Merge branch 'master' into gg/ggml-common-decl

Author: ggerganov

CMakeLists.txt

@@ -116,6 +116,7 @@ option(LLAMA_MPI                             "llama: use MPI"
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
+set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
 option(LLAMA_CPU_HBM                         "llama: use memkind for CPU HBM"                   OFF)
 
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
@@ -207,7 +208,7 @@ if (LLAMA_METAL)
         enable_language(ASM)
         add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
 
-        set(METALLIB_SOURCE "${CMAKE_SOURCE_DIR}/ggml-metal.metal")
+        set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
         file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
         set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
 
@@ -534,6 +535,10 @@ if (LLAMA_HIPBLAS)
 endif()
 
 if (LLAMA_SYCL)
+    if (NOT LLAMA_SYCL_TARGET MATCHES "^(INTEL|NVIDIA)$")
+        message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL or NVIDIA")
+    endif()
+
     if ( NOT DEFINED ENV{ONEAPI_ROOT})
         message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
     endif()
@@ -555,14 +560,21 @@ if (LLAMA_SYCL)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
+    if (LLAMA_SYCL_TARGET STREQUAL "NVIDIA")
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
+    endif()
 
     set(GGML_HEADERS_SYCL ggml-sycl.h)
     set(GGML_SOURCES_SYCL ggml-sycl.cpp)
 
     if (WIN32)
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl sycl7 OpenCL mkl_sycl_blas_dll.lib mkl_intel_ilp64_dll.lib mkl_sequential_dll.lib mkl_core_dll.lib)
     else()
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+        if (LLAMA_SYCL_TARGET STREQUAL "INTEL")
+            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+        elseif (LLAMA_SYCL_TARGET STREQUAL "NVIDIA")
+            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl pthread m dl onemkl)
+        endif()
     endif()
 endif()
 

README-sycl.md

@@ -73,6 +73,29 @@ For iGPU, please make sure the shared memory from host memory is enough. For lla
 
 For dGPU, please make sure the device memory is enough. For llama-2-7b.Q4_0, recommend the device memory is 4GB+.
 
+## Nvidia GPU
+
+### Verified
+
+|Intel GPU| Status | Verified Model|
+|-|-|-|
+|Ampere Series| Support| A100|
+
+### oneMKL
+
+The current oneMKL release does not contain the oneMKL cuBlas backend.
+As a result for Nvidia GPU's oneMKL must be built from source.
+
+```
+git clone https://github.com/oneapi-src/oneMKL
+cd oneMKL
+mkdir build
+cd build
+cmake -G Ninja .. -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON
+ninja
+// Add paths as necessary
+```
+
 ## Docker
 
 Note:
@@ -186,6 +209,9 @@ source /opt/intel/oneapi/setvars.sh
 # Or, for FP32:
 cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
+# For Nvidia GPUs
+cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
 # Build example/main only
 #cmake --build . --config Release --target main
 

README.md

@@ -8,11 +8,6 @@
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
-> [!IMPORTANT]
-> **Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962**
->
-> Vote for which quantization type provides better responses, all other parameters being the same.
-
 ### Recent API changes
 
 - [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_max_seq()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
@@ -21,6 +16,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
+- Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981
+- Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962
 - Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
 
 ----

examples/llama.android/app/src/main/cpp/llama-android.cpp

@@ -33,6 +33,45 @@ jclass la_int_var;
 jmethodID la_int_var_value;
 jmethodID la_int_var_inc;
 
+std::string cached_token_chars;
+
+bool is_valid_utf8(const char * string) {
+    if (!string) {
+        return true;
+    }
+
+    const unsigned char * bytes = (const unsigned char *)string;
+    int num;
+
+    while (*bytes != 0x00) {
+        if ((*bytes & 0x80) == 0x00) {
+            // U+0000 to U+007F
+            num = 1;
+        } else if ((*bytes & 0xE0) == 0xC0) {
+            // U+0080 to U+07FF
+            num = 2;
+        } else if ((*bytes & 0xF0) == 0xE0) {
+            // U+0800 to U+FFFF
+            num = 3;
+        } else if ((*bytes & 0xF8) == 0xF0) {
+            // U+10000 to U+10FFFF
+            num = 4;
+        } else {
+            return false;
+        }
+
+        bytes += 1;
+        for (int i = 1; i < num; ++i) {
+            if ((*bytes & 0xC0) != 0x80) {
+                return false;
+            }
+            bytes += 1;
+        }
+    }
+
+    return true;
+}
+
 static void log_callback(ggml_log_level level, const char * fmt, void * data) {
     if (level == GGML_LOG_LEVEL_ERROR)     __android_log_print(ANDROID_LOG_ERROR, TAG, fmt, data);
     else if (level == GGML_LOG_LEVEL_INFO) __android_log_print(ANDROID_LOG_INFO, TAG, fmt, data);
@@ -295,6 +334,8 @@ Java_com_example_llama_Llm_completion_1init(
         jint n_len
     ) {
 
