@fugood/llama.node 0.4.7 → 0.5.0

package/src/llama.cpp/tools/mtmd/mtmd.h

@@ -39,6 +39,7 @@
 #    define MTMD_API
 #endif
 
+// deprecated marker, use mtmd_default_marker() instead
 #define MTMD_DEFAULT_IMAGE_MARKER "<__image__>"
 
 #ifdef __cplusplus
@@ -48,6 +49,7 @@ extern "C" {
 enum mtmd_input_chunk_type {
     MTMD_INPUT_CHUNK_TYPE_TEXT,
     MTMD_INPUT_CHUNK_TYPE_IMAGE,
+    MTMD_INPUT_CHUNK_TYPE_AUDIO,
 };
 
 // opaque types
@@ -79,9 +81,12 @@ struct mtmd_context_params {
     bool print_timings;
     int n_threads;
     enum ggml_log_level verbosity;
-    const char * image_marker;
+    const char * image_marker; // deprecated, use media_marker instead
+    const char * media_marker;
 };
 
+MTMD_API const char * mtmd_default_marker(void);
+
 MTMD_API struct mtmd_context_params mtmd_context_params_default(void);
 
 // initialize the mtmd context
@@ -98,18 +103,28 @@ MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx);
 // whether the current model use M-RoPE for llama_decode
 MTMD_API bool mtmd_decode_use_mrope(mtmd_context * ctx);
 
+// whether the current model supports vision input
+MTMD_API bool mtmd_support_vision(mtmd_context * ctx);
+
+// whether the current model supports audio input
+MTMD_API bool mtmd_support_audio(mtmd_context * ctx);
 
 // mtmd_bitmap
 //
-// length of data must be nx * ny * 3
-// the data is in RGBRGBRGB... format
-MTMD_API mtmd_bitmap *         mtmd_bitmap_init    (uint32_t nx,
-                                                    uint32_t ny,
-                                                    const unsigned char * data);
-MTMD_API uint32_t              mtmd_bitmap_get_nx  (const mtmd_bitmap * bitmap);
-MTMD_API uint32_t              mtmd_bitmap_get_ny  (const mtmd_bitmap * bitmap);
-MTMD_API const unsigned char * mtmd_bitmap_get_data(const mtmd_bitmap * bitmap);
-MTMD_API void                  mtmd_bitmap_free    (mtmd_bitmap * bitmap);
+// if bitmap is image:
+//     length of data must be nx * ny * 3
+//     the data is in RGBRGBRGB... format
+// if bitmap is audio:
+//     length of data must be n_samples * sizeof(float)
+//     the data is in float format (PCM F32)
+MTMD_API mtmd_bitmap *         mtmd_bitmap_init           (uint32_t nx, uint32_t ny, const unsigned char * data);
+MTMD_API mtmd_bitmap *         mtmd_bitmap_init_from_audio(size_t n_samples,         const float         * data);
+MTMD_API uint32_t              mtmd_bitmap_get_nx     (const mtmd_bitmap * bitmap);
+MTMD_API uint32_t              mtmd_bitmap_get_ny     (const mtmd_bitmap * bitmap);
+MTMD_API const unsigned char * mtmd_bitmap_get_data   (const mtmd_bitmap * bitmap);
+MTMD_API size_t                mtmd_bitmap_get_n_bytes(const mtmd_bitmap * bitmap);
+MTMD_API bool                  mtmd_bitmap_is_audio   (const mtmd_bitmap * bitmap);
+MTMD_API void                  mtmd_bitmap_free       (mtmd_bitmap * bitmap);
 // bitmap ID is optional, but useful for KV cache tracking
 // these getters/setters are dedicated functions, so you can for example calculate the hash of the image based on mtmd_bitmap_get_data()
 MTMD_API const char * mtmd_bitmap_get_id(const mtmd_bitmap * bitmap);
@@ -132,6 +147,11 @@ MTMD_API void                     mtmd_input_chunks_free(mtmd_input_chunks * chu
 MTMD_API enum mtmd_input_chunk_type mtmd_input_chunk_get_type        (const mtmd_input_chunk * chunk);
 MTMD_API const llama_token *        mtmd_input_chunk_get_tokens_text (const mtmd_input_chunk * chunk, size_t * n_tokens_output);
 MTMD_API const mtmd_image_tokens *  mtmd_input_chunk_get_tokens_image(const mtmd_input_chunk * chunk);
+MTMD_API size_t                     mtmd_input_chunk_get_n_tokens    (const mtmd_input_chunk * chunk);
+// returns nullptr for ID on text chunk
+MTMD_API const char *               mtmd_input_chunk_get_id          (const mtmd_input_chunk * chunk);
+// number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
+MTMD_API llama_pos                  mtmd_input_chunk_get_n_pos       (const mtmd_input_chunk * chunk);
 
