Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
kimi_k3: keep K3_CUDA the only engine switch; coli maps --gpu onto it
JustVugg2bc08d0ea452
diff --git a/c/kimi_k3.c b/c/kimi_k3.cindex 549d1eb8..a05f6336 100644--- a/c/kimi_k3.c+++ b/c/kimi_k3.c@@ -533,13 +533,6 @@ static int g_k3_mmap=0; /* prepared U8/F32 tensors stay file-bastatic uint64_t g_k3_mmap_bytes=0, g_k3_mmap_views=0;#ifdef COLI_CUDAstatic int g_k3_cuda=0; /* K3_CUDA=1: MXFP4 routed experts on CUDA at decode */-/* K3_CUDA is the explicit override. COLI_CUDA is what `coli --gpu` writes. */-static int k3_cuda_requested(void){- const char *e=getenv("K3_CUDA");- if(e && *e) return atoi(e);- e=getenv("COLI_CUDA");- return e ? atoi(e) : 0;-}#endifstatic int g_k3_direct=-1; /* K3_DIRECT: O_DIRECT expert reads */static int g_k3_idot=1; /* K3_IDOT: int8-activation expert matmuls */@@ -847,7 +840,7 @@ static void model_init_range(Model *m, const char *snap, int layer_begin,{ const char *ev=getenv("K3_VK"); vk_requested=!ev||atoi(ev); }#endif#ifdef COLI_CUDA- cuda_requested=k3_cuda_requested();+ cuda_requested=getenv("K3_CUDA")&&atoi(getenv("K3_CUDA"));#endifif(!k3_mmap_backend_allowed(1,vk_requested,cuda_requested)){fprintf(stderr,"K3_MMAP=1 is CPU-only; set K3_VK=0 and K3_CUDA=0 -- refusing before model load\n");@@ -856,7 +849,7 @@ static void model_init_range(Model *m, const char *snap, int layer_begin,fprintf(stderr,"[K3-MMAP] prepared tensor mapping enabled (CPU-only, no conversion fallback)\n");}#ifdef COLI_CUDA- g_k3_cuda = k3_cuda_requested();+ g_k3_cuda = getenv("K3_CUDA") ? atoi(getenv("K3_CUDA")) : 0;if(g_k3_cuda){int dev0 = 0;if(!coli_cuda_init(&dev0, 1)){