@fugood/llama.node 1.4.13 → 1.4.15

package/src/llama.cpp/common/download.h

@@ -1,12 +1,27 @@
 #pragma once
 
 #include <string>
+#include <vector>
 
 struct common_params_model;
 
-//
-// download functionalities
-//
+using common_header      = std::pair<std::string, std::string>;
+using common_header_list = std::vector<common_header>;
+
+struct common_remote_params {
+    common_header_list headers;
+    long timeout  = 0;           // in seconds, 0 means no timeout
+    long max_size = 0;           // unlimited if 0
+};
+
+// get remote file content, returns <http_code, raw_response_body>
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params);
+
+// split HF repo with tag into <repo, tag>
+// for example: "user/model:tag" -> <"user/model", "tag">
+// if tag is not present, default to "latest"
+// example: "user/model" -> <"user/model", "latest">
+std::pair<std::string, std::string> common_download_split_repo_tag(const std::string & hf_repo_with_tag);
 
 struct common_cached_model_info {
     std::string manifest_path;
@@ -41,17 +56,29 @@ struct common_hf_file_res {
 common_hf_file_res common_get_hf_file(
     const std::string & hf_repo_with_tag,
     const std::string & bearer_token,
-    bool offline);
+    bool offline,
+    const common_header_list & headers = {}
+);
 
 // returns true if download succeeded
 bool common_download_model(
     const common_params_model & model,
     const std::string & bearer_token,
-    bool offline);
+    bool offline,
+    const common_header_list & headers = {}
+);
 
 // returns list of cached models
 std::vector<common_cached_model_info> common_list_cached_models();
 
+// download single file from url to local path
+// returns status code or -1 on error
+int common_download_file_single(const std::string & url,
+                                const std::string & path,
+                                const std::string & bearer_token,
+                                bool offline,
+                                const common_header_list & headers = {});
+
 // resolve and download model from Docker registry
 // return local path to downloaded model file
 std::string common_docker_resolve_model(const std::string & docker);

package/src/llama.cpp/common/preset.cpp

@@ -16,6 +16,48 @@ static std::string rm_leading_dashes(const std::string & str) {
     return str.substr(pos);
 }
 
+// only allow a subset of args for remote presets for security reasons
+// do not add more args unless absolutely necessary
+// args that output to files are strictly prohibited
+static std::set<std::string> get_remote_preset_whitelist(const std::map<std::string, common_arg> & key_to_opt) {
+    static const std::set<std::string> allowed_options = {
+        "model-url",
+        "hf-repo",
+        "hf-repo-draft",
+        "hf-repo-v", // vocoder
+        "hf-file-v", // vocoder
+        "mmproj-url",
+        "pooling",
+        "jinja",
+        "batch-size",
+        "ubatch-size",
+        "cache-reuse",
+        "chat-template-kwargs",
+        "mmap",
+        // note: sampling params are automatically allowed by default
+        // negated args will be added automatically if the positive arg is specified above
+    };
+
+    std::set<std::string> allowed_keys;
+
+    for (const auto & it : key_to_opt) {
+        const std::string & key = it.first;
+        const common_arg & opt = it.second;
+        if (allowed_options.find(key) != allowed_options.end() || opt.is_sparam) {
+            allowed_keys.insert(key);
+            // also add variant keys (args without leading dashes and env vars)
+            for (const auto & arg : opt.get_args()) {
+                allowed_keys.insert(rm_leading_dashes(arg));
+            }
+            for (const auto & env : opt.get_env()) {
+                allowed_keys.insert(env);
+            }
+        }
+    }
+
+    return allowed_keys;
+}
+
 std::vector<std::string> common_preset::to_args(const std::string & bin_path) const {
     std::vector<std::string> args;
 
@@ -121,6 +163,29 @@ void common_preset::merge(const common_preset & other) {
     }
 }
 
