code wiki / _hdl_build / nx_fnet_model_gate.nx
nx_fnet_model_gate.nx source
↩ module page · 285 lines · 13133 B
1// nx_fnet_model_gate.nx -- GATE for MODEL-001: the FIRST end-to-end sovereign SUB-QUADRATIC MODEL. A real
2// sequence classifier with no attention anywhere:
3// tokens -> EMBED (trained) -> FNET token-mix (sub-quadratic) -> relu FFN -> linear -> softmax-CE
4// Trained by AdamW JOINTLY over the embedding table + FFN/head weights. The embedding has no attention matrix
5// and no quadratic cost -- token mixing is the parameter-free Fourier transform. Embeddings are trained without
6// a dedicated op: the tape yields dL/dx at the input leaf, which is SCATTERED back into the embedding rows.
7//
8// Task: classify whether token[0] > token[3] over 8 length-4 sequences (vocab 4). This genuinely needs the FNet
9// to MIX positions 0 and 3 and the relu to compare -- a bag-of-words linear model cannot do it.
10//
11// G_train the model learns: accuracy >= 7/8 AND final loss < first loss.
12// G_repro bit-exact: train twice -> identical accuracy + final loss bits.
13//
14// Evidence -> knowledge/status/fnet_model.log (FNETMODELGATE authored=organ ... verdict=GREEN). license_tier: ORIGINAL
15import "nx_autograd_tensor.nx" // ta_* incl ta_fnet + transitively the f32 tower + fft_f32
16import "nx_syscalls.nx"
17
18const MV: i64 = 4 // vocab
19const MN: i64 = 4 // sequence length
20const MD: i64 = 4 // d_model
21const MH: i64 = 8 // FFN hidden
22const MC: i64 = 2 // classes
23const MND: i64 = 16 // MN*MD (flattened FNet output)
24
25const MO_LOG: *u8 = "knowledge/status/fnet_model.log"
26
27func mo_w(fd: i64, s: *u8) -> i64 { var n: i64 = 0; while s[n] != (0 as u8) { n = n + 1 } sys_write(fd, s, n); return 0 }
28func mo_wn(fd: i64, v: i64) -> i64 {
29 let bb: *u8 = sys_mmap(28); var m: i64 = v
30 if m < 0 { m = 0 - m; sys_write(fd, "-" as *u8, 1) }
31 let t: *u8 = sys_mmap(28); var k: i64 = 0
32 if m == 0 { t[0] = 48; k = 1 }
33 while m > 0 { t[k] = (48 + (m % 10)) as u8; m = m / 10; k = k + 1 }
34 var i: i64 = 0
35 while i < k { bb[i] = t[k - 1 - i]; i = i + 1 }
36 sys_write(fd, bb, k); return 0
37}
38
39// build x[MN*MD] by gathering embedding rows for the token sequence seq[MN]
40func mo_embed(E: *i64, seq: *i64, soff: i64, xout: *i64) -> i64 {
41 var i: i64 = 0
42 while i < MN {
43 let tok: i64 = seq[soff + i]
44 var j: i64 = 0
45 while j < MD { xout[i * MD + j] = E[tok * MD + j]; j = j + 1 }
46 i = i + 1
47 }
48 return 0
49}
50
51// one forward example -> logits node. caller has already placed embeddings into x.
52func mo_fwd(tape: *i64, vals: *i64, st: *i64, x: *i64, nW1: i64, nb1: i64, nW2: i64, nb2: i64) -> i64 {
53 let xl: i64 = ta_leaf(tape, vals, st, MN, MD, x, 0)
54 let m: i64 = ta_fnet(tape, vals, st, xl)
55 let h: i64 = ta_relu(tape, vals, st, ta_vadd(tape, vals, st, ta_matvec(tape, vals, st, nW1, m), nb1))
56 let lo: i64 = ta_vadd(tape, vals, st, ta_matvec(tape, vals, st, nW2, h), nb2)
57 return lo
58}
59
60// full-batch forward over 8 examples; writes shared-leaf indices wb[0..3] and per-example input-leaf idx xl8[0..7].