+    cached_token_chars.clear();
+
     const auto text = env->GetStringUTFChars(jtext, 0);
     const auto context = reinterpret_cast<llama_context *>(context_pointer);
     const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
@@ -372,8 +413,16 @@ Java_com_example_llama_Llm_completion_1loop(
     }
 
     auto new_token_chars = llama_token_to_piece(context, new_token_id);
-    LOGi("new_token_chars: `%s`", new_token_chars.c_str());
-    auto new_token = env->NewStringUTF(new_token_chars.c_str());
+    cached_token_chars += new_token_chars;
+
+    jstring new_token = nullptr;
+    if (is_valid_utf8(cached_token_chars.c_str())) {
+        new_token = env->NewStringUTF(cached_token_chars.c_str());
+        LOGi("cached: %s, new_token_chars: `%s`, id: %d", cached_token_chars.c_str(), new_token_chars.c_str(), new_token_id);
+        cached_token_chars.clear();
+    } else {
+        new_token = env->NewStringUTF("");
+    }
 
     llama_batch_clear(*batch);
     llama_batch_add(*batch, new_token_id, n_cur, { 0 }, true);

examples/llama.android/app/src/main/java/com/example/llama/Llm.kt

@@ -71,7 +71,7 @@ class Llm {
         batch: Long,
         nLen: Int,
         ncur: IntVar
-    ): String
+    ): String?
 
     private external fun kv_cache_clear(context: Long)
 
@@ -115,7 +115,7 @@ class Llm {
                 val ncur = IntVar(completion_init(state.context, state.batch, message, nlen))
                 while (ncur.value <= nlen) {
                     val str = completion_loop(state.context, state.batch, nlen, ncur)
-                    if (str.isEmpty()) {
+                    if (str == null) {
                         break
                     }
                     emit(str)

examples/server/server.cpp

@@ -3195,11 +3195,12 @@ int main(int argc, char ** argv) {
         ctx_server.queue_results.add_waiting_task_id(id_task);
         ctx_server.request_completion(id_task, -1, data, false, false);
 
+        const auto completion_id = gen_chatcmplid();
         if (!json_value(data, "stream", false)) {
             server_task_result result = ctx_server.queue_results.recv(id_task);
 
             if (!result.error && result.stop) {
-                json result_oai = format_final_response_oaicompat(data, result.data);
+                json result_oai = format_final_response_oaicompat(data, result.data, completion_id);
 
                 res.set_content(result_oai.dump(-1, ' ', false, json::error_handler_t::replace), "application/json; charset=utf-8");
             } else {
@@ -3208,11 +3209,11 @@ int main(int argc, char ** argv) {
             }
             ctx_server.queue_results.remove_waiting_task_id(id_task);
         } else {
-            const auto chunked_content_provider = [id_task, &ctx_server](size_t, httplib::DataSink & sink) {
+            const auto chunked_content_provider = [id_task, &ctx_server, completion_id](size_t, httplib::DataSink & sink) {
                 while (true) {
                     server_task_result result = ctx_server.queue_results.recv(id_task);
                     if (!result.error) {
-                        std::vector<json> result_array = format_partial_response_oaicompat(result.data);
+                        std::vector<json> result_array = format_partial_response_oaicompat(result.data, completion_id);
 
                         for (auto it = result_array.begin(); it != result_array.end(); ++it) {
                             if (!it->empty()) {

examples/server/utils.hpp

@@ -378,7 +378,7 @@ static json oaicompat_completion_params_parse(
     return llama_params;
 }
 
-static json format_final_response_oaicompat(const json & request, json result, bool streaming = false) {
+static json format_final_response_oaicompat(const json & request, json result, const std::string & completion_id, bool streaming = false) {
     bool stopped_word        = result.count("stopped_word") != 0;
     bool stopped_eos         = json_value(result, "stopped_eos", false);
     int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
@@ -412,7 +412,7 @@ static json format_final_response_oaicompat(const json & request, json result, b
             {"prompt_tokens",     num_prompt_tokens},
             {"total_tokens",      num_tokens_predicted + num_prompt_tokens}
         }},
-        {"id", gen_chatcmplid()}
+        {"id", completion_id}
     };
 
     if (server_verbose) {
@@ -427,7 +427,7 @@ static json format_final_response_oaicompat(const json & request, json result, b
 }
 
 // return value is vector as there is one case where we might need to generate two responses
-static std::vector<json> format_partial_response_oaicompat(json result) {
+static std::vector<json> format_partial_response_oaicompat(json result, const std::string & completion_id) {
     if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
         return std::vector<json>({result});
     }
@@ -471,7 +471,7 @@ static std::vector<json> format_partial_response_oaicompat(json result) {
                                             {"role", "assistant"}
                                         }}}})},
                             {"created", t},
-                            {"id", gen_chatcmplid()},
+                            {"id", completion_id},
                             {"model", modelname},
                             {"object", "chat.completion.chunk"}};
 
@@ -482,7 +482,7 @@ static std::vector<json> format_partial_response_oaicompat(json result) {
                                                             {"content", content}}}
                                                             }})},
                             {"created", t},
-                            {"id", gen_chatcmplid()},
+                            {"id", completion_id},
                             {"model", modelname},
                             {"object", "chat.completion.chunk"}};
 
@@ -509,7 +509,7 @@ static std::vector<json> format_partial_response_oaicompat(json result) {
     json ret = json {
         {"choices", choices},
         {"created", t},
-        {"id",      gen_chatcmplid()},
+        {"id",      completion_id},
         {"model",   modelname},
         {"object",  "chat.completion.chunk"}
     };