code wiki / _hdl_build / nx_nofloat_lm_gate.nx
`nx_nofloat_lm_gate.nx` source

↩ module page · 295 lines · 17415 B
1// nx_nofloat_lm_gate.nx -- THE CAPSTONE: a tiny TRANSFORMER LANGUAGE MODEL trains END-TO-END on next-token
2// prediction, in PURE INTEGER Q16 (CAP-NF-LM). Full LM forward:
3//   X = embed(E, ids)  ->  pre-norm transformer block (attn+RoPE+causal-softmax+SwiGLU-FFN+residuals)
4//   ->  rmsnorm_rows  ->  logits = H_n . W_lm  ->  loss = softmax-cross-entropy(logits, next-token ids)
5// Composed ENTIRELY from gradcheck-verified ops + the two new LM-head ops (embed gather/scatter, fused softce).
6//
7//   A1 embed gradcheck   : dE (gather backward = scatter-add) == finite difference.
8//   A2 softce gradcheck  : dlogits (fused CE identity softmax-onehot) == finite difference.
9//   A3 full-LM gradcheck wrt W_lm : the LM-head gradient through CE == finite diff.
10//   A4 full-LM gradcheck wrt E    : the LONGEST chain -- E's gradient flows back through head+norm+FFN+attention
11//      +RoPE+softmax+embed (the WHOLE model) == finite diff. This proves end-to-end backprop of the LM.
12//   D  neg-control teeth ; C bit-exact.
13//   B  THE LM FITS (memorizes a FIXED sequence)     : train E + W_lm (through the full forward; gradient traverses the whole block) on a
14//      next-token task (ids 0,1,2,3 -> 1,2,3,0); assert CE loss drops substantially from ~ln(V).
15//
16// Evidence -> knowledge/status/nofloat_lm.log. Sovereign: nx_nofloat_autograd + nx_syscalls (pure integer).
17// HONEST scope: single-head, single-block, tiny vocab; multi-head/GQA + deeper stack + scale remain. expect_exit: 0
18import "nx_nofloat_autograd.nx"
19import "nx_syscalls.nx"
20import "nx_gate_emit_lib.nx"
21
22const LLOG: *u8 = "knowledge/status/nofloat_lm.log"
23const Q16:  i64 = 65536
24
25
26func g_abs(v: i64) -> i64 { if v < 0 { return 0 - v } return v }
27func l_ws(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 l_wn(fd: i64, v: i64) -> i64 { let b: *u8=sys_mmap(28); var m: i64=v; if m<0{sys_write(fd,"-" as *u8,1);m=0-m} let t: *u8=sys_mmap(28); var k: i64=0; if m==0{t[0]=48;k=1} while m>0{t[k]=(48+(m%10)) as u8;m=m/10;k=k+1} var i: i64=0; while i<k{b[i]=t[k-1-i];i=i+1} sys_write(fd,b,k); return 0 }
29// deterministic small nonzero init (zero-init = dead-gradient trap): values in ~[-0.25, 0.25] Q16
30func lm_init(arr: *i64, n: i64, seed: i64) -> i64 { var i: i64=0; while i<n { arr[i] = (((i*7 + seed*13 + 1) % 11) - 5) * 3277; i=i+1 } return 0 }
31
32// ---- full LM forward; W = ptr-array of 9 weight arrays [E,Wq,Wk,Wv,Wo,Wg,Wu,Wd,Wlm]; leaves[i]<->W[i], leaves[9]=logits ----
