Hip: disable VMM on hip as it seams that it dosent work in some configurations (llama/11420)

Author: IMbackK

ggml/CMakeLists.txt

@@ -155,6 +155,7 @@ option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp on
 
 option(GGML_HIP                             "ggml: use HIP"                                   OFF)
 option(GGML_HIP_GRAPHS                      "ggml: use HIP graph, experimental, slow"         OFF)
+option(GGML_HIP_NO_VMM                      "ggml: do not try to use HIP VMM"                 ON)
 option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
 option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
 option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)

ggml/src/ggml-cuda/common.cuh

@@ -131,6 +131,10 @@ typedef float dfloat; // dequantize float
 typedef float2 dfloat2;
 #endif // GGML_CUDA_F16
 
+#if (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
+#define GGML_USE_VMM
+#endif // (!defined(GGML_USE_HIP) && !defined(GGML_CUDA_NO_VMM)) || (defined(GGML_USE_HIP) && !defined(GGML_HIP_NO_VMM))
+
 #if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL
 #define FP16_AVAILABLE
 #endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -152,7 +152,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
     for (int id = 0; id < info.device_count; ++id) {
         int device_vmm = 0;
 
-#if !defined(GGML_CUDA_NO_VMM)
+#if defined(GGML_USE_VMM)
         CUdevice device;
         CU_CHECK(cuDeviceGet(&device, id));
         CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
@@ -164,7 +164,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
             alloc_prop.location.id = id;
             CU_CHECK(cuMemGetAllocationGranularity(&info.devices[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
         }
-#endif // !defined(GGML_CUDA_NO_VMM)
+#endif // defined(GGML_USE_VMM)
         info.devices[id].vmm = !!device_vmm;
 
         cudaDeviceProp prop;
@@ -300,7 +300,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
 };
 
 // pool with virtual memory
-#if !defined(GGML_CUDA_NO_VMM)
+#if defined(GGML_USE_VMM)
 struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
     static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
 
@@ -408,14 +408,14 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
         GGML_ASSERT(ptr == (void *) ((char *)(pool_addr) + pool_used));
     }
 };
-#endif // !defined(GGML_CUDA_NO_VMM)
+#endif // defined(GGML_USE_VMM)
 
 std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
-#if !defined(GGML_CUDA_NO_VMM)
+#if defined(GGML_USE_VMM)
     if (ggml_cuda_info().devices[device].vmm) {
         return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
     }
-#endif // !defined(GGML_CUDA_NO_VMM)
+#endif // defined(GGML_USE_VMM)
     return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }
 
@@ -3250,7 +3250,7 @@ static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t
         features.push_back({ "FORCE_CUBLAS", "1" });
     #endif
 
-    #ifdef GGML_CUDA_NO_VMM
+    #ifndef GGML_USE_VMM
         features.push_back({ "NO_VMM", "1" });
     #endif
 

ggml/src/ggml-hip/CMakeLists.txt

@@ -96,8 +96,8 @@ if (GGML_HIP_GRAPHS)
     add_compile_definitions(GGML_HIP_GRAPHS)
 endif()
 
-if (GGML_CUDA_NO_VMM)
-    add_compile_definitions(GGML_CUDA_NO_VMM)
+if (GGML_HIP_NO_VMM)
+    add_compile_definitions(GGML_HIP_NO_VMM)
 endif()
 
 if (CXX_IS_HIPCC)