Merge branch 'upstream' into concedo_experimental

# Conflicts:
#	README.md
#	docs/build-s390x.md
#	examples/llama.vim
#	ggml/src/ggml-cann/aclnn_ops.cpp
#	ggml/src/ggml-cann/common.h
#	scripts/compare-llama-bench.py
#	src/CMakeLists.txt
#	tests/test-backend-ops.cpp
#	tools/llama-bench/README.md
#	tools/llama-bench/llama-bench.cpp
#	tools/server/README.md

tools/server/bench/bench.py

@@ -274,7 +274,6 @@ def start_server_background(args):
     server_args.extend(['--batch-size', args.batch_size])
     server_args.extend(['--ubatch-size', args.ubatch_size])
     server_args.extend(['--n-predict', args.max_tokens * 2])
-    server_args.extend(['--defrag-thold', "0.1"])
     server_args.append('--cont-batching')
     server_args.append('--metrics')
     server_args.append('--flash-attn')

tools/server/server.cpp

@@ -4309,6 +4309,7 @@ int main(int argc, char ** argv) {
     };
 
     const auto handle_api_show = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
+        bool has_mtmd = ctx_server.mctx != nullptr;
         json data = {
             {
                 "template", common_chat_templates_source(ctx_server.chat_templates.get()),
@@ -4330,7 +4331,7 @@ int main(int argc, char ** argv) {
                 {"quantization_level", ""}
             }},
             {"model_info", ""},
-            {"capabilities", {"completion"}}
+            {"capabilities", has_mtmd ? json({"completion","multimodal"}) : json({"completion"})}
         };
 
         res_ok(res, data);
@@ -4356,56 +4357,15 @@ int main(int argc, char ** argv) {
             // TODO: this log can become very long, put it behind a flag or think about a more compact format
             //SRV_DBG("Prompt: %s\n", prompt.is_string() ? prompt.get<std::string>().c_str() : prompt.dump(2).c_str());
 
-            // process files
-            mtmd::bitmaps bitmaps;
-            const bool has_mtmd = ctx_server.mctx != nullptr;
-            {
-                if (!has_mtmd && !files.empty()) {
-                    throw std::runtime_error("This server does not support multimodal");
-                }
-                for (auto & file : files) {
-                    mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_buf(ctx_server.mctx, file.data(), file.size()));
-                    if (!bmp.ptr) {
-                        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.n_bytes());
-                    bmp.set_id(hash.c_str());
-                    bitmaps.entries.push_back(std::move(bmp));
-                }
-            }
-
             // process prompt
             std::vector<server_tokens> inputs;
 
-            if (oaicompat && has_mtmd) {
-                // multimodal
-                std::string prompt_str = prompt.get<std::string>();
-                mtmd_input_text inp_txt = {
-                    prompt_str.c_str(),
-                    /* add_special */   true,
-                    /* parse_special */ true,
-                };
-                mtmd::input_chunks chunks(mtmd_input_chunks_init());
-                auto bitmaps_c_ptr = bitmaps.c_ptr();
-                int32_t tokenized = mtmd_tokenize(ctx_server.mctx,
-                                                    chunks.ptr.get(),
-                                                    &inp_txt,
-                                                    bitmaps_c_ptr.data(),
-                                                    bitmaps_c_ptr.size());
-                if (tokenized != 0) {
-                    throw std::runtime_error("Failed to tokenize prompt");
-                }
-
-                server_tokens tmp(chunks, true);
-                inputs.push_back(std::move(tmp));
+            if (oaicompat && ctx_server.mctx != nullptr) {
+                // This is the case used by OAI compatible chat path with MTMD. TODO It can be moved to the path below.
+                inputs.push_back(process_mtmd_prompt(ctx_server.mctx, prompt.get<std::string>(), files));
             } else {
-                // non-multimodal version
-                auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
-                for (auto & p : tokenized_prompts) {
-                    auto tmp = server_tokens(p, ctx_server.mctx != nullptr);
-                    inputs.push_back(std::move(tmp));
-                }
+                // Everything else, including multimodal completions.
+                inputs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true);
             }
 
