code wiki / _hdl_build / nx_intfp_block_gradcheck_gate.nx
nx_intfp_block_gradcheck_gate.nx
buildroot/runtime/_hdl_build/nx_intfp_block_gradcheck_gate.nx
about
nx_intfp_block_gradcheck_gate.nx -- CAPSTONE assembly pattern: a pre-norm residual block y = x + Attn(RMSNorm(x))
with the FULL end-to-end integer backward, validating the two NEW assembly risks the isolated gates didn't cover:
(1) INPUT-gradient propagation dX through a sublayer (attention dh), and
(2) RESIDUAL gradient accumulation (dx = dx_residual + dx_norm-path).
Gradchecked (integer finite-diff, max-|grad| metric) on gamma, Wq, Wv AND the input X (which flows through BOTH
the residual and the norm->attention path), then TRAINED (loss drops). If this holds, the full transformer layer
is the same pattern twice (attn block + FFN block). Q20 integer, no float. Ptr-array plumbing to fit arg budget.
slots: 0x 1gamma 2Wq 3Wk 4Wv 5Y | 6h 7Q 8K 9V 10A 11O 12y 13rms | 14dgamma 15dWq 16dWk 17dWv 18dX
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
| 16 | const S: i64 = 1048576 // Q20 |
| 17 | const T: i64 = 3 |
| 18 | const DM: i64 = 4 |
| 19 | const SCALE: i64 = 524288 // 1/sqrt(4)=0.5 |
| 20 | const EPS: i64 = 1048576 // eps in Q40 (tiny vs ms) |
functions
| 12 | 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 } |
| 13 | 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 } |
| 14 | func iabs(v: i64) -> i64 { if v<0 { return 0-v } return v } called by 1: gcheck |
| 22 | func isqrt(n: i64) -> i64 called by 1: block_fwd |
| 29 | func fp_exp(xq: i64) -> i64 called by 1: block_fwd |
| 42 | func block_fwd(P: *i64) -> i64 |
| 67 | func block_bwd(P: *i64) -> i64 |
| 103 | func gcheck(name: *u8, P: *i64, tgtslot: i64, dW: *i64, ncell: i64) -> i64 |
| 122 | func main() -> i64 |