 // in case you want to use custom logic to handle the chunk (i.e. KV cache management)
 // you can move the chunk ownership to your own code by copying it
@@ -144,27 +164,28 @@ MTMD_API void               mtmd_input_chunk_free(mtmd_input_chunk * chunk);
 //
 // the instance will be constructed via mtmd_tokenize()
 // it will be freed along with mtmd_input_chunk
-MTMD_API size_t       mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens);
+MTMD_API size_t       mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens); // TODO: deprecate
 MTMD_API size_t       mtmd_image_tokens_get_nx      (const mtmd_image_tokens * image_tokens);
 MTMD_API size_t       mtmd_image_tokens_get_ny      (const mtmd_image_tokens * image_tokens);
-MTMD_API const char * mtmd_image_tokens_get_id      (const mtmd_image_tokens * image_tokens);
+MTMD_API const char * mtmd_image_tokens_get_id      (const mtmd_image_tokens * image_tokens); // TODO: deprecate
 // number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
-MTMD_API llama_pos    mtmd_image_tokens_get_n_pos   (const mtmd_image_tokens * image_tokens);
+MTMD_API llama_pos    mtmd_image_tokens_get_n_pos   (const mtmd_image_tokens * image_tokens); // TODO: deprecate
 
-// tokenize an input text prompt and an image
-// the prompt must have the input image marker (default: "<__image__>") in it
-// the marker will be replaced with the image tokens
+// tokenize an input text prompt and a list of bitmaps (images/audio)
+// the prompt must have the input image marker (default: "<__media__>") in it
+// the default marker is defined by mtmd_default_marker()
+// the marker will be replaced with the image/audio chunk
 // for example:
-//   "here is an image: <__image__>\ndescribe it in detail."
+//   "here is an image: <__media__>\ndescribe it in detail."
 //   this will gives 3 chunks:
 //   1. "here is an image: <start_of_image>"
-//   2. (image tokens)
+//   2. (image/audio tokens)
 //   3. "<end_of_image>\ndescribe it in detail."
-// number of bitmaps must be equal to the number of image markers in the prompt
+// number of bitmaps must be equal to the number of markers in the prompt
 // this function is thread-safe (shared ctx)
 // return values:
 //   0 on success
-//   1 on number of images not matching the number of markers
+//   1 on number of bitmaps not matching the number of markers
 //   2 on image preprocessing error
 MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
                                mtmd_input_chunks * output,
@@ -173,9 +194,14 @@ MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
                                size_t n_bitmaps);
 
 // returns 0 on success
+// TODO: deprecate
 MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
                              const mtmd_image_tokens * image_tokens);
 
+// returns 0 on success
+MTMD_API int32_t mtmd_encode_chunk(mtmd_context * ctx,
+                                   const mtmd_input_chunk * chunk);
+
 // get output embeddings from the last encode pass
 MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
 
