cortiq_engine/nystrom.rs
1//! Nyström (landmark) attention kernel — streaming per-GQA-group runtime
2//! for long-context `attn_type: nystrom` layers.
3//!
4//! Attention splits into an EXACT sliding window (last `w` keys) and a
5//! landmark-skeleton far field sharing ONE joint denominator:
6//!
7//! ```text
8//! out(q_t) = (Σ_{j>t-w} e_j·v_j + F·M·T_far) / (Σ_{j>t-w} e_j + F·M·Z_far)
9//! e_j = exp(q_t·k_j/√d) exact near weights
10//! F_i = exp(q_t·k̃_i/√d) scores vs landmark keys
11//! M = pinv_reg(exp(Q̃·K̃ᵀ/√d)) fixed after prefill
12//! T_far = Σ_{j≤t-w} exp(Q̃·k_j/√d)·v_jᵀ [m × dv]
13//! Z_far = Σ_{j≤t-w} exp(Q̃·k_j/√d) [m]
14//! ```
15//!
16//! exp(q·k) is a PSD kernel, so the UNNORMALIZED skeleton (classic
17//! Nyström/CUR) is legal. Do NOT row-softmax the factors and do NOT
18//! normalize the key scores over landmarks — both "simplifications"
19//! measurably collapse quality (validated in the torch matrix probes).
20//!
21//! Boundary discipline: key j enters T/Z at the exact step it LEAVES
22//! the window (t = j+w) — delayed insertion, no overlap, no hole; the
23//! near mass stays exact rather than Nyström-estimated.
24//!
25//! Sink tokens (spec §5b, StreamingLLM discipline): the first `sink`
26//! keys of the sequence are PERMANENT exact keys — the near mask is
27//! (t-j < w) OR (j < sink) — and must never enter the far accumulators.
28//! Here they never enter the ring window in the first place (they live
29//! in a dedicated buffer), so delayed insertion cannot see them: no
30//! double count, no gap. Measured: sinks make the full 28/28-layer
31//! O(1) conversion viable — the default mode.
32//!
33//! Quality of THIS kernel, measured through it (`cortiq ppl --o1 all`,
34//! Qwen3-0.6B, all 28 layers, m=32 W=128 sink=4, wikitext-2 val, 12×512
35//! windows, landmarks frozen at a 256-token prefill): ×1.296 vs exact
36//! attention over the same scored tokens (28.04 vs 21.63).
37//!
38//! The older ×1.177 figure is NOT this operator: it comes from the torch
39//! matrix probe, which (a) rectifies every per-(t,j) weight — impossible
40//! to stream, the weights are never materialized — (b) builds landmarks
41//! from the FULL sequence rather than the prefill, and (c) averages in
42//! the first W positions, which are pure-exact and cost nothing. Quote
43//! ×1.296 for the runtime; ×1.177 is an upper bound the runtime cannot
44//! reach by construction.
45//!
46//! fp32 numerics: raw exp overflows on real logits, so shifts are
47//! absorbed into diagonals. T̂[i]/Ẑ[i] live at scale e^{-m_i} with a
48//! per-landmark running max m_i (flash-style rescale on growth); each
49//! token's landmark row uses its own shift f; near and far are brought
50//! to one common scale before the single joint division.
51
52/// Which rectifier keeps the skeleton's estimated far mass non-negative.
53///
54/// pinv(exp(Q̃K̃ᵀ/√d)) is violently ill-conditioned, so M is indefinite
55/// and the raw skeleton estimates negative weights for a large minority
56/// of keys (measured on Qwen3-0.6B: 23.5% of far weights negative,
57/// carrying 24.5% of the absolute far mass). Unrectified, the joint
58/// denominator goes near-zero/negative and the model collapses (×510).
59///
60/// The matrix probe rectifies every estimated weight — `west =
61/// ((Fu@Mu)@E).clamp_min(0)`. A STREAMING kernel cannot do that: the
62/// per-(t, j) weights are never materialized, they exist only already
63/// contracted against the accumulators. Two streaming-legal stand-ins:
64///
65/// MEASURED (Qwen3-0.6B, all 28 layers, W=128, sink=4, wikitext-2 val,
66/// 12×512 windows, landmarks frozen at a 256-token prefill — i.e. the
67/// runtime's real discipline, `cortiq ppl --o1`):
68///
69/// ```text
70/// m=8 m=16 m=32
71/// agg 28.51 (×1.318) 28.82 (×1.332) 28.04 (×1.296) ← default
72/// fm 28.97 (×1.340) 29.69 (×1.373) 30.58 (×1.414)
73/// ```
74///
75/// `Aggregate` wins at every m, so it stays the default. `Fm` is kept
76/// selectable because its per-key guarantee is the intuitively "correct"
77/// fix and someone will re-derive it: this table is the evidence that it
78/// costs quality HERE, and the reason is that the guarantee is bought by
79/// destroying signal — clamping a landmark's coefficient zeroes its
80/// contribution to EVERY far key, including the majority where the
81/// weighted sum was already positive and accurate.
82#[derive(Clone, Copy, Debug, PartialEq, Eq)]
83pub enum O1Rect {
84 /// Clamp only the AGGREGATE far denominator: a row whose skeleton
85 /// denominator comes out negative drops its far field entirely.
86 /// Coarse — negative per-key mass survives untouched whenever the
87 /// row sum happens to stay positive — but measured BEST (see above):
88 /// the surviving negatives are apparently error-cancelling, not
89 /// error-causing.