61func mo_build(tape: *i64, vals: *i64, st: *i64, E: *i64, W1: *i64, b1: *i64, W2: *i64, b2: *i64,
62 seqs: *i64, labels: *i64, c8: *i64, wb: *i64, xl8: *i64) -> i64 {
63 st[0] = 0; st[1] = 0
64 let nW1: i64 = ta_leaf(tape, vals, st, MH, MND, W1, 0)
65 let nb1: i64 = ta_leaf(tape, vals, st, MH, 1, b1, 0)
66 let nW2: i64 = ta_leaf(tape, vals, st, MC, MH, W2, 0)
67 let nb2: i64 = ta_leaf(tape, vals, st, MC, 1, b2, 0)
68 wb[0] = nW1; wb[1] = nb1; wb[2] = nW2; wb[3] = nb2
69 let x: *i64 = (sys_mmap(MND * 8)) as *i64
70 let th: *i64 = (sys_mmap(MC * 8)) as *i64
71 var sumn: i64 = 0 - 1
72 var s: i64 = 0
73 while s < 8 {
74 mo_embed(E, seqs, s * MN, x)
75 let xl: i64 = ta_leaf(tape, vals, st, MN, MD, x, 0)
76 xl8[s] = xl
77 let m: i64 = ta_fnet(tape, vals, st, xl)
78 let h: i64 = ta_relu(tape, vals, st, ta_vadd(tape, vals, st, ta_matvec(tape, vals, st, nW1, m), nb1))
79 let lo: i64 = ta_vadd(tape, vals, st, ta_matvec(tape, vals, st, nW2, h), nb2)
80 var j: i64 = 0
81 while j < MC { th[j] = TA_F32_ZERO; j = j + 1 }
82 th[labels[s]] = TA_F32_ONE
83 let tgt: i64 = ta_leaf(tape, vals, st, MC, 1, th, 0)
84 let ls: i64 = ta_softce(tape, vals, st, lo, tgt)
85 if s == 0 { sumn = ls } else { sumn = ta_vadd(tape, vals, st, sumn, ls) }
86 s = s + 1
87 }
88 let inv8: i64 = ta_leaf(tape, vals, st, 1, 1, c8, 0)
89 return ta_matvec(tape, vals, st, inv8, sumn)
90}
91
92func mo_predict(tape: *i64, vals: *i64, st: *i64, E: *i64, W1: *i64, b1: *i64, W2: *i64, b2: *i64, seqs: *i64, s: i64) -> i64 {
93 st[0] = 0; st[1] = 0
94 let nW1: i64 = ta_leaf(tape, vals, st, MH, MND, W1, 0)
95 let nb1: i64 = ta_leaf(tape, vals, st, MH, 1, b1, 0)
96 let nW2: i64 = ta_leaf(tape, vals, st, MC, MH, W2, 0)
97 let nb2: i64 = ta_leaf(tape, vals, st, MC, 1, b2, 0)
98 let x: *i64 = (sys_mmap(MND * 8)) as *i64
99 mo_embed(E, seqs, s * MN, x)
100 let lo: i64 = mo_fwd(tape, vals, st, x, nW1, nb1, nW2, nb2)
101 if nx_f32_gt(ta_val(tape, vals, lo, 1), ta_val(tape, vals, lo, 0)) == 1 { return 1 }
102 return 0
103}
104
105// AdamW update of n params in place
106func mo_adamw(p: *i64, m: *i64, v: *i64, g: *i64, n: i64, lr: i64, beta1: i64, beta2: i64, om1: i64, om2: i64, eps: i64, c1: i64, c2: i64) -> i64 {
107 var i: i64 = 0
108 while i < n {
109 let gi: i64 = g[i]
110 m[i] = nx_f32_add(nx_f32_mul(beta1, m[i]), nx_f32_mul(om1, gi))
111 v[i] = nx_f32_add(nx_f32_mul(beta2, v[i]), nx_f32_mul(om2, nx_f32_mul(gi, gi)))
112 let mhat: i64 = nx_f32_div(m[i], c1)
113 let vhat: i64 = nx_f32_div(v[i], c2)
114 p[i] = nx_f32_sub(p[i], nx_f32_mul(lr, nx_f32_div(mhat, nx_f32_add(nx_f32_sqrt(vhat), eps))))
115 i = i + 1
116 }
117 return 0
118}
119
120// train the whole model. writes final E/W1/b1/W2/b2 in place; returns nothing (params are the outputs).
