Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

score: validate complete input rows before scoring

Rudy Celeklic8a48a7fd8cb
diff --git a/c/colibri.c b/c/colibri.c
index 0c19116b..7a684380 100644
--- a/c/colibri.c
+++ b/c/colibri.c
@@ -8262,6 +8262,35 @@ static int forward_all(Model *m, const int *ids, int S, int *pred, const int *re
* input: file con righe "<ctxlen> <contlen> <id0> .. <id_{T-1}>" (T=ctxlen+contlen)
* output: riga "<logprob_continuazione> <contlen> <greedy 0/1>" per richiesta.
* Un solo forward per richiesta (teacher-forcing): niente generazione -> fattibile a bassa velocita'. */
+/* SCORE rows are a whitespace-separated pair of lengths and exactly that
+ * many vocabulary ids. Check the entire file before allocating inference
+ * buffers or emitting any score; a missing id must not become token zero. */
+static int score_int(char **cursor, int *value){
+ char *p=*cursor, *end;
+ while(isspace((unsigned char)*p)) p++;
+ if(!*p) return 0;
+ errno=0;
+ long n=strtol(p,&end,10);
+ if(p==end || errno==ERANGE || n<0 || n>INT_MAX ||
+ (*end && !isspace((unsigned char)*end))) return 0;
+ *value=(int)n; *cursor=end; return 1;
+}
+static int score_row(char *line, size_t length, int vocab, int reserve,
+ int *ctx, int *cont, int *ids, int capacity){
+ if(memchr(line,0,length)) return 0;
+ char *p=line;
+ if(!score_int(&p,ctx) || !score_int(&p,cont) || *ctx<1 ||
+ *ctx>INT_MAX-reserve || *cont>INT_MAX-reserve-*ctx) return 0;
+ int total=*ctx+*cont;
+ if(ids && total>capacity-reserve) return 0;
+ for(int i=0;i<total;i++){
+ int id;
+ if(!score_int(&p,&id) || id>=vocab) return 0;
+ if(ids) ids[i]=id;
+ }
+ while(isspace((unsigned char)*p)) p++;
+ return !*p;
+}
static void run_score(Model *m, const char *snap, const char *path){
Cfg *c=&m->c; int D=c->hidden;
/* prefisso GLM (#108): il modello vede [gMASK]<sop> in testa a OGNI sequenza di training —
@@ -8285,18 +8314,29 @@ static void run_score(Model *m, const char *snap, const char *path){
free(ar);
}
FILE *f=fopen(path,"rb"); if(!f){perror(path);exit(1);}
- int maxT=1; { char *ln=NULL; size_t cp=0;
- while(getline(&ln,&cp,f)>0){ int a,b; if(sscanf(ln,"%d %d",&a,&b)==2 && a+b>maxT) maxT=a+b; }
+ int maxT=1, reserve=pfx_on?2:0; { char *ln=NULL; size_t cp=0; ssize_t length; size_t lineno=0;
+ while((length=getline(&ln,&cp,f))>0){
+ int ctx,cont; lineno++;
+ if(!score_row(ln,(size_t)length,c->vocab,reserve,&ctx,&cont,NULL,0)){
+ fprintf(stderr,"[SCORE] invalid request at line %zu\n",lineno); exit(1);
+ }
+ if(ctx+cont>maxT) maxT=ctx+cont;
+ }
+ if(ferror(f)){ perror(path); exit(1); }
free(ln); }
if(pfx_on) maxT+=2; /* le richieste senza prefisso crescono di 2 token */
kv_alloc(m,maxT);
float *x=falloc((int64_t)maxT*D), *lo=falloc(c->vocab), *row=falloc(D);
- int *ids=malloc(maxT*sizeof(int));
- rewind(f); char *ln=NULL; size_t cp=0; int nreq=0; double t0=now_s();
- while(getline(&ln,&cp,f)>0){
- char *p=ln; int ctxlen=strtol(p,&p,10), contlen=strtol(p,&p,10), T=ctxlen+contlen;
- if(T<=0||ctxlen<1){ printf("0 0 0\n"); fflush(stdout); continue; }
- for(int i=0;i<T;i++) ids[i]=strtol(p,&p,10);
+ int *ids=malloc((size_t)maxT*sizeof(int));
+ if(!ids){ perror("SCORE token buffer"); exit(1); }
+ rewind(f); char *ln=NULL; size_t cp=0; ssize_t length; size_t lineno=0;
+ int nreq=0; double t0=now_s();
+ while((length=getline(&ln,&cp,f))>0){
+ int ctxlen,contlen; lineno++;
+ if(!score_row(ln,(size_t)length,c->vocab,reserve,&ctxlen,&contlen,ids,maxT)){
+ fprintf(stderr,"[SCORE] invalid request at line %zu\n",lineno); exit(1);
+ }
+ int T=ctxlen+contlen;
if(pfx_on && !(T>=2 && ids[0]==pfx[0] && ids[1]==pfx[1])){ /* gia' prefissato -> intatto */
memmove(ids+2,ids,(size_t)T*sizeof(int));
ids[0]=pfx[0]; ids[1]=pfx[1]; ctxlen+=2; T+=2;
@@ -8313,6 +8353,7 @@ static void run_score(Model *m, const char *snap, const char *path){
if(++nreq%5==0) fprintf(stderr,"[score %d req | %.1fs | RSS %.2f GB | hit %.0f%%]\n",
nreq, now_s()-t0, rss_gb(), (m->hits+m->miss)?100.0*m->hits/(m->hits+m->miss):0.0);
}
+ if(ferror(f)){ perror(path); exit(1); }
free(ln); free(ids); free(x); free(lo); free(row); fclose(f);
}
diff --git a/c/tests/test_score_rows.py b/c/tests/test_score_rows.py
new file mode 100644
index 00000000..7d71a07f
--- /dev/null
+++ b/c/tests/test_score_rows.py
@@ -0,0 +1,61 @@
+"""Real-process SCORE rows must not silently invent or ignore token ids."""
