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

fix(hip): map __trap to abort in the GPU compat header

monotophica541e77a6260
diff --git a/c/backend_gpu_compat.h b/c/backend_gpu_compat.h
index 7f22863f..f30a60ed 100644
--- a/c/backend_gpu_compat.h
+++ b/c/backend_gpu_compat.h
@@ -89,6 +89,14 @@ namespace nvcuda { namespace wmma = ::rocwmma; }
#define cudaMemcpyPeer hipMemcpyPeer
#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
#define cudaMemsetAsync hipMemsetAsync
+/* Device-side kernel abort (weight_at's undecodable-format backstop,
+ * backend_cuda.cu). CUDA's __trap() intrinsic has no HIP name-twin; HIP
+ * device code aborts via abort() -- the same mapping hipify applies. NVCC
+ * never sees this define, so the CUDA-side semantics the fmt trap test pins
+ * (kernel aborts, context poisoned) are untouched. Under HIP this makes the
+ * refusal COMPILE; its runtime behavior on AMD silicon is not verified by
+ * this repo's tests (same hardware caveat as the rest of this HIP arm). */
+#define __trap abort
#else
#include <cuda_runtime.h>
#include <mma.h>