121func mo_train(tape: *i64, vals: *i64, grads: *i64, st: *i64, E: *i64, W1: *i64, b1: *i64, W2: *i64, b2: *i64,
122 seqs: *i64, labels: *i64, epochs: i64, lf: *i64, ll: *i64) -> i64 {
123 // init params (deterministic symmetry-breaking)
124 ta_det_init(E, MV * MD, 5)
125 ta_det_init(W1, MH * MND, 3)
126 ta_det_init(W2, MC * MH, 7)
127 var z: i64 = 0
128 while z < MH { b1[z] = TA_F32_ZERO; z = z + 1 }
129 z = 0
130 while z < MC { b2[z] = TA_F32_ZERO; z = z + 1 }
131 // adamw state
132 let mE: *i64 = (sys_mmap(MV * MD * 8)) as *i64; let vE: *i64 = (sys_mmap(MV * MD * 8)) as *i64
133 let mW1: *i64 = (sys_mmap(MH * MND * 8)) as *i64; let vW1: *i64 = (sys_mmap(MH * MND * 8)) as *i64
134 let mb1: *i64 = (sys_mmap(MH * 8)) as *i64; let vb1: *i64 = (sys_mmap(MH * 8)) as *i64
135 let mW2: *i64 = (sys_mmap(MC * MH * 8)) as *i64; let vW2: *i64 = (sys_mmap(MC * MH * 8)) as *i64
136 let mb2: *i64 = (sys_mmap(MC * 8)) as *i64; let vb2: *i64 = (sys_mmap(MC * 8)) as *i64
137 z = 0
138 while z < MV * MD { mE[z] = TA_F32_ZERO; vE[z] = TA_F32_ZERO; z = z + 1 }
139 z = 0
140 while z < MH * MND { mW1[z] = TA_F32_ZERO; vW1[z] = TA_F32_ZERO; z = z + 1 }
141 z = 0
142 while z < MH { mb1[z] = TA_F32_ZERO; vb1[z] = TA_F32_ZERO; z = z + 1 }
143 z = 0
144 while z < MC * MH { mW2[z] = TA_F32_ZERO; vW2[z] = TA_F32_ZERO; z = z + 1 }
145 z = 0
146 while z < MC { mb2[z] = TA_F32_ZERO; vb2[z] = TA_F32_ZERO; z = z + 1 }
147
148 let beta1: i64 = ta_constf(9, 10); let beta2: i64 = ta_constf(999, 1000)
149 let om1: i64 = ta_constf(1, 10); let om2: i64 = ta_constf(1, 1000)
150 let lr: i64 = ta_constf(1, 50); let eps: i64 = ta_constf(1, 100000000)
151 var b1t: i64 = TA_F32_ONE; var b2t: i64 = TA_F32_ONE
152
153 let c8: *i64 = (sys_mmap(8)) as *i64; c8[0] = ta_constf(1, 8)
154 let wb: *i64 = (sys_mmap(4 * 8)) as *i64
155 let xl8: *i64 = (sys_mmap(8 * 8)) as *i64
156 let gW1: *i64 = (sys_mmap(MH * MND * 8)) as *i64
157 let gb1: *i64 = (sys_mmap(MH * 8)) as *i64
158 let gW2: *i64 = (sys_mmap(MC * MH * 8)) as *i64
159 let gb2: *i64 = (sys_mmap(MC * 8)) as *i64
160 let dE: *i64 = (sys_mmap(MV * MD * 8)) as *i64
161
162 var ep: i64 = 0
163 while ep < epochs {
164 let loss: i64 = mo_build(tape, vals, st, E, W1, b1, W2, b2, seqs, labels, c8, wb, xl8)
165 ta_backward(tape, vals, grads, st[0], loss)
166 if ep == 0 { *lf = ta_val(tape, vals, loss, 0) }
