tests: re-enable MiniMax M3 in test-llama-archs (#26633 )

mtmd: Unlimited-OCR fix max_tiles, setting in converter (#25614 )
grammar : degrade max repetition >= 2000 to unbounded (#26613 )
2026-08-05 19:28:03 +02:00 · 2026-08-05 17:58:34 +02:00 · 2026-08-05 15:30:14 +02:00 · 2026-08-05 07:39:10 -05:00 · 2026-08-05 13:34:52 +02:00 · 2026-08-05 13:29:45 +02:00
155 changed files with 11742 additions and 4062 deletions
@@ -60,10 +60,10 @@ jobs:
            -DCMAKE_BUILD_RPATH="@loader_path" \
            -DLLAMA_FATAL_WARNINGS=ON \
            -DLLAMA_BUILD_BORINGSSL=ON \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=OFF \
            -DGGML_METAL_SHADER_DEBUG=ON \
-            -DGGML_RPC=ON
+            -DGGML_RPC=ON \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
          leaks -atExit -- ./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1

@@ -126,7 +126,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_OPENSSL=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -177,7 +176,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_COMMON=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -211,7 +209,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_BUILD_COMMON=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -256,7 +253,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_OPENSSL=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -119,6 +119,7 @@ jobs:
        run: |
          source ./vulkan_sdk/setup-env.sh
          cmake -B build \
+            -DGGML_NATIVE=OFF \
            -DGGML_VULKAN=ON
          cmake --build build --config Release -j $(nproc)

@@ -93,13 +93,13 @@ jobs:
          - build: 'arm64'
            arch: 'arm64'
            os: macos-26
-            defines: "-DGGML_METAL_USE_BF16=ON -DGGML_METAL_EMBED_LIBRARY=ON"
+            defines: "-DGGML_METAL_EMBED_LIBRARY=ON -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3"
          # TODO: this build is disabled to save Github Actions resources (https://github.com/ggml-org/llama.cpp/pull/23780)
          #       in order to enable it again, we have to provision dedicated runners  to run it
          #- build: 'arm64-kleidiai'
          #  arch: 'arm64'
          #  os: macos-14
-          #  defines: "-DGGML_METAL_USE_BF16=ON -DGGML_METAL_EMBED_LIBRARY=ON -DGGML_CPU_KLEIDIAI=ON"
+          #  defines: "-DGGML_METAL_EMBED_LIBRARY=ON -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3 -DGGML_CPU_KLEIDIAI=ON"
          - build: 'x64'
            arch: 'x64'
            os: macos-15-intel
@@ -1402,7 +1402,6 @@ jobs:
        run: |
          sysctl -a
          cmake -B build -G Xcode \
-            -DGGML_METAL_USE_BF16=ON \
            -DGGML_METAL_EMBED_LIBRARY=ON \
            -DLLAMA_OPENSSL=OFF \
            -DLLAMA_BUILD_APP=OFF \
@@ -21,11 +21,18 @@ Please disclose it as a private [security advisory](https://github.com/ggml-org/

 A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.

+### AI-powered code scan
+
+llama.cpp has an AI security scanner that scans the code periodically. The full prompts and tool set can be found in [ggml-org/security-scan-prompt](https://github.com/ggml-org/security-scan-prompt).
+
+We greatly appreciate reports that reflect genuine research effort, and we are happy to spend our time reviewing them. Findings that an autonomous AI agent can surface on its own add little on top of the scans we already run.
+
 ### Requirements

 Before submitting your report, ensure you meet the following requirements:

 - You have read this policy and fully understand it.
+- You have searched for existing discussions of the issue. If it has already been reported, your report will likely be rejected as a duplicate.
 - AI is only permitted in an assistive capacity as stated in [AGENTS.md](AGENTS.md). We do not accept reports that are written exclusively by AI.
 - Your report must include a working Proof-of-Concept in the form of a script and/or attached files.

@@ -46,6 +53,8 @@ Only vulnerabilities that fall within these parts of the project are considered

 Note that none of the topics under [Using llama.cpp securely](#using-llamacpp-securely) are considered vulnerabilities in LLaMA C++.

+Denial-of-Service (DoS) bugs are generally not treated as vulnerabilities. We don't reject them outright, but we look at them case-by-case and only accept those that are genuinely worth fixing.
+
 For vulnerabilities that fall within the `vendor` directory, please report them directly to the third-party project.

 ## Using llama.cpp securely
@@ -17,7 +17,6 @@ LLAMA_BUILD_MTMD=ON
 GGML_METAL=ON
 GGML_METAL_EMBED_LIBRARY=ON
 GGML_BLAS_DEFAULT=ON
-GGML_METAL_USE_BF16=ON
 GGML_OPENMP=OFF

 COMMON_C_FLAGS="-Wno-macro-redefined -Wno-shorten-64-to-32 -Wno-unused-command-line-argument -g"
@@ -44,7 +43,6 @@ COMMON_CMAKE_ARGS=(
    -DGGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY}
    -DGGML_BLAS_DEFAULT=${GGML_BLAS_DEFAULT}
    -DGGML_METAL=${GGML_METAL}
-    -DGGML_METAL_USE_BF16=${GGML_METAL_USE_BF16}
    -DGGML_NATIVE=OFF
    -DGGML_OPENMP=${GGML_OPENMP}
 )
@@ -61,6 +61,7 @@ static std::initializer_list<enum llama_example> mmproj_examples = {
    LLAMA_EXAMPLE_MTMD,
    LLAMA_EXAMPLE_SERVER,
    LLAMA_EXAMPLE_CLI,
+    LLAMA_EXAMPLE_TTS,
 };

 static std::string read_file(const std::string & fname) {
@@ -360,7 +361,6 @@ static bool spec_types_is_default(const common_params & params) {
 common_models_handler common_models_handler_init(const common_params & params, llama_example curr_ex) {
    common_download_hf_plan plan;
    common_download_hf_plan plan_spec;
-    common_download_hf_plan plan_voc;
    common_download_opts opts;

    const bool spec_type_draft_mtp = std::find(params.speculative.types.begin(),
@@ -413,11 +413,7 @@ common_models_handler common_models_handler_init(const common_params & params, l
        plan_spec = common_download_get_hf_plan(params.speculative.draft.mparams, opts_spec);
    }

-    if (!params.vocoder.model.hf_repo.empty()) {
-        plan_voc = common_download_get_hf_plan(params.vocoder.model, opts);
-    }
-
-    return common_models_handler{plan, plan_spec, plan_voc, opts};
+    return common_models_handler{plan, plan_spec, opts};
 }

 bool common_models_handler_is_preset_repo(const common_models_handler & handler) {
@@ -467,7 +463,6 @@ void common_models_handler_apply(common_models_handler & handler, common_params

    auto & plan      = handler.plan;
    auto & plan_spec = handler.plan_spec;
-    auto & plan_voc  = handler.plan_voc;

    auto opts = handler.opts; // copy
    opts.callback = callback;
@@ -482,7 +477,6 @@ void common_models_handler_apply(common_models_handler & handler, common_params
    };
    handle_url(params.model);
    handle_url(params.mmproj);
-    handle_url(params.vocoder.model);
    handle_url(params.speculative.draft.mparams);

    // optionally, if docker repo is set, resolve it
@@ -510,14 +504,6 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        task.opts       = opts;
        tasks.push_back(task);
    }
-    if (!params.vocoder.model.url.empty()) {
-        common_download_task task;
-        task.url        = params.vocoder.model.url;
-        task.local_path = params.vocoder.model.path;
-        task.opts       = opts;
-        tasks.push_back(task);
-    }
-
    bool had_spec_url = false;
    if (!params.speculative.draft.mparams.url.empty()) {
        common_download_task task;
@@ -631,11 +617,6 @@ void common_models_handler_apply(common_models_handler & handler, common_params
        had_spec_url = true;
    }

-    // handle vocoder plan (e.g. --hf-repo-v)
-    if (!plan_voc.model_files.empty()) {
-        add_tasks(plan_voc.model_files, plan_voc.primary, params.vocoder.model);
-    }
-
    if (!plan.model_files.empty()) {
        add_tasks(plan.model_files, plan.primary, params.model);
    }
@@ -1361,6 +1342,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
        params.n_parallel = -1;     // auto by default
    } else if (ex == LLAMA_EXAMPLE_TOKENIZE) {
        params.parse_special = true; // parse special tokens by default, like the old tokenize tool
+    } else if (ex == LLAMA_EXAMPLE_TTS) {
+        params.out_file = "output.wav";
+        params.sampling.penalty_repeat = 1.05f;
+        params.sampling.penalty_last_n = -1;
    }

    params.use_color = tty_can_use_colors();
@@ -2023,9 +2008,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_sampling());
    add_opt(common_arg(
        {"--repeat-last-n"}, "N",
-        string_format("last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)", params.sampling.penalty_last_n),
+        string_format("last n tokens to consider for penalize (default: %d, 0 = disabled)", params.sampling.penalty_last_n),
        [](common_params & params, int value) {
-            if (value < -1) {
+            if (value < 0) {
                throw std::runtime_error(string_format("error: invalid repeat-last-n = %d\n", value));
            }
            params.sampling.penalty_last_n = value;
@@ -2096,9 +2081,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_sampling());
    add_opt(common_arg(
        {"--dry-penalty-last-n"}, "N",
-        string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable, -1 = context size)", params.sampling.dry_penalty_last_n),
+        string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable)", params.sampling.dry_penalty_last_n),
        [](common_params & params, int value) {
-            if (value < -1) {
+            if (value < 0) {
                throw std::runtime_error(string_format("error: invalid dry-penalty-last-n = %d\n", value));
            }
            params.sampling.dry_penalty_last_n = value;
@@ -2983,20 +2968,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            params.model.hf_file = value;
        }
    ).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_DOWNLOAD, LLAMA_EXAMPLE_TOKENIZE}).set_env("LLAMA_ARG_HF_FILE"));
-    add_opt(common_arg(
-        {"-hfv", "-hfrv", "--hf-repo-v"}, "<user>/<model>[:quant]",
-        "Hugging Face model repository for the vocoder model (default: unused)",
-        [](common_params & params, const std::string & value) {
-            params.vocoder.model.hf_repo = value;
-        }
-    ).set_env("LLAMA_ARG_HF_REPO_V"));
-    add_opt(common_arg(
-        {"-hffv", "--hf-file-v"}, "FILE",
-        "Hugging Face model file for the vocoder model (default: unused)",
-        [](common_params & params, const std::string & value) {
-            params.vocoder.model.hf_file = value;
-        }
-    ).set_env("LLAMA_ARG_HF_FILE_V"));
    add_opt(common_arg(
        {"-hft", "--hf-token"}, "TOKEN",
        "Hugging Face access token (default: value from HF_TOKEN environment variable)",
@@ -4272,24 +4243,18 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    //

    add_opt(common_arg(
-        {"-mv", "--model-vocoder"}, "FNAME",
-        "vocoder model for audio generation (default: unused)",
+        {"--tts-lang"}, "FNAME",
+        "language (ISO 639-1) for audio generation\n"
+        "see tts/README.md for per-model usage notes",
        [](common_params & params, const std::string & value) {
-            params.vocoder.model.path = value;
+            params.tts_lang = value;
        }
-    ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
-     add_opt(common_arg(
-        {"--tts-use-guide-tokens"},
-        "Use guide tokens to improve TTS word recall",
-        [](common_params & params) {
-            params.vocoder.use_guide_tokens = true;
-        }
-    ).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_TTS}));
    add_opt(common_arg(
        {"--tts-speaker-file"}, "FNAME",
        "speaker file path for audio generation",
        [](common_params & params, const std::string & value) {
-            params.vocoder.speaker_file = value;
+            params.tts_speaker_file = value;
        }
    ).set_examples({LLAMA_EXAMPLE_TTS}));

@@ -4409,16 +4374,6 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    ).set_examples({LLAMA_EXAMPLE_DEBUG}));

    // presets
-    add_opt(common_arg(
-        {"--tts-oute-default"},
-        string_format("use default OuteTTS models (note: can download weights from the internet)"),
-        [](common_params & params) {
-            params.model.hf_repo = "OuteAI/OuteTTS-0.2-500M-GGUF";
-            params.model.hf_file = "OuteTTS-0.2-500M-Q8_0.gguf";
-            params.vocoder.model.hf_repo = "ggml-org/WavTokenizer";
-            params.vocoder.model.hf_file = "WavTokenizer-Large-75-F16.gguf";
-        }
-    ).set_examples({LLAMA_EXAMPLE_TTS}));

    add_opt(common_arg(
        {"--embd-gemma-default"},
@@ -137,7 +137,6 @@ void common_params_add_preset_options(std::vector<common_arg> & args);
 struct common_models_handler {
    common_download_hf_plan plan;
    common_download_hf_plan plan_spec;
-    common_download_hf_plan plan_voc;
    common_download_opts opts;
 };

@@ -1302,23 +1302,12 @@ common_init_result::common_init_result(common_params & params, bool model_only)
                params.sampling.logit_bias_eog.begin(), params.sampling.logit_bias_eog.end());
    }

-    //if (params.sampling.penalty_last_n == -1) {
-    //    LOG_TRC("%s: setting penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-    //    params.sampling.penalty_last_n = llama_n_ctx(lctx);
-    //}
-
-    //if (params.sampling.dry_penalty_last_n == -1) {
-    //    LOG_TRC("%s: setting dry_penalty_last_n to ctx_size = %d\n", __func__, llama_n_ctx(lctx));
-    //    params.sampling.dry_penalty_last_n = llama_n_ctx(lctx);
-    //}
-
    // init the backend samplers as part of the context creation
    pimpl->samplers.resize(cparams.n_seq_max);
    pimpl->samplers_seq_config.resize(cparams.n_seq_max);

-    const int32_t n_ctx = cparams.n_ctx > 0 ? (int32_t) cparams.n_ctx : llama_model_n_ctx_train(model);
    for (int i = 0; i < (int) cparams.n_seq_max; ++i) {
-        pimpl->samplers[i].reset(common_sampler_init(model, params.sampling, n_ctx));
+        pimpl->samplers[i].reset(common_sampler_init(model, params.sampling));
        pimpl->samplers_seq_config[i] = { i, common_sampler_get(pimpl->samplers[i].get()) };
    }

@@ -235,14 +235,14 @@ struct common_params_sampling {
    float   temp               = 0.80f;  // <= 0.0 to sample greedily, 0.0 to not output probabilities
    float   dynatemp_range     = 0.00f;  // 0.0 = disabled
    float   dynatemp_exponent  = 1.00f;  // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t penalty_last_n     = 64;     // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    int32_t penalty_last_n     = 64;     // last n tokens to penalize (0 = disable penalty)
    float   penalty_repeat     = 1.00f;  // 1.0 = disabled
    float   penalty_freq       = 0.00f;  // 0.0 = disabled
    float   penalty_present    = 0.00f;  // 0.0 = disabled
    float   dry_multiplier     = 0.0f;   // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
    float   dry_base           = 1.75f;  // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
    int32_t dry_allowed_length = 2;      // tokens extending repetitions beyond this receive penalty
-    int32_t dry_penalty_last_n = -1;     // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
+    int32_t dry_penalty_last_n = 64;     // how many tokens to scan for repetitions (0 = disable penalty)
    float   adaptive_target    = -1.0f;  // select tokens near this probability (valid range 0.0 to 1.0; negative = disabled)
    float   adaptive_decay     = 0.90f;  // EMA decay for adaptation; history ≈ 1/(1-decay) tokens (0.0 - 0.99)
    int32_t mirostat           = 0;      // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
@@ -392,14 +392,6 @@ struct common_params_speculative {
    }
 };

-struct common_params_vocoder {
-    struct common_params_model model;
-
-    std::string speaker_file; // speaker file path
-
-    bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy
-};
-
 struct common_params_diffusion {
    int32_t steps         = 128;
    bool    visual_mode   = false;
@@ -497,7 +489,6 @@ struct common_params {

    struct common_params_sampling    sampling;
    struct common_params_speculative speculative;
-    struct common_params_vocoder     vocoder;
    struct common_params_diffusion   diffusion;

    struct common_params_model model;
@@ -740,6 +731,10 @@ struct common_params {
    void *                  load_progress_callback_user_data = NULL;
    bool no_alloc = false; // Don't allocate model buffers

+    // TTS params
+    std::string tts_lang = "";
+    std::string tts_speaker_file = "";
+
    bool is_gen_docs = false; // whether we are running inside llama-gen-docs
 };

@@ -136,7 +136,10 @@ static std::vector<llama_device_memory_data> common_get_device_memory_data_impl(
        devs.push_back(llama_model_get_device(model, i));
    }

-    hp_ngl         = llama_model_n_layer(model) + llama_model_n_layer_nextn(model);
+    hp_ngl         = llama_model_n_layer(model);
+    if (mparams->load_mtp) {
+        hp_ngl    += llama_model_n_layer_nextn(model);
+    }
    hp_n_ctx_train = llama_model_n_ctx_train(model);
    hp_n_expert    = llama_model_n_expert(model);

@@ -186,8 +186,7 @@ std::string common_params_sampling::print() const {

 struct common_sampler * common_sampler_init(
        const struct llama_model * model,
-        struct common_params_sampling & params,
-        int32_t n_ctx) {
+        struct common_params_sampling & params) {
    if (!std::isfinite(params.penalty_repeat) ||
        params.penalty_repeat <= 0.0f ||
        !std::isfinite(1.0f/params.penalty_repeat)) {
@@ -199,10 +198,6 @@ struct common_sampler * common_sampler_init(
    if (!std::isfinite(params.penalty_present)) {
        throw std::invalid_argument("penalty_present must be finite");
    }
-    if (params.penalty_last_n == -1) {
-        params.penalty_last_n = n_ctx > 0 ? n_ctx : llama_model_n_ctx_train(model);
-    }
-
    const llama_vocab * vocab = llama_model_get_vocab(model);
    llama_sampler_chain_params lparams = llama_sampler_chain_default_params();

@@ -355,7 +350,7 @@ struct common_sampler * common_sampler_init(
                        for (const auto & str : params.dry_sequence_breakers) {
                            c_breakers.push_back(str.c_str());
                        }
-                        samplers.push_back(llama_sampler_init_dry(vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                        samplers.push_back(llama_sampler_init_dry(vocab, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
                    }
                    break;
                case COMMON_SAMPLER_TYPE_TOP_K:
@@ -383,7 +378,7 @@ struct common_sampler * common_sampler_init(
                    samplers.push_back(llama_sampler_init_infill(vocab));
                    break;
                case COMMON_SAMPLER_TYPE_PENALTIES:
-                    samplers.push_back(llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
+                    samplers.push_back(llama_sampler_init_penalties(llama_vocab_n_tokens(vocab), params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
                    break;
                case COMMON_SAMPLER_TYPE_ADAPTIVE_P:
                    // the `adaptive-p` sampler is like `dist` and `mirostat` in that it selects
@@ -39,8 +39,7 @@ struct common_sampler;
 // note: can mutate params in some cases
 struct common_sampler * common_sampler_init(
        const struct llama_model * model,
-        struct common_params_sampling & params,
-        int32_t n_ctx = 0);
+        struct common_params_sampling & params);

 void common_sampler_free(struct common_sampler * gsmpl);

@@ -2385,57 +2385,28 @@ common_speculative * common_speculative_init(common_params_speculative & params,
    {
        uint32_t enabled_configs = common_get_enabled_speculative_configs(params.types);

-        bool has_draft_simple = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE));
-        bool has_draft_eagle3 = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3)) && params.draft.ctx_dft != nullptr;
-        bool has_draft_mtp    = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_MTP))    && params.draft.ctx_dft != nullptr;
-        bool has_draft_dflash = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH)) && params.draft.ctx_dft != nullptr;
-        bool has_draft_dspark = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK)) && params.draft.ctx_dft != nullptr;
-
-
-
-        bool has_ngram_cache   = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_CACHE));
-        bool has_ngram_simple  = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE));
-        bool has_ngram_map_k   = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K));
-        bool has_ngram_map_k4v = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V));
-        bool has_ngram_mod     = (enabled_configs & (1u << COMMON_SPECULATIVE_TYPE_NGRAM_MOD));
+        auto add_config_if_enabled = [&](common_speculative_type type, bool available = true) {
+            if (available && (enabled_configs & (1u << type))) {
+                configs.emplace_back(type, params);
+            }
+        };

        // when adding a new type - update here the logic above
        static_assert(COMMON_SPECULATIVE_TYPE_COUNT == 11);

        // this list here defines the priority of the speculators
        // the one with highest priority are listed first
-        if (has_ngram_simple) {
-            // This implementation can guess a lot of tokens without any draft model.
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE, params));
-        }
-        if (has_ngram_map_k) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K, params));
-        }
-        if (has_ngram_map_k4v) {
-            // This implementation can guess tokens with high acceptance rate but is more expensive.
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V, params));
-        }
-        if (has_ngram_mod) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_MOD, params));
-        }
-        if (has_ngram_cache) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE, params));
-        }
-        if (has_draft_simple) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE, params));
-        }
-        if (has_draft_eagle3) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, params));
-        }
-        if (has_draft_mtp) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_MTP, params));
-        }
-        if (has_draft_dflash) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, params));
-        }
-        if (has_draft_dspark) {
-            configs.push_back(common_speculative_config(COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK, params));
-        }
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_SIMPLE);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_MAP_K4V);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_MOD);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_NGRAM_CACHE);
+
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_SIMPLE);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3, params.draft.ctx_dft != nullptr);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_MTP,    params.draft.ctx_dft != nullptr);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH, params.draft.ctx_dft != nullptr);
+        add_config_if_enabled(COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK, params.draft.ctx_dft != nullptr);
    }

    std::vector<std::unique_ptr<common_speculative_impl>> impls = {};
@@ -210,6 +210,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "Qwen3MoeForCausalLM": "qwen",
    "Qwen3NextForCausalLM": "qwen",
    "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
+    "Qwen3TTSForConditionalGeneration": "qwen3tts",
    "Qwen3VLForConditionalGeneration": "qwen3vl",
    "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
    "Qwen3_5ForCausalLM": "qwen",
@@ -304,6 +305,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "Qwen2_5_VLForConditionalGeneration": "qwenvl",
    "Qwen3ASRForConditionalGeneration": "qwen3vl",
    "Qwen3OmniMoeForConditionalGeneration": "qwen3vl",
+    "Qwen3TTSForConditionalGeneration": "qwen3tts",
    "Qwen3VLForConditionalGeneration": "qwen3vl",
    "Qwen3VLMoeForConditionalGeneration": "qwen3vl",
    "Qwen3_5ForConditionalGeneration": "qwen3vl",
@@ -81,7 +81,7 @@ class ChatGLMModel(TextModel):

    @staticmethod
    def token_bytes_to_string(b):
-        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode  # ty: ignore[unresolved-import]
+        from transformers.convert_slow_tokenizer import bytes_to_unicode
        byte_encoder = bytes_to_unicode()
        return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])

@@ -17,8 +17,11 @@ from .base import LazyTorchTensor, MmprojModel, ModelBase, TextModel, gguf, logg
 from .qwen import QwenModel


-@ModelBase.register("DeepseekOCRForCausalLM", "UnlimitedOCRForCausalLM")
+@ModelBase.register("DeepseekOCRForCausalLM")
 class DeepseekOCRVisionModel(MmprojModel):
+    # HF dynamic_preprocess() max_num, which differs per model
+    preproc_max_tiles = 9
+
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR
@@ -43,6 +46,9 @@ class DeepseekOCRVisionModel(MmprojModel):
        # @bluebread: there's no window_size in config but just add it here anyway
        self.gguf_writer.add_vision_window_size(self.hparams.get("window_size", 14))

+        self.gguf_writer.add_vision_preproc_min_tiles(2)
+        self.gguf_writer.add_vision_preproc_max_tiles(self.preproc_max_tiles)
+
        # SAM configuration
        sam_hparams = hparams['sam']
        self.gguf_writer.add_vision_sam_layers_count(sam_hparams['layers'])
@@ -93,8 +99,15 @@ class DeepseekOCRVisionModel(MmprojModel):
        return super().filter_tensors((name, gen))


+@ModelBase.register("UnlimitedOCRForCausalLM")
+class UnlimitedOCRVisionModel(DeepseekOCRVisionModel):
+    preproc_max_tiles = 32
+
+
@ModelBase.register("DeepseekOCR2ForCausalLM")
 class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):
+    preproc_max_tiles = 6
+
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR2
@@ -119,7 +119,7 @@ class LlamaModel(TextModel):
        path_tekken_json = self.dir_model / "tekken.json"
        path_tokenizer_json = self.dir_model / "tokenizer.json"
        if path_tekken_json.is_file() and not path_tokenizer_json.is_file():
-            self._set_vocab_mistral()
+            return self._set_vocab_mistral()

        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
        if tokenizer_config_file.is_file():
@@ -18,7 +18,7 @@ class QwenModel(TextModel):

    @staticmethod
    def token_bytes_to_string(b):
-        from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode  # ty: ignore[unresolved-import]
+        from transformers.convert_slow_tokenizer import bytes_to_unicode
        byte_encoder = bytes_to_unicode()
        return ''.join([byte_encoder[ord(char)] for char in b.decode('latin-1')])

