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

hip: map the peer-copy/event-wait surface the resident path uses

ZacharyZcRed66f5ccd4eb
diff --git a/c/backend_gpu_compat.h b/c/backend_gpu_compat.h
index 98ec98e0..990462af 100644
--- a/c/backend_gpu_compat.h
+++ b/c/backend_gpu_compat.h
@@ -41,9 +41,12 @@
#define cudaStreamNonBlocking hipStreamNonBlocking
#define cudaStreamDestroy hipStreamDestroy
#define cudaStreamSynchronize hipStreamSynchronize
+#define cudaStreamWaitEvent hipStreamWaitEvent
#define cudaDeviceSynchronize hipDeviceSynchronize
#define cudaEvent_t hipEvent_t
#define cudaEventCreate hipEventCreate
+#define cudaEventCreateWithFlags hipEventCreateWithFlags
+#define cudaEventDisableTiming hipEventDisableTiming
#define cudaEventDestroy hipEventDestroy
#define cudaEventRecord hipEventRecord
#define cudaEventSynchronize hipEventSynchronize
@@ -52,6 +55,7 @@
#define cudaFreeHost hipHostFree
#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
#define cudaMemcpyPeer hipMemcpyPeer
+#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
#define cudaMemsetAsync hipMemsetAsync
#else
#include <cuda_runtime.h>