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>>> =
26 OnceLock::new();
27
28pub fn validate_active_count(n_used: usize, active_count: usize) -> Result<(), String> {
30 if active_count < n_used {
31 return Err(format!(
32 "sigmoid router requires active_count >= n_used: active_count={active_count}, n_used={n_used}",
33 ));
34 }
35 Ok(())
36}
37
38pub fn verify_host_expf() -> Result<(), String> {
40 static RESULT: OnceLock<Result<(), String>> = OnceLock::new();
41 RESULT
42 .get_or_init(|| {
43 const CASES: &[(u32, u32)] = &[
44 (0xc0f4_bca4, 0x39fa_13b9), (0x4152_4c00, 0x48f9_5e97),
45 (0xc100_cdee, 0x39a7_4153), (0xc07a_9e68, 0x3ca3_33fa),
46 (0x412b_32ca, 0x472d_3f68), (0xc153_f532, 0x35ec_e4cd),
47 (0x4031_a0d8, 0x4180_5da2), (0xc0e3_fbc0, 0x3a53_10be),
48 (0x4141_525e, 0x482c_a0b2), (0x40f0_a9a8, 0x44e6_bc14),
49 (0xc158_4cca, 0x35b4_96ea), (0xc123_1dda, 0x381c_b7e3),
50 (0x3f72_ed60, 0x4025_4f28), (0x4179_f55a, 0x4ab9_e5a0),
51 (0x415c_15c0, 0x4965_e07c), (0x3fa0_5800, 0x405f_fb8f),
52 (0x416d_e91a, 0x4a2f_193b), (0x417c_082a, 0x4ad3_9e5b),
53 (0x4133_a8f6, 0x4792_ffbd), (0xc17b_4c4a, 0x3422_1cc9),
54 (0x4049_f238, 0x41bb_b376), (0x4122_c18a, 0x46cc_6dd0),
55 (0x40c3_8904, 0x43e1_46c6), (0x411f_3fd2, 0x46a4_31c2),
56 ];
57 for (case, &(input_bits, expected_bits)) in CASES.iter().enumerate() {
58 let input = std::hint::black_box(f32::from_bits(input_bits));
59 let actual_bits = input.exp().to_bits();
60 if actual_bits != expected_bits {
61 return Err(format!(
62 "host expf byte probe mismatch at case {case}: input=0x{input_bits:08x}, expected=0x{expected_bits:08x}, actual=0x{actual_bits:08x}",
63 ));
64 }
65 }
66 Ok(())
67 })
68 .clone()
69}
70
71pub fn served_logit_trace_enabled() -> bool {
72 std::env::var_os("MEMRA_SIG_ROUTER_LOGIT_TRACE").is_some()
73}
74
75#[allow(clippy::too_many_arguments)]
78pub fn capture_served_logits(
79 layer: u32,
80 tokens: usize,
81 n_expert: usize,
82 n_used: usize,
83 scaling_factor: f32,
84 route_norm: bool,
85 active: &[u8],
86 bias: &[f32],
87 logits: &[f32],
88) -> Result<(), String> {
89 use std::io::Write as _;
90
91 let Some(path) = std::env::var_os("MEMRA_SIG_ROUTER_LOGIT_TRACE") else {
92 return Ok(());
93 };
94 if tokens != 1 {
95 return Ok(());
96 }
97 if active.len() != n_expert || bias.len() != n_expert || logits.len() < n_expert {
98 return Err(format!(
99 "served sigmoid-logit trace shape mismatch at layer {layer}: active={} bias={} logits={} n_expert={n_expert}",
100 active.len(), bias.len(), logits.len(),
101 ));
102 }
103 let path = std::path::PathBuf::from(path);
104 let state = SERVED_LOGIT_WRITER.get_or_init(|| std::sync::Mutex::new(None));
105 let mut state = state
106 .lock()
107 .map_err(|_| "served sigmoid-logit trace writer lock is poisoned".to_string())?;
108 if state.is_none() {
109 let mut file = std::fs::OpenOptions::new()
110 .create(true)
111 .truncate(true)
112 .write(true)
113 .open(&path)
114 .map_err(|error| format!("cannot create served sigmoid-logit trace {}: {error}", path.display()))?;
115 file.write_all(SERVED_LOGIT_MAGIC)
116 .map_err(|error| format!("cannot write served sigmoid-logit trace header: {error}"))?;
117 *state = Some(ServedLogitWriter {
118 path: path.clone(),
119 file,
120 seen_layers: std::collections::HashSet::new(),
121 });
122 }
123 let writer = state.as_mut().unwrap();
124 if writer.path != path {
125 return Err("MEMRA_SIG_ROUTER_LOGIT_TRACE changed after capture started".into());
126 }
