talk-llama : sync llama.cpp

examples/talk-llama/llama-arch.cpp

@@ -179,6 +179,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_HF_JSON,              "tokenizer.huggingface.json"              },
     { LLM_KV_TOKENIZER_RWKV,                 "tokenizer.rwkv.world"                    },
     { LLM_KV_TOKENIZER_CHAT_TEMPLATE,        "tokenizer.chat_template"                 },
+    { LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,      "tokenizer.chat_template.%s"              },
     { LLM_KV_TOKENIZER_FIM_PRE_ID,           "tokenizer.ggml.fim_pre_token_id"         },
     { LLM_KV_TOKENIZER_FIM_SUF_ID,           "tokenizer.ggml.fim_suf_token_id"         },
     { LLM_KV_TOKENIZER_FIM_MID_ID,           "tokenizer.ggml.fim_mid_token_id"         },
@@ -1023,6 +1024,9 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_OUTPUT,          "output" },
             { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
             { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
             { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
             { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
@@ -1443,10 +1447,11 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_CONVNEXT_GAMMA,             {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 };
 
-LLM_KV::LLM_KV(llm_arch arch) : arch(arch) {}
+LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}
 
 std::string LLM_KV::operator()(llm_kv kv) const {
-    return ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
+    return suffix ? ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch), suffix)
+        : ::format(LLM_KV_NAMES.at(kv), LLM_ARCH_NAMES.at(arch));
 }
 
 std::string LLM_TN_IMPL::str() const {

examples/talk-llama/llama-arch.h

@@ -177,6 +177,7 @@ enum llm_kv {
     LLM_KV_TOKENIZER_HF_JSON,
     LLM_KV_TOKENIZER_RWKV,
     LLM_KV_TOKENIZER_CHAT_TEMPLATE,
+    LLM_KV_TOKENIZER_CHAT_TEMPLATE_N,
     LLM_KV_TOKENIZER_FIM_PRE_ID,
     LLM_KV_TOKENIZER_FIM_SUF_ID,
     LLM_KV_TOKENIZER_FIM_MID_ID,
@@ -335,9 +336,10 @@ enum llm_tensor_layer {
 };
 
 struct LLM_KV {
-    LLM_KV(llm_arch arch);
+    LLM_KV(llm_arch arch, const char * suffix = nullptr);
 
     llm_arch arch;
+    const char * suffix;
 
     std::string operator()(llm_kv kv) const;
 };

examples/talk-llama/llama-chat.cpp

@@ -51,6 +51,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "llama3",            LLM_CHAT_TEMPLATE_LLAMA_3           },
     { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGML_3         },
     { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGML_4         },
+    { "glmedge",           LLM_CHAT_TEMPLATE_GLMEDGE           },
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
     { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
@@ -115,7 +116,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
         return LLM_CHAT_TEMPLATE_PHI_3;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
-        return LLM_CHAT_TEMPLATE_FALCON_3;
+        return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
     } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
         return LLM_CHAT_TEMPLATE_ZEPHYR;
     } else if (tmpl_contains("bos_token + message['role']")) {
@@ -152,7 +153,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_MINICPM;
     } else if (tmpl_contains("'Assistant: ' + message['content'] + eos_token")) {
         return LLM_CHAT_TEMPLATE_DEEPSEEK_2;
-    } else if (tmpl_contains(LU8("'<｜Assistant｜>' + message['content'] + '<｜end▁of▁sentence｜>'"))) {
+    } else if (tmpl_contains(LU8("<｜Assistant｜>")) && tmpl_contains(LU8("<｜User｜>")) && tmpl_contains(LU8("<｜end▁of▁sentence｜>"))) {
         return LLM_CHAT_TEMPLATE_DEEPSEEK_3;
     } else if (tmpl_contains("[|system|]") && tmpl_contains("[|assistant|]") && tmpl_contains("[|endofturn|]")) {
         // ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct/discussions/8#66bae61b1893d14ee8ed85bb
@@ -440,6 +441,14 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_GLMEDGE) {
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>" << "\n" << message->content;
+        }
+        if (add_ass) {
+            ss << "<|assistant|>";
+        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_MINICPM) {
         // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
         for (auto message : chat) {

examples/talk-llama/llama-chat.h

@@ -31,6 +31,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_LLAMA_3,
     LLM_CHAT_TEMPLATE_CHATGML_3,
     LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_GLMEDGE,
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
     LLM_CHAT_TEMPLATE_RWKV_WORLD,

examples/talk-llama/llama-grammar.cpp

@@ -560,7 +560,7 @@ bool llama_grammar_parser::parse(const char * src) {
             }
         }
     } catch (const std::exception & err) {
-        fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
+        fprintf(stderr, "%s: error parsing grammar: %s\n\n%s\n", __func__, err.what(), src);
         rules.clear();
         return false;
     }
@@ -960,10 +960,28 @@ struct llama_grammar * llama_grammar_init_impl(
     // Important: vec_rules has to be moved here, not copied, because stacks contains
     // pointers to elements of vec_rules. If vec_rules were copied into llama_grammar
     // then the pointers would be invalidated when the local vec_rules goes out of scope.
-    return new llama_grammar { vocab, std::move(vec_rules), std::move(stacks), {}, };
+    return new llama_grammar {
+        vocab,
+        std::move(vec_rules),
+        std::move(stacks),
+        /* .partial_utf8 = */     {},
+        /* .lazy =*/              false,
+        /* .awaiting_trigger = */ false,
+        /* .trigger_buffer = */   "",
+        /* .trigger_tokens   = */ {},
+        /* .trigger_words    = */ {},
+    };
 }
 
-struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root) {
+struct llama_grammar * llama_grammar_init_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens) {
     llama_grammar_parser parser;
 
     // if there is a grammar, parse it
@@ -1035,10 +1053,31 @@ struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab,
         }
     } while (true);
 
+    std::vector<llama_token>    vec_trigger_tokens;
+    std::vector<std::string> vec_trigger_words;
+    for (size_t i = 0; i < num_trigger_tokens; i++) {
+        GGML_ASSERT(trigger_tokens != nullptr);
+        vec_trigger_tokens.push_back(trigger_tokens[i]);
+    }
+    for (size_t i = 0; i < num_trigger_words; i++) {
+        GGML_ASSERT(trigger_words != nullptr);
+        vec_trigger_words.push_back(trigger_words[i]);
+    }
+
     // Important: vec_rules has to be moved here, not copied, because stacks contains
     // pointers to elements of vec_rules. If vec_rules were copied into llama_grammar
     // then the pointers would be invalidated when the local vec_rules goes out of scope.
-    return new llama_grammar { vocab, std::move(vec_rules), std::move(stacks), {}, };
+    return new llama_grammar {
+        vocab,
+        std::move(vec_rules),
+        std::move(stacks),
+        /* .partial_utf8 = */     {},
+        /* .lazy = */             lazy,
+        /* .awaiting_trigger = */ lazy,
+        /* .trigger_buffer = */   "",
+        std::move(vec_trigger_tokens),
+        std::move(vec_trigger_words),
+    };
 }
 
 void llama_grammar_free_impl(struct llama_grammar * grammar) {
@@ -1055,6 +1094,11 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
         grammar.rules,
         grammar.stacks,
         grammar.partial_utf8,
+        grammar.lazy,
+        grammar.awaiting_trigger,
+        grammar.trigger_buffer,
+        grammar.trigger_tokens,
+        grammar.trigger_words,
     };
 
