llama : offload to RPC in addition to other backends (#7640)

* llama : offload to RPC in addition to other backends

* - fix copy_tensor being called on the src buffer instead of the dst buffer

- always initialize views in the view_src buffer

- add RPC backend to Makefile build

- add endpoint to all RPC object names

* add rpc-server to Makefile

* Update llama.cpp

Co-authored-by: slaren <slarengh@gmail.com>

---------

Co-authored-by: slaren <slarengh@gmail.com>

Makefile

@@ -69,6 +69,10 @@ ifeq ($(UNAME_S),Darwin)
 	endif
 endif
 
+ifdef LLAMA_RPC
+	BUILD_TARGETS += rpc-server
+endif
+
 default: $(BUILD_TARGETS)
 
 test: $(TEST_TARGETS)
@@ -429,6 +433,11 @@ ifdef LLAMA_BLIS
 	MK_LDFLAGS  += -lblis -L/usr/local/lib
 endif # LLAMA_BLIS
 
+ifdef LLAMA_RPC
+	MK_CPPFLAGS   += -DGGML_USE_RPC
+	OBJS          += ggml-rpc.o
+endif # LLAMA_RPC
+
 ifdef LLAMA_CUBLAS
 # LLAMA_CUBLAS is deprecated and will be removed in the future
 	LLAMA_CUDA := 1
@@ -654,11 +663,26 @@ ggml-metal-embed.o: ggml-metal.metal ggml-common.h
 endif
 endif # LLAMA_METAL
 
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
+COMMON_H_DEPS = common/common.h common/sampling.h common/log.h llama.h
+COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o json-schema-to-grammar.o
+
 ifndef LLAMA_NO_LLAMAFILE
 sgemm.o: sgemm.cpp sgemm.h ggml.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif
 
+ifdef LLAMA_RPC
+ggml-rpc.o: ggml-rpc.cpp ggml-rpc.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+rpc-server.o: examples/rpc/rpc-server.cpp ggml-rpc.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+rpc-server: rpc-server.o ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+endif # LLAMA_RPC
+
 GF_CC := $(CC)
 include scripts/get-flags.mk
 
@@ -738,14 +762,9 @@ unicode.o: unicode.cpp unicode.h
 unicode-data.o: unicode-data.cpp unicode-data.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
-
 llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h common/log.h llama.h
-COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o json-schema-to-grammar.o
-
 common.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 

ggml-alloc.c

@@ -750,7 +750,7 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
                 // this tensor was allocated without ggml-backend
                 return;
             }
-            ggml_backend_view_init(galloc->buffers[buffer_id], tensor);
+            ggml_backend_view_init(tensor);
         }
     } else {
         if (tensor->data == NULL) {
@@ -899,12 +899,12 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
             if (t->view_src == NULL) {
                 ggml_tallocr_alloc(&tallocr, t);
             } else if (t->buffer == NULL) {
-                ggml_backend_view_init(buffer, t);
+                ggml_backend_view_init(t);
             }
         } else {
             if (t->view_src != NULL && t->buffer == NULL) {
                 // view of a pre-allocated tensor
-                ggml_backend_view_init(buffer, t);
+                ggml_backend_view_init(t);
             }
         }
     }

ggml-backend.c

@@ -151,7 +151,7 @@ void ggml_backend_buffer_reset(ggml_backend_buffer_t buffer) {
 bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst) {
     ggml_backend_buffer_t dst_buf = dst->view_src ? dst->view_src->buffer : dst->buffer;
     if (dst_buf->iface.cpy_tensor) {
-        return src->buffer->iface.cpy_tensor(dst_buf, src, dst);
+        return dst_buf->iface.cpy_tensor(dst_buf, src, dst);
     }
     return false;
 }
@@ -1887,15 +1887,15 @@ ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched,
 
 // utils
 
-void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+void ggml_backend_view_init(struct ggml_tensor * tensor) {
     GGML_ASSERT(tensor->buffer == NULL);
     GGML_ASSERT(tensor->view_src != NULL);
     GGML_ASSERT(tensor->view_src->buffer != NULL);
     GGML_ASSERT(tensor->view_src->data != NULL);
 
