code wiki / _hdl_build / nx_intfp_attention_gradcheck_gate.nx
nx_intfp_attention_gradcheck_gate.nx
buildroot/runtime/_hdl_build/nx_intfp_attention_gradcheck_gate.nx
about
nx_intfp_attention_gradcheck_gate.nx -- THE COMPOSITION PROOF: a full single-head CAUSAL self-attention block
(Q,K,V = X@W ; scores = scale*QK^T ; A = softmax(scores) causal ; O = A@V ; L = sum(O^2)) with the COMPLETE
backward, done ENTIRELY in Q16 integer, gradchecked (integer finite-diff) on all three projection matrices
Wq/Wk/Wv. Wq/Wk gradients flow through the softmax Jacobian AND two matmul-transpose backprops -- if THESE
pass, the integer ops compose correctly and the full transformer tape is assemblable. No float anywhere.
scale = 1/sqrt(d); d=4 -> scale=0.5 exactly (Q16 32768). license_tier: ORIGINAL
dependencies 1 imports · 0 importers
imports: nx_syscalls.nx
imported by: nobody (leaf or entry point)
call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown
structs
| none |
consts
| 13 | const S: i64 = 1048576 // Q20 -- the BACKWARD needs more than Q16 for small gradients through the softmax Jacobian |
| 14 | const T: i64 = 3 |
| 15 | const DM: i64 = 4 |
| 16 | const SCALE: i64 = 524288 // 1/sqrt(4) = 0.5 in Q20 |
functions
| 9 | func w(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 } |
| 10 | func wn(v: i64) -> i64 { if v==0 { sys_write(1,"0" as *u8,1); return 0 } var m: i64=v; if m<0{sys_write(1,"-" as *u8,1);m=0-m} let t: *u8=sys_mmap(24); var k: i64=0; while m>0{t[k]=(48+(m%10)) as u8;m=m/10;k=k+1} let o: *u8=sys_mmap(24); var q: i64=k-1; var i: i64=0; while q>=0{o[i]=t[q];i=i+1;q=q-1} sys_write(1,o,i); return 0 } |
| 11 | func iabs(v: i64) -> i64 { if v<0 { return 0-v } return v } called by 1: gcheck |
| 19 | func fp_exp_q16(xq: i64) -> i64 called by 1: attn_fwd |
| 32 | func attn_fwd(X: *i64, Wq: *i64, Wk: *i64, Wv: *i64, Q: *i64, K: *i64, V: *i64, A: *i64, O: *i64) -> i64 |
| 53 | func attn_bwd(X: *i64, Q: *i64, K: *i64, V: *i64, A: *i64, O: *i64, dWq: *i64, dWk: *i64, dWv: *i64) -> i64 |
| 75 | func gcheck(name: *u8, X: *i64, Wq: *i64, Wk: *i64, Wv: *i64, Wtarget: *i64, dW: *i64, Q: *i64, K: *i64, V: *i64, A: *i64, O: *i64) -> i64 |
| 93 | func main() -> i64 |