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

test: reconcile test_backend_metal with post-rebase dev (qgs=0 at fp8/e8 call sites, rename local fp8_nblk to avoid quant.h clash)

Doug Sharp184b3a56320a
diff --git a/c/tests/test_backend_metal.mm b/c/tests/test_backend_metal.mm
index 372e2bac..6b66cd71 100644
--- a/c/tests/test_backend_metal.mm
+++ b/c/tests/test_backend_metal.mm
@@ -269,7 +269,7 @@ static int run_fp8_moe_gate(const char *name) {
const float *gs[1] = {(const float*)bad}, *us[1] = {(const float*)bad}, *ds[1] = {(const float*)bad};
float xg[8]={0}, out[8]={0}, rw[1]={1.0f};
int xoff[1]={0}, nr[1]={1}, rows[1]={0};
- int rc = coli_metal_moe_block(1, 8, 8, FP8, g, u, d, gs, us, ds, xg, xoff, nr, rows, rw, out, 1);
+ int rc = coli_metal_moe_block(1, 8, 8, FP8, 0, g, u, d, gs, us, ds, xg, xoff, nr, rows, rw, out, 1);
int ok = (rc == 0);
printf(" %-42s rc=%d (expect 0/CPU-fallback) %s\n", name, rc, ok?"ok":"*** MISMATCH (should have refused)");
return ok?0:1;
@@ -372,7 +372,7 @@ static int run_moe_e8(const std::vector<int>& nrv, const char* name) {
std::vector<float> xg_gpu(xg);
for(int gr=0;gr<R;gr++) e8_rot_rows(&xg_gpu[(size_t)gr*D],1,D);
std::vector<float> gout((size_t)S*D,0.f);
- int ok = coli_metal_moe_block(nb,D,I,fmt,g.data(),u.data(),d.data(),gs.data(),us.data(),ds.data(),
+ int ok = coli_metal_moe_block(nb,D,I,fmt,0,g.data(),u.data(),d.data(),gs.data(),us.data(),ds.data(),
xg_gpu.data(),xoff.data(),nr.data(),rows.data(),rw.data(),gout.data(),S);
double maxabs=0,ymax=0; for(size_t i=0;i<gout.size();i++){ maxabs=fmax(maxabs,fabs(gout[i]-refout[i])); ymax=fmax(ymax,fabs(refout[i])); }
double nerr=maxabs/(ymax+1e-9); int pass = ok && nerr<1e-4;