@@ -0,0 +1,471 @@
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any, Callable, Iterable, TYPE_CHECKING
+
+import torch
+import torch.nn.functional as F
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, MmprojModel, TextModel, gguf
+
+# Tricks being used to support this model via existing llama.cpp code paths:
+# - Text projection MLP is folded into the embedding table
+# - codec_embedding is concat to the text embedding table, vocab is extended
+#   example: codec_bos_id(2149) --> "<|codec_bos|>"
+#            codec_eos_token_id(2150) --> "<|codec_eos_token|>"
+#            codec_language_id.chinese(2055) --> "<|codec_language_chinese|>"
+#            other rows --> "<|codec_0|>", "<|codec_1|>", ..., "<|codec_1023|>"
+# - output tensor codec_head is smaller than vocab, so logits will be padded at inference time
+# - suppress_tokens is used to limit the backbone to only sample either semantic or EOS (stop) token
+
+# pipeline stage mapping:
+#   speaker reference encoder --> mapped to normal mtmd audio encoder
+#   backbone --> mapped to normal libllama text model (autoregressive)
+#   code_predictor --> MTMD_GEN_PROCESS_TYPE_GEN_CODE
+#   code2wav --> MTMD_GEN_PROCESS_TYPE_GEN_WAV
+
+# torch activation functions used by Qwen3TTSTalkerResizeMLP (config's hidden_act)
+_ACT2FN = {
+    "silu": F.silu,
+    "gelu": F.gelu,
+    "relu": F.relu,
+}
+
+
+@ModelBase.register("Qwen3TTSForConditionalGeneration")
+class Qwen3TTSTalkerModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.QWEN3TTS
+
+    _TEXT_PROJ_KEYS = (
+        "model.text_embedding.weight",
+        "text_projection.linear_fc1.weight",
+        "text_projection.linear_fc1.bias",
+        "text_projection.linear_fc2.weight",
+        "text_projection.linear_fc2.bias",
+    )
+
+    _text_proj_buffer: dict[str, Tensor]
+    _folded_text_embed: Tensor | None
+    _codec_embed: Tensor | None
+
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(dir_model, is_mistral_format=False)
+        raw_talker_config = dict(hparams["talker_config"])
+        self._talker_config = raw_talker_config
+        self.n_codec_vocab = raw_talker_config["vocab_size"]
+        talker_config = dict(raw_talker_config)
+        talker_config["vocab_size"] = talker_config["text_vocab_size"]
+        hparams["text_config"] = talker_config
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+        self._text_proj_buffer = {}
+        self._folded_text_embed = None
+        self._codec_embed = None
+
+    def _codec_token_names(self) -> list[str]:
+        # start every row with a generic name, then override the ones with a
+        # known meaning (bos/eos/language/etc, derived from the *_id fields
+        # of talker_config) with a more descriptive one
+        names = [f"<|codec_{i}|>" for i in range(self.n_codec_vocab)]
+        for key, val in self._talker_config.items():
+            if not key.endswith("_id"):
+                continue
+            prefix = key[:-len("_id")]
+            if isinstance(val, int):
+                names[val] = f"<|{prefix}|>"
+            elif isinstance(val, dict):
+                for subkey, subval in val.items():
+                    names[subval] = f"<|{prefix}_{subkey}|>"
+        return names
+
+    def set_vocab(self):
+        codec_tokens = self._codec_token_names()
+        codec_toktypes = [gguf.TokenType.CONTROL] * len(codec_tokens)
+
+        try:
+            tokens, scores, toktypes = self._create_vocab_sentencepiece()
+            self.gguf_writer.add_tokenizer_model("llama")
+            self.gguf_writer.add_tokenizer_pre("default")
+            tokens += [t.encode("utf-8") for t in codec_tokens]
+            scores += [0.0] * len(codec_tokens)
+            toktypes += codec_toktypes
+            self.gguf_writer.add_token_list(tokens)
+            self.gguf_writer.add_token_scores(scores)
+            self.gguf_writer.add_token_types(toktypes)
+            special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+            special_vocab.add_to_gguf(self.gguf_writer)
+            return
+        except FileNotFoundError:
+            pass
+
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        tokens += codec_tokens
+        toktypes += codec_toktypes
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        # make sure that the model has no chat template, so chat will be disabled
+        self.gguf_writer.add_chat_template(None)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # note: final vocab layout is [text_vocab | codec_vocab], with text_vocab is actually padded with -inf in cgraph
+        # for codec_vocab, only first 2048 rows can be sampled for semantic code
+        # plus codec_eos_token_id that used for signaling end of generation
+        # ref: https://github.com/QwenLM/Qwen3-TTS/blob/022e286b98fbec7e1e916cb940cdf532cd9f488e/qwen_tts/core/models/modeling_qwen3_tts.py#L2059-L2063
+
+        vocab_size = self.hparams["vocab_size"] + self.n_codec_vocab
+        codec_eos_token_id = self.hparams["vocab_size"] + self._talker_config["codec_eos_token_id"]
+        self.gguf_writer.add_suppress_tokens([
+            i for i in range(vocab_size - 1024, vocab_size)
+            if i != codec_eos_token_id
+        ])
+        self.gguf_writer.add_eos_token_id(codec_eos_token_id)
+        self.gguf_writer.add_add_eos_token(False)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+
+        if not name.startswith("talker.") or name.startswith("talker.code_predictor."):
+            return None
+
+        name = name[len("talker."):]
+        return super().filter_tensors((name, gen))
+
+    def _maybe_emit_token_embd(self) -> Iterable[tuple[str, Tensor]]:
+        if self._folded_text_embed is None or self._codec_embed is None:
+            return
+        combined = torch.cat([self._folded_text_embed, self._codec_embed], dim=0)
+        yield (self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD), combined)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # codec_embedding rows are appended after the text vocab, extending the embedding table
+        if name == "model.codec_embedding.weight":
+            self._codec_embed = data_torch
+            yield from self._maybe_emit_token_embd()
+            return
+
+        # codec_head is the output head for the (smaller) codec vocab; logits get padded to
+        # the extended vocab size at inference time
+        if name == "codec_head.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch)
+            return
+
+        if name in self._TEXT_PROJ_KEYS:
+            self._text_proj_buffer[name] = data_torch
+            if len(self._text_proj_buffer) < len(self._TEXT_PROJ_KEYS):
+                return
+
+            # fold MLP into the embedding table at conversion time, MLP won't be used at inference time anyway
+            act_fn = _ACT2FN[self.hparams["hidden_act"]]
+            embed = self._text_proj_buffer["model.text_embedding.weight"]
+            hidden = act_fn(F.linear(embed,
+                                     self._text_proj_buffer["text_projection.linear_fc1.weight"],
+                                     self._text_proj_buffer["text_projection.linear_fc1.bias"]))
+            folded = F.linear(hidden,
+                              self._text_proj_buffer["text_projection.linear_fc2.weight"],
+                              self._text_proj_buffer["text_projection.linear_fc2.bias"])
+            self._folded_text_embed = folded
+            yield from self._maybe_emit_token_embd()
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Qwen3TTSForConditionalGeneration")
+class Qwen3TTSSpeakerEncoderModel(MmprojModel):
+    has_vision_encoder = False
+    has_audio_encoder = True
+
+    # talker.code_predictor.model.layers.{bid}.<key> -> A_GEN_CODE_*
+    # bypass tensor_mapping.py for now to make it simple
+    _CODE_LAYER_TENSOR_MAP = {
+        "input_layernorm": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_NORM,
+        "self_attn.q_proj": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_Q,
+        "self_attn.q_norm": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_Q_NORM,
+        "self_attn.k_proj": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_K,
+        "self_attn.k_norm": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_K_NORM,
+        "self_attn.v_proj": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_V,
+        "self_attn.o_proj": gguf.MODEL_TENSOR.A_GEN_CODE_ATTN_OUT,
+        "post_attention_layernorm": gguf.MODEL_TENSOR.A_GEN_CODE_FFN_NORM,
+        "mlp.gate_proj": gguf.MODEL_TENSOR.A_GEN_CODE_FFN_GATE,
+        "mlp.up_proj": gguf.MODEL_TENSOR.A_GEN_CODE_FFN_UP,
+        "mlp.down_proj": gguf.MODEL_TENSOR.A_GEN_CODE_FFN_DOWN,
+    }
+
+    # note: codebook pages will be stacked to 3D
+    _CODE_GEN_N_CODEBOOKS = 15
+    _code_embed_buffer: dict[int, Tensor] = {}
+    _code_head_buffer: dict[int, Tensor] = {}
+    _wav_config_cache: dict[str, Any] | None = None
+
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(dir_model, is_mistral_format=False)
+        hparams["text_config"] = {"hidden_size": hparams["talker_config"]["hidden_size"]}
+        # ECAPA-TDNN has a fixed 4-stage backbone, but MmprojModel.__init__ needs a n_block_keys
+        hparams["speaker_encoder_config"]["n_layers"] = 4
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+        self._wav_config_cache = None
+
+    def get_audio_config(self) -> dict[str, Any] | None:
+        return self.global_config.get("speaker_encoder_config")
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_clip_has_audio_encoder(True)
+        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.QWEN3TTS_SPKENC)
+
+        # handle speaker encoder config
+        self.gguf_writer.add_audio_projection_dim(self.n_embd_text)
+        # mel_spectrogram() front-end: sr=24000, n_fft=1024, hop=256, n_mels=128, fmin=0, fmax=12000 (=sr/2, the clip.cpp default)
+        self.gguf_writer.add_audio_num_mel_bins(128)
+        # 3 SE-Res2Net stages; the stem conv, mfa, asp and fc are not counted here
+        self.gguf_writer.add_audio_block_count(3)
+        # ECAPA-TDNN has no attention/FFN, these are dummy to allow clip.cpp to load it
+        self.gguf_writer.add_audio_embedding_length(1536)
+        self.gguf_writer.add_audio_head_count(1)
+        self.gguf_writer.add_audio_feed_forward_length(1536)
+        self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+
+        # handle code predictor config
+        self.gguf_writer.add_clip_has_gen_audio_encoder(True)
+        self.gguf_writer.add_clip_gen_audio_projector_type(gguf.VisionProjectorType.QWEN3TTS_GEN)
+        code_predictor_config = self.global_config["talker_config"]["code_predictor_config"]
+        self.gguf_writer.add_gen_audio_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_gen_audio_embedding_length(code_predictor_config["hidden_size"])
+        self.gguf_writer.add_gen_audio_feed_forward_length(code_predictor_config["intermediate_size"])
+        self.gguf_writer.add_gen_audio_block_count(code_predictor_config["num_hidden_layers"])
+        self.gguf_writer.add_gen_audio_head_count(code_predictor_config["num_attention_heads"])
+        self.gguf_writer.add_gen_audio_head_count_kv(code_predictor_config["num_key_value_heads"])
+        self.gguf_writer.add_gen_audio_attention_layernorm_eps(code_predictor_config["rms_norm_eps"])
+        # note: code2wav hparams are hardcoded on the mtmd/clip.cpp side for now, not written here
+
+    def _wav_decoder_config(self) -> dict[str, Any] | None:
+        # code2wav has its own config.json, inside the speech_tokenizer dir
+        if self._wav_config_cache is None:
+            path = self.dir_model / "speech_tokenizer" / "config.json"
+            with open(path, "r", encoding="utf-8") as f:
+                cfg = json.load(f)
+            self._wav_config_cache = cfg["decoder_config"]
+        return self._wav_config_cache
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        # conv1d/conv1d_dw kernels must be F16, ggml_conv_1d(_dw) has no BF16 path
+        if new_name.endswith(".weight") and (
+            new_name in ("a.gen.wav.pre_conv.weight", "a.gen.wav.dac.entry.weight", "a.gen.wav.dac.post_conv.weight")
+            or (".up.blk." in new_name and new_name.endswith(".dwconv.weight"))
+            or (".dac.blk." in new_name and (new_name.endswith(".conv1.weight") or new_name.endswith(".conv2.weight")))
+        ):
+            return gguf.GGMLQuantizationType.F16
+        # ConvTranspose1d kernels: only F16/F32 are implemented, no BF16
+        if new_name.endswith(".conv.weight") and (".up.blk." in new_name or ".dac.blk." in new_name):
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+
+        if not (
+            name.startswith("speaker_encoder.")
+            or name.startswith("talker.code_predictor.")
+            or name == "talker.model.codec_embedding.weight"
+        ):
+            return None
+
+        return super().filter_tensors((name, gen))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # code2wav tensors are already named by generate_extra_tensors(), pass them through
+        if name.startswith("a.gen.wav."):
+            yield (name, data_torch)
+            return
+
+        # codebook-0 embedding, fed back to the talker backbone (codebooks 1-15 live in code_predictor)
+        if name == "talker.model.codec_embedding.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_GEN_CODE_OUT_EMBD), data_torch)
+            return
+
+        if name == "talker.code_predictor.model.norm.weight":
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_GEN_CODE_OUTPUT_NORM), data_torch)
+            return
+
+        if name.startswith("talker.code_predictor.small_to_mtp_projection."):
+            suffix = "." + name.rsplit(".", 1)[1]
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_GEN_CODE_PROJ_IN, suffix=suffix), data_torch)
+            return
+
+        if name.startswith("talker.code_predictor.model.codec_embedding."):
+            idx = int(name.split("codec_embedding.")[1].split(".")[0])
+            self._code_embed_buffer[idx] = data_torch
+            if len(self._code_embed_buffer) < self._CODE_GEN_N_CODEBOOKS:
+                return
+            stacked = torch.stack([self._code_embed_buffer.pop(i) for i in range(self._CODE_GEN_N_CODEBOOKS)], dim=0)
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_GEN_CODE_EMBD), stacked)
+            return
+
+        if name.startswith("talker.code_predictor.lm_head."):
+            idx = int(name.split("lm_head.")[1].split(".")[0])
+            self._code_head_buffer[idx] = data_torch
+            if len(self._code_head_buffer) < self._CODE_GEN_N_CODEBOOKS:
+                return
+            stacked = torch.stack([self._code_head_buffer.pop(i) for i in range(self._CODE_GEN_N_CODEBOOKS)], dim=0)
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.A_GEN_CODE_HEAD), stacked)
+            return
+
+        if name.startswith("talker.code_predictor.model.layers."):
+            rest = name.split("model.layers.")[1]        # "{bid}.<key>.weight"
+            _, key_with_suffix = rest.split(".", 1)       # "<key>.weight"
+            key = key_with_suffix.rsplit(".", 1)[0]        # "<key>"
+            tensor = self._CODE_LAYER_TENSOR_MAP.get(key)
+            if tensor is not None:
+                yield (self.format_tensor_name(tensor, bid), data_torch)
+                return
+
+        if "res2net_block.blocks." in name:
+            assert bid is not None  # the outer stage index, picked up from the tensor name automatically
+            xid = int(name.split("res2net_block.blocks.")[1].split(".")[0])
+            suffix = "." + name.rsplit(".", 1)[1]
+            new_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.A_ENC_CONV_RES2].format(bid=bid, xid=xid) + suffix
+            yield (new_name, data_torch)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        yield from self._generate_code2wav_tensors()
+
+    def _generate_code2wav_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        # code2wav weights live in speech_tokenizer/model.safetensors, not the main safetensors
+        from safetensors.torch import load_file
+
+        wav_config = self._wav_decoder_config()
+        state_dict = load_file(self.dir_model / "speech_tokenizer" / "model.safetensors")
+
+        def get(name: str) -> Tensor:
+            return state_dict[name]
+
+        def snake_fold(alpha: Tensor, beta: Tensor) -> tuple[Tensor, Tensor]:
+            # fold SnakeBeta's exp()/reciprocal here, so the graph is only mul/sin/sqr/mul/add
+            return torch.exp(alpha), 1.0 / (torch.exp(beta) + 1e-9)
+
+        def rvq_codebook(prefix: str, n_layers: int) -> Tensor:
+            # checkpoint has EMA accumulators, so codebook[i] = embedding_sum[i] / cluster_usage[i]
+            books = []
+            for i in range(n_layers):
+                embedding_sum = get(f"{prefix}.vq.layers.{i}._codebook.embedding_sum")
+                cluster_usage = get(f"{prefix}.vq.layers.{i}._codebook.cluster_usage")
+                books.append(embedding_sum / cluster_usage.clamp_min(1e-5).unsqueeze(-1))
+            return torch.stack(books, dim=0) if n_layers > 1 else books[0]
+
+        T = gguf.MODEL_TENSOR
+
+        # --- quantizer: RVQ codebook decode ---
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_FIRST_IN), get("decoder.quantizer.rvq_first.input_proj.weight").squeeze(-1))
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_FIRST_OUT), get("decoder.quantizer.rvq_first.output_proj.weight").squeeze(-1))
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_FIRST_CB), rvq_codebook("decoder.quantizer.rvq_first", 1))
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_REST_IN), get("decoder.quantizer.rvq_rest.input_proj.weight").squeeze(-1))
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_REST_OUT), get("decoder.quantizer.rvq_rest.output_proj.weight").squeeze(-1))
+        yield (self.format_tensor_name(T.A_GEN_WAV_QUANT_REST_CB), rvq_codebook("decoder.quantizer.rvq_rest", self._CODE_GEN_N_CODEBOOKS))
+
+        # --- pre_conv ---
+        yield (self.format_tensor_name(T.A_GEN_WAV_PRE_CONV, suffix=".weight"), get("decoder.pre_conv.conv.weight"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_PRE_CONV, suffix=".bias"), get("decoder.pre_conv.conv.bias"))
+
+        # --- pre_transformer ---
+        yield (self.format_tensor_name(T.A_GEN_WAV_TFM_IN_PROJ, suffix=".weight"), get("decoder.pre_transformer.input_proj.weight"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_TFM_IN_PROJ, suffix=".bias"), get("decoder.pre_transformer.input_proj.bias"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_TFM_OUT_PROJ, suffix=".weight"), get("decoder.pre_transformer.output_proj.weight"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_TFM_OUT_PROJ, suffix=".bias"), get("decoder.pre_transformer.output_proj.bias"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_TFM_OUTPUT_NORM), get("decoder.pre_transformer.norm.weight"))
+
+        tfm_layer_map = {
+            "input_layernorm.weight":         T.A_GEN_WAV_TFM_ATTN_NORM,
+            "self_attn.q_proj.weight":        T.A_GEN_WAV_TFM_ATTN_Q,
+            "self_attn.k_proj.weight":        T.A_GEN_WAV_TFM_ATTN_K,
+            "self_attn.v_proj.weight":        T.A_GEN_WAV_TFM_ATTN_V,
+            "self_attn.o_proj.weight":        T.A_GEN_WAV_TFM_ATTN_OUT,
+            "self_attn_layer_scale.scale":    T.A_GEN_WAV_TFM_ATTN_SCALE,
+            "post_attention_layernorm.weight": T.A_GEN_WAV_TFM_FFN_NORM,
+            "mlp.gate_proj.weight":           T.A_GEN_WAV_TFM_FFN_GATE,
+            "mlp.up_proj.weight":             T.A_GEN_WAV_TFM_FFN_UP,
+            "mlp.down_proj.weight":           T.A_GEN_WAV_TFM_FFN_DOWN,
+            "mlp_layer_scale.scale":          T.A_GEN_WAV_TFM_FFN_SCALE,
+        }
+        assert wav_config is not None
+        for bid in range(wav_config["num_hidden_layers"]):
+            for key, tensor_id in tfm_layer_map.items():
+                yield (self.format_tensor_name(tensor_id, bid), get(f"decoder.pre_transformer.layers.{bid}.{key}"))
+
+        # --- upsample: 2x (causal ConvTranspose1d + ConvNeXt block) ---
+        up_map = {
+            "0.conv.weight":     (T.A_GEN_WAV_UP_CONV, ".weight"),
+            "0.conv.bias":       (T.A_GEN_WAV_UP_CONV, ".bias"),
+            "1.dwconv.conv.weight": (T.A_GEN_WAV_UP_DWCONV, ".weight"),
+            "1.dwconv.conv.bias":   (T.A_GEN_WAV_UP_DWCONV, ".bias"),
+            "1.norm.weight":     (T.A_GEN_WAV_UP_NORM, ".weight"),
+            "1.norm.bias":       (T.A_GEN_WAV_UP_NORM, ".bias"),
+            "1.pwconv1.weight":  (T.A_GEN_WAV_UP_PW1, ".weight"),
+            "1.pwconv1.bias":    (T.A_GEN_WAV_UP_PW1, ".bias"),
+            "1.pwconv2.weight":  (T.A_GEN_WAV_UP_PW2, ".weight"),
+            "1.pwconv2.bias":    (T.A_GEN_WAV_UP_PW2, ".bias"),
+            "1.gamma":           (T.A_GEN_WAV_UP_GAMMA, ""),
+        }
+        for bid in range(len(wav_config["upsampling_ratios"])):
+            for key, (tensor_id, suffix) in up_map.items():
+                yield (self.format_tensor_name(tensor_id, bid, suffix=suffix), get(f"decoder.upsample.{bid}.{key}"))
+
+        # --- DAC decoder ---
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_ENTRY, suffix=".weight"), get("decoder.decoder.0.conv.weight"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_ENTRY, suffix=".bias"), get("decoder.decoder.0.conv.bias"))
+
+        n_dac_blocks = len(wav_config["upsample_rates"])
+        for bid in range(n_dac_blocks):
+            py = bid + 1  # decoder.decoder.0 is the entry conv, blocks start at 1
+
+            a, b = snake_fold(get(f"decoder.decoder.{py}.block.0.alpha"), get(f"decoder.decoder.{py}.block.0.beta"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_SNAKE, bid, suffix=".alpha"), a)
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_SNAKE, bid, suffix=".beta"), b)
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_CONV, bid, suffix=".weight"), get(f"decoder.decoder.{py}.block.1.conv.weight"))
+            yield (self.format_tensor_name(T.A_GEN_WAV_DAC_UP_CONV, bid, suffix=".bias"), get(f"decoder.decoder.{py}.block.1.conv.bias"))
+
+            for xid in range(3):
+                ridx = xid + 2  # block.2/3/4 are the 3 residual units
+
+                a1, b1 = snake_fold(get(f"decoder.decoder.{py}.block.{ridx}.act1.alpha"), get(f"decoder.decoder.{py}.block.{ridx}.act1.beta"))
+                name1 = gguf.TENSOR_NAMES[T.A_GEN_WAV_DAC_RES_ACT1].format(bid=bid, xid=xid)
+                yield (name1 + ".alpha", a1)
+                yield (name1 + ".beta", b1)
+
+                name_conv1 = gguf.TENSOR_NAMES[T.A_GEN_WAV_DAC_RES_CONV1].format(bid=bid, xid=xid)
+                yield (name_conv1 + ".weight", get(f"decoder.decoder.{py}.block.{ridx}.conv1.conv.weight"))
+                yield (name_conv1 + ".bias", get(f"decoder.decoder.{py}.block.{ridx}.conv1.conv.bias"))
+
+                a2, b2 = snake_fold(get(f"decoder.decoder.{py}.block.{ridx}.act2.alpha"), get(f"decoder.decoder.{py}.block.{ridx}.act2.beta"))
+                name2 = gguf.TENSOR_NAMES[T.A_GEN_WAV_DAC_RES_ACT2].format(bid=bid, xid=xid)
+                yield (name2 + ".alpha", a2)
+                yield (name2 + ".beta", b2)
+
+                name_conv2 = gguf.TENSOR_NAMES[T.A_GEN_WAV_DAC_RES_CONV2].format(bid=bid, xid=xid)
+                yield (name_conv2 + ".weight", get(f"decoder.decoder.{py}.block.{ridx}.conv2.conv.weight"))
+                yield (name_conv2 + ".bias", get(f"decoder.decoder.{py}.block.{ridx}.conv2.conv.bias"))
+
+        a5, b5 = snake_fold(get("decoder.decoder.5.alpha"), get("decoder.decoder.5.beta"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_SNAKE, suffix=".alpha"), a5)
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_SNAKE, suffix=".beta"), b5)
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_CONV, suffix=".weight"), get("decoder.decoder.6.conv.weight"))
+        yield (self.format_tensor_name(T.A_GEN_WAV_DAC_POST_CONV, suffix=".bias"), get("decoder.decoder.6.conv.bias"))
@@ -133,6 +133,7 @@ Note:
 - To debug the multimodal preprocessor and encoder, you can use [llama-mtmd-debug](tools/mtmd/debug/mtmd-debug.cpp).
 - Adding a model-specific API or CLI is an anti-pattern in `libmtmd`. The goal of `libmtmd` is to provide an easy-to-use, model-agnostic library for multimodal pipeline.
 - In most cases, `llama-mtmd-cli` should not be modified. If a model requires a specific prompt, either let the user provide it or bake it into the Jinja chat template.
+- For audio generation models, see `tools/mtmd/README-dev.md`

 ## Tips and tricks

@@ -23,16 +23,16 @@ Legend:
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                        COL2IM_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        COL2IM_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
@@ -51,8 +51,8 @@ Legend:
 |                             FILL | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |                            FLOOR | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | 🟡 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -60,14 +60,14 @@ Legend:
 |                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                LIGHTNING_INDEXER | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -76,13 +76,13 @@ Legend:
 |                 MUL_MAT_HADAMARD | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ❌ |
+|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | 🟡 |
+|                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 |
 |                              PAD | ❌ | 🟡 | ✅ | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -103,13 +103,13 @@ Legend:
 |                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -5,7 +5,7 @@ project("ggml" C CXX ASM)
 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
 set(GGML_VERSION_MINOR 18)
-set(GGML_VERSION_PATCH 0)
+set(GGML_VERSION_PATCH 1)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")

 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
@@ -127,7 +127,15 @@ static void concat_T_sycl_non_cont(
    int64_t ne2, int64_t ne3, uint64_t nb0, uint64_t nb1, uint64_t nb2,
    uint64_t nb3, int32_t dim) {
  sycl::range<3> gridDim(ne3, ne2, ne1);
-  stream->parallel_for(sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+
+  // Avoid oversubscribing device when there is not enough elements along the innermost dim to
+  // fill a full SYCL_CONCAT_BLOCK_SIZE. For larger # of elements, the full SYCL_CONCAT_BLOCK_SIZE
+  // is used.
+  const int64_t ne0_pad   = GGML_PAD(ne0, WARP_SIZE);
+  const int64_t block_ne0 = ne0_pad < SYCL_CONCAT_BLOCK_SIZE ? ne0_pad : (int64_t) SYCL_CONCAT_BLOCK_SIZE;
+  sycl::range<3> blockDim(1, 1, block_ne0);
+
+  stream->parallel_for(sycl::nd_range<3>(gridDim * blockDim, blockDim), [=](sycl::nd_item<3> item_ct1) {
      int64_t i3 = item_ct1.get_group(0);
      int64_t i2 = item_ct1.get_group(1);
      int64_t i1 = item_ct1.get_group(2);
@@ -2,11 +2,13 @@
 #include <cstdio>
 #include <cstring>
 #include <string>
+#include <optional>
 #include <unordered_map>
 #include <vector>

 #include "fattn-onednn.hpp"
 #include "fattn-tile.hpp"
+#include "convert.hpp"

 // set minimum query length to treat as prefill (32)
 #define GGML_SYCL_FA_ONEDNN_MIN_Q 32
@@ -33,10 +35,30 @@ bool ggml_sycl_flash_attn_ext_onednn_supported(const ggml_tensor * dst) {
    const ggml_tensor * mask  = dst->src[3];
    const ggml_tensor * sinks = dst->src[4];

-    // gate for f16 KV only for now
-    // need to implement quantized KV
+    // F16 KV: native SDPA at any KV length.
+    // Non-F16: dequant to F16 then SDPA at prefill lengths. Only the
+    // standard quantized KV cache types (Q4_0-Q8_0) and F32 are accepted
+    // because their to_fp16_sycl conversion is verified. BF16 and IQ*
+    // are excluded: BF16 needs a strided conversion kernel that does not
+    // exist yet; IQ types are model-weight-only quants with no dequant
+    // registration and are never used as KV caches.
    if (K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16) {
-        return false;
+        auto kt = K->type, vt = V->type;
+        bool k_ok = kt == GGML_TYPE_F32 || kt == GGML_TYPE_Q4_0 || kt == GGML_TYPE_Q4_1 ||
+                    kt == GGML_TYPE_Q5_0 || kt == GGML_TYPE_Q5_1 || kt == GGML_TYPE_Q8_0;
+        bool v_ok = vt == GGML_TYPE_F32 || vt == GGML_TYPE_Q4_0 || vt == GGML_TYPE_Q4_1 ||
+                    vt == GGML_TYPE_Q5_0 || vt == GGML_TYPE_Q5_1 || vt == GGML_TYPE_Q8_0;
+        if (!k_ok || !v_ok) {
+            return false;
+        }
+        if (Q->ne[1] < 32 || K->ne[1] < 1024) {
+            return false;
+        }
+        for (const ggml_tensor * t : {K, V}) {
+            if (t->type == GGML_TYPE_F16 && t->nb[1] % (t->ne[0] * 2) != 0) {
+                return false;
+            }
+        }
    }
    // Optional KV-length ceiling (GGML_SYCL_FA_ONEDNN_MAX_KV, 0 = unlimited). Escape hatch:
    // very long sequences make the fused SDPA slow enough to risk the xe driver watchdog on
@@ -205,13 +227,101 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso
    dnnl::engine    eng    = ctx.engine_dnnl(stream);
    dnnl::stream    strm   = ctx.stream_dnnl(stream);

-    // cont/cast inputs to contiguous f16 (head-major) -- the layout the fast systolic path wants.
-    ggml_sycl_pool_alloc<sycl::half> Qf(ctx.pool(), (size_t) H   * q   * d);
-    ggml_sycl_pool_alloc<sycl::half> Kf(ctx.pool(), (size_t) Hkv * seq * d);
-    ggml_sycl_pool_alloc<sycl::half> Vf(ctx.pool(), (size_t) Hkv * seq * d);
-    cont_to_f16_sycl<float>     ((const char *) Q->data, Qf.get(), d, q,   H,   mb, Q->nb[1], Q->nb[2], Q->nb[3], stream);
-    cont_to_f16_sycl<sycl::half>((const char *) K->data, Kf.get(), d, seq, Hkv, mb, K->nb[1], K->nb[2], K->nb[3], stream);
-    cont_to_f16_sycl<sycl::half>((const char *) V->data, Vf.get(), d, seq, Hkv, mb, V->nb[1], V->nb[2], V->nb[3], stream);
+    // Q: always f32 -- copy to dense f16.
+    ggml_sycl_pool_alloc<sycl::half> Qf(ctx.pool(), (size_t) H * q * d);
+    cont_to_f16_sycl<float>((const char *) Q->data, Qf.get(), d, q, H, mb, Q->nb[1], Q->nb[2], Q->nb[3], stream);
+
+    // K/V: use pool-alloc for both F16 and dequant paths.
+    sycl::half * K_ptr = nullptr;
+    sycl::half * V_ptr = nullptr;
+    std::optional<ggml_sycl_pool_alloc<sycl::half>> Kf_pool;
+    std::optional<ggml_sycl_pool_alloc<sycl::half>> Vf_pool;
+
+    if (K->type == GGML_TYPE_F16 && V->type == GGML_TYPE_F16) {
+        Kf_pool.emplace(ctx.pool(), (size_t) Hkv * seq * d);
+        Vf_pool.emplace(ctx.pool(), (size_t) Hkv * seq * d);
+        cont_to_f16_sycl<sycl::half>((const char *) K->data, Kf_pool->get(), d, seq, Hkv, mb, K->nb[1], K->nb[2], K->nb[3], stream);
+        cont_to_f16_sycl<sycl::half>((const char *) V->data, Vf_pool->get(), d, seq, Hkv, mb, V->nb[1], V->nb[2], V->nb[3], stream);
+        K_ptr = Kf_pool->get();
+        V_ptr = Vf_pool->get();
+    } else if (ggml_is_quantized(K->type)) {
+        // Quantized K/V: dequant to dense F16 using pool, same lifetime as F16 path.
+        Kf_pool.emplace(ctx.pool(), ggml_nelements(K));
+        K_ptr = Kf_pool->get();
+        {
+            const char * K_data = (const char *)K->data;
+            const bool k_non_dense = ((int64_t)K->ne[1] * K->nb[1] != K->nb[2]) && K->ne[2] > 1;
+            const bool k_gemma = k_non_dense &&
+                ((int64_t)K->nb[2] < (int64_t)K->ne[1] * (int64_t)K->nb[1]);
+            if (ggml_is_contiguously_allocated(K) && !k_non_dense) {
+                to_fp16_sycl_t to_fp16 = ggml_get_to_fp16_sycl(K->type, dst);
+                to_fp16(K_data, K_ptr, ggml_nelements(K), stream);
+            } else {
+                const size_t bs = ggml_blck_size(K->type);
+                const size_t ts = ggml_type_size(K->type);
+                to_fp16_nc_sycl_t to_fp16 = ggml_get_to_fp16_nc_sycl(K->type);
+                int64_t s01, s02, s03;
+                if (k_gemma) {
+                    const int64_t blk_per_row = (int64_t)K->ne[0] / bs;
+                    s01 = (int64_t)Hkv * blk_per_row;
+                    s02 = blk_per_row;
+                    s03 = (int64_t)K->ne[1] * s01;
+                } else {
+                    s01 = (int64_t)K->nb[1] / ts;
+                    s02 = (int64_t)K->nb[2] / ts;
+                    s03 = (int64_t)K->nb[3] / ts;
+                }
+                to_fp16(K_data, K_ptr,
+                        K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                        s01, s02, s03, stream);
+            }
+        }
+        // Quantized V: always dequant separately. Even when K and V share
+        // the same underlying allocation (V is a view of K with the same
+        // data pointer), their logical values differ because the quantized
+        // elements at different positions/offsets represent different K/V
+        // data. Master's F16 path also never aliases K and V.
+        Vf_pool.emplace(ctx.pool(), ggml_nelements(V));
+        V_ptr = Vf_pool->get();
+        {
+            const char * V_data = (const char *)V->data;
+            const bool v_non_dense = ((int64_t)V->ne[1] * V->nb[1] != V->nb[2]) && V->ne[2] > 1;
+            const bool v_gemma = v_non_dense &&
+                ((int64_t)V->nb[2] < (int64_t)V->ne[1] * (int64_t)V->nb[1]);
+            if (ggml_is_contiguously_allocated(V) && !v_non_dense) {
+                to_fp16_sycl_t to_fp16 = ggml_get_to_fp16_sycl(V->type, dst);
+                to_fp16(V_data, V_ptr, ggml_nelements(V), stream);
+            } else {
+                const size_t bs = ggml_blck_size(V->type);
+                const size_t ts = ggml_type_size(V->type);
+                to_fp16_nc_sycl_t to_fp16 = ggml_get_to_fp16_nc_sycl(V->type);
+                int64_t s01, s02, s03;
+                if (v_gemma) {
+                    const int64_t blk_per_row = (int64_t)V->ne[0] / bs;
+                    s01 = (int64_t)V->ne[2] * blk_per_row;
+                    s02 = blk_per_row;
+                    s03 = (int64_t)V->ne[1] * s01;
+                } else {
+                    s01 = (int64_t)V->nb[1] / ts;
+                    s02 = (int64_t)V->nb[2] / ts;
+                    s03 = (int64_t)V->nb[3] / ts;
+                }
+                to_fp16(V_data, V_ptr,
+                        V->ne[0], V->ne[1], V->ne[2], V->ne[3],
+                        s01, s02, s03, stream);
+            }
+        }
+    } else {
+        // F32: strided copy to dense F16 via cont_to_f16_sycl<float>.
+        Kf_pool.emplace(ctx.pool(), ggml_nelements(K));
+        K_ptr = Kf_pool->get();
+        cont_to_f16_sycl<float>((const char *) K->data, K_ptr, K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+                                K->nb[1], K->nb[2], K->nb[3], stream);
+        Vf_pool.emplace(ctx.pool(), ggml_nelements(V));
+        V_ptr = Vf_pool->get();
+        cont_to_f16_sycl<float>((const char *) V->data, V_ptr, V->ne[0], V->ne[1], V->ne[2], V->ne[3],
+                                V->nb[1], V->nb[2], V->nb[3], stream);
+    }

