nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 18 Packages Projects Releases Wiki Activity

rpc : use graph uid instead of graph cache (#22701)

Browse Source 
Store the last graph uid and compare against it to determine if the same
graph is being computed.
This commit is contained in:
Radoslav Gerganov
2026-05-05 13:47:13 +03:00
committed by GitHub
parent 2635ac76e8
commit d5003b6e4d
1 changed files with 7 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-rpc/ggml-rpc.cpp
+4 -28
View File
@@ -207,35 +207,11 @@ struct ggml_backend_rpc_buffer_type_context {
size_t max_size; size_t max_size;
}; };
struct graph_cache {
bool is_cached(const ggml_cgraph * cgraph) {
if ((int)last_graph.size() != cgraph->n_nodes) {
return false;
}
for (int i = 0; i < cgraph->n_nodes; i++) {
if (memcmp(&last_graph[i], cgraph->nodes[i], sizeof(ggml_tensor)) != 0) {
return false;
}
}
return true;
}
void add(const ggml_cgraph * cgraph) {
last_graph.resize(cgraph->n_nodes);
for (int i = 0; i < cgraph->n_nodes; i++) {
memcpy(&last_graph[i], cgraph->nodes[i], sizeof(ggml_tensor));
}
}
std::vector<ggml_tensor> last_graph;
};
struct ggml_backend_rpc_context { struct ggml_backend_rpc_context {
std::string endpoint; std::string endpoint;
uint32_t device; uint32_t device;
std::string name; std::string name;
graph_cache gc; uint64_t last_graph_uid;
}; };
struct ggml_backend_rpc_buffer_context { struct ggml_backend_rpc_buffer_context {
@@ -717,7 +693,7 @@ static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, g
ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context; ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
GGML_ASSERT(cgraph->n_nodes > 0); GGML_ASSERT(cgraph->n_nodes > 0);
bool reuse = rpc_ctx->gc.is_cached(cgraph); bool reuse = cgraph->uid != 0 && rpc_ctx->last_graph_uid == cgraph->uid;
if (reuse) { if (reuse) {
rpc_msg_graph_recompute_req request; rpc_msg_graph_recompute_req request;
request.device = rpc_ctx->device; request.device = rpc_ctx->device;
@@ -725,7 +701,7 @@ static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, g
bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_RECOMPUTE, &request, sizeof(request)); bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_RECOMPUTE, &request, sizeof(request));
RPC_STATUS_ASSERT(status); RPC_STATUS_ASSERT(status);
} else { } else {
rpc_ctx->gc.add(cgraph); rpc_ctx->last_graph_uid = cgraph->uid;
std::vector<uint8_t> input; std::vector<uint8_t> input;
serialize_graph(rpc_ctx->device, cgraph, input); serialize_graph(rpc_ctx->device, cgraph, input);
auto sock = get_socket(rpc_ctx->endpoint); auto sock = get_socket(rpc_ctx->endpoint);
@@ -794,7 +770,7 @@ ggml_backend_t ggml_backend_rpc_init(const char * endpoint, uint32_t device) {
/* .endpoint = */ endpoint, /* .endpoint = */ endpoint,
/* .device = */ device, /* .device = */ device,
/* .name = */ dev_name, /* .name = */ dev_name,
/* .gc = */ {}, /* .last_graph_uid = */ 0,
}; };
auto reg = ggml_backend_rpc_add_server(endpoint); auto reg = ggml_backend_rpc_add_server(endpoint);
ggml_backend_t backend = new ggml_backend { ggml_backend_t backend = new ggml_backend {