167 *ll = ta_val(tape, vals, loss, 0)
168 // read shared-leaf grads
169 var i: i64 = 0
170 while i < MH * MND { gW1[i] = ta_grad(tape, grads, wb[0], i); i = i + 1 }
171 i = 0
172 while i < MH { gb1[i] = ta_grad(tape, grads, wb[1], i); i = i + 1 }
173 i = 0
174 while i < MC * MH { gW2[i] = ta_grad(tape, grads, wb[2], i); i = i + 1 }
175 i = 0
176 while i < MC { gb2[i] = ta_grad(tape, grads, wb[3], i); i = i + 1 }
177 // scatter input-leaf grads into the embedding table
178 i = 0
179 while i < MV * MD { dE[i] = TA_F32_ZERO; i = i + 1 }
180 var s: i64 = 0
181 while s < 8 {
182 var pos: i64 = 0
183 while pos < MN {
184 let tok: i64 = seqs[s * MN + pos]
185 var j: i64 = 0
186 while j < MD {
187 dE[tok * MD + j] = nx_f32_add(dE[tok * MD + j], ta_grad(tape, grads, xl8[s], pos * MD + j))
188 j = j + 1
189 }
190 pos = pos + 1
191 }
192 s = s + 1
193 }
194 // adamw step
195 b1t = nx_f32_mul(b1t, beta1); b2t = nx_f32_mul(b2t, beta2)
196 let c1: i64 = nx_f32_sub(TA_F32_ONE, b1t); let c2: i64 = nx_f32_sub(TA_F32_ONE, b2t)
197 mo_adamw(E, mE, vE, dE, MV * MD, lr, beta1, beta2, om1, om2, eps, c1, c2)
198 mo_adamw(W1, mW1, vW1, gW1, MH * MND, lr, beta1, beta2, om1, om2, eps, c1, c2)
199 mo_adamw(b1, mb1, vb1, gb1, MH, lr, beta1, beta2, om1, om2, eps, c1, c2)
200 mo_adamw(W2, mW2, vW2, gW2, MC * MH, lr, beta1, beta2, om1, om2, eps, c1, c2)
201 mo_adamw(b2, mb2, vb2, gb2, MC, lr, beta1, beta2, om1, om2, eps, c1, c2)
202 ep = ep + 1
203 }
204 return 0
205}
206
207func main() -> i64 {
208 var ok: i64 = 1
209 let tape: *i64 = (sys_mmap(2048 * 7 * 8)) as *i64
210 let vals: *i64 = (sys_mmap(8192 * 8)) as *i64
211 let grads: *i64 = (sys_mmap(8192 * 8)) as *i64
212 let st: *i64 = (sys_mmap(2 * 8)) as *i64
213
214 // dataset: 8 sequences, label = (token[0] > token[3]) ? 1 : 0
215 let seqs: *i64 = (sys_mmap(32 * 8)) as *i64
216 seqs[0]=0; seqs[1]=1; seqs[2]=2; seqs[3]=3
217 seqs[4]=3; seqs[5]=2; seqs[6]=1; seqs[7]=0
218 seqs[8]=1; seqs[9]=1; seqs[10]=2; seqs[11]=2
219 seqs[12]=2; seqs[13]=2; seqs[14]=1; seqs[15]=1
220 seqs[16]=0; seqs[17]=0; seqs[18]=3; seqs[19]=3
221 seqs[20]=3; seqs[21]=3; seqs[22]=0; seqs[23]=0
222 seqs[24]=1; seqs[25]=2; seqs[26]=3; seqs[27]=0
223 seqs[28]=2; seqs[29]=1; seqs[30]=0; seqs[31]=3
224 let labels: *i64 = (sys_mmap(8 * 8)) as *i64