90 Aggregate,
91 /// Clamp FM = F_u·M_u (an m-vector, per query row) at zero.
92 /// ŵ(t,j) = Σ_b FM[b]·E[b,j] and E = exp(·) ≥ 0 ELEMENTWISE, so
93 /// FM ≥ 0 is SUFFICIENT for every far weight to be non-negative —
94 /// a per-key guarantee bought with O(m) work on a vector the row
95 /// already materializes, state untouched. It is strictly stronger
96 /// than the probe's clamp (a negative landmark is dropped for every
97 /// key, not only where the sum would go negative), so this is a
98 /// DIFFERENT operator, not an emulation of the matrix reference —
99 /// and, measured, a worse one. Opt in with `--o1-rect fm`.
100 Fm,
101}
102
103/// Ridge factor for the regularized pseudo-inverse of the landmark
104/// kernel: λ = RIDGE_REL · mean(diag(AᵀA)).
105const RIDGE_REL: f64 = 1e-6;
106/// Floor for the joint denominator (mirrors the reference probe).
107const DEN_EPS: f32 = 1e-30;
108/// Prompts of length ≤ w + EXACT_SLACK skip the skeleton entirely:
109/// tiny prefills duplicate segment-mean landmarks (singular Au).
110const EXACT_SLACK: usize = 8;
111
112/// Streaming Nyström attention state for ONE GQA group.
113///
114/// State splits along the GQA grain, because the operator does:
115///
116/// * SHARED per KV group (`NystromGroup`) — the exact window ring, the
117/// sink buffer and the key landmarks K̃. Under GQA every Q head of a
118/// group reads the SAME k/v rows, so all three are bit-identical
119/// across the group; storing them once per group instead of once per
120/// Q head is the point of this split (identical arithmetic,
121/// ×heads_per_kv less window memory). K̃ = seg_means(ks, t, d, m_eff)
122/// is a pure function of the group's keys and of `t` (which fixes
123/// m_eff), so it is shareable for the same reason the keys are.
124/// * PRIVATE per Q head (`NystromHead`) — the far accumulators T̂/Ẑ and
125/// their per-landmark running maxima, the QUERY landmarks Q̃, and the
126/// mixing matrix M = pinv(exp(Q̃K̃ᵀ/√d)). Q̃ is built from that head's
127/// own queries, so M and the far field it drives are per-Q-head and
128/// cannot be shared: the far mass a head accumulates is contracted
129/// against its own query landmarks.
130///
131/// Lifecycle: `new(m, w, sink)` → `prefill(prompt)` once → `step()` per
132/// decode token (single-head façade), or `new_group`/`prefill_group`/
133/// `step_group` for a whole GQA group at once. All buffers are flat
134/// `Vec<f32>`, row-major; the skeleton path performs no allocations
135/// inside `step()`.
136#[derive(Clone, Debug)]
137pub struct NystromState {
138 group: NystromGroup,
139 heads: Vec<NystromHead>,
140}
141
142/// The part of the state a GQA group shares: everything derived from
143/// the group's KEYS and VALUES alone (see `NystromState`).
144#[derive(Clone, Debug)]
145struct NystromGroup {
146 /// Landmark budget (m) — effective count may be lower (`m_eff`).
147 m: usize,
148 /// Exact-window width in keys.
149 w: usize,
150 /// Permanent exact sink keys at positions 0..sink (spec §5b).
151 sink: usize,
152 d: usize,
153 dv: usize,
154 /// Effective landmark count: clamp(t/8, 4, m) at prefill. Derived
155 /// from the prompt length, hence equal for every head of the group.
156 m_eff: usize,
157 /// Short-prompt mode: window holds ALL keys, no skeleton. The
158 /// buffer grows on decode, so this mode may allocate in `step()` —
159 /// acceptable for the ≤ w+8-token degenerate case.
160 exact_only: bool,
161 scale: f32,
162 /// Window keys `[cap][d]` — ring buffer in skeleton mode (cap = w),
163 /// append-only in exact-only mode.
164 win_k: Vec<f32>,
165 /// Window values `[cap][dv]`.
166 win_v: Vec<f32>,
167 win_len: usize,
168 /// Ring slot of the OLDEST window entry (0 while not yet full).
169 win_head: usize,
170 /// Sink keys `[sink_len][d]` — filled once at prefill, immutable.
171 sink_k: Vec<f32>,
172 /// Sink values `[sink_len][dv]`.
173 sink_v: Vec<f32>,
174 /// Number of stored sink tokens (0 in exact-only mode, where every
175 /// key is permanent-exact anyway).
176 sink_len: usize,
177 /// Key landmarks `[m_eff][d]` — segment means of the group's keys.
178 k_tilde: Vec<f32>,
179}
180
181/// The part of the state that is private to one Q head: everything that
182/// touches that head's QUERIES (see `NystromState`).
183#[derive(Clone, Debug)]
184struct NystromHead {
185 /// How the indefinite skeleton is rectified (see `O1Rect`).
186 rect: O1Rect,
187 /// Far numerator `[m_eff][dv]`, stored at scale e^{-m_max[i]}.
188 t_hat: Vec<f32>,
189 /// Far denominator `[m_eff]`, same scale.
190 z_hat: Vec<f32>,
191 /// Per-landmark running max of far logits q̃_i·k_j/√d.
192 m_max: Vec<f32>,
193 /// Number of keys absorbed into the far field.
194 far_len: usize,
195 /// Query landmarks `[m_eff][d]` (segment means of the prefill).