127 if !writer.seen_layers.insert(layer) {
128 return Ok(());
129 }
130
131 for value in [
132 layer,
133 tokens as u32,
134 n_expert as u32,
135 n_used as u32,
136 scaling_factor.to_bits(),
137 u32::from(route_norm),
138 ] {
139 writer.file.write_all(&value.to_le_bytes())
140 .map_err(|error| format!("cannot write served sigmoid-logit trace row: {error}"))?;
141 }
142 writer.file.write_all(active)
143 .map_err(|error| format!("cannot write served sigmoid-logit active mask: {error}"))?;
144 for value in bias.iter().chain(logits[..n_expert].iter()) {
145 writer.file.write_all(&value.to_bits().to_le_bytes())
146 .map_err(|error| format!("cannot write served sigmoid-logit f32 row: {error}"))?;
147 }
148 writer.file.flush()
149 .map_err(|error| format!("cannot flush served sigmoid-logit trace: {error}"))?;
150 Ok(())
151}
152
153fn read_u32(bytes: &[u8], cursor: &mut usize) -> Result<u32, String> {
154 let end = cursor.saturating_add(4);
155 let raw: [u8; 4] = bytes
156 .get(*cursor..end)
157 .ok_or_else(|| "truncated served sigmoid-logit trace".to_string())?
158 .try_into()
159 .unwrap();
160 *cursor = end;
161 Ok(u32::from_le_bytes(raw))
162}
163
164pub fn read_served_logits(path: &std::path::Path) -> Result<Vec<ServedLogitRecord>, String> {
165 let bytes = std::fs::read(path)
166 .map_err(|error| format!("cannot read served sigmoid-logit trace {}: {error}", path.display()))?;
167 if bytes.get(..SERVED_LOGIT_MAGIC.len()) != Some(SERVED_LOGIT_MAGIC) {
168 return Err("served sigmoid-logit trace has wrong or missing v1 header".into());
169 }
170 let mut cursor = SERVED_LOGIT_MAGIC.len();
171 let mut records = Vec::new();
172 while cursor < bytes.len() {
173 let layer = read_u32(&bytes, &mut cursor)?;
174 let tokens = read_u32(&bytes, &mut cursor)? as usize;
175 let n_expert = read_u32(&bytes, &mut cursor)? as usize;
176 let n_used = read_u32(&bytes, &mut cursor)? as usize;
177 let scaling_factor = f32::from_bits(read_u32(&bytes, &mut cursor)?);
178 let route_norm = match read_u32(&bytes, &mut cursor)? {
179 0 => false,
180 1 => true,
181 value => return Err(format!("invalid route_norm={value} in served sigmoid-logit trace")),
182 };
183 if tokens != 1 || n_expert == 0 || n_used == 0 || n_expert > 1024 {
184 return Err(format!(
185 "invalid served sigmoid-logit record shape: layer={layer} tokens={tokens} n_expert={n_expert} n_used={n_used}",
186 ));
187 }
188 let active_end = cursor.saturating_add(n_expert);
189 let active = bytes
190 .get(cursor..active_end)
191 .ok_or_else(|| "truncated served sigmoid-logit active mask".to_string())?
192 .to_vec();
193 cursor = active_end;
194 let mut read_f32_row = || -> Result<Vec<f32>, String> {
195 (0..n_expert)
196 .map(|_| read_u32(&bytes, &mut cursor).map(f32::from_bits))
197 .collect()
198 };
199 let bias = read_f32_row()?;
200 let logits = read_f32_row()?;
201 records.push(ServedLogitRecord {
202 layer,
203 tokens,
204 n_expert,
205 n_used,
206 scaling_factor,
207 route_norm,
208 active,
209 bias,
210 logits,
211 });
212 }
213 Ok(records)
214}
215
216#[cfg(test)]
217mod tests {
218 use super::*;
219
220 #[test]
221 fn active_count_error_quotes_both_counts() {
222 assert_eq!(
223 validate_active_count(8, 7).unwrap_err(),
224 "sigmoid router requires active_count >= n_used: active_count=7, n_used=8",
225 );
226 assert!(validate_active_count(8, 8).is_ok());
227 }
228
229 #[test]
230 fn pinned_host_expf_matches() {
231 verify_host_expf().unwrap();
232 }
233}