+void common_preset::apply_to_params(common_params & params) const {
+    for (const auto & [opt, val] : options) {
+        // apply each option to params
+        if (opt.handler_string) {
+            opt.handler_string(params, val);
+        } else if (opt.handler_int) {
+            opt.handler_int(params, std::stoi(val));
+        } else if (opt.handler_bool) {
+            opt.handler_bool(params, common_arg_utils::is_truthy(val));
+        } else if (opt.handler_str_str) {
+            // not supported yet
+            throw std::runtime_error(string_format(
+                "%s: option with two values is not supported yet",
+                __func__
+            ));
+        } else if (opt.handler_void) {
+            opt.handler_void(params);
+        } else {
+            GGML_ABORT("unknown handler type");
+        }
+    }
+}
+
 static std::map<std::string, std::map<std::string, std::string>> parse_ini_from_file(const std::string & path) {
     std::map<std::string, std::map<std::string, std::string>> parsed;
 
@@ -230,10 +295,16 @@ static std::string parse_bool_arg(const common_arg & arg, const std::string & ke
     return value;
 }
 
-common_preset_context::common_preset_context(llama_example ex)
+common_preset_context::common_preset_context(llama_example ex, bool only_remote_allowed)
         : ctx_params(common_params_parser_init(default_params, ex)) {
     common_params_add_preset_options(ctx_params.options);
     key_to_opt = get_map_key_opt(ctx_params);
+
+    // setup allowed keys if only_remote_allowed is true
+    if (only_remote_allowed) {
+        filter_allowed_keys = true;
+        allowed_keys = get_remote_preset_whitelist(key_to_opt);
+    }
 }
 
 common_presets common_preset_context::load_from_ini(const std::string & path, common_preset & global) const {
@@ -249,7 +320,18 @@ common_presets common_preset_context::load_from_ini(const std::string & path, co
         }
         LOG_DBG("loading preset: %s\n", preset.name.c_str());
         for (const auto & [key, value] : section.second) {
+            if (key == "version") {
+                // skip version key (reserved for future use)
+                continue;
+            }
+
             LOG_DBG("option: %s = %s\n", key.c_str(), value.c_str());
+            if (filter_allowed_keys && allowed_keys.find(key) == allowed_keys.end()) {
+                throw std::runtime_error(string_format(
+                    "option '%s' is not allowed in remote presets",
+                    key.c_str()
+                ));
+            }
             if (key_to_opt.find(key) != key_to_opt.end()) {
                 const auto & opt = key_to_opt.at(key);
                 if (is_bool_arg(opt)) {
@@ -259,7 +341,10 @@ common_presets common_preset_context::load_from_ini(const std::string & path, co
                 }
                 LOG_DBG("accepted option: %s = %s\n", key.c_str(), preset.options[opt].c_str());
             } else {
-                // TODO: maybe warn about unknown key?
+                throw std::runtime_error(string_format(
+                    "option '%s' not recognized in preset '%s'",
+                    key.c_str(), preset.name.c_str()
+                ));
             }
         }
 

package/src/llama.cpp/common/preset.h

@@ -6,6 +6,7 @@
 #include <string>
 #include <vector>
 #include <map>
+#include <set>
 
 //
 // INI preset parser and writer
@@ -40,6 +41,9 @@ struct common_preset {
 
     // merge another preset into this one, overwriting existing options
     void merge(const common_preset & other);
+
+    // apply preset options to common_params
+    void apply_to_params(common_params & params) const;
 };
 
 // interface for multiple presets in one file
@@ -50,7 +54,12 @@ struct common_preset_context {
     common_params default_params; // unused for now
     common_params_context ctx_params;
     std::map<std::string, common_arg> key_to_opt;
-    common_preset_context(llama_example ex);
+
+    bool filter_allowed_keys = false;
+    std::set<std::string> allowed_keys;
+
+    // if only_remote_allowed is true, only accept whitelisted keys
+    common_preset_context(llama_example ex, bool only_remote_allowed = false);
 