196 q_tilde: Vec<f32>,
197 /// Regularized pseudo-inverse of Au = exp(Q̃·K̃ᵀ/√d), `[m_eff][m_eff]`.
198 mu: Vec<f32>,
199 // Scratch preallocated at prefill so skeleton-mode step() is
200 // allocation-free. Per head rather than per group: the heads of a
201 // group write it independently, and it is ~0.5 KB.
202 scr_s: Vec<f32>,
203 scr_fh: Vec<f32>,
204 scr_u: Vec<f32>,
205 scr_l: Vec<f32>,
206}
207
208impl NystromState {
209 /// Single-head state (`heads_per_kv == 1`, and the shape the kernel
210 /// unit tests use).
211 ///
212 /// `m` — landmark budget (≥ 4; see `O1_DEFAULT_M`),
213 /// `w` — exact window width (validated setting is 128),
214 /// `sink` — permanent exact sink keys (validated default is 4;
215 /// 0 reproduces the sink-free kernel bit-for-bit).
216 /// Rectifier defaults to `O1_DEFAULT_RECT`; override with
217 /// `with_rect` (the golden-parity test pins it explicitly).
218 pub fn new(m: usize, w: usize, sink: usize) -> Self {
219 Self::new_group(m, w, sink, 1)
220 }
221
222 /// State for one GQA group of `q_heads` query heads sharing a KV
223 /// head. The window/sink/K̃ are stored ONCE for the group; each Q
224 /// head keeps its own far field, Q̃ and M.
225 pub fn new_group(m: usize, w: usize, sink: usize, q_heads: usize) -> Self {
226 assert!(m >= 4, "landmark budget must be at least 4");
227 assert!(w >= 1, "window must hold at least one key");
228 assert!(q_heads >= 1, "a GQA group needs at least one query head");
229 NystromState {
230 group: NystromGroup {
231 m,
232 w,
233 sink,
234 d: 0,
235 dv: 0,
236 m_eff: 0,
237 exact_only: true,
238 scale: 0.0,
239 win_k: Vec::new(),
240 win_v: Vec::new(),
241 win_len: 0,
242 win_head: 0,
243 sink_k: Vec::new(),
244 sink_v: Vec::new(),
245 sink_len: 0,
246 k_tilde: Vec::new(),
247 },
248 heads: (0..q_heads).map(|_| NystromHead::new()).collect(),
249 }
250 }
251
252 /// Select the skeleton rectifier for every head of the group
253 /// (builder; see `O1Rect`).
254 pub fn with_rect(mut self, rect: O1Rect) -> Self {
255 for h in &mut self.heads {
256 h.rect = rect;
257 }
258 self
259 }
260
261 /// Query heads in this group.
262 pub fn num_q_heads(&self) -> usize {
263 self.heads.len()
264 }
265
266 /// Keys absorbed into head `head`'s far field. Exposed for the
267 /// delayed-insertion invariant test: eviction is a GROUP event, but
268 /// each head must absorb the evicted key EXACTLY once, so this must
269 /// equal the number of evictions — never a multiple of it.
270 pub fn far_len(&self, head: usize) -> usize {
271 self.heads[head].far_len
272 }
273
274 /// Absorb the whole prompt for a single-head state — see
275 /// `prefill_group`.
276 pub fn prefill(&mut self, qs: &[f32], ks: &[f32], vs: &[f32], t: usize, d: usize, dv: usize) {
277 assert_eq!(self.heads.len(), 1, "use prefill_group for a GQA group");
278 self.prefill_group(&[qs], ks, vs, t, d, dv);
279 }
280
281 /// Absorb the whole prompt for a GQA group: freeze each head's
282 /// landmarks and M, then replay the prompt through the step() state
283 /// semantics (window fill + delayed far insertion). `qs[h]` is that
284 /// head's `[t][d]` query block; `ks` is `[t][d]` and `vs` is
285 /// `[t][dv]` — the group's shared keys/values, row-major.
286 pub fn prefill_group(
287 &mut self,
288 qs: &[&[f32]],
289 ks: &[f32],
290 vs: &[f32],
291 t: usize,
292 d: usize,
293 dv: usize,
294 ) {
295 assert_eq!(qs.len(), self.heads.len(), "one query block per head");
296 for q in qs {
297 assert_eq!(q.len(), t * d);
298 }
299 assert_eq!(ks.len(), t * d);
300 assert_eq!(vs.len(), t * dv);
301
302 let Some(k_tilde64) = self.group.prefill_shared(ks, vs, t, d, dv) else {
303 // exact-only: no skeleton, no far field — nothing per head
304 // beyond the score scratch.
305 for h in &mut self.heads {
306 h.seal_exact(t);
307 }
308 return;
309 };
310 for (h, q) in self.heads.iter_mut().zip(qs) {
311 h.seal(&self.group, q, t, &k_tilde64);
312 }
313 // Replay the post-sink prompt ONCE for the group: each key
314 // enters the shared window, evicting the (j-w)-th into every
315 // head's far field.
316 for j in self.group.sink..t {
317 Self::advance(
318 &mut self.group,
319 &mut self.heads,
320 &ks[j * d..(j + 1) * d],
321 &vs[j * dv..(j + 1) * dv],
322 );
323 }
324 }
325
326 /// One decode step for a single-head state — see `step_group`.