    // divide-by-(1/scale) reproduces ggml's score *= kq_scale on the proven probe graph.
    //
@@ -244,8 +354,8 @@ void ggml_sycl_flash_attn_ext_onednn(ggml_backend_sycl_context & ctx, ggml_tenso

    auto id2ptr = [&](size_t r) -> void * {
        if (r == E.id_q)     return Qf.get();
-        if (r == E.id_k)     return Kf.get();
-        if (r == E.id_v)     return Vf.get();
+        if (r == E.id_k)     return K_ptr;
+        if (r == E.id_v)     return V_ptr;
        if (r == E.id_scale) return scale_dev;
        if (r == E.id_mask)  return (void *) mask->data;
        return nullptr;
@@ -97,7 +97,7 @@ static void ggml_sycl_flash_attn_ext_vec(ggml_backend_sycl_context & ctx, ggml_t
 enum best_fattn_kernel {
    BEST_FATTN_KERNEL_NONE     =   0,
    BEST_FATTN_KERNEL_VEC      = 100,
-    BEST_FATTN_KERNEL_ONEDNN   = 150, // added enum for onednn==150
+    BEST_FATTN_KERNEL_ONEDNN   = 150, // oneDNN SDPA: native F16 (PR #25222)
    BEST_FATTN_KERNEL_TILE     = 200,
    BEST_FATTN_KERNEL_MKL      = 300,
 };
@@ -130,6 +130,14 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const

    bool gqa_opt_applies = gqa_ratio >= 2 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;

+    // XMX-accelerated path: oneDNN SDPA (native F16 and dequant+non-F16).
+    // ONEDNN requires min 32 query tokens — short-circuit decode to avoid
+    // calling _supported() on every decode FA call.
+    if (Q->ne[1] >= 32
+        && ggml_sycl_flash_attn_ext_onednn_supported(dst)) {
+        return BEST_FATTN_KERNEL_ONEDNN;
+    }
+
    // MKL path: XMX-accelerated GEMM for prompt processing (all KV cache types).
    // The MKL kernel converts non-F16 K/V to F16 via to_fp16_sycl before GEMM,
    // so quantized, F16, BF16, and F32 caches all benefit from XMX acceleration.
@@ -167,7 +175,6 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
            return BEST_FATTN_KERNEL_MKL;
        }
    }
-
    for (const ggml_tensor * t : {Q, K, V, mask}) {
        if (t == nullptr || ggml_is_quantized(t->type)) {
            continue;
@@ -215,6 +222,7 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
    switch (K->type) {
        case GGML_TYPE_F32:
        case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
            break;
        case GGML_TYPE_Q4_1:
        case GGML_TYPE_Q5_0:
@@ -233,8 +241,11 @@ static best_fattn_kernel ggml_sycl_get_best_fattn_kernel(const int device, const
        return BEST_FATTN_KERNEL_NONE;
    }

-    // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
-    const bool can_use_vector_kernel = Q->ne[0] <= 512 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes.
+    // BF16 is excluded: the VEC kernel has no BF16 template (it needs GGML_SYCL_FA_ALL_QUANTS for non-F16/Q4_0/Q8_0).
+    const bool has_bf16 = (K->type == GGML_TYPE_BF16 || V->type == GGML_TYPE_BF16);
+    const bool can_use_vector_kernel = Q->ne[0] <= 512 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0
+        && !has_bf16;

    // Fused-XMX path: oneDNN Graph SDPA (flash attention). Strictly
    // additive -- taken only when statically supported, otherwise falls through to VEC/TILE below.
@@ -276,6 +287,7 @@ void ggml_sycl_flash_attn_ext(ggml_backend_sycl_context & ctx, ggml_tensor * dst
        const char * kname = "TILE";
        best_fattn_kernel k = ggml_sycl_get_best_fattn_kernel(ctx.device, dst);
        if (k == BEST_FATTN_KERNEL_MKL)  kname = "MKL";
+        if (k == BEST_FATTN_KERNEL_ONEDNN)  kname = "ONEDNN";
        if (k == BEST_FATTN_KERNEL_VEC)  kname = "VEC";
        int64_t delta = 0;
        if (Dk == 256) {
@@ -292,7 +304,8 @@ void ggml_sycl_flash_attn_ext(ggml_backend_sycl_context & ctx, ggml_tensor * dst
                (long long)V_dbg->ne[1]);
    }

-    switch (ggml_sycl_get_best_fattn_kernel(ggml_sycl_get_device(), dst)) {
+    const best_fattn_kernel fk = ggml_sycl_get_best_fattn_kernel(ggml_sycl_get_device(), dst);
+    switch (fk) {
        case BEST_FATTN_KERNEL_NONE:
            GGML_ABORT("Not support Flash-Attention");
        case BEST_FATTN_KERNEL_ONEDNN:
@@ -331,6 +344,7 @@ void ggml_sycl_flash_attn_ext(ggml_backend_sycl_context & ctx, ggml_tensor * dst
            q->wait();
            const char * kname = "???";
            best_fattn_kernel kb = ggml_sycl_get_best_fattn_kernel(ctx.device, dst);
+            if (kb == BEST_FATTN_KERNEL_ONEDNN) kname = "ONEDNN";
            if (kb == BEST_FATTN_KERNEL_MKL) kname = "MKL";
            if (kb == BEST_FATTN_KERNEL_TILE) kname = "TILE";
            if (kb == BEST_FATTN_KERNEL_VEC) kname = "VEC";
@@ -354,6 +368,7 @@ void ggml_sycl_flash_attn_ext(ggml_backend_sycl_context & ctx, ggml_tensor * dst
            }
        }
    }
+
 }

 bool ggml_sycl_flash_attn_ext_supported(int device, const ggml_tensor * dst) {
@@ -1026,6 +1026,7 @@ struct vk_device_struct {
    vk_pipeline pipeline_pool2d_f32;
    vk_pipeline pipeline_rwkv_wkv6_f32;
    vk_pipeline pipeline_rwkv_wkv7_f32;
+    vk_pipeline pipeline_gated_linear_attn_f32;
    // [size_idx][kda] where size_idx: 0=d16, 1=d32, 2=d64, 3=d128
    vk_pipeline pipeline_gated_delta_net[4][2];
    vk_pipeline pipeline_ssm_scan_f32_d128;
@@ -1747,6 +1748,13 @@ struct vk_op_rwkv_wkv7_push_constants {
    uint32_t C;
    uint32_t H;
 };
+struct vk_op_gated_linear_attn_push_constants {
+    uint32_t B;
+    uint32_t T;
+    uint32_t C;
+    uint32_t H;
+    float scale;
+};
 struct vk_op_gated_delta_net_push_constants {
    uint32_t H;
    uint32_t n_tokens;
@@ -5665,6 +5673,8 @@ static void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {

    ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv7_f32, "rwkv_wkv7_f32", rwkv_wkv7_f32_len, rwkv_wkv7_f32_data, "main", 8, sizeof(vk_op_rwkv_wkv7_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);

+    ggml_vk_create_pipeline(device, device->pipeline_gated_linear_attn_f32, "gated_linear_attn_f32", gated_linear_attn_f32_len, gated_linear_attn_f32_data, "main", 6, sizeof(vk_op_gated_linear_attn_push_constants), {1, 1, 1}, {}, 1);
+
    {
        const uint32_t gdn_sizes[] = {16, 32, 64, 128};
        const char * gdn_names[][2] = {
@@ -11392,6 +11402,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
            return ctx->device->pipeline_rwkv_wkv7_f32;
        }
        return nullptr;
+    case GGML_OP_GATED_LINEAR_ATTN:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_gated_linear_attn_f32;
+        }
+        return nullptr;
    case GGML_OP_GATED_DELTA_NET:
        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
            const uint32_t S_v = dst->src[2]->ne[0];
@@ -12422,6 +12437,41 @@ static void ggml_vk_rwkv_wkv7(ggml_backend_vk_context * ctx, vk_context& subctx,
    );
 }

+static void ggml_vk_gated_linear_attn(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
+    const size_t seq_length = dst->src[0]->ne[2];
+    const size_t n_embed    = dst->ne[0];
+    const size_t n_heads    = dst->src[0]->ne[1];
+    const size_t n_seqs     = dst->src[4]->ne[1];
+
+    float scale;
+    memcpy(&scale, dst->op_params, sizeof(float));
+
+    GGML_ASSERT(dst->buffer != nullptr);
+
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, dst->src[0], dst->src[1], dst->src[2], dst, dst->op);
+    GGML_ASSERT(pipeline != nullptr);
+
+    ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+
+    vk_subbuffer dst_buf = ggml_vk_tensor_subbuffer(ctx, dst);
+    vk_subbuffer src_buf[5] = {};
+    for (int i = 0; i < 5; i++) {
+        src_buf[i] = ggml_vk_tensor_subbuffer(ctx, dst->src[i]);
+    }
+
+    const vk_op_gated_linear_attn_push_constants pc = {
+        (uint32_t)n_seqs,
+        (uint32_t)seq_length,
+        (uint32_t)n_embed,
+        (uint32_t)n_heads,
+        scale,
+    };
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+        {src_buf[0], src_buf[1], src_buf[2], src_buf[3], src_buf[4], dst_buf},
+        pc, { (uint32_t)(n_seqs * n_heads), 1, 1 });
+}
+
 static void ggml_vk_gated_delta_net(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst) {
    const ggml_tensor * src_q     = dst->src[0];
    const ggml_tensor * src_v     = dst->src[2];
@@ -15421,6 +15471,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr

        break;

+    case GGML_OP_GATED_LINEAR_ATTN:
+        ggml_vk_gated_linear_attn(ctx, compute_ctx, node);
+
+        break;
+
    case GGML_OP_GATED_DELTA_NET:
        ggml_vk_gated_delta_net(ctx, compute_ctx, node);

@@ -18128,6 +18183,9 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
        case GGML_OP_RWKV_WKV6:
        case GGML_OP_RWKV_WKV7:
            return true; // all inputs are contiguous, see ggml.c
+        case GGML_OP_GATED_LINEAR_ATTN:
+            // the shader block size is hardcoded to head_size 64
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 && op->src[0]->ne[0] == 64;
        case GGML_OP_GATED_DELTA_NET:
            {
                const uint32_t S_v = op->src[2]->ne[0];
@@ -19117,6 +19175,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
        } else if (tensor->op == GGML_OP_RWKV_WKV7) {
            tensor_clone = ggml_rwkv_wkv7(ggml_ctx, src_clone[0], src_clone[1], src_clone[2], src_clone[3],
            src_clone[4], src_clone[5], src_clone[6]);
+        } else if (tensor->op == GGML_OP_GATED_LINEAR_ATTN) {
+            const float * op_params = (const float *)tensor->op_params;
+            tensor_clone = ggml_gated_linear_attn(ggml_ctx, src_clone[0], src_clone[1],
+            src_clone[2], src_clone[3], src_clone[4], op_params[0]);
        } else if (tensor->op == GGML_OP_GATED_DELTA_NET) {
            tensor_clone = ggml_gated_delta_net(ggml_ctx, src_clone[0], src_clone[1],
            src_clone[2], src_clone[3], src_clone[4], src_clone[5],
@@ -0,0 +1,82 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : require
+
+#define BLOCK_SIZE 64
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout(push_constant) uniform Parameters {
+    uint B;
+    uint T;
+    uint C;
+    uint H;
+    float scale;
+};
+
+layout(binding = 0) readonly buffer KBuf { A_TYPE k[]; };
+layout(binding = 1) readonly buffer VBuf { A_TYPE v[]; };
+layout(binding = 2) readonly buffer QBuf { A_TYPE q[]; };
+layout(binding = 3) readonly buffer GBuf { A_TYPE g[]; };
+layout(binding = 4) readonly buffer StateBuf { A_TYPE state_in[]; };
+layout(binding = 5) buffer DstBuf { A_TYPE dst[]; };
+
+shared A_TYPE _k[BLOCK_SIZE], _q[BLOCK_SIZE], _g[BLOCK_SIZE];
+
+void main() {
+    const uint head_size = BLOCK_SIZE;
+    const uint batch_id = gl_WorkGroupID.x / H;
+    const uint head_id = gl_WorkGroupID.x % H;
+    const uint tid = gl_LocalInvocationID.x;
+
+    const uint state_size = C * head_size;
+    const uint n_seq_tokens = T / B;
+
+    if (batch_id >= B || head_id >= H) {
+        return;
+    }
+
+    // state[i] holds column tid of this head's state matrix: S[i][tid]
+    A_TYPE state[BLOCK_SIZE];
+    [[unroll]] for (uint i = 0; i < head_size; i++) {
+        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
+                          + i * head_size + tid];
+    }
+
+    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
+    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
+
+    for (uint t = start_t; t < end_t; t += C) {
+        barrier();
+        _k[tid] = k[t];
+        _q[tid] = q[t];
+        _g[tid] = g[t];
+        barrier();
+
+        const A_TYPE v_val = v[t];
+        A_TYPE y = 0.0;
+
+        [[unroll]] for (uint i = 0; i < head_size; i += 4) {
+            vec4 k_vec = vec4(_k[i], _k[i+1], _k[i+2], _k[i+3]);
+            vec4 q_vec = vec4(_q[i], _q[i+1], _q[i+2], _q[i+3]);
+            vec4 g_vec = vec4(_g[i], _g[i+1], _g[i+2], _g[i+3]);
+            vec4 s_vec = vec4(state[i], state[i+1], state[i+2], state[i+3]);
+
+            vec4 kv = k_vec * v_val;
+
+            s_vec = s_vec * g_vec + kv;
+            y += dot(q_vec, s_vec);
+
+            state[i]   = s_vec.x;
+            state[i+1] = s_vec.y;
+            state[i+2] = s_vec.z;
+            state[i+3] = s_vec.w;
+        }
+
+        dst[t] = y * scale;
+    }
+
+    [[unroll]] for (uint i = 0; i < head_size; i++) {
+        dst[T * C + batch_id * state_size + head_id * head_size * head_size
+            + i * head_size + tid] = state[i];
+    }
+}
@@ -1057,6 +1057,8 @@ void process_shaders() {

    string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));

+    string_to_spv("gated_linear_attn_f32", "gla.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
+
    string_to_spv("rwkv_wkv7_f32", "wkv7.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));

    string_to_spv("gated_delta_net_f32", "gated_delta_net.comp", merge_maps(base_dict, {{"FLOAT_TYPE", "float"}, {"USE_SUBGROUP_ADD", "1"}, {"USE_SUBGROUP_CLUSTERED", "1"}}));
@@ -11,6 +11,7 @@ GGUF_MAGIC             = 0x46554747  # "GGUF"
 GGUF_VERSION           = 3
 GGUF_DEFAULT_ALIGNMENT = 32
 GGML_QUANT_VERSION     = 2  # GGML_QNT_VERSION from ggml.h
+GGML_MAX_DIMS          = 4  # GGML_MAX_DIMS from ggml.h

 #
 # metadata keys
@@ -322,6 +323,7 @@ class Keys:
        PROJECTOR_TYPE        = "clip.projector_type"
        HAS_VISION_ENCODER    = "clip.has_vision_encoder"
        HAS_AUDIO_ENCODER     = "clip.has_audio_encoder"
+        HAS_GEN_AUDIO_ENCODER = "clip.has_gen_audio_encoder"
        HAS_LLAVA_PROJECTOR   = "clip.has_llava_projector"

    class ClipVision:
@@ -396,6 +398,18 @@ class Keys:
            DOWNSAMPLE_RATE = "clip.audio.projector.downsample_rate"
            HEAD_COUNT      = "clip.audio.projector.head_count"

+    class ClipGenAudio:
+        PROJECTOR_TYPE      = "clip.gen.audio.projector_type" # for mixed modality models
+        EMBEDDING_LENGTH    = "clip.gen.audio.embedding_length"
+        FEED_FORWARD_LENGTH = "clip.gen.audio.feed_forward_length"
+        BLOCK_COUNT         = "clip.gen.audio.block_count"
+        PROJECTION_DIM      = "clip.gen.audio.projection_dim"
+
+        class Attention:
+            HEAD_COUNT      = "clip.gen.audio.attention.head_count"
+            HEAD_COUNT_KV   = "clip.gen.audio.attention.head_count_kv"
+            LAYERNORM_EPS   = "clip.gen.audio.attention.layer_norm_epsilon"
+
    class Diffusion:
        SHIFT_LOGITS        = "diffusion.shift_logits"

@@ -557,6 +571,7 @@ class MODEL_ARCH(IntEnum):
    TALKIE           = auto()
    MELLUM           = auto()
    NANBEIGE         = auto()
+    QWEN3TTS         = auto()


 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -957,6 +972,65 @@ class MODEL_TENSOR(IntEnum):
    A_ENC_DOWNSAMPLE_CONV = auto() # mimo-audio-tokenizer: post-transformer downsample conv
    A_ENC_DOWNSAMPLE_NORM = auto() # mimo-audio-tokenizer: post-transformer downsample norm
    A_ENC_RVQ_CODEBOOK    = auto() # mimo-audio-tokenizer: residual vector quantizer codebook, per quantizer index
+    A_ENC_CONV_RES2       = auto() # qwen3tts
+    A_ENC_SE_CONV1        = auto() # qwen3tts
+    A_ENC_SE_CONV2        = auto() # qwen3tts
+    A_ENC_ASP_ATTN        = auto() # qwen3tts
+    A_ENC_ASP_TDNN        = auto() # qwen3tts
+    # qwen3tts code_predictor: predicts the remaining RVQ codebooks
+    A_GEN_CODE_PROJ_IN     = auto() # small_to_mtp_projection
+    A_GEN_CODE_EMBD        = auto() # per-codebook embedding table, merged 3D [n_codebooks, vocab, dim]
+    A_GEN_CODE_HEAD        = auto() # per-codebook output head, merged 3D [n_codebooks, vocab, dim]
+    A_GEN_CODE_OUT_EMBD    = auto() # codebook-0 embedding, re-fed into the talker backbone (talker.model.codec_embedding)
+    A_GEN_CODE_ATTN_NORM   = auto()
+    A_GEN_CODE_ATTN_Q      = auto()
+    A_GEN_CODE_ATTN_Q_NORM = auto()
+    A_GEN_CODE_ATTN_K      = auto()
+    A_GEN_CODE_ATTN_K_NORM = auto()
+    A_GEN_CODE_ATTN_V      = auto()
+    A_GEN_CODE_ATTN_OUT    = auto()
+    A_GEN_CODE_FFN_NORM    = auto()
+    A_GEN_CODE_FFN_GATE    = auto()
+    A_GEN_CODE_FFN_UP      = auto()
+    A_GEN_CODE_FFN_DOWN    = auto()
+    A_GEN_CODE_OUTPUT_NORM = auto()
+    # qwen3tts code2wav: RVQ codes -> raw PCM
+    A_GEN_WAV_QUANT_FIRST_IN       = auto() # semantic RVQ, in_proj (1x1 conv, loaded as 2D)
+    A_GEN_WAV_QUANT_FIRST_OUT      = auto() # semantic RVQ, out_proj
+    A_GEN_WAV_QUANT_FIRST_CB = auto() # semantic RVQ codebook (1 layer), folded from embedding_sum/cluster_usage
+    A_GEN_WAV_QUANT_REST_IN        = auto() # acoustic RVQ, in_proj
+    A_GEN_WAV_QUANT_REST_OUT       = auto() # acoustic RVQ, out_proj
+    A_GEN_WAV_QUANT_REST_CB  = auto() # acoustic RVQ codebooks, merged 3D [15, vocab, dim]
+    A_GEN_WAV_PRE_CONV             = auto()
+    A_GEN_WAV_TFM_IN_PROJ          = auto()
+    A_GEN_WAV_TFM_OUT_PROJ         = auto()
+    A_GEN_WAV_TFM_OUTPUT_NORM      = auto()
+    A_GEN_WAV_TFM_ATTN_NORM        = auto()
+    A_GEN_WAV_TFM_ATTN_Q           = auto()
+    A_GEN_WAV_TFM_ATTN_K           = auto()
+    A_GEN_WAV_TFM_ATTN_V           = auto()
+    A_GEN_WAV_TFM_ATTN_OUT         = auto()
+    A_GEN_WAV_TFM_ATTN_SCALE       = auto() # layer scale (gamma) on the attn output
+    A_GEN_WAV_TFM_FFN_NORM         = auto()
+    A_GEN_WAV_TFM_FFN_GATE         = auto()
+    A_GEN_WAV_TFM_FFN_UP           = auto()
+    A_GEN_WAV_TFM_FFN_DOWN         = auto()
+    A_GEN_WAV_TFM_FFN_SCALE        = auto() # layer scale (gamma) on the FFN output
+    A_GEN_WAV_UP_CONV              = auto() # causal ConvTranspose1d, 2x upsample
+    A_GEN_WAV_UP_DWCONV            = auto() # ConvNeXt depthwise conv
+    A_GEN_WAV_UP_NORM              = auto() # ConvNeXt LayerNorm
+    A_GEN_WAV_UP_PW1               = auto() # ConvNeXt pointwise conv 1 (expand)
+    A_GEN_WAV_UP_PW2               = auto() # ConvNeXt pointwise conv 2 (project)
+    A_GEN_WAV_UP_GAMMA             = auto() # ConvNeXt layer scale
+    A_GEN_WAV_DAC_ENTRY            = auto() # DAC conv_pre
+    A_GEN_WAV_DAC_UP_SNAKE         = auto() # DAC per-block SnakeBeta before the upsample conv
+    A_GEN_WAV_DAC_UP_CONV          = auto() # DAC per-block causal ConvTranspose1d
+    A_GEN_WAV_DAC_RES_ACT1         = auto() # DAC residual unit, SnakeBeta before conv1
+    A_GEN_WAV_DAC_RES_CONV1        = auto() # DAC residual unit, dilated causal conv
+    A_GEN_WAV_DAC_RES_ACT2         = auto() # DAC residual unit, SnakeBeta before conv2
+    A_GEN_WAV_DAC_RES_CONV2        = auto() # DAC residual unit, pointwise causal conv
+    A_GEN_WAV_DAC_POST_SNAKE       = auto() # DAC final SnakeBeta
+    A_GEN_WAV_DAC_POST_CONV        = auto() # DAC conv_post -> 1-channel PCM
    A_MMPROJ              = auto()
    A_MMPROJ_FC           = auto()
    A_MM_NORM_PRE         = auto()
@@ -1169,6 +1243,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
    MODEL_ARCH.TALKIE:           "talkie",
    MODEL_ARCH.MELLUM:           "mellum",
    MODEL_ARCH.NANBEIGE:         "nanbeige",
+    MODEL_ARCH.QWEN3TTS:         "qwen3tts",
 }

