nx_nofloat_qwen_diag_gate.nx
buildroot/runtime/nx_nofloat_qwen_diag_gate.nx
about
nx_nofloat_qwen_diag_gate.nx -- FAITHFULNESS LOCALIZER (memory-safe: ONE no-float forward, lazy per-layer load).
The config is verified correct (RoPE ln(1e6), QKV bias, scale 0.125) and the matmul agrees with f32 (<64 ulps),
yet the composed 24-layer forward predicts garbage. This instruments the SAME forward `run_stack` does (inlined,
calling the canonical attn_sublayer/ffn_sublayer per layer) and prints per-layer max|hidden| (blow-up / vanish
check) + the final top-5 logits with decoded tokens. Light footprint = single forward, no f32 stack, no OOM.
- per-layer max|hidden| trend: stable => systematic bug; growing/vanishing => Q16 precision/overflow compounding
- top-5 logits: a sharp WRONG peak => systematic; near-uniform => signal washed out (precision)
No hw writes (Rule 26). expect_exit: 0 license_tier: ORIGINAL
dependencies 9 imports · 0 importers
imports: nx_syscalls.nxnx_tier.nxnx_le.nxnx_tensor.nxnx_gguf.nxnx_gguf_load.nxnx_gguf_meta.nxnx_nofloat_llm.nxnx_nofloat_tok.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
| none |
functions
| 19 | func dg_puts(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 } |
| 20 | func dg_num(v: i64) -> i64 { let b: *u8=sys_mmap(28); var m: i64=v; if m<0{m=0-m;sys_write(1,"-" as *u8,1)} let t: *u8=sys_mmap(28); var k: i64=0; if m==0{t[0]=48 as u8;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(1,b,k); return 0 } |
| 21 | func dg_slen(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} return n } called by 1: main |
| 22 | func maxabs(a: *i64, n: i64) -> i64 { var mx: i64=0; var i: i64=0; while i<n { var v: i64=a[i]; if v<0 { v=0-v } if v>mx { mx=v } i=i+1 } return mx } called by 1: main |
| 24 | func main() -> i64 |