327 pub fn step(&mut self, q: &[f32], k: &[f32], v: &[f32], out: &mut [f32]) {
328 assert_eq!(self.heads.len(), 1, "use step_group for a GQA group");
329 self.step_group(q, k, v, out);
330 }
331
332 /// One decode step for the whole GQA group. Inserts the group's
333 /// (k, v) ONCE, evicting the oldest window key into every head's far
334 /// accumulators, then writes each head's attention output.
335 /// `q_all` is `[q_heads][d]`, `out_all` is `[q_heads][dv]`.
336 pub fn step_group(&mut self, q_all: &[f32], k: &[f32], v: &[f32], out_all: &mut [f32]) {
337 let (d, dv) = (self.group.d, self.group.dv);
338 assert!(d > 0, "prefill() must run before step()");
339 let nh = self.heads.len();
340 assert_eq!(q_all.len(), nh * d);
341 assert_eq!(k.len(), d);
342 assert_eq!(v.len(), dv);
343 assert_eq!(out_all.len(), nh * dv);
344 // The current token is part of its own near window (t-j = 0),
345 // so insertion happens BEFORE any output is computed.
346 Self::advance(&mut self.group, &mut self.heads, k, v);
347 for (h, head) in self.heads.iter_mut().enumerate() {
348 head.step(
349 &self.group,
350 &q_all[h * d..(h + 1) * d],
351 &mut out_all[h * dv..(h + 1) * dv],
352 );
353 }
354 }
355
356 /// Heap bytes held by this group's state (shared window + sinks +
357 /// K̃, plus each head's skeleton and scratch) — feeds the honest
358 /// "KV+state" memory line, same discipline as counting
359 /// `linear_state` for the linear core.
360 pub fn memory_bytes(&self) -> usize {
361 self.group.memory_bytes()
362 + self
363 .heads
364 .iter()
365 .map(NystromHead::memory_bytes)
366 .sum::<usize>()
367 }
368
369 /// Push the group's (k, v) into the shared window. In skeleton mode
370 /// a full ring first evicts its oldest key (delayed insertion — the
371 /// key leaves the exact window at this very step).
372 ///
373 /// The eviction is a GROUP event: the window is shared, so there is
374 /// exactly ONE eviction per position, not one per Q head. The far
375 /// accumulators are per head, though, so that single evicted key is
376 /// absorbed once into EACH head — one eviction, `q_heads`
377 /// insertions. Getting this wrong in either direction breaks the
378 /// boundary invariant (a key enters the far field at exactly the
379 /// step it leaves the window: no double count, no hole).
380 fn advance(g: &mut NystromGroup, heads: &mut [NystromHead], k: &[f32], v: &[f32]) {
381 let (d, dv) = (g.d, g.dv);
382 if !g.exact_only && g.win_len == g.w {
383 let slot = g.win_head;
384 // Every head absorbs the outgoing key BEFORE the slot is
385 // overwritten by the incoming one.
386 for h in heads.iter_mut() {
387 h.far_insert(g, slot);
388 }
389 g.win_k[slot * d..(slot + 1) * d].copy_from_slice(k);
390 g.win_v[slot * dv..(slot + 1) * dv].copy_from_slice(v);
391 g.win_head = (g.win_head + 1) % g.w;
392 } else if g.exact_only {
393 g.win_k.extend_from_slice(k);
394 g.win_v.extend_from_slice(v);
395 g.win_len += 1;
396 } else {
397 g.win_k[g.win_len * d..(g.win_len + 1) * d].copy_from_slice(k);
398 g.win_v[g.win_len * dv..(g.win_len + 1) * dv].copy_from_slice(v);
399 g.win_len += 1;
400 }
401 }
402}
403
404impl NystromGroup {
405 /// Freeze the group-shared geometry from the prompt's keys/values.
406 /// Returns the f64 key landmarks (which the heads need at full
407 /// precision to build Au), or None in exact-only mode.
408 fn prefill_shared(
409 &mut self,
410 ks: &[f32],
411 vs: &[f32],
412 t: usize,
413 d: usize,
414 dv: usize,
415 ) -> Option<Vec<f64>> {
416 self.d = d;
417 self.dv = dv;
418 self.scale = 1.0 / (d as f32).sqrt();
419 self.win_len = 0;
420 self.win_head = 0;
421 self.sink_len = 0;
422 self.exact_only = t <= self.w + self.sink + EXACT_SLACK;
423
424 if self.exact_only {
425 // Everything fits in the exact window (plus slack for a few
426 // decode steps before Vec growth); no skeleton is built and
427 // no separate sink buffer is needed — every key is already
428 // a permanent exact key in this mode.
429 self.win_k = Vec::with_capacity((t + 64) * d);
430 self.win_v = Vec::with_capacity((t + 64) * dv);
431 self.win_k.extend_from_slice(ks);
432 self.win_v.extend_from_slice(vs);
433 self.win_len = t;
434 return None;
435 }
436
437 // Sink tokens: positions 0..sink become permanent exact keys.
438 // They bypass the ring window entirely, so the delayed-insertion
439 // path can never move them into the far accumulators.
440 self.sink_len = self.sink; // skeleton mode guarantees t > sink
441 self.sink_k = ks[..self.sink * d].to_vec();
442 self.sink_v = vs[..self.sink * dv].to_vec();
443
444 // Landmarks: contiguous segment means of the prompt. The
445 // integer split (i·t)/m matches the reference probe; the clamp
446 // keeps tiny prompts from producing duplicate landmarks.