 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -1566,6 +1641,63 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
    MODEL_TENSOR.A_ENC_DOWNSAMPLE_CONV:     "a.downsample.conv",
    MODEL_TENSOR.A_ENC_DOWNSAMPLE_NORM:     "a.downsample.norm",
    MODEL_TENSOR.A_ENC_RVQ_CODEBOOK:        "a.rvq.codebook",
+    MODEL_TENSOR.A_ENC_CONV_RES2:           "a.blk.{bid}.res2.{xid}",
+    MODEL_TENSOR.A_ENC_SE_CONV1:            "a.blk.{bid}.se_conv1",
+    MODEL_TENSOR.A_ENC_SE_CONV2:            "a.blk.{bid}.se_conv2",
+    MODEL_TENSOR.A_ENC_ASP_ATTN:            "a.asp_attn",
+    MODEL_TENSOR.A_ENC_ASP_TDNN:            "a.asp_tdnn",
+    MODEL_TENSOR.A_GEN_CODE_PROJ_IN:        "a.gen.code.proj_in",
+    MODEL_TENSOR.A_GEN_CODE_EMBD:           "a.gen.code.embd",
+    MODEL_TENSOR.A_GEN_CODE_HEAD:           "a.gen.code.head",
+    MODEL_TENSOR.A_GEN_CODE_OUT_EMBD:       "a.gen.code.out_embd",
+    MODEL_TENSOR.A_GEN_CODE_ATTN_NORM:      "a.gen.code.blk.{bid}.ln1", # reuses the generic clip.cpp block loader (TN_LN_1)
+    MODEL_TENSOR.A_GEN_CODE_ATTN_Q:         "a.gen.code.blk.{bid}.attn_q",
+    MODEL_TENSOR.A_GEN_CODE_ATTN_Q_NORM:    "a.gen.code.blk.{bid}.attn_q_norm",
+    MODEL_TENSOR.A_GEN_CODE_ATTN_K:         "a.gen.code.blk.{bid}.attn_k",
+    MODEL_TENSOR.A_GEN_CODE_ATTN_K_NORM:    "a.gen.code.blk.{bid}.attn_k_norm",
+    MODEL_TENSOR.A_GEN_CODE_ATTN_V:         "a.gen.code.blk.{bid}.attn_v",
+    MODEL_TENSOR.A_GEN_CODE_ATTN_OUT:       "a.gen.code.blk.{bid}.attn_out",
+    MODEL_TENSOR.A_GEN_CODE_FFN_NORM:       "a.gen.code.blk.{bid}.ln2", # reuses the generic clip.cpp block loader (TN_LN_2)
+    MODEL_TENSOR.A_GEN_CODE_FFN_GATE:       "a.gen.code.blk.{bid}.ffn_gate",
+    MODEL_TENSOR.A_GEN_CODE_FFN_UP:         "a.gen.code.blk.{bid}.ffn_up",
+    MODEL_TENSOR.A_GEN_CODE_FFN_DOWN:       "a.gen.code.blk.{bid}.ffn_down",
+    MODEL_TENSOR.A_GEN_CODE_OUTPUT_NORM:    "a.gen.code.output_norm",
+    MODEL_TENSOR.A_GEN_WAV_QUANT_FIRST_IN:  "a.gen.wav.quant.first.in_proj",
+    MODEL_TENSOR.A_GEN_WAV_QUANT_FIRST_OUT: "a.gen.wav.quant.first.out_proj",
+    MODEL_TENSOR.A_GEN_WAV_QUANT_FIRST_CB:  "a.gen.wav.quant.first.codebook",
+    MODEL_TENSOR.A_GEN_WAV_QUANT_REST_IN:   "a.gen.wav.quant.rest.in_proj",
+    MODEL_TENSOR.A_GEN_WAV_QUANT_REST_OUT:  "a.gen.wav.quant.rest.out_proj",
+    MODEL_TENSOR.A_GEN_WAV_QUANT_REST_CB:   "a.gen.wav.quant.rest.codebook",
+    MODEL_TENSOR.A_GEN_WAV_PRE_CONV:        "a.gen.wav.pre_conv",
+    MODEL_TENSOR.A_GEN_WAV_TFM_IN_PROJ:     "a.gen.wav.tfm.in_proj",
+    MODEL_TENSOR.A_GEN_WAV_TFM_OUT_PROJ:    "a.gen.wav.tfm.out_proj",
+    MODEL_TENSOR.A_GEN_WAV_TFM_OUTPUT_NORM: "a.gen.wav.tfm.output_norm",
+    MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_NORM:   "a.gen.wav.tfm.blk.{bid}.ln1",
+    MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_Q:      "a.gen.wav.tfm.blk.{bid}.attn_q",
+    MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_K:      "a.gen.wav.tfm.blk.{bid}.attn_k",
+    MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_V:      "a.gen.wav.tfm.blk.{bid}.attn_v",
+    MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_OUT:    "a.gen.wav.tfm.blk.{bid}.attn_out",
+    MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_SCALE:  "a.gen.wav.tfm.blk.{bid}.ls1",
+    MODEL_TENSOR.A_GEN_WAV_TFM_FFN_NORM:    "a.gen.wav.tfm.blk.{bid}.ln2",
+    MODEL_TENSOR.A_GEN_WAV_TFM_FFN_GATE:    "a.gen.wav.tfm.blk.{bid}.ffn_gate",
+    MODEL_TENSOR.A_GEN_WAV_TFM_FFN_UP:      "a.gen.wav.tfm.blk.{bid}.ffn_up",
+    MODEL_TENSOR.A_GEN_WAV_TFM_FFN_DOWN:    "a.gen.wav.tfm.blk.{bid}.ffn_down",
+    MODEL_TENSOR.A_GEN_WAV_TFM_FFN_SCALE:   "a.gen.wav.tfm.blk.{bid}.ls2",
+    MODEL_TENSOR.A_GEN_WAV_UP_CONV:         "a.gen.wav.up.blk.{bid}.conv",
+    MODEL_TENSOR.A_GEN_WAV_UP_DWCONV:       "a.gen.wav.up.blk.{bid}.dwconv",
+    MODEL_TENSOR.A_GEN_WAV_UP_NORM:         "a.gen.wav.up.blk.{bid}.norm",
+    MODEL_TENSOR.A_GEN_WAV_UP_PW1:          "a.gen.wav.up.blk.{bid}.pw1",
+    MODEL_TENSOR.A_GEN_WAV_UP_PW2:          "a.gen.wav.up.blk.{bid}.pw2",
+    MODEL_TENSOR.A_GEN_WAV_UP_GAMMA:        "a.gen.wav.up.blk.{bid}.gamma",
+    MODEL_TENSOR.A_GEN_WAV_DAC_ENTRY:       "a.gen.wav.dac.entry",
+    MODEL_TENSOR.A_GEN_WAV_DAC_UP_SNAKE:    "a.gen.wav.dac.blk.{bid}.snake",
+    MODEL_TENSOR.A_GEN_WAV_DAC_UP_CONV:     "a.gen.wav.dac.blk.{bid}.conv",
+    MODEL_TENSOR.A_GEN_WAV_DAC_RES_ACT1:    "a.gen.wav.dac.blk.{bid}.res.{xid}.act1",
+    MODEL_TENSOR.A_GEN_WAV_DAC_RES_CONV1:   "a.gen.wav.dac.blk.{bid}.res.{xid}.conv1",
+    MODEL_TENSOR.A_GEN_WAV_DAC_RES_ACT2:    "a.gen.wav.dac.blk.{bid}.res.{xid}.act2",
+    MODEL_TENSOR.A_GEN_WAV_DAC_RES_CONV2:   "a.gen.wav.dac.blk.{bid}.res.{xid}.conv2",
+    MODEL_TENSOR.A_GEN_WAV_DAC_POST_SNAKE:  "a.gen.wav.dac.post_snake",
+    MODEL_TENSOR.A_GEN_WAV_DAC_POST_CONV:   "a.gen.wav.dac.post_conv",
    MODEL_TENSOR.A_MMPROJ:                  "mm.a.mlp.{bid}",
    MODEL_TENSOR.A_MMPROJ_FC:               "mm.a.fc",
    MODEL_TENSOR.A_MM_NORM_PRE:             "mm.a.norm_pre",
@@ -1820,6 +1952,63 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.A_ENC_CONV_NORM,
        MODEL_TENSOR.A_ENC_CONV_PW1,
        MODEL_TENSOR.A_ENC_CONV_PW2,
+        MODEL_TENSOR.A_ENC_CONV_RES2,
+        MODEL_TENSOR.A_ENC_SE_CONV1,
+        MODEL_TENSOR.A_ENC_SE_CONV2,
+        MODEL_TENSOR.A_ENC_ASP_ATTN,
+        MODEL_TENSOR.A_ENC_ASP_TDNN,
+        MODEL_TENSOR.A_GEN_CODE_PROJ_IN,
+        MODEL_TENSOR.A_GEN_CODE_EMBD,
+        MODEL_TENSOR.A_GEN_CODE_HEAD,
+        MODEL_TENSOR.A_GEN_CODE_OUT_EMBD,
+        MODEL_TENSOR.A_GEN_CODE_ATTN_NORM,
+        MODEL_TENSOR.A_GEN_CODE_ATTN_Q,
+        MODEL_TENSOR.A_GEN_CODE_ATTN_Q_NORM,
+        MODEL_TENSOR.A_GEN_CODE_ATTN_K,
+        MODEL_TENSOR.A_GEN_CODE_ATTN_K_NORM,
+        MODEL_TENSOR.A_GEN_CODE_ATTN_V,
+        MODEL_TENSOR.A_GEN_CODE_ATTN_OUT,
+        MODEL_TENSOR.A_GEN_CODE_FFN_NORM,
+        MODEL_TENSOR.A_GEN_CODE_FFN_GATE,
+        MODEL_TENSOR.A_GEN_CODE_FFN_UP,
+        MODEL_TENSOR.A_GEN_CODE_FFN_DOWN,
+        MODEL_TENSOR.A_GEN_CODE_OUTPUT_NORM,
+        MODEL_TENSOR.A_GEN_WAV_QUANT_FIRST_IN,
+        MODEL_TENSOR.A_GEN_WAV_QUANT_FIRST_OUT,
+        MODEL_TENSOR.A_GEN_WAV_QUANT_FIRST_CB,
+        MODEL_TENSOR.A_GEN_WAV_QUANT_REST_IN,
+        MODEL_TENSOR.A_GEN_WAV_QUANT_REST_OUT,
+        MODEL_TENSOR.A_GEN_WAV_QUANT_REST_CB,
+        MODEL_TENSOR.A_GEN_WAV_PRE_CONV,
+        MODEL_TENSOR.A_GEN_WAV_TFM_IN_PROJ,
+        MODEL_TENSOR.A_GEN_WAV_TFM_OUT_PROJ,
+        MODEL_TENSOR.A_GEN_WAV_TFM_OUTPUT_NORM,
+        MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_NORM,
+        MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_Q,
+        MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_K,
+        MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_V,
+        MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_OUT,
+        MODEL_TENSOR.A_GEN_WAV_TFM_ATTN_SCALE,
+        MODEL_TENSOR.A_GEN_WAV_TFM_FFN_NORM,
+        MODEL_TENSOR.A_GEN_WAV_TFM_FFN_GATE,
+        MODEL_TENSOR.A_GEN_WAV_TFM_FFN_UP,
+        MODEL_TENSOR.A_GEN_WAV_TFM_FFN_DOWN,
+        MODEL_TENSOR.A_GEN_WAV_TFM_FFN_SCALE,
+        MODEL_TENSOR.A_GEN_WAV_UP_CONV,
+        MODEL_TENSOR.A_GEN_WAV_UP_DWCONV,
+        MODEL_TENSOR.A_GEN_WAV_UP_NORM,
+        MODEL_TENSOR.A_GEN_WAV_UP_PW1,
+        MODEL_TENSOR.A_GEN_WAV_UP_PW2,
+        MODEL_TENSOR.A_GEN_WAV_UP_GAMMA,
+        MODEL_TENSOR.A_GEN_WAV_DAC_ENTRY,
+        MODEL_TENSOR.A_GEN_WAV_DAC_UP_SNAKE,
+        MODEL_TENSOR.A_GEN_WAV_DAC_UP_CONV,
+        MODEL_TENSOR.A_GEN_WAV_DAC_RES_ACT1,
+        MODEL_TENSOR.A_GEN_WAV_DAC_RES_CONV1,
+        MODEL_TENSOR.A_GEN_WAV_DAC_RES_ACT2,
+        MODEL_TENSOR.A_GEN_WAV_DAC_RES_CONV2,
+        MODEL_TENSOR.A_GEN_WAV_DAC_POST_SNAKE,
+        MODEL_TENSOR.A_GEN_WAV_DAC_POST_CONV,
        MODEL_TENSOR.A_ENC_CONV_NORM_MEAN,
        MODEL_TENSOR.A_ENC_CONV_NORM_VAR,
        MODEL_TENSOR.A_ENC_MEL_FILTERS,
@@ -4647,6 +4836,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.FFN_DOWN,
        MODEL_TENSOR.FFN_UP,
    ],
+    MODEL_ARCH.QWEN3TTS: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
 }

 # tensors that will not be serialized
@@ -4921,6 +5126,8 @@ class VisionProjectorType:
    GLM4V = "glm4v"
    YOUTUVL = "youtuvl"
    NEMOTRON_V2_VL = "nemotron_v2_vl"
+    QWEN3TTS_SPKENC = "qwen3tts_spkenc" # audio: ECAPA-TDNN speaker encoder
+    QWEN3TTS_GEN = "qwen3tts_gen" # audio generation: code_predictor
    HUNYUANVL      = "hunyuanvl"
    PARAKEET       = "parakeet"  # audio
    MINIMAXM3      = "minimax_m3"
@@ -22,6 +22,7 @@ if __name__ == "__main__":
    sys.path.insert(0, str(Path(__file__).parent.parent))

 from gguf.constants import (
+    GGML_MAX_DIMS,
    GGML_QUANT_SIZES,
    GGUF_DEFAULT_ALIGNMENT,
    GGUF_MAGIC,
@@ -266,6 +267,8 @@ class GGUFReader:
        # Get Tensor Dimensions Count
        n_dims = self._get(offs, np.uint32)
        offs += int(n_dims.nbytes)
+        if n_dims[0] > GGML_MAX_DIMS:
+            raise ValueError(f'Tensor dimensions count {n_dims[0]} exceeds GGML_MAX_DIMS ({GGML_MAX_DIMS})')

        # Get Tensor Dimension Array
        dims = self._get(offs, np.uint64, n_dims[0])
@@ -326,7 +329,10 @@ class GGUFReader:
                raise ValueError(f'Found duplicated tensor with name {tensor_name}')
            tensor_names.add(tensor_name)
            ggml_type = GGMLQuantizationType(raw_dtype[0])
-            n_elems = int(np.prod(dims))
+            # use Python ints: np.prod on uint64 wraps silently on overflow
+            n_elems = 1
+            for dim in dims.tolist():
+                n_elems *= int(dim)
            np_dims = tuple(reversed(dims.tolist()))
            block_size, type_size = GGML_QUANT_SIZES[ggml_type]
            n_bytes = n_elems * type_size // block_size
@@ -280,6 +280,10 @@ class GGUFWriter:

        self.kv_data[0][key] = GGUFValue(value=val, type=vtype, sub_type=sub_type)

+    def remove_key(self, key: str) -> None:
+        for kv_data in self.kv_data:
+            kv_data.pop(key, None)
+
    def add_uint8(self, key: str, val: int) -> None:
        self.add_key_value(key,val, GGUFValueType.UINT8)

@@ -1144,7 +1148,11 @@ class GGUFWriter:
    def add_precompiled_charsmap(self, charsmap: bytes) -> None:
        self.add_array(Keys.Tokenizer.PRECOMPILED_CHARSMAP, charsmap)

-    def add_chat_template(self, value: str | Sequence[Mapping[str, str]]) -> None:
+    def add_chat_template(self, value: str | Sequence[Mapping[str, str]] | None) -> None:
+        if value is None:
+            self.remove_key(Keys.Tokenizer.CHAT_TEMPLATE)
+            return
+
        if not isinstance(value, str):
            template_default = None
            template_names = set()
@@ -1199,6 +1207,9 @@ class GGUFWriter:
    def add_clip_has_audio_encoder(self, value: bool) -> None:
        self.add_bool(Keys.Clip.HAS_AUDIO_ENCODER, value)

+    def add_clip_has_gen_audio_encoder(self, value: bool) -> None:
+        self.add_bool(Keys.Clip.HAS_GEN_AUDIO_ENCODER, value)
+
    def add_clip_projector_type(self, value: str) -> None:
        self.add_string(Keys.Clip.PROJECTOR_TYPE, value)

@@ -1401,6 +1412,32 @@ class GGUFWriter:
    def add_audio_projector_head_count(self, value: int) -> None:
        self.add_uint32(Keys.ClipAudio.Projector.HEAD_COUNT, value)

+    # audio generation (mmproj)
+
+    def add_clip_gen_audio_projector_type(self, value: str) -> None:
+        self.add_string(Keys.ClipGenAudio.PROJECTOR_TYPE, value)
+
+    def add_gen_audio_projection_dim(self, value: int) -> None:
+        self.add_uint32(Keys.ClipGenAudio.PROJECTION_DIM, value)
+
+    def add_gen_audio_embedding_length(self, value: int) -> None:
+        self.add_uint32(Keys.ClipGenAudio.EMBEDDING_LENGTH, value)
+
+    def add_gen_audio_feed_forward_length(self, value: int) -> None:
+        self.add_uint32(Keys.ClipGenAudio.FEED_FORWARD_LENGTH, value)
+
+    def add_gen_audio_block_count(self, value: int) -> None:
+        self.add_uint32(Keys.ClipGenAudio.BLOCK_COUNT, value)
+
+    def add_gen_audio_head_count(self, value: int) -> None:
+        self.add_uint32(Keys.ClipGenAudio.Attention.HEAD_COUNT, value)
+
+    def add_gen_audio_head_count_kv(self, value: int) -> None:
+        self.add_uint32(Keys.ClipGenAudio.Attention.HEAD_COUNT_KV, value)
+
+    def add_gen_audio_attention_layernorm_eps(self, value: float) -> None:
+        self.add_float32(Keys.ClipGenAudio.Attention.LAYERNORM_EPS, value)
+
    def add_xielu_alpha_p(self, values: Sequence[float]):
        self.add_array(Keys.xIELU.ALPHA_P, values)

@@ -59,11 +59,29 @@ def byteswap_q6_k(tensor, block_offs):
    delta.byteswap(inplace=True)


+def byteswap_q1_0(tensor, block_offs):
+    # Each block_q1_0 consists of an f16 delta followed by 16 int8 quantizations.
+
+    # Byte-Swap f16 sized delta field
+    delta = tensor.data[block_offs:block_offs + 2].view(dtype=np.uint16)
+    delta.byteswap(inplace=True)
+
+
+def byteswap_tq2_0(tensor, block_offs):
+    # Each block_tq2_0 consists of 64 int8 values followed by 1 f16 value.
+
+    # Byte-Swap f16 sized field
+    delta = tensor.data[block_offs + 64:block_offs + 66].view(dtype=np.uint16)
+    delta.byteswap(inplace=True)
+
+
 byteswap_tensors = {
+    gguf.GGMLQuantizationType.Q1_0:  byteswap_q1_0,
    gguf.GGMLQuantizationType.Q4_0:  byteswap_q4_0,
    gguf.GGMLQuantizationType.Q8_0:  byteswap_q8_0,
    gguf.GGMLQuantizationType.Q4_K:  byteswap_q4_k,
    gguf.GGMLQuantizationType.Q6_K:  byteswap_q6_k,
+    gguf.GGMLQuantizationType.TQ2_0: byteswap_tq2_0,
    gguf.GGMLQuantizationType.MXFP4: byteswap_noop,
    gguf.GGMLQuantizationType.NVFP4: byteswap_noop,
 }
@@ -2109,6 +2109,7 @@ class TensorNameMap:
            "conformer.subsample_conv_projection.layer{bid}.conv", # gemma4
            "sound_encoder.encoder.subsampling.layers.{bid}", # parakeet
            "encoder.conv{bid}", # mimo-audio-tokenizer
+            "speaker_encoder.blocks.{bid}.conv", # qwen3tts speaker encoder (only bid=0, the stem TDNN)
        ),

        MODEL_TENSOR.A_ENC_CONV1D_NORM: (
@@ -2126,6 +2127,7 @@ class TensorNameMap:

        MODEL_TENSOR.A_ENC_CONV_OUT: (
            "audio_tower.conv_out", # qwen3omni
+            "speaker_encoder.mfa.conv", # qwen3tts speaker encoder: multi-layer feature aggregation
        ),

        MODEL_TENSOR.A_PRE_NORM: (),
@@ -2336,7 +2338,8 @@ class TensorNameMap:
        MODEL_TENSOR.A_MMPROJ_FC: (
            "audio.multi_modal_projector.linear", # qwen2audio
            "audio_tower.proj", # qwen2omni
-            "model.audio_tower.output_proj" # gemma4
+            "model.audio_tower.output_proj", # gemma4
+            "speaker_encoder.fc", # qwen3tts speaker encoder: final speaker embedding projection
        ),

        MODEL_TENSOR.A_MM_NORM_PRE: (
@@ -2411,6 +2414,7 @@ class TensorNameMap:
            "conformer.layers.{bid}.lconv1d.linear_start", # gemma3n
            "sound_encoder.encoder.layers.{bid}.conv.pointwise_conv1", # parakeet
            "encoder.layers.{bid}.conv.up_conv", # granite_speech
+            "speaker_encoder.blocks.{bid}.tdnn1.conv", # qwen3tts speaker encoder
        ),

        MODEL_TENSOR.A_ENC_CONV_PW2: (
@@ -2418,6 +2422,23 @@ class TensorNameMap:
            "conformer.layers.{bid}.lconv1d.linear_end", # gemma3n
            "sound_encoder.encoder.layers.{bid}.conv.pointwise_conv2", # parakeet
            "encoder.layers.{bid}.conv.down_conv", # granite_speech
+            "speaker_encoder.blocks.{bid}.tdnn2.conv", # qwen3tts speaker encoder
+        ),
+
+        MODEL_TENSOR.A_ENC_SE_CONV1: (
+            "speaker_encoder.blocks.{bid}.se_block.conv1", # qwen3tts
+        ),
+
+        MODEL_TENSOR.A_ENC_SE_CONV2: (
+            "speaker_encoder.blocks.{bid}.se_block.conv2", # qwen3tts
+        ),
+
+        MODEL_TENSOR.A_ENC_ASP_ATTN: (
+            "speaker_encoder.asp.conv", # qwen3tts
+        ),
+
+        MODEL_TENSOR.A_ENC_ASP_TDNN: (
+            "speaker_encoder.asp.tdnn.conv", # qwen3tts
        ),

        MODEL_TENSOR.A_ENC_NORM_CONV: (
@@ -0,0 +1,37 @@
+import struct
+import numpy as np
+import pytest
+
+from gguf.gguf_reader import GGUFReader
+
+
+def _write_gguf(path, n_dims_field, dims):
+    buf = b'GGUF' + struct.pack('<IQQ', 3, 1, 0)  # version 3, 1 tensor, 0 kv
+    name = b'bad_tensor'
+    buf += struct.pack('<Q', len(name)) + name
+    buf += struct.pack('<I', n_dims_field)
+    for d in dims:
+        buf += struct.pack('<Q', d)
+    buf += struct.pack('<I', 0)  # dtype F32
+    buf += struct.pack('<Q', 0)  # tensor offset
+    buf += b'\x00' * 64
+    path.write_bytes(buf)
+
+
+def test_n_dims_upper_bound(tmp_path):
+    # crafted file claims 1_000_000 dims; must be rejected, not read past EOF
+    p = tmp_path / 'evil_ndims.gguf'
+    _write_gguf(p, 1_000_000, [1] * 8)
+    with pytest.raises(ValueError, match='exceeds GGML_MAX_DIMS'):
+        GGUFReader(p)
+
+
+def test_dims_product_no_uint64_wraparound(tmp_path):
+    # dims whose true product overflows uint64; np.prod would wrap to 4 and
+    # silently pass an undersized read. The reader must not accept it.
+    dims = [4194305, 4194305, 211106198978564]
+    assert int(np.prod(np.array(dims, dtype=np.uint64))) == 4  # the wrap bug
+    p = tmp_path / 'evil_overflow.gguf'
+    _write_gguf(p, len(dims), dims)
+    with pytest.raises(ValueError):
+        GGUFReader(p)
@@ -1256,7 +1256,6 @@ extern "C" {
        struct ggml_tensor * probs;
        struct ggml_tensor * sampled;
        struct ggml_tensor * candidates;
-        int64_t              n_vocab;
    };

    // user code can implement the interface below in order to create custom llama_sampler
@@ -1425,7 +1424,8 @@ extern "C" {

    /// NOTE: Avoid using on the full vocabulary as searching for repeated tokens can become slow. For example, apply top-k or top-p sampling first.
    LLAMA_API struct llama_sampler * llama_sampler_init_penalties(
-                             int32_t   penalty_last_n,   // last n tokens to penalize (0 = disable penalty, -1 = context size)
+                             int32_t   n_vocab,
+                             int32_t   penalty_last_n,   // last n tokens to penalize (0 = disable penalty)
                               float   penalty_repeat,   // must be > 0.0, 1.0 = disabled
                               float   penalty_freq,     // must be finite, 0.0 = disabled
                               float   penalty_present); // must be finite, 0.0 = disabled
@@ -1433,11 +1433,10 @@ extern "C" {
    ///  @details DRY sampler, designed by p-e-w, as described in: https://github.com/oobabooga/text-generation-webui/pull/5677, porting Koboldcpp implementation authored by pi6am: https://github.com/LostRuins/koboldcpp/pull/982
    LLAMA_API struct llama_sampler * llama_sampler_init_dry(
            const struct llama_vocab *  vocab,
-                             int32_t    n_ctx_train,
                               float    dry_multiplier,
                               float    dry_base,
                             int32_t    dry_allowed_length,
-                             int32_t    dry_penalty_last_n,
+                             int32_t    dry_penalty_last_n, // last n tokens to penalize (0 = disable penalty)
                          const char ** seq_breakers,
                              size_t    num_breakers);

@@ -1 +1 @@
-06ca97616793248fadb410ea8d69c7511b2005e4
+90951f99af1fbebef3fbdd58ff5b8715b0bb9c43
@@ -24,10 +24,31 @@ vendor = {
    "https://raw.githubusercontent.com/sheredom/subprocess.h/8671cee1fc09f11a70ce3782a0ee13177c3aa387/subprocess.h": "vendor/sheredom/subprocess.h",
 }

+# TODO @ngxson : this is temporary, to be removed in the future
+patches = [
+    # https://github.com/sheredom/subprocess.h/pull/102
+    "vendor/sheredom/patch-bsd.patch",
+    # https://github.com/sheredom/subprocess.h/pull/101
+    "vendor/sheredom/patch-windows-quote-backslash.patch",
+    # https://github.com/sheredom/subprocess.h/pull/104
+    # note: must be applied after patch-bsd.patch, they touch adjacent lines
+    "vendor/sheredom/patch-glibc-older-than-2.29.patch",
+]
+
 for url, filename in vendor.items():
    print(f"downloading {url} to {filename}") # noqa: NP100
    urllib.request.urlretrieve(url, filename)

+for patch in patches:
+    print(f"applying {patch}") # noqa: NP100
+    try:
+        subprocess.check_call([
+            "git", "apply", "--directory", os.path.dirname(patch), patch
+        ])
+    except Exception as e:
+        print(f"Error: {e}") # noqa: NP100
+        sys.exit(1)
+
 print("Splitting httplib.h...") # noqa: NP100
 try:
    subprocess.check_call([
@@ -123,15 +123,15 @@ function(npm_build out_var)
    endif()

    if(need_install)
-        message(STATUS "UI: running npm install")
+        message(STATUS "UI: running npm ci")
        execute_process(
-            COMMAND ${NPM_EXECUTABLE} install
+            COMMAND ${NPM_EXECUTABLE} ci
            WORKING_DIRECTORY "${WORK_DIR}"
            RESULT_VARIABLE rc
            ERROR_VARIABLE  err
        )
        if(NOT rc EQUAL 0)
-            message(STATUS "UI: npm install failed (${rc})")
+            message(STATUS "UI: npm ci failed (${rc})")
            message(STATUS "  stderr: ${err}")
            return()
        endif()
@@ -119,6 +119,7 @@ Public API changes carry a higher bar than internal ones (`CONTRIBUTING.md`). Re
 - In most cases, `build_vit` should be enough to build the transformer graph for vision models. Do not add a loop to build the transformer graph manually, unless you have a very good reason to do so. If you do, please explain why in the PR description.
 - If you need a dedicated preprocessor, there is a high chance that it can be a derived class from one of the existing preprocessors. Check carefully before adding a new preprocessor class.
 - If the model need a new public API in `mtmd.h`, open a discussion first.
+- For audio generation models, see `tools/mtmd/README-dev.md`

 ## General (always)

@@ -144,6 +144,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
    { LLM_ARCH_TALKIE,           "talkie"           },
    { LLM_ARCH_MELLUM,           "mellum"           },
    { LLM_ARCH_NANBEIGE,         "nanbeige"         },
+    { LLM_ARCH_QWEN3TTS,         "qwen3tts"         },
    { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };

@@ -1026,6 +1027,7 @@ bool llm_arch_supports_sm_tensor(const llm_arch & arch) {
        case LLM_ARCH_MINIMAX_M3:
        case LLM_ARCH_MISTRAL4:
        case LLM_ARCH_KIMI_LINEAR:
+        case LLM_ARCH_QWEN3TTS:
            return false;
        default:
            return true;
@@ -149,6 +149,7 @@ enum llm_arch {
    LLM_ARCH_MINIMAX_M3,
    LLM_ARCH_DFLASH,
    LLM_ARCH_NANBEIGE,
+    LLM_ARCH_QWEN3TTS,
    LLM_ARCH_UNKNOWN,
 };

@@ -124,3 +124,9 @@ LLAMA_API llama_context * llama_get_ctx_other(struct llama_context * ctx);
 LLAMA_API const int32_t * llama_model_target_layer_ids  (const struct llama_model * model);
 // returns the number of extracted layers from target model
 LLAMA_API uint32_t        llama_model_target_layer_ids_n(const struct llama_model * model);
+
+// retrieves the whole token embedding matrix in F32 format (n_embd * n_vocab)
+// returns total number of elements or 0 on error
+// if out is nullptr, returns the number of tokens without writing to out
+// caller must allocate enough memory for out before calling
+LLAMA_API uint32_t llama_model_get_tok_embd(const struct llama_model * model, float * out);
@@ -648,10 +648,12 @@ const char * llama_grammar_parser::parse_sequence(
            } else {
                throw std::runtime_error(std::string("expecting ',' at ") + pos);
            }
-            bool has_max = max_times != UINT64_MAX;
-            if (min_times > MAX_REPETITION_THRESHOLD || (has_max && max_times > MAX_REPETITION_THRESHOLD)) {
+            if (min_times > MAX_REPETITION_THRESHOLD) {
                throw std::runtime_error(std::string("number of repetitions exceeds sane defaults, please reduce the number of repetitions"));
            }
+            if (max_times != UINT64_MAX && max_times > MAX_REPETITION_THRESHOLD) {
+                max_times = UINT64_MAX;
+            }
            handle_repetitions(min_times, max_times);
        } else {
            break;
@@ -3683,7 +3683,6 @@ void llm_graph_context::build_sampling() const {
            /*.probs       =*/ nullptr,
            /*.sampled     =*/ nullptr,
            /*.candidates  =*/ nullptr,
-            /*.n_vocab     =*/ logits_seq->ne[0],
        };

        assert(sampler->iface->backend_apply);
@@ -857,7 +857,11 @@ struct ggml_tensor * llama_model_loader::require_tensor_meta(const std::string &
    return tensor;
 }

-const struct ggml_tensor * llama_model_loader::check_tensor_dims(const std::string & name, const std::vector<int64_t> & ne, bool required) const {
+const struct ggml_tensor * llama_model_loader::check_tensor_dims(
+        const std::string & name,
+        const std::vector<int64_t> & ne,
+        bool required,
+        bool allow_reshape) const {
    const struct ggml_tensor * cur = get_tensor_meta(name.c_str());

    if (cur == NULL) {
@@ -867,21 +871,33 @@ const struct ggml_tensor * llama_model_loader::check_tensor_dims(const std::stri
        throw std::runtime_error(format("%s: tensor '%s' not found", __func__, name.c_str()));
    }

-    {
-        bool is_ok = true;
+    bool is_ok = true;
+
+    if (allow_reshape) {
+        // check total number of elements only
+        const int64_t ncur = ggml_nelements(cur);
+        int64_t nexp = 1;
+        for (size_t i = 0; i < ne.size(); ++i) {
+            nexp *= ne[i];
+        }
+        if (ncur != nexp) {
+            is_ok = false;
+        }
+    } else {
        for (size_t i = 0; i < GGML_MAX_DIMS; ++i) {
            if ((i < ne.size() && ne[i] != cur->ne[i]) || (i >= ne.size() && cur->ne[i] != 1)) {
                is_ok = false;
                break;
            }
        }
-        if (!is_ok) {
-            throw std::runtime_error(
-                    format("%s: tensor '%s' has wrong shape; expected %s, got %s",
-                        __func__, name.c_str(),
-                        llama_format_tensor_shape(ne).c_str(),
-                        llama_format_tensor_shape(cur).c_str()));
-        }
+    }
+
+    if (!is_ok) {
+        throw std::runtime_error(
+                format("%s: tensor '%s' has wrong shape; expected %s, got %s",
+                    __func__, name.c_str(),
+                    llama_format_tensor_shape(ne).c_str(),
+                    llama_format_tensor_shape(cur).c_str()));
    }

    return cur;
@@ -1246,11 +1262,25 @@ struct ggml_tensor * llama_model_loader::create_tensor(
        return ret;
    }

