code wiki / (root) / nx_gpu_serve_gate.nx

nx_gpu_serve_gate.nx

buildroot/runtime/nx_gpu_serve_gate.nx

10610 B149 linesdepth 8pulls 19 transitivereach 0 importersview sourcekind gate/prooftopic gpu
docsdependenciesstructsconstsfunctions

about

nx_gpu_serve_gate.nx -- #23 KEYSTONE: the SOVEREIGN serve-side GPU client, gated byte-exact. Loads the CODER model (nx_coder_model.gguf), generates NGEN greedy tokens on the proven CPU path (sequential decode_step_kv_cached_i8 + i32 head -- the serve math), then repeats the SAME generation via the GPU logits server on unix socket /home/elderwesto/nx_stage/nx_gpu.sock: per token the client sends [pos: i64][x1: ne i64 embedding row] and receives [logits: vocab i64], argmaxes on CPU (lowest-id tie-break = head_argmax convention). GATE: token ids IDENTICAL CPU vs GPU for the prompt-next + all NGEN steps. The C server is the BOOTSTRAP ORACLE (doctrine: 3rd parties = benchmarks); THIS client is the sovereign integration shape for nx_nofloat_serve_core. Requires the server running: /tmp/gls --serve. expect_exit: 0 license_tier: ORIGINAL

dependencies 10 imports · 0 importers

nx_syscalls.nx nx_tier.nx nx_le.nx nx_tensor.nx nx_gguf.nx nx_gguf_load.nx nx_gguf_meta.nx nx_nofloat_llm.nx nx_nofloat_tok.nx nx_gate_verdict.nx nx_gpu_serve_gate.nx

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

main gs_puts sys_write sys_mmap sys_read_file sys_openat_rd sys_lseek sys_mmap ↻ sys_read sys_close nx_gguf_parse nx_le_read_u32 nx_le_read_u64 nx_le_read_u32 ↻ _gguf_skip_value nx_le_read_u64 ↻ nx_le_read_u32 ↻ _gguf_skip_value ↻ sys_mmap ↻ nx_gguf_meta_find nx_le_read_u64 ↻ nx_le_read_u32 ↻ _gmeta_bytes_equal nx_gguf_skip_value _gguf_skip_value ↻ gs_slen nx_gguf_meta_array_count nx_le_read_u64 ↻ nx_gguf_meta_array_first_e nx_gguf_find_tensor nx_gguf_tensor_at nx_gguf_name_equals nx_gguf_tensor_at ↻ load_named_q16 nx_gguf_find_tensor ↻ nx_gguf_tensor_at ↻ nx_gguf_tensor_n_values dequant_to_q16 f32_to_q16 nx_le_read_u32 ↻

structs

none

consts

none

functions

21func gs_puts(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} sys_write(1,s,n); return 0 }
called by 1: main calls 1: sys_write
22func gs_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 }
called by 1: main calls 2: sys_mmapsys_write
23func gs_slen(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} return n }
called by 2: gs_connectmain
25func gs_wall(fd: i64, buf: *u8, count: i64) -> i64 { var off: i64=0; while off<count { let w: i64=sys_write(fd, ((buf as i64)+off) as *u8, count-off); if w<=0 { return 0-1 } off=off+w } return 0 }
called by 1: main calls 1: sys_write
26func gs_rall(fd: i64, buf: *u8, count: i64) -> i64 { var off: i64=0; while off<count { let r: i64=sys_read(fd, ((buf as i64)+off) as *u8, count-off); if r<=0 { return 0-1 } off=off+r } return 0 }
called by 1: main calls 1: sys_read
28func gs_connect(path: *u8) -> i64
called by 1: main calls 4: sys_socketsys_mmapgs_slensys_connect
42func gs_argmax(lgv: *i64, vocab: i64) -> i64 { var best: i64=lgv[0]; var bid: i64=0; var v: i64=1; while v<vocab { if lgv[v]>best { best=lgv[v]; bid=v } v=v+1 } return bid }
called by 1: main
44func main() -> i64
calls 32: gs_putssys_mmapsys_read_filenx_gguf_parsenx_gguf_meta_findgs_slen+26