cmake : add pkg-config spec file for llama.cpp (#7702)

* support lm_head

* remove the code block

---------

Co-authored-by: zhangkaihuo <zhangkaihuo@modelbest.cn>

CMakeLists.txt

@@ -1373,6 +1373,13 @@ if (LLAMA_METAL)
     endif()
 endif()
 
+configure_file(cmake/llama.pc.in
+        "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
+        @ONLY)
+
+install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
+        DESTINATION lib/pkgconfig)
+
 #
 # programs, examples and tests
 #

cmake/llama.pc.in

@@ -0,0 +1,10 @@
+prefix=@CMAKE_INSTALL_PREFIX@
+exec_prefix=${prefix}
+libdir=${exec_prefix}/lib
+includedir=${prefix}/include
+
+Name: llama
+Description: Port of Facebook's LLaMA model in C/C++
+Version: @PROJECT_VERSION@
+Libs: -L${libdir} -lllama
+Cflags: -I${includedir}

gguf-py/gguf/constants.py

@@ -645,6 +645,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
     ],
     MODEL_ARCH.MINICPM: [
         MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
         MODEL_TENSOR.OUTPUT_NORM,
         MODEL_TENSOR.ROPE_FREQS,
         MODEL_TENSOR.ATTN_NORM,

llama.cpp

@@ -2164,8 +2164,7 @@ struct llama_vocab {
     std::vector<token_data>       id_to_token;
 
     std::vector<id>    cache_special_tokens;
-    std::vector<token> cache_token_to_piece;         // llama_token_to_piece(special = false);
-    std::vector<token> cache_token_to_piece_special; // llama_token_to_piece(special = true);
+    std::vector<token> cache_token_to_piece; // llama_token_to_piece(special = true);
 
     std::map<std::pair<std::string, std::string>, int> bpe_ranks;
 
@@ -4845,23 +4844,19 @@ static void llm_load_vocab(
         LLAMA_LOG_INFO("%s: special tokens cache size = %u\n", __func__, (uint32_t)vocab.cache_special_tokens.size());
     }
 
-    // build token to piece caches
+    // build token to piece cache
     {
         size_t size_cache = 0;
 
-        std::vector<llama_vocab::token> cache_token_to_piece        (n_vocab);
-        std::vector<llama_vocab::token> cache_token_to_piece_special(n_vocab);
+        std::vector<llama_vocab::token> cache_token_to_piece(n_vocab);
 
         for (uint32_t id = 0; id < n_vocab; ++id) {
-            cache_token_to_piece[id]         = llama_token_to_piece(&model, id, false);
-            cache_token_to_piece_special[id] = llama_token_to_piece(&model, id, true);
+            cache_token_to_piece[id] = llama_token_to_piece(&model, id, true);
 
             size_cache += cache_token_to_piece[id].size();
-            size_cache += cache_token_to_piece_special[id].size();
         }
 
-        std::swap(vocab.cache_token_to_piece,         cache_token_to_piece);
-        std::swap(vocab.cache_token_to_piece_special, cache_token_to_piece_special);
+        std::swap(vocab.cache_token_to_piece, cache_token_to_piece);
 
         LLAMA_LOG_INFO("%s: token to piece cache size = %.4f MB\n", __func__, size_cache / 1024.0 / 1024.0);
     }
@@ -5129,12 +5124,10 @@ static bool llm_load_tensors(
                     // output
                     {
                         model.output_norm = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
-                        if (model.arch != LLM_ARCH_MINICPM){
-                            model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
-                            // if output is NULL, init from the input tok embed
-                            if (model.output == NULL) {
-                                model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
-                            }
+                        model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        // if output is NULL, init from the input tok embed
+                        if (model.output == NULL) {
+                            model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED);
                         }
                     }
 
@@ -10217,7 +10210,7 @@ struct llm_build_context {
         cb(cur, "lmhead_scaling", -1);
 
         // lm_head
-        cur = ggml_mul_mat(ctx0, model.tok_embd, cur);
+        cur = ggml_mul_mat(ctx0, model.output, cur);
         cb(cur, "result_output", -1);
 
         ggml_build_forward_expand(gf, cur);
@@ -18318,9 +18311,14 @@ static std::string llama_decode_text(const std::string & text) {
 
 // does not write null-terminator to buf
 int32_t llama_token_to_piece(const struct llama_model * model, llama_token token, char * buf, int32_t length, bool special) {
+    // ref: https://github.com/ggerganov/llama.cpp/pull/7587#discussion_r1620983843
+    if (!special && llama_is_control_token(model->vocab, token)) {
+        return 0;
+    }
+
     // if we have a cache - use it
     {
-        const auto & cache = special ? model->vocab.cache_token_to_piece_special : model->vocab.cache_token_to_piece;
+        const auto & cache = model->vocab.cache_token_to_piece;
 
         if (!cache.empty()) {
             const auto & res = cache.at(token);