-    ggml_tensor * t_meta = get_tensor_meta(tn.str().c_str());
-    ggml_backend_buffer_type_t buft = buft_for_tensor(t_meta);
-    if (buft == nullptr) {
-        return nullptr; // return type is ggml_tensor *
+    LLAMA_LOG_DEBUG("%s: loading tensor %s\n", __func__, tn.str().c_str());
+    const struct ggml_tensor * cur = check_tensor_dims(tn.str(), ne, !(flags & TENSOR_NOT_REQUIRED), flags & TENSOR_ALLOW_RESHAPE);
+    if (cur == NULL) {
+        return NULL;
    }
+
+    ggml_tensor t_meta = *cur;
+    if (flags & TENSOR_ALLOW_RESHAPE) {
+        for (size_t dim = 0; dim < GGML_MAX_DIMS; dim++) {
+            t_meta.ne[dim] = dim < ne.size() ? ne.begin()[dim] : 1;
+            t_meta.nb[dim] = dim == 0 ? ggml_type_size(t_meta.type) : t_meta.ne[dim-1]*t_meta.nb[dim-1];
+        }
+    }
+
+    ggml_backend_buffer_type_t buft = buft_for_tensor(&t_meta);
+    if (buft == nullptr) {
+        return nullptr;
+    }
+
    ggml_context * ctx = ctx_for_buft(buft);

    // if duplicated, check if the original tensor was allocated in the same buffer type context and avoid creating a new one
@@ -1261,20 +1291,13 @@ struct ggml_tensor * llama_model_loader::create_tensor(
        }
    }

-    LLAMA_LOG_DEBUG("%s: loading tensor %s\n", __func__, tn.str().c_str());
-    const struct ggml_tensor * cur = check_tensor_dims(tn.str(), ne, !(flags & TENSOR_NOT_REQUIRED));
-
-    if (cur == NULL) {
-        return NULL;
-    }
-
    const bool duplicated = flags & TENSOR_DUPLICATED;

-    struct ggml_tensor * tensor = ggml_dup_tensor(ctx, cur);
-    ggml_set_name(tensor, ggml_get_name(cur));
+    struct ggml_tensor * tensor = ggml_dup_tensor(ctx, &t_meta);
+    ggml_set_name(tensor, ggml_get_name(&t_meta));

    if (duplicated) {
-        size_data += ggml_nbytes(cur);
+        size_data += ggml_nbytes(&t_meta);
    } else {
        n_created++;
    }
@@ -1282,34 +1305,6 @@ struct ggml_tensor * llama_model_loader::create_tensor(
    return tensor;
 }

-struct ggml_tensor * llama_model_loader::create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::initializer_list<int64_t> & ne, size_t offset, bool required) {
-    const struct ggml_tensor * cur = check_tensor_dims(name, ne, required);
-
-    if (cur == NULL) {
-        return NULL;
-    }
-
-    if (cur->type != base->type) {
-        throw std::runtime_error(format("%s: tensor '%s' has wrong type; expected %s, got %s", __func__, name.c_str(), ggml_type_name(base->type), ggml_type_name(cur->type)));
-    }
-
-    std::array<int64_t, GGML_MAX_DIMS> dims;
-    for (size_t i = 0; i < GGML_MAX_DIMS; ++i) {
-        dims[i] = i < ne.size() ? ne.begin()[i] : 1;
-    }
-
-    struct ggml_tensor * tensor = ggml_view_4d(ctx, base,
-                                    dims[0], dims[1], dims[2], dims[3],
-                                    cur->nb[1], cur->nb[2], cur->nb[3],
-                                    offset);
-
-    ggml_set_name(tensor, name.c_str());
-
-    n_created++;
-
-    return tensor;
-}
-
 void llama_model_loader::done_getting_tensors(bool partial) const {
    if (n_created > n_tensors) {
        throw std::runtime_error(format("%s: too many tensors created; expected %d, got %d", __func__, n_tensors, n_created));
@@ -67,6 +67,7 @@ struct llama_model_loader {
    static const int TENSOR_DUPLICATED      = 1 << 1;
    static const int TENSOR_SKIP            = 1 << 2;
    static const int TENSOR_SKIP_IF_VIRTUAL = 1 << 3;
+    static const int TENSOR_ALLOW_RESHAPE   = 1 << 4;

    int n_kv      = 0;
    int n_tensors = 0;
@@ -177,14 +178,16 @@ struct llama_model_loader {

    struct ggml_tensor * require_tensor_meta(const std::string & name) const;

-    const struct ggml_tensor * check_tensor_dims(const std::string & name, const std::vector<int64_t> & ne, bool required) const;
+    const struct ggml_tensor * check_tensor_dims(
+            const std::string & name,
+            const std::vector<int64_t> & ne,
+            bool required,
+            bool allow_reshape) const;

    struct ggml_tensor * create_tensor(
        const llama_hparams & hparams, const buft_list_t * buft_list_cpu, const buft_list_t * buft_list_input, const buft_list_t * buft_list_output,
        const buft_list_t * buft_list_layer, const LLM_TN_IMPL & tn, const std::initializer_list<int64_t> & ne, int flags);

-    struct ggml_tensor * create_tensor_as_view(struct ggml_context * ctx, struct ggml_tensor * base, const std::string & name, const std::initializer_list<int64_t> & ne, size_t offset, bool required = true);
-
    void done_getting_tensors(bool partial = false) const;

    void init_mappings(bool prefetch = true, llama_mlocks * mlock_mmaps = nullptr);
@@ -112,6 +112,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
            return new llama_model_qwen3vl(params);
        case LLM_ARCH_QWEN3VLMOE:
            return new llama_model_qwen3vlmoe(params);
+        case LLM_ARCH_QWEN3TTS:
+            return new llama_model_qwen3tts(params);
        case LLM_ARCH_PHI2:
            return new llama_model_phi2(params);
        case LLM_ARCH_PHI3:
@@ -2693,6 +2695,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
        case LLM_ARCH_QWEN3VLMOE:
        case LLM_ARCH_QWEN35:
        case LLM_ARCH_QWEN35MOE:
+        case LLM_ARCH_QWEN3TTS:
            return LLAMA_ROPE_TYPE_IMROPE;

        case LLM_ARCH_GLM4:
@@ -2867,7 +2870,8 @@ llama_model_base::llama_model_base(const struct llama_model_params & params) : l
    TENSOR_DUPLICATED     (llama_model_loader::TENSOR_DUPLICATED),
    TENSOR_NOT_REQUIRED   (llama_model_loader::TENSOR_NOT_REQUIRED),
    TENSOR_SKIP           (llama_model_loader::TENSOR_SKIP),
-    TENSOR_SKIP_IF_VIRTUAL(llama_model_loader::TENSOR_SKIP_IF_VIRTUAL) {}
+    TENSOR_SKIP_IF_VIRTUAL(llama_model_loader::TENSOR_SKIP_IF_VIRTUAL),
+    TENSOR_ALLOW_RESHAPE  (llama_model_loader::TENSOR_ALLOW_RESHAPE) {}

 ggml_tensor * llama_model_base::create_tensor(const LLM_TN_IMPL & tn, const std::initializer_list<int64_t> & ne, int flags) {
    GGML_ASSERT(ml != nullptr);
@@ -2886,6 +2890,21 @@ void llama_model_base::create_tensor_qkv(llama_layer & layer, int bid,
        int64_t n_embd_, int64_t n_embd_q_, int64_t n_embd_k_, int64_t n_embd_v_,
        int flags) {
    const int64_t n_embd_qkv = n_embd_q_ + n_embd_k_ + n_embd_v_;
+
+    if (flags & TENSOR_SKIP) {
+        const int skip = TENSOR_NOT_REQUIRED | TENSOR_SKIP;
+
+        create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", bid), {n_embd_, n_embd_qkv}, skip | TENSOR_SKIP_IF_VIRTUAL);
+        create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias",   bid), {n_embd_qkv},          skip | TENSOR_SKIP_IF_VIRTUAL);
+        create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", bid), {n_embd_, n_embd_q_},  skip);
+        create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", bid), {n_embd_, n_embd_k_},  skip);
+        create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", bid), {n_embd_, n_embd_v_},  skip);
+        create_tensor(tn(LLM_TENSOR_ATTN_Q,   "bias",   bid), {n_embd_q_},           skip);
+        create_tensor(tn(LLM_TENSOR_ATTN_K,   "bias",   bid), {n_embd_k_},           skip);
+        create_tensor(tn(LLM_TENSOR_ATTN_V,   "bias",   bid), {n_embd_v_},           skip);
+        return;
+    }
+
    layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", bid), {n_embd_, n_embd_qkv}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
    if (layer.wqkv) {
        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", bid), {n_embd_qkv}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
@@ -2907,3 +2926,38 @@ const int32_t * llama_model_target_layer_ids(const struct llama_model * model) {
 uint32_t llama_model_target_layer_ids_n(const struct llama_model * model) {
    return (uint32_t) model->target_layer_ids.size();
 }
+
+uint32_t llama_model_get_tok_embd(const struct llama_model * model, float * out) {
+    if (model->vocab.n_tokens() == 0 || model->tok_embd == nullptr) {
+        return 0;
+    }
+
+    const ggml_tensor * tensor = model->tok_embd;
+    const size_t nelements = ggml_nelements(tensor);
+    GGML_ASSERT(nelements <= UINT32_MAX); // for the return type
+
+    if (out == nullptr) {
+        return (uint32_t) nelements;
+    }
+
+    if (tensor->type == GGML_TYPE_F32) {
+        ggml_backend_tensor_get(tensor, out, 0, nelements * sizeof(float));
+        return (uint32_t) nelements;
+    }
+
+    std::vector<uint8_t> buf(ggml_nbytes(tensor));
+    ggml_backend_tensor_get(tensor, buf.data(), 0, buf.size());
+
+    const ggml_type_traits * traits = ggml_get_type_traits(tensor->type);
+    if (tensor->type == GGML_TYPE_F16) {
+        ggml_fp16_to_fp32_row((const ggml_fp16_t *) buf.data(), out, nelements);
+    } else if (tensor->type == GGML_TYPE_BF16) {
+        ggml_bf16_to_fp32_row((const ggml_bf16_t *) buf.data(), out, nelements);
+    } else if (ggml_is_quantized(tensor->type) && traits->to_float != nullptr) {
+        traits->to_float(buf.data(), out, nelements);
+    } else {
+        GGML_ABORT("unsupported tensor type for dequantization: %s", ggml_type_name(tensor->type));
+    }
+
+    return (uint32_t) nelements;
+}
@@ -719,6 +719,7 @@ struct llama_model_base : public llama_model {
    const int TENSOR_NOT_REQUIRED;
    const int TENSOR_SKIP;
    const int TENSOR_SKIP_IF_VIRTUAL;
+    const int TENSOR_ALLOW_RESHAPE;

    explicit llama_model_base(const llama_model_params & params);
    virtual ~llama_model_base() = default;
@@ -589,7 +589,6 @@ static bool llama_sampler_backend_support(
        /*.probs      = */ nullptr,
        /*.sampled    = */ nullptr,
        /*.candidates = */ ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n),
-        /*.n_vocab    = */ n,
    };

    ggml_cgraph * gf = ggml_new_graph(ctx);
@@ -2640,6 +2639,7 @@ struct llama_sampler * llama_sampler_init_grammar_lazy_patterns(
 // penalties

 struct llama_sampler_penalties : public llama_sampler_backend {
+    const int32_t n_vocab;
    const int32_t penalty_last_n;
    const float   penalty_repeat;
    const float   penalty_freq;
@@ -2655,7 +2655,6 @@ struct llama_sampler_penalties : public llama_sampler_backend {
    ggml_tensor * inp_counts    = nullptr;

    // backend helpers
-    int32_t n_vocab = 0;
    int32_t n_max   = 0;
    bool has_candidates = false;

@@ -2676,11 +2675,13 @@ struct llama_sampler_penalties : public llama_sampler_backend {
    }

    llama_sampler_penalties(
+            int32_t n_vocab,
            int32_t penalty_last_n,
            float   penalty_repeat,
            float   penalty_freq,
            float   penalty_present)
        : llama_sampler_backend("penalties")
+        , n_vocab         (n_vocab)
        , penalty_last_n  (penalty_last_n)
        , penalty_repeat  (penalty_repeat)
        , penalty_freq    (penalty_freq)
@@ -2766,6 +2767,7 @@ static void llama_sampler_penalties_reset(struct llama_sampler * smpl) {
 static struct llama_sampler * llama_sampler_penalties_clone(const struct llama_sampler * smpl) {
    const auto * ctx = (const llama_sampler_penalties *) smpl->ctx;
    auto * result = llama_sampler_init_penalties(
+            ctx->n_vocab,
            ctx->penalty_last_n,
            ctx->penalty_repeat,
            ctx->penalty_freq,
@@ -2811,10 +2813,9 @@ static void llama_sampler_penalties_backend_apply(
        return;
    }

-    GGML_ASSERT(data->n_vocab > 0 && data->n_vocab <= INT32_MAX);
+    GGML_ASSERT(sctx->n_vocab > 0);

    sctx->has_candidates = data->candidates != nullptr;
-    sctx->n_vocab = (int32_t) data->n_vocab;
    sctx->n_max   = std::min(sctx->penalty_last_n, sctx->n_vocab);

    sctx->inp_token_ids = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sctx->n_max);
@@ -2965,6 +2966,7 @@ static struct llama_sampler_i llama_sampler_penalties_i = {
 };

 struct llama_sampler * llama_sampler_init_penalties(
+        int32_t n_vocab,
        int32_t penalty_last_n,
        float penalty_repeat,
        float penalty_freq,
@@ -2979,6 +2981,7 @@ struct llama_sampler * llama_sampler_init_penalties(
    return llama_sampler_init(
        /* .iface = */ &llama_sampler_penalties_i,
        /* .ctx   = */ new llama_sampler_penalties(
+            n_vocab,
            penalty_last_n,
            penalty_repeat,
            penalty_freq,
@@ -3075,8 +3078,6 @@ struct llama_sampler * llama_sampler_init_top_n_sigma(float n) {
 // DRY

 struct llama_sampler_dry {
-    int32_t total_context_size;
-
    const float   dry_multiplier;
    const float   dry_base;
    const int32_t dry_allowed_length;
@@ -3152,8 +3153,7 @@ static void llama_sampler_dry_apply(struct llama_sampler * smpl, llama_token_dat
        return;
    }

-    int32_t effective_dry_penalty_last_n = (ctx->dry_penalty_last_n == -1) ? ctx->total_context_size : std::max(ctx->dry_penalty_last_n, 0);
-    int last_n_repeat = std::min(std::min((int)ctx->last_tokens.size(), effective_dry_penalty_last_n), ctx->total_context_size);
+    int last_n_repeat = std::min((int) ctx->last_tokens.size(), ctx->dry_penalty_last_n);

    if (last_n_repeat <= ctx->dry_allowed_length) {
        return;
@@ -3366,7 +3366,7 @@ static struct llama_sampler * llama_sampler_dry_clone(const struct llama_sampler
    llama_vocab dummy_vocab;

    // dummy vocab is passed because it is only needed for raw sequence breaker processing, which we have already done and will simply be copying
-    auto * result = llama_sampler_init_dry(&dummy_vocab, ctx->total_context_size, ctx->dry_multiplier, ctx->dry_base, ctx->dry_allowed_length, ctx->dry_penalty_last_n, NULL, 0);
+    auto * result = llama_sampler_init_dry(&dummy_vocab, ctx->dry_multiplier, ctx->dry_base, ctx->dry_allowed_length, ctx->dry_penalty_last_n, NULL, 0);

    // Copy the state, including the processed breakers
    {
@@ -3397,8 +3397,8 @@ static struct llama_sampler_i llama_sampler_dry_i = {
    /* .backend_set_input = */ nullptr,
 };

-struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab, int32_t n_ctx_train, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
-    int32_t effective_dry_penalty_last_n = (dry_penalty_last_n == -1) ? n_ctx_train : std::max(dry_penalty_last_n, 0);
+struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const char** seq_breakers, size_t num_breakers) {
+    dry_penalty_last_n = std::max(dry_penalty_last_n, 0);
    std::unordered_multimap<llama_token, std::vector<llama_token>> processed_breakers;
    const int MAX_CHAR_LEN = 40;
    const int MAX_SEQ_LEN = 20;
@@ -3435,23 +3435,22 @@ struct llama_sampler * llama_sampler_init_dry(const struct llama_vocab * vocab,
    return llama_sampler_init(
        /* .iface = */ &llama_sampler_dry_i,
        /* .ctx   = */ new llama_sampler_dry {
-            /* .total_context_size     = */ n_ctx_train,
            /* .dry_multiplier         = */ dry_multiplier,
            /* .dry_base               = */ dry_base,
            /* .dry_allowed_length     = */ dry_allowed_length,
            /* .dry_penalty_last_n     = */ dry_penalty_last_n,
            /* .dry_processed_breakers = */ std::move(processed_breakers),
-            /* .dry_repeat_count       = */ dry_enabled ? std::vector<int>(effective_dry_penalty_last_n, 0) : std::vector<int>{},
+            /* .dry_repeat_count       = */ dry_enabled ? std::vector<int>(dry_penalty_last_n, 0) : std::vector<int>{},
            /* .dry_max_token_repeat   = */ {},
-            /* .last_tokens            = */ dry_enabled ? ring_buffer<llama_token>(effective_dry_penalty_last_n) : ring_buffer<llama_token>(0),
+            /* .last_tokens            = */ dry_enabled ? ring_buffer<llama_token>(dry_penalty_last_n) : ring_buffer<llama_token>(0),
        }
    );
 }

 // wrapper for test-sampling.cpp
-struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers) {
+struct llama_sampler * llama_sampler_init_dry_testing(float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers) {
    llama_vocab dummy_vocab;
-    auto * result = llama_sampler_init_dry(&dummy_vocab, context_size, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, NULL, 0);
+    auto * result = llama_sampler_init_dry(&dummy_vocab, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, NULL, 0);
    auto * ctx = (llama_sampler_dry *) result->ctx;

    // Process the token-based sequence breakers
@@ -34,7 +34,6 @@ struct llama_sampler_chain {
 };

 struct llama_sampler * llama_sampler_init_dry_testing(
-        int32_t context_size,
        float   dry_multiplier,
        float   dry_base,
        int32_t dry_allowed_length,
@@ -1373,8 +1373,10 @@ struct llm_tokenizer_plamo2 : llm_tokenizer {
            if (vocab.is_byte(token_id)) {
                if (entry.text.length() == 6 && entry.text.substr(0, 3) == "<0x" && entry.text.back() == '>') {
                    std::string hex_str = entry.text.substr(3, 2);
-                    int byte_val = std::stoi(hex_str, nullptr, 16);
-                    bytes_[byte_val] = static_cast<llama_token>(token_id);
+                    if (std::isxdigit(static_cast<unsigned char>(hex_str[0])) && std::isxdigit(static_cast<unsigned char>(hex_str[1]))) {
+                        int byte_val = std::stoi(hex_str, nullptr, 16);
+                        bytes_[byte_val] = static_cast<llama_token>(token_id);
+                    }
                }
                continue;
            }
@@ -3625,12 +3627,15 @@ int32_t llama_vocab::impl::token_to_piece(llama_token token, char * buf, int32_t
                if (vocab.is_byte(token)) {
                    // Handle byte tokens like <0xXX>
                    if (token_text.length() == 6 && token_text.substr(0, 3) == "<0x" && token_text.back() == '>') {
-                        int hex_val = std::stoi(token_text.substr(3, 2), nullptr, 16);
-                        if (length < 1) {
-                            return -1;
+                        std::string hex_str = token_text.substr(3, 2);
+                        if (std::isxdigit(static_cast<unsigned char>(hex_str[0])) && std::isxdigit(static_cast<unsigned char>(hex_str[1]))) {
+                            int hex_val = std::stoi(hex_str, nullptr, 16);
+                            if (length < 1) {
+                                return -1;
+                            }
+                            buf[0] = static_cast<char>(hex_val);
+                            return 1;
                        }
-                        buf[0] = static_cast<char>(hex_val);
-                        return 1;
                    }
                }

@@ -114,7 +114,9 @@ void llama_model_deepseek4::load_arch_tensors(llama_model_loader & ml) {
        layer.wq_b          = create_tensor(tn(LLM_TENSOR_ATTN_Q_B,      "weight", i), {q_lora_rank, n_head * n_embd_head}, flags);
        layer.wkv           = create_tensor(tn(LLM_TENSOR_ATTN_KV,       "weight", i), {n_embd, n_embd_head}, flags);
        layer.attn_kv_norm  = create_tensor(tn(LLM_TENSOR_ATTN_KV_NORM,  "weight", i), {n_embd_head}, flags);
-        layer.wo_a          = create_tensor(tn(LLM_TENSOR_ATTN_OUT_A,    "weight", i), {n_head * n_embd_head / o_groups, o_lora_rank * o_groups}, flags);
+        // for wo_a, the shape in the file is (n_head * n_embd_head / o_groups, o_lora_rank*o_groups)
+        // so we reshape here, to avoid reshaping the tensor in the graph
+        layer.wo_a          = create_tensor(tn(LLM_TENSOR_ATTN_OUT_A,    "weight", i), {n_head * n_embd_head / o_groups, o_lora_rank, o_groups}, flags | TENSOR_ALLOW_RESHAPE);
        layer.wo_b          = create_tensor(tn(LLM_TENSOR_ATTN_OUT_B,    "weight", i), {o_groups * o_lora_rank, n_embd}, flags);

        layer.hc_attn_fn    = create_tensor(tn(LLM_TENSOR_HC_ATTN_FN,    "weight", i), {hc_dim, hc_mix_dim}, flags);
@@ -1258,7 +1260,7 @@ ggml_tensor * llama_model_deepseek4::graph::build_attention_impl(

    out = ggml_reshape_3d(ctx0, out, o_group_dim, n_groups, nt);
    out = ggml_permute(ctx0, out, 0, 2, 1, 3);
-    ggml_tensor * oa = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, layer.wo_a, layer.wo_a->ne[0], o_lora_rank, n_groups), out);
+    ggml_tensor * oa = ggml_mul_mat(ctx0, layer.wo_a, out);
    cb(oa, "attn_wo_a", il);
    oa = ggml_permute(ctx0, oa, 0, 2, 1, 3);
    oa = ggml_cont_2d(ctx0, oa, o_lora_rank*n_groups, nt);
@@ -125,7 +125,7 @@ void llama_model_dflash::load_arch_tensors(llama_model_loader &) {
            layer.wq_b          = create_tensor(tn(LLM_TENSOR_ATTN_Q_B,      "weight", i), {q_lora_rank, n_head * n_embd_head}, 0);
            layer.wkv           = create_tensor(tn(LLM_TENSOR_ATTN_KV,       "weight", i), {n_embd, n_embd_head}, 0);
            layer.attn_kv_norm  = create_tensor(tn(LLM_TENSOR_ATTN_KV_NORM,  "weight", i), {n_embd_head}, 0);
-            layer.wo_a          = create_tensor(tn(LLM_TENSOR_ATTN_OUT_A,    "weight", i), {n_head * n_embd_head / o_groups, o_lora_rank * o_groups}, 0);
+            layer.wo_a          = create_tensor(tn(LLM_TENSOR_ATTN_OUT_A,    "weight", i), {n_head * n_embd_head / o_groups, o_lora_rank, o_groups}, TENSOR_ALLOW_RESHAPE);
            layer.wo_b          = create_tensor(tn(LLM_TENSOR_ATTN_OUT_B,    "weight", i), {o_groups * o_lora_rank, n_embd}, 0);

            layer.hc_attn_fn    = create_tensor(tn(LLM_TENSOR_HC_ATTN_FN,    "weight", i), {hc_dim, hc_mix_dim}, 0);
@@ -596,6 +596,11 @@ struct llama_model_qwen3vlmoe : public llama_model_base {
 };


+struct llama_model_qwen3tts : public llama_model_qwen3vl {
+    llama_model_qwen3tts(const struct llama_model_params & params) : llama_model_qwen3vl(params) {}
+};
+
+
 struct llama_model_phi2 : public llama_model_base {
    llama_model_phi2(const struct llama_model_params & params) : llama_model_base(params) {}
    void load_arch_hparams(llama_model_loader & ml) override;
@@ -0,0 +1,3 @@
+#include "models.h"
+
+// llama_model_qwen3tts reuses llama_model_qwen3vl's hparams/tensors/graph logic
@@ -16,11 +16,16 @@ void llama_model_qwen3vl::load_arch_hparams(llama_model_loader & ml) {
 void llama_model_qwen3vl::load_arch_tensors(llama_model_loader &) {
    LLAMA_LOAD_LOCALS;

+    int64_t n_vocab_out = n_vocab;
+    if (arch == LLM_ARCH_QWEN3TTS) {
+        n_vocab_out = 3072;
+    }
+
    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);

    // output
    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
-    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab_out}, TENSOR_NOT_REQUIRED);
    // if output is NULL, init from the input tok embed
    if (output == NULL) {
        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
@@ -166,6 +171,24 @@ llama_model_qwen3vl::graph::graph(const llama_model & model, const llm_graph_par
    // lm_head
    cur = build_lora_mm(model.output, cur, model.output_s);

+    int64_t n_vocab_in  = model.tok_embd->ne[1];
+    int64_t n_vocab_out = model.output->ne[1];
+    if (n_vocab_in > n_vocab_out) {
+        // case: Qwen3TTS model with codec_head as output
+        GGML_ASSERT(model.output_norm);
+        int64_t pad = n_vocab_in - n_vocab_out;
+
+        // using this trick to get a scalar -inf tensor to pad the output
+        ggml_tensor * neg_inf = ggml_scale_bias(ctx0,
+                ggml_view_1d(ctx0, model.output_norm, 1, 0),
+                0.0f, -INFINITY);
+        neg_inf = ggml_repeat_4d(ctx0, neg_inf, pad, cur->ne[1], 1, 1);
+        cur = ggml_concat(ctx0, neg_inf, cur, 0); // [padded .. n_vocab_out, n_stream]
+
+    } else if (n_vocab_in < n_vocab_out) {
+        GGML_ABORT("invalid case");
+    }
+
    cb(cur, "result_output", -1);
    res->t_logits = cur;

@@ -101,6 +101,14 @@ static void test(void) {

    {
        common_params penalty_params;
+        assert(penalty_params.sampling.penalty_last_n == 64);
+        assert(penalty_params.sampling.dry_penalty_last_n == 64);
+
+        argv = {"binary_name", "--repeat-last-n", "-1"};
+        assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), penalty_params, LLAMA_EXAMPLE_COMMON));
+
+        argv = {"binary_name", "--dry-penalty-last-n", "-1"};
+        assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), penalty_params, LLAMA_EXAMPLE_COMMON));

        argv = {"binary_name", "--repeat-penalty", "0"};
        assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), penalty_params, LLAMA_EXAMPLE_COMMON));
@@ -823,6 +823,7 @@ enum class penalties_position {
 static void add_filter_and_penalties(
        llama_sampler * chain,
        const sampler_init_fn & init_filter,
+        int32_t n_vocab,
        int32_t penalty_last_n,
        float penalty_repeat,
        float penalty_freq,
@@ -830,7 +831,7 @@ static void add_filter_and_penalties(
        penalties_position position) {
    const auto add_penalties = [&]() {
        llama_sampler_chain_add(chain, llama_sampler_init_penalties(
-                    penalty_last_n, penalty_repeat, penalty_freq, penalty_present));
+                    n_vocab, penalty_last_n, penalty_repeat, penalty_freq, penalty_present));
    };

    if (position == penalties_position::before_filter) {
@@ -1006,7 +1007,7 @@ static sampler_comparison_output run_penalties_comparison(
    const std::vector<float> raw_logits = decode_raw_logits(params, prompt);
    const auto add_samplers = [&](llama_sampler * chain) {
        llama_sampler_chain_add(chain, llama_sampler_init_penalties(
-                    penalty_last_n, penalty_repeat, penalty_freq, penalty_present));
+                    llama_vocab_n_tokens(vocab), penalty_last_n, penalty_repeat, penalty_freq, penalty_present));
    };
    const auto accept_history = [&](llama_sampler * chain) {
        accept_prompt(chain, vocab, prompt);
@@ -1105,7 +1106,7 @@ static void compare_top_k_penalties_logits(
    GGML_ASSERT(excluded_history_token != LLAMA_TOKEN_NULL);

    const auto add_samplers = [&](llama_sampler * chain) {
-        add_filter_and_penalties(chain, init_top_k,
+        add_filter_and_penalties(chain, init_top_k, n_vocab,
                penalty_last_n, penalty_repeat, penalty_freq, penalty_present, position);
    };

@@ -1190,7 +1191,7 @@ static void compare_masking_penalties_logits(
    GGML_ASSERT(masked_token != LLAMA_TOKEN_NULL);

    const auto add_samplers = [&](llama_sampler * chain) {
-        add_filter_and_penalties(chain, init_filter,
+        add_filter_and_penalties(chain, init_filter, n_vocab,
                penalty_last_n, penalty_repeat, penalty_freq, penalty_present, position);
    };
    auto accept_history = [&](llama_sampler * smpl) {
@@ -1218,7 +1219,7 @@ static void compare_masking_penalties_logits(
            GGML_ASSERT(fabsf(expected_logits.at(penalized_token) - raw_logits[penalized_token]) > 1e-6f);
        } else {
            llama_sampler_ptr penalties(llama_sampler_init_penalties(
-                        penalty_last_n, penalty_repeat, penalty_freq, penalty_present));
+                        n_vocab, penalty_last_n, penalty_repeat, penalty_freq, penalty_present));
            accept_history(penalties.get());
            const std::unordered_map<llama_token, float> penalized_logits =
                map_logits(apply_cpu_sampler(raw_logits, penalties.get()));
@@ -153,6 +153,53 @@ int main()
        root ::= "a"{,10}"
    )""");