@@ -189,12 +215,16 @@ MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
 //
 
 // helper function to construct a mtmd_bitmap from a file
+// it calls mtmd_helper_bitmap_init_from_buf() internally
 // returns nullptr on failure
 // this function is thread-safe
 MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_file(const char * fname);
 
 // helper function to construct a mtmd_bitmap from a buffer containing a file
-// the file content must be an image in format supported by stb_image (jpg, png, bmp, gif, etc.)
+// supported formats:
+//     image: formats supported by stb_image: jpg, png, bmp, gif, etc.
+//     audio: formats supported by miniaudio: wav, mp3, flac
+// note: audio files will be auto-detected based on magic bytes
 // returns nullptr on failure
 // this function is thread-safe
 MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(const unsigned char * buf, size_t len);
@@ -293,6 +323,7 @@ struct bitmap {
     uint32_t nx() { return mtmd_bitmap_get_nx(ptr.get()); }
     uint32_t ny() { return mtmd_bitmap_get_ny(ptr.get()); }
     const unsigned char * data() { return mtmd_bitmap_get_data(ptr.get()); }
+    size_t n_bytes() { return mtmd_bitmap_get_n_bytes(ptr.get()); }
     std::string id() { return mtmd_bitmap_get_id(ptr.get()); }
     void set_id(const char * id) { mtmd_bitmap_set_id(ptr.get(), id); }
 };

package/src/llama.cpp/tools/run/run.cpp

@@ -936,7 +936,7 @@ static int apply_chat_template(const struct common_chat_templates * tmpls, Llama
 // Function to tokenize the prompt
 static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt,
                            std::vector<llama_token> & prompt_tokens, const LlamaData & llama_data) {
-    const bool is_first = llama_kv_self_used_cells(llama_data.context.get()) == 0;
+    const bool is_first = llama_kv_self_seq_pos_max(llama_data.context.get(), 0) == 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);
@@ -952,7 +952,7 @@ static int tokenize_prompt(const llama_vocab * vocab, const std::string & prompt
 // Check if we have enough space in the context to evaluate this batch
 static int check_context_size(const llama_context_ptr & ctx, const llama_batch & batch) {
     const int n_ctx      = llama_n_ctx(ctx.get());
-    const int n_ctx_used = llama_kv_self_used_cells(ctx.get());
+    const int n_ctx_used = llama_kv_self_seq_pos_max(ctx.get(), 0);
     if (n_ctx_used + batch.n_tokens > n_ctx) {
         printf(LOG_COL_DEFAULT "\n");
         printe("context size exceeded\n");

package/src/llama.cpp/tools/server/server.cpp

@@ -951,7 +951,7 @@ struct server_task_result_cmpl_partial : server_task_result {
     }
 
     json to_json_oaicompat_chat() {
-        bool first = n_decoded == 0;
+        bool first = n_decoded == 1;
         std::time_t t = std::time(0);
         json choices;
 
@@ -962,15 +962,18 @@ struct server_task_result_cmpl_partial : server_task_result {
                                             {"delta", json{{"role", "assistant"}}}}});
             } else {
                 // We have to send this as two updates to conform to openai behavior
+                // initial_ret is the role message for stream=True
                 json initial_ret = json{{"choices", json::array({json{
                                         {"finish_reason", nullptr},
                                         {"index", 0},
                                         {"delta", json{
-                                            {"role", "assistant"}
+                                            {"role", "assistant"},
+                                            {"content", ""}
                                         }}}})},
                             {"created", t},
                             {"id", oaicompat_cmpl_id},
                             {"model", oaicompat_model},
+                            {"system_fingerprint", build_info},
                             {"object", "chat.completion.chunk"}};
 
                 json second_ret = json{
@@ -982,8 +985,19 @@ struct server_task_result_cmpl_partial : server_task_result {
                             {"created", t},
                             {"id", oaicompat_cmpl_id},
                             {"model", oaicompat_model},
+                            {"system_fingerprint", build_info},
                             {"object", "chat.completion.chunk"}};
 
