llama : make loras compatible with repacking

ggml-ci

Author: ggerganov

src/llama-adapter.cpp

@@ -247,6 +247,27 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
         }
     }
 
+    // get extra buffer types of the CPU
+    std::vector<ggml_backend_buffer_type_t> buft_extra;
+    {
+        auto * cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        auto * cpu_reg = ggml_backend_dev_backend_reg(cpu_dev);
+
+        // add the default CPU buffer type which will be used as a fallback if the lora needs to be loaded to an extra buft
+        buft_extra.emplace_back(ggml_backend_dev_buffer_type(cpu_dev));
+
+        auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t)
+            ggml_backend_reg_get_proc_address(cpu_reg, "ggml_backend_dev_get_extra_bufts");
+
+        if (ggml_backend_dev_get_extra_bufts_fn) {
+            ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(cpu_dev);
+            while (extra_bufts && *extra_bufts) {
+                buft_extra.emplace_back(*extra_bufts);
+                ++extra_bufts;
+            }
+        }
+    }
+
     // add tensors
     for (auto & it : ab_map) {
         const std::string & name = it.first;
@@ -263,7 +284,20 @@ static void llama_adapter_lora_init_impl(llama_model & model, const char * path_
             throw std::runtime_error("LoRA tensor '" + name + "' does not exist in base model (hint: maybe wrong base model?)");
         }
 
-        ggml_context * dev_ctx = ctx_for_buft(ggml_backend_buffer_get_type(model_tensor->buffer));
+        auto * buft = ggml_backend_buffer_get_type(model_tensor->buffer);
+
+        // do not load loras to extra buffer types (i.e. bufts for repacking) -> use the CPU in that case
+        for (auto & ex : buft_extra) {
+            if (ex == buft) {
+                LLAMA_LOG_WARN("%s: lora for '%s' cannot use buft '%s'\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
+                buft = buft_extra[0];
+                break;
+            }
+        }
+
+        LLAMA_LOG_DEBUG("%s: lora for '%s' -> '%s'\n", __func__, model_tensor->name, ggml_backend_buft_name(buft));
+
+        ggml_context * dev_ctx = ctx_for_buft(buft);
         // validate tensor shape
         if (is_token_embd) {
             // expect B to be non-transposed, A and B are flipped; see llm_build_inp_embd()