+    verify_failure(R"""(
+        root ::= "a"{5000}
+    )""");
+
+    verify_failure(R"""(
+        root ::= "a"{5000,}
+    )""");
+
+    verify_failure(R"""(
+        root ::= "a"{5000,6000}
+    )""");
+
+    verify_parsing(R"""(
+        root ::= "a"{0,5000}
+    )""", {
+        {"root", 0},
+        {"root_1", 1},
+    }, {
+        // root (index 0)
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_END, 0},
+        // root_1 (index 1)
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_END, 0},
+    });
+
+    verify_parsing(R"""(
+        root ::= "a"{3,5000}
+    )""", {
+        {"root", 0},
+        {"root_1", 1},
+    }, {
+        // root (index 0)
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_END, 0},
+        // root_1 (index 1)
+        {LLAMA_GRETYPE_CHAR, 'a'},
+        {LLAMA_GRETYPE_RULE_REF, /* root_1 */ 1},
+        {LLAMA_GRETYPE_ALT, 0},
+        {LLAMA_GRETYPE_END, 0},
+    });
+
    verify_parsing(R"""(
        root  ::= "a"
    )""", {
@@ -113,6 +113,8 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
        n_layer = 3;
    } else if (arch == LLM_ARCH_CHAMELEON) {
        n_vocab = 10240;
+    } else if (arch == LLM_ARCH_QWEN3TTS) {
+        n_vocab = 4096; // must be >= the hard-coded codec head size (3072)
    }

    const uint32_t n_embd_head = n_embd / n_head;
@@ -430,7 +432,7 @@ static bool arch_supported(const llm_arch arch) {

    // FIXME: these hit scheduler/view-backed-output issues with WebGPU on CI.
 #ifdef GGML_USE_WEBGPU
-    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA || arch == LLM_ARCH_MINIMAX_M3) {
+    if (arch == LLM_ARCH_DEEPSEEK32 || arch == LLM_ARCH_GLM_DSA) {
        return false;
    }
 #endif // GGML_USE_WEBGPU
@@ -10,7 +10,7 @@
 #include <string>
 #include <vector>

-extern struct llama_sampler * llama_sampler_init_dry_testing(int32_t context_size, float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers);
+extern struct llama_sampler * llama_sampler_init_dry_testing(float dry_multiplier, float dry_base, int32_t dry_allowed_length, int32_t dry_penalty_last_n, const std::vector<std::vector<llama_token>>& seq_breakers);

 static void dump(const llama_token_data_array * cur_p) {
    for (size_t i = 0; i < cur_p->size; i++) {
@@ -144,7 +144,7 @@ static void test_penalties(

    sampler_tester tester(probs, probs_expected);

-    auto * sampler = llama_sampler_init_penalties(last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence);
+    auto * sampler = llama_sampler_init_penalties((int32_t) probs.size(), (int32_t) last_tokens.size(), repeat_penalty, alpha_frequency, alpha_presence);

    for (size_t i = 0; i < last_tokens.size(); i++) {
        llama_sampler_accept(sampler, last_tokens[i]);
@@ -168,7 +168,7 @@ static void test_dry(

    sampler_tester tester(probs, expected_probs);

-    auto * sampler = llama_sampler_init_dry_testing(1024, dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, seq_breakers);
+    auto * sampler = llama_sampler_init_dry_testing(dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n, seq_breakers);

    for (size_t i = 0; i < last_tokens.size(); i++) {
        llama_sampler_accept(sampler, last_tokens[i]);
@@ -116,14 +116,14 @@
 | `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) |
 | `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) |
 | `--typical, --typical-p N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
-| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
+| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled) |
 | `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) |
 | `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) |
 | `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) |
 | `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) |
 | `--dry-base N` | set DRY sampling base value (default: 1.75) |
 | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
-| `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
+| `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: 64, 0 = disable) |
 | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
 | `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) |
 | `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.<br/>(valid range 0.0 to 0.99) (default: 0.90) |
@@ -199,14 +199,14 @@ llama-completion.exe -m models\gemma-1.1-7b-it.Q4_K_M.gguf --ignore-eos -n -1
 | `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) |
 | `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) |
 | `--typical, --typical-p N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
-| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
+| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled) |
 | `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) |
 | `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) |
 | `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) |
 | `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) |
 | `--dry-base N` | set DRY sampling base value (default: 1.75) |
 | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
-| `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
+| `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: 64, 0 = disable) |
 | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
 | `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) |
 | `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.<br/>(valid range 0.0 to 0.99) (default: 0.90) |
@@ -388,11 +388,11 @@ Example usage: `--temp 0`
 ### Repeat Penalty

 -   `--repeat-penalty N`: Control the repetition of token sequences in the generated text default: 1.0, 1.0 = disabled).
-   `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled, -1 = ctx-size).
+-   `--repeat-last-n N`: Last n tokens to consider for penalizing repetition (default: 64, 0 = disabled).

 The `repeat-penalty` option helps prevent the model from generating repetitive or monotonous text. A higher value (e.g., 1.5) will penalize repetitions more strongly, while a lower value (e.g., 0.9) will be more lenient. The default value is 1.

-The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty, and a value of -1 sets the number of tokens considered equal to the context size (`ctx-size`).
+The `repeat-last-n` option controls the number of tokens in the history to consider for penalizing repetition. A larger value will look further back in the generated text to prevent repetitions, while a smaller value will only consider recent tokens. A value of 0 disables the penalty.

 ### DRY Repetition Penalty

@@ -401,7 +401,7 @@ DRY (Don't Repeat Yourself) sampling is an effective technique for reducing repe
 - `--dry-multiplier N`: Set the DRY sampling multiplier (default: 0.0, 0.0 = disabled).
 - `--dry-base N`: Set the DRY sampling base value (default: 1.75).
 - `--dry-allowed-length N`: Set the allowed length for DRY sampling (default: 2).
- `--dry-penalty-last-n N`: Set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size).
+- `--dry-penalty-last-n N`: Set DRY penalty for the last n tokens (default: 64, 0 = disable).
 - `--dry-sequence-breaker STRING`: Add a sequence breaker for DRY sampling. Can be used more than once to add multiple sequence breakers. Using this clears out the default breakers, which consist of: `['\n', ':', '"', '*']`. If the string `"none"` is supplied, no sequence breakers are used.

 The `dry-multiplier` option controls the strength of the DRY sampling effect. A value of 0.0 disables DRY sampling, while higher values increase its influence. A typical recommended value is 0.8.
@@ -410,13 +410,13 @@ The `dry-base` option sets the base value for the exponential penalty calculatio

 The `dry-allowed-length` option sets the maximum length of repeated sequences that will not be penalized. Repetitions shorter than or equal to this length are not penalized, allowing for natural repetitions of short phrases or common words.

-The `dry-penalty-last-n` option controls how many recent tokens to consider when applying the DRY penalty. A value of -1 considers the entire context. Use a positive value to limit the consideration to a specific number of recent tokens.
+The `dry-penalty-last-n` option controls how many recent tokens to consider when applying the DRY penalty. A value of 0 disables the penalty. Use a positive value to limit the consideration to a specific number of recent tokens.

 The `dry-sequence-breaker` option adds a single sequence breaker and can be used more than once to specify multiple sequence breakers. Sequence breakers interrupt sequence matching and break the input into parts where matching can be applied.

 DRY sampling provides more nuanced control over text generation, particularly for reducing long-range repetitions and maintaining global coherence.

-Example usage: `--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 --dry-penalty-last-n -1 --dry-sequence-breaker "—" --dry-sequence-breaker "##"`
+Example usage: `--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 --dry-penalty-last-n 64 --dry-sequence-breaker "—" --dry-sequence-breaker "##"`

 ### Top-K Sampling

@@ -47,6 +47,7 @@ struct split_params {
    std::string output;
    bool no_tensor_first_split = false;
    bool dry_run = false;
+    bool delete_splits = false;
 };

 static void split_print_usage(const char * executable) {
@@ -65,6 +66,7 @@ static void split_print_usage(const char * executable) {
    printf("  --split-max-size N(M|G) max size per split\n");
    printf("  --no-tensor-first-split do not add tensors to the first split (disabled by default)\n");
    printf("  --dry-run               only print out a split plan and exit, without writing any new files\n");
+    printf("  --delete-splits         delete the split files during merge to free up disk space WARNING: this option is unsafe and will leave you in an unrecoverable state if something fails during the merge\n");
    printf("\n");
 }

@@ -147,6 +149,9 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p
            }
            params.mode = MODE_SIZE;
            params.n_bytes_split = split_str_to_n_bytes(argv[arg_idx]);
+        } else if (arg == "--delete-splits") {
+            arg_found = true;
+            params.delete_splits = true;
        }

        if (!arg_found) {
@@ -509,6 +514,7 @@ static void gguf_merge(const split_params & split_params) {
    }

    // Write tensors data
+    bool merge_error = false;
    for (int i_split = 0; i_split < n_split; i_split++) {
        llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split);
        std::ifstream f_input(split_path, std::ios::binary);
@@ -554,6 +560,16 @@ static void gguf_merge(const split_params & split_params) {
        ggml_free(ctx_meta);
        f_input.close();
        fprintf(stderr, "\033[3Ddone\n");
+
+        if (!split_params.dry_run && split_params.delete_splits) {
+            int delete_result = std::remove(split_path);
+            if (delete_result != 0) {
+                merge_error = true;
+                fprintf(stderr, "error: failed to delete %s\n", split_path);
+            } else {
+                fprintf(stderr, "%s: deleted file %s\n", __func__, split_path);
+            }
+        }
    }

    if (!split_params.dry_run) {
@@ -568,6 +584,10 @@ static void gguf_merge(const split_params & split_params) {

    fprintf(stderr, "%s: %s merged from %d split with %d tensors.\n",
            __func__, split_params.output.c_str(), n_split, total_tensors);
+
+    if (merge_error) {
+        exit(EXIT_FAILURE);
+    }
 }

 int main(int argc, const char ** argv) {
@@ -66,12 +66,12 @@ echo PASS
 echo

 # 5. Merge
-#$SPLIT --merge $WORK_PATH/ggml-model-split-32-tensors-00001-of-00012.gguf $WORK_PATH/ggml-model-merge-2.gguf
+#$SPLIT --merge $WORK_PATH/ggml-model-split-32-tensors-00001-of-00011.gguf $WORK_PATH/ggml-model-merge-2.gguf
 #echo PASS
 #echo

 # 5b. Test the merged model is loading properly
-#$MAIN -no-cnv --model $WORK_PATH/ggml-model-merge-2.gguf --n-predict 32
+#$MAIN -no-cnv --model $WORK_PATH/ggml-model-merge-2.gguf -p "I believe the meaning of life is" --n-predict 32
 #echo PASS
 #echo

@@ -85,5 +85,25 @@ $MAIN -no-cnv --model $WORK_PATH/ggml-model-split-500M-00001-of-00002.gguf -p "I
 echo PASS
 echo

+# 7. Merge with delete splits
+#for i in $(seq -w 1 11); do
+#    cp "$WORK_PATH/ggml-model-split-32-tensors-000${i}-of-00011.gguf" "$WORK_PATH/ggml-model-split-32-tensors-copy-000${i}-of-00011.gguf"
+#done
+#$SPLIT --merge --delete-splits $WORK_PATH/ggml-model-split-32-tensors-copy-00001-of-00011.gguf $WORK_PATH/ggml-model-merge-3.gguf
+#echo PASS
+#echo
+
+# 7b. Test the merged model is loading properly
+#$MAIN -no-cnv --model $WORK_PATH/ggml-model-merge-3.gguf -p "I believe the meaning of life is" --n-predict 32
+#echo PASS
+#echo
+
+# 7c. Test the files were deleted
+#for i in $(seq -w 1 11); do
+#    test ! -f "$WORK_PATH/ggml-model-split-32-tensors-copy-000${i}-of-00011.gguf"
+#done
+#echo PASS
+#echo
+
 # Clean up
 rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-merge*.gguf
@@ -18,6 +18,8 @@ add_library(mtmd
            mtmd-image.cpp
            mtmd.h
            mtmd-helper.cpp
+            mtmd-helper-gen.cpp
+            mtmd-helper-common.h
            mtmd-helper.h
            clip.cpp
            clip.h
@@ -52,6 +54,8 @@ add_library(mtmd
            models/mimovl.cpp
            models/qwen3a.cpp
            models/mimo-audio.cpp
+            models/qwen3tts-spkenc.cpp
+            models/qwen3tts-gen.cpp
            models/step3vl.cpp
            models/siglip.cpp
            models/whisper-enc.cpp
@@ -33,3 +33,52 @@ A typical pipeline of the core libmtmd is as follows:
 We provide a set of helper functions via `mtmd_helper` to make using libmtmd easier. The helper provides:
 - Image, audio and video file decoding (for example, decode raw JPEG into RGB bitmap)
 - Manage `llama_batch` and calls to `llama_decode`
+
+## Audio generation support
+
+Audio generation is added to mtmd in PR [#26254](https://github.com/ggml-org/llama.cpp/pull/26254)
+
+Currently, we support the 3-stage pipeline below which should cover most TTS models:
+- Stage 1: Backbone / Semantic Stage: Backbone model accepts text prompt and reference voice as input
+- Stage 2: Acoustic Detail Generator: A model takes the hidden state from backbone and generate audio details (usually as audio codes or mel-spectrogram)
+- Stage 3: Waveform Reconstruction: Convert the semantic and acoustic data from previous stages to the final waveform
+
+For example, Qwen3-TTS:
+- Reference voice is encoded using ECAPA-TDNN speaker encoder (`speaker_encoder`)
+- Text prompt and reference voice are processed via a backbone (`talker.model`)
+- A model converts sampled semantic token and hidden state from stage 2 into a list of 15 acoustic codes (`talker.code_predictor`)
+- 16 generated codes are converted into waveform (`code2wav`)
+
+### API design constraints
+
+Due to wide variety of audio generation pipelines, the `mtmd_gen_audio` system is designed to be flexible and reusable by new models.
+
+`mtmd_gen_audio` is split into 2 main API:
+- Core API `mtmd.h`: handles main inference. Important: the API surface must be stateless; caller must handle state management and audio frame accumulation.
+- Helper API `mtmd-helper.h`: provides a model-agnostic stateful API. Usage example can be found in the `tools/tts` directory.
+
+### Checklist for porting new audio generation models to mtmd
+
+1. Establish a list of reusable and missing components from the current mtmd implementation.
+2. For GGUF conversion:
+    - Backbone model should be converted to a normal text model (loadable via `libllama`)
+        - If model used hard-coded embedding row ID, append them to token embeddings and assign token name for them (see `qwen3tts.py`)
+        - If model have a specific output logits head for audio codes (usually semantic code), keep the head as-is and pad the logits at inference time (see `src/models/qwen3vl.cpp`)
+    - Sidecar models (code2wav, bigvgan, etc) must live inside the mmproj GGUF (but can be in different `clip_context` if necessary)
+        - Note: it should use `ggml_build_forward_select` to select graphs if multiple graphs living in the same context
+    - Reuse existing GGUF metadata key name and tensor name whenever possible; think twice before adding extensive changes to GGUF writer. For example, Qwen3-TTS hard-code part of the hparams to `clip.cpp` as they won't likely to change.
+    - For tensor naming:
+        - Prefixed with `a.*` for tensors used by speaker encoder pipeline
+        - Prefixed with `a.gen.*` for generation stages (code / mel-spectrogram / PCM generation)
+3. Make sure most of the changes happen inside `mtmd-helper-gen.cpp`. A good PR looks like this:
+    - 10-20% changes is to add new backbone (text) model and conversion
+    - 60% changes inside `mtmd-helper-gen.cpp`
+    - 10% changes inside `libmtmd` and `clip.cpp` systems
+    - The rest downstream code (CLI, server) should have no changes at all
+4. Update usage documentation in `tools/tts/README.md`
+
+IMPORTANT: If your model needs changes that don't fit the existing infrastructure, **open an issue first for discussion**.
+
+No-go checklist (these will get the PR rejected and require discussion before proceeding):
+- Violating the API design constraints stated above
+- Adding a new model-specific binary: the API and binary surface must stay model-agnostic
@@ -54,6 +54,9 @@ struct clip_graph {

    clip_graph(clip_ctx * ctx, const clip_image_f32 & img);

+    // build sub-graph, reuse buf from parent
+    clip_graph(const clip_graph & parent);
+
    virtual ~clip_graph() = default;
    virtual ggml_cgraph * build() = 0;

@@ -32,6 +32,7 @@
 #define KEY_PROJ_TYPE           "clip.projector_type"
 #define KEY_HAS_AUDIO_ENC       "clip.has_audio_encoder"
 #define KEY_HAS_VISION_ENC      "clip.has_vision_encoder"
+#define KEY_HAS_GEN_AUDIO_ENC   "clip.has_gen_audio_encoder"
 #define KEY_USE_GELU            "clip.use_gelu"
 #define KEY_USE_SILU            "clip.use_silu"

@@ -89,6 +90,8 @@
 #define KEY_A_ATTN_WINDOW_SIZE     "clip.audio.window_size"          // mimo-audio-tokenizer: sliding-window radius
 #define KEY_A_LOCAL_BLOCK_COUNT    "clip.audio.local_block_count"    // mimo-v2.5: input_local_transformer layer count
 #define KEY_A_LOCAL_GROUP_SIZE     "clip.audio.local_group_size"     // mimo-v2.5: input_local_transformer grouping size
+// audio generation (gen-audio)-specific
+#define KEY_GEN_AUDIO_PROJ_TYPE    "clip.gen.audio.projector_type" // for models with mixed modalities
 #define KEY_AUDIO_SUBSAMPLING_FACTOR "clip.audio.subsampling_factor"

 //
@@ -201,6 +204,48 @@
 #define TN_MM_A_LOCAL_LN2      "mm.a.local_blk.%d.ln2.%s"
 #define TN_MM_A_LOCAL_NORM     "mm.a.local_norm.%s"

+// qwen3tts speaker encoder (ECAPA-TDNN)
+#define TN_A_SE_CONV1  "a.blk.%d.se_conv1.%s"
+#define TN_A_SE_CONV2  "a.blk.%d.se_conv2.%s"
+#define TN_A_CONV_RES2 "a.blk.%d.res2.%d.%s"
+#define TN_A_ASP_ATTN  "a.asp_attn.%s"
+#define TN_A_ASP_TDNN  "a.asp_tdnn.%s"
+
+// qwen3tts code_predictor
+#define TN_A_GEN_CODE_PROJ_IN  "a.gen.code.proj_in.%s"
+#define TN_A_GEN_CODE_EMBD     "a.gen.code.embd.%s"
+#define TN_A_GEN_CODE_HEAD     "a.gen.code.head.%s"
+#define TN_A_GEN_CODE_OUT_EMBD "a.gen.code.out_embd.%s"
+#define TN_A_GEN_CODE_NORM     "a.gen.code.output_norm.%s"
+
+// qwen3tts code2wav (RVQ codes -> raw PCM)
+// pre_transformer layers use the generic TN_ATTN_*/TN_FFN_*/TN_LN_*/TN_LS_* macros, prefix "a.gen.wav.tfm"
+#define TN_A_GEN_WAV_QUANT_FIRST_IN  "a.gen.wav.quant.first.in_proj.%s"
+#define TN_A_GEN_WAV_QUANT_FIRST_OUT "a.gen.wav.quant.first.out_proj.%s"
+#define TN_A_GEN_WAV_QUANT_FIRST_CB  "a.gen.wav.quant.first.codebook.%s"
+#define TN_A_GEN_WAV_QUANT_REST_IN   "a.gen.wav.quant.rest.in_proj.%s"
+#define TN_A_GEN_WAV_QUANT_REST_OUT  "a.gen.wav.quant.rest.out_proj.%s"
+#define TN_A_GEN_WAV_QUANT_REST_CB   "a.gen.wav.quant.rest.codebook.%s"
+#define TN_A_GEN_WAV_PRE_CONV        "a.gen.wav.pre_conv.%s"
+#define TN_A_GEN_WAV_TFM_IN_PROJ     "a.gen.wav.tfm.in_proj.%s"
+#define TN_A_GEN_WAV_TFM_OUT_PROJ    "a.gen.wav.tfm.out_proj.%s"
+#define TN_A_GEN_WAV_TFM_OUT_NORM    "a.gen.wav.tfm.output_norm.%s"
+#define TN_A_GEN_WAV_UP_CONV         "a.gen.wav.up.blk.%d.conv.%s"
+#define TN_A_GEN_WAV_UP_DWCONV       "a.gen.wav.up.blk.%d.dwconv.%s"
+#define TN_A_GEN_WAV_UP_NORM         "a.gen.wav.up.blk.%d.norm.%s"
+#define TN_A_GEN_WAV_UP_PW1          "a.gen.wav.up.blk.%d.pw1.%s"
+#define TN_A_GEN_WAV_UP_PW2          "a.gen.wav.up.blk.%d.pw2.%s"
+#define TN_A_GEN_WAV_UP_GAMMA        "a.gen.wav.up.blk.%d.gamma"
+#define TN_A_GEN_WAV_DAC_ENTRY       "a.gen.wav.dac.entry.%s"
+#define TN_A_GEN_WAV_DAC_SNAKE       "a.gen.wav.dac.blk.%d.snake.%s"
+#define TN_A_GEN_WAV_DAC_CONV        "a.gen.wav.dac.blk.%d.conv.%s"
+#define TN_A_GEN_WAV_DAC_RES_ACT1    "a.gen.wav.dac.blk.%d.res.%d.act1.%s"
+#define TN_A_GEN_WAV_DAC_RES_CONV1   "a.gen.wav.dac.blk.%d.res.%d.conv1.%s"
+#define TN_A_GEN_WAV_DAC_RES_ACT2    "a.gen.wav.dac.blk.%d.res.%d.act2.%s"
+#define TN_A_GEN_WAV_DAC_RES_CONV2   "a.gen.wav.dac.blk.%d.res.%d.conv2.%s"
+#define TN_A_GEN_WAV_DAC_POST_SNAKE  "a.gen.wav.dac.post_snake.%s"
+#define TN_A_GEN_WAV_DAC_POST_CONV   "a.gen.wav.dac.post_conv.%s"
+
 // cogvlm
 #define TN_MM_POST_FC_NORM "mm.post_fc_norm.%s"
 #define TN_MM_H_TO_4H      "mm.up.%s"
@@ -408,6 +453,8 @@ enum projector_type {
    PROJECTOR_TYPE_MINIMAX_M3,
    PROJECTOR_TYPE_GRANITE4_VISION,
    PROJECTOR_TYPE_MIMO_AUDIO,
+    PROJECTOR_TYPE_QWEN3TTS_SPKENC,
+    PROJECTOR_TYPE_QWEN3TTS_GEN,
    PROJECTOR_TYPE_UNKNOWN,
 };

@@ -465,6 +512,8 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
    { PROJECTOR_TYPE_GRANITE4_VISION,   "granite4_vision"},
    { PROJECTOR_TYPE_MIMO_AUDIO,        "mimo_audio"},
    { PROJECTOR_TYPE_PARAKEET,          "parakeet"},
+    { PROJECTOR_TYPE_QWEN3TTS_SPKENC,   "qwen3tts_spkenc"},
+    { PROJECTOR_TYPE_QWEN3TTS_GEN,      "qwen3tts_gen"},
 };

 static projector_type clip_projector_type_from_string(const std::string & str) {
@@ -136,6 +136,19 @@ struct clip_hparams {
    int32_t rvq_num_quantizers = 0;
    std::vector<int32_t> rvq_codebook_size; // per-quantizer bin count (ragged, e.g. 1024/1024/256/128x17)

+    // qwen3tts code2wav
+    int32_t wav_tfm_n_layer      = 0;
+    int32_t wav_tfm_n_embd       = 0;
+    int32_t wav_tfm_n_ff         = 0;
+    int32_t wav_tfm_n_head       = 0;
+    int32_t wav_tfm_n_head_kv    = 0;
+    float   wav_tfm_eps          = 1e-5f;
+    float   wav_tfm_rope_theta   = 10000.0f;
+    int32_t wav_upsample_n_block = 0;
+    int32_t wav_dac_n_block      = 0;
+    int32_t wav_dac_n_res        = 0;
+    int32_t wav_tfm_swa          = 0; // pre_transformer's KV cache size, in frames
+
    // mimo-v2.5: LLM-side connector (input_local_transformer)
    int32_t audio_local_n_layer = 0;
    int32_t audio_local_group_size = 0;
@@ -286,6 +299,14 @@ struct clip_layer {
    ggml_tensor * cross_attn_norm_w = nullptr;
    ggml_tensor * cross_attn_norm_b = nullptr;

+    // qwen3tts speaker encoder: SE-Res2Net block, tdnn1/tdnn2 reuse conv_pw1_w/b and conv_pw2_w/b above
+    ggml_tensor * se_conv1_w = nullptr;
+    ggml_tensor * se_conv1_b = nullptr;
+    ggml_tensor * se_conv2_w = nullptr;
+    ggml_tensor * se_conv2_b = nullptr;
+    std::vector<ggml_tensor *> res2_conv_w; // Res2Net hierarchical branches
+    std::vector<ggml_tensor *> res2_conv_b;
+
    bool has_deepstack() const {
        return deepstack_fc1_w != nullptr;
    }
@@ -365,6 +386,73 @@ struct qf_block {
    std::vector<clip_layer> qf_proj_layers;
 };

+// qwen3tts code2wav: RVQ codes -> raw PCM
+struct clip_code2wav {
+    // "upsample" stage: one ConvNeXt block plus the causal ConvTranspose1d before it
+    struct upsample_block {
+        ggml_tensor * conv_w   = nullptr; // causal ConvTranspose1d, 2x
+        ggml_tensor * conv_b   = nullptr;
+        ggml_tensor * dwconv_w = nullptr; // depthwise causal conv, k=7
+        ggml_tensor * dwconv_b = nullptr;
+        ggml_tensor * norm_w   = nullptr; // LayerNorm
+        ggml_tensor * norm_b   = nullptr;
+        ggml_tensor * pw1_w    = nullptr; // pointwise expand
+        ggml_tensor * pw1_b    = nullptr;
+        ggml_tensor * pw2_w    = nullptr; // pointwise project
+        ggml_tensor * pw2_b    = nullptr;
+        ggml_tensor * gamma    = nullptr; // layer scale
+    };
+
+    // one DAC residual unit: SnakeBeta -> dilated causal conv -> SnakeBeta -> pointwise causal conv
+    struct dac_res {
+        ggml_tensor * act1_alpha = nullptr;
+        ggml_tensor * act1_beta  = nullptr;
+        ggml_tensor * conv1_w    = nullptr;
+        ggml_tensor * conv1_b    = nullptr;
+        ggml_tensor * act2_alpha = nullptr;
+        ggml_tensor * act2_beta  = nullptr;
+        ggml_tensor * conv2_w    = nullptr;
+        ggml_tensor * conv2_b    = nullptr;
+    };
+
+    // one DAC upsample block (SnakeBeta -> causal ConvTranspose1d -> 3 residual units)
+    struct dac_block {
+        ggml_tensor * snake_alpha = nullptr;
+        ggml_tensor * snake_beta  = nullptr;
+        ggml_tensor * conv_w      = nullptr; // causal ConvTranspose1d
+        ggml_tensor * conv_b      = nullptr;
+        std::vector<dac_res> res;
+    };
+
+    // quantizer: RVQ codebook decode
+    ggml_tensor * quant_first_in_w  = nullptr; // semantic RVQ, in_proj (1x1 conv, loaded as 2D)
+    ggml_tensor * quant_first_out_w = nullptr;
+    ggml_tensor * quant_first_cb_w  = nullptr; // codebook (1 layer)
+    ggml_tensor * quant_rest_in_w   = nullptr; // acoustic RVQ
+    ggml_tensor * quant_rest_out_w  = nullptr;
+    ggml_tensor * quant_rest_cb_w   = nullptr; // codebooks, merged 3D [15, vocab, dim]
+
+    ggml_tensor * pre_conv_w = nullptr;
+    ggml_tensor * pre_conv_b = nullptr;
+
+    ggml_tensor * tfm_in_proj_w     = nullptr;
+    ggml_tensor * tfm_in_proj_b     = nullptr;
+    ggml_tensor * tfm_out_proj_w    = nullptr;
+    ggml_tensor * tfm_out_proj_b    = nullptr;
+    ggml_tensor * tfm_output_norm_w = nullptr;
+    std::vector<clip_layer> tfm_layers; // reuses the generic block fields (ln_1/attn/ln_2/ffn/ls_1/ls_2)
+
+    std::vector<upsample_block> upsample;
+
+    ggml_tensor * dac_entry_w = nullptr;
+    ggml_tensor * dac_entry_b = nullptr;
+    std::vector<dac_block> dac;
+    ggml_tensor * dac_post_snake_alpha = nullptr;
+    ggml_tensor * dac_post_snake_beta  = nullptr;
+    ggml_tensor * dac_post_conv_w      = nullptr;
+    ggml_tensor * dac_post_conv_b      = nullptr;
+};
+
 struct clip_model {
    clip_modality modality = CLIP_MODALITY_VISION;
    projector_type proj_type = PROJECTOR_TYPE_MLP;
@@ -577,6 +665,24 @@ struct clip_model {
    ggml_tensor * conv2d_3_w = nullptr;
    ggml_tensor * conv2d_3_b = nullptr;