+                if (prob_output.probs.size() > 0) {
+                    second_ret["choices"][0]["logprobs"] = json{
+                        {"content", completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs)},
+                    };
+                }
+
+                if (timings.prompt_n >= 0) {
+                    second_ret.push_back({"timings", timings.to_json()});
+                }
+
                 return std::vector<json>({initial_ret, second_ret});
             }
         } else {
@@ -1137,9 +1151,6 @@ struct server_task_result_metrics : server_task_result {
     int n_tasks_deferred;
     int64_t t_start;
 
-    int32_t kv_cache_tokens_count;
-    int32_t kv_cache_used_cells;
-
     // TODO: somehow reuse server_metrics in the future, instead of duplicating the fields
     uint64_t n_prompt_tokens_processed_total = 0;
     uint64_t t_prompt_processing_total       = 0;
@@ -1179,9 +1190,6 @@ struct server_task_result_metrics : server_task_result {
             { "n_decode_total",                  n_decode_total },
             { "n_busy_slots_total",              n_busy_slots_total },
 
-            { "kv_cache_tokens_count",           kv_cache_tokens_count },
-            { "kv_cache_used_cells",             kv_cache_used_cells },
-
             { "slots",                           slots_data },
         };
     }
@@ -1883,6 +1891,7 @@ struct server_context {
     float slot_prompt_similarity = 0.0f;
 
     common_chat_templates_ptr chat_templates;
+    oaicompat_parser_options  oai_parser_opt;
 
     ~server_context() {
         mtmd_free(mctx);
@@ -2004,6 +2013,23 @@ struct server_context {
             }
         }
 
+        if (!llama_kv_self_can_shift(ctx)) {
+            if (params_base.ctx_shift) {
+                params_base.ctx_shift = false;
+                SRV_WRN("%s\n", "ctx_shift is not supported by this context, it will be disabled");
+            }
+
+            if (params_base.n_cache_reuse) {
+                params_base.n_cache_reuse = 0;
+                SRV_WRN("%s\n", "cache_reuse is not supported by this context, it will be disabled");
+            }
+
+            if (!params_base.speculative.model.path.empty()) {
+                SRV_ERR("%s\n", "err: speculative decode is not supported by this context");
+                return false;
+            }
+        }
+
         return true;
     }
 
@@ -2061,6 +2087,15 @@ struct server_context {
         }
 
         metrics.init();
+
+        oai_parser_opt = {
+            /* use_jinja             */ params_base.use_jinja,
+            /* prefill_assistant     */ params_base.prefill_assistant,
+            /* reasoning_format      */ params_base.reasoning_format,
+            /* common_chat_templates */ chat_templates.get(),
+            /* allow_image           */ mctx ? mtmd_support_vision(mctx) : false,
+            /* allow_audio           */ mctx ? mtmd_support_audio (mctx) : false,
+        };
     }
 
     server_slot * get_slot_by_id(int id) {
@@ -2754,9 +2789,6 @@ struct server_context {
                     res->n_tasks_deferred    = queue_tasks.queue_tasks_deferred.size();
                     res->t_start             = metrics.t_start;
 
-                    res->kv_cache_tokens_count = llama_kv_self_n_tokens(ctx);
-                    res->kv_cache_used_cells   = llama_kv_self_used_cells(ctx);
-
                     res->n_prompt_tokens_processed_total = metrics.n_prompt_tokens_processed_total;
                     res->t_prompt_processing_total       = metrics.t_prompt_processing_total;
                     res->n_tokens_predicted_total        = metrics.n_tokens_predicted_total;
@@ -3181,7 +3213,15 @@ struct server_context {
                                 // if we don't cache the prompt, we have to remove the entire KV cache
                                 llama_kv_self_seq_rm(ctx, slot.id, 0, -1);
                                 slot.n_past = 0;
-                                slot.cache_tokens.clear();
+                                slot.cache_tokens.clear(); // TODO: not needed, will be cleared later via "keep_first()"
+                            }
+
+                            if (slot.n_past > 0 && slot.n_past < (int) slot.cache_tokens.size()) {
+                                if (llama_kv_self_seq_pos_min(ctx, slot.id) > 0) {
+                                    SLT_WRN(slot, "forcing full prompt re-processing due to lack of cache data (likely due to SWA, see %s)\n",
+                                            "https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055");
+                                    slot.n_past = 0;
+                                }
                             }
                         }
 