33func lm_fwd(tape: *i64, vals: *i64, st: *i64, W: *i64, ids: *i64, tgt: *i64, T: i64, dm: i64, ffn: i64, V: i64, scale: i64, leaves: *i64) -> i64 {
  let E: *i64 = W[0] as *i64; let Wq: *i64 = W[1] as *i64; let Wk: *i64 = W[2] as *i64; let Wv: *i64 = W[3] as *i64; let Wo: *i64 = W[4] as *i64
  let Wg: *i64 = W[5] as *i64; let Wu: *i64 = W[6] as *i64; let Wd: *i64 = W[7] as *i64; let Wlm: *i64 = W[8] as *i64
  st[0]=0; st[1]=0
  let nE:  i64 = nfa_leaf(tape,vals,st,V,dm,E,0)
  let nWq: i64 = nfa_leaf(tape,vals,st,dm,dm,Wq,0)
  let nWk: i64 = nfa_leaf(tape,vals,st,dm,dm,Wk,0)
  let nWv: i64 = nfa_leaf(tape,vals,st,dm,dm,Wv,0)
  let nWo: i64 = nfa_leaf(tape,vals,st,dm,dm,Wo,0)
  let nWg: i64 = nfa_leaf(tape,vals,st,dm,ffn,Wg,0)
  let nWu: i64 = nfa_leaf(tape,vals,st,dm,ffn,Wu,0)
  let nWd: i64 = nfa_leaf(tape,vals,st,ffn,dm,Wd,0)
  let nWlm:i64 = nfa_leaf(tape,vals,st,dm,V,Wlm,0)
  let nX:  i64 = nfa_embed(tape,vals,st,nE,ids,T)
  let nXn: i64 = nfa_rmsnorm_rows(tape,vals,st,nX)
  let nQ:  i64 = nfa_matmul(tape,vals,st,nXn,nWq)
  let nK:  i64 = nfa_matmul(tape,vals,st,nXn,nWk)
  let nV:  i64 = nfa_matmul(tape,vals,st,nXn,nWv)
  let nQr: i64 = nfa_rope(tape,vals,st,nQ)
  let nKr: i64 = nfa_rope(tape,vals,st,nK)
  let nS:  i64 = nfa_matmul_nt(tape,vals,st,nQr,nKr)
  let nSs: i64 = nfa_cmul(tape,vals,st,nS,scale)
  let nA:  i64 = nfa_softmax_rows(tape,vals,st,nSs,1)
  let nO:  i64 = nfa_matmul(tape,vals,st,nA,nV)
  let nOp: i64 = nfa_matmul(tape,vals,st,nO,nWo)
  let nH:  i64 = nfa_vadd(tape,vals,st,nX,nOp)
  let nHn: i64 = nfa_rmsnorm_rows(tape,vals,st,nH)
  let nG:  i64 = nfa_matmul(tape,vals,st,nHn,nWg)
  let nU:  i64 = nfa_matmul(tape,vals,st,nHn,nWu)
  let nSg: i64 = nfa_silu(tape,vals,st,nG)
  let nHs: i64 = nfa_hadamard(tape,vals,st,nSg,nU)
  let nDp: i64 = nfa_matmul(tape,vals,st,nHs,nWd)
  let nY:  i64 = nfa_vadd(tape,vals,st,nH,nDp)
  let nYn: i64 = nfa_rmsnorm_rows(tape,vals,st,nY)
  let nLg: i64 = nfa_matmul(tape,vals,st,nYn,nWlm)
  let nLoss: i64 = nfa_softce_rows(tape,vals,st,nLg,tgt)
  leaves[0]=nE; leaves[1]=nWq; leaves[2]=nWk; leaves[3]=nWv; leaves[4]=nWo
  leaves[5]=nWg; leaves[6]=nWu; leaves[7]=nWd; leaves[8]=nWlm; leaves[9]=nLg
  return nLoss
72}
73func lm_lossval(tape: *i64, vals: *i64, st: *i64, W: *i64, ids: *i64, tgt: *i64, T: i64, dm: i64, ffn: i64, V: i64, scale: i64) -> i64 {
  let lv: *i64 = sys_mmap(10*8) as *i64
  let nLoss: i64 = lm_fwd(tape,vals,st,W,ids,tgt,T,dm,ffn,V,scale,lv)
  return nfa_val(tape,vals,nLoss,0)
77}
78// gradcheck wrt weight array widx (leaves[widx]); extract ALL analytic grads first (lm_lossval rebuilds the tape).