+    // qwen3tts speaker encoder (ECAPA-TDNN)
+    // reused tensors: stem conv is conv1d_1_w/b, feature aggregation is conv_out_w/b, output proj is mm_fc_w/b
+    ggml_tensor * spk_asp_attn_w = nullptr;
+    ggml_tensor * spk_asp_attn_b = nullptr;
+    ggml_tensor * spk_asp_tdnn_w = nullptr;
+    ggml_tensor * spk_asp_tdnn_b = nullptr;
+
+    // qwen3tts code_predictor
+    ggml_tensor * gen_code_proj_in_w  = nullptr; // small_to_mtp_projection
+    ggml_tensor * gen_code_proj_in_b  = nullptr;
+    ggml_tensor * gen_code_embd_w     = nullptr; // per-codebook embedding, merged 3D
+    ggml_tensor * gen_code_head_w     = nullptr; // per-codebook output head, merged 3D
+    ggml_tensor * gen_code_out_embd_w = nullptr; // codebook-0 embedding, fed back into the talker
+    ggml_tensor * gen_code_norm_w     = nullptr; // final norm
+
+    // qwen3tts code2wav: RVQ codes -> raw PCM
+    clip_code2wav c2w;
+
    // cogvlm
    ggml_tensor * mm_post_fc_norm_w = nullptr;
    ggml_tensor * mm_post_fc_norm_b = nullptr;
@@ -17,6 +17,7 @@
 #include <cstring>
 #include <fstream>
 #include <map>
+#include <random>
 #include <stdexcept>
 #include <unordered_set>
 #include <vector>
@@ -269,6 +270,29 @@ clip_graph::clip_graph(clip_ctx * ctx, const clip_image_f32 & img) :
    gf = ggml_new_graph_custom(ctx0, ctx->max_nodes, false);
 }

+clip_graph::clip_graph(const clip_graph & parent) :
+        model(parent.model),
+        hparams(parent.hparams),
+        proj_type(parent.proj_type),
+        img(parent.img),
+        patch_size(parent.patch_size),
+        n_patches_x(parent.n_patches_x),
+        n_patches_y(parent.n_patches_y),
+        n_patches(parent.n_patches),
+        n_embd(parent.n_embd),
+        n_head(parent.n_head),
+        n_head_kv(parent.n_head_kv),
+        d_head(parent.d_head),
+        n_layer(parent.n_layer),
+        n_mmproj_embd(parent.n_mmproj_embd),
+        eps(parent.eps),
+        kq_scale(parent.kq_scale),
+        flash_attn_type(parent.flash_attn_type) {
+    // reuse from parent
+    ctx0 = parent.ctx0;
+    gf   = parent.gf;
+}
+
 ggml_tensor * clip_graph::build_mm(ggml_tensor * w, ggml_tensor * x) const {
    return ggml_mul_mat(ctx0, w, x);
 }
@@ -873,7 +897,8 @@ ggml_tensor * clip_graph::build_patch_merge_permute(ggml_tensor * cur, int scale
    return cur;
 }

-static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const clip_image_f32_batch & imgs) {
+static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const clip_image_f32_batch & imgs,
+                                                            const clip_encode_params * params = nullptr) {
    const clip_image_f32 & img = imgs.entries[0];
    std::unique_ptr<clip_graph> builder;

@@ -1025,6 +1050,17 @@ static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const
            {
                builder = std::make_unique<clip_graph_mimo_audio>(ctx, img);
            } break;
+        case PROJECTOR_TYPE_QWEN3TTS_SPKENC:
+            {
+                builder = std::make_unique<clip_graph_qwen3tts_spkenc>(ctx, img);
+            } break;
+        case PROJECTOR_TYPE_QWEN3TTS_GEN:
+            {
+                const auto  gen_process = params ? params->gen_process : CLIP_GEN_PROCESS_GEN_CODE;
+                const int   top_k = params ? params->top_k : 50;
+                const float top_p = params ? params->top_p : 1.0f;
+                builder = std::make_unique<clip_graph_qwen3tts_gen>(ctx, img, gen_process, top_k, top_p);
+            } break;
        case PROJECTOR_TYPE_YOUTUVL:
            {
                builder = std::make_unique<clip_graph_youtuvl>(ctx, img);
@@ -1065,8 +1101,9 @@ struct clip_model_loader {

    size_t model_size = 0; // in bytes

-    bool has_vision = false;
-    bool has_audio  = false;
+    bool has_vision    = false;
+    bool has_audio     = false;
+    bool has_gen_audio = false;

    mtmd_progress_callback progress_callback = nullptr;
    void * progress_callback_user_data = nullptr;
@@ -1112,8 +1149,9 @@ struct clip_model_loader {

        // modalities
        {
-            get_bool(KEY_HAS_VISION_ENC, has_vision, false);
-            get_bool(KEY_HAS_AUDIO_ENC,  has_audio,  false);
+            get_bool(KEY_HAS_VISION_ENC,    has_vision,    false);
+            get_bool(KEY_HAS_AUDIO_ENC,     has_audio,     false);
+            get_bool(KEY_HAS_GEN_AUDIO_ENC, has_gen_audio, false);

            if (has_vision) {
                LOG_INF("%s: has vision encoder\n", __func__);
@@ -1121,6 +1159,9 @@ struct clip_model_loader {
            if (has_audio) {
                LOG_INF("%s: has audio encoder\n", __func__);
            }
+            if (has_gen_audio) {
+                LOG_INF("%s: has audio generation (gen) encoder\n", __func__);
+            }
        }

        // tensors
@@ -1147,6 +1188,8 @@ struct clip_model_loader {
            GGML_ASSERT(has_vision);
        } else if (modality == CLIP_MODALITY_AUDIO) {
            GGML_ASSERT(has_audio);
+        } else if (modality == CLIP_MODALITY_GEN_AUDIO) {
+            GGML_ASSERT(has_gen_audio);
        }
        model.modality = modality;

@@ -1163,6 +1206,8 @@ struct clip_model_loader {
                    get_string(KEY_VISION_PROJ_TYPE, proj_type, false);
                } else if (modality == CLIP_MODALITY_AUDIO) {
                    get_string(KEY_AUDIO_PROJ_TYPE, proj_type, false);
+                } else if (modality == CLIP_MODALITY_GEN_AUDIO) {
+                    get_string(KEY_GEN_AUDIO_PROJ_TYPE, proj_type, false);
                } else {
                    GGML_ABORT("unknown modality");
                }
@@ -1182,12 +1227,13 @@ struct clip_model_loader {
            }
        }

-        const bool is_vision = model.modality == CLIP_MODALITY_VISION;
-        const bool is_audio  = model.modality == CLIP_MODALITY_AUDIO;
+        const bool is_vision    = model.modality == CLIP_MODALITY_VISION;
+        const bool is_audio     = model.modality == CLIP_MODALITY_AUDIO;
+        const bool is_gen_audio = model.modality == CLIP_MODALITY_GEN_AUDIO;

        // other hparams
        {
-            const char * prefix = is_vision ? "vision" : "audio";
+            const char * prefix = is_vision ? "vision" : (is_audio ? "audio" : "gen.audio");
            get_u32(string_format(KEY_N_EMBD,         prefix), hparams.n_embd);
            get_u32(string_format(KEY_N_HEAD,         prefix), hparams.n_head);
            get_u32(string_format(KEY_N_EMBD_HEAD,    prefix), hparams.n_embd_head, false);
@@ -1198,6 +1244,7 @@ struct clip_model_loader {

            // n_head_kv is optional (for GQA), default to n_head
            hparams.n_head_kv = hparams.n_head;
+            get_u32(string_format(KEY_N_HEAD_KV, prefix), hparams.n_head_kv, false);

            if (is_vision) {
                get_u32(KEY_IMAGE_SIZE, hparams.image_size);
@@ -1226,6 +1273,11 @@ struct clip_model_loader {
                hparams.image_size = 0;
                hparams.patch_size = 1;

+            } else if (is_gen_audio) {
+                // these are unused, but still need to be set to avoid issues
+                hparams.image_size = 0;
+                hparams.patch_size = 1;
+
            } else {
                GGML_ASSERT(false && "unknown modality");
            }
@@ -1647,6 +1699,33 @@ struct clip_model_loader {
                                "%s: mimo_audio: %s must be > 0\n", __func__, KEY_A_LOCAL_GROUP_SIZE));
                        }
                    } break;
+                case PROJECTOR_TYPE_QWEN3TTS_SPKENC:
+                    {
+                        // ECAPA-TDNN speaker encoder, mel front-end uses the Slaney default (fmin=0, fmax=sr/2)
+                        hparams.audio_sample_rate = 24000;
+                        hparams.audio_n_fft       = 1024;
+                        hparams.audio_window_len  = 1024;
+                        hparams.audio_hop_len     = 256;
+                    } break;
+                case PROJECTOR_TYPE_QWEN3TTS_GEN:
+                    {
+                        // TODO: hardcoded for now, read from code_predictor_config instead
+                        hparams.rope_theta = 1000000.0f;
+
+                        // code2wav params
+                        hparams.wav_tfm_n_layer      = 8;
+                        hparams.wav_tfm_n_embd       = 512;
+                        hparams.wav_tfm_n_ff         = 1024;
+                        hparams.wav_tfm_n_head       = 16;
+                        hparams.wav_tfm_n_head_kv    = 16;
+                        hparams.wav_tfm_eps          = 1e-5f;
+                        hparams.wav_tfm_rope_theta   = 10000.0f;
+                        hparams.wav_upsample_n_block = 2;
+                        hparams.wav_dac_n_block      = 4;
+                        hparams.wav_dac_n_res        = 3;
+                        // matches the reference decoder's sliding_window (speech_tokenizer/config.json)
+                        hparams.wav_tfm_swa = 72;
+                    } break;
                case PROJECTOR_TYPE_PADDLEOCR:
                    {
                        hparams.n_merge = 2;
@@ -1682,6 +1761,10 @@ struct clip_model_loader {
                            // qwen2 encoder is GQA, requires KEY_N_HEAD_KV
                            get_u32(string_format(KEY_N_HEAD_KV, "vision"), hparams.n_head_kv);
                        }
+                        // unlimited-ocr shares the v1 projector but tiles up to 32
+                        get_u32(KEY_PREPROC_MIN_TILES, hparams.preproc_min_tiles, false);
+                        get_u32(KEY_PREPROC_MAX_TILES, hparams.preproc_max_tiles, false);
+                        GGML_ASSERT(hparams.preproc_min_tiles <= hparams.preproc_max_tiles);
                     } break;
                case PROJECTOR_TYPE_HUNYUANVL:
                    {
@@ -1830,6 +1913,9 @@ struct clip_model_loader {
                if (hparams.image_max_pixels > 0) {
                    LOG_INF("%s: image_max_pixels:   %d%s\n", __func__, hparams.image_max_pixels, hparams.custom_image_max_tokens > 0 ? " (custom value)" : "");
                }
+                if (hparams.preproc_max_tiles > 0) {
+                    LOG_INF("%s: preproc_tiles:      %d - %d\n", __func__, hparams.preproc_min_tiles, hparams.preproc_max_tiles);
+                }
            } else if (is_audio) {
                LOG_INF("\n--- audio hparams ---\n");
                LOG_INF("%s: n_mel_bins:         %d\n", __func__, hparams.n_mel_bins);
@@ -1871,7 +1957,9 @@ struct clip_model_loader {
        }

        // TODO @ngxson : support both audio and video in the future
-        const char * prefix = model.modality == CLIP_MODALITY_AUDIO ? "a" : "v";
+        const char * prefix = model.modality == CLIP_MODALITY_AUDIO ? "a"
+                             : model.modality == CLIP_MODALITY_GEN_AUDIO ? "a.gen.code"
+                             : "v";

        // get offsets
        for (int64_t i = 0; i < gguf_get_n_tensors(ctx_gguf.get()); ++i) {
@@ -1973,7 +2061,8 @@ struct clip_model_loader {
        model.position_embeddings = get_tensor(string_format(TN_POS_EMBD, prefix), false);

        const bool has_standard_layers = (
-            model.proj_type != PROJECTOR_TYPE_GEMMA3NV);
+            model.proj_type != PROJECTOR_TYPE_GEMMA3NV &&
+            model.proj_type != PROJECTOR_TYPE_QWEN3TTS_SPKENC);

        // layers
        const int n_layers_to_load = has_standard_layers ? hparams.n_layer : 0;
@@ -2599,6 +2688,144 @@ struct clip_model_loader {
                    model.mm_1_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 1, "weight"));
                    model.mm_2_w = get_tensor(string_format(TN_MM_AUDIO_MLP, 2, "weight"));
                } break;
+            case PROJECTOR_TYPE_QWEN3TTS_SPKENC:
+                {
+                    // stem TDNN (block 0)
+                    model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 0, "weight"));
+                    model.conv1d_1_b = get_tensor(string_format(TN_CONV1D, 0, "bias"));
+
+                    // SE-Res2Net blocks (GGUF bid 1..3, one per hparams.n_layer)
+                    model.layers.resize(hparams.n_layer);
+                    for (int il = 0; il < hparams.n_layer; il++) {
+                        auto & layer = model.layers[il];
+                        int bid = il + 1;
+                        layer.conv_pw1_w = get_tensor(string_format(TN_CONV_PW1, prefix, bid, "weight"));
+                        layer.conv_pw1_b = get_tensor(string_format(TN_CONV_PW1, prefix, bid, "bias"));
+                        layer.conv_pw2_w = get_tensor(string_format(TN_CONV_PW2, prefix, bid, "weight"));
+                        layer.conv_pw2_b = get_tensor(string_format(TN_CONV_PW2, prefix, bid, "bias"));
+                        layer.se_conv1_w = get_tensor(string_format(TN_A_SE_CONV1, bid, "weight"));
+                        layer.se_conv1_b = get_tensor(string_format(TN_A_SE_CONV1, bid, "bias"));
+                        layer.se_conv2_w = get_tensor(string_format(TN_A_SE_CONV2, bid, "weight"));
+                        layer.se_conv2_b = get_tensor(string_format(TN_A_SE_CONV2, bid, "bias"));
+                        layer.res2_conv_w.resize(7);
+                        layer.res2_conv_b.resize(7);
+                        for (int xid = 0; xid < 7; xid++) {
+                            layer.res2_conv_w[xid] = get_tensor(string_format(TN_A_CONV_RES2, bid, xid, "weight"));
+                            layer.res2_conv_b[xid] = get_tensor(string_format(TN_A_CONV_RES2, bid, xid, "bias"));
+                        }
+                    }
+
+                    // multi-layer feature aggregation
+                    model.conv_out_w = get_tensor(string_format(TN_CONV_OUT, "weight"));
+                    model.conv_out_b = get_tensor(string_format(TN_CONV_OUT, "bias"));
+
+                    // attentive statistics pooling
+                    model.spk_asp_attn_w = get_tensor(string_format(TN_A_ASP_ATTN, "weight"));
+                    model.spk_asp_attn_b = get_tensor(string_format(TN_A_ASP_ATTN, "bias"));
+                    model.spk_asp_tdnn_w = get_tensor(string_format(TN_A_ASP_TDNN, "weight"));
+                    model.spk_asp_tdnn_b = get_tensor(string_format(TN_A_ASP_TDNN, "bias"));
+
+                    // final speaker embedding projection
+                    model.mm_fc_w = get_tensor(string_format(TN_MM_AUDIO_FC, "weight"));
+                    model.mm_fc_b = get_tensor(string_format(TN_MM_AUDIO_FC, "bias"));
+                } break;
+            case PROJECTOR_TYPE_QWEN3TTS_GEN:
+                {
+                    // code_predictor
+                    model.gen_code_proj_in_w = get_tensor(string_format(TN_A_GEN_CODE_PROJ_IN, "weight"));
+                    model.gen_code_proj_in_b = get_tensor(string_format(TN_A_GEN_CODE_PROJ_IN, "bias"));
+                    model.gen_code_embd_w     = get_tensor(string_format(TN_A_GEN_CODE_EMBD,     "weight"));
+                    model.gen_code_head_w     = get_tensor(string_format(TN_A_GEN_CODE_HEAD,     "weight"));
+                    model.gen_code_out_embd_w = get_tensor(string_format(TN_A_GEN_CODE_OUT_EMBD, "weight"));
+                    model.gen_code_norm_w     = get_tensor(string_format(TN_A_GEN_CODE_NORM,     "weight"));
+
+                    // code2wav: RVQ codes -> raw PCM, lives in the same ctx as code_predictor
+                    {
+                        auto & c2w = model.c2w;
+
+                        c2w.quant_first_in_w  = get_tensor(string_format(TN_A_GEN_WAV_QUANT_FIRST_IN,  "weight"));
+                        c2w.quant_first_out_w = get_tensor(string_format(TN_A_GEN_WAV_QUANT_FIRST_OUT, "weight"));
+                        c2w.quant_first_cb_w  = get_tensor(string_format(TN_A_GEN_WAV_QUANT_FIRST_CB,  "weight"));
+                        c2w.quant_rest_in_w   = get_tensor(string_format(TN_A_GEN_WAV_QUANT_REST_IN,   "weight"));
+                        c2w.quant_rest_out_w  = get_tensor(string_format(TN_A_GEN_WAV_QUANT_REST_OUT,  "weight"));
+                        c2w.quant_rest_cb_w   = get_tensor(string_format(TN_A_GEN_WAV_QUANT_REST_CB,   "weight"));
+
+                        c2w.pre_conv_w = get_tensor(string_format(TN_A_GEN_WAV_PRE_CONV, "weight"));
+                        c2w.pre_conv_b = get_tensor(string_format(TN_A_GEN_WAV_PRE_CONV, "bias"));
+
+                        c2w.tfm_in_proj_w     = get_tensor(string_format(TN_A_GEN_WAV_TFM_IN_PROJ,  "weight"));
+                        c2w.tfm_in_proj_b     = get_tensor(string_format(TN_A_GEN_WAV_TFM_IN_PROJ,  "bias"));
+                        c2w.tfm_out_proj_w    = get_tensor(string_format(TN_A_GEN_WAV_TFM_OUT_PROJ, "weight"));
+                        c2w.tfm_out_proj_b    = get_tensor(string_format(TN_A_GEN_WAV_TFM_OUT_PROJ, "bias"));
+                        c2w.tfm_output_norm_w = get_tensor(string_format(TN_A_GEN_WAV_TFM_OUT_NORM, "weight"));
+
+                        // loaded manually, the generic model.layers loop is taken by code_predictor
+                        c2w.tfm_layers.resize(hparams.wav_tfm_n_layer);
+                        for (int il = 0; il < hparams.wav_tfm_n_layer; il++) {
+                            auto & layer = c2w.tfm_layers[il];
+                            const char * p = "a.gen.wav.tfm";
+                            layer.q_w      = get_tensor(string_format(TN_ATTN_Q,      p, il, "weight"));
+                            layer.k_w      = get_tensor(string_format(TN_ATTN_K,      p, il, "weight"));
+                            layer.v_w      = get_tensor(string_format(TN_ATTN_V,      p, il, "weight"));
+                            layer.o_w      = get_tensor(string_format(TN_ATTN_OUTPUT, p, il, "weight"));
+                            layer.ln_1_w   = get_tensor(string_format(TN_LN_1,        p, il, "weight"));
+                            layer.ln_2_w   = get_tensor(string_format(TN_LN_2,        p, il, "weight"));
+                            layer.ls_1_w   = get_tensor(string_format(TN_LS_1,        p, il, "weight"));
+                            layer.ls_2_w   = get_tensor(string_format(TN_LS_2,        p, il, "weight"));
+                            layer.ff_gate_w = get_tensor(string_format(TN_FFN_GATE,   p, il, "weight"));
+                            layer.ff_up_w   = get_tensor(string_format(TN_FFN_UP,     p, il, "weight"));
+                            layer.ff_down_w = get_tensor(string_format(TN_FFN_DOWN,   p, il, "weight"));
+                        }
+
+                        // upsample: 2x (causal ConvTranspose1d + ConvNeXt block)
+                        c2w.upsample.resize(hparams.wav_upsample_n_block);
+                        for (int il = 0; il < hparams.wav_upsample_n_block; il++) {
+                            auto & up = c2w.upsample[il];
+                            up.conv_w   = get_tensor(string_format(TN_A_GEN_WAV_UP_CONV,   il, "weight"));
+                            up.conv_b   = get_tensor(string_format(TN_A_GEN_WAV_UP_CONV,   il, "bias"));
+                            up.dwconv_w = get_tensor(string_format(TN_A_GEN_WAV_UP_DWCONV, il, "weight"));
+                            up.dwconv_b = get_tensor(string_format(TN_A_GEN_WAV_UP_DWCONV, il, "bias"));
+                            up.norm_w   = get_tensor(string_format(TN_A_GEN_WAV_UP_NORM,   il, "weight"));
+                            up.norm_b   = get_tensor(string_format(TN_A_GEN_WAV_UP_NORM,   il, "bias"));
+                            up.pw1_w    = get_tensor(string_format(TN_A_GEN_WAV_UP_PW1,    il, "weight"));
+                            up.pw1_b    = get_tensor(string_format(TN_A_GEN_WAV_UP_PW1,    il, "bias"));
+                            up.pw2_w    = get_tensor(string_format(TN_A_GEN_WAV_UP_PW2,    il, "weight"));
+                            up.pw2_b    = get_tensor(string_format(TN_A_GEN_WAV_UP_PW2,    il, "bias"));
+                            up.gamma    = get_tensor(string_format(TN_A_GEN_WAV_UP_GAMMA,  il));
+                        }
+
+                        // DAC decoder: conv_pre + n upsample blocks (each with n_res residual units) + conv_post
+                        c2w.dac_entry_w = get_tensor(string_format(TN_A_GEN_WAV_DAC_ENTRY, "weight"));
+                        c2w.dac_entry_b = get_tensor(string_format(TN_A_GEN_WAV_DAC_ENTRY, "bias"));
+
+                        c2w.dac.resize(hparams.wav_dac_n_block);
+                        for (int il = 0; il < hparams.wav_dac_n_block; il++) {
+                            auto & blk = c2w.dac[il];
+                            blk.snake_alpha = get_tensor(string_format(TN_A_GEN_WAV_DAC_SNAKE, il, "alpha"));
+                            blk.snake_beta  = get_tensor(string_format(TN_A_GEN_WAV_DAC_SNAKE, il, "beta"));
+                            blk.conv_w      = get_tensor(string_format(TN_A_GEN_WAV_DAC_CONV,  il, "weight"));
+                            blk.conv_b      = get_tensor(string_format(TN_A_GEN_WAV_DAC_CONV,  il, "bias"));
+
+                            blk.res.resize(hparams.wav_dac_n_res);
+                            for (int ir = 0; ir < hparams.wav_dac_n_res; ir++) {
+                                auto & res = blk.res[ir];
+                                res.act1_alpha = get_tensor(string_format(TN_A_GEN_WAV_DAC_RES_ACT1,  il, ir, "alpha"));
+                                res.act1_beta  = get_tensor(string_format(TN_A_GEN_WAV_DAC_RES_ACT1,  il, ir, "beta"));
+                                res.conv1_w    = get_tensor(string_format(TN_A_GEN_WAV_DAC_RES_CONV1, il, ir, "weight"));
+                                res.conv1_b    = get_tensor(string_format(TN_A_GEN_WAV_DAC_RES_CONV1, il, ir, "bias"));
+                                res.act2_alpha = get_tensor(string_format(TN_A_GEN_WAV_DAC_RES_ACT2,  il, ir, "alpha"));
+                                res.act2_beta  = get_tensor(string_format(TN_A_GEN_WAV_DAC_RES_ACT2,  il, ir, "beta"));
+                                res.conv2_w    = get_tensor(string_format(TN_A_GEN_WAV_DAC_RES_CONV2, il, ir, "weight"));
+                                res.conv2_b    = get_tensor(string_format(TN_A_GEN_WAV_DAC_RES_CONV2, il, ir, "bias"));
+                            }
+                        }
+
+                        c2w.dac_post_snake_alpha = get_tensor(string_format(TN_A_GEN_WAV_DAC_POST_SNAKE, "alpha"));
+                        c2w.dac_post_snake_beta  = get_tensor(string_format(TN_A_GEN_WAV_DAC_POST_SNAKE, "beta"));
+                        c2w.dac_post_conv_w      = get_tensor(string_format(TN_A_GEN_WAV_DAC_POST_CONV,  "weight"));
+                        c2w.dac_post_conv_b      = get_tensor(string_format(TN_A_GEN_WAV_DAC_POST_CONV,  "bias"));
+                    }
+                } break;
            case PROJECTOR_TYPE_VOXTRAL:
                {
                    model.conv1d_1_w = get_tensor(string_format(TN_CONV1D, 1, "weight"));
@@ -3427,6 +3654,7 @@ struct clip_model_loader {
 struct clip_init_result clip_init(const char * fname, struct clip_context_params ctx_params) {
    clip_ctx * ctx_vision = nullptr;
    clip_ctx * ctx_audio = nullptr;
+    clip_ctx * ctx_gen_audio = nullptr;

    try {
        clip_model_loader loader(fname,
@@ -3459,16 +3687,25 @@ struct clip_init_result clip_init(const char * fname, struct clip_context_params
            }
        }

+        if (loader.has_gen_audio) {
+            ctx_gen_audio = new clip_ctx(ctx_params);
+            loader.load_hparams(ctx_gen_audio->model, CLIP_MODALITY_GEN_AUDIO);
+            loader.load_tensors(*ctx_gen_audio);
+            // TODO: fix warmup
+            ctx_gen_audio->buf_compute_meta.resize(ctx_gen_audio->max_nodes * ggml_tensor_overhead() + ggml_graph_overhead());
+        }
+
    } catch (const std::exception & e) {
        LOG_ERR("%s: failed to load model '%s': %s\n", __func__, fname, e.what());

        delete ctx_vision;
        delete ctx_audio;
+        delete ctx_gen_audio;

-        return {nullptr, nullptr};
+        return {nullptr, nullptr, nullptr};
    }

-    return {ctx_vision, ctx_audio};
+    return {ctx_vision, ctx_audio, ctx_gen_audio};
 }

 struct clip_cap clip_get_cap(const char * fname) {
@@ -3784,6 +4021,16 @@ int clip_n_output_tokens(const clip_ctx * ctx, const clip_image_f32 * img) {
                const int ds = ctx->model.hparams.audio_proj_downsample_rate;
                n_patches = ((img->nx() + ws - 1) / ws) * (ws / ds);
            } break;
+        case PROJECTOR_TYPE_QWEN3TTS_SPKENC:
+            {
+                // pooling gives one speaker embedding, whatever the clip length is
+                n_patches = 1;
+            } break;
+        case PROJECTOR_TYPE_QWEN3TTS_GEN:
+            {
+                // one hidden-state vector fed back to the talker per call
+                n_patches = 1;
+            } break;
        case PROJECTOR_TYPE_GRANITE4_VISION:
            {
                // Per-tile output token count: each projector block outputs
@@ -3817,7 +4064,16 @@ bool clip_image_encode(struct clip_ctx * ctx, int n_threads, const clip_image_f3
 }

 bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32_batch * imgs_c_ptr, std::vector<float> & out_batch_embd) {
-    const clip_image_f32_batch & imgs = *imgs_c_ptr;
+    clip_encode_params params;
+    params.imgs = imgs_c_ptr;
+    params.n_threads = n_threads;
+    params.out_embd = &out_batch_embd;
+
+    return clip_encode(ctx, &params);
+}
+
+bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
+    const clip_image_f32_batch & imgs = *params->imgs;
    int n_batch_cur = imgs.entries.size();

    // [QWEN_VIDEO] for video models, the batch dimension is used as temporal dimension for merged frames
@@ -3828,12 +4084,12 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32

    // if buffers are not allocated, we need to do a warmup run to allocate them
    if (!ctx->is_allocated) {
-        clip_model_loader::warmup(*ctx, *imgs_c_ptr);
+        clip_model_loader::warmup(*ctx, *params->imgs);
    }

    // build the inference graph
    ggml_backend_sched_reset(ctx->sched.get());
-    ggml_cgraph * gf = clip_get_graph_builder(ctx, imgs)->build();
+    ggml_cgraph * gf = clip_get_graph_builder(ctx, imgs, params)->build();
    ggml_backend_sched_alloc_graph(ctx->sched.get(), gf);

    // set inputs
@@ -3918,8 +4174,8 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32
        }
        set_input_f32("inp_raw", inp_raw);

