len=17, na=3, k=3)
trim_proposer_state(s3@2:len=17, na=2, stream=0)
run_mtp_catchup_batched(tokens=[[27, 28], [38, 39]], slots=[4, 8], pos=[13, 15]) -> 0
run_mtp_propose_batched(tokens=[29, 40], positions=[15, 17], stash=[1, 2], n=3, stream=0) -> [[30, 31, 1], [41, 63, 63]]
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s1@0:len=18)
free_sequence(s1@0:len=18, blocks=2)
ep_broadcast_cmd_for_seq(seq=0, cmd=0xfffffff1)
fx apply CompactSlot{target=0}
compact_sequence(s3@2:len=17, new_slot=0)
num_free_blocks() -> 996
-- tick 5
mtp_slot_draft_capacity(1)
mtp_slot_draft_capacity(0)
can_batch_verify([4, 3]) -> true
sync_secondary()
decode_verify_batched(tokens=[29, 30, 31, 1, 40, 41, 63], ks=[4, 3], seqs=["s2@1:len=19", "s3@0:len=20"], stream=0, woa=true) -> [30, 31, 63, 1, 41, 63, 63]
stash_verify_hidden_rows([2, 6], stream=0)
stash_verify_catchup_rows([(0, 0), (1, 1), (4, 4), (5, 5)])
gdn_fold_accepted(slots=[1, 0], accepted=[3, 3], k=4) -> true
trim_proposer_state(s2@1:len=18, na=2, stream=0)
commit_accepted_prefix(s2@1:len=18, na=3, k=4)
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s2@1:len=18)
free_sequence(s2@1:len=18, blocks=2)
ep_broadcast_cmd_for_seq(seq=1, cmd=0xfffffff1)
fx apply ReleaseSeq{cache=true, finish_sequence}
cache_sequence(s3@0:len=20)
free_sequence(s3@0:len=20, blocks=2)
ep_broadcast_cmd_for_seq(seq=0, cmd=0xfffffff1)
num_free_blocks() -> 1000
-- tick 6
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 T16 T17 T18 T19 T20 T21 Done(stop, n=13, think=0, cached=0, acc=9, guard=None)
out s2: T20 T21 T22 T23 T24 T25 T26 T27 T28 T29 T30 T31 Done(stop, n=13, think=0, cached=0, acc=7, guard=None)
out s3: T30 T31 T32 T33 T34 T35 T36 T37 T38 T39 T40 T41 Done(stop, n=13, think=0, cached=0, acc=8, guard=None)