Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	.gitignore
#	CMakeLists.txt
#	flake.lock
#	llama.cpp

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

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-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));

llama.cpp

@@ -2190,8 +2190,7 @@ 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; // llama_token_to_piece(special = true);
-    std::vector<token> cache_token_to_piece_special; // llama_token_to_piece(special = true);
 
     std::map<std::pair<std::string, std::string>, int> bpe_ranks;
 
@@ -4908,23 +4907,19 @@ 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(n_vocab);
-        std::vector<llama_vocab::token> cache_token_to_piece_special(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[id] = llama_token_to_piece(&model, id, true);
-            cache_token_to_piece_special[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);
 
         LLAMA_LOG_INFO("%s: token to piece cache size = %.4f MB\n", __func__, size_cache / 1024.0 / 1024.0);
     }
@@ -18638,9 +18633,14 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token
         return llama_token_to_piece_old(model, token, buf, length);
     }
 
+    // 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);