Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
fix(metal-test): rename test-local fp8_nblk, which quant.h now also declares
JustVuggc7aaaf7e0fbd
diff --git a/c/tests/test_backend_metal.mm b/c/tests/test_backend_metal.mmindex 07b82613..2c1b4f10 100644--- a/c/tests/test_backend_metal.mm+++ b/c/tests/test_backend_metal.mm@@ -96,11 +96,17 @@ static float ref_e4m3(uint8_t b) {: (1.0f+(float)mant*(1.0f/8.0f))*powf(2.0f,(float)exp-7.0f);return sign ? -val : val;}-static int fp8_nblk(int n){ return (n+127)/128; }+/* Renamed from fp8_nblk: quant.h, included further down this same+ * translation unit, now declares `static inline int64_t fp8_nblk`.+ * Matching the return type alone still collides (inline vs non-inline),+ * and this file keeps its own reference implementations on purpose --+ * see the E4M3 note above -- so it takes the ref_ prefix the other+ * independent helpers here already use. (#838) */+static int64_t ref_fp8_nblk(int n){ return (n+127)/128; }static void cpu_ref_fp8(const uint8_t *q8, const float *bscale, const float *x,double *y, double *mag, int S, int I, int O) {- int nblkI = fp8_nblk(I);+ int nblkI = ref_fp8_nblk(I);for (int o=0;o<O;o++){const uint8_t *w = q8 + (size_t)o*I;const float *scl = bscale + (size_t)(o/128)*nblkI;@@ -172,7 +178,7 @@ static int run_grouped(int O, int I, int gs, int S, int outlier, const char *nam// the block-scale accumulation, avoiding cancellation-noise false positives on// near-zero results).static int run_fp8(int O, int I, int S, const char *name) {- int nblkO=fp8_nblk(O), nblkI=fp8_nblk(I), nblk=nblkO*nblkI;+ int nblkO=ref_fp8_nblk(O), nblkI=ref_fp8_nblk(I), nblk=nblkO*nblkI;std::vector<uint8_t> W((size_t)O*I);std::vector<float> scale((size_t)nblk), x((size_t)S*I), yg((size_t)S*O);std::vector<double> yr((size_t)S*O), mag((size_t)S*O);@@ -213,7 +219,7 @@ static int run_fp8(int O, int I, int S, const char *name) {static int run_fp8_lut(const char *name) {enum { O=256, I=1 };std::vector<uint8_t> W(O*I); for (int b=0;b<O;b++) W[b]=(uint8_t)b;- std::vector<float> scale(fp8_nblk(O)*fp8_nblk(I), 1.0f); // nblkO=2,nblkI=1 -> both blocks scale=1+ std::vector<float> scale(ref_fp8_nblk(O)*ref_fp8_nblk(I), 1.0f); // nblkO=2,nblkI=1 -> both blocks scale=1std::vector<float> x(I, 1.0f), yg(O);ColiMetalTensor *t=nullptr;if (!coli_metal_matmul(&t, yg.data(), x.data(), W.data(), scale.data(), FP8, 1, I, O, 0)) {