79func lm_gc_w(tape: *i64, vals: *i64, grads: *i64, st: *i64, W: *i64, ids: *i64, tgt: *i64, T: i64, dm: i64, ffn: i64, V: i64, scale: i64, leaves: *i64, widx: i64, h: i64, tol_q: i64, floor_q: i64, worst: *i64) -> i64 {
  let nLoss: i64 = lm_fwd(tape,vals,st,W,ids,tgt,T,dm,ffn,V,scale,leaves)
  nfa_backward(tape,vals,grads,st[0],nLoss)
  let nnode: i64 = leaves[widx]
  let cnt: i64 = tape[7*nnode+3]*tape[7*nnode+4]
  let ana: *i64 = sys_mmap(64*8) as *i64
  var c: i64=0
  while c<cnt { ana[c]=nfa_grad(tape,grads,nnode,c); c=c+1 }
  let arr: *i64 = W[widx] as *i64
  var ok: i64=1; worst[0]=0
  var i: i64=0
  while i<cnt {
      let old: i64 = arr[i]
      arr[i]=old+h; let lp: i64 = lm_lossval(tape,vals,st,W,ids,tgt,T,dm,ffn,V,scale)
      arr[i]=old-h; let lm2: i64 = lm_lossval(tape,vals,st,W,ids,tgt,T,dm,ffn,V,scale)
      arr[i]=old
      let fd: i64 = ((lp-lm2)*Q16)/(2*h)
      let num: i64 = g_abs(fd-ana[i])
      var den: i64 = g_abs(ana[i]); if den<floor_q { den=floor_q }
      if num >= ((tol_q*den)>>16) { ok=0 }
      let rel: i64 = (num*1000)/den
      if rel>worst[0] { worst[0]=rel }
      i=i+1
  }
  return ok
104}
105
106// ---- A1 embed standalone: loss = mse(embed(E,ids), tgt) ----
107func emb_loss(tape: *i64, vals: *i64, st: *i64, E: *i64, ids: *i64, Tg: *i64, V: i64, dm: i64, T: i64, leaves: *i64) -> i64 {
  st[0]=0; st[1]=0
  let nE: i64 = nfa_leaf(tape,vals,st,V,dm,E,0)
  let nX: i64 = nfa_embed(tape,vals,st,nE,ids,T)
  let nt: i64 = nfa_leaf(tape,vals,st,T,dm,Tg,0)
  let loss: i64 = nfa_mse(tape,vals,st,nX,nt)
  leaves[0]=nE
  return loss
115}
116func emb_lossval(tape: *i64, vals: *i64, st: *i64, E: *i64, ids: *i64, Tg: *i64, V: i64, dm: i64, T: i64) -> i64 {
  let lv: *i64 = sys_mmap(8) as *i64
  let loss: i64 = emb_loss(tape,vals,st,E,ids,Tg,V,dm,T,lv)
  return nfa_val(tape,vals,loss,0)
120}
121
122// ---- A2 softce standalone: loss = softce(logits_leaf, tgt) ----
123func ce_loss(tape: *i64, vals: *i64, st: *i64, L: *i64, tgt: *i64, T: i64, V: i64, leaves: *i64) -> i64 {
  st[0]=0; st[1]=0
  let nL: i64 = nfa_leaf(tape,vals,st,T,V,L,0)
  let loss: i64 = nfa_softce_rows(tape,vals,st,nL,tgt)
  leaves[0]=nL
  return loss
129}
130func ce_lossval(tape: *i64, vals: *i64, st: *i64, L: *i64, tgt: *i64, T: i64, V: i64) -> i64 {
  let lv: *i64 = sys_mmap(8) as *i64
  let loss: i64 = ce_loss(tape,vals,st,L,tgt,T,V,lv)
  return nfa_val(tape,vals,loss,0)
134}
135
136func main() -> i64 {
  g_puts("nx_nofloat_lm gate (a tiny TRANSFORMER LM trains end-to-end on next-token, PURE INTEGER Q16)\n" as *u8)
  var pass: i64=0; var total: i64=0
  let tape:  *i64 = sys_mmap(512*7*8) as *i64
  let vals:  *i64 = sys_mmap(16384*8) as *i64
  let grads: *i64 = sys_mmap(16384*8) as *i64
  let st:    *i64 = sys_mmap(2*8) as *i64
  let h: i64 = 512; let floor_q: i64 = 4096
  let worst: *i64 = sys_mmap(8) as *i64
145
  // ---- A1: embed gradcheck (V=3,dm=2,T=3; ids gather rows 0,2,1; one repeats nothing here) ----
  let eE: *i64 = sys_mmap(6*8) as *i64; lm_init(eE,6,1)