             tasks.reserve(inputs.size());
@@ -4574,7 +4534,7 @@ int main(int argc, char ** argv) {
         data["input_extra"] = input_extra; // default to empty array if it's not exist
 
         std::string prompt = json_value(data, "prompt", std::string());
-        std::vector<llama_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, false, true);
+        std::vector<server_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, false, true);
         SRV_DBG("creating infill tasks, n_prompts = %d\n", (int) tokenized_prompts.size());
         data["prompt"] = format_infill(
             ctx_server.vocab,
@@ -4585,7 +4545,7 @@ int main(int argc, char ** argv) {
             ctx_server.params_base.n_predict,
             ctx_server.slots[0].n_ctx, // TODO: there should be a better way
             ctx_server.params_base.spm_infill,
-            tokenized_prompts[0]
+            tokenized_prompts[0].get_text_tokens() // TODO: this could maybe be multimodal.
         );
 
         std::vector<raw_buffer> files; // dummy
@@ -4634,7 +4594,7 @@ int main(int argc, char ** argv) {
         if (current_state == SERVER_STATE_READY) {
             model_meta = ctx_server.model_meta();
         }
-
+        bool has_mtmd = ctx_server.mctx != nullptr;
         json models = {
             {"models", {
                 {
@@ -4646,7 +4606,7 @@ int main(int argc, char ** argv) {
                     {"type", "model"},
                     {"description", ""},
                     {"tags", {""}},
-                    {"capabilities", {"completion"}},
+                    {"capabilities", has_mtmd ? json({"completion","multimodal"}) : json({"completion"})},
                     {"parameters", ""},
                     {"details", {
                         {"parent_model", ""},
@@ -4763,7 +4723,7 @@ int main(int argc, char ** argv) {
             }
         }
 
-        auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, prompt, true, true);
+        auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true);
         for (const auto & tokens : tokenized_prompts) {
             // this check is necessary for models that do not add BOS token to the input
             if (tokens.empty()) {
@@ -4791,7 +4751,7 @@ int main(int argc, char ** argv) {
 
                 task.id            = ctx_server.queue_tasks.get_new_id();
                 task.index         = i;
-                task.prompt_tokens = server_tokens(tokenized_prompts[i], ctx_server.mctx != nullptr);
+                task.prompt_tokens = std::move(tokenized_prompts[i]);
 
                 // OAI-compat
                 task.params.oaicompat = oaicompat;
@@ -4878,7 +4838,10 @@ int main(int argc, char ** argv) {
             return;
         }
 
-        llama_tokens tokenized_query = tokenize_input_prompts(ctx_server.vocab, query, /* add_special */ false, true)[0];
+        std::vector<server_tokens> tokenized_queries = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, query, /* add_special */ false, true);
+        if (tokenized_queries.size() != 1) {
+            res_error(res, format_error_response("\"query\" must contain only a single prompt", ERROR_TYPE_INVALID_REQUEST));
+        }
 
         // create and queue the task
         json responses = json::array();
@@ -4886,14 +4849,14 @@ int main(int argc, char ** argv) {
         std::unordered_set<int> task_ids;
         {
             std::vector<server_task> tasks;
-            auto tokenized_docs = tokenize_input_prompts(ctx_server.vocab, documents, /* add_special */ false, true);
+            auto tokenized_docs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, documents, /* add_special */ false, true);
             tasks.reserve(tokenized_docs.size());
             for (size_t i = 0; i < tokenized_docs.size(); i++) {
-                auto tmp = format_rerank(ctx_server.vocab, tokenized_query, tokenized_docs[i]);
+                auto tmp = format_rerank(ctx_server.vocab, tokenized_queries[0], tokenized_docs[i]);
                 server_task task   = server_task(SERVER_TASK_TYPE_RERANK);
                 task.id            = ctx_server.queue_tasks.get_new_id();
                 task.index         = i;
-                task.prompt_tokens = server_tokens(tmp, ctx_server.mctx != nullptr);
+                task.prompt_tokens = std::move(tmp);
                 tasks.push_back(std::move(task));
             }
 

tools/server/tests/unit/test_completion.py

@@ -6,6 +6,8 @@ from utils import *
 
 server = ServerPreset.tinyllama2()
 