     // redirect elements in stacks to point to new rules
@@ -1076,6 +1120,10 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
 void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_data_array * cur_p) {
     GGML_ASSERT(grammar.vocab != nullptr);
 
+    if (grammar.awaiting_trigger) {
+        return;
+    }
+
     bool allow_eog = false;
     for (const auto & stack : grammar.stacks) {
         if (stack.empty()) {
@@ -1115,6 +1163,34 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
 void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token) {
     GGML_ASSERT(grammar.vocab != nullptr);
 
+    const auto & piece = grammar.vocab->token_to_piece(token);
+
+    if (grammar.awaiting_trigger) {
+        if (std::find(grammar.trigger_tokens.begin(), grammar.trigger_tokens.end(), token) != grammar.trigger_tokens.end()) {
+            grammar.awaiting_trigger = false;
+            grammar.trigger_buffer.clear();
+            llama_grammar_accept_str(grammar, piece);
+            LLAMA_LOG_DEBUG("Grammar triggered on token %u (`%s`)", token, piece.c_str());
+            return;
+        } else {
+            // TODO: consider a smarter incremental substring search algorithm (store last position to search from).
+            grammar.trigger_buffer += piece;
+            for (const auto & word : grammar.trigger_words) {
+                auto pos = grammar.trigger_buffer.find(word);
+                if (pos != std::string::npos) {
+                    grammar.awaiting_trigger = false;
+                    auto constrained_str = grammar.trigger_buffer.substr(pos);
+                    grammar.trigger_buffer.clear();
+                    llama_grammar_accept_str(grammar, constrained_str);
+                    LLAMA_LOG_DEBUG("Grammar triggered on word `%s`", word.c_str());
+                    return;
+                }
+            }
+            LLAMA_LOG_DEBUG("Grammar still awaiting trigger after token %d (`%s`) (buffer: `%s`)\n", token, piece.c_str(), grammar.trigger_buffer.c_str());
+            return;
+        }
+    }
+
     if (grammar.vocab->is_eog(token)) {
         for (const auto & stack : grammar.stacks) {
             if (stack.empty()) {
@@ -1124,8 +1200,10 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
         GGML_ABORT("fatal error");
     }
 
-    const std::string & piece = grammar.vocab->token_to_piece(token);
+    llama_grammar_accept_str(grammar, piece);
+}
 
+void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string & piece) {
     // Note terminating 0 in decoded string
     const auto   decoded     = decode_utf8(piece, grammar.partial_utf8);
     const auto & code_points = decoded.first;
@@ -1135,5 +1213,7 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
     }
 
     grammar.partial_utf8 = decoded.second;
-    GGML_ASSERT(!grammar.stacks.empty());
+    if (grammar.stacks.empty()) {
+        throw std::runtime_error("Unexpected empty grammar stack after accepting piece: " + piece);
+    }
 }

examples/talk-llama/llama-grammar.h

@@ -114,6 +114,15 @@ struct llama_grammar {
 
     // buffer for partially generated UTF-8 sequence from accepted tokens
     llama_partial_utf8 partial_utf8;
+
+    // lazy grammars wait for trigger words or tokens before constraining the sampling.
+    // we still ahve trigger_tokens for non-lazy grammars to force printing of special trigger tokens.
+    // (useful e.g. for tool_choice=required)
+    bool                     lazy             = false;
+    bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only
+    std::string              trigger_buffer;           // Output buffered by lazy grammar. Will be cleared once trigger is found.
+    std::vector<llama_token> trigger_tokens;           // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
+    std::vector<std::string> trigger_words;
 };
 
 //
@@ -127,7 +136,15 @@ struct llama_grammar * llama_grammar_init_impl(
         size_t n_rules,
         size_t start_rule_index);
 
-struct llama_grammar * llama_grammar_init_impl(const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root);
+struct llama_grammar * llama_grammar_init_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens);
 
 void llama_grammar_free_impl(struct llama_grammar * grammar);
 
@@ -141,3 +158,7 @@ void llama_grammar_apply_impl(
 void llama_grammar_accept_impl(
               struct llama_grammar & grammar,
                        llama_token   token);
+
+void llama_grammar_accept_str(
+              struct llama_grammar & grammar,
+                 const std::string & piece);

examples/talk-llama/llama-mmap.cpp

@@ -7,6 +7,7 @@
 #include <cstring>
 #include <climits>
 #include <stdexcept>
+#include <cerrno>
 
 #ifdef __has_include
     #if __has_include(<unistd.h>)

examples/talk-llama/llama-model-loader.cpp

@@ -64,6 +64,33 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
     }
 }
 
+// return a list of splits for a given path
+// for example, given "<name>-00002-of-00004.gguf", returns list of all 4 splits
+static std::vector<std::string> llama_get_list_splits(const std::string & path, const int idx, const int n_split) {
+    std::vector<std::string> paths;
+    std::string split_prefix;
+    std::vector<char> buf(llama_path_max(), 0);
+
+    {
+        int ret = llama_split_prefix(buf.data(), buf.size(), path.c_str(), idx, n_split);
+        if (!ret) {
+            throw std::runtime_error(format("invalid split file name: %s", path.c_str()));
+        }
+        split_prefix = std::string(buf.data(), ret);
+    }
+
+    if (split_prefix.empty()) {
+        throw std::runtime_error(format("invalid split file: %s", path.c_str()));
+    }
+
+    for (int idx = 0; idx < n_split; ++idx) {
+        int ret = llama_split_path(buf.data(), buf.size(), split_prefix.c_str(), idx, n_split);
+        paths.push_back(std::string(buf.data(), ret));
+    }
+
+    return paths;
+}
+
 namespace GGUFMeta {
     template <typename T, gguf_type gt_, T (*gfun)(const gguf_context *, const int64_t)>
     struct GKV_Base_Type {
@@ -413,7 +440,12 @@ namespace GGUFMeta {
     template bool llama_model_loader::get_key_or_arr<std::array<int, 4>>(enum llm_kv kid, std::array<int, 4> & result, uint32_t n, bool required);
     template bool llama_model_loader::get_key_or_arr<std::array<uint32_t, 512>>(enum llm_kv kid, std::array<uint32_t, 512> & result, uint32_t n, bool required);
 
