nx_q8_0_linear_verify.nx source

1// nx_q8_0_linear_verify.nx -- verify a full Q8_0 linear at REAL DiT scale vs a Python ggml golden. 2// 3// Reads the real Z-Image DiT layers.0.attention.out.weight (Q8_0, [3840,3840], dumped to dit_out_weight.q8) 4// and computes output rows 0..3 = W[o,:] . x with x[i]=(i%7)+1, confirming the row stride (120 blocks * 5// 34B = 4080B/row) + the Q8_0 dot at scale. Golden (Python): -20.6074, -1.0092, 22.8396, -16.1888. 6// out_q34 / 2^24 = out in Q10 = out*1024. Golden*1024 ~= -21102, -1033, 23388, -16577. 7// license_tier: ORIGINAL 8import "nx_syscalls.nx" 9import "nx_tier.nx" 10import "nx_le.nx" 11import "nx_strconv.nx" 12import "nx_gguf_load.nx" 13 14func q8v_dot(buf: *u8, block_off: i64, n_blocks: i64, col_q10: *i64) -> i64 { 15 var dot: i64 = 0 16 var blk: i64 = 0 17 while blk < n_blocks { 18 let bo: i64 = block_off + blk * 34 19 let d_q24: i64 = _gguf_f16_to_q24(nx_le_read_u16(buf, bo)) 20 var i: i64 = 0 21 while i < 32 { 22 var q: i64 = nx_le_read_u8(buf, bo + 2 + i) 23 if q >= 128 { q = q - 256 } 24 dot = dot + (d_q24 * q) * col_q10[blk * 32 + i] 25 i = i + 1 26 } 27 blk = blk + 1 28 } 29 return dot 30} 31 32func q8v_emit(fd: i64, o: i64, v: i64) -> i64 { 33 let dec: *u8 = sys_mmap(32) 34 sys_write(fd, "row" as *u8, 3) 35 let nd0: i64 = nx_strconv_format_i64(o, dec) 36 sys_write(fd, dec, nd0) 37 sys_write(fd, "=" as *u8, 1) 38 let nd: i64 = nx_strconv_format_i64(v, dec) 39 sys_write(fd, dec, nd) 40 sys_write(fd, "\n" as *u8, 1) 41 return 0 42} 43 44func main() -> i64 { 45 let fd: i64 = sys_openat_rd("/mnt/c/Users/elder/nishi-core/nxc2/dit_out_weight.q8" as *u8) 46 if fd < 0 { return 30 } 47 let CAP: i64 = 15667200 48 let buf: *u8 = sys_mmap(CAP + 64) 49 var total: i64 = 0 50 var go: i64 = 1 51 while go == 1 { 52 let r: i64 = sys_read(fd, ((buf as i64) + total) as *u8, CAP - total) 53 if r <= 0 { go = 0 } else { total = total + r; if total >= CAP { go = 0 } } 54 } 55 sys_close(fd) 56 if total != CAP { return 31 } 57 58 let IN: i64 = 3840 59 let col: *i64 = sys_mmap(IN * 8) as *i64 60 var i: i64 = 0 61 while i < IN { col[i] = ((i - (i / 7) * 7) + 1) * 1024; i = i + 1 } 62 63 let gold: *i64 = sys_mmap(4 * 8) as *i64 64 gold[0] = 0 - 21102 65 gold[1] = 0 - 1033 66 gold[2] = 23388 67 gold[3] = 0 - 16577 68 69 let div: i64 = 1 << 24 70 let ofd: i64 = sys_openat_wr("/tmp/dit_linear.txt" as *u8, 0x1a4) 71 var o: i64 = 0 72 while o < 4 { 73 let dq34: i64 = q8v_dot(buf, o * 4080, 120, col) 74 let q10: i64 = dq34 / div 75 if ofd >= 0 { q8v_emit(ofd, o, q10) } 76 var diff: i64 = q10 - gold[o] 77 if diff < 0 { diff = 0 - diff } 78 if diff > 15 { if ofd >= 0 { sys_close(ofd) } return 40 + o } 79 o = o + 1 80 } 81 if ofd >= 0 { sys_close(ofd) } 82 return 0 83}