447 let m_eff = (t / 8).clamp(4, self.m);
448 self.m_eff = m_eff;
449 let k_tilde64 = seg_means(ks, t, d, m_eff);
450 self.k_tilde = k_tilde64.iter().map(|&x| x as f32).collect();
451
452 self.win_k = vec![0.0; self.w * d];
453 self.win_v = vec![0.0; self.w * dv];
454 Some(k_tilde64)
455 }
456
457 fn memory_bytes(&self) -> usize {
458 (self.win_k.len()
459 + self.win_v.len()
460 + self.sink_k.len()
461 + self.sink_v.len()
462 + self.k_tilde.len())
463 * std::mem::size_of::<f32>()
464 }
465}
466
467impl NystromHead {
468 fn new() -> Self {
469 NystromHead {
470 rect: O1_DEFAULT_RECT,
471 t_hat: Vec::new(),
472 z_hat: Vec::new(),
473 m_max: Vec::new(),
474 far_len: 0,
475 q_tilde: Vec::new(),
476 mu: Vec::new(),
477 scr_s: Vec::new(),
478 scr_fh: Vec::new(),
479 scr_u: Vec::new(),
480 scr_l: Vec::new(),
481 }
482 }
483
484 /// exact-only mode: no skeleton state at all, just room to score the
485 /// growing window.
486 fn seal_exact(&mut self, t: usize) {
487 self.far_len = 0;
488 self.scr_s = Vec::with_capacity(t + 64);
489 }
490
491 /// Freeze this head's query landmarks and mixing matrix against the
492 /// group's (already frozen) key landmarks.
493 fn seal(&mut self, g: &NystromGroup, qs: &[f32], t: usize, k_tilde64: &[f64]) {
494 let (d, dv, m_eff) = (g.d, g.dv, g.m_eff);
495 self.far_len = 0;
496 let q_tilde64 = seg_means(qs, t, d, m_eff);
497 self.q_tilde = q_tilde64.iter().map(|&x| x as f32).collect();
498
499 // Au and its regularized pseudo-inverse in f64 — one-off m×m
500 // work at prefill only; the hot path stays f32.
501 let mut au = vec![0.0f64; m_eff * m_eff];
502 for i in 0..m_eff {
503 for j in 0..m_eff {
504 let mut s = 0.0f64;
505 for c in 0..d {
506 s += q_tilde64[i * d + c] * k_tilde64[j * d + c];
507 }
508 au[i * m_eff + j] = (s * g.scale as f64).exp();
509 }
510 }
511 let mu64 = ridge_pinv(&au, m_eff);
512 self.mu = mu64.iter().map(|&x| x as f32).collect();
513
514 self.t_hat = vec![0.0; m_eff * dv];
515 self.z_hat = vec![0.0; m_eff];
516 self.m_max = vec![f32::NEG_INFINITY; m_eff];
517 self.scr_s = vec![0.0; g.sink + g.w];
518 self.scr_fh = vec![0.0; m_eff];
519 self.scr_u = vec![0.0; m_eff];
520 self.scr_l = vec![0.0; m_eff];
521 }
522
523 /// This head's output for `q` against the group's current window and
524 /// sinks and its own far field. The window insertion for this
525 /// position already happened at group level (`NystromState::advance`).
526 fn step(&mut self, g: &NystromGroup, q: &[f32], out: &mut [f32]) {
527 let (d, dv) = (g.d, g.dv);
528 assert_eq!(q.len(), d);
529 assert_eq!(out.len(), dv);
530
531 // Near field: exact logits over sinks + window, one shared
532 // shift. Sinks are permanent exact keys (near mask §5b:
533 // t-j < w OR j < sink); sink_len = 0 in exact-only mode.
534 let ns = g.sink_len;
535 let n = ns + g.win_len;
536 self.scr_s.resize(n, 0.0);
537 let mut c = f32::NEG_INFINITY;
538 for s in 0..ns {
539 let lg = dot(q, &g.sink_k[s * d..(s + 1) * d]) * g.scale;
540 self.scr_s[s] = lg;
541 c = c.max(lg);
542 }
543 // Window scores are the decode hot loop — NEON dot (same
544 // products, regrouped sums; parity-gated by the golden tests).
545 for s in 0..g.win_len {
546 let lg = crate::attention::dot_f32(q, &g.win_k[s * d..(s + 1) * d]) * g.scale;
547 self.scr_s[ns + s] = lg;
548 c = c.max(lg);
549 }
550
551 // Far field: shifted skeleton (spec §3). All exp arguments are
552 // ≤ 0 relative to the joint shift c_all, so nothing overflows.
553 let mut far_den = 0.0f32;
554 let mut c_all = c;
555 let mut have_far = false;
556 if self.far_len > 0 {
557 // Per-token row shift f over landmark scores.
558 let mut f = f32::NEG_INFINITY;
559 for a in 0..g.m_eff {
560 let s = crate::attention::dot_f32(q, &g.k_tilde[a * d..(a + 1) * d]) * g.scale;
561 self.scr_fh[a] = s;
562 f = f.max(s);
563 }
564 for a in 0..g.m_eff {
565 self.scr_fh[a] = (self.scr_fh[a] - f).exp();
566 }
567 // u = (F·e^{-f}) · M — the landmark mixing row (= FM, up to
568 // the positive factor e^{-f}).
569 for b in 0..g.m_eff {
570 let mut s = 0.0f32;
571 for a in 0..g.m_eff {
572 s += self.scr_fh[a] * self.mu[a * g.m_eff + b];
573 }
574 // FM rectifier: every far weight is Σ_b FM[b]·E[b,j]
575 // with E ≥ 0 elementwise, so clamping this m-vector is
576 // enough to make all of them non-negative — the per-key
577 // guarantee the streaming form otherwise cannot state.