-llama_model_loader::llama_model_loader(const std::string & fname, bool use_mmap, bool check_tensors, const struct llama_model_kv_override * param_overrides_p) {
+llama_model_loader::llama_model_loader(
+        const std::string & fname,
+        std::vector<std::string> & splits,
+        bool use_mmap,
+        bool check_tensors,
+        const struct llama_model_kv_override * param_overrides_p) {
     int trace = 0;
     if (getenv("LLAMA_TRACE")) {
         trace = atoi(getenv("LLAMA_TRACE"));
@@ -425,6 +457,7 @@ llama_model_loader::llama_model_loader(const std::string & fname, bool use_mmap,
         }
     }
 
+    // Load the main GGUF
     struct ggml_context * ctx = NULL;
     struct gguf_init_params params = {
         /*.no_alloc = */ true,
@@ -460,35 +493,54 @@ llama_model_loader::llama_model_loader(const std::string & fname, bool use_mmap,
 
     // Load additional GGML contexts
     if (n_split > 1) {
+        // make sure the main file is loaded first
         uint16_t idx = 0;
-        get_key(llm_kv(LLM_KV_SPLIT_NO), idx);
+        const std::string kv_split_no = llm_kv(LLM_KV_SPLIT_NO);
+        get_key(kv_split_no, idx);
         if (idx != 0) {
-            throw std::runtime_error(format("illegal split file: %d, model must be loaded with the first split", idx));
+            throw std::runtime_error(format("illegal split file idx: %d (file: %s), model must be loaded with the first split", idx, fname.c_str()));
+        }
+
+        // generate list of splits if needed
+        if (splits.empty()) {
+            splits = llama_get_list_splits(fname, idx, n_split);
         }
 
-        std::vector<char> split_prefix(llama_path_max(), 0);
-        if (!llama_split_prefix(split_prefix.data(), split_prefix.size(), fname.c_str(), idx, n_split)) {
-            throw std::runtime_error(format("invalid split file: %s", fname.c_str()));
+        // in case user give a custom list of splits, check if it matches the expected number
+        if (n_split != (uint16_t)splits.size()) {
+            throw std::runtime_error(format("invalid split count, given: %zu splits, but expected %d", splits.size(), n_split));
         }
 
         if (trace > 0) {
             LLAMA_LOG_INFO("%s: loading additional %d GGUFs\n", __func__, n_split);
         }
 
-        std::vector<char> split_path(llama_path_max(), 0);
+        // load other splits
         for (idx = 1; idx < n_split; idx++) {
-            llama_split_path(split_path.data(), split_path.size(), split_prefix.data(), idx, n_split);
+            const char * fname_split = splits[idx].c_str();
 
             struct gguf_init_params split_params = {
                 /*.no_alloc = */ true,
                 /*.ctx      = */ &ctx,
             };
-            gguf_context_ptr ctx_gguf { gguf_init_from_file(split_path.data(), split_params) };
+            gguf_context_ptr ctx_gguf { gguf_init_from_file(fname_split, split_params) };
             if (!ctx_gguf) {
-                throw std::runtime_error(format("%s: failed to load GGUF split from %s\n", __func__, split_path.data()));
+                throw std::runtime_error(format("%s: failed to load GGUF split from %s\n", __func__, fname_split));
+            }
+
+            // check idx
+            {
+                const int kid = gguf_find_key(ctx_gguf.get(), kv_split_no.c_str());
+                if (kid < 0) {
+                    throw std::runtime_error(format("missing key %s in GGUF split %s", kv_split_no.c_str(), fname_split));
+                }
+                int idx_gguf = gguf_get_val_u16(ctx_gguf.get(), kid);
+                if (idx_gguf != idx) {
+                    throw std::runtime_error(format("invalid split file idx: %d (file: %s), expected %d", idx_gguf, fname_split, idx));
+                }
             }
 
-            files.emplace_back(new llama_file(split_path.data(), "rb"));
+            files.emplace_back(new llama_file(fname_split, "rb"));
             contexts.emplace_back(ctx);
 
             // Save tensors data offset info of the shard.
@@ -767,7 +819,7 @@ void llama_model_loader::init_mappings(bool prefetch, llama_mlocks * mlock_mmaps
         for (const auto & file : files) {
             auto * reg = ggml_backend_dev_backend_reg(ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU));
             auto * is_numa_fn = (decltype(ggml_is_numa) *) ggml_backend_reg_get_proc_address(reg, "ggml_backend_cpu_is_numa");
-            std::unique_ptr<llama_mmap> mapping(new llama_mmap(file.get(), prefetch ? -1 : 0, is_numa_fn()));
+            std::unique_ptr<llama_mmap> mapping = std::make_unique<llama_mmap>(file.get(), prefetch ? -1 : 0, is_numa_fn());
             mmaps_used.emplace_back(mapping->size(), 0);
             if (mlock_mmaps) {
                 std::unique_ptr<llama_mlock> mlock_mmap(new llama_mlock());

examples/talk-llama/llama-model-loader.h

@@ -90,7 +90,12 @@ struct llama_model_loader {
     size_t size_data = 0;
     std::vector<std::pair<size_t, size_t>> mmaps_used;
 
-    llama_model_loader(const std::string & fname, bool use_mmap, bool check_tensors, const struct llama_model_kv_override * param_overrides_p);
+    llama_model_loader(
+        const std::string & fname,
+        std::vector<std::string> & splits, // optional, only need if the split does not follow naming scheme
+        bool use_mmap,
+        bool check_tensors,
+        const struct llama_model_kv_override * param_overrides_p);
 
     template<typename T>
     typename std::enable_if<std::is_integral<T>::value, bool>::type

examples/talk-llama/llama-model.cpp

@@ -1093,8 +1093,20 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
                 switch (hparams.n_layer) {
-                    case 28: type = LLM_TYPE_6B; break;
-                    case 40: type = LLM_TYPE_9B; break;
+                    case 28: {
+                        if (hparams.n_head(0) == 16) {
+                            type = LLM_TYPE_1_5B;
+                        } else {
+                            type = LLM_TYPE_6B;
+                        }
+                    } break;
+                    case 40: {
+                        if (hparams.n_head(0) == 24) {
+                            type = LLM_TYPE_4B;
+                        } else {
+                            type = LLM_TYPE_9B;
+                        }
+                    } break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -1303,10 +1315,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
     const int act_gpu_layers = devices.empty() ? 0 : std::min(n_gpu_layers, (int)n_layer + 1);
     auto get_layer_buft_list = [&](int il) -> llama_model::impl::layer_dev {
         if (il < i_gpu_start || (il - i_gpu_start) >= act_gpu_layers) {
+            LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s\n", il, ggml_backend_dev_name(cpu_dev));
             return {cpu_dev, &pimpl->cpu_buft_list};
         }
         const int layer_gpu = std::upper_bound(splits.begin(), splits.begin() + n_devices(), float(il - i_gpu_start)/act_gpu_layers) - splits.begin();
         auto * dev = devices.at(layer_gpu);
+        LLAMA_LOG_DEBUG("load_tensors: layer %3d assigned to device %s\n", il, ggml_backend_dev_name(dev));
         return {dev, &pimpl->gpu_buft_list.at(dev)};
     };
 
