llama : offload to RPC in addition to other backends

Author: rgerganov

ggml-backend.c

@@ -321,7 +321,9 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
         ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
     } else if (ggml_backend_buffer_is_host(dst->buffer)) {
         ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
-    } else if (!ggml_backend_buffer_copy_tensor(src, dst)) {
+    }
+    bool same_backend = strcmp(ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer)) == 0;
+    if (!same_backend || !ggml_backend_buffer_copy_tensor(src, dst)) {
 #ifndef NDEBUG
         fprintf(stderr, "%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
 #endif

llama.cpp

@@ -2371,13 +2371,34 @@ struct llama_context {
     struct llama_control_vector cvec;
 };
 
+static size_t llama_get_device_count(const llama_model & model) {
+    size_t count = 1;
+#if defined(GGML_USE_CUDA)
+    count = ggml_backend_cuda_get_device_count();
+#elif defined(GGML_USE_SYCL)
+    count = ggml_backend_sycl_get_device_count();
+#elif defined(GGML_USE_VULKAN)
+    count = ggml_backend_vk_get_device_count();
+#endif
+#if defined(GGML_USE_RPC)
+    count += model.rpc_servers.size();
+#endif
+    return count;
+    GGML_UNUSED(model);
+}
+
 static ggml_backend_buffer_type_t llama_default_buffer_type_offload(const llama_model & model, int gpu) {
     ggml_backend_buffer_type_t buft = nullptr;
 
-#ifdef GGML_USE_RPC
-    std::string endpoint = model.rpc_servers[gpu];
-    buft = ggml_backend_rpc_buffer_type(endpoint.c_str());
-#elif defined(GGML_USE_METAL)
+#if defined(GGML_USE_RPC)
+    int dev_count = (int)llama_get_device_count(model);
+    int rpc_count = (int)model.rpc_servers.size();
+    if (gpu >= dev_count - rpc_count) {
+        const char * endpoint = model.rpc_servers[gpu - dev_count + rpc_count].c_str();
+        return ggml_backend_rpc_buffer_type(endpoint);
+    }
+#endif
+#if defined(GGML_USE_METAL)
     buft = ggml_backend_metal_buffer_type();
 #elif defined(GGML_USE_CUDA)
     buft = ggml_backend_cuda_buffer_type(gpu);
@@ -2425,29 +2446,19 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_split(const llama_mo
     GGML_UNUSED(tensor_split);
 }
 
-static size_t llama_get_device_count(const llama_model & model) {
-#if defined(GGML_USE_RPC)
-    return model.rpc_servers.size();
-#elif defined(GGML_USE_CUDA)
-    return ggml_backend_cuda_get_device_count();
-#elif defined(GGML_USE_SYCL)
-    return ggml_backend_sycl_get_device_count();
-#elif defined(GGML_USE_VULKAN)
-    return ggml_backend_vk_get_device_count();
-#else
-    return 1;
-#endif
-    GGML_UNUSED(model);
-}
-
 static size_t llama_get_device_memory(const llama_model & model, int device) {
 #if defined(GGML_USE_RPC)
-    size_t total;
-    size_t free;
-    std::string endpoint = model.rpc_servers[device];
-    ggml_backend_rpc_get_device_memory(endpoint.c_str(), &free, &total);
-    return free;
-#elif defined(GGML_USE_CUDA)
+    int dev_count = (int)llama_get_device_count(model);
+    int rpc_count = (int)model.rpc_servers.size();
+    if (device >= dev_count - rpc_count) {
+        size_t total;
+        size_t free;
+        const char * endpoint = model.rpc_servers[device - dev_count + rpc_count].c_str();
+        ggml_backend_rpc_get_device_memory(endpoint, &free, &total);
+        return free;
+    }
+#endif
+#if defined(GGML_USE_CUDA)
     size_t total;
     size_t free;
     ggml_backend_cuda_get_device_memory(device, &free, &total);
@@ -16162,7 +16173,7 @@ struct llama_model * llama_load_model_from_file(
             return true;
         };
     }
-    if (params.rpc_servers != nullptr) {
+    if (params.rpc_servers != nullptr && params.rpc_servers[0] != '\0') {
         // split the servers set them into model->rpc_servers
         std::string servers(params.rpc_servers);
         size_t pos = 0;
@@ -16325,17 +16336,7 @@ struct llama_context * llama_new_context_with_model(
 
     if (!hparams.vocab_only) {
         // initialize backends
-#if defined(GGML_USE_RPC)
-        for (auto & server : model->rpc_servers) {
-            ggml_backend_t backend = ggml_backend_rpc_init(server.c_str());
-            if (backend == nullptr) {
-                LLAMA_LOG_ERROR("%s: failed to connect RPC backend to %s\n", __func__, server.c_str());
-                llama_free(ctx);
-                return nullptr;
-            }
-            ctx->backends.push_back(backend);
-        }
-#elif defined(GGML_USE_METAL)
+#if defined(GGML_USE_METAL)
         if (model->n_gpu_layers > 0) {
             ctx->backend_metal = ggml_backend_metal_init();
             if (ctx->backend_metal == nullptr) {
@@ -16427,6 +16428,18 @@ struct llama_context * llama_new_context_with_model(
             }
             ctx->backends.push_back(backend);
         }
+#endif
+#if defined(GGML_USE_RPC)
+        for (int i = 0; i < (int)model->rpc_servers.size(); i++) {
+            const char * endpoint = model->rpc_servers[i].c_str();
+            ggml_backend_t backend = ggml_backend_rpc_init(endpoint);
+            if (backend == nullptr) {
+                LLAMA_LOG_ERROR("%s: failed to initialize RPC to '%s'\n", __func__, endpoint);
+                llama_free(ctx);
+                return nullptr;
+            }
+            ctx->backends.push_back(backend);
+        }
 #endif
         ctx->backend_cpu = ggml_backend_cpu_init();
         if (ctx->backend_cpu == nullptr) {