@@ -3311,6 +3351,37 @@ struct server_context {
             common_set_adapter_lora(ctx, slot_batched->lora);
         }
 
+        const bool do_encode = (params_base.embedding || params_base.reranking);
+
+        // pad the batch so that batch.n_tokens >= n_slots
+        // TODO: temporary workaround for https://github.com/ggml-org/llama.cpp/issues/13689
+        if (do_encode) {
+            const int n_slots = slots.size();
+
+            if (batch.n_tokens < n_slots) {
+                std::set<llama_seq_id> seq_ids;
+                for (int j = 0; j < batch.n_tokens; ++j) {
+                    seq_ids.insert(batch.seq_id[j][0]);
+                }
+
+                // find unused sequence id
+                llama_seq_id seq_id = -1;
+                for (int i = 0; i < n_slots; ++i) {
+                    if (seq_ids.find(i) == seq_ids.end()) {
+                        seq_id = i;
+                    }
+                }
+
+                const int n_add = n_slots - batch.n_tokens;
+
+                SRV_WRN("adding %d dummy tokens to the batch, seq_id = %d\n", n_add, seq_id);
+
+                for (int j = 0; j < n_add; ++j) {
+                    common_batch_add(batch, 0, j, { seq_id }, false);
+                }
+            }
+        }
+
         // process the created batch of tokens
         for (int32_t i = 0; i < batch.n_tokens; i += n_batch) {
             const int32_t n_tokens = std::min(n_batch, batch.n_tokens - i);
@@ -3327,7 +3398,7 @@ struct server_context {
 
             int ret = 0;
 
-            if (params_base.embedding || params_base.reranking) {
+            if (do_encode) {
                 ret = llama_encode(ctx, batch_view);
             } else {
                 ret = llama_decode(ctx, batch_view);
@@ -3336,14 +3407,29 @@ struct server_context {
             metrics.on_decoded(slots);
 
             if (ret != 0) {
-                if (n_batch == 1 || ret < 0) {
-                    // if you get here, it means the KV cache is full - try increasing it via the context size
-                    SRV_ERR("failed to decode the batch: KV cache is full - try increasing it via the context size, i = %d, n_batch = %d, ret = %d\n", i, n_batch, ret);
-                    for (auto & slot : slots) {
-                        slot.release();
-                        send_error(slot, "Input prompt is too big compared to KV size. Please try increasing KV size.");
+                {
+                    std::string err;
+
+                    if (n_batch == 1 && ret == 1) {
+                        err = "Context size has been exceeded.";
+                    }
+
+                    if (ret == -1) {
+                        err = "Invalid input batch.";
+                    }
+
+                    if (ret < -1) {
+                        err = "Compute error.";
+                    }
+
+                    if (!err.empty()) {
+                        SRV_ERR("%s, i = %d, n_batch = %d, ret = %d\n", err.c_str(), i, n_batch, ret);
+                        for (auto & slot : slots) {
+                            slot.release();
+                            send_error(slot, err);
+                        }
+                        break;
                     }
-                    break; // break loop of n_batch
                 }
 
                 // retry with half the batch size to try to find a free slot in the KV cache
@@ -3677,6 +3763,7 @@ int main(int argc, char ** argv) {
             "/health",
             "/models",
             "/v1/models",
+            "/api/tags"
         };
 