@@ -2203,6 +2217,50 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
                     }
                 } break;
+            case LLM_ARCH_PHIMOE:
+                {
+                    const int64_t n_embd_head = n_embd / n_head;
+
+                    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+                    // output
+                    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+                    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+                    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), { n_embd, n_vocab }, 0);
+                    output_b      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "bias"),   { n_vocab }, 0);
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        auto & layer = layers[i];
+
+                        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
+                        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias",   i), { n_embd }, 0);
+
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, n_embd + 2 * n_embd_gqa }, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        if (layer.wqkv == nullptr) {
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias",   i), {n_embd}, 0);
+
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa}, 0);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias",   i), {n_embd_gqa}, 0);
+
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa}, 0);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias",   i), {n_embd_gqa}, 0);
+                        }
+                        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd }, 0);
+                        layer.bo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), { n_embd }, 0);
+
+                        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
+                        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias",   i), { n_embd }, 0);
+
+                        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert},         0);
+                        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert}, 0);
+                        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert}, 0);
+                        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert}, 0);
+
+                        layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                        layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_embd_head/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+                     }
+                } break;
             case LLM_ARCH_PLAMO:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -3022,9 +3080,17 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         auto & layer = layers[i];
 
                         layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
 
-                        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
-                        layer.bqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, 0);
+                        if (layer.wqkv == nullptr) {
+                            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+                            layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k_gqa}, 0);
+                            layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v_gqa}, 0);
+                            layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias", i), {n_embd},     llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                            layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias", i), {n_embd_gqa}, llama_model_loader::TENSOR_NOT_REQUIRED);
+                        }
 
                         layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
 
@@ -3717,7 +3783,6 @@ struct llama_model_params llama_model_default_params() {
         /*.split_mode                  =*/ LLAMA_SPLIT_MODE_LAYER,
         /*.main_gpu                    =*/ 0,
         /*.tensor_split                =*/ nullptr,
-        /*.rpc_servers                 =*/ nullptr,
         /*.progress_callback           =*/ nullptr,
         /*.progress_callback_user_data =*/ nullptr,
         /*.kv_overrides                =*/ nullptr,
@@ -3912,8 +3977,10 @@ uint64_t llama_model_size(const struct llama_model * model) {
     return model->size();
 }
 
-const char * llama_model_chat_template(const struct llama_model * model) {
-    const auto & it = model->gguf_kv.find(LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE));
+const char * llama_model_chat_template(const struct llama_model * model, const char * name) {
+    const auto key = name ? LLM_KV(model->arch, name)(LLM_KV_TOKENIZER_CHAT_TEMPLATE_N)
+        : LLM_KV(model->arch)(LLM_KV_TOKENIZER_CHAT_TEMPLATE);
+    const auto & it = model->gguf_kv.find(key);
     if (it == model->gguf_kv.end()) {
         return nullptr;
     }

examples/talk-llama/llama-model.h

@@ -323,8 +323,6 @@ struct llama_model {
     // gguf metadata
     std::unordered_map<std::string, std::string> gguf_kv;
 
-    std::vector<std::string> rpc_servers;
-
     // list of devices used in this model
     std::vector<ggml_backend_dev_t> devices;
 

examples/talk-llama/llama-quant.cpp

@@ -526,7 +526,8 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         kv_overrides = v->data();
     }
 
-    llama_model_loader ml(fname_inp, use_mmap, /*check_tensors*/ true, kv_overrides);
+    std::vector<std::string> splits = {};
+    llama_model_loader ml(fname_inp, splits, use_mmap, /*check_tensors*/ true, kv_overrides);
     ml.init_mappings(false); // no prefetching
 
     llama_model model(llama_model_default_params());

examples/talk-llama/llama-sampling.cpp

@@ -1433,13 +1433,30 @@ static void llama_sampler_grammar_apply(struct llama_sampler * smpl, llama_token
     }
 }
 
+// Fwd declare to break reset --> init_impl --> llama_sampler_grammar_i --> reset cycle.
+static struct llama_sampler * llama_sampler_init_grammar_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens);
+
 static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
     auto * ctx = (llama_sampler_grammar *) smpl->ctx;
     if (!ctx->grammar) {
         return;
     }
 
-    auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str());
+    std::vector<const char *>  trigger_words;
+    for (auto & word : ctx->grammar->trigger_words) {
+        trigger_words.push_back(word.c_str());
+    }
+    auto * grammar_new = llama_grammar_init_impl(ctx->grammar->vocab, ctx->grammar_str.c_str(), ctx->grammar_root.c_str(),
+                                                 ctx->grammar->lazy, trigger_words.data(), trigger_words.size(),
+                                                 ctx->grammar->trigger_tokens.data(), ctx->grammar->trigger_tokens.size());
 
     llama_grammar_free_impl(ctx->grammar);
     ctx->grammar = grammar_new;
@@ -1448,7 +1465,7 @@ static void llama_sampler_grammar_reset(struct llama_sampler * smpl) {
 static struct llama_sampler * llama_sampler_grammar_clone(const struct llama_sampler * smpl) {
     const auto * ctx = (const llama_sampler_grammar *) smpl->ctx;
 
-    auto * result = llama_sampler_init_grammar(ctx->vocab, nullptr, nullptr);
+    auto * result = llama_sampler_init_grammar_impl(ctx->vocab, nullptr, nullptr, false, nullptr, 0, nullptr, 0);
 
     // copy the state
     {
@@ -1484,7 +1501,15 @@ static struct llama_sampler_i llama_sampler_grammar_i = {
     /* .free   = */ llama_sampler_grammar_free,
 };
 
-struct llama_sampler * llama_sampler_init_grammar(const struct llama_vocab * vocab, const char * grammar_str, const char * grammar_root) {
+static struct llama_sampler * llama_sampler_init_grammar_impl(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                              bool lazy,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens) {
     auto * ctx = new llama_sampler_grammar;
 
     if (grammar_str != nullptr && grammar_str[0] != '\0') {
@@ -1492,7 +1517,7 @@ struct llama_sampler * llama_sampler_init_grammar(const struct llama_vocab * voc
             /* .vocab        = */ vocab,
             /* .grammar_str  = */ grammar_str,
             /* .grammar_root = */ grammar_root,
-            /* .grammar      = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root),
+            /* .grammar      = */ llama_grammar_init_impl(vocab, grammar_str, grammar_root, lazy, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens),
         };
     } else {
         *ctx = {
@@ -1509,6 +1534,24 @@ struct llama_sampler * llama_sampler_init_grammar(const struct llama_vocab * voc
     };
 }
 