+import os
+from pathlib import Path
+import subprocess
+import tempfile
+import unittest
+
+C_DIR = Path(__file__).resolve().parents[1]
+ENGINE = C_DIR / ("colibri.exe" if os.name == "nt" else "colibri")
+SNAP = C_DIR / "glm_tiny"
+
+
+class ScoreRowsTest(unittest.TestCase):
+ @classmethod
+ def setUpClass(cls):
+ if not ENGINE.is_file() or not (SNAP / "model.safetensors").is_file():
+ raise unittest.SkipTest("build colibri and generate the GLM tiny fixture first")
+
+ def run_score(self, rows):
+ keep = {"PATH", "HOME", "SYSTEMROOT", "WINDIR", "TEMP", "TMP", "USERPROFILE",
+ "LD_LIBRARY_PATH", "DYLD_LIBRARY_PATH"}
+ env = {k: v for k, v in os.environ.items() if k in keep}
+ with tempfile.TemporaryDirectory() as tmp:
+ path = Path(tmp) / "rows.txt"
+ path.write_bytes(rows.encode() if isinstance(rows, str) else rows)
+ env.update(SNAP=str(SNAP), SCORE=str(path), SCORE_PREFIX="0", OMP_NUM_THREADS="4")
+ return subprocess.run([str(ENGINE), "64", "16", "16"], cwd=C_DIR,
+ env=env, capture_output=True, text=True, timeout=30)
+
+ def test_valid_rows_keep_scores_and_empty_continuations(self):
+ result = self.run_score("1 1 3 14\n1 0 3\n")
+ self.assertEqual(result.returncode, 0, result.stderr)
+ rows = result.stdout.splitlines()[-2:]
+ self.assertEqual(rows[0].split()[1:], ["1", "0"])
+ self.assertAlmostEqual(float(rows[0].split()[0]), -5.152055, places=4)
+ self.assertEqual(rows[1], "0.000000 0 1")
+
+ def test_missing_negative_out_of_vocabulary_and_extra_ids(self):
+ for row in ("1 1 3", "1 1 3 -1", "1 1 3 10000000", "1 1 3 14 7"):
+ with self.subTest(row=row):
+ result = self.run_score(row + "\n")
+ self.assertEqual(result.returncode, 1, result.stderr)
+ self.assertIn("[SCORE] invalid request at line 1", result.stderr)
+
+ def test_invalid_lengths_and_noninteger_fields(self):
+ for row in ("0 1 3", "1 -1", "1.0 1 3 14", "1 1 3 14junk",
+ "2147483647 1", "999999999999999999999 0", "1 1 3 14\0ignored"):
+ with self.subTest(row=row):
+ result = self.run_score(row + "\n")
+ self.assertEqual(result.returncode, 1, result.stderr)
+ self.assertIn("[SCORE] invalid request at line 1", result.stderr)
+
+ def test_whole_file_validation_precedes_any_score(self):
+ result = self.run_score("1 1 3 14\n1 1 3\n")
+ self.assertEqual(result.returncode, 1, result.stderr)
+ self.assertIn("[SCORE] invalid request at line 2", result.stderr)
+ self.assertNotIn("-5.152055 1 0", result.stdout)
+
+
+if __name__ == "__main__":
+ unittest.main()