nx_nofloat_qwen_fastgen_gate.nx source
↩ module page · 159 lines · 10910 B
1// nx_nofloat_qwen_fastgen_gate.nx -- KV-cached generation with a DEQUANT-ONCE weight+head cache.
2// The dqprobe gate MEASURED that ~88% of a decode step is re-dequantizing frozen weights (1213/1363 ms), and
3// the streamed LM head re-dequantizes all 152K rows every token (~650 ms). Both are removed here by dequanting
4// ONCE up front (nf_dequant_all_layers + nf_dequant_head_all) and reusing the cache every step. This is LOSSLESS:
5// identical dequant, identical accumulation order -> the tokens MUST still be [12095,13,1084,374] (=' Paris' '.'
6// ' It' ' is'), pinned as teeth. Reports one-time cache cost + per-token decode/head split. No hw writes (Rule 26).
7// expect_exit: 0 license_tier: ORIGINAL
8import "nx_syscalls.nx"
9import "nx_tier.nx"
10import "nx_le.nx"
11import "nx_tensor.nx"
12import "nx_gguf.nx"
13import "nx_gguf_load.nx"
14import "nx_gguf_meta.nx"
15import "nx_nofloat_llm.nx"
16import "nx_nofloat_tok.nx"
17import "nx_gate_verdict.nx"
18
19func 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 }
20func 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 }
21func fg_slen(s: *u8) -> i64 { var n: i64=0; while s[n]!=(0 as u8){n=n+1} return n }
22
23func main() -> i64 {
24 fg_puts("FAST NO-FLOAT GENERATION: dequant-once weight+head cache (LOSSLESS), flat single-token decode\n\n" as *u8)
25 let path: *u8 = "/home/elderwesto/nx_stage/nx_real_model.gguf\x00" as *u8
26 let len_out: *i64 = sys_mmap(8) as *i64; len_out[0]=0
27 let buf: *u8 = sys_read_file(path, len_out)
28 if buf == (0 as *u8) { fg_puts("MODEL ABSENT\n" as *u8); return 1 }
29 let hdr: *NxGgufHeader = sys_mmap(NX_GGUF_HDR_BYTES) as *NxGgufHeader
30 if nx_gguf_parse(buf, len_out[0], hdr) != NX_GGUF_OK { fg_puts("PARSE FAIL\n" as *u8); return 1 }
31
32 let ne: i64=896; let qd: i64=896; let kvd: i64=128; let fd: i64=4864; let MAXT: i64=16
33 let voff: *i64=sys_mmap(8) as *i64; let vty: *i64=sys_mmap(8) as *i64
34 var mfirst: i64=0; var nm_c: i64=0; var vfirst: i64=0; var vocab: i64=0
35 let km: *u8="tokenizer.ggml.merges\x00" as *u8; let kt: *u8="tokenizer.ggml.tokens\x00" as *u8
36 if nx_gguf_meta_find(buf, len_out[0], hdr, km, fg_slen(km), voff, vty)==NX_GMETA_OK { nm_c=nx_gguf_meta_array_count(buf, voff[0]); mfirst=nx_gguf_meta_array_first_elt_off(buf, voff[0]) }
37 if nx_gguf_meta_find(buf, len_out[0], hdr, kt, fg_slen(kt), voff, vty)==NX_GMETA_OK { vocab=nx_gguf_meta_array_count(buf, voff[0]); vfirst=nx_gguf_meta_array_first_elt_off(buf, voff[0]) }
38 let nt: *u8="token_embd.weight\x00" as *u8; let no: *u8="output.weight\x00" as *u8; let nn: *u8="output_norm.weight\x00" as *u8