-    tensor->buffer = buffer;
+    tensor->buffer = tensor->view_src->buffer;
     tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
-    ggml_backend_buffer_init_tensor(buffer, tensor);
+    ggml_backend_buffer_init_tensor(tensor->buffer, tensor);
 }
 
 void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
@@ -1954,7 +1954,7 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
     struct ggml_tensor * dst = node_copies[id];
     if (dst->view_src != NULL) {
         graph_copy_init_tensor(hash_set, node_copies, node_init, src->view_src);
-        ggml_backend_view_init(dst->view_src->buffer, dst);
+        ggml_backend_view_init(dst);
     }
     else {
         ggml_backend_tensor_copy(src, dst);

ggml-backend.h

@@ -225,7 +225,7 @@ extern "C" {
 
     // Tensor initialization
     GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
-    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
 
 
 #ifdef  __cplusplus

ggml-rpc.cpp

@@ -491,7 +491,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer
     if (remote_ptr != 0) {
         ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
             ggml_backend_rpc_buffer_interface,
-            new ggml_backend_rpc_buffer_context{sock, {}, remote_ptr, "RPC"},
+            new ggml_backend_rpc_buffer_context{sock, {}, remote_ptr, "RPC[" + std::string(buft_ctx->endpoint) + "]"},
             remote_size);
         return buffer;
     } else {
@@ -692,7 +692,7 @@ GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const
 GGML_CALL ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
     ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context {
         /* .endpoint  = */ endpoint,
-        /* .name      = */ "RPC",
+        /* .name      = */ "RPC[" + std::string(endpoint) + "]",
     };
 
     ggml_backend_t backend = new ggml_backend {

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
+#if defined(GGML_USE_RPC)
-    std::string endpoint = model.rpc_servers[gpu];
+    int dev_count = (int)llama_get_device_count(model);
-    buft = ggml_backend_rpc_buffer_type(endpoint.c_str());
+    int rpc_count = (int)model.rpc_servers.size();
-#elif defined(GGML_USE_METAL)
+    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)
+    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;
-    std::string endpoint = model.rpc_servers[device];
+        const char * endpoint = model.rpc_servers[device - dev_count + rpc_count].c_str();
-    ggml_backend_rpc_get_device_memory(endpoint.c_str(), &free, &total);
+        ggml_backend_rpc_get_device_memory(endpoint, &free, &total);
         return free;
-#elif defined(GGML_USE_CUDA)
+    }
+#endif
+#if defined(GGML_USE_CUDA)
     size_t total;
     size_t free;
     ggml_backend_cuda_get_device_memory(device, &free, &total);
@@ -16160,7 +16171,7 @@ 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;
@@ -16323,17 +16334,7 @@ struct llama_context * llama_new_context_with_model(
 
     if (!hparams.vocab_only) {
         // initialize backends
-#if defined(GGML_USE_RPC)
+#if defined(GGML_USE_METAL)
-        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 (model->n_gpu_layers > 0) {
             ctx->backend_metal = ggml_backend_metal_init();
             if (ctx->backend_metal == nullptr) {
@@ -16425,6 +16426,19 @@ struct llama_context * llama_new_context_with_model(
             }
             ctx->backends.push_back(backend);
         }
+#endif
+#if defined(GGML_USE_RPC)
+        if (model->n_gpu_layers > 0) {
+            for (const auto & endpoint : model->rpc_servers) {
+                ggml_backend_t backend = ggml_backend_rpc_init(endpoint.c_str());
+                if (backend == nullptr) {
+                    LLAMA_LOG_ERROR("%s: failed to initialize RPC to '%s'\n", __func__, endpoint.c_str());
+                    llama_free(ctx);
+                    return nullptr;
+                }
+                ctx->backends.push_back(backend);
+            }
+        }
 #endif
         ctx->backend_cpu = ggml_backend_cpu_init();
         if (ctx->backend_cpu == nullptr) {