Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
fix(cuda): wire the two new symbols through the Windows DLL loader
Unknown-Findout7b4a236e3de9
diff --git a/c/backend_loader.c b/c/backend_loader.cindex 833e8e40..a4d6c2b3 100644--- a/c/backend_loader.c+++ b/c/backend_loader.c@@ -101,6 +101,8 @@ typedef int (*fn_matmul)(ColiCudaTensor **tensor, float *y, const floint fmt, int S, int I, int O, int device, int gs);typedef void (*fn_tensor_free)(ColiCudaTensor *tensor);typedef size_t (*fn_tensor_bytes)(const ColiCudaTensor *tensor);+typedef size_t (*fn_tensor_vram)(const ColiCudaTensor *tensor);+typedef size_t (*fn_alloc_footprint)(size_t bytes);typedef int (*fn_tensor_device)(const ColiCudaTensor *tensor);/* --- #111 GPU resident pipeline additions (matched to backend_cuda.h) --- */@@ -174,6 +176,8 @@ static struct {fn_matmul matmul;fn_tensor_free tensor_free;fn_tensor_bytes tensor_bytes;+ fn_tensor_vram tensor_vram;+ fn_alloc_footprint alloc_footprint;fn_tensor_device tensor_device;fn_attention_absorb_batch attention_absorb_batch;@@ -1418,6 +1422,13 @@ static int coli_cuda_load(void){RESOLVE(matmul, fn_matmul)RESOLVE(tensor_free, fn_tensor_free)RESOLVE(tensor_bytes, fn_tensor_bytes)+ /* Optional, same reasoning as e8_set_grid above: a DLL predating #687+ * leaves these NULL and the wrappers fall back to the logical byte+ * count, which is exactly the behaviour that shipped before. Using+ * RESOLVE here instead would unload the whole CUDA backend over one+ * missing symbol - a far worse failure than the over-commit it fixes. */+ RESOLVE_OPT(tensor_vram, fn_tensor_vram)+ RESOLVE_OPT(alloc_footprint, fn_alloc_footprint)RESOLVE(tensor_device, fn_tensor_device)RESOLVE(attention_absorb_batch, fn_attention_absorb_batch)@@ -1619,6 +1630,22 @@ size_t coli_cuda_tensor_bytes(const ColiCudaTensor *tensor){return g_cuda.tensor_bytes(tensor);}+/* Falls back to the LOGICAL size when the DLL predates #687. That is an+ * under-count, and it is the pre-existing behaviour rather than a new+ * failure mode: the tier over-commits exactly as much as it always did.+ * Returning 0 here would be far worse - the caller would charge nothing+ * for an expert it just placed. */+size_t coli_cuda_tensor_vram(const ColiCudaTensor *tensor){+ if(!g_cuda.available) return 0;+ if(!g_cuda.tensor_vram) return g_cuda.tensor_bytes(tensor);+ return g_cuda.tensor_vram(tensor);+}++size_t coli_cuda_alloc_footprint(size_t bytes){+ if(!g_cuda.available || !g_cuda.alloc_footprint) return bytes;+ return g_cuda.alloc_footprint(bytes);+}+int coli_cuda_tensor_device(const ColiCudaTensor *tensor){if(!g_cuda.available) return -1;return g_cuda.tensor_device(tensor);