     // load presets from INI file
     common_presets load_from_ini(const std::string & path, common_preset & global) const;

package/src/llama.cpp/ggml/include/ggml.h

@@ -234,6 +234,11 @@
 
 #if UINTPTR_MAX == 0xFFFFFFFF
     #define GGML_MEM_ALIGN 4
+#elif defined(__EMSCRIPTEN__)
+// emscripten uses max_align_t == 8, so we need GGML_MEM_ALIGN == 8 for 64-bit wasm.
+// (for 32-bit wasm, the first conditional is true and GGML_MEM_ALIGN stays 4.)
+// ref: https://github.com/ggml-org/llama.cpp/pull/18628
+    #define GGML_MEM_ALIGN 8
 #else
     #define GGML_MEM_ALIGN 16
 #endif

package/src/llama.cpp/include/llama.h

@@ -309,6 +309,7 @@ extern "C" {
         // Keep the booleans together to avoid misalignment during copy-by-value.
         bool vocab_only;      // only load the vocabulary, no weights
         bool use_mmap;        // use mmap if possible
+        bool use_direct_io;   // use direct io, takes precedence over use_mmap
         bool use_mlock;       // force system to keep model in RAM
         bool check_tensors;   // validate model tensor data
         bool use_extra_bufts; // use extra buffer types (used for weight repacking)
@@ -494,7 +495,7 @@ extern "C" {
                     struct llama_context_params * cparams,
                                           float * tensor_split,          // writable buffer for tensor split, needs at least llama_max_devices elements
         struct llama_model_tensor_buft_override * tensor_buft_overrides, // writable buffer for overrides, needs at least llama_max_tensor_buft_overrides elements
-                                         size_t   margin,                // margin of memory to leave per device in bytes
+                                         size_t * margins,               // margins of memory to leave per device in bytes
                                        uint32_t   n_ctx_min,             // minimum context size to set when trying to reduce memory use
                             enum ggml_log_level   log_level);            // minimum log level to print during fitting, lower levels go to debug log
 
@@ -1291,7 +1292,9 @@ extern "C" {
     // available samplers:
 
     LLAMA_API struct llama_sampler * llama_sampler_init_greedy(void);
-    LLAMA_API struct llama_sampler * llama_sampler_init_dist  (uint32_t seed);
+
+    /// seed == LLAMA_DEFAULT_SEED to use a random seed.
+    LLAMA_API struct llama_sampler * llama_sampler_init_dist(uint32_t seed);
 
     /// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
     /// Setting k <= 0 makes this a noop

package/src/llama.cpp/src/CMakeLists.txt

@@ -62,6 +62,7 @@ add_library(llama
             models/ernie4-5.cpp
             models/exaone.cpp
             models/exaone4.cpp
+            models/exaone-moe.cpp
             models/falcon-h1.cpp
             models/falcon.cpp
             models/gemma-embedding.cpp

package/src/llama.cpp/src/llama-arch.cpp

@@ -81,6 +81,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_NEMOTRON_H_MOE,   "nemotron_h_moe"   },
     { LLM_ARCH_EXAONE,           "exaone"           },
     { LLM_ARCH_EXAONE4,          "exaone4"          },
+    { LLM_ARCH_EXAONE_MOE,       "exaone-moe"       },
     { LLM_ARCH_RWKV6,            "rwkv6"            },
     { LLM_ARCH_RWKV6QWEN2,       "rwkv6qwen2"       },
     { LLM_ARCH_RWKV7,            "rwkv7"            },
@@ -950,6 +951,8 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_ATTN_K_NORM,
                 LLM_TENSOR_ATTN_V,
                 LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_ATTN_QKV,
+                LLM_TENSOR_ATTN_GATE,
                 LLM_TENSOR_FFN_NORM,
                 LLM_TENSOR_FFN_GATE_INP,
                 LLM_TENSOR_FFN_GATE_EXPS,
@@ -1726,6 +1729,38 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_FFN_UP,
                 LLM_TENSOR_FFN_POST_NORM,
             };
+        case LLM_ARCH_EXAONE_MOE:
+            return {
+                LLM_TENSOR_TOKEN_EMBD,
+                LLM_TENSOR_OUTPUT_NORM,
+                LLM_TENSOR_OUTPUT,
+                LLM_TENSOR_ROPE_FREQS,
+                LLM_TENSOR_ATTN_NORM,
+                LLM_TENSOR_ATTN_Q,
+                LLM_TENSOR_ATTN_Q_NORM,
+                LLM_TENSOR_ATTN_K,
+                LLM_TENSOR_ATTN_K_NORM,
+                LLM_TENSOR_ATTN_V,
+                LLM_TENSOR_ATTN_OUT,
+                LLM_TENSOR_FFN_NORM,
+                LLM_TENSOR_FFN_GATE,
+                LLM_TENSOR_FFN_DOWN,
+                LLM_TENSOR_FFN_UP,
+                LLM_TENSOR_FFN_GATE_INP,
+                LLM_TENSOR_FFN_GATE_EXPS,
+                LLM_TENSOR_FFN_DOWN_EXPS,
+                LLM_TENSOR_FFN_UP_EXPS,
+                LLM_TENSOR_FFN_GATE_SHEXP,
+                LLM_TENSOR_FFN_UP_SHEXP,
+                LLM_TENSOR_FFN_DOWN_SHEXP,
+                LLM_TENSOR_FFN_EXP_PROBS_B,
+                LLM_TENSOR_NEXTN_EH_PROJ,
+                LLM_TENSOR_NEXTN_EMBED_TOKENS,
+                LLM_TENSOR_NEXTN_ENORM,
+                LLM_TENSOR_NEXTN_HNORM,
+                LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,
+                LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,
+            };
         case LLM_ARCH_RWKV6:
             return {
                 LLM_TENSOR_TOKEN_EMBD,