+JSON_MULTIMODAL_KEY = "multimodal_data"
+JSON_PROMPT_STRING_KEY = "prompt_string"
 
 @pytest.fixture(autouse=True)
 def create_server():
@@ -231,6 +233,28 @@ def test_nocache_long_input_prompt():
     })
     assert res.status_code == 400
 
+def test_json_prompt_no_mtmd():
+    global server
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "prompt": { JSON_PROMPT_STRING_KEY: "I believe the meaning of life is" },
+        "seed": 42,
+        "temperature": 1.0,
+        "cache_prompt": False,
+    })
+    assert res.status_code == 200
+
+def test_json_prompt_mtm_error_when_not_supported():
+    global server
+    server.start()
+    res = server.make_request("POST", "/completion", data={
+        "prompt": { JSON_PROMPT_STRING_KEY: "I believe the meaning of life is <__media__>", JSON_MULTIMODAL_KEY: "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=" },
+        "seed": 42,
+        "temperature": 1.0,
+        "cache_prompt": False,
+    })
+    # MTMD is disabled on this model, so this should fail.
+    assert res.status_code != 200
 
 def test_completion_with_tokens_input():
     global server
@@ -269,6 +293,20 @@ def test_completion_with_tokens_input():
     assert len(res.body) == 2
     assert res.body[0]["content"] == res.body[1]["content"]
 
+    # mixed JSON and tokens
+    res = server.make_request("POST", "/completion", data={
+        "prompt": [
+            tokens,
+            {
+                JSON_PROMPT_STRING_KEY: "I believe the meaning of life is",
+            },
+        ],
+    })
+    assert res.status_code == 200
+    assert type(res.body) == list
+    assert len(res.body) == 2
+    assert res.body[0]["content"] == res.body[1]["content"]
+
     # mixed string and tokens in one sequence
     res = server.make_request("POST", "/completion", data={
         "prompt": [1, 2, 3, 4, 5, 6, prompt_str, 7, 8, 9, 10, prompt_str],

tools/server/tests/unit/test_vision_api.py

@@ -10,21 +10,48 @@ IMG_URL_1 = "https://huggingface.co/ggml-org/tinygemma3-GGUF/resolve/main/test/9
 
 response = requests.get(IMG_URL_0)
 response.raise_for_status() # Raise an exception for bad status codes
-IMG_BASE64_0 = "data:image/png;base64," + base64.b64encode(response.content).decode("utf-8")
+IMG_BASE64_URI_0 = "data:image/png;base64," + base64.b64encode(response.content).decode("utf-8")
+IMG_BASE64_0 = base64.b64encode(response.content).decode("utf-8")
 
+response = requests.get(IMG_URL_1)
+response.raise_for_status() # Raise an exception for bad status codes
+IMG_BASE64_URI_1 = "data:image/png;base64," + base64.b64encode(response.content).decode("utf-8")
+IMG_BASE64_1 = base64.b64encode(response.content).decode("utf-8")
+
+JSON_MULTIMODAL_KEY = "multimodal_data"
+JSON_PROMPT_STRING_KEY = "prompt_string"
 
 @pytest.fixture(autouse=True)
 def create_server():
     global server
     server = ServerPreset.tinygemma3()
 
+def test_models_supports_multimodal_capability():
+    global server
+    server.start() # vision model may take longer to load due to download size
+    res = server.make_request("GET", "/models", data={})
+    assert res.status_code == 200
+    model_info = res.body["models"][0]
+    print(model_info)
+    assert "completion" in model_info["capabilities"]
+    assert "multimodal" in model_info["capabilities"]
+
+def test_v1_models_supports_multimodal_capability():
+    global server
+    server.start() # vision model may take longer to load due to download size
+    res = server.make_request("GET", "/v1/models", data={})
+    assert res.status_code == 200
+    model_info = res.body["models"][0]
+    print(model_info)
+    assert "completion" in model_info["capabilities"]
+    assert "multimodal" in model_info["capabilities"]
 
