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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
pub(crate) mod debug;
pub(crate) mod gguf_metadata;
pub(crate) mod memory_usage;
pub(crate) mod model_config;
pub(crate) mod normal;
pub(crate) mod progress;
pub(crate) mod tiktoken;
pub(crate) mod tokenizer;
pub(crate) mod tokens;
pub(crate) mod unvarbuilder;
pub(crate) mod varbuilder_utils;
#[doc(hidden)]
#[macro_export]
macro_rules! get_mut_arcmutex {
($thing:expr) => {
loop {
if let Ok(inner) = $thing.try_lock() {
break inner;
}
// Yield to allow other threads to make progress and release the lock.
// This prevents deadlock when a spawned async task busy-loops while
// another task holds the lock across an await point.
std::thread::yield_now();
}
};
}
#[doc(hidden)]
#[macro_export]
macro_rules! handle_seq_error {
($fallible:expr, $response:expr) => {
match $fallible {
Ok(v) => v,
Err(e) => {
use $crate::response::Response;
if let Err(_) = $response.send(Response::InternalError(e.into())).await {
tracing::warn!("Receiver disconnected");
}
return;
}
}
};
}
#[doc(hidden)]
#[macro_export]
macro_rules! handle_seq_error_ok {
($fallible:expr, $response:expr) => {
match $fallible {
Ok(v) => v,
Err(e) => {
use $crate::response::Response;
if let Err(_) = $response.send(Response::InternalError(e.into())).await {
tracing::warn!("Receiver disconnected");
}
return Ok(());
}
}
};
}
#[doc(hidden)]
#[macro_export]
macro_rules! handle_seq_error_stateaware_ok {
($fallible:expr, $seq:expr) => {
match $fallible {
Ok(v) => v,
Err(e) => {
use $crate::response::Response;
use $crate::sequence::SequenceState;
if let Err(_) = $seq
.responder()
.send(Response::InternalError(e.into()))
.await
{
tracing::warn!("Receiver disconnected");
}
$seq.set_state(SequenceState::Error);
return Ok(());
}
}
};
}
#[doc(hidden)]
#[macro_export]
macro_rules! handle_pipeline_forward_error {
($stage: tt, $fallible:expr, $seq_slice:expr, $pipeline:expr, $label:tt, $prefix_cacher:expr) => {
match $fallible {
Ok(v) => v,
Err(e) => {
// Auto-retry on iOS Metal background GPU error: when the iOS app
// goes to background, Metal rejects command buffers. We detect this,
// reset cache, sleep, and let the engine loop retry. Sequences stay
// in the scheduler (still in Running state) and are re-scheduled.
#[cfg(feature = "metal")]
{
let err_str = e.to_string();
if err_str.contains("Insufficient Permission")
|| err_str.contains("BackgroundExecutionNotPermitted")
{
tracing::warn!(
"Metal GPU background error detected (iOS app likely in background). \
Pausing 1s before retry..."
);
{
let p = get_mut_arcmutex!($pipeline);
p.set_none_cache($seq_slice, true, true, false);
}
get_mut_arcmutex!($prefix_cacher).evict_all_caches().unwrap();
tokio::time::sleep(std::time::Duration::from_secs(1)).await;
continue $label;
}
}
let (tokenizer, pipeline_name) = {
let pipeline = get_mut_arcmutex!($pipeline);
let pipeline_name = pipeline.name();
let tokenizer = pipeline.tokenizer();
(tokenizer, pipeline_name)
};
use $crate::response::Response;
use $crate::sequence::SequenceState;
use $crate::response::SYSTEM_FINGERPRINT;
use tracing::error;
error!("{} - Model failed with error: {:?}", $stage, &e);
for seq in $seq_slice.iter_mut() {
// Step 1: Add all choices to groups
let start = seq.prompt_tokens().min(seq.get_toks().len());
let res = match &tokenizer {
Some(tok) => match tok.decode(&seq.get_toks()[start..], false) {
Ok(t) => t,
Err(_) => "".to_string(),
},
None => "".to_string(),
};
if seq.get_mut_group().is_chat {
let choice = Choice {
finish_reason: "error".to_string(),
index: seq.get_response_index(),
message: ResponseMessage {
content: Some(res),
role: "assistant".to_string(),
tool_calls: None,
reasoning_content: None,
},
logprobs: None,
};
seq.add_choice_to_group(choice);
} else {
let choice = CompletionChoice {
finish_reason: "error".to_string(),
index: seq.get_response_index(),
text: res,
logprobs: None,
};
seq.add_completion_choice_to_group(choice);
}
}
for seq in $seq_slice.iter_mut() {
// Step 2: Respond with all groups
let group = seq.get_mut_group();
if group.is_chat {
let partial_completion_response = ChatCompletionResponse {
id: seq.id().to_string(),
choices: group.get_choices().to_vec(),
created: seq.creation_time(),
model: pipeline_name.clone(),
system_fingerprint: SYSTEM_FINGERPRINT.to_string(),
object: "chat.completion".to_string(),
usage: group.get_usage(),
};
seq.responder()
.send(Response::ModelError(
e.to_string(),
partial_completion_response
))
.await
.unwrap();
} else {
let partial_completion_response = CompletionResponse {
id: seq.id().to_string(),
choices: group.get_completion_choices().to_vec(),
created: seq.creation_time(),
model: pipeline_name.clone(),
system_fingerprint: SYSTEM_FINGERPRINT.to_string(),
object: "text_completion".to_string(),
usage: group.get_usage(),
};
seq.responder()
.send(Response::CompletionModelError(
e.to_string(),
partial_completion_response
))
.await
.unwrap();
}
}
for seq in $seq_slice.iter_mut() {
// Step 3: Set state - This cannot be done in Step 2 as `group` is locking the refcell
seq.set_state(SequenceState::Error);
}
let p = get_mut_arcmutex!($pipeline);
// Also reset non granular state because:
// - The sequence is gone
// - We should reset the state then, including draft.
p.set_none_cache($seq_slice, true, true, false);
get_mut_arcmutex!($prefix_cacher).evict_all_caches().unwrap();
continue $label;
}
}
};
}
#[doc(hidden)]
#[macro_export]
macro_rules! get_mut_group {
($this:expr) => {
loop {
if let Ok(inner) = $this.group.try_lock() {
break inner;
}
// Yield to allow other threads to make progress and release the lock.
std::thread::yield_now();
}
};
}
#[doc(hidden)]
#[macro_export]
macro_rules! serde_default_fn {
($t:ty, $name:ident, $v:expr) => {
fn $name() -> $t {
$v
}
};
}
/// `true` if built with CUDA (requires Unix) /Metal
#[cfg(any(all(feature = "cuda", target_family = "unix"), feature = "metal"))]
pub const fn paged_attn_supported() -> bool {
true
}
/// `true` if built with CUDA (requires Unix) /Metal
#[cfg(not(any(all(feature = "cuda", target_family = "unix"), feature = "metal")))]
pub const fn paged_attn_supported() -> bool {
false
}
/// `true` if built with the `flash-attn` or `flash-attn-v3` features, false otherwise.
#[cfg(not(any(feature = "flash-attn", feature = "flash-attn-v3")))]
pub const fn using_flash_attn() -> bool {
false
}
/// `true` if built with the `flash-attn` or `flash-attn-v3` features, false otherwise.
#[cfg(any(feature = "flash-attn", feature = "flash-attn-v3"))]
pub const fn using_flash_attn() -> bool {
true
}