39 let ie: nx_int=nx_gguf_find_tensor(hdr, nt, 17); let io: nx_int=nx_gguf_find_tensor(hdr, no, 13); let inn: nx_int=nx_gguf_find_tensor(hdr, nn, 18)
40 if ie<0 { return 1 } if io<0 { return 1 } if inn<0 { return 1 }
41 let te: *NxGgufTensorInfo=nx_gguf_tensor_at(hdr, ie); let te_base: i64=hdr.data_off+te.offset; let te_ty: i64=te.ggml_type
42 let oh: *NxGgufTensorInfo=nx_gguf_tensor_at(hdr, io); let oh_base: i64=hdr.data_off+oh.offset; let oh_ty: i64=oh.ggml_type
43 let gout: *i64=sys_mmap(ne*8) as *i64; load_named_q16(buf, hdr, nn, 18, gout, ne)
44
45 let wb: *i64=sys_mmap(12*8) as *i64
46 wb[0]=sys_mmap(ne*8) as i64; wb[1]=sys_mmap(qd*ne*8) as i64; wb[2]=sys_mmap(kvd*ne*8) as i64; wb[3]=sys_mmap(kvd*ne*8) as i64; wb[4]=sys_mmap(ne*qd*8) as i64
47 wb[5]=sys_mmap(ne*8) as i64; wb[6]=sys_mmap(ne*fd*8) as i64; wb[7]=sys_mmap(ne*fd*8) as i64; wb[8]=sys_mmap(fd*ne*8) as i64
48 wb[9]=sys_mmap(qd*8) as i64; wb[10]=sys_mmap(kvd*8) as i64; wb[11]=sys_mmap(kvd*8) as i64
49 let sb: *i64=sys_mmap(14*8) as *i64
50 sb[0]=sys_mmap(MAXT*ne*8) as i64; sb[1]=sys_mmap(MAXT*qd*8) as i64; sb[2]=sys_mmap(MAXT*kvd*8) as i64; sb[3]=sys_mmap(MAXT*kvd*8) as i64; sb[4]=sys_mmap(MAXT*qd*8) as i64
51 sb[5]=sys_mmap(MAXT*8) as i64; sb[6]=sys_mmap(MAXT*8) as i64; sb[7]=sys_mmap(MAXT*ne*8) as i64; sb[8]=sys_mmap(MAXT*fd*8) as i64; sb[9]=sys_mmap(MAXT*fd*8) as i64
52 sb[10]=sys_mmap(MAXT*fd*8) as i64; sb[11]=sys_mmap(MAXT*ne*8) as i64; sb[12]=sys_mmap(MAXT*ne*8) as i64; sb[13]=sys_mmap(MAXT*ne*8) as i64
53 let kvc: *i64=sys_mmap(48*8) as *i64
54 var kl: i64=0
55 while kl<24 { kvc[2*kl]=sys_mmap(MAXT*kvd*8) as i64; kvc[2*kl+1]=sys_mmap(MAXT*kvd*8) as i64; kl=kl+1 }
56 let nmbuf: *u8=sys_mmap(64); let freqs: *i64=sys_mmap(32*8) as *i64; rope_freqs(freqs, 64)
57 let tmp: *i64=sys_mmap(64*256*8) as *i64
58 let x: *i64=sys_mmap(MAXT*ne*8) as *i64
59 let hout: *i64=sys_mmap(MAXT*ne*8) as *i64
60 let x1: *i64=sys_mmap(ne*8) as *i64
61 let h1: *i64=sys_mmap(ne*8) as *i64
62 let cfgA: *i64=sys_mmap(8*8) as *i64; cfgA[1]=ne; cfgA[2]=14; cfgA[3]=2; cfgA[4]=64; cfgA[5]=qd; cfgA[6]=kvd; cfgA[7]=8192
63 let cfgF: *i64=sys_mmap(4*8) as *i64; cfgF[1]=ne; cfgF[2]=fd
64
65 let input: *u8="The capital of France is\x00" as *u8
66 let ids: *i64=sys_mmap(MAXT*8) as *i64; let tokp: *i64=sys_mmap(MAXT*8) as *i64; let tokl: *i64=sys_mmap(MAXT*8) as *i64
67 let nprompt: i64=tk_bpe_encode(buf, mfirst, nm_c, vfirst, vocab, input, fg_slen(input), tokp, tokl, ids)
68 fg_puts("prompt tokens: "); fg_num(nprompt); fg_puts("\n")
69
70 let normed: *i64=sys_mmap(ne*8) as *i64
71 let lgout: *i64=sys_mmap(8) as *i64
