llama : offload to RPC in addition to other backends (llama/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 <redacted>
---------

Co-authored-by: slaren <redacted>

diff --git a/include/ggml/ggml-backend.h b/include/ggml/ggml-backend.h
index 744b6a77457d730a1abdb849d994f7c3d5cc81fe..c582b06850ed1a62147352ccffaab72ec00fce12 100644 (file)
--- a/include/ggml/ggml-backend.h
+++ b/include/ggml/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

diff --git a/src/ggml-alloc.c b/src/ggml-alloc.c
index 0146946ebd7649a8f77515bf01bf479867a6b837..73a3c15756ba1eb201993600cd458bce0f10b185 100644 (file)
--- a/src/ggml-alloc.c
+++ b/src/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);
             }
         }
     }

diff --git a/src/ggml-backend.c b/src/ggml-backend.c
index 9e35ce98d7ace898338369bdfd36ef26d3213bfb..05737ed69695406a3492543ab9164917e515730e 100644 (file)
--- a/src/ggml-backend.c
+++ b/src/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);

diff --git a/src/ggml-rpc.cpp b/src/ggml-rpc.cpp
index 49a20df4bd85e9511d7176256a7b7a641c07676a..679ce4f280c5f05e6a2d9003adeb91dff80104d0 100644 (file)
--- a/src/ggml-rpc.cpp
+++ b/src/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 {