+struct llama_sampler * llama_sampler_init_grammar(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root) {
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ false, nullptr, 0, nullptr, 0);
+}
+
+struct llama_sampler * llama_sampler_init_grammar_lazy(
+        const struct llama_vocab * vocab,
+                      const char * grammar_str,
+                      const char * grammar_root,
+                     const char ** trigger_words,
+                            size_t num_trigger_words,
+               const llama_token * trigger_tokens,
+                            size_t num_trigger_tokens) {
+    return llama_sampler_init_grammar_impl(vocab, grammar_str, grammar_root, /* lazy= */ true, trigger_words, num_trigger_words, trigger_tokens, num_trigger_tokens);
+}
+
 // penalties
 
 struct llama_sampler_penalties {

examples/talk-llama/llama-vocab.cpp

@@ -439,7 +439,7 @@ struct llm_tokenizer_bpe_session {
                 "also starts with a BOS token. So now the final prompt starts with 2 BOS tokens. "
                 "Are you sure this is what you want?\n", __FUNCTION__);
         }
-        if (vocab.get_add_bos() && output.size() >= 2 && *(output.end()-2) == vocab.token_eos()) {
+        if (vocab.get_add_eos() && output.size() >= 2 && *(output.end()-2) == vocab.token_eos()) {
             LLAMA_LOG_WARN(
                 "%s: Added a EOS token to the prompt as specified by the model but the prompt "
                 "also ends with a EOS token. So now the final prompt ends with 2 EOS tokens. "
@@ -1245,8 +1245,13 @@ struct llama_vocab::impl {
 
     std::vector<llama_token> cache_special_tokens;
     std::vector<std::string> cache_token_to_piece; // llama_token_to_piece(special = true);
-
-    std::map<std::pair<std::string, std::string>, int> bpe_ranks;
+    struct pair_hash {
+        size_t operator()(const std::pair<std::string, std::string> & p) const {
+            return std::hash<std::string>{}(p.first) ^  //create some hash for pair
+                   (std::hash<std::string>{}(p.second) << 1);
+        }
+    };
+    std::unordered_map<std::pair<std::string, std::string>, int, pair_hash> bpe_ranks;
 
     // set of all tokens that cause "end of generation"
     std::set<llama_token> special_eog_ids;
@@ -1356,8 +1361,9 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
             // read vocab size from metadata
             uint32_t n_tokens = 0;
-            if (!ml.get_key(LLM_KV_VOCAB_SIZE, n_tokens, false)) {
-                LLAMA_LOG_WARN("%s: there is no vocab_size in metadata\n", __func__);
+            if (ml.get_key(LLM_KV_VOCAB_SIZE, n_tokens, false)) {
+                LLAMA_LOG_WARN("%s: adding %u dummy tokens\n", __func__, n_tokens);
+                id_to_token.resize(n_tokens);
             }
 
             return;
@@ -1522,7 +1528,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_COMMAND_R;
                 clean_spaces = false;
             } else if (
-                tokenizer_pre == "qwen2") {
+                    tokenizer_pre == "qwen2" ||
+                    tokenizer_pre == "deepseek-r1-qwen") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2;
                 clean_spaces = false;
             } else if (
@@ -1685,7 +1692,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
         GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
         linefeed_id = ids[0];
     } else {
-        const std::vector<int> ids = tokenize("\xC4\x8A", false); // U+010A
+        const std::vector<int> ids = tokenize("\n", false);
 
         //GGML_ASSERT(!ids.empty() && "model vocab missing newline token");
         if (ids.empty()) {

examples/talk-llama/llama.cpp

@@ -31,7 +31,7 @@
 #endif
 
 // Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
-static int llama_model_load(const std::string & fname, llama_model & model, llama_model_params & params) {
+static int llama_model_load(const std::string & fname, std::vector<std::string> & splits, llama_model & model, llama_model_params & params) {
     // loading time will be recalculated after the first eval, so
     // we take page faults deferred by mmap() into consideration
     model.t_load_us = 0;
@@ -40,7 +40,7 @@ static int llama_model_load(const std::string & fname, llama_model & model, llam
     model.t_start_us = tm.t_start_us;
 
     try {
-        llama_model_loader ml(fname, params.use_mmap, params.check_tensors, params.kv_overrides);
+        llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.kv_overrides);
 
         ml.print_info();
 
@@ -4642,7 +4642,7 @@ struct llm_build_context {
                     0);
                 cb(v_states, "v_states", il);
 
-                q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
+                q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
                 q_pe = ggml_rope_ext(
                     ctx0, q_pe, inp_pos, rope_factors,
                     n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -4651,7 +4651,7 @@ struct llm_build_context {
                 cb(q_pe, "q_pe", il);
 
                 // shared RoPE key
-                k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
+                k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
                 k_pe = ggml_rope_ext(
                     ctx0, k_pe, inp_pos, rope_factors,
                     n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -6496,7 +6496,7 @@ struct llm_build_context {
                     0);
                 cb(v_states, "v_states", il);
 
-                q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
+                q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
                 q_pe = ggml_rope_ext(
                     ctx0, q_pe, inp_pos, nullptr,
                     n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -6505,7 +6505,7 @@ struct llm_build_context {
                 cb(q_pe, "q_pe", il);
 
                 // shared RoPE key
-                k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend does not support non-contiguous RoPE
+                k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this
                 k_pe = ggml_rope_ext(
                     ctx0, k_pe, inp_pos, nullptr,
                     n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
@@ -7215,17 +7215,30 @@ struct llm_build_context {
                 struct ggml_tensor * Qcur = nullptr;
                 struct ggml_tensor * Kcur = nullptr;
                 struct ggml_tensor * Vcur = nullptr;
-
-                cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
-                cb(cur, "wqkv", il);
-
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
-
-                Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
-
+                if (model.type == LLM_TYPE_1_5B || model.type == LLM_TYPE_4B || model.type == LLM_TYPE_9B) {
+                    Qcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, cur);
+                    if (model.layers[il].bq) {
+                        Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+                    }
+                    Kcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, cur);
+                    if (model.layers[il].bk) {
+                        Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+                    }
+                    Vcur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, cur);
+                    if (model.layers[il].bv) {
+                        Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                    }
+                } else {
+                    cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, cur);
+                    cb(cur, "wqkv", il);
+                    if (model.layers[il].bqkv) {
+                        cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                        cb(cur, "bqkv", il);
+                    }
+                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                }
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
                 cb(Vcur, "Vcur", il);
@@ -7700,17 +7713,13 @@ struct llm_build_context {
                 1
             );
 
+            struct ggml_tensor * last_norm_att = ggml_view_3d(ctx0, x_norm_att, n_embd, 1, n_seqs, x_norm_att->nb[1], x_norm_att->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(x_norm_att));
             ggml_build_forward_expand(
                 gf,
                 ggml_cpy(
                     ctx0,
-                    wkv_states,
-                    ggml_view_1d(
-                        ctx0,
-                        kv_self.v_l[il],
-                        hparams.n_embd_v_s() * n_seqs,
-                        hparams.n_embd_v_s() * kv_head * ggml_element_size(kv_self.v_l[il])
-                    )
+                    ggml_view_1d(ctx0, last_norm_att, n_embd * n_seqs, 0),
+                    ggml_view_1d(ctx0, kv_self.k_l[il], hparams.n_embd_k_s() * n_seqs, hparams.n_embd_k_s() * kv_head * ggml_element_size(kv_self.k_l[il]))
                 )
             );
 
