Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
Merge pull request #1677 from ZacharyZcR/perf/qwen36-dnproj-prefill
Vincenzo Fornaro94cef8ce66af
diff --cc c/qwen36_tier.cindex 0d00ed1e,98cd4dca..742769d9--- a/c/qwen36_tier.c+++ b/c/qwen36_tier.c@@@ -839,14 -808,14 +839,17 @@@ int qt_dense_matmul_batch(int h, floatG_dense[h].on = 0;return 0;}+int qt_dense_matmul(int h, float *y, const float *x, int I, int O){+ return qt_dense_matmul_batch(h, y, x, 1, I, O);+}int qt_dense_count(void){ return G_dense_n; }- int qt_dnproj_matmul(int layer, float *y, const float *x, int I, int O){- if(layer < 0 || layer >= QT_DN_MAX_LAYERS || !G_dnp[layer].on) return 0;- if(coli_cuda_matmul(&G_dnp[layer].t,y,x,NULL,NULL,1,1,I,O,G_dnp[layer].dev,0))+ int qt_dnproj_ready(int layer){+ return layer >= 0 && layer < QT_DN_MAX_LAYERS && G_dnp[layer].on;+ }+ int qt_dnproj_matmul_batch(int layer, float *y, const float *x, int S, int I, int O){+ if(!qt_dnproj_ready(layer) || S <= 0) return 0;+ if(coli_cuda_matmul(&G_dnp[layer].t,y,x,NULL,NULL,1,S,I,O,G_dnp[layer].dev,0))return 1;fprintf(stderr,"[dnp] layer %d GPU matmul failed; CPU from here on\n", layer);G_dnp[layer].on = 0;diff --cc c/qwen36_tier.hindex 25a67b34,1098a0c3..41a73bf7--- a/c/qwen36_tier.h+++ b/c/qwen36_tier.h@@@ -151,9 -147,10 +153,11 @@@ static inline int qt_place_is_auto(voistatic inline void qt_trunk_withdraw(const char*a){(void)a;}static inline int qt_dnproj_init(int a,const int8_t*b,const float*c,int d,int e,int f){(void)a;(void)b;(void)c;(void)d;(void)e;(void)f;return 0;}static inline int qt_dnproj_matmul(int a,float*b,const float*c,int d,int e){(void)a;(void)b;(void)c;(void)d;(void)e;return 0;}+ static inline int qt_dnproj_ready(int a){(void)a;return 0;}+ static inline int qt_dnproj_matmul_batch(int a,float*b,const float*c,int d,int e,int f){(void)a;(void)b;(void)c;(void)d;(void)e;(void)f;return 0;}static inline int qt_dense_init(const int8_t*a,const float*b,int c,int d,int e){(void)a;(void)b;(void)c;(void)d;(void)e;return -1;}static inline int qt_dense_matmul(int a,float*b,const float*c,int d,int e){(void)a;(void)b;(void)c;(void)d;(void)e;return 0;}+static inline int qt_dense_matmul_batch(int a,float*b,const float*c,int d,int e,int f){(void)a;(void)b;(void)c;(void)d;(void)e;(void)f;return 0;}static inline int qt_dense_count(void){return 0;}static inline int qt_ready(void){return 0;}static inline int qt_is_resident(int a,int b){(void)a;(void)b;return 0;}