@fugood/llama.node 0.3.9 → 0.3.11

package/src/llama.cpp/examples/run/linenoise.cpp/linenoise.h

@@ -45,6 +45,7 @@ extern "C" {
 #endif
 
 #include <stddef.h> /* For size_t. */
+#include <stdlib.h>
 
 extern const char *linenoiseEditMore;
 
@@ -69,10 +70,23 @@ struct linenoiseState {
     int history_index;  /* The history index we are currently editing. */
 };
 
-typedef struct linenoiseCompletions {
-  size_t len;
-  char **cvec;
-} linenoiseCompletions;
+struct linenoiseCompletions {
+    size_t  len     = 0;
+    char ** cvec    = nullptr;
+    bool    to_free = true;
+
+    ~linenoiseCompletions() {
+        if (!to_free) {
+            return;
+        }
+
+        for (size_t i = 0; i < len; ++i) {
+            free(cvec[i]);
+        }
+
+        free(cvec);
+    }
+};
 
 /* Non blocking API. */
 int linenoiseEditStart(struct linenoiseState *l, int stdin_fd, int stdout_fd, char *buf, size_t buflen, const char *prompt);

package/src/llama.cpp/examples/run/run.cpp

@@ -24,14 +24,16 @@
 #include <string>
 #include <vector>
 
+#include "chat-template.hpp"
 #include "common.h"
 #include "json.hpp"
 #include "linenoise.cpp/linenoise.h"
 #include "llama-cpp.h"
+#include "log.h"
 
 #if defined(__unix__) || (defined(__APPLE__) && defined(__MACH__)) || defined(_WIN32)
 [[noreturn]] static void sigint_handler(int) {
-    printf("\n\033[0m");
+    printf("\n" LOG_COL_DEFAULT);
     exit(0);  // not ideal, but it's the only way to guarantee exit in all cases
 }
 #endif
@@ -64,6 +66,13 @@ static int printe(const char * fmt, ...) {
     return ret;
 }
 
+static std::string strftime_fmt(const char * fmt, const std::tm & tm) {
+    std::ostringstream oss;
+    oss << std::put_time(&tm, fmt);
+
+    return oss.str();
+}
+
 class Opt {
   public:
     int init(int argc, const char ** argv) {
@@ -105,6 +114,7 @@ class Opt {
     llama_model_params   model_params;
     std::string model_;
     std::string          user;
+    bool                 use_jinja   = false;
     int                  context_size = -1, ngl = -1;
     float                temperature = -1;
     bool                 verbose     = false;
@@ -145,7 +155,8 @@ class Opt {
                 if (handle_option_with_value(argc, argv, i, context_size) == 1) {
                     return 1;
                 }
-            } else if (options_parsing && (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "--ngl") == 0)) {
+            } else if (options_parsing &&
+                       (strcmp(argv[i], "-n") == 0 || strcmp(argv[i], "-ngl") == 0 || strcmp(argv[i], "--ngl") == 0)) {
                 if (handle_option_with_value(argc, argv, i, ngl) == 1) {
                     return 1;
                 }
@@ -156,6 +167,8 @@ class Opt {
             } else if (options_parsing &&
                        (parse_flag(argv, i, "-v", "--verbose") || parse_flag(argv, i, "-v", "--log-verbose"))) {
                 verbose = true;
+            } else if (options_parsing && strcmp(argv[i], "--jinja") == 0) {
+                use_jinja = true;
             } else if (options_parsing && parse_flag(argv, i, "-h", "--help")) {
                 help = true;
                 return 0;
@@ -176,6 +189,10 @@ class Opt {
             }
         }
 
+        if (model_.empty()){
+            return 1;
+        }
+
         return 0;
     }
 
@@ -190,7 +207,7 @@ class Opt {
             "Options:\n"
             "  -c, --context-size <value>\n"
             "      Context size (default: %d)\n"
-            "  -n, --ngl <value>\n"
+            "  -n, -ngl, --ngl <value>\n"
             "      Number of GPU layers (default: %d)\n"
             "  --temp <value>\n"
             "      Temperature (default: %.1f)\n"
@@ -314,6 +331,10 @@ class HttpClient {
   public:
     int init(const std::string & url, const std::vector<std::string> & headers, const std::string & output_file,
              const bool progress, std::string * response_str = nullptr) {
+        if (std::filesystem::exists(output_file)) {
+            return 0;
+        }
+
         std::string output_file_partial;
         curl = curl_easy_init();
         if (!curl) {
@@ -341,7 +362,11 @@ class HttpClient {
         data.file_size = set_resume_point(output_file_partial);
         set_progress_options(progress, data);
         set_headers(headers);
-        perform(url);
+        CURLcode res = perform(url);
+        if (res != CURLE_OK){
+            printe("Fetching resource '%s' failed: %s\n", url.c_str(), curl_easy_strerror(res));
+            return 1;
+        }
         if (!output_file.empty()) {
             std::filesystem::rename(output_file_partial, output_file);
         }
@@ -406,16 +431,12 @@ class HttpClient {
         }
     }
 
-    void perform(const std::string & url) {
-        CURLcode res;
+    CURLcode perform(const std::string & url) {
         curl_easy_setopt(curl, CURLOPT_URL, url.c_str());
         curl_easy_setopt(curl, CURLOPT_FOLLOWLOCATION, 1L);
         curl_easy_setopt(curl, CURLOPT_DEFAULT_PROTOCOL, "https");
         curl_easy_setopt(curl, CURLOPT_FAILONERROR, 1L);
-        res = curl_easy_perform(curl);
-        if (res != CURLE_OK) {
-            printe("curl_easy_perform() failed: %s\n", curl_easy_strerror(res));
-        }
+        return curl_easy_perform(curl);
     }
 
     static std::string human_readable_time(double seconds) {
@@ -553,13 +574,14 @@ class LlamaData {
         }
 
         sampler = initialize_sampler(opt);
+
         return 0;
     }
 
   private:
 #ifdef LLAMA_USE_CURL
-    int download(const std::string & url, const std::vector<std::string> & headers, const std::string & output_file,
-                 const bool progress, std::string * response_str = nullptr) {
+    int download(const std::string & url, const std::string & output_file, const bool progress,
+                 const std::vector<std::string> & headers = {}, std::string * response_str = nullptr) {
         HttpClient http;
         if (http.init(url, headers, output_file, progress, response_str)) {
             return 1;
@@ -568,48 +590,85 @@ class LlamaData {
         return 0;
     }
 #else
-    int download(const std::string &, const std::vector<std::string> &, const std::string &, const bool,
+    int download(const std::string &, const std::string &, const bool, const std::vector<std::string> & = {},
                  std::string * = nullptr) {
         printe("%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
+
         return 1;
     }
 #endif
 
-    int huggingface_dl(const std::string & model, const std::vector<std::string> headers, const std::string & bn) {
+    // Helper function to handle model tag extraction and URL construction
+    std::pair<std::string, std::string> extract_model_and_tag(std::string & model, const std::string & base_url) {
+        std::string  model_tag = "latest";
+        const size_t colon_pos = model.find(':');
+        if (colon_pos != std::string::npos) {
+            model_tag = model.substr(colon_pos + 1);
+            model     = model.substr(0, colon_pos);
+        }
+
+        std::string url = base_url + model + "/manifests/" + model_tag;
+
+        return { model, url };
+    }
+
+    // Helper function to download and parse the manifest
+    int download_and_parse_manifest(const std::string & url, const std::vector<std::string> & headers,
+                                    nlohmann::json & manifest) {
+        std::string manifest_str;
+        int         ret = download(url, "", false, headers, &manifest_str);
+        if (ret) {
+            return ret;
+        }
+
+        manifest = nlohmann::json::parse(manifest_str);
+
+        return 0;
+    }
+
+    int huggingface_dl(std::string & model, const std::string & bn) {
         // Find the second occurrence of '/' after protocol string
         size_t pos = model.find('/');
         pos        = model.find('/', pos + 1);
+        std::string              hfr, hff;
+        std::vector<std::string> headers = { "User-Agent: llama-cpp", "Accept: application/json" };
+        std::string              url;
+
         if (pos == std::string::npos) {
-            return 1;
+            auto [model_name, manifest_url] = extract_model_and_tag(model, "https://huggingface.co/v2/");
+            hfr                             = model_name;
+
+            nlohmann::json manifest;
+            int            ret = download_and_parse_manifest(manifest_url, headers, manifest);
+            if (ret) {
+                return ret;
+            }
+
+            hff = manifest["ggufFile"]["rfilename"];
+        } else {
+            hfr = model.substr(0, pos);
+            hff = model.substr(pos + 1);
         }
 
-        const std::string hfr = model.substr(0, pos);
-        const std::string hff = model.substr(pos + 1);
-        const std::string url = "https://huggingface.co/" + hfr + "/resolve/main/" + hff;
-        return download(url, headers, bn, true);
+        url = "https://huggingface.co/" + hfr + "/resolve/main/" + hff;
+
+        return download(url, bn, true, headers);
     }
 
-    int ollama_dl(std::string & model, const std::vector<std::string> headers, const std::string & bn) {
+    int ollama_dl(std::string & model, const std::string & bn) {
+        const std::vector<std::string> headers = { "Accept: application/vnd.docker.distribution.manifest.v2+json" };
         if (model.find('/') == std::string::npos) {
             model = "library/" + model;
         }
 
-        std::string model_tag = "latest";
-        size_t      colon_pos = model.find(':');
-        if (colon_pos != std::string::npos) {
-            model_tag = model.substr(colon_pos + 1);
-            model     = model.substr(0, colon_pos);
-        }
-
-        std::string manifest_url = "https://registry.ollama.ai/v2/" + model + "/manifests/" + model_tag;
-        std::string manifest_str;
-        const int   ret = download(manifest_url, headers, "", false, &manifest_str);
+        auto [model_name, manifest_url] = extract_model_and_tag(model, "https://registry.ollama.ai/v2/");
+        nlohmann::json manifest;
+        int            ret = download_and_parse_manifest(manifest_url, {}, manifest);
         if (ret) {
             return ret;
         }
 
-        nlohmann::json manifest = nlohmann::json::parse(manifest_str);
-        std::string    layer;
+        std::string layer;
         for (const auto & l : manifest["layers"]) {
             if (l["mediaType"] == "application/vnd.ollama.image.model") {
                 layer = l["digest"];
@@ -617,8 +676,67 @@ class LlamaData {
             }
         }
 
-        std::string blob_url = "https://registry.ollama.ai/v2/" + model + "/blobs/" + layer;
-        return download(blob_url, headers, bn, true);
+        std::string blob_url = "https://registry.ollama.ai/v2/" + model_name + "/blobs/" + layer;
+
+        return download(blob_url, bn, true, headers);
+    }
+
+    int github_dl(const std::string & model, const std::string & bn) {
+        std::string  repository = model;
+        std::string  branch     = "main";
+        const size_t at_pos     = model.find('@');
+        if (at_pos != std::string::npos) {
+            repository = model.substr(0, at_pos);
+            branch     = model.substr(at_pos + 1);
+        }
+
+        const std::vector<std::string> repo_parts = string_split(repository, "/");
+        if (repo_parts.size() < 3) {
+            printe("Invalid GitHub repository format\n");
+            return 1;
+        }
+
+        const std::string & org          = repo_parts[0];
+        const std::string & project      = repo_parts[1];
+        std::string         url          = "https://raw.githubusercontent.com/" + org + "/" + project + "/" + branch;
+        for (size_t i = 2; i < repo_parts.size(); ++i) {
+            url += "/" + repo_parts[i];
+        }
+
+        return download(url, bn, true);
+    }
+
+    int s3_dl(const std::string & model, const std::string & bn) {
+        const size_t slash_pos = model.find('/');
+        if (slash_pos == std::string::npos) {
+            return 1;
+        }
+
+        const std::string bucket     = model.substr(0, slash_pos);
+        const std::string key        = model.substr(slash_pos + 1);
+        const char * access_key = std::getenv("AWS_ACCESS_KEY_ID");
+        const char * secret_key = std::getenv("AWS_SECRET_ACCESS_KEY");
+        if (!access_key || !secret_key) {
+            printe("AWS credentials not found in environment\n");
+            return 1;
+        }
+
+        // Generate AWS Signature Version 4 headers
+        // (Implementation requires HMAC-SHA256 and date handling)
+        // Get current timestamp
+        const time_t                   now     = time(nullptr);
+        const tm                       tm      = *gmtime(&now);
+        const std::string              date     = strftime_fmt("%Y%m%d", tm);
+        const std::string              datetime = strftime_fmt("%Y%m%dT%H%M%SZ", tm);
+        const std::vector<std::string> headers  = {
+            "Authorization: AWS4-HMAC-SHA256 Credential=" + std::string(access_key) + "/" + date +
+                "/us-east-1/s3/aws4_request",
+            "x-amz-content-sha256: UNSIGNED-PAYLOAD", "x-amz-date: " + datetime
+        };
+
+        const std::string url = "https://" + bucket + ".s3.amazonaws.com/" + key;
+
+        return download(url, bn, true, headers);
     }
 
     std::string basename(const std::string & path) {
@@ -630,37 +748,44 @@ class LlamaData {
         return path.substr(pos + 1);
     }
 
-    int remove_proto(std::string & model_) {
-        const std::string::size_type pos = model_.find("://");
+    int rm_until_substring(std::string & model_, const std::string & substring) {
+        const std::string::size_type pos = model_.find(substring);
         if (pos == std::string::npos) {
             return 1;
         }
 
-        model_ = model_.substr(pos + 3);  // Skip past "://"
+        model_ = model_.substr(pos + substring.size());  // Skip past the substring
         return 0;
     }
 
     int resolve_model(std::string & model_) {
         int                            ret     = 0;
         if (string_starts_with(model_, "file://") || std::filesystem::exists(model_)) {
-            remove_proto(model_);
+            rm_until_substring(model_, "://");
 
             return ret;
         }
 
-        const std::string              bn      = basename(model_);
-        const std::vector<std::string> headers = { "--header",
-                                                   "Accept: application/vnd.docker.distribution.manifest.v2+json" };
-        if (string_starts_with(model_, "hf://") || string_starts_with(model_, "huggingface://")) {
-            remove_proto(model_);
-            ret = huggingface_dl(model_, headers, bn);
-        } else if (string_starts_with(model_, "ollama://")) {
-            remove_proto(model_);
-            ret = ollama_dl(model_, headers, bn);
-        } else if (string_starts_with(model_, "https://")) {
-            download(model_, headers, bn, true);
-        } else {
-            ret = ollama_dl(model_, headers, bn);
+        const std::string bn = basename(model_);
+        if (string_starts_with(model_, "hf://") || string_starts_with(model_, "huggingface://") ||
+            string_starts_with(model_, "hf.co/")) {
+            rm_until_substring(model_, "hf.co/");
+            rm_until_substring(model_, "://");
+            ret = huggingface_dl(model_, bn);
+        } else if ((string_starts_with(model_, "https://") || string_starts_with(model_, "http://")) &&
+                   !string_starts_with(model_, "https://ollama.com/library/")) {
+            ret = download(model_, bn, true);
+        } else if (string_starts_with(model_, "github:") || string_starts_with(model_, "github://")) {
+            rm_until_substring(model_, "github:");
+            rm_until_substring(model_, "://");
+            ret = github_dl(model_, bn);
+        } else if (string_starts_with(model_, "s3://")) {
+            rm_until_substring(model_, "://");
+            ret = s3_dl(model_, bn);
+        } else {  // ollama:// or nothing
+            rm_until_substring(model_, "ollama.com/library/");
+            rm_until_substring(model_, "://");
+            ret = ollama_dl(model_, bn);
         }
 
         model_ = bn;
@@ -713,13 +838,39 @@ static void add_message(const char * role, const std::string & text, LlamaData &
 }
 
 // Function to apply the chat template and resize `formatted` if needed
-static int apply_chat_template(LlamaData & llama_data, const bool append) {
+static int apply_chat_template(const common_chat_template & tmpl, LlamaData & llama_data, const bool append, bool use_jinja) {
+    if (use_jinja) {
+        json messages = json::array();
+        for (const auto & msg : llama_data.messages) {
+            messages.push_back({
+                {"role", msg.role},
+                {"content", msg.content},
+            });
+        }
+        try {
+            minja::chat_template_inputs tmpl_inputs;
+            tmpl_inputs.messages = messages;
+            tmpl_inputs.add_generation_prompt = append;
+
+            minja::chat_template_options tmpl_opts;
+            tmpl_opts.use_bos_token = false;
+            tmpl_opts.use_eos_token = false;
+
+            auto result = tmpl.apply(tmpl_inputs, tmpl_opts);
+            llama_data.fmtted.resize(result.size() + 1);
+            memcpy(llama_data.fmtted.data(), result.c_str(), result.size() + 1);
+            return result.size();
+        } catch (const std::exception & e) {
+            printe("failed to render the chat template: %s\n", e.what());
+            return -1;
+        }
+    }
     int result = llama_chat_apply_template(
-        llama_model_chat_template(llama_data.model.get()), llama_data.messages.data(), llama_data.messages.size(), append,
+        tmpl.source().c_str(), llama_data.messages.data(), llama_data.messages.size(), append,
         append ? llama_data.fmtted.data() : nullptr, append ? llama_data.fmtted.size() : 0);
     if (append && result > static_cast<int>(llama_data.fmtted.size())) {
         llama_data.fmtted.resize(result);
-        result = llama_chat_apply_template(llama_model_chat_template(llama_data.model.get()), llama_data.messages.data(),
+        result = llama_chat_apply_template(tmpl.source().c_str(), llama_data.messages.data(),
                                            llama_data.messages.size(), append, llama_data.fmtted.data(),
                                            llama_data.fmtted.size());
     }
@@ -729,10 +880,12 @@ static int apply_chat_template(LlamaData & llama_data, const bool append) {
 
 // Function to tokenize the prompt
 static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt,
-                           std::vector<llama_token> & prompt_tokens) {
-    const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, true, true);
+                           std::vector<llama_token> & prompt_tokens, const LlamaData & llama_data) {
+    const bool is_first = llama_get_kv_cache_used_cells(llama_data.context.get()) == 0;
+
+    const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
     prompt_tokens.resize(n_prompt_tokens);
-    if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true,
+    if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), is_first,
                        true) < 0) {
         printe("failed to tokenize the prompt\n");
         return -1;
@@ -746,7 +899,7 @@ static int check_context_size(const llama_context_ptr & ctx, const llama_batch &
     const int n_ctx      = llama_n_ctx(ctx.get());
     const int n_ctx_used = llama_get_kv_cache_used_cells(ctx.get());
     if (n_ctx_used + batch.n_tokens > n_ctx) {
-        printf("\033[0m\n");
+        printf(LOG_COL_DEFAULT "\n");
         printe("context size exceeded\n");
         return 1;
     }
@@ -778,7 +931,7 @@ static int generate(LlamaData & llama_data, const std::string & prompt, std::str
     const llama_vocab * vocab = llama_model_get_vocab(llama_data.model.get());
 
     std::vector<llama_token> tokens;
-    if (tokenize_prompt(vocab, prompt, tokens) < 0) {
+    if (tokenize_prompt(vocab, prompt, tokens, llama_data) < 0) {
         return 1;
     }
 
@@ -809,7 +962,7 @@ static int generate(LlamaData & llama_data, const std::string & prompt, std::str
         batch = llama_batch_get_one(&new_token_id, 1);
     }
 
-    printf("\033[0m");
+    printf(LOG_COL_DEFAULT);
     return 0;
 }
 
@@ -818,7 +971,7 @@ static int read_user_input(std::string & user_input) {
 #ifdef WIN32
     printf(
         "\r%*s"
-        "\r\033[0m%s",
+        "\r" LOG_COL_DEFAULT "%s",
         get_terminal_width(), " ", prompt_prefix);
 
     std::getline(std::cin, user_input);
@@ -855,7 +1008,7 @@ static int generate_response(LlamaData & llama_data, const std::string & prompt,
                              const bool stdout_a_terminal) {
     // Set response color
     if (stdout_a_terminal) {
-        printf("\033[33m");
+        printf(LOG_COL_YELLOW);
     }
 
     if (generate(llama_data, prompt, response)) {
@@ -864,13 +1017,13 @@ static int generate_response(LlamaData & llama_data, const std::string & prompt,
     }
 
     // End response with color reset and newline
-    printf("\n%s", stdout_a_terminal ? "\033[0m" : "");
+    printf("\n%s", stdout_a_terminal ? LOG_COL_DEFAULT : "");
     return 0;
 }
 
 // Helper function to apply the chat template and handle errors
-static int apply_chat_template_with_error_handling(LlamaData & llama_data, const bool append, int & output_length) {
-    const int new_len = apply_chat_template(llama_data, append);
+static int apply_chat_template_with_error_handling(const common_chat_template & tmpl, LlamaData & llama_data, const bool append, int & output_length, bool use_jinja) {
+    const int new_len = apply_chat_template(tmpl, llama_data, append, use_jinja);
     if (new_len < 0) {
         printe("failed to apply the chat template\n");
         return -1;
@@ -929,9 +1082,11 @@ static int get_user_input(std::string & user_input, const std::string & user) {
 }
 
 // Main chat loop function
-static int chat_loop(LlamaData & llama_data, const std::string & user) {
+static int chat_loop(LlamaData & llama_data, const std::string & user, bool use_jinja) {
     int prev_len = 0;
     llama_data.fmtted.resize(llama_n_ctx(llama_data.context.get()));
+    auto chat_templates = common_chat_templates_from_model(llama_data.model.get(), "");
+    GGML_ASSERT(chat_templates.template_default);
     static const bool stdout_a_terminal = is_stdout_a_terminal();
     while (true) {
         // Get user input
@@ -942,7 +1097,7 @@ static int chat_loop(LlamaData & llama_data, const std::string & user) {
 
         add_message("user", user.empty() ? user_input : user, llama_data);
         int new_len;
-        if (apply_chat_template_with_error_handling(llama_data, true, new_len) < 0) {
+        if (apply_chat_template_with_error_handling(*chat_templates.template_default, llama_data, true, new_len, use_jinja) < 0) {
             return 1;
         }
 
@@ -957,7 +1112,7 @@ static int chat_loop(LlamaData & llama_data, const std::string & user) {
         }
 
         add_message("assistant", response, llama_data);
-        if (apply_chat_template_with_error_handling(llama_data, false, prev_len) < 0) {
+        if (apply_chat_template_with_error_handling(*chat_templates.template_default, llama_data, false, prev_len, use_jinja) < 0) {
             return 1;
         }
     }
@@ -1017,7 +1172,7 @@ int main(int argc, const char ** argv) {
         return 1;
     }
 
-    if (chat_loop(llama_data, opt.user)) {
+    if (chat_loop(llama_data, opt.user, opt.use_jinja)) {
         return 1;
     }