72 let idout: *i64=sys_mmap(8) as *i64
73
74 // ---- ONE-TIME dequant caches: all 24 layers + the full output head ----
75 let c0: i64=sys_now_ms()
76 let wcache: *i64=sys_mmap(24*8) as *i64
77 nf_dequant_all_layers(buf, hdr, wcache, 24, ne, qd, kvd, fd)
78 let c1: i64=sys_now_ms()
79 let hcache: *i64=nf_dequant_head_all(buf, oh_base, oh_ty, vocab, ne)
80 let c2: i64=sys_now_ms()
81 if (hcache as i64)==0 { fg_puts("HEAD CACHE OOM\n" as *u8); return 1 }
82 fg_puts("one-time cache: layers="); fg_num(c1-c0); fg_puts(" ms, head="); fg_num(c2-c1); fg_puts(" ms (amortized over the whole sequence)\n\n")
83
84 let hcp: *i64=sys_mmap(6*8) as *i64
85 hcp[0]=hcache as i64; hcp[1]=normed as i64; hcp[2]=vocab; hcp[3]=ne; hcp[4]=idout as i64; hcp[5]=lgout as i64
86
87 // ---- PREFILL: full prompt once, capturing all layers' K/V (per-step dequant here is one-time, fine) ----
88 let t0: i64=sys_now_ms()
89 var ei: i64=0; while ei<nprompt { dequant_row(buf, te_base, te_ty, ids[ei], ne, ((x as i64)+ei*ne*8) as *i64, tmp); ei=ei+1 }
90 cfgA[0]=nprompt; cfgF[0]=nprompt
91 run_stack_prefill_kv(buf, hdr, x, hout, wb, sb, nmbuf, freqs, kvc, cfgA, cfgF, 24)
92 rmsnorm_gamma_row_q24(hout, gout, (nprompt-1)*ne, ne, normed, 0)
93 var next: i64=head_argmax_cached(hcp)
94 let t1: i64=sys_now_ms()
95 fg_puts(" prefill ("); fg_num(nprompt); fg_puts(" tok): next id="); fg_num(next); fg_puts(" logit="); fg_num(lgout[0]); fg_puts(" ["); fg_num(t1-t0); fg_puts(" ms]\n")
96 let prefill_first: i64=next
97 var T: i64=nprompt
98 ids[T]=next; T=T+1
99 var ngen: i64=1
100
101 // ---- DECODE: cached single-token steps, timing decode vs head split ----
102 let MAXNEW: i64=6
103 var stop: i64=0
104 if next==151643 { stop=1 }
105 if next==151645 { stop=1 }
106 var dec_total: i64=0; var head_total: i64=0; var ndec: i64=0
107 while stop==0 {
108 if ngen>=MAXNEW { stop=1 } else { if T>=MAXT { stop=1 } else {
109 let pos: i64=T-1
110 dequant_row(buf, te_base, te_ty, ids[pos], ne, x1, tmp)
111 let d0: i64=sys_now_ms()
112 decode_step_kv_cached(buf, hdr, x1, h1, wcache, sb, nmbuf, freqs, kvc, pos, cfgA, cfgF, 24)
113 let d1: i64=sys_now_ms()
114 rmsnorm_gamma_row_q24(h1, gout, 0, ne, normed, 0)
115 next=head_argmax_cached(hcp)
116 let d2: i64=sys_now_ms()
117 dec_total=dec_total+(d1-d0); head_total=head_total+(d2-d1); ndec=ndec+1
118 fg_puts(" decode pos="); fg_num(pos); fg_puts(": id="); fg_num(next); fg_puts(" piece='")
119 let off: i64=tk_decode_off(buf, vfirst, next); let pl: i64=nx_gguf_meta_read_string_len(buf, off)
120 if pl>0 { sys_write(1, nx_gguf_meta_read_string_ptr(buf, off), pl) }
121 fg_puts("' [decode="); fg_num(d1-d0); fg_puts(" head="); fg_num(d2-d1); fg_puts(" ms]\n")