         // If API key is not set, skip validation
@@ -3715,7 +3802,7 @@ int main(int argc, char ** argv) {
             if (req.path == "/" || tmp.back() == "html") {
                 res.set_content(reinterpret_cast<const char*>(loading_html), loading_html_len, "text/html; charset=utf-8");
                 res.status = 503;
-            } else if (req.path == "/models" || req.path == "/v1/models") {
+            } else if (req.path == "/models" || req.path == "/v1/models" || req.path == "/api/tags") {
                 // allow the models endpoint to be accessed during loading
                 return true;
             } else {
@@ -3858,14 +3945,6 @@ int main(int argc, char ** argv) {
                     {"name",  "predicted_tokens_seconds"},
                     {"help",  "Average generation throughput in tokens/s."},
                     {"value",  res_metrics->n_tokens_predicted ? 1.e3 / res_metrics->t_tokens_generation * res_metrics->n_tokens_predicted : 0.}
-            },{
-                    {"name",  "kv_cache_usage_ratio"},
-                    {"help",  "KV-cache usage. 1 means 100 percent usage."},
-                    {"value",  1. * res_metrics->kv_cache_used_cells / params.n_ctx}
-            },{
-                    {"name",  "kv_cache_tokens"},
-                    {"help",  "KV-cache tokens."},
-                    {"value",  (uint64_t) res_metrics->kv_cache_tokens_count}
             },{
                     {"name",  "requests_processing"},
                     {"help",  "Number of requests processing."},
@@ -4023,7 +4102,10 @@ int main(int argc, char ** argv) {
             { "default_generation_settings", ctx_server.default_generation_settings_for_props },
             { "total_slots",                 ctx_server.params_base.n_parallel },
             { "model_path",                  ctx_server.params_base.model.path },
-            { "modalities",                  json{{"vision", ctx_server.mctx != nullptr}} }, // TODO: add more in the future
+            { "modalities",                  json{
+                {"vision", ctx_server.oai_parser_opt.allow_image},
+                {"audio",  ctx_server.oai_parser_opt.allow_audio},
+            } },
             { "chat_template",               common_chat_templates_source(ctx_server.chat_templates.get()) },
             { "bos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_bos(ctx_server.vocab), /* special= */ true)},
             { "eos_token",                   common_token_to_piece(ctx_server.ctx, llama_vocab_eos(ctx_server.vocab), /* special= */ true)},
@@ -4061,6 +4143,19 @@ int main(int argc, char ** argv) {
                     { "llama.context_length", ctx_server.slots.back().n_ctx, },
                 }
             },
+            {"modelfile", ""},
+            {"parameters", ""},
+            {"template", common_chat_templates_source(ctx_server.chat_templates.get())},
+            {"details", {
+                {"parent_model", ""},
+                {"format", "gguf"},
+                {"family", ""},
+                {"families", {""}},
+                {"parameter_size", ""},
+                {"quantization_level", ""}
+            }},
+            {"model_info", ""},
+            {"capabilities", {"completion"}}
         };
 
         res_ok(res, data);
@@ -4101,10 +4196,10 @@ int main(int argc, char ** argv) {
                 for (auto & file : files) {
                     mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_buf(file.data(), file.size()));
                     if (!bmp.ptr) {
-                        throw std::runtime_error("Failed to load image");
+                        throw std::runtime_error("Failed to load image or audio file");
                     }
                     // calculate bitmap hash (for KV caching)
-                    std::string hash = fnv_hash(bmp.data(), bmp.nx()*bmp.ny()*3);
+                    std::string hash = fnv_hash(bmp.data(), bmp.n_bytes());
                     bmp.set_id(hash.c_str());
                     bitmaps.entries.push_back(std::move(bmp));
                 }
@@ -4336,7 +4431,7 @@ int main(int argc, char ** argv) {
             OAICOMPAT_TYPE_NONE); // infill is not OAI compatible
     };
 