@@ -8432,74 +8441,33 @@ static enum ggml_status llama_graph_compute(
     return status;
 }
 
-// decode a batch of tokens by evaluating the transformer
-// in case of unsuccessful decoding (error or warning),
-// the kv_cache state will be returned to its original state
-// (for non-recurrent models) or cleaned (for recurrent models)
-//
-//   - lctx:      llama context
-//   - batch:     batch to evaluate
-//
-// return 0 on success
-// return positive int on warning
-// return negative int on error
-//
-static int llama_decode_impl(
-         llama_context & lctx,
-           llama_batch   inp_batch) {
-
-    lctx.is_encoding = false;
-
-    if (inp_batch.n_tokens == 0) {
-        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
-        return -1;
-    }
-
-    // temporary allocate memory for the input batch if needed
-    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : lctx.kv_self.max_pos() + 1);
-
-    const llama_batch & batch = batch_allocr.batch;
-    const uint32_t n_tokens_all = batch.n_tokens;
-
+static int llama_prepare_sbatch(
+        llama_context     & lctx,
+        const llama_batch & batch,
+        uint32_t          & n_outputs) {
     const auto & model   = lctx.model;
-    const auto & vocab   = model.vocab;
     const auto & hparams = model.hparams;
     const auto & cparams = lctx.cparams;
 
-    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
+    const uint32_t n_tokens_all = batch.n_tokens;
+    const  int64_t n_embd       = hparams.n_embd;
+
+    // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
+    const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
 
+    GGML_ASSERT((!batch.token && batch.embd) || (batch.token && !batch.embd)); // NOLINT
     if (batch.token) {
         for (uint32_t i = 0; i < n_tokens_all; ++i) {
-            if (batch.token[i] < 0 || (uint32_t) batch.token[i] >= model.vocab.n_tokens()) {
+            if (batch.token[i] < 0 || uint32_t(batch.token[i]) >= model.vocab.n_tokens()) {
                 LLAMA_LOG_ERROR("%s: invalid token[%d] = %d\n", __func__, i, batch.token[i]);
                 return -1;
             }
         }
     }
-
     GGML_ASSERT(n_tokens_all <= cparams.n_batch);
-
     GGML_ASSERT((cparams.causal_attn || cparams.n_ubatch >= n_tokens_all) && "non-causal attention requires n_ubatch >= n_tokens");
 
-    if (lctx.t_compute_start_us == 0) {
-        lctx.t_compute_start_us = ggml_time_us();
-    }
     lctx.n_queued_tokens += n_tokens_all;
-
-    auto & kv_self = lctx.kv_self;
-    llama_kv_slot_restorer kv_slot_restorer(kv_self);
-
-    const int64_t n_embd  = hparams.n_embd;
-    const int64_t n_vocab = vocab.n_tokens();
-
-    uint32_t n_outputs = 0;
-    uint32_t n_outputs_prev = 0;
-
-    const auto n_ubatch = cparams.n_ubatch;
-
-    // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
-    const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
-
     lctx.embd_seq.clear();
 
     // count outputs
@@ -8515,7 +8483,7 @@ static int llama_decode_impl(
     }
 
     lctx.sbatch.from_batch(batch, n_embd,
-        /* simple_split */ !kv_self.recurrent,
+        /* simple_split */ !lctx.kv_self.recurrent,
         /* logits_all   */ n_outputs == n_tokens_all);
 
     // reserve output buffer
@@ -8524,70 +8492,148 @@ static int llama_decode_impl(
         return -2;
     };
 
-    while (lctx.sbatch.n_tokens > 0) {
-        llama_ubatch ubatch;
-        if (kv_self.recurrent) {
-            if (embd_pooled) {
-                // Pooled embeddings cannot be split across ubatches (yet)
-                ubatch = lctx.sbatch.split_seq(n_ubatch);
-            } else {
-                // recurrent model architectures are easier to implement
-                // with equal-length sequences
-                ubatch = lctx.sbatch.split_equal(n_ubatch);
-            }
+    return 0;
+}
+
+static int llama_prepare_ubatch(
+        llama_context          & lctx,
+        llama_kv_slot_restorer & kv_slot_restorer,
+        llama_ubatch           & ubatch,
+        const uint32_t           n_outputs,
+        const uint32_t           n_tokens_all) {
+    GGML_ASSERT(lctx.sbatch.n_tokens > 0);
+
+    auto       & kv_self = lctx.kv_self;
+    const auto & cparams = lctx.cparams;
+    const auto & hparams = lctx.model.hparams;
+
+    // this indicates we are doing pooled embedding, so we ignore batch.logits and output all tokens
+    const bool embd_pooled = cparams.embeddings && cparams.pooling_type != LLAMA_POOLING_TYPE_NONE;
+
+    if (lctx.kv_self.recurrent) {
+        if (embd_pooled) {
+            // Pooled embeddings cannot be split across ubatches (yet)
+            ubatch = lctx.sbatch.split_seq(cparams.n_ubatch);
         } else {
-            ubatch = lctx.sbatch.split_simple(n_ubatch);
+            // recurrent model architectures are easier to implement
+            // with equal-length sequences
+            ubatch = lctx.sbatch.split_equal(cparams.n_ubatch);
         }
-        const uint32_t n_tokens = ubatch.n_tokens;
+    } else {
+        ubatch = lctx.sbatch.split_simple(cparams.n_ubatch);
+    }
 
-        // count the outputs in this u_batch
-        {
-            int32_t n_outputs_new = 0;
+    // count the outputs in this u_batch
+    {
+        int32_t n_outputs_new = 0;
 
-            if (n_outputs == n_tokens_all) {
-                n_outputs_new = n_tokens;
-            } else {
-                GGML_ASSERT(ubatch.output);
-                for (uint32_t i = 0; i < n_tokens; i++) {
-                    n_outputs_new += (int32_t) (ubatch.output[i] != 0);
-                }
+        if (n_outputs == n_tokens_all) {
+            n_outputs_new = ubatch.n_tokens;
+        } else {
+            GGML_ASSERT(ubatch.output);
+            for (uint32_t i = 0; i < ubatch.n_tokens; i++) {
+                n_outputs_new += int32_t(ubatch.output[i] != 0);
             }
+        }
+
+        // needs to happen before the graph is built
+        lctx.n_outputs = n_outputs_new;
+    }
+
+    // non-causal masks do not use the KV cache
+    if (hparams.causal_attn) {
+        llama_kv_cache_update(&lctx);
 
-            // needs to happen before the graph is built
-            lctx.n_outputs = n_outputs_new;
+        // if we have enough unused cells before the current head ->
+        //   better to start searching from the beginning of the cache, hoping to fill it
+        if (kv_self.head > kv_self.used + 2*ubatch.n_tokens) {
+            kv_self.head = 0;
         }
 
-        int n_threads = n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch;
-        ggml_threadpool_t threadpool = n_tokens == 1 ? lctx.threadpool : lctx.threadpool_batch;
+        const auto slot = llama_kv_cache_find_slot(kv_self, ubatch);
+        if (!slot) {
+            return 1;
+        }
+        kv_slot_restorer.save(slot);
 