225 labels[0]=0; labels[1]=1; labels[2]=0; labels[3]=1; labels[4]=0; labels[5]=1; labels[6]=1; labels[7]=0
226
227 let E: *i64 = (sys_mmap(MV * MD * 8)) as *i64
228 let W1: *i64 = (sys_mmap(MH * MND * 8)) as *i64
229 let b1: *i64 = (sys_mmap(MH * 8)) as *i64
230 let W2: *i64 = (sys_mmap(MC * MH * 8)) as *i64
231 let b2: *i64 = (sys_mmap(MC * 8)) as *i64
232 let lf: *i64 = (sys_mmap(8)) as *i64
233 let ll: *i64 = (sys_mmap(8)) as *i64
234 mo_train(tape, vals, grads, st, E, W1, b1, W2, b2, seqs, labels, 2500, lf, ll)
235
236 var acc: i64 = 0
237 var s: i64 = 0
238 while s < 8 {
239 if mo_predict(tape, vals, st, E, W1, b1, W2, b2, seqs, s) == labels[s] { acc = acc + 1 }
240 s = s + 1
241 }
242 var trainPass: i64 = 1
243 if acc < 7 { trainPass = 0 }
244 if nx_f32_lt(*ll, *lf) != 1 { trainPass = 0 }
245 if trainPass != 1 { ok = 0 }
246
247 // bit-exact reproducibility
248 let E2: *i64 = (sys_mmap(MV * MD * 8)) as *i64
249 let W1b: *i64 = (sys_mmap(MH * MND * 8)) as *i64
250 let b1b: *i64 = (sys_mmap(MH * 8)) as *i64
251 let W2b: *i64 = (sys_mmap(MC * MH * 8)) as *i64
252 let b2b: *i64 = (sys_mmap(MC * 8)) as *i64
253 let lf2: *i64 = (sys_mmap(8)) as *i64
254 let ll2: *i64 = (sys_mmap(8)) as *i64
255 mo_train(tape, vals, grads, st, E2, W1b, b1b, W2b, b2b, seqs, labels, 2500, lf2, ll2)
256 var acc2: i64 = 0
257 s = 0
258 while s < 8 {
259 if mo_predict(tape, vals, st, E2, W1b, b1b, W2b, b2b, seqs, s) == labels[s] { acc2 = acc2 + 1 }
260 s = s + 1
261 }
262 var reproPass: i64 = 1
263 if acc2 != acc { reproPass = 0 }
264 if *ll2 != *ll { reproPass = 0 }
265 if reproPass != 1 { ok = 0 }
266
267 var fdi: i64 = 1
268 while fdi >= 0 {
269 var out: i64 = 1
270 if fdi == 0 { out = sys_openat_append(MO_LOG, 420) }
271 if out >= 0 {
272 mo_w(out, "FNETMODELGATE authored=organ model=embed+fnet-mix+relu-ffn+softmaxCE task=t0>t3 no-attention" as *u8)
273 mo_w(out, " | accuracy=" as *u8); mo_wn(out, acc); mo_w(out, "/8" as *u8)
274 mo_w(out, " loss_first_milli=" as *u8); mo_wn(out, ta_f32_to_milli(*lf))
275 mo_w(out, " loss_last_milli=" as *u8); mo_wn(out, ta_f32_to_milli(*ll))
276 mo_w(out, " | bitexact_repro=" as *u8); mo_wn(out, reproPass)
277 if ok == 1 { mo_w(out, " verdict=GREEN\n" as *u8) } else { mo_w(out, " verdict=RED\n" as *u8) }
278 if fdi == 0 { sys_close(out) }
279 }
280 fdi = fdi - 1
281 }
282
283 if ok == 1 { return 0 }
284 return 1
285}