package/src/llama.cpp/src/llama-arch.h

@@ -85,6 +85,7 @@ enum llm_arch {
     LLM_ARCH_NEMOTRON_H_MOE,
     LLM_ARCH_EXAONE,
     LLM_ARCH_EXAONE4,
+    LLM_ARCH_EXAONE_MOE,
     LLM_ARCH_RWKV6,
     LLM_ARCH_RWKV6QWEN2,
     LLM_ARCH_RWKV7,

package/src/llama.cpp/src/llama-chat.cpp

@@ -57,6 +57,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
     { "exaone4",           LLM_CHAT_TEMPLATE_EXAONE_4          },
+    { "exaone-moe",        LLM_CHAT_TEMPLATE_EXAONE_MOE        },
     { "rwkv-world",        LLM_CHAT_TEMPLATE_RWKV_WORLD        },
     { "granite",           LLM_CHAT_TEMPLATE_GRANITE           },
     { "gigachat",          LLM_CHAT_TEMPLATE_GIGACHAT          },
@@ -137,6 +138,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     } else if (tmpl_contains("[gMASK]<sop>")) {
         return LLM_CHAT_TEMPLATE_CHATGLM_4;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
+        if (tmpl_contains("<|tool_declare|>")) {
+            return LLM_CHAT_TEMPLATE_EXAONE_MOE;
+        }
         return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
     } else if (tmpl_contains("<|{{ item['role'] }}|>") && tmpl_contains("<|begin_of_image|>")) {
         return LLM_CHAT_TEMPLATE_GLMEDGE;
@@ -576,6 +580,22 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "[|assistant|]";
         }
+    } else if (tmpl == LLM_CHAT_TEMPLATE_EXAONE_MOE) {
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << "<|system|>\n" << trim(message->content) << "<|endofturn|>\n";
+            } else if (role == "user") {
+                ss << "<|user|>\n" << trim(message->content) << "<|endofturn|>\n";
+            } else if (role == "assistant") {
+                ss << "<|assistant|>\n" << trim(message->content) << "<|endofturn|>\n";
+            } else if (role == "tool") {
+                ss << "<|tool|>\n" << trim(message->content) << "<|endofturn|>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "<|assistant|>\n";
+        }
     } else if (tmpl == LLM_CHAT_TEMPLATE_RWKV_WORLD) {
         // this template requires the model to have "\n\n" as EOT token
         for (size_t i = 0; i < chat.size(); i++) {

package/src/llama.cpp/src/llama-chat.h

@@ -36,6 +36,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
     LLM_CHAT_TEMPLATE_EXAONE_4,
+    LLM_CHAT_TEMPLATE_EXAONE_MOE,
     LLM_CHAT_TEMPLATE_RWKV_WORLD,
     LLM_CHAT_TEMPLATE_GRANITE,
     LLM_CHAT_TEMPLATE_GIGACHAT,

package/src/llama.cpp/src/llama-graph.cpp

@@ -96,11 +96,9 @@ void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) {
 
         int32_t * data = (int32_t *) pos_bucket->data;
 