 @pytest.mark.parametrize(
     "prompt, image_url, success, re_content",
     [
         # test model is trained on CIFAR-10, but it's quite dumb due to small size
         ("What is this:\n", IMG_URL_0,                True, "(cat)+"),
-        ("What is this:\n", "IMG_BASE64_0",           True, "(cat)+"), # exceptional, so that we don't cog up the log
+        ("What is this:\n", "IMG_BASE64_URI_0",       True, "(cat)+"), # exceptional, so that we don't cog up the log
         ("What is this:\n", IMG_URL_1,                True, "(frog)+"),
         ("Test test\n",     IMG_URL_1,                True, "(frog)+"), # test invalidate cache
         ("What is this:\n", "malformed",              False, None),
@@ -36,8 +63,8 @@ def create_server():
 def test_vision_chat_completion(prompt, image_url, success, re_content):
     global server
     server.start(timeout_seconds=60) # vision model may take longer to load due to download size
-    if image_url == "IMG_BASE64_0":
-        image_url = IMG_BASE64_0
+    if image_url == "IMG_BASE64_URI_0":
+        image_url = IMG_BASE64_URI_0
     res = server.make_request("POST", "/chat/completions", data={
         "temperature": 0.0,
         "top_k": 1,
@@ -58,3 +85,61 @@ def test_vision_chat_completion(prompt, image_url, success, re_content):
     else:
         assert res.status_code != 200
 
+
+@pytest.mark.parametrize(
+    "prompt, image_data, success, re_content",
+    [
+        # test model is trained on CIFAR-10, but it's quite dumb due to small size
+        ("What is this: <__media__>\n", IMG_BASE64_0,           True, "(cat)+"),
+        ("What is this: <__media__>\n", IMG_BASE64_1,           True, "(frog)+"),
+        ("What is this: <__media__>\n", "malformed",            False, None), # non-image data
+        ("What is this:\n",             "",                     False, None), # empty string
+    ]
+)
+def test_vision_completion(prompt, image_data, success, re_content):
+    global server
+    server.start() # vision model may take longer to load due to download size
+    res = server.make_request("POST", "/completions", data={
+        "temperature": 0.0,
+        "top_k": 1,
+        "prompt": { JSON_PROMPT_STRING_KEY: prompt, JSON_MULTIMODAL_KEY: [ image_data ] },
+    })
+    if success:
+        assert res.status_code == 200
+        content = res.body["content"]
+        assert match_regex(re_content, content)
+    else:
+        assert res.status_code != 200
+
+
+@pytest.mark.parametrize(
+    "prompt, image_data, success",
+    [
+        # test model is trained on CIFAR-10, but it's quite dumb due to small size
+        ("What is this: <__media__>\n", IMG_BASE64_0,           True), # exceptional, so that we don't cog up the log
+        ("What is this: <__media__>\n", IMG_BASE64_1,           True),
+        ("What is this: <__media__>\n", "malformed",            False), # non-image data
+        ("What is this:\n",             "base64",               False), # non-image data
+    ]
+)
+def test_vision_embeddings(prompt, image_data, success):
+    global server
+    server.server_embeddings=True
+    server.n_batch=512
+    server.start() # vision model may take longer to load due to download size
+    res = server.make_request("POST", "/embeddings", data={
+        "content": [
+            { JSON_PROMPT_STRING_KEY: prompt, JSON_MULTIMODAL_KEY: [ image_data ] },
+            { JSON_PROMPT_STRING_KEY: prompt, JSON_MULTIMODAL_KEY: [ image_data ] },
+            { JSON_PROMPT_STRING_KEY: prompt, },
+        ],
+    })
+    if success:
+        assert res.status_code == 200
+        content = res.body
+        # Ensure embeddings are stable when multimodal.
+        assert content[0]['embedding'] == content[1]['embedding']
+        # Ensure embeddings without multimodal but same prompt do not match multimodal embeddings.
+        assert content[0]['embedding'] != content[2]['embedding']
+    else:
+        assert res.status_code != 200

tools/server/utils.hpp

@@ -123,6 +123,19 @@ static bool json_is_array_of_mixed_numbers_strings(const json & data) {
     return false;
 }
 