578 // The row shift e^{-f} and the flash factors below are
579 // strictly positive, so clamping here or after the
580 // rescale is the same predicate.
581 self.scr_u[b] = if self.rect == O1Rect::Fm {
582 s.max(0.0)
583 } else {
584 s
585 };
586 }
587 // Joint scale: the far term b carries e^{f + m_max[b]}, the
588 // near term e^{c}; take the max so every factor is ≤ 1.
589 for b in 0..g.m_eff {
590 c_all = c_all.max(f + self.m_max[b]);
591 }
592 for b in 0..g.m_eff {
593 let gain = self.scr_u[b] * (f + self.m_max[b] - c_all).exp();
594 self.scr_u[b] = gain;
595 far_den += gain * self.z_hat[b];
596 }
597 // Aggregate guard — the rectifier of `O1Rect::Aggregate`,
598 // and a second line of defence under `Fm` (where far_den is
599 // a sum of non-negative terms, so this can only fire on
600 // rounding): a negative denominator means the skeleton
601 // estimate is unusable for this row — drop the far field.
602 if far_den >= 0.0 {
603 have_far = true;
604 } else {
605 far_den = 0.0;
606 }
607 }
608
609 for o in out.iter_mut() {
610 *o = 0.0;
611 }
612 if have_far {
613 for b in 0..g.m_eff {
614 crate::attention::axpy_f32(out, &self.t_hat[b * dv..(b + 1) * dv], self.scr_u[b]);
615 }
616 }
617 let mut den = far_den;
618 for s in 0..n {
619 let p = (self.scr_s[s] - c_all).exp();
620 den += p;
621 // scr_s rows 0..ns are sinks, the rest are window entries.
622 let vv = if s < ns {
623 &g.sink_v[s * dv..(s + 1) * dv]
624 } else {
625 &g.win_v[(s - ns) * dv..(s - ns + 1) * dv]
626 };
627 crate::attention::axpy_f32(out, vv, p);
628 }
629 let den = den.max(DEN_EPS);
630 for o in out.iter_mut() {
631 *o /= den;
632 }
633 }
634
635 /// Absorb the group's window slot into THIS head's far accumulators
636 /// with the per-landmark flash shift: T̂[i]/Ẑ[i] live at scale
637 /// e^{-m_max[i]}; when a new logit raises the max, existing mass is
638 /// rescaled by e^{old-new} (exactly 0 on first insertion, since
639 /// m_max = -inf).
640 fn far_insert(&mut self, g: &NystromGroup, slot: usize) {
641 let (d, dv) = (g.d, g.dv);
642 // Runs once per evicted key per head — NEON dot/axpy like the
643 // decode loop (same products, regrouped sums).
644 for i in 0..g.m_eff {
645 self.scr_l[i] = crate::attention::dot_f32(
646 &self.q_tilde[i * d..(i + 1) * d],
647 &g.win_k[slot * d..(slot + 1) * d],
648 ) * g.scale;
649 }
650 for i in 0..g.m_eff {
651 let l = self.scr_l[i];
652 if l > self.m_max[i] {
653 let r = (self.m_max[i] - l).exp();
654 self.z_hat[i] *= r;
655 for e in self.t_hat[i * dv..(i + 1) * dv].iter_mut() {
656 *e *= r;
657 }
658 self.m_max[i] = l;
659 }
660 let e = (l - self.m_max[i]).exp();
661 self.z_hat[i] += e;
662 crate::attention::axpy_f32(
663 &mut self.t_hat[i * dv..(i + 1) * dv],
664 &g.win_v[slot * dv..(slot + 1) * dv],
665 e,
666 );
667 }
668 self.far_len += 1;
669 }
670
671 fn memory_bytes(&self) -> usize {
672 (self.t_hat.len()
673 + self.z_hat.len()
674 + self.m_max.len()
675 + self.q_tilde.len()
676 + self.mu.len()
677 + self.scr_s.len()
678 + self.scr_fh.len()
679 + self.scr_u.len()
680 + self.scr_l.len())
681 * std::mem::size_of::<f32>()
682 }
683}
684
685/// Contiguous segment means (the Nyströmformer landmark recipe), f64
686/// accumulation. The split (i·t)/m matches the Python reference.
687fn seg_means(xs: &[f32], t: usize, d: usize, m: usize) -> Vec<f64> {
688 let mut out = vec![0.0f64; m * d];
689 for i in 0..m {
690 let lo = i * t / m;
691 let hi = (i + 1) * t / m;
692 for j in lo..hi {
693 for c in 0..d {
694 out[i * d + c] += xs[j * d + c] as f64;
695 }
696 }
697 let inv = 1.0 / (hi - lo) as f64;
698 for c in 0..d {
699 out[i * d + c] *= inv;
700 }
701 }
702 out
703}
704
705fn dot(a: &[f32], b: &[f32]) -> f32 {
706 let mut s = 0.0f32;
707 for (x, y) in a.iter().zip(b) {
708 s += x * y;
709 }
710 s
711}
712
713/// Regularized pseudo-inverse M = (AᵀA + λI)⁻¹ Aᵀ of a square matrix,
714/// λ = RIDGE_REL·mean(diag(AᵀA)), solved via Cholesky. f64 internal —
715/// this runs once per prefill on an m×m matrix (m ≤ 32). If Cholesky
716/// fails (Au numerically singular despite the m_eff clamp), λ grows
717/// tenfold — the jitter fallback of the reference probe.