  let eids: *i64 = sys_mmap(3*8) as *i64; eids[0]=0; eids[1]=2; eids[2]=1
  let eTg: *i64 = sys_mmap(6*8) as *i64; eTg[0]=13107; eTg[1]=0-6554; eTg[2]=19661; eTg[3]=6554; eTg[4]=0-26214; eTg[5]=32768
  let elv: *i64 = sys_mmap(8) as *i64
  let el: i64 = emb_loss(tape,vals,st,eE,eids,eTg,3,2,3,elv)
  nfa_backward(tape,vals,grads,st[0],el)
  let enode: i64 = elv[0]
  let eana: *i64 = sys_mmap(8*8) as *i64
  var ec: i64=0
  while ec<6 { eana[ec]=nfa_grad(tape,grads,enode,ec); ec=ec+1 }
  var e_ok: i64=1; var e_worst: i64=0
  var ei: i64=0
  while ei<6 {
      let old: i64=eE[ei]; eE[ei]=old+h; let lp: i64=emb_lossval(tape,vals,st,eE,eids,eTg,3,2,3); eE[ei]=old-h; let lm2: i64=emb_lossval(tape,vals,st,eE,eids,eTg,3,2,3); eE[ei]=old
      let fd: i64=((lp-lm2)*Q16)/(2*h); let num: i64=g_abs(fd-eana[ei]); var den: i64=g_abs(eana[ei]); if den<floor_q{den=floor_q}
      if num >= ((4096*den)>>16) { e_ok=0 }
      let rel: i64=(num*1000)/den; if rel>e_worst{e_worst=rel}
      ei=ei+1
  }
  g_puts("  [measure] embed worst rel grad err = " as *u8); g_pn(e_worst); g_puts(" /1000 (tol=62)\n" as *u8)
  pass=pass+g_check("A1: embed gradcheck -- gather backward (scatter-add) == finite differences" as *u8, e_ok); total=total+1
168
  // ---- A2: softce gradcheck (T=2,V=3) ----
  let cL: *i64 = sys_mmap(6*8) as *i64; cL[0]=32768; cL[1]=0-16384; cL[2]=49152; cL[3]=16384; cL[4]=65536; cL[5]=0-32768
  let ctg: *i64 = sys_mmap(2*8) as *i64; ctg[0]=1; ctg[1]=2
  let clv: *i64 = sys_mmap(8) as *i64
  let cl: i64 = ce_loss(tape,vals,st,cL,ctg,2,3,clv)
  nfa_backward(tape,vals,grads,st[0],cl)
  let cnode: i64 = clv[0]
  let cana: *i64 = sys_mmap(8*8) as *i64
  var cc: i64=0
  while cc<6 { cana[cc]=nfa_grad(tape,grads,cnode,cc); cc=cc+1 }
  var c_ok: i64=1; var c_worst: i64=0
  var ci: i64=0
  while ci<6 {
      let old: i64=cL[ci]; cL[ci]=old+h; let lp: i64=ce_lossval(tape,vals,st,cL,ctg,2,3); cL[ci]=old-h; let lm2: i64=ce_lossval(tape,vals,st,cL,ctg,2,3); cL[ci]=old
      let fd: i64=((lp-lm2)*Q16)/(2*h); let num: i64=g_abs(fd-cana[ci]); var den: i64=g_abs(cana[ci]); if den<floor_q{den=floor_q}
      if num >= ((8192*den)>>16) { c_ok=0 }       // tol 1/8 (ln/exp fixed-point approximation)
      let rel: i64=(num*1000)/den; if rel>c_worst{c_worst=rel}
      ci=ci+1
  }
  g_puts("  [measure] softce worst rel grad err = " as *u8); g_pn(c_worst); g_puts(" /1000 (tol=125)\n" as *u8)
  pass=pass+g_check("A2: softce gradcheck -- fused softmax-CE (softmax-onehot) backward == finite differences" as *u8, c_ok); total=total+1
190
  // ---- LM dims + weights ----
  let T: i64=4; let dm: i64=4; let ffn: i64=8; let V: i64=4; let scale: i64=32768   // 1/sqrt(4)=0.5
  let E:  *i64 = sys_mmap(V*dm*8) as *i64;  lm_init(E,V*dm,1)
  let Wq: *i64 = sys_mmap(dm*dm*8) as *i64; lm_init(Wq,dm*dm,2)
  let Wk: *i64 = sys_mmap(dm*dm*8) as *i64; lm_init(Wk,dm*dm,3)
  let Wv: *i64 = sys_mmap(dm*dm*8) as *i64; lm_init(Wv,dm*dm,4)
  let Wo: *i64 = sys_mmap(dm*dm*8) as *i64; lm_init(Wo,dm*dm,5)
  let Wg: *i64 = sys_mmap(dm*ffn*8) as *i64; lm_init(Wg,dm*ffn,6)
  let Wu: *i64 = sys_mmap(dm*ffn*8) as *i64; lm_init(Wu,dm*ffn,7)
  let Wd: *i64 = sys_mmap(ffn*dm*8) as *i64; lm_init(Wd,ffn*dm,8)
  let Wlm:*i64 = sys_mmap(dm*V*8) as *i64;  lm_init(Wlm,dm*V,9)
  let W: *i64 = sys_mmap(9*8) as *i64
  W[0]=E as i64; W[1]=Wq as i64; W[2]=Wk as i64; W[3]=Wv as i64; W[4]=Wo as i64; W[5]=Wg as i64; W[6]=Wu as i64; W[7]=Wd as i64; W[8]=Wlm as i64
  let ids: *i64 = sys_mmap(T*8) as *i64; ids[0]=0; ids[1]=1; ids[2]=2; ids[3]=3
  let tgt: *i64 = sys_mmap(T*8) as *i64; tgt[0]=1; tgt[1]=2; tgt[2]=3; tgt[3]=0
  let leaves: *i64 = sys_mmap(10*8) as *i64
207
  // ---- A3: full-LM gradcheck wrt Wlm (widx 8) ----
  let a3_ok: i64 = lm_gc_w(tape,vals,grads,st,W,ids,tgt,T,dm,ffn,V,scale,leaves,8,h,8192,floor_q,worst)
  g_puts("  [measure] full-LM dL/dWlm worst rel grad err = " as *u8); g_pn(worst[0]); g_puts(" /1000 (tol=125)\n" as *u8)
  pass=pass+g_check("A3: full-LM gradcheck wrt Wlm (LM head through CE) == finite differences" as *u8, a3_ok); total=total+1
212
  // ---- A4: full-LM gradcheck wrt E (widx 0) -- the LONGEST chain (head+norm+FFN+attn+RoPE+softmax+embed) ----
  let a4_ok: i64 = lm_gc_w(tape,vals,grads,st,W,ids,tgt,T,dm,ffn,V,scale,leaves,0,h,16384,floor_q,worst)
  g_puts("  [measure] full-LM dL/dE worst rel grad err = " as *u8); g_pn(worst[0]); g_puts(" /1000 (tol=250)\n" as *u8)
  pass=pass+g_check("A4: full-LM gradcheck wrt E through the WHOLE model == finite differences (end-to-end backprop)" as *u8, a4_ok); total=total+1
217
  // ---- D: neg-control teeth (Wlm grad) ----
  let nLd: i64 = lm_fwd(tape,vals,st,W,ids,tgt,T,dm,ffn,V,scale,leaves)
  nfa_backward(tape,vals,grads,st[0],nLd)
  let dana: i64 = nfa_grad(tape,grads,leaves[8],0)
  let old0: i64 = Wlm[0]; Wlm[0]=old0+h; let lpd: i64=lm_lossval(tape,vals,st,W,ids,tgt,T,dm,ffn,V,scale); Wlm[0]=old0-h; let lmd: i64=lm_lossval(tape,vals,st,W,ids,tgt,T,dm,ffn,V,scale); Wlm[0]=old0
  let dfd: i64 = ((lpd-lmd)*Q16)/(2*h)
  let dbad: i64 = 0 - dana
  var dden: i64 = g_abs(dana); if dden<floor_q { dden=floor_q }
  var caught: i64 = 1
  if g_abs(dfd-dbad) < ((8192*dden)>>16) { caught=0 }
  pass=pass+g_check("D: neg-control -- a WRONG LM gradient is rejected (teeth)" as *u8, caught); total=total+1
229
  // ---- B: the LM FITS (memorizes a FIXED sequence) next-token (train E + Wlm through the full forward; grad traverses the whole block) ----
  let gbuf: *i64 = sys_mmap(64*8) as *i64
  var lf: i64=0; var ll: i64=0
  var ep: i64=0
  while ep < 8000 {
      let nLoss: i64 = lm_fwd(tape,vals,st,W,ids,tgt,T,dm,ffn,V,scale,leaves)
      nfa_backward(tape,vals,grads,st[0],nLoss)
      if ep==0 { lf = nfa_val(tape,vals,nLoss,0) }
      ll = nfa_val(tape,vals,nLoss,0)
      // update E (widx 0) and Wlm (widx 8)
      let nE2: i64 = leaves[0]; let cE: i64 = V*dm
      var z: i64=0
      while z<cE { gbuf[z]=nfa_grad(tape,grads,nE2,z); z=z+1 }
      nfa_sgd(E, gbuf, cE, 2048)
      let nWl: i64 = leaves[8]; let cWl: i64 = dm*V