+// does array have any individual integers/tokens?
+static bool json_is_array_and_contains_numbers(const json & data) {
+    if (data.is_array()) {
+        for (const auto & e : data) {
+            if (e.is_number_integer()) {
+                return true;
+            }
+        }
+        return false;
+    }
+    return false;
+}
+
 // get value by path(key1 / key2)
 static json json_get_nested_values(const std::vector<std::string> & paths, const json & js) {
     json result = json::object();
@@ -186,48 +199,6 @@ static llama_tokens tokenize_mixed(const llama_vocab * vocab, const json & json_
     return prompt_tokens;
 }
 
-/**
- * break the input "prompt" object into multiple prompt if needed, then tokenize them
- * this supports these cases:
- * - "prompt": "string"
- * - "prompt": [12, 34, 56]
- * - "prompt": [12, 34, "string", 56, 78]
- * and multiple prompts (multi-tasks):
- * - "prompt": ["string1", "string2"]
- * - "prompt": ["string1", [12, 34, 56]]
- * - "prompt": [[12, 34, 56], [78, 90, 12]]
- * - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56]]
- */
-static std::vector<llama_tokens> tokenize_input_prompts(const llama_vocab * vocab, const json & json_prompt, bool add_special, bool parse_special) {
-    std::vector<llama_tokens> result;
-    if (json_prompt.is_string() || json_is_array_of_mixed_numbers_strings(json_prompt)) {
-        // string or mixed
-        result.push_back(tokenize_mixed(vocab, json_prompt, add_special, parse_special));
-    } else if (json_is_array_of_numbers(json_prompt)) {
-        // array of tokens
-        result.push_back(json_prompt.get<llama_tokens>());
-    } else if (json_prompt.is_array()) {
-        // array of prompts
-        result.reserve(json_prompt.size());
-        for (const auto & p : json_prompt) {
-            if (p.is_string() || json_is_array_of_mixed_numbers_strings(p)) {
-                result.push_back(tokenize_mixed(vocab, p, add_special, parse_special));
-            } else if (json_is_array_of_numbers(p)) {
-                // array of tokens
-                result.push_back(p.get<llama_tokens>());
-            } else {
-                throw std::runtime_error("element of \"prompt\" must be a string, an list of tokens, or a list of mixed strings & tokens");
-            }
-        }
-    } else {
-        throw std::runtime_error("\"prompt\" must be a string, an list of tokens, a list of mixed strings & tokens, or a list of prompts");
-    }
-    if (result.empty()) {
-        throw std::runtime_error("\"prompt\" must not be empty");
-    }
-    return result;
-}
-
 // return the last index of character that can form a valid string
 // if the last character is potentially cut in half, return the index before the cut
 // if validate_utf8(text) == text.size(), then the whole text is valid utf8
@@ -262,35 +233,6 @@ static size_t validate_utf8(const std::string& text) {
 // template utils
 //
 
-// format rerank task: [BOS]query[EOS][SEP]doc[EOS]
-static llama_tokens format_rerank(const struct llama_vocab * vocab, const llama_tokens & query, const llama_tokens & doc) {
-    llama_tokens result;
-
-    // Get EOS token - use SEP token as fallback if EOS is not available
-    llama_token eos_token = llama_vocab_eos(vocab);
-    if (eos_token == LLAMA_TOKEN_NULL) {
-        eos_token = llama_vocab_sep(vocab);
-    }
-
-    result.reserve(doc.size() + query.size() + 4);
-    if (llama_vocab_get_add_bos(vocab)) {
-        result.push_back(llama_vocab_bos(vocab));
-    }
-    result.insert(result.end(), query.begin(), query.end());
-    if (llama_vocab_get_add_eos(vocab)) {
-        result.push_back(eos_token);
-    }
-    if (llama_vocab_get_add_sep(vocab)) {
-        result.push_back(llama_vocab_sep(vocab));
-    }
-    result.insert(result.end(), doc.begin(), doc.end());
-    if (llama_vocab_get_add_eos(vocab)) {
-        result.push_back(eos_token);
-    }
-
-    return result;
-}
-
 // format infill task
 static llama_tokens format_infill(
         const llama_vocab * vocab,
@@ -1186,6 +1128,24 @@ public:
         }
     }
 
