danielkrause/colibri-strix
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

sampling defaults for int4 reality: temp 0.7, nucleus 0.90 (official 1.0/0.95 tuned for full precision; the int4 tail is quantization noise)

Browse Source 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
JustVugg
2026-07-06 19:30:34 +02:00
parent 1a8d2bcff3
commit 2dd6800aea
1 changed files with 5 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
c/glm.c
+5 -3
View File
@@ -1061,7 +1061,8 @@ static void run_text(Model *m, const char *snap, const char *prompt, int ngen){
Tok T; tok_load(&T,tkp); Tok T; tok_load(&T,tkp);
int eos=tok_id_of(&T,"<|endoftext|>"); int eos=tok_id_of(&T,"<|endoftext|>");
stops_arm(&m->c, eos); stops_arm(&m->c, eos);
if(g_temp<0) g_temp=1.0f; /* auto: sampling ufficiale (temp 1.0, top-p 0.95) */ if(g_temp<0) g_temp=0.7f; /* auto: 0.7, NON l'1.0 ufficiale — la coda della
* distribuzione int4 e' rumore di quantizzazione */
int cap=(int)strlen(prompt)+16; int *pids=malloc(cap*sizeof(int)); int cap=(int)strlen(prompt)+16; int *pids=malloc(cap*sizeof(int));
int np=tok_encode(&T,prompt,(int)strlen(prompt),pids,cap); int np=tok_encode(&T,prompt,(int)strlen(prompt),pids,cap);
if(np<1){ fprintf(stderr,"prompt vuoto dopo tokenizzazione\n"); return; } if(np<1){ fprintf(stderr,"prompt vuoto dopo tokenizzazione\n"); return; }
@@ -1101,7 +1102,8 @@ static void run_serve(Model *m, const char *snap){
Tok T; tok_load(&T,tkp); Tok T; tok_load(&T,tkp);
int eos=tok_id_of(&T,"<|endoftext|>"); int eos=tok_id_of(&T,"<|endoftext|>");
stops_arm(&m->c, eos); stops_arm(&m->c, eos);
if(g_temp<0) g_temp=1.0f; /* auto: sampling ufficiale (temp 1.0, top-p 0.95) */ if(g_temp<0) g_temp=0.7f; /* auto: 0.7, NON l'1.0 ufficiale — la coda della
* distribuzione int4 e' rumore di quantizzazione */
int ngen=getenv("NGEN")?atoi(getenv("NGEN")):256; int ngen=getenv("NGEN")?atoi(getenv("NGEN")):256;
int maxctx=getenv("CTX")?atoi(getenv("CTX")):4096; int maxctx=getenv("CTX")?atoi(getenv("CTX")):4096;
int templ=getenv("CHAT_TEMPLATE")?atoi(getenv("CHAT_TEMPLATE")):1; int templ=getenv("CHAT_TEMPLATE")?atoi(getenv("CHAT_TEMPLATE")):1;
@@ -1328,7 +1330,7 @@ int main(int argc, char **argv){
g_draft = getenv("DRAFT")?atoi(getenv("DRAFT")):-1; /* -1 = auto: 3 se MTP, 0 senza */ g_draft = getenv("DRAFT")?atoi(getenv("DRAFT")):-1; /* -1 = auto: 3 se MTP, 0 senza */
g_direct = getenv("DIRECT")?atoi(getenv("DIRECT")):0; g_direct = getenv("DIRECT")?atoi(getenv("DIRECT")):0;
g_temp = getenv("TEMP")?atof(getenv("TEMP")):-1; /* -1 = auto (1.0 chat/testo, greedy altrove) */ g_temp = getenv("TEMP")?atof(getenv("TEMP")):-1; /* -1 = auto (1.0 chat/testo, greedy altrove) */
g_nuc = getenv("NUCLEUS")?atof(getenv("NUCLEUS")):0.95f; g_nuc = getenv("NUCLEUS")?atof(getenv("NUCLEUS")):0.90f; /* piu' stretto dell'ufficiale 0.95: la coda int4 e' rumore */
if(getenv("SEED")) g_rng = (uint64_t)atoll(getenv("SEED"))*0x9E3779B97F4A7C15ULL+1; if(getenv("SEED")) g_rng = (uint64_t)atoll(getenv("SEED"))*0x9E3779B97F4A7C15ULL+1;
else { struct timespec ts; clock_gettime(CLOCK_MONOTONIC,&ts); g_rng ^= (uint64_t)ts.tv_nsec<<20 ^ (uint64_t)getpid(); } else { struct timespec ts; clock_gettime(CLOCK_MONOTONIC,&ts); g_rng ^= (uint64_t)ts.tv_nsec<<20 ^ (uint64_t)getpid(); }
if(g_draft>63) g_draft=63; /* -1 = auto, risolto dopo model_init */ if(g_draft>63) g_draft=63; /* -1 = auto, risolto dopo model_init */