1use std::sync::OnceLock;
4
5const SERVED_LOGIT_MAGIC: &[u8; 8] = b"MSIGRPL1";
6
7pub struct ServedLogitRecord {
8 pub layer: u32,
9 pub tokens: usize,
10 pub n_expert: usize,
11 pub n_used: usize,
12 pub scaling_factor: f32,
13 pub route_norm: bool,
14 pub active: Vec<u8>,
15 pub bias: Vec<f32>,
16 pub logits: Vec<f32>,
17}
18
19struct ServedLogitWriter {
20 path: std::path::PathBuf,
21 file: std::fs::File,
22 seen_layers: std::collections::HashSet<u32>,
23}
24
25static SERVED_LOGIT_WRITER: OnceLock<std::sync::Mutex<Option<ServedLogitWriter>>> = OnceLock::new();
26
27pub fn validate_active_count(n_used: usize, active_count: usize) -> Result<(), String> {
29 if active_count < n_used {
30 return Err(format!(
31 "sigmoid router requires active_count >= n_used: active_count={active_count}, n_used={n_used}",
32 ));
33 }
34 Ok(())
35}
36
37pub fn verify_host_expf() -> Result<(), String> {
39 static RESULT: OnceLock<Result<(), String>> = OnceLock::new();
40 RESULT
41 .get_or_init(|| {
42 const CASES: &[(u32, u32)] = &[
43 (0xc0f4_bca4, 0x39fa_13b9), (0x4152_4c00, 0x48f9_5e97),
44 (0xc100_cdee, 0x39a7_4153), (0xc07a_9e68, 0x3ca3_33fa),
45 (0x412b_32ca, 0x472d_3f68), (0xc153_f532, 0x35ec_e4cd),
46 (0x4031_a0d8, 0x4180_5da2), (0xc0e3_fbc0, 0x3a53_10be),
47 (0x4141_525e, 0x482c_a0b2), (0x40f0_a9a8, 0x44e6_bc14),
48 (0xc158_4cca, 0x35b4_96ea), (0xc123_1dda, 0x381c_b7e3),
49 (0x3f72_ed60, 0x4025_4f28), (0x4179_f55a, 0x4ab9_e5a0),
50 (0x415c_15c0, 0x4965_e07c), (0x3fa0_5800, 0x405f_fb8f),
51 (0x416d_e91a, 0x4a2f_193b), (0x417c_082a, 0x4ad3_9e5b),
52 (0x4133_a8f6, 0x4792_ffbd), (0xc17b_4c4a, 0x3422_1cc9),
53 (0x4049_f238, 0x41bb_b376), (0x4122_c18a, 0x46cc_6dd0),
54 (0x40c3_8904, 0x43e1_46c6), (0x411f_3fd2, 0x46a4_31c2),
55 ];
56 for (case, &(input_bits, expected_bits)) in CASES.iter().enumerate() {
57 let input = std::hint::black_box(f32::from_bits(input_bits));
58 let actual_bits = input.exp().to_bits();
59 if actual_bits != expected_bits {
60 return Err(format!(
61 "host expf byte probe mismatch at case {case}: input=0x{input_bits:08x}, expected=0x{expected_bits:08x}, actual=0x{actual_bits:08x}",
62 ));
63 }
64 }
65 Ok(())
66 })
67 .clone()
68}
69
70pub fn served_logit_trace_enabled() -> bool {
71 std::env::var_os("MEMRA_SIG_ROUTER_LOGIT_TRACE").is_some()
72}
73
74#[allow(clippy::too_many_arguments)]
77pub fn capture_served_logits(
78 layer: u32,
79 tokens: usize,
80 n_expert: usize,
81 n_used: usize,
82 scaling_factor: f32,
83 route_norm: bool,
84 active: &[u8],
85 bias: &[f32],
86 logits: &[f32],
87) -> Result<(), String> {
88 use std::io::Write as _;
89
90 let Some(path) = std::env::var_os("MEMRA_SIG_ROUTER_LOGIT_TRACE") else {
91 return Ok(());
92 };
93 if tokens != 1 {
94 return Ok(());
95 }
96 if active.len() != n_expert || bias.len() != n_expert || logits.len() < n_expert {
97 return Err(format!(
98 "served sigmoid-logit trace shape mismatch at layer {layer}: active={} bias={} logits={} n_expert={n_expert}",
99 active.len(),
100 bias.len(),
101 logits.len(),
102 ));
103 }
104 let path = std::path::PathBuf::from(path);
105 let state = SERVED_LOGIT_WRITER.get_or_init(|| std::sync::Mutex::new(None));
106 let mut state = state
107 .lock()
108 .map_err(|_| "served sigmoid-logit trace writer lock is poisoned".to_string())?;
109 if state.is_none() {
110 let mut file = std::fs::OpenOptions::new()
111 .create(true)
112 .truncate(true)
113 .write(true)
114 .open(&path)
115 .map_err(|error| {
116 format!(
117 "cannot create served sigmoid-logit trace {}: {error}",
118 path.display()
119 )
120 })?;
121 file.write_all(SERVED_LOGIT_MAGIC)
122 .map_err(|error| format!("cannot write served sigmoid-logit trace header: {error}"))?;
123 *state = Some(ServedLogitWriter {
124 path: path.clone(),
125 file,