+    // appends server tokens, updates the media map. copies media chunks.
+    void push_back(server_tokens & tokens) {
+        size_t start_pos = size();
+        for (size_t i = 0; i < tokens.size(); i++) {
+            push_back(tokens[i]);
+        }
+        if (tokens.has_mtmd) {
+            // Assert if we are copying MTMD chunks to a server_tokens that does not have mtmd.
+            // We could also just check, but this will prevent silently dropping MTMD data.
+            GGML_ASSERT(has_mtmd);
+            for (auto it = tokens.map_pos_to_media.begin(); it != tokens.map_pos_to_media.end(); ) {
+                auto chunk = tokens.map_pos_to_media[it->first].get();
+                mtmd::input_chunk_ptr new_chunk(mtmd_input_chunk_copy(chunk));
+                map_pos_to_media[start_pos+it->first] = std::move(new_chunk);
+            }
+        }
+    }
+
     // for compatibility with context shift and prompt truncation
     void insert(const llama_tokens & inp_tokens) {
         GGML_ASSERT(!has_mtmd); // only allow this if mtmd is disabled
@@ -1356,3 +1316,137 @@ static std::string fnv_hash(const uint8_t * data, size_t len) {
     }
     return std::to_string(hash);
 }
+
+
+// format rerank task: [BOS]query[EOS][SEP]doc[EOS].
+static server_tokens format_rerank(const struct llama_vocab * vocab, server_tokens & query, server_tokens & doc) {
+    server_tokens result = {};
+
+    // Get EOS token - use SEP token as fallback if EOS is not available
+    llama_token eos_token = llama_vocab_eos(vocab);
+    if (eos_token == LLAMA_TOKEN_NULL) {
+        eos_token = llama_vocab_sep(vocab);
+    }
+    if (llama_vocab_get_add_bos(vocab)) {
+        result.push_back(llama_vocab_bos(vocab));
+    }
+    result.push_back(query);
+    if (llama_vocab_get_add_eos(vocab)) {
+        result.push_back(eos_token);
+    }
+    if (llama_vocab_get_add_sep(vocab)) {
+        result.push_back(llama_vocab_sep(vocab));
+    }
+    result.push_back(doc);
+    if (llama_vocab_get_add_eos(vocab)) {
+        result.push_back(eos_token);
+    }
+    return result;
+}
+
+
+static server_tokens process_mtmd_prompt(mtmd_context * mctx, std::string prompt, std::vector<raw_buffer> files) {
+    mtmd::bitmaps bitmaps;
+    for (auto & file : files) {
+        mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_buf(mctx, file.data(), file.size()));
+        if (!bmp.ptr) {
+            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.n_bytes());
+        bmp.set_id(hash.c_str());
+        bitmaps.entries.push_back(std::move(bmp));
+    }
+    // process prompt
+    std::vector<server_tokens> inputs;
+    // multimodal
+    mtmd_input_text inp_txt = {
+        prompt.c_str(),
+        /* add_special */   true,
+        /* parse_special */ true,
+    };
+    mtmd::input_chunks chunks(mtmd_input_chunks_init());
+    auto bitmaps_c_ptr = bitmaps.c_ptr();
+    int32_t tokenized = mtmd_tokenize(mctx,
+                                      chunks.ptr.get(),
+                                      &inp_txt,
+                                      bitmaps_c_ptr.data(),
+                                      bitmaps_c_ptr.size());
+    if (tokenized != 0) {
+        throw std::runtime_error("Failed to tokenize prompt");
+    }
+    auto result = server_tokens(chunks, true);
+    return result;
+}
+
+/**
+ * break the input "prompt" object into multiple prompt if needed, then tokenize them
+ * use tokenize_input_prompts() if the input could be an array.
+ * this supports these cases:
+ * - "prompt": "string"
+ * - "prompt": [12, 34, 56]
+ * - "prompt": [12, 34, "string", 56, 78]
+ * - "prompt": { "prompt_string": "string", "multimodal_data": [ "base64" ] }
+ */
+static server_tokens tokenize_input_subprompt(const llama_vocab * vocab, mtmd_context * mctx, const json & json_prompt, bool add_special, bool parse_special) {
+    constexpr char JSON_STRING_PROMPT_KEY[] = "prompt_string";
+    constexpr char JSON_MTMD_DATA_KEY[] = "multimodal_data";
+    const bool has_mtmd = mctx != nullptr;
+    if (json_prompt.is_string() || json_is_array_of_mixed_numbers_strings(json_prompt)) {
+        // string or mixed
+        llama_tokens tmp = tokenize_mixed(vocab, json_prompt, add_special, parse_special);
+        return server_tokens(tmp, false);
+    } else if (json_is_array_of_numbers(json_prompt)) {
+        // array of tokens
+        llama_tokens tmp = json_prompt.get<llama_tokens>();
+        return server_tokens(tmp, false);
+    } else if (json_prompt.contains(JSON_STRING_PROMPT_KEY)) {
+        // JSON object with prompt key.
+        if (json_prompt.contains(JSON_MTMD_DATA_KEY)) {
+            if (!has_mtmd)
+                throw std::runtime_error("Multimodal data provided, but model does not support multimodal requests.");
+
+            // JSON object with prompt and multimodal key.
+            std::vector<raw_buffer> files;
+            for (const auto & entry : json_prompt.at(JSON_MTMD_DATA_KEY)) {
+                files.push_back(base64_decode(entry));
+            }
+            return process_mtmd_prompt(mctx, json_prompt.at(JSON_STRING_PROMPT_KEY), files);
+        } else {
+            // Not multimodal, but contains a subobject.
+            llama_tokens tmp = tokenize_mixed(vocab, json_prompt.at(JSON_STRING_PROMPT_KEY), add_special, parse_special);
+            return server_tokens(tmp, false);
+        }
+   } else {
+       throw std::runtime_error("\"prompt\" elements must be a string, a list of tokens, a JSON object containing a prompt string, or a list of mixed strings & tokens.");
+   }
+}
+
+/**
+ * break the input "prompt" object into multiple prompt if needed, then tokenize them
+ * this supports these cases:
+ * - "prompt": "string"
+ * - "prompt": [12, 34, 56]
+ * - "prompt": [12, 34, "string", 56, 78]
+ * - "prompt": { "prompt_string": "string", "multimodal_data": [ "base64" ] }
+ * and multiple prompts (multi-tasks):
+ * - "prompt": ["string1", "string2"]
+ * - "prompt": ["string1", [12, 34, 56]]
+ * - "prompt": [[12, 34, 56], [78, 90, 12]]
+ * - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56], { "prompt_string": "string", "multimodal_data": [ "base64" ]}]
+ */
+static std::vector<server_tokens> tokenize_input_prompts(const llama_vocab * vocab, mtmd_context * mctx, const json & json_prompt, bool add_special, bool parse_special) {
+    std::vector<server_tokens> result;
+    if (json_prompt.is_array() && !json_is_array_and_contains_numbers(json_prompt)) {
+        result.reserve(json_prompt.size());
+        for (const auto & p : json_prompt) {
+            result.push_back(tokenize_input_subprompt(vocab, mctx, p,add_special, parse_special));
+        }
+    } else {
+        result.push_back(tokenize_input_subprompt(vocab, mctx, json_prompt, add_special, parse_special));
+    }
+    if (result.empty()) {
+        throw std::runtime_error("\"prompt\" must not be empty");
+    }
+    return result;
+}