718/// pub(crate): the FCD polish trainer builds its (constant-in-backward)
719/// mixing matrix with the SAME solver the runtime seals with.
720pub(crate) fn ridge_pinv(a: &[f64], n: usize) -> Vec<f64> {
721 let mut ata = vec![0.0f64; n * n];
722 for i in 0..n {
723 for j in 0..n {
724 let mut s = 0.0;
725 for k in 0..n {
726 s += a[k * n + i] * a[k * n + j];
727 }
728 ata[i * n + j] = s;
729 }
730 }
731 let mean_diag: f64 = (0..n).map(|i| ata[i * n + i]).sum::<f64>() / n as f64;
732 let mut lambda = RIDGE_REL * mean_diag.max(f64::MIN_POSITIVE);
733 for _ in 0..12 {
734 let mut g = ata.clone();
735 for i in 0..n {
736 g[i * n + i] += lambda;
737 }
738 if let Some(l) = cholesky(&mut g, n) {
739 // Solve G·M = Aᵀ column by column; column j of Aᵀ is row j
740 // of A.
741 let mut m_out = vec![0.0f64; n * n];
742 let mut x = vec![0.0f64; n];
743 for j in 0..n {
744 let rhs = &a[j * n..(j + 1) * n];
745 // Forward: L·y = rhs.
746 for i in 0..n {
747 let mut s = rhs[i];
748 for k in 0..i {
749 s -= l[i * n + k] * x[k];
750 }
751 x[i] = s / l[i * n + i];
752 }
753 // Backward: Lᵀ·x = y.
754 for i in (0..n).rev() {
755 let mut s = x[i];
756 for k in i + 1..n {
757 s -= l[k * n + i] * x[k];
758 }
759 x[i] = s / l[i * n + i];
760 }
761 for i in 0..n {
762 m_out[i * n + j] = x[i];
763 }
764 }
765 return m_out;
766 }
767 lambda *= 10.0;
768 }
769 // Unreachable in practice: λ eventually dominates the diagonal.
770 // Degrade to a scaled identity rather than poison the output.
771 let mut fallback = vec![0.0f64; n * n];
772 for i in 0..n {
773 fallback[i * n + i] = 1.0 / mean_diag.max(f64::MIN_POSITIVE);
774 }
775 fallback
776}
777
778// ── Runtime configuration (v1: runtime-level, NOT a format change) ──
779//
780// A layer set + {m, w, sink}, resolved in priority order:
781// 1. CLI flag (`--o1` on run/serve/bench) — explicit user intent;
782// 2. env `CMF_O1` (all | deepN | i,j,k | off) with CMF_O1_M /
783// CMF_O1_WINDOW / CMF_O1_SINK parameter overrides;
784// 3. converter hint in the header JSON (`provenance.o1_attn`,
785// written by `cortiq convert --o1`) — additive metadata, the
786// binary envelope is untouched.
787
788/// Validated defaults (spec: m=32, W=128, sink=4; sink ablation ×2.39).
789pub const O1_DEFAULT_M: usize = 32;
790pub const O1_DEFAULT_W: usize = 128;
791pub const O1_DEFAULT_SINK: usize = 4;
792/// Rectifier default (see `O1Rect`).
793pub const O1_DEFAULT_RECT: O1Rect = O1Rect::Aggregate;
794
795/// Which layers run the O(1) kernel.
796#[derive(Clone, Debug, PartialEq, Eq)]
797pub enum O1Layers {
798 All,
799 /// The N deepest layers (deep-N ladder of the price map; the
800 /// early stack is the most sink-dependent, depth converts best).
801 Deep(usize),
802 /// Explicit layer indices.
803 List(Vec<usize>),
804}
805
806/// Per-model O(1)-attention setting.
807#[derive(Clone, Debug)]
808pub struct O1Cfg {
809 pub layers: O1Layers,
810 /// Landmark budget (≥ 4; m=64 measured WORSE — collinear segment
811 /// means poison the pinv, so don't "help" by raising it).
812 pub m: usize,
813 /// Exact-window width — the main quality lever.
814 pub w: usize,
815 /// Permanent exact sink keys (StreamingLLM discipline, spec §5b).
816 pub sink: usize,
817 /// Skeleton rectifier (see `O1Rect`).
818 pub rect: O1Rect,
819}
820
821/// Three-state env reading: unset falls through to the header hint,
822/// `off`/`0` force-disables even a header hint (the escape hatch).
823pub enum O1Env {
824 Unset,
825 Off,
826 On(O1Cfg),
827}
828
829impl O1Cfg {
830 /// Parse a layer spec: `all` | `deepN` | `i,j,k`. None = not a spec
831 /// (also used for `off`/`0`/empty).
832 pub fn parse_layers(spec: &str) -> Option<O1Layers> {
833 let s = spec.trim();
834 match s {
835 "" | "off" | "0" | "none" => None,
836 "all" => Some(O1Layers::All),
837 _ => {
838 if let Some(n) = s.strip_prefix("deep") {
839 return n
840 .parse::<usize>()
841 .ok()
842 .filter(|&n| n > 0)
843 .map(O1Layers::Deep);
844 }
845 let idx: Result<Vec<usize>, _> =
846 s.split(',').map(|p| p.trim().parse::<usize>()).collect();
847 idx.ok().filter(|v| !v.is_empty()).map(O1Layers::List)
848 }
849 }
850 }
851
852 /// Parse a rectifier spec: `agg`/`aggregate` | `fm`. None = not a
853 /// spec.