-        for (int h = 0; h < 1; ++h) {
-            for (int j = 0; j < n_tokens; ++j) {
-                for (int i = 0; i < n_tokens; ++i) {
-                    data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(ubatch->pos[i], ubatch->pos[j], hparams.n_rel_attn_bkts, true);
-                }
+        for (int j = 0; j < n_tokens; ++j) {
+            for (int i = 0; i < n_tokens; ++i) {
+                data[j*n_tokens + i] = llama_relative_position_bucket(ubatch->pos[i], ubatch->pos[j], hparams.n_rel_attn_bkts, true);
             }
         }
     }
@@ -323,34 +321,32 @@ void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) {
     const int64_t n_tokens = ubatch->n_tokens;
 
     const auto fill_mask = [&](float * data, int n_swa, llama_swa_type swa_type) {
-        for (int h = 0; h < 1; ++h) {
-            for (int i1 = 0; i1 < n_tokens; ++i1) {
-                const llama_seq_id s1 = ubatch->seq_id[i1][0];
-                const llama_pos    p1 = ubatch->pos[i1];
+        for (int i1 = 0; i1 < n_tokens; ++i1) {
+            const llama_seq_id s1 = ubatch->seq_id[i1][0];
+            const llama_pos    p1 = ubatch->pos[i1];
 
-                const uint64_t idst = h*(n_kv*n_tokens) + i1*n_kv;
+            const uint64_t idst = i1*n_kv;
 
-                for (int i0 = 0; i0 < n_tokens; ++i0) {
-                    const llama_seq_id s0 = ubatch->seq_id[i0][0];
-                    const llama_pos p0    = ubatch->pos[i0];
-
-                    // mask different sequences
-                    if (s0 != s1) {
-                        continue;
-                    }
+            for (int i0 = 0; i0 < n_tokens; ++i0) {
+                const llama_seq_id s0 = ubatch->seq_id[i0][0];
+                const llama_pos p0    = ubatch->pos[i0];
 
-                    // mask future tokens
-                    if (cparams.causal_attn && p0 > p1) {
-                        continue;
-                    }
+                // mask different sequences
+                if (s0 != s1) {
+                    continue;
+                }
 
-                    // apply SWA if any
-                    if (llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1)) {
-                        continue;
-                    }
+                // mask future tokens
+                if (cparams.causal_attn && p0 > p1) {
+                    continue;
+                }
 
-                    data[idst + i0] = hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f;
+                // apply SWA if any
+                if (llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1)) {
+                    continue;
                 }
+
+                data[idst + i0] = hparams.use_alibi ? -std::abs(p0 - p1) : 0.0f;
             }
         }
     };
@@ -454,27 +450,19 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
 
     float * data = (float *) cross_kq_mask->data;
 
-    for (int h = 0; h < 1; ++h) {
-        for (int i = 0; i < n_tokens; ++i) {
-            for (int j = 0; j < n_enc; ++j) {
-                float f = -INFINITY;
+    for (int i = 0; i < n_tokens; ++i) {
+        for (int j = 0; j < n_enc; ++j) {
+            float f = -INFINITY;
 
-                for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
-                    const llama_seq_id seq_id = ubatch->seq_id[i][s];
+            for (int s = 0; s < ubatch->n_seq_id[i]; ++s) {
+                const llama_seq_id seq_id = ubatch->seq_id[i][s];
 
-                    if (cross->seq_ids_enc[j].find(seq_id) != cross->seq_ids_enc[j].end()) {
-                        f = 0.0f;
-                    }
+                if (cross->seq_ids_enc[j].find(seq_id) != cross->seq_ids_enc[j].end()) {
+                    f = 0.0f;
                 }
-
-                data[h*(n_enc*n_tokens) + i*n_enc + j] = f;
             }
-        }
 
-        for (int i = n_tokens; i < n_tokens; ++i) {
-            for (int j = 0; j < n_enc; ++j) {
-                data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY;
-            }
+            data[i*n_enc + j] = f;
         }
     }
 }