move graph_uids to rpc_dispatcher

ggml/include/ggml-rpc.h

@@ -6,7 +6,7 @@
 extern "C" {
 #endif
 
-#define RPC_PROTO_MAJOR_VERSION    6
+#define RPC_PROTO_MAJOR_VERSION    7
 #define RPC_PROTO_MINOR_VERSION    0
 #define RPC_PROTO_PATCH_VERSION    0
 

ggml/src/ggml-backend-meta.cpp

@@ -865,7 +865,12 @@ static struct ggml_backend_meta_split_state ggml_backend_meta_get_split_state(
         ggml_backend_meta_split_state split_state;
         switch (tensor->op) {
             case GGML_OP_NONE: {
-                split_state = {GGML_BACKEND_SPLIT_AXIS_MIRRORED, {0}, {1}, 1};
+                if (tensor->view_src != nullptr) {
+                    // full-tensor view created with ggml_view_tensor, transparent for the split state
+                    split_state = ggml_backend_meta_get_split_state(stc, tensor->view_src, assume_sync);
+                } else {
+                    split_state = {GGML_BACKEND_SPLIT_AXIS_MIRRORED, {0}, {1}, 1};
+                }
             } break;
             case GGML_OP_DUP: {
                 split_state = handle_generic(src_ss, /*scalar_only =*/ true);
@@ -2283,6 +2288,14 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
                 cgraph_ij->uid = ggml_graph_next_uid();
             }
         }
+
+        // Aux graph contents are rewritten on every compute but are identical across calls while the subgraphs are reused,
+        // so they can get stable uids on rebuild. Only safe without a comm backend, where the fallback usage is deterministic.
+        if (backend_ctx->comm_ctx == nullptr) {
+            for (ggml_cgraph * cgraph_aux : backend_ctx->cgraphs_aux) {
+                cgraph_aux->uid = ggml_graph_next_uid();
+            }
+        }
     }
 
     size_t iga = 0; // i graph aux

tools/server/server-models.cpp

@@ -1129,7 +1129,8 @@ void server_models::load(const std::string & name, const load_options & opts) {
     // exceeding models_max. Without this, the window between unload_lru()
     // releasing its lock and this lock_guard acquiring allows multiple
     // threads to each observe capacity and all proceed to load.
-    if (base_params.models_max > 0) {
+    // Download workers do not use models_max slots.
+    if (opts.mode == SERVER_CHILD_MODE_NORMAL && base_params.models_max > 0) {
         size_t count_active = 0;
         for (const auto & m : mapping) {
             if (m.second.meta.is_running()) {

tools/server/tests/unit/test_router.py

@@ -540,13 +540,17 @@ def _wait_for_sse_event(collected: list, event_type: str, model: str, timeout: i
 
 
 def test_router_download_model():
-    """Case 1: download a model, verify SSE events and GET /models."""
+    """Case 1: download a model at the model limit, verify SSE events and GET /models."""
     global server
+    server.models_max = 1
     server.start()
 
     # Ensure the model is not present before we start
     server.make_request("DELETE", f"/models?model={MODEL_DOWNLOAD_ID}")
 
+    # A download worker must not consume or evict a model slot
+    _load_model_and_wait(MODEL_B, timeout=120)
+
     sse_events: list = []
     stop = threading.Event()
     sse_ready = threading.Event()
@@ -580,6 +584,7 @@ def test_router_download_model():
     # Model should now appear in GET /models
     ids = _get_model_ids(is_reload=False)
     assert MODEL_DOWNLOAD_ID in ids, f"{MODEL_DOWNLOAD_ID} not found in /models after download"
+    assert _get_model_status(MODEL_B) == "loaded"
 
 
 def test_router_delete_model():