-        GGML_ASSERT(n_threads > 0);
+        if (!kv_self.recurrent) {
+            // a heuristic, to avoid attending the full cache if it is not yet utilized
+            // after enough generations, the benefit from this heuristic disappears
+            // if we start defragmenting the cache, the benefit from this will be more important
+            const uint32_t pad = llama_kv_cache_get_padding(cparams);
+            kv_self.n = std::min(kv_self.size, std::max(pad, GGML_PAD(llama_kv_cache_cell_max(kv_self), pad)));
+            //kv_self.n = llama_kv_cache_cell_max(kv_self);
+        }
+    }
 
-        // non-causal masks do not use the KV cache
-        if (hparams.causal_attn) {
-            llama_kv_cache_update(&lctx);
+    return 0;
+}
 
-            // if we have enough unused cells before the current head ->
-            //   better to start searching from the beginning of the cache, hoping to fill it
-            if (kv_self.head > kv_self.used + 2*n_tokens) {
-                kv_self.head = 0;
-            }
+// decode a batch of tokens by evaluating the transformer
+// in case of unsuccessful decoding (error or warning),
+// the kv_cache state will be returned to its original state
+// (for non-recurrent models) or cleaned (for recurrent models)
+//
+//   - lctx:      llama context
+//   - inp_batch: batch to evaluate
+//
+// return 0 on success
+// return positive int on warning
+// return negative int on error
+//
+static int llama_decode_impl(
+         llama_context & lctx,
+           llama_batch   inp_batch) {
 
-            const auto slot = llama_kv_cache_find_slot(kv_self, ubatch);
-            if (!slot) {
-                return 1;
-            }
-            kv_slot_restorer.save(slot);
+    lctx.is_encoding = false;
+
+    if (inp_batch.n_tokens == 0) {
+        LLAMA_LOG_ERROR("%s: n_tokens == 0\n", __func__);
+        return -1;
+    }
+
+    // temporarily allocate memory for the input batch if needed
+    llama_batch_allocr batch_allocr(inp_batch, inp_batch.pos ? -1 : lctx.kv_self.max_pos() + 1);
+    const llama_batch & batch = batch_allocr.batch;
+
+    const auto & model   = lctx.model;
+    const auto & vocab   = model.vocab;
+    const auto & hparams = model.hparams;
+    const auto & cparams = lctx.cparams;
 
-            if (!kv_self.recurrent) {
-                // a heuristic, to avoid attending the full cache if it is not yet utilized
-                // after enough generations, the benefit from this heuristic disappears
-                // if we start defragmenting the cache, the benefit from this will be more important
-                const uint32_t pad = llama_kv_cache_get_padding(cparams);
-                kv_self.n = std::min(kv_self.size, std::max(pad, GGML_PAD(llama_kv_cache_cell_max(kv_self), pad)));
-                //kv_self.n = llama_kv_cache_cell_max(kv_self);
+    if (lctx.t_compute_start_us == 0) {
+        lctx.t_compute_start_us = ggml_time_us();
+    }
+    auto & kv_self = lctx.kv_self;
+    llama_kv_slot_restorer kv_slot_restorer(kv_self);
+
+    const int64_t n_embd  = hparams.n_embd;
+    const int64_t n_vocab = vocab.n_tokens();
+
+    uint32_t n_outputs = 0;
+    uint32_t n_outputs_prev = 0;
+
+    {
+        const int ret = llama_prepare_sbatch(lctx, batch, n_outputs);
+        if (ret != 0) {
+            return ret;
+        }
+    }
+
+    while (lctx.sbatch.n_tokens > 0) {
+        llama_ubatch ubatch;
+        {
+            const int ret = llama_prepare_ubatch(lctx, kv_slot_restorer, ubatch, n_outputs, batch.n_tokens);
+            if (ret != 0) {
+                return ret;
             }
         }
 
+        const int         n_threads  = ubatch.n_tokens == 1 ? cparams.n_threads : cparams.n_threads_batch;
+        ggml_threadpool_t threadpool = ubatch.n_tokens == 1 ? lctx.threadpool   : lctx.threadpool_batch;
+
+        GGML_ASSERT(n_threads > 0);
+
         //printf("kv_self.n = %5d, kv_self.used = %5d, kv_self.head = %5d\n", kv_self.n, kv_self.used, kv_self.head);
 
         ggml_backend_sched_reset(lctx.sched.get());
@@ -8640,7 +8686,7 @@ static int llama_decode_impl(
 
         // update the kv ring buffer
         {
-            kv_self.head += n_tokens;
+            kv_self.head += ubatch.n_tokens;
 
             // Ensure kv cache head points to a valid index.
             if (kv_self.head >= kv_self.size) {
@@ -9374,14 +9420,9 @@ int64_t llama_time_us(void) {
     return ggml_time_us();
 }
 
-struct llama_model * llama_load_model_from_file(
-        const char * path_model,
-        struct llama_model_params params) {
-    return llama_model_load_from_file(path_model, params);
-}
-
-struct llama_model * llama_model_load_from_file(
-        const char * path_model,
+static struct llama_model * llama_model_load_from_file_impl(
+        const std::string & path_model,
+        std::vector<std::string> & splits,
         struct llama_model_params params) {
     ggml_time_init();
 
@@ -9404,53 +9445,13 @@ struct llama_model * llama_model_load_from_file(
         };
     }
 
-    if (params.rpc_servers != nullptr && params.rpc_servers[0] != '\0') {
-        // split the servers set them into model->rpc_servers
-        std::string servers(params.rpc_servers);
-        size_t pos = 0;
-        while ((pos = servers.find(',')) != std::string::npos) {
-            std::string server = servers.substr(0, pos);
-            model->rpc_servers.push_back(server);
-            servers.erase(0, pos + 1);
-        }
-        model->rpc_servers.push_back(servers);
-    }
-
-    // add RPC devices
-    if (!model->rpc_servers.empty()) {
-        ggml_backend_reg_t rpc_reg = ggml_backend_reg_by_name("RPC");
-        if (!rpc_reg) {
-            LLAMA_LOG_ERROR("%s: failed to find RPC backend\n", __func__);
-            llama_model_free(model);
-            return nullptr;
-        }
-
-        typedef ggml_backend_dev_t (*ggml_backend_rpc_add_device_t)(const char * endpoint);
-        ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
-        if (!ggml_backend_rpc_add_device_fn) {
-            LLAMA_LOG_ERROR("%s: failed to find RPC device add function\n", __func__);
-            llama_model_free(model);
-            return nullptr;
-        }
-
-        for (const std::string & server : model->rpc_servers) {
-            ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
-            if (dev) {
-                model->devices.push_back(dev);
-            } else {
-                LLAMA_LOG_ERROR("%s: failed to add RPC device for server '%s'\n", __func__, server.c_str());
-                llama_model_free(model);
-                return nullptr;
-            }
-        }
-    }
-
     // create list of devices to use with this model
     if (params.devices) {
         for (ggml_backend_dev_t * dev = params.devices; *dev; ++dev) {
             model->devices.push_back(*dev);
         }
     } else {
+        std::vector<ggml_backend_dev_t> rpc_servers;
         // use all available devices
         for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
             ggml_backend_dev_t dev = ggml_backend_dev_get(i);
@@ -9461,10 +9462,19 @@ struct llama_model * llama_model_load_from_file(
                     break;
 
                 case GGML_BACKEND_DEVICE_TYPE_GPU:
-                    model->devices.push_back(dev);
+                    ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
+                    if (ggml_backend_reg_name(reg) == std::string("RPC")) {
+                        rpc_servers.push_back(dev);
+                    } else {
+                        model->devices.push_back(dev);
+                    }
                     break;
             }
         }
+        // add RPC servers at the front of the list
+        if (!rpc_servers.empty()) {
+            model->devices.insert(model->devices.begin(), rpc_servers.begin(), rpc_servers.end());
+        }
     }
 