854 pub fn parse_rect(spec: &str) -> Option<O1Rect> {
855 match spec.trim() {
856 "agg" | "aggregate" => Some(O1Rect::Aggregate),
857 "fm" => Some(O1Rect::Fm),
858 _ => None,
859 }
860 }
861
862 /// Rectifier from an explicit value, else `CMF_O1_RECT`, else the
863 /// default.
864 fn rect_or_env(rect: Option<O1Rect>) -> O1Rect {
865 rect.or_else(|| {
866 std::env::var("CMF_O1_RECT")
867 .ok()
868 .as_deref()
869 .and_then(Self::parse_rect)
870 })
871 .unwrap_or(O1_DEFAULT_RECT)
872 }
873
874 /// Build from an explicit spec (CLI path). None = `off` or malformed.
875 /// Explicit m/w/sink/rect beat env overrides beat validated defaults.
876 pub fn from_spec(
877 spec: &str,
878 m: Option<usize>,
879 w: Option<usize>,
880 sink: Option<usize>,
881 rect: Option<O1Rect>,
882 ) -> Option<O1Cfg> {
883 let layers = Self::parse_layers(spec)?;
884 let env = |k: &str| std::env::var(k).ok().and_then(|v| v.parse::<usize>().ok());
885 Some(O1Cfg {
886 layers,
887 // NystromState asserts m ≥ 4 and w ≥ 1 — clamp rather than
888 // panic deep in the first prefill.
889 m: m.or_else(|| env("CMF_O1_M")).unwrap_or(O1_DEFAULT_M).max(4),
890 w: w.or_else(|| env("CMF_O1_WINDOW"))
891 .unwrap_or(O1_DEFAULT_W)
892 .max(1),
893 sink: sink
894 .or_else(|| env("CMF_O1_SINK"))
895 .unwrap_or(O1_DEFAULT_SINK),
896 rect: Self::rect_or_env(rect),
897 })
898 }
899
900 /// Converter hint from the header JSON: `{"layers": "all"|[i,…],
901 /// "m": …, "w": …, "sink": …}`. Env parameter overrides still apply
902 /// (the operator's knob wins over the file's suggestion).
903 pub fn from_json(v: &serde_json::Value) -> Option<O1Cfg> {
904 let layers = match v.get("layers") {
905 Some(serde_json::Value::String(s)) => Self::parse_layers(s)?,
906 Some(serde_json::Value::Array(a)) => O1Layers::List(
907 a.iter()
908 .filter_map(|x| x.as_u64().map(|n| n as usize))
909 .collect(),
910 ),
911 _ => return None,
912 };
913 let f = |k: &str| v.get(k).and_then(|x| x.as_u64()).map(|n| n as usize);
914 let env = |k: &str| std::env::var(k).ok().and_then(|s| s.parse::<usize>().ok());
915 Some(O1Cfg {
916 layers,
917 m: env("CMF_O1_M")
918 .or_else(|| f("m"))
919 .unwrap_or(O1_DEFAULT_M)
920 .max(4),
921 w: env("CMF_O1_WINDOW")
922 .or_else(|| f("w"))
923 .unwrap_or(O1_DEFAULT_W)
924 .max(1),
925 sink: env("CMF_O1_SINK")
926 .or_else(|| f("sink"))
927 .unwrap_or(O1_DEFAULT_SINK),
928 // The rectifier is a runtime property of the kernel, not a
929 // property of the weights — a file hint cannot pin it.
930 rect: Self::rect_or_env(None),
931 })
932 }
933
934 /// Per-layer flags over `num_layers` (indices past the end are
935 /// silently dropped; the pipeline additionally filters non-Full
936 /// layers — a linear layer keeps its own operator).
937 pub fn layer_flags(&self, num_layers: usize) -> Vec<bool> {
938 let mut flags = vec![false; num_layers];
939 match &self.layers {
940 O1Layers::All => flags.iter_mut().for_each(|f| *f = true),
941 O1Layers::Deep(n) => {
942 for f in flags.iter_mut().skip(num_layers.saturating_sub(*n)) {
943 *f = true;
944 }
945 }
946 O1Layers::List(idx) => {
947 for &i in idx {
948 if i < num_layers {
949 flags[i] = true;
950 }
951 }
952 }
953 }
954 flags
955 }
956}
957
958/// Read `CMF_O1` (+ parameter overrides) — the embedding-friendly path
959/// for hosts that don't go through the CLI flags.
960pub fn o1_from_env() -> O1Env {
961 match std::env::var("CMF_O1") {
962 Err(_) => O1Env::Unset,
963 Ok(s) => match O1Cfg::from_spec(&s, None, None, None, None) {
964 Some(cfg) => O1Env::On(cfg),
965 None => O1Env::Off,
966 },
967 }
968}
969
970/// In-place lower Cholesky of an SPD matrix; None if a pivot fails.
971fn cholesky(g: &mut [f64], n: usize) -> Option<&[f64]> {
972 for i in 0..n {
973 for j in 0..=i {
974 let mut s = g[i * n + j];
975 for k in 0..j {
976 s -= g[i * n + k] * g[j * n + k];
977 }
978 if i == j {
979 if s <= 0.0 || !s.is_finite() {
980 return None;
981 }
982 g[i * n + i] = s.sqrt();
983 } else {
984 g[i * n + j] = s / g[j * n + j];
985 }
986 }
987 }
988 Some(g)
989}