-    const auto handle_chat_completions = [&ctx_server, &params, &res_error, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_chat_completions = [&ctx_server, &res_error, &handle_completions_impl](const httplib::Request & req, httplib::Response & res) {
         LOG_DBG("request: %s\n", req.body.c_str());
         if (ctx_server.params_base.embedding) {
             res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
@@ -4345,13 +4440,9 @@ int main(int argc, char ** argv) {
 
         auto body = json::parse(req.body);
         std::vector<raw_buffer> files;
-        json data = oaicompat_completion_params_parse(
+        json data = oaicompat_chat_params_parse(
             body,
-            params.use_jinja,
-            params.prefill_assistant,
-            params.reasoning_format,
-            ctx_server.chat_templates.get(),
-            ctx_server.mctx,
+            ctx_server.oai_parser_opt,
             files);
 
         handle_completions_impl(
@@ -4364,16 +4455,12 @@ int main(int argc, char ** argv) {
     };
 
     // same with handle_chat_completions, but without inference part
-    const auto handle_apply_template = [&ctx_server, &params, &res_ok](const httplib::Request & req, httplib::Response & res) {
+    const auto handle_apply_template = [&ctx_server, &res_ok](const httplib::Request & req, httplib::Response & res) {
         auto body = json::parse(req.body);
         std::vector<raw_buffer> files; // dummy, unused
-        json data = oaicompat_completion_params_parse(
+        json data = oaicompat_chat_params_parse(
             body,
-            params.use_jinja,
-            params.prefill_assistant,
-            params.reasoning_format,
-            ctx_server.chat_templates.get(),
-            ctx_server.mctx,
+            ctx_server.oai_parser_opt,
             files);
         res_ok(res, {{ "prompt", std::move(data.at("prompt")) }});
     };
@@ -4386,6 +4473,28 @@ int main(int argc, char ** argv) {
         }
 
         json models = {
+            {"models", {
+                {
+                    {"name", params.model_alias.empty() ? params.model.path : params.model_alias},
+                    {"model", params.model_alias.empty() ? params.model.path : params.model_alias},
+                    {"modified_at", ""},
+                    {"size", ""},
+                    {"digest", ""}, // dummy value, llama.cpp does not support managing model file's hash
+                    {"type", "model"},
+                    {"description", ""},
+                    {"tags", {""}},
+                    {"capabilities", {"completion"}},
+                    {"parameters", ""},
+                    {"details", {
+                        {"parent_model", ""},
+                        {"format", "gguf"},
+                        {"family", ""},
+                        {"families", {""}},
+                        {"parameter_size", ""},
+                        {"quantization_level", ""}
+                    }}
+                }
+            }},
             {"object", "list"},
             {"data", {
                 {
@@ -4395,7 +4504,7 @@ int main(int argc, char ** argv) {
                     {"owned_by", "llamacpp"},
                     {"meta",     model_meta},
                 },
-             }}
+            }}
         };
 
         res_ok(res, models);
@@ -4723,11 +4832,13 @@ int main(int argc, char ** argv) {
     svr->Post("/api/show",            handle_api_show);
     svr->Get ("/models",              handle_models); // public endpoint (no API key check)
     svr->Get ("/v1/models",           handle_models); // public endpoint (no API key check)
+    svr->Get ("/api/tags",            handle_models); // ollama specific endpoint. public endpoint (no API key check)
     svr->Post("/completion",          handle_completions); // legacy
     svr->Post("/completions",         handle_completions);
     svr->Post("/v1/completions",      handle_completions_oai);
     svr->Post("/chat/completions",    handle_chat_completions);
     svr->Post("/v1/chat/completions", handle_chat_completions);
+    svr->Post("/api/chat",            handle_chat_completions); // ollama specific endpoint
     svr->Post("/infill",              handle_infill);
     svr->Post("/embedding",           handle_embeddings); // legacy
     svr->Post("/embeddings",          handle_embeddings);