     // if using single GPU mode, remove all except the main GPU
@@ -9485,7 +9495,7 @@ struct llama_model * llama_model_load_from_file(
         LLAMA_LOG_INFO("%s: using device %s (%s) - %zu MiB free\n", __func__, ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), free/1024/1024);
     }
 
-    const int status = llama_model_load(path_model, *model, params);
+    const int status = llama_model_load(path_model, splits, *model, params);
     GGML_ASSERT(status <= 0);
     if (status < 0) {
         if (status == -1) {
@@ -9501,6 +9511,35 @@ struct llama_model * llama_model_load_from_file(
     return model;
 }
 
+// deprecated
+struct llama_model * llama_load_model_from_file(
+        const char * path_model,
+        struct llama_model_params params) {
+    return llama_model_load_from_file(path_model, params);
+}
+
+struct llama_model * llama_model_load_from_file(
+        const char * path_model,
+        struct llama_model_params params) {
+    std::vector<std::string> splits = {};
+    return llama_model_load_from_file_impl(path_model, splits, params);
+}
+
+struct llama_model * llama_model_load_from_splits(
+        const char ** paths,
+        size_t n_paths,
+        struct llama_model_params params) {
+    std::vector<std::string> splits;
+    if (n_paths == 0) {
+        LLAMA_LOG_ERROR("%s: list of splits is empty\n", __func__);
+        return nullptr;
+    }
+    for (size_t i = 0; i < n_paths; ++i) {
+        splits.push_back(paths[i]);
+    }
+    return llama_model_load_from_file_impl(splits.front(), splits, params);
+}
+
 struct llama_context * llama_init_from_model(
                  struct llama_model * model,
         struct llama_context_params   params) {

examples/talk-llama/llama.h

@@ -288,9 +288,6 @@ extern "C" {
         // proportion of the model (layers or rows) to offload to each GPU, size: llama_max_devices()
         const float * tensor_split;
 
-        // comma separated list of RPC servers to use for offloading
-        const char * rpc_servers;
-
         // Called with a progress value between 0.0 and 1.0. Pass NULL to disable.
         // If the provided progress_callback returns true, model loading continues.
         // If it returns false, model loading is immediately aborted.
@@ -418,10 +415,20 @@ extern "C" {
               struct llama_model_params   params),
             "use llama_model_load_from_file instead");
 
+    // Load the model from a file
+    // If the file is split into multiple parts, the file name must follow this pattern: <name>-%05d-of-%05d.gguf
+    // If the split file name does not follow this pattern, use llama_model_load_from_splits
     LLAMA_API struct llama_model * llama_model_load_from_file(
                              const char * path_model,
               struct llama_model_params   params);
 
+    // Load the model from multiple splits (support custom naming scheme)
+    // The paths must be in the correct order
+    LLAMA_API struct llama_model * llama_model_load_from_splits(
+                             const char ** paths,
+                                 size_t    n_paths,
+              struct llama_model_params    params);
+
     DEPRECATED(LLAMA_API void llama_free_model(struct llama_model * model),
             "use llama_model_free instead");
 
@@ -503,7 +510,8 @@ extern "C" {
     LLAMA_API uint64_t llama_model_size(const struct llama_model * model);
 
     // Get the default chat template. Returns nullptr if not available
-    LLAMA_API const char * llama_model_chat_template(const struct llama_model * model);
+    // If name is NULL, returns the default chat template
+    LLAMA_API const char * llama_model_chat_template(const struct llama_model * model, const char * name);
 
     // Returns the total number of parameters in the model
     LLAMA_API uint64_t llama_model_n_params(const struct llama_model * model);
@@ -951,7 +959,7 @@ extern "C" {
     LLAMA_API llama_token llama_vocab_fim_rep(const struct llama_vocab * vocab);
     LLAMA_API llama_token llama_vocab_fim_sep(const struct llama_vocab * vocab);
 
-    DEPRECATED(LLAMA_API const char * llama_token_get_text(const struct llama_vocab * vocab, llama_token token), "use llama_vocabable_get_text instead");
+    DEPRECATED(LLAMA_API const char * llama_token_get_text(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_text instead");
     DEPRECATED(LLAMA_API float llama_token_get_score(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_score instead");
     DEPRECATED(LLAMA_API enum llama_token_attr llama_token_get_attr(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_get_attr instead");
     DEPRECATED(LLAMA_API bool llama_token_is_eog(const struct llama_vocab * vocab, llama_token token), "use llama_vocab_is_eog instead");
@@ -1191,6 +1199,18 @@ extern "C" {
                           const char * grammar_str,
                           const char * grammar_root);
 
+    /// @details Lazy grammar sampler, introduced in https://github.com/ggerganov/llama.cpp/pull/9639
+    /// @param trigger_words A list of words that will trigger the grammar sampler. This may be updated to a loose regex syntax (w/ ^) in a near future.
+    /// @param trigger_tokens A list of tokens that will trigger the grammar sampler.
+    LLAMA_API struct llama_sampler * llama_sampler_init_grammar_lazy(
+            const struct llama_vocab * vocab,
+                          const char * grammar_str,
+                          const char * grammar_root,
+                         const char ** trigger_words,
+                                size_t num_trigger_words,
+                   const llama_token * trigger_tokens,
+                                size_t num_trigger_tokens);
+
     /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first.
     LLAMA_API struct llama_sampler * llama_sampler_init_penalties(
                              int32_t   penalty_last_n,   // last n tokens to penalize (0 = disable penalty, -1 = context size)

examples/talk-llama/unicode.cpp

@@ -7,18 +7,17 @@
 
 #include <algorithm>
 #include <cassert>
+#include <codecvt>
 #include <cstddef>
 #include <cstdint>
+#include <locale>
 #include <map>
 #include <regex>
 #include <stdexcept>
 #include <string>
 #include <unordered_map>
-#include <unordered_set>
 #include <utility>
 #include <vector>
-#include <locale>
-#include <codecvt>
 
 size_t unicode_len_utf8(char src) {
     const size_t lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };