nx_nofloat_k3block.nx source

1// nx_nofloat_k3block.nx -- LOGICALLY-INTEGRATED K3-style transformer BLOCK in no-float (operator 2026-07-19 2// "logically integrated"). Proves the scattered K3 primitives COMPOSE into one deterministic block, not just 3// exist as separate gates. Minimal K3 block = the KDA/linear-attention sublayer + a SwiGLU-ish FFN sublayer, 4// both with residual connections (the DeepSeek-V3/Moonlight skeleton the real K3 uses): 5// attn_in = x ; q,k,v = x.Wq, x.Wk, x.Wv ; a = linattn(q,k,v) ; h = x + a.Wo (KDA sublayer + residual) 6// f = relu(h.W1).W2 ; out = h + f (FFN sublayer + residual) 7// All PURE INTEGER Q16 (matvec accumulate-then-shift + the proven linear-attn kernel) = bit-exact 8// deterministic K3-style block. Composes nx_nofloat_linattn (the KDA foundation). Grounded arXiv 2510.26692. 9// license_tier: ORIGINAL No hw writes (Rule 26). 10import "nx_nofloat_linattn.nx" 11import "nx_syscalls.nx" 12 13const K3_QBITS: i64 = 16 14const K3_DMAX: i64 = 16 15const K3_FFMAX: i64 = 32 16 17// Y[T x n] = X[T x d] . W[d x n] (Q16, accumulate i64 >>16) 18func k3_matmul(x: *i64, w: *i64, y: *i64, t: i64, d: i64, n: i64) -> i64 { 19 var ti: i64 = 0 20 while ti < t { 21 var j: i64 = 0 22 while j < n { 23 var acc: i64 = 0 24 var kk: i64 = 0 25 while kk < d { acc = acc + x[ti*d+kk] * w[kk*n+j]; kk = kk + 1 } 26 y[ti*n+j] = acc >> K3_QBITS 27 j = j + 1 28 } 29 ti = ti + 1 30 } 31 return 0 32} 33// out = a + b (T x d), elementwise 34func k3_add(a: *i64, b: *i64, out: *i64, t: i64, d: i64) -> i64 { 35 var i: i64 = 0 36 while i < t*d { out[i] = a[i] + b[i]; i = i + 1 } 37 return 0 38} 39// ReLU in place (T x d) 40func k3_relu(a: *i64, t: i64, d: i64) -> i64 { 41 var i: i64 = 0 42 while i < t*d { if a[i] < 0 { a[i] = 0 } i = i + 1 } 43 return 0 44} 45// one K3-style block. x: T x d (Q16). weights: Wq,Wk,Wv,Wo [d x d], W1 [d x dff], W2 [dff x d]. out: T x d. 46func k3_block(x: *i64, wq: *i64, wk: *i64, wv: *i64, wo: *i64, w1: *i64, w2: *i64, out: *i64, t: i64, d: i64, dff: i64) -> i64 { 47 let q: *i64 = sys_mmap(K3_DMAX*K3_DMAX*8) as *i64 48 let k: *i64 = sys_mmap(K3_DMAX*K3_DMAX*8) as *i64 49 let v: *i64 = sys_mmap(K3_DMAX*K3_DMAX*8) as *i64 50 let a: *i64 = sys_mmap(K3_DMAX*K3_DMAX*8) as *i64 51 let ao: *i64 = sys_mmap(K3_DMAX*K3_DMAX*8) as *i64 52 let h: *i64 = sys_mmap(K3_DMAX*K3_DMAX*8) as *i64 53 let f1: *i64 = sys_mmap(K3_DMAX*K3_FFMAX*8) as *i64 54 let f2: *i64 = sys_mmap(K3_DMAX*K3_DMAX*8) as *i64 55 // KDA sublayer 56 k3_matmul(x, wq, q, t, d, d) 57 k3_matmul(x, wk, k, t, d, d) 58 k3_matmul(x, wv, v, t, d, d) 59 la_forward(q, k, v, a, t, d) // linear attention (KDA kernel) 60 k3_matmul(a, wo, ao, t, d, d) // output projection 61 k3_add(x, ao, h, t, d) // residual: h = x + attn 62 // FFN sublayer 63 k3_matmul(h, w1, f1, t, d, dff) 64 k3_relu(f1, t, dff) 65 k3_matmul(f1, w2, f2, t, dff, d) 66 k3_add(h, f2, out, t, d) // residual: out = h + ffn 67 return 0 68}