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

* support lm_head

* remove the code block

---------

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

.gitignore

@@ -105,6 +105,7 @@ examples/jeopardy/results.txt
 examples/server/*.html.hpp
 examples/server/*.js.hpp
 examples/server/*.mjs.hpp
+examples/server/*.css.hpp
 
 poetry.lock
 poetry.toml

CMakeLists.txt

@@ -777,6 +777,7 @@ if (LLAMA_KOMPUTE)
             kompute-shaders/op_mul_mat_q4_0.comp
             kompute-shaders/op_mul_mat_q4_1.comp
             kompute-shaders/op_mul_mat_q6_k.comp
+            kompute-shaders/op_getrows_f32.comp
             kompute-shaders/op_getrows_f16.comp
             kompute-shaders/op_getrows_q4_0.comp
             kompute-shaders/op_getrows_q4_1.comp
@@ -809,6 +810,7 @@ if (LLAMA_KOMPUTE)
             shaderop_mul_mat_q4_0.h
             shaderop_mul_mat_q4_1.h
             shaderop_mul_mat_q6_k.h
+            shaderop_getrows_f32.h
             shaderop_getrows_f16.h
             shaderop_getrows_q4_0.h
             shaderop_getrows_q4_1.h
@@ -1371,6 +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}

examples/rpc/rpc-server.cpp

@@ -6,6 +6,10 @@
 #include "ggml-metal.h"
 #endif
 
+#ifdef GGML_USE_SYCL
+#include "ggml-sycl.h"
+#endif
+
 #include "ggml-rpc.h"
 #ifdef _WIN32
 #  include <windows.h>
@@ -79,6 +83,12 @@ static ggml_backend_t create_backend() {
     if (!backend) {
         fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
     }
+#elif GGML_USE_SYCL
+    fprintf(stderr, "%s: using SYCL backend\n", __func__);
+    backend = ggml_backend_sycl_init(0); // init device 0
+    if (!backend) {
+        fprintf(stderr, "%s: ggml_backend_sycl_init() failed\n", __func__);
+    }
 #endif
 
     // if there aren't GPU Backends fallback to CPU backend

flake.lock

@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1715865404,
-        "narHash": "sha256-/GJvTdTpuDjNn84j82cU6bXztE0MSkdnTWClUCRub78=",
+        "lastModified": 1717285511,
+        "narHash": "sha256-iKzJcpdXih14qYVcZ9QC9XuZYnPc6T8YImb6dX166kw=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "8dc45382d5206bd292f9c2768b8058a8fd8311d9",
+        "rev": "2a55567fcf15b1b1c7ed712a2c6fadaec7412ea8",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1716509168,
-        "narHash": "sha256-4zSIhSRRIoEBwjbPm3YiGtbd8HDWzFxJjw5DYSDy1n8=",
+        "lastModified": 1716948383,
+        "narHash": "sha256-SzDKxseEcHR5KzPXLwsemyTR/kaM9whxeiJohbL04rs=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "bfb7a882678e518398ce9a31a881538679f6f092",
+        "rev": "ad57eef4ef0659193044870c731987a6df5cf56b",
         "type": "github"
       },
       "original": {
@@ -36,14 +36,14 @@
     },
     "nixpkgs-lib": {
       "locked": {
-        "lastModified": 1714640452,
-        "narHash": "sha256-QBx10+k6JWz6u7VsohfSw8g8hjdBZEf8CFzXH1/1Z94=",
+        "lastModified": 1717284937,
+        "narHash": "sha256-lIbdfCsf8LMFloheeE6N31+BMIeixqyQWbSr2vk79EQ=",
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/50eb7ecf4cd0a5756d7275c8ba36790e5bd53e33.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/eb9ceca17df2ea50a250b6b27f7bf6ab0186f198.tar.gz"
       },
       "original": {
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/50eb7ecf4cd0a5756d7275c8ba36790e5bd53e33.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/eb9ceca17df2ea50a250b6b27f7bf6ab0186f198.tar.gz"
       }
     },
     "root": {

ggml-alloc.c

@@ -377,7 +377,7 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
     galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
     GGML_ASSERT(galloc->bufts != NULL);
 