126 seen_layers: std::collections::HashSet::new(),
127 });
128 }
129 let writer = state.as_mut().unwrap();
130 if writer.path != path {
131 return Err("MEMRA_SIG_ROUTER_LOGIT_TRACE changed after capture started".into());
132 }
133 if !writer.seen_layers.insert(layer) {
134 return Ok(());
135 }
136
137 for value in [
138 layer,
139 tokens as u32,
140 n_expert as u32,
141 n_used as u32,
142 scaling_factor.to_bits(),
143 u32::from(route_norm),
144 ] {
145 writer
146 .file
147 .write_all(&value.to_le_bytes())
148 .map_err(|error| format!("cannot write served sigmoid-logit trace row: {error}"))?;
149 }
150 writer
151 .file
152 .write_all(active)
153 .map_err(|error| format!("cannot write served sigmoid-logit active mask: {error}"))?;
154 for value in bias.iter().chain(logits[..n_expert].iter()) {
155 writer
156 .file
157 .write_all(&value.to_bits().to_le_bytes())
158 .map_err(|error| format!("cannot write served sigmoid-logit f32 row: {error}"))?;
159 }
160 writer
161 .file
162 .flush()
163 .map_err(|error| format!("cannot flush served sigmoid-logit trace: {error}"))?;
164 Ok(())
165}
166
167fn read_u32(bytes: &[u8], cursor: &mut usize) -> Result<u32, String> {
168 let end = cursor.saturating_add(4);
169 let raw: [u8; 4] = bytes
170 .get(*cursor..end)
171 .ok_or_else(|| "truncated served sigmoid-logit trace".to_string())?
172 .try_into()
173 .unwrap();
174 *cursor = end;
175 Ok(u32::from_le_bytes(raw))
176}
177
178pub fn read_served_logits(path: &std::path::Path) -> Result<Vec<ServedLogitRecord>, String> {
179 let bytes = std::fs::read(path).map_err(|error| {
180 format!(
181 "cannot read served sigmoid-logit trace {}: {error}",
182 path.display()
183 )
184 })?;
185 if bytes.get(..SERVED_LOGIT_MAGIC.len()) != Some(SERVED_LOGIT_MAGIC) {
186 return Err("served sigmoid-logit trace has wrong or missing v1 header".into());
187 }
188 let mut cursor = SERVED_LOGIT_MAGIC.len();
189 let mut records = Vec::new();
190 while cursor < bytes.len() {
191 let layer = read_u32(&bytes, &mut cursor)?;
192 let tokens = read_u32(&bytes, &mut cursor)? as usize;
193 let n_expert = read_u32(&bytes, &mut cursor)? as usize;
194 let n_used = read_u32(&bytes, &mut cursor)? as usize;
195 let scaling_factor = f32::from_bits(read_u32(&bytes, &mut cursor)?);
196 let route_norm = match read_u32(&bytes, &mut cursor)? {
197 0 => false,
198 1 => true,
199 value => {
200 return Err(format!(
201 "invalid route_norm={value} in served sigmoid-logit trace"
202 ));
203 }
204 };
205 if tokens != 1 || n_expert == 0 || n_used == 0 || n_expert > 1024 {
206 return Err(format!(
207 "invalid served sigmoid-logit record shape: layer={layer} tokens={tokens} n_expert={n_expert} n_used={n_used}",
208 ));
209 }
210 let active_end = cursor.saturating_add(n_expert);
211 let active = bytes
212 .get(cursor..active_end)
213 .ok_or_else(|| "truncated served sigmoid-logit active mask".to_string())?
214 .to_vec();
215 cursor = active_end;
216 let mut read_f32_row = || -> Result<Vec<f32>, String> {
217 (0..n_expert)
218 .map(|_| read_u32(&bytes, &mut cursor).map(f32::from_bits))
219 .collect()
220 };
221 let bias = read_f32_row()?;
222 let logits = read_f32_row()?;
223 records.push(ServedLogitRecord {
224 layer,
225 tokens,
226 n_expert,
227 n_used,
228 scaling_factor,
229 route_norm,
230 active,
231 bias,
232 logits,
233 });
234 }
235 Ok(records)
236}
237
238#[cfg(test)]
239mod tests {
240 use super::*;
241
242 #[test]
243 fn active_count_error_quotes_both_counts() {
244 assert_eq!(
245 validate_active_count(8, 7).unwrap_err(),
246 "sigmoid router requires active_count >= n_used: active_count=7, n_used=8",
247 );
248 assert!(validate_active_count(8, 8).is_ok());
249 }
250
251 #[test]
252 fn pinned_host_expf_matches() {
253 verify_host_expf().unwrap();
254 }
255}