Merge 3b82c3d3e7 into 39bf692af1

common/common.cpp

@@ -1344,12 +1344,15 @@ std::string get_model_endpoint() {
 }
 
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
-    llama_clear_adapter_lora(ctx);
-    for (auto & la : lora) {
-        if (la.scale != 0.0f) {
-            llama_set_adapter_lora(ctx, la.ptr, la.scale);
-        }
+    std::vector<llama_adapter_lora*> loras;
+    std::vector<float> scales;
+
+    for (auto & la: lora) {
+        loras.push_back(la.ptr);
+        scales.push_back(la.scale);
     }
+
+    llama_put_adapter_loras(ctx, loras.size(), loras.data(), scales.data());
 }
 
 struct llama_model_params common_model_params_to_llama(common_params & params) {

include/llama.h

@@ -672,6 +672,9 @@ extern "C" {
     // Remove all LoRA adapters from given context
     LLAMA_API void llama_clear_adapter_lora(struct llama_context * ctx);
 
+    // Set LoRa adapters on the context. Will only modify if the adapters currently in context are different.
+    LLAMA_API void llama_put_adapter_loras(struct llama_context * ctx, size_t num_adapters, struct llama_adapter_lora ** adapters, float * scales);
+
     // Apply a loaded control vector to a llama_context, or if data is NULL, clear
     // the currently loaded vector.
     // n_embd should be the size of a single layer's control, and data should point

src/llama-context.cpp

@@ -1089,6 +1089,40 @@ bool llama_context::rm_adapter_lora(
     return false;
 }
 
+void llama_context::put_adapter_loras(size_t num_adapters, llama_adapter_lora ** adapters, float * scales) {
+    LLAMA_LOG_DEBUG("%s: adapters = %p\n", __func__, (void *) adapters);
+
+    if (are_adapter_loras_same(num_adapters, adapters, scales)) {
+        return;
+    }
+
+    clear_adapter_lora();
+
+    for (size_t i = 0; i < num_adapters; i ++) {
+        if (scales[i] != 0.0f) {
+            set_adapter_lora(adapters[i], scales[i]);
+        }
+    }
+}
+
+bool llama_context::are_adapter_loras_same(size_t num_adapters, llama_adapter_lora ** adapters, float * scales) {
+    LLAMA_LOG_DEBUG("%s: adapters = %p\n", __func__, (void *) adapters);
+
+    if (num_adapters != loras.size()) {
+        return false;
+    }
+
+    for (size_t i = 0; i < num_adapters; i ++) {
+        auto it = loras.find(adapters[i]);
+
+        if (it == loras.end() || it->second != scales[i]) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
 void llama_context::clear_adapter_lora() {
     LLAMA_LOG_DEBUG("%s: call\n", __func__);
 
@@ -3239,6 +3273,10 @@ void llama_clear_adapter_lora(llama_context * ctx) {
     ctx->clear_adapter_lora();
 }
 
+void llama_put_adapter_loras(llama_context * ctx, size_t num_adapters, llama_adapter_lora ** adapters, float * scales) {
+    ctx->put_adapter_loras(num_adapters, adapters, scales);
+}
+
 int32_t llama_apply_adapter_cvec(
         llama_context * ctx,
                  const float * data,

src/llama-context.h

@@ -111,6 +111,10 @@ struct llama_context {
     bool rm_adapter_lora(
             llama_adapter_lora * adapter);
 
+    void put_adapter_loras(size_t num_adapters, llama_adapter_lora ** adapters, float * scales);
+
+    bool are_adapter_loras_same(size_t num_adapters, llama_adapter_lora ** adapters, float * scales);
+
     void clear_adapter_lora();
 
     bool apply_adapter_cvec(