-    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t) * n_bufs);
+    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t));
     GGML_ASSERT(galloc->buffers != NULL);
 
     galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));

ggml-cuda/fattn-vec-f32.cuh

@@ -278,14 +278,10 @@ void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx,
 
 template <int D, ggml_type type_K, ggml_type type_V>
 void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    ggml_tensor * KQV = dst;
     ggml_tensor * Q   = dst->src[0];
     ggml_tensor * K   = dst->src[1];
     ggml_tensor * V   = dst->src[2];
 
-    const int32_t precision = KQV->op_params[2];
-    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
-
     GGML_ASSERT(K->type == type_K);
     GGML_ASSERT(V->type == type_V);
 

ggml-kompute.cpp

@@ -22,6 +22,7 @@
 #include "shaderop_mul_mat_q4_1.h"
 #include "shaderop_mul_mat_q6_k.h"
 #include "shaderop_mul_mat_mat_f32.h"
+#include "shaderop_getrows_f32.h"
 #include "shaderop_getrows_f16.h"
 #include "shaderop_getrows_q4_0.h"
 #include "shaderop_getrows_q4_1.h"
@@ -1146,6 +1147,14 @@ static void ggml_vk_get_rows(
     seq.record<kp::OpAlgoDispatch>(s_algo);
 }
 
+template <typename... Args>
+static void ggml_vk_get_rows_f32(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f32_comp_spv,
+        kp::shader_data::op_getrows_f32_comp_spv_len);
+
+    ggml_vk_get_rows(spirv, "f32", sizeof(float), 0, std::forward<Args>(args)...);
+}
+
 template <typename... Args>
 static void ggml_vk_get_rows_f16(Args&&... args) {
     const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f16_comp_spv,
@@ -1371,6 +1380,7 @@ static bool ggml_vk_supports_op(const struct ggml_tensor * op) {
             return op->ne[3] == 1;
         case GGML_OP_GET_ROWS:
             switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_Q4_0:
                 case GGML_TYPE_Q4_1:
@@ -1661,7 +1671,9 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
                     } break;
                 case GGML_OP_GET_ROWS:
                     {
-                        if (src0t == GGML_TYPE_F16) {
+                        if (src0t == GGML_TYPE_F32) {
+                            ggml_vk_get_rows_f32(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
+                        } else if (src0t == GGML_TYPE_F16) {
                             ggml_vk_get_rows_f16(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
                         } else if (src0t == GGML_TYPE_Q4_0) {
                             ggml_vk_get_rows_q4_0(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));

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,

kompute-shaders/op_getrows_f32.comp

@@ -0,0 +1,31 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { float inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+void dequantize_row_f32(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
+    for (int j = 0; j < k; j++) {
+        out_[y + j] = inA[x + j];
+    }
+}
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const int r = inB[i + pcs.inBOff];
+
+    dequantize_row_f32(r*pcs.nb01/4 + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
+}

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);

tests/test-backend-ops.cpp

@@ -1584,9 +1584,11 @@ struct test_flash_attn_ext : public test_case {
         : hs(hs), nh(nh), kv(kv), nb(nb), mask(mask), max_bias(max_bias), type_KV(type_KV) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
-        ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs, nb, nh, 1);
-        ggml_tensor * k = ggml_new_tensor_4d(ctx, type_KV,       hs, kv, nh, 1);
-        ggml_tensor * v = ggml_new_tensor_4d(ctx, type_KV,       hs, kv, nh, 1);
+        const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV));
+
+        ggml_tensor * q = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, hs_padded, nb, nh, 1);
+        ggml_tensor * k = ggml_new_tensor_4d(ctx, type_KV,       hs_padded, kv, nh, 1);
+        ggml_tensor * v = ggml_new_tensor_4d(ctx, type_KV,       hs_padded, kv, nh, 1);
         ggml_tensor * m = mask ? ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kv, GGML_PAD(nb, GGML_KQ_MASK_PAD), 1, 1) : nullptr;
         ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias);
         return out;