Merge commit '64eda5deb9859e87a020e56bab5d2f9ca956f1de' into concedo_experimental

# Conflicts:
#	.devops/cuda.Dockerfile
#	.devops/intel.Dockerfile
#	.devops/llama-cli-cann.Dockerfile
#	.devops/musa.Dockerfile
#	.devops/rocm.Dockerfile
#	.devops/vulkan.Dockerfile
#	.github/workflows/build.yml
#	.github/workflows/docker.yml
#	README.md
#	docs/backend/SYCL.md
#	examples/llava/clip.cpp
#	examples/server_embd.py
#	ggml/src/ggml-cann/acl_tensor.cpp
#	ggml/src/ggml-cann/aclnn_ops.cpp
#	ggml/src/ggml-cann/aclnn_ops.h
#	ggml/src/ggml-cann/ggml-cann.cpp
#	src/CMakeLists.txt
#	tests/test-chat-template.cpp

examples/llava/clip.cpp

@@ -411,6 +411,7 @@ struct clip_ctx {
         gguf_free(ctx_gguf);
         ggml_backend_buffer_free(buf);
         ggml_backend_free(backend);
+        clip_image_size_free(load_image_size);
     }
 };
 
@@ -1255,9 +1256,7 @@ struct clip_model_loader {
 
         // print gguf info
         try {
-            int ftype = -1;
-            get_u32(KEY_FTYPE, ftype, false);
-            const std::string ftype_str = ggml_type_name(static_cast<ggml_type>(ftype));
+        
             std::string name;
             get_string(KEY_NAME, name, false);
             std::string description;
@@ -1268,7 +1267,6 @@ struct clip_model_loader {
             LOG_INF("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx_gguf.get()));
             LOG_INF("%s: n_tensors:    %d\n",  __func__, n_tensors);
             LOG_INF("%s: n_kv:         %d\n",  __func__, (int)gguf_get_n_kv(ctx_gguf.get()));
-            LOG_INF("%s: ftype:        %s\n",  __func__, ftype_str.c_str());
             LOG_INF("\n");
         } catch (std::runtime_error & /*e*/) {
             LOG_INF("Could not list CLIP model properties.\n");
@@ -1740,6 +1738,12 @@ struct clip_image_f32 * clip_image_f32_init() {
     return new clip_image_f32();
 }
 
+void clip_image_size_free(struct clip_image_size * load_image_size) {
+    if (load_image_size == nullptr) {
+        return;
+    }
+    delete load_image_size;
+}
 void clip_image_u8_free(struct clip_image_u8  * img) { delete img; }
 void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
 void clip_image_u8_batch_free(struct clip_image_u8_batch  * batch) {
@@ -2479,6 +2483,9 @@ ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
 }
 
 void clip_free(clip_ctx * ctx) {
+    if (ctx == nullptr) {
+        return;
+    }
     delete ctx;
 }
 
@@ -3192,10 +3199,19 @@ int clip_is_minicpmv(const struct clip_ctx * ctx) {
 bool clip_is_glm(const struct clip_ctx * ctx) {
     return ctx->has_glm_projector;
 }
+
 bool clip_is_qwen2vl(const struct clip_ctx * ctx) {
     return ctx->has_qwen2vl_merger;
 }
 
+bool clip_is_llava(const struct clip_ctx * ctx) {
+    return ctx->has_llava_projector;
+}
+
+bool clip_is_gemma3(const struct clip_ctx * ctx) {
+    return ctx->proj_type == PROJECTOR_TYPE_GEMMA3;
+}
+
 // Determine the number of encoder layers to iterate over
 int get_deepest_feature_layer(const struct clip_ctx * ctx) {
     // Get the index of the second to last layer; this is the

examples/llava/clip.h

@@ -77,6 +77,7 @@ CLIP_API struct clip_image_size * clip_image_size_init();
 CLIP_API struct clip_image_u8  * clip_image_u8_init ();
 CLIP_API struct clip_image_f32 * clip_image_f32_init();
 
+CLIP_API void clip_image_size_free (struct clip_image_size * img_size);
 CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
 CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
 CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
@@ -106,6 +107,8 @@ CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out
 CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_glm(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_llava(const struct clip_ctx * ctx);
+CLIP_API bool clip_is_gemma3(const struct clip_ctx * ctx);
 
 CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);
 

examples/server/server.cpp

@@ -1705,6 +1705,8 @@ private:
 };
 
 struct server_response {
+    bool running = true;
+
     // for keeping track of all tasks waiting for the result
     std::unordered_set<int> waiting_task_ids;
 
@@ -1759,6 +1761,10 @@ struct server_response {
         while (true) {
             std::unique_lock<std::mutex> lock(mutex_results);
             condition_results.wait(lock, [&]{
+                if (!running) {
+                    SRV_DBG("%s : queue result stop\n", __func__);
+                    std::terminate(); // we cannot return here since the caller is HTTP code
+                }
                 return !queue_results.empty();
             });
 
@@ -1789,6 +1795,10 @@ struct server_response {
             }
 
             std::cv_status cr_res = condition_results.wait_for(lock, std::chrono::seconds(timeout));
+            if (!running) {
+                SRV_DBG("%s : queue result stop\n", __func__);
+                std::terminate(); // we cannot return here since the caller is HTTP code
+            }
             if (cr_res == std::cv_status::timeout) {
                 return nullptr;
             }
@@ -1818,6 +1828,12 @@ struct server_response {
             }
         }
     }
+
+    // terminate the waiting loop
+    void terminate() {
+        running = false;
+        condition_results.notify_all();
+    }
 };
 
 struct server_context {
@@ -4491,9 +4507,10 @@ int main(int argc, char ** argv) {
     svr->new_task_queue = [&params] { return new httplib::ThreadPool(params.n_threads_http); };
 
     // clean up function, to be called before exit
-    auto clean_up = [&svr]() {
+    auto clean_up = [&svr, &ctx_server]() {
         SRV_INF("%s: cleaning up before exit...\n", __func__);
         svr->stop();
+        ctx_server.queue_results.terminate();
         llama_backend_free();
     };
 
@@ -4534,7 +4551,7 @@ int main(int argc, char ** argv) {
 
     if (!ctx_server.load_model(params)) {
         clean_up();
-        // t.join(); // FIXME: see below
+        t.join();
         LOG_ERR("%s: exiting due to model loading error\n", __func__);
         return 1;
     }
@@ -4582,7 +4599,7 @@ int main(int argc, char ** argv) {
     ctx_server.queue_tasks.start_loop();
 
     clean_up();
-    // t.join(); // FIXME: http thread may stuck if there is an on-going request. we don't need to care about this for now as the HTTP connection will already be closed at this point, but it's better to fix this
+    t.join();
 
     return 0;
 }

examples/server/tests/unit/test_embedding.py

@@ -49,6 +49,26 @@ def test_embedding_multiple():
         assert len(d['embedding']) > 1
 
 
+def test_embedding_multiple_with_fa():
+    server = ServerPreset.bert_bge_small_with_fa()
+    server.pooling = 'last'
+    server.start()
+    # one of these should trigger the FA branch (i.e. context size % 256 == 0)
+    res = server.make_request("POST", "/v1/embeddings", data={
+        "input": [
+            "a "*253,
+            "b "*254,
+            "c "*255,
+            "d "*256,
+        ],
+    })
+    assert res.status_code == 200
+    assert len(res.body['data']) == 4
+    for d in res.body['data']:
+        assert 'embedding' in d
+        assert len(d['embedding']) > 1
+
+
 @pytest.mark.parametrize(
     "input,is_multi_prompt",
     [

examples/server/tests/utils.py

@@ -323,6 +323,21 @@ class ServerPreset:
         server.server_embeddings = True
         return server
 
+    @staticmethod
+    def bert_bge_small_with_fa() -> ServerProcess:
+        server = ServerProcess()
+        server.model_hf_repo = "ggml-org/models"
+        server.model_hf_file = "bert-bge-small/ggml-model-f16.gguf"
+        server.model_alias = "bert-bge-small"
+        server.n_ctx = 1024
+        server.n_batch = 300
+        server.n_ubatch = 300
+        server.n_slots = 2
+        server.fa = True
+        server.seed = 42
+        server.server_embeddings = True
+        return server
+
     @staticmethod
     def tinyllama_infill() -> ServerProcess:
         server = ServerProcess()

examples/server/utils.hpp

@@ -3,7 +3,7 @@
 #include "common.h"
 #include "log.h"
 #include "llama.h"
-#include "common/base64.hpp"
+#include "base64.hpp"
 
 // increase max payload length to allow use of larger context size
 #define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576