122 ids[T]=next; T=T+1; ngen=ngen+1
123 if next==151643 { stop=1 }
124 if next==151645 { stop=1 }
125 } }
126 }
127
128 fg_puts("\nCOMPLETION: '" as *u8)
129 var gi: i64=nprompt
130 var outbytes: i64=0
131 while gi<T { let o2: i64=tk_decode_off(buf, vfirst, ids[gi]); let p2: i64=nx_gguf_meta_read_string_len(buf, o2); if p2>0 { sys_write(1, nx_gguf_meta_read_string_ptr(buf, o2), p2); outbytes=outbytes+p2 } gi=gi+1 }
132 fg_puts("'\n" as *u8)
133 fg_puts("timing: mean decode="); if ndec>0 { fg_num(dec_total/ndec) } else { fg_num(0) }
134 fg_puts(" ms + head="); if ndec>0 { fg_num(head_total/ndec) } else { fg_num(0) }
135 fg_puts(" ms = "); if ndec>0 { fg_num((dec_total+head_total)/ndec) } else { fg_num(0) } fg_puts(" ms/token (vs 2014 ms/token uncached-dequant)\n\n" as *u8)
136
137 var pass: i64=0; var ttl: i64=0
138 ttl=ttl+1; fg_puts(" T1 cached prefill argmax == 12095 (ĠParis): "); if prefill_first==12095 { pass=pass+1; fg_puts("PASS\n") } else { fg_puts("FAIL\n") }
139 ttl=ttl+1; fg_puts(" T2 LOSSLESS: first 4 generated == [12095,13,1084,374] (bit-match uncached-dequant path): ")
140 var ok4: i64=1
141 if ids[nprompt]!=12095 { ok4=0 }
142 if ids[nprompt+1]!=13 { ok4=0 }
143 if ids[nprompt+2]!=1084 { ok4=0 }
144 if ids[nprompt+3]!=374 { ok4=0 }
145 if ok4==1 { pass=pass+1; fg_puts("PASS\n") } else { fg_puts("FAIL\n") }
146 ttl=ttl+1; fg_puts(" T3 generated "); fg_num(ngen); fg_puts(" tokens, all valid ids: "); var okids: i64=1; var ci: i64=nprompt; while ci<T { if ids[ci]<0 { okids=0 } if ids[ci]>=vocab { okids=0 } ci=ci+1 } if ngen>=5 { if okids==1 { pass=pass+1; fg_puts("PASS\n") } else { fg_puts("FAIL\n") } } else { fg_puts("FAIL\n") }
147 ttl=ttl+1; fg_puts(" T4 mean decode+head < 1000 ms/token (interactive floor cleared): "); var mt: i64=0; if ndec>0 { mt=(dec_total+head_total)/ndec } if mt<1000 { pass=pass+1; fg_puts("PASS ("); fg_num(mt); fg_puts(" ms)\n") } else { fg_puts("FAIL ("); fg_num(mt); fg_puts(" ms)\n") }
148
149 fg_puts("NX-NOFLOAT-QWEN-FASTGEN-GATE passed "); fg_num(pass); fg_puts("/"); fg_num(ttl)
150 // MIGRATED onto nx_gate_verdict by nx_gate_dry_apply (D001, minimal form): every check
151 // row above is untouched, so the PASS/FAIL vector cannot change; only the hand-rolled
152 // verdict emission is replaced by the ONE shared base class. Proven by nx_gate_migrate verify.
153 let ctr__dry: *i64 = gv_ctr()
154 ctr__dry[0] = pass
155 ctr__dry[1] = ttl
156 let rc__dry: i64 = gv_verdict("NOFLOAT-QWEN-FASTGEN-GATE" as *u8, ctr__dry, "dequant-once cache: LOSSLESS + sub-second/token)" as *u8)
157 sys_exit(rc__dry)
158 return rc__dry
159}