      z=0
      while z<cWl { gbuf[z]=nfa_grad(tape,grads,nWl,z); z=z+1 }
      nfa_sgd(Wlm, gbuf, cWl, 2048)
      ep=ep+1
  }
  g_puts("  [measure] LM next-token CE loss: start=" as *u8); g_pn(lf); g_puts(" end=" as *u8); g_pn(ll); g_puts("  (ln(V)=ln4 ~ 90852 Q16)\n" as *u8)
  var fits: i64 = 1
  if ll*2 > lf { fits=0 }            // >= 50% CE reduction = it fit the fixed sequence (memorization, not generalization)
  if lf <= 0 { fits=0 }
  pass=pass+g_check("B: the LM FITS (memorizes a FIXED sequence) -- next-token CE loss drops >=50% via end-to-end backprop (E+Wlm trained)" as *u8, fits); total=total+1
255
  // ---- C: bit-exact ----
  let E2: *i64 = sys_mmap(V*dm*8) as *i64;  lm_init(E2,V*dm,1)
  let Wlm2:*i64 = sys_mmap(dm*V*8) as *i64;  lm_init(Wlm2,dm*V,9)
  let W2: *i64 = sys_mmap(9*8) as *i64
  W2[0]=E2 as i64; W2[1]=Wq as i64; W2[2]=Wk as i64; W2[3]=Wv as i64; W2[4]=Wo as i64; W2[5]=Wg as i64; W2[6]=Wu as i64; W2[7]=Wd as i64; W2[8]=Wlm2 as i64
  let lvc: *i64 = sys_mmap(10*8) as *i64
  let gbuf2: *i64 = sys_mmap(64*8) as *i64
  var ep2: i64=0
  while ep2 < 8000 {
      let nLoss: i64 = lm_fwd(tape,vals,st,W2,ids,tgt,T,dm,ffn,V,scale,lvc)
      nfa_backward(tape,vals,grads,st[0],nLoss)
      let nE2: i64=lvc[0]; var z: i64=0
      while z<V*dm { gbuf2[z]=nfa_grad(tape,grads,nE2,z); z=z+1 }
      nfa_sgd(E2, gbuf2, V*dm, 2048)
      let nWl: i64=lvc[8]; z=0
      while z<dm*V { gbuf2[z]=nfa_grad(tape,grads,nWl,z); z=z+1 }
      nfa_sgd(Wlm2, gbuf2, dm*V, 2048)
      ep2=ep2+1
  }
  var bitexact: i64 = 1
  var zc: i64=0
  while zc<V*dm { if E2[zc]!=E[zc] { bitexact=0 } zc=zc+1 }
  zc=0
  while zc<dm*V { if Wlm2[zc]!=Wlm[zc] { bitexact=0 } zc=zc+1 }
  pass=pass+g_check("C: bit-exact -- training the LM twice gives IDENTICAL integer weights (determinism)" as *u8, bitexact); total=total+1
281
  // ---- emit ----
  var okall: i64=0; if pass==total { okall=1 }
  let logf: i64 = sys_openat_append(LLOG, 420)
  if logf >= 0 {
      l_ws(logf,"NOFLOATLM 1-block 1-head V=4 A1_embed=" as *u8); l_wn(logf,e_ok); l_ws(logf," A2_softce=" as *u8); l_wn(logf,c_ok)
      l_ws(logf," A3_Wlm=" as *u8); l_wn(logf,a3_ok); l_ws(logf," A4_E_fullchain=" as *u8); l_wn(logf,a4_ok); l_ws(logf," D=" as *u8); l_wn(logf,caught)
      l_ws(logf," B_fits=" as *u8); l_wn(logf,fits); l_ws(logf," CE_start=" as *u8); l_wn(logf,lf); l_ws(logf," CE_end=" as *u8); l_wn(logf,ll); l_ws(logf," C_bitexact=" as *u8); l_wn(logf,bitexact)
      if okall==1 { l_ws(logf," verdict=GREEN\n" as *u8) } else { l_ws(logf," verdict=RED\n" as *u8) }
      sys_close(logf)
  }
  g_puts("---- nofloat_lm gate: passed " as *u8); g_pn(pass); g_puts(" / " as *u8); g_pn(total); g_puts(" ----\n" as *u8)
  if okall==1 { g_puts("verdict=GREEN (a tiny transformer LM trains end-to-end on next-token prediction in pure integer Q16)\n" as *u8); sys_exit(0); return 0 }
  g_puts("verdict=RED\n" as *u8); sys_exit(1); return 1
295}
nx_nofloat_lm_gate.nx source

`nx_nofloat_lm_gate.nx` source
