nx_nofloat_qwen_fastgen_gate.nx
buildroot/runtime/nx_nofloat_qwen_fastgen_gate.nx
about
nx_nofloat_qwen_fastgen_gate.nx -- KV-cached generation with a DEQUANT-ONCE weight+head cache.
The dqprobe gate MEASURED that ~88% of a decode step is re-dequantizing frozen weights (1213/1363 ms), and
the streamed LM head re-dequantizes all 152K rows every token (~650 ms). Both are removed here by dequanting
ONCE up front (nf_dequant_all_layers + nf_dequant_head_all) and reusing the cache every step. This is LOSSLESS:
identical dequant, identical accumulation order -> the tokens MUST still be [12095,13,1084,374] (=' Paris' '.'
' It' ' is'), pinned as teeth. Reports one-time cache cost + per-token decode/head split. No hw writes (Rule 26).
expect_exit: 0 license_tier: ORIGINAL
dependencies 10 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.nxnx_gate_verdict.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 fg_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 fg_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 fg_slen(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} return n } called by 1: main |
| 23 | func main() -> i64 |