CUDA/HIP: Share the same unified memory allocation logic. (llama/12934)

Replace compile-time `GGML_HIP_UMA` with environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY`. This unifies the usage on NVIDIA and AMD GPUs, and allows a single binary to be shared between integrated and dedicated GPUs.

diff --git a/ggml/CMakeLists.txt b/ggml/CMakeLists.txt
index d33f843b417cfd8e03fd7556fb78fe3120eec4bd..438c2a7309191984613614afa068ab6fa3b70a1f 100644 (file)
--- a/ggml/CMakeLists.txt
+++ b/ggml/CMakeLists.txt
@@ -170,7 +170,6 @@ option(GGML_HIP                             "ggml: use HIP"
 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_ROCWMMA_FATTN               "ggml: enable rocWMMA for FlashAttention"         OFF)
-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)
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 4af1897017567205d842dfa0ffca75d87de78e24..9ced4665127888bcf77b1cb62b90fd0054f8de63 100644 (file)
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -96,31 +96,32 @@ int ggml_cuda_get_device() {
 
 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
     ggml_cuda_set_device(device);
-#if defined(GGML_USE_HIP) && defined(GGML_HIP_UMA)
-    auto res = hipMallocManaged(ptr, size);
-    if (res == hipSuccess) {
-        // if error we "need" to know why...
-        CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
-    }
-    return res;
-#else
-
-#if !defined(GGML_USE_HIP)
     cudaError_t err;
     if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
     {
         err = cudaMallocManaged(ptr, size);
+#if defined(GGML_USE_HIP)
+        if (err == hipSuccess) {
+            CUDA_CHECK(cudaMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
+        }
+
+        // fall back to cudaMalloc if not supported (e.g. on Windows)
+        if (err == hipErrorNotSupported) {
+            static bool warned_unsupported = false;
+            if (!warned_unsupported) {
+                GGML_LOG_WARN("hipMallocManaged unsupported, falling back to hipMalloc.\n");
+                warned_unsupported = true;
+            }
+
+            err = cudaMalloc(ptr, size);
+        }
+#endif // defined(GGML_USE_HIP)
     }
     else
     {
         err = cudaMalloc(ptr, size);
     }
     return err;
-#else
-    return cudaMalloc(ptr, size);
-#endif // !defined(GGML_USE_HIP)
-
-#endif
 }
 
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)

diff --git a/ggml/src/ggml-cuda/vendors/hip.h b/ggml/src/ggml-cuda/vendors/hip.h
index 420b41b8d652d6b6c15b780fd28e348a63643bf0..1a28831b7a96b74f880363a05450819cb2d66a5a 100644 (file)
--- a/ggml/src/ggml-cuda/vendors/hip.h
+++ b/ggml/src/ggml-cuda/vendors/hip.h
@@ -71,6 +71,8 @@
 #define cudaLaunchHostFunc hipLaunchHostFunc
 #define cudaMalloc hipMalloc
 #define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMallocManaged hipMallocManaged
+#define cudaMemAdvise hipMemAdvise
 #define cudaMemcpy hipMemcpy
 #define cudaMemcpyAsync hipMemcpyAsync
 #define cudaMemcpyPeerAsync hipMemcpyPeerAsync

diff --git a/ggml/src/ggml-hip/CMakeLists.txt b/ggml/src/ggml-hip/CMakeLists.txt
index e3762649fd27574001ec1e5bb3f51cbdc739c2c8..1fe8fe3b8d079167267d7b4efb9d935a13b57c31 100644 (file)
--- a/ggml/src/ggml-hip/CMakeLists.txt
+++ b/ggml/src/ggml-hip/CMakeLists.txt
@@ -89,10 +89,6 @@ endif()
 
 add_compile_definitions(GGML_USE_HIP)
 
-if (GGML_HIP_UMA)
-    add_compile_definitions(GGML_HIP_UMA)
-endif()
-
 if (GGML_CUDA_FORCE_MMQ)
     add_compile_definitions(GGML_CUDA_FORCE_MMQ)
 endif()