-    } else {
-        // audio input
+    } else if (!(ctx->proj_type() == PROJECTOR_TYPE_QWEN3TTS_GEN && params->gen_process == CLIP_GEN_PROCESS_GEN_WAV)) {
+        // audio input, code2wav is not here: its only input is "inp_codes", set in the switch below
        GGML_ASSERT(imgs.entries.size() == 1);

        const auto & mel_inp = imgs.entries[0];
@@ -4475,9 +4731,77 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32
        case PROJECTOR_TYPE_COGVLM:
        case PROJECTOR_TYPE_YASA2:
        case PROJECTOR_TYPE_GEMMA4UA:
+        case PROJECTOR_TYPE_QWEN3TTS_SPKENC:
            {
                // do nothing
            } break;
+        case PROJECTOR_TYPE_QWEN3TTS_GEN:
+            {
+                if (params->gen_process == CLIP_GEN_PROCESS_GEN_WAV) {
+                    GGML_ASSERT(params->codes != nullptr);
+
+                    // frame-major input to group-major, rear-padded with code 0 up to one window
+                    const int64_t n_codes  = model.gen_code_head_w->ne[2] + 1;
+                    const int64_t n_frames_w = hparams.wav_tfm_swa;
+                    const int64_t n_frames   = (int64_t) params->codes->size() / n_codes;
+                    GGML_ASSERT(n_frames > 0 && n_frames <= n_frames_w);
+
+                    // codes are used as ggml_get_rows indices, so check them against the codebook vocab
+                    const int64_t vocab_first = model.c2w.quant_first_cb_w->ne[1];
+                    const int64_t vocab_rest  = model.c2w.quant_rest_cb_w->ne[1];
+                    for (int64_t f = 0; f < n_frames; f++) {
+                        for (int64_t g = 0; g < n_codes; g++) {
+                            const int32_t c = (*params->codes)[f * n_codes + g];
+                            const int64_t vocab = (g == 0) ? vocab_first : vocab_rest;
+                            if (c < 0 || (int64_t) c >= vocab) {
+                                LOG_ERR("%s: code out of range (frame %lld, group %lld, code %d, vocab %lld)\n",
+                                        __func__, (long long) f, (long long) g, c, (long long) vocab);
+                                return false;
+                            }
+                        }
+                    }
+
+                    std::vector<int32_t> codes(n_frames_w * n_codes, 0);
+                    for (int64_t f = 0; f < n_frames; f++) {
+                        for (int64_t g = 0; g < n_codes; g++) {
+                            codes[g * n_frames_w + f] = (*params->codes)[f * n_codes + g];
+                        }
+                    }
+                    set_input_i32("inp_codes", codes);
+
+                    // upload the state from the previous call, or zero-fill on a cold start
+                    size_t offset = 0;
+                    for (const auto & slot : list_c2w_state_slots(hparams, model)) {
+                        ggml_tensor * t = get_inp_tensor(("state_in_" + slot.name).c_str());
+                        const size_t nb = ggml_nbytes(t);
+                        if (params->state_in && params->state_in->size() >= offset + nb) {
+                            ggml_backend_tensor_set(t, params->state_in->data() + offset, 0, nb);
+                        } else {
+                            std::vector<uint8_t> zeros(nb, 0);
+                            ggml_backend_tensor_set(t, zeros.data(), 0, nb);
+                        }
+                        offset += nb;
+                    }
+                } else {
+                    // code0 indexes gen_code_out_embd_w via ggml_get_rows; bound it
+                    const int64_t vocab0 = model.gen_code_out_embd_w->ne[1];
+                    if (params->code0 < 0 || (int64_t) params->code0 >= vocab0) {
+                        LOG_ERR("%s: code0 out of range (%d, vocab %lld)\n", __func__, params->code0, (long long) vocab0);
+                        return false;
+                    }
+                    std::vector<int32_t> code0 = { params->code0 };
+                    set_input_i32("inp_code0", code0);
+
+                    // one uniform(0,1) draw per codebook, used by do_sampling()
+                    static std::mt19937 rng{ std::random_device{}() };
+                    std::uniform_real_distribution<float> dist(0.0f, 1.0f);
+                    const int64_t n_acoustic = model.gen_code_head_w->ne[2];
+                    for (int64_t g = 0; g < n_acoustic; g++) {
+                        std::vector<float> r = { dist(rng) };
+                        set_input_f32(("inp_rand_" + std::to_string(g)).c_str(), r);
+                    }
+                }
+            } break;
        case PROJECTOR_TYPE_HUNYUANVL:
            {
                // Compute the HunyuanVL 2D position embedding on CPU (with the
@@ -4883,7 +5207,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32
    if (reg) {
        auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t) ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
        if (ggml_backend_set_n_threads_fn) {
-            ggml_backend_set_n_threads_fn(ctx->backend_cpu, n_threads);
+            ggml_backend_set_n_threads_fn(ctx->backend_cpu, params->n_threads);
        }
    }

@@ -4893,34 +5217,90 @@ bool clip_image_batch_encode(clip_ctx * ctx, int n_threads, const clip_image_f32
        return false;
    }

-    // the last node is the embedding tensor
-    ggml_tensor * embeddings = ggml_graph_node(gf, -1);
+    // the last node is the embedding tensor, code2wav has no out_embd
+    ggml_tensor * embeddings = params->out_embd ? ggml_graph_node(gf, -1) : nullptr;

-    // sanity check (assuming that all images in batch have the same number of tokens, so we only check the first one)
-    const int n_tokens_out = embeddings->ne[1];
-    const int expected_n_tokens_out = clip_n_output_tokens(ctx, &imgs.entries[0]);
-    if (n_tokens_out != expected_n_tokens_out) {
-        LOG_ERR("%s: expected output %d tokens, got %d\n", __func__, expected_n_tokens_out, n_tokens_out);
-        GGML_ABORT("Invalid number of output tokens");
-    }
+    if (embeddings != nullptr) {
+        // sanity check (assuming that all images in batch have the same number of tokens, so we only check the first one)
+        const int n_tokens_out = embeddings->ne[1];
+        const int expected_n_tokens_out = clip_n_output_tokens(ctx, &imgs.entries[0]);
+        if (n_tokens_out != expected_n_tokens_out) {
+            LOG_ERR("%s: expected output %d tokens, got %d\n", __func__, expected_n_tokens_out, n_tokens_out);
+            GGML_ABORT("Invalid number of output tokens");
+        }

-    LOG_DBG("%s: output embedding shape [%d, %d, %d]\n", __func__,
-        (int)embeddings->ne[0], (int)embeddings->ne[1], (int)embeddings->ne[2]);
+        LOG_DBG("%s: output embedding shape [%d, %d, %d]\n", __func__,
+            (int)embeddings->ne[0], (int)embeddings->ne[1], (int)embeddings->ne[2]);

-    // copy output to user buffer if provided
-    // if output is empty, skip the copy
-    if (!out_batch_embd.empty()) {
-        if (out_batch_embd.size() != (size_t)ggml_nelements(embeddings)) {
-            LOG_ERR("%s: output buffer has %zu elements but expected %zu\n", __func__, out_batch_embd.size(), (size_t)ggml_nelements(embeddings));
-            GGML_ABORT("Output buffer size mismatch");
+        // copy output to user buffer if provided
+        // if output is empty, skip the copy
+        auto & out_batch_embd = *params->out_embd;
+        if (!out_batch_embd.empty()) {
+            if (out_batch_embd.size() != (size_t)ggml_nelements(embeddings)) {
+                LOG_ERR("%s: output buffer has %zu elements but expected %zu\n", __func__, out_batch_embd.size(), (size_t)ggml_nelements(embeddings));
+                GGML_ABORT("Output buffer size mismatch");
+            }
+            ggml_backend_tensor_get(embeddings, out_batch_embd.data(), 0, ggml_nbytes(embeddings));
+        } else {
+            LOG_WRN("%s: output buffer is empty, skipping copy\n", __func__);
        }
-        ggml_backend_tensor_get(embeddings, out_batch_embd.data(), 0, ggml_nbytes(embeddings));
-    } else {
-        LOG_WRN("%s: output buffer is empty, skipping copy\n", __func__);
    }

+    //
+    // for audio gen models
+    //
+
+    if (params->out_codes != nullptr) {
+        ggml_tensor * codes = ggml_graph_get_tensor(gf, "out_codes");
+        if (codes == nullptr) {
+            GGML_ABORT("out_codes requested but graph has no \"out_codes\" tensor");
+        }
+        auto & out_codes = *params->out_codes;
+        out_codes.resize(ggml_nelements(codes));
+        ggml_backend_tensor_get(codes, out_codes.data(), 0, ggml_nbytes(codes));
+    }
+    if (params->out_audio != nullptr) {
+        ggml_tensor * audio = ggml_graph_get_tensor(gf, "out_audio");
+        if (audio == nullptr) {
+            GGML_ABORT("out_audio requested but graph has no \"out_audio\" tensor");
+        }
+        auto & out_audio = *params->out_audio;
+        out_audio.resize(ggml_nelements(audio));
+        ggml_backend_tensor_get(audio, out_audio.data(), 0, ggml_nbytes(audio));
+
+        // drop the tail audio that comes from the code-0 rear padding
+        const int64_t n_codes    = model.gen_code_head_w->ne[2] + 1;
+        const int64_t n_frames_w = hparams.wav_tfm_swa;
+        const int64_t n_frames   = (int64_t) params->codes->size() / n_codes;
+        if (n_frames < n_frames_w) {
+            const size_t hop = out_audio.size() / n_frames_w;
+            out_audio.resize((size_t) n_frames * hop);
+        }
+    }
+    if (params->state_out != nullptr) {
+        auto & state_out = *params->state_out;
+        size_t total = 0;
+        for (const auto & slot : list_c2w_state_slots(hparams, model)) {
+            total += (size_t) (slot.ne0 * slot.ne1) * sizeof(float);
+        }
+        state_out.resize(total);
+        size_t offset = 0;
+        for (const auto & slot : list_c2w_state_slots(hparams, model)) {
+            ggml_tensor * t = ggml_graph_get_tensor(gf, ("state_out_" + slot.name).c_str());
+            if (t == nullptr) {
+                GGML_ABORT("state_out requested but graph has no \"state_out_%s\" tensor", slot.name.c_str());
+            }
+            const size_t nb = ggml_nbytes(t);
+            ggml_backend_tensor_get(t, state_out.data() + offset, 0, nb);
+            offset += nb;
+        }
+    }
+
+    //
    // Debug: dump final embeddings if MTMD_DEBUG_EMBEDDINGS is set
-    if (ctx->debug_output_embeddings) {
+    //
+
+    if (ctx->debug_output_embeddings && embeddings != nullptr) {
        const int64_t n_embd = embeddings->ne[0];
        const int64_t n_tokens = embeddings->ne[1];
        std::vector<float> emb_data(ggml_nelements(embeddings));
@@ -5047,6 +5427,10 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
            return ctx->model.mm_ffn_down_w->ne[1];
        case PROJECTOR_TYPE_MIMO_AUDIO:
            return ctx->model.mm_2_w->ne[1];
+        case PROJECTOR_TYPE_QWEN3TTS_SPKENC:
+            return ctx->model.mm_fc_w->ne[2];
+        case PROJECTOR_TYPE_QWEN3TTS_GEN:
+            return ctx->model.gen_code_out_embd_w->ne[0];
        case PROJECTOR_TYPE_PARAKEET:
            return ctx->model.mm_1_w->ne[1];
        default:
@@ -37,6 +37,7 @@ struct clip_image_f32_batch;
 enum clip_modality {
    CLIP_MODALITY_VISION,
    CLIP_MODALITY_AUDIO,
+    CLIP_MODALITY_GEN_AUDIO,
 };

 enum clip_flash_attn_type {
@@ -61,6 +62,7 @@ struct clip_context_params {
 struct clip_init_result {
    struct clip_ctx * ctx_v; // vision context
    struct clip_ctx * ctx_a; // audio context
+    struct clip_ctx * ctx_gen_a; // audio generation context
 };

 struct clip_init_result clip_init(const char * fname, struct clip_context_params ctx_params);
@@ -84,6 +86,33 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, const clip_image_f32 * img, std::vector<float> & out_vec);
 bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, std::vector<float> & out_batch_embd);

+enum clip_gen_process_type {
+    CLIP_GEN_PROCESS_GEN_UNKNOWN,
+    CLIP_GEN_PROCESS_GEN_CODE, // h_state to codes
+    CLIP_GEN_PROCESS_GEN_WAV,  // codes to raw PCM audio
+};
+struct clip_encode_params {
+    int n_threads = 1;
+    const clip_image_f32_batch * imgs = nullptr;
+    std::vector<float> * out_embd = nullptr;
+
+    // for audio gen, imgs has exactly one entry: hidden state from backbone (GEN_CODE) or unused (GEN_WAV)
+    clip_gen_process_type gen_process = CLIP_GEN_PROCESS_GEN_UNKNOWN;
+
+    // GEN_CODE: out_embd receives the embd to feed back to the backbone
+    int32_t code0 = 0; // semantic code sampled by the backbone
+    int32_t top_k = 50;
+    float   top_p = 1.0f;
+    std::vector<int32_t> * out_codes = nullptr; // this frame's 16 sampled codes
+
+    // GEN_WAV
+    const std::vector<int32_t> * codes = nullptr;     // this frame's 16 RVQ codes
+    std::vector<float> * out_audio = nullptr;         // decoded PCM samples, F32
+    const std::vector<uint8_t> * state_in  = nullptr; // state from previous call, null or wrong size means cold start
+    std::vector<uint8_t> *       state_out = nullptr; // state for the next call
+};
+bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params);
+
 bool clip_is_llava(const struct clip_ctx * ctx);
 // note for contributor: this clip_is_(model) pattern is deprecated
 //                       do NOT add new functions like this
@@ -253,6 +253,9 @@ ggml_cgraph * clip_graph_deepseekocr::build() {

    bool is_overview = img.add_viewsep;
    int n_tiles_per_row = 0;
+    // number of separate "row" images batched together in this graph call
+    // (captured now, before n_batch below gets repurposed as the SAM/ViT batch size)
+    const int n_rows_batch = n_batch;

    // note: we expect either a batch of rows or a batch of overviews, but not a mix of both

@@ -272,16 +275,18 @@ ggml_cgraph * clip_graph_deepseekocr::build() {
        GGML_ASSERT(img.ny() % img.nx() == 0);
        n_tiles_per_row = img.ny() / img.nx();

-        // input shape: [tile_size, tile_size * n_tiles_per_row, 3]
-        // we want to reshape it to [tile_size, tile_size, 3, n_tiles_per_row]
-        inp_raw = ggml_reshape_4d(ctx0, inp_raw, img.nx(), img.nx(), n_tiles_per_row, 3);
-        inp_raw = ggml_cont(ctx0, ggml_permute(ctx0, inp_raw, 0, 1, 3, 2));
+        // each entry is one "row" image of shape [tile_size, tile_size * n_tiles_per_row, 3];
+        // merge the tile axis into the batch axis, giving a combined SAM input of shape
+        // [tile_size, tile_size, 3, n_tiles_per_row * n_rows_batch] (tile fast, row slow)
+        inp_raw = ggml_reshape_4d(ctx0, inp_raw, img.nx() * img.nx(), n_tiles_per_row, 3, n_rows_batch);
+        inp_raw = ggml_cont(ctx0, ggml_permute(ctx0, inp_raw, 0, 2, 1, 3));
+        inp_raw = ggml_reshape_4d(ctx0, inp_raw, img.nx(), img.nx(), 3, n_tiles_per_row * n_rows_batch);
    }

    ggml_tensor * sam_out = build_sam(inp_raw);

    if (!is_overview) {
-        n_batch = n_tiles_per_row;
+        n_batch = n_tiles_per_row * n_rows_batch;
    }

    const int clip_n_patches = sam_out->ne[0] * sam_out->ne[1];
@@ -354,34 +359,36 @@ ggml_cgraph * clip_graph_deepseekocr::build() {
        const auto w     = h;
        const auto n_dim = cur->ne[0];

-        ggml_tensor * imgnl = ggml_repeat_4d(ctx0, model.image_newline, n_dim, 1, h, 1);
-        cur = ggml_reshape_3d(ctx0, cur, n_dim, w, h);
-        cur = ggml_reshape_2d(ctx0, ggml_concat(ctx0, cur, imgnl, 1), n_dim, (w + 1) * h);
-        cur = ggml_concat(ctx0, cur, model.view_seperator, 1);  // (n_dim, h*(w+1) + 1)
+        ggml_tensor * imgnl = ggml_repeat_4d(ctx0, model.image_newline, n_dim, 1, h, n_batch);
+        cur = ggml_reshape_4d(ctx0, cur, n_dim, w, h, n_batch);
+        cur = ggml_reshape_3d(ctx0, ggml_concat(ctx0, cur, imgnl, 1), n_dim, (w + 1) * h, n_batch);
+        ggml_tensor * vs = ggml_repeat_4d(ctx0, model.view_seperator, n_dim, 1, n_batch, 1);
+        cur = ggml_concat(ctx0, cur, vs, 1);  // (n_dim, h*(w+1) + 1, n_batch)
    } else {
        // tile row: interleave tiles within each row, add newline per row
-        const int grid_x      = static_cast<int>(std::sqrt(static_cast<float>(clip_n_patches)));
-        const int grid_y      = grid_x;
-        const auto n_dim      = cur->ne[0];
+        const int  grid_x = static_cast<int>(std::sqrt(static_cast<float>(clip_n_patches)));
+        const int  grid_y = grid_x;
+        const auto n_dim  = cur->ne[0];

-        // (n_dim, clip_n_patches, n_batch) -> (n_dim, grid_x, grid_y, n_batch)
-        cur = ggml_reshape_4d(ctx0, cur, n_dim, grid_x, grid_y, n_batch);
+        // merge n_dim into the grid_x axis, freeing the 4th axis for n_rows_batch
+        // (n_dim, clip_n_patches, n_tiles_per_row * n_rows_batch) -> (n_dim*grid_x, grid_y, n_tiles_per_row, n_rows_batch)
+        cur = ggml_reshape_4d(ctx0, cur, n_dim * grid_x, grid_y, n_tiles_per_row, n_rows_batch);

        // tiles: re-order from A.row0 A.row1 B.row0 B.row1 ...
        //        to A.row0 B.row0 A.row1 B.row1 ...
        //        then add nl: A.row0 B.row0 [nl] A.row1 B.row1 [nl] ...
-        // interleave tiles: (n_dim, grid_x, grid_y, n_batch) -> (n_dim, grid_x, n_batch, grid_y)
-        cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 0, 1, 3, 2));
+        // interleave tiles: -> (n_dim*grid_x, n_tiles_per_row, grid_y, n_rows_batch)
+        cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 0, 2, 1, 3));

-        // merge: (n_dim, grid_x, n_batch, grid_y) -> (n_dim, grid_x*n_batch, grid_y, 1)
-        cur = ggml_reshape_4d(ctx0, cur, n_dim, grid_x * n_batch, grid_y, 1);
+        // merge: -> (n_dim, grid_x*n_tiles_per_row, grid_y, n_rows_batch)
+        cur = ggml_reshape_4d(ctx0, cur, n_dim, grid_x * n_tiles_per_row, grid_y, n_rows_batch);

-        // append newline per row: (n_dim, grid_x*n_batch+1, grid_y, 1)
-        ggml_tensor * imgnl = ggml_repeat_4d(ctx0, model.image_newline, n_dim, 1, grid_y, 1);
+        // append newline per row: (n_dim, grid_x*n_tiles_per_row+1, grid_y, n_rows_batch)
+        ggml_tensor * imgnl = ggml_repeat_4d(ctx0, model.image_newline, n_dim, 1, grid_y, n_rows_batch);
        cur = ggml_concat(ctx0, cur, imgnl, 1);

-        // flatten: (n_dim, (grid_x*n_batch+1)*grid_y)
-        cur = ggml_reshape_2d(ctx0, cur, n_dim, (grid_x * n_batch + 1) * grid_y);
+        // flatten: (n_dim, (grid_x*n_tiles_per_row+1)*grid_y, n_rows_batch)
+        cur = ggml_reshape_3d(ctx0, cur, n_dim, (grid_x * n_tiles_per_row + 1) * grid_y, n_rows_batch);
    }

    cb(cur, "dsocr_output", -1);
@@ -14,8 +14,9 @@ ggml_cgraph * clip_graph_deepseekocr2::build() {
    {
        ggml_tensor * inp;

-        inp = ggml_reshape_2d(ctx0, sam_out, sam_out->ne[0] * sam_out->ne[1], sam_out->ne[2]); // H*W, C
-        inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
+        // H*W, C, B
+        inp = ggml_reshape_3d(ctx0, sam_out, sam_out->ne[0] * sam_out->ne[1], sam_out->ne[2], sam_out->ne[3]);
+        inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3)); // C, H*W, B

        auto num_image_tokens = inp->ne[1]; // H*W
        GGML_ASSERT(num_image_tokens == 144 || num_image_tokens == 256);
@@ -32,8 +33,10 @@ ggml_cgraph * clip_graph_deepseekocr2::build() {
            num_queries = 144;
        }

-        // (B, num_image_tokens + num_queries, C)
-        inp = ggml_concat(ctx0, inp, ggml_cast(ctx0, query_embed, inp->type), 1);
+        // repeat the query embedding per batch item, then append: (C, num_image_tokens + num_queries, B)
+        query_embed = ggml_cast(ctx0, query_embed, inp->type);
+        query_embed = ggml_repeat_4d(ctx0, query_embed, query_embed->ne[0], num_queries, inp->ne[2], 1);
+        inp = ggml_concat(ctx0, inp, query_embed, 1);

        auto seq_len = inp->ne[1];

@@ -57,7 +60,7 @@ ggml_cgraph * clip_graph_deepseekocr2::build() {
                                      /* learned_pos_embd */ nullptr, add_rope, vit_opts);

        cur = ggml_cont(ctx0,
-                        ggml_view_2d(ctx0, cur, cur->ne[0], num_queries, cur->nb[1],
+                        ggml_view_3d(ctx0, cur, cur->ne[0], num_queries, cur->ne[2], cur->nb[1], cur->nb[2],
                                     cur->nb[1] * (cur->ne[1] - num_queries))); // only take query tokens for output

        ggml_build_forward_expand(gf, cur);
@@ -71,7 +74,8 @@ ggml_cgraph * clip_graph_deepseekocr2::build() {

    // view_seperator only after the global view
    if (img.add_viewsep) {
-        cur = ggml_concat(ctx0, cur, model.view_seperator, 1); // (n_dim, 257)
+        ggml_tensor * vs = ggml_repeat_4d(ctx0, model.view_seperator, model.view_seperator->ne[0], 1, cur->ne[2], 1);
+        cur = ggml_concat(ctx0, cur, vs, 1); // (n_dim, 257, n_batch)
    }

    cb(cur, "dsocr2_output", -1);
@@ -2,6 +2,11 @@

 #include "../clip-graph.h"

+#include <map>
+#include <string>
+#include <utility>
+#include <vector>
+
 /*
 * IMPORTANT: The mtmd module does NOT accept pull requests that are fully or predominantly AI-generated.
 * We encourage human contributors to ensure the quality and reliability of the codebase.
@@ -133,12 +138,13 @@ struct clip_graph_deepseekocr : clip_graph {
    clip_graph_deepseekocr(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
    ggml_cgraph * build() override;
    ggml_tensor * build_sam(ggml_tensor * inp); // build the SAM model
-    // bool support_batch() const override { return true; } // TODO: support batch for DeepSeek-OCR v1
+    bool support_batch() const override { return true; }
 };

 struct clip_graph_deepseekocr2 : clip_graph_deepseekocr {
    clip_graph_deepseekocr2(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph_deepseekocr(ctx, img) {}
    ggml_cgraph * build() override; // reuses build_sam() from base
+    bool support_batch() const override { return true; }
 };

 struct clip_graph_conformer : clip_graph {
@@ -215,6 +221,111 @@ struct clip_graph_mimo_audio : clip_graph {
    ggml_cgraph * build() override;
 };

+struct clip_graph_qwen3tts_spkenc : clip_graph {
+    clip_graph_qwen3tts_spkenc(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+
+    ggml_tensor * conv1d_same(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int dilation) const;
+    ggml_tensor * res2net(ggml_tensor * x, const clip_layer & layer, int dilation, int scale) const;
+    ggml_tensor * se_block(ggml_tensor * x, const clip_layer & layer) const;
+    ggml_tensor * se_res2net_block(ggml_tensor * x, const clip_layer & layer, int dilation, int scale) const;
+    ggml_tensor * attentive_stats_pool(ggml_tensor * x) const;
+};
+
+struct clip_graph_qwen3tts_gen : clip_graph {
+    clip_graph_qwen3tts_gen(clip_ctx * ctx, const clip_image_f32 & img, clip_gen_process_type gen_process, int top_k, float top_p)
+        : clip_graph(ctx, img), gen_process(gen_process), top_k(top_k), top_p(top_p) {}
+    ggml_cgraph * build() override;
+
+    // which sub-graph build() constructs, fixed at graph-build time
+    clip_gen_process_type gen_process;
+
+    // sampling params, fixed at graph-build time (GEN_CODE only)
+    int   top_k;
+    float top_p;
+
+    //
+    // code_gen: backbone hidden state + sampled code0 -> 16 RVQ codes
+    // MTP-style code predictor, one token per codebook
+    //
+    struct code_gen : clip_graph {
+        code_gen(const clip_graph & parent, int top_k, float top_p)
+            : clip_graph(parent), top_k(top_k), top_p(top_p) {}
+        ggml_cgraph * build() override { GGML_ABORT("call prefill()/step() instead"); }
+
+        int   top_k;
+        float top_p;
+
+        ggml_tensor * cache_set(ggml_tensor * cache, int row_idx, ggml_tensor * value) const;
+        ggml_tensor * do_sampling(ggml_tensor * logits, ggml_tensor * inp_rand) const;
+
+        ggml_tensor * const_i32(ggml_tensor * anchor, float value) const;
+        ggml_tensor * causal_mask_row(int64_t n_kv_pad, int pos) const;
+        ggml_tensor * project_in(ggml_tensor * cur) const;
+
+        ggml_tensor * layer_forward(
+                ggml_tensor * cur,
+                const clip_layer & layer,
+                ggml_tensor * inp_pos,
+                ggml_tensor * kq_mask,
+                ggml_tensor *& k_cache_layer,
+                ggml_tensor *& v_cache_layer,
+                int64_t n_kv_pad,
+                int pos,
+                int il) const;
+
+        void prefill(
+                std::vector<ggml_tensor *> & k_cache,
+                std::vector<ggml_tensor *> & v_cache,
+                ggml_tensor *& out_code_cache,
+                ggml_tensor * h_state,
+                ggml_tensor * code0_embd,
+                ggml_tensor * inp_rand) const;
+
+        ggml_tensor * step(
+                std::vector<ggml_tensor *> & k_cache,
+                std::vector<ggml_tensor *> & v_cache,
+                ggml_tensor * out_code_cache,
+                ggml_tensor * inp_rand,
+                int step_idx) const;
+    };
+
+    //
+    // code2wav: RVQ codes -> raw PCM (quantizer + pre_conv + pre_transformer + upsample + DAC).
+    //
+    struct code2wav : clip_graph {
+        code2wav(const clip_graph & parent) : clip_graph(parent) {}
+        ggml_cgraph * build() override { GGML_ABORT("call decode() instead"); }
+
+        // state_in: previous call's persisted state, by slot name (see list_c2w_state_slots())
+        std::map<std::string, ggml_tensor *> state_in;
+        // state_out: this call's state to persist, added to the graph outputs by build()
+        mutable std::vector<std::pair<std::string, ggml_tensor *>> state_out;
+
+        // stateful conv ops: read/update their state via state_in/state_out[state_name]
+        ggml_tensor * causal_conv1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int dilation, const std::string & state_name) const;
+        ggml_tensor * causal_conv1d_dw(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, const std::string & state_name) const;
+        ggml_tensor * causal_conv_transpose1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride, const std::string & state_name) const;
+        ggml_tensor * snake(ggml_tensor * x, ggml_tensor * alpha, ggml_tensor * beta) const;
+
+        ggml_tensor * quant_decode(ggml_tensor * inp_codes) const;
+        ggml_tensor * tfm_layer_forward(ggml_tensor * cur, const clip_layer & layer, int il) const;
+        ggml_tensor * convnext_block(ggml_tensor * x, const clip_code2wav::upsample_block & blk, const std::string & state_prefix) const;
+        ggml_tensor * dac_res_unit(ggml_tensor * x, const clip_code2wav::dac_res & res, int dilation, const std::string & state_name) const;
+
+        // inp_codes [1, n_codes] I32 -> this frame's audio samples [n_samples] F32, clamped to [-1, 1]
+        ggml_tensor * decode(ggml_tensor * inp_codes) const;
+    };
+};
+
+// one persisted state buffer used by code2wav, see qwen3tts-gen.cpp
+struct c2w_state_slot {
+    std::string name;
+    int64_t     ne0;
+    int64_t     ne1;
+};
+std::vector<c2w_state_slot> list_c2w_state_slots(const clip_hparams & hparams, const clip_model & model);
+
 struct clip_graph_kimik25 : clip_graph {
    clip_graph_kimik25(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
    ggml_cgraph * build() override;
@@ -0,0 +1,766 @@
+#include "models.h"
+
+#include <string>
+
+// on-device sampling: top-k, top-p, then a random draw
+ggml_tensor * clip_graph_qwen3tts_gen::code_gen::do_sampling(ggml_tensor * logits, ggml_tensor * inp_rand) const {
+    logits = ggml_reshape_1d(ctx0, logits, ggml_nelements(logits));
+    const int64_t n_vocab = logits->ne[0];
+
+    // sort a's rows by idx
+    auto sort_by = [this](ggml_tensor * a, ggml_tensor * idx) {
+        ggml_tensor * a2d = ggml_reshape_2d(ctx0, a, 1, a->ne[0]);
+        return ggml_reshape_1d(ctx0, ggml_get_rows(ctx0, a2d, idx), idx->ne[0]);
+    };
+
+    ggml_tensor * cur        = logits;
+    ggml_tensor * candidates = nullptr; // maps row index back to vocab id
+
+    if (top_k > 0 && top_k < n_vocab) {
+        ggml_tensor * idx = ggml_top_k(ctx0, cur, top_k);
+        candidates = idx;
+        cur        = sort_by(cur, idx);
+        cb(cur, "sample_top_k_logits", -1);
+    }
+
+    if (top_p < 1.0f) {
+        ggml_tensor * sorted_idx    = ggml_argsort(ctx0, cur, GGML_SORT_ORDER_DESC);
+        ggml_tensor * sorted_logits = sort_by(cur, sorted_idx);
+        candidates = candidates ? sort_by(candidates, sorted_idx) : sorted_idx;
+
+        ggml_tensor * probs = ggml_soft_max(ctx0, sorted_logits);
+        ggml_tensor * cdf   = ggml_cumsum(ctx0, probs);
+
+        // keep_mask[i] = 1 once cdf[i] crosses top_p
+        ggml_tensor * cdf_scaled = ggml_scale_bias(ctx0, cdf, -1.0f, top_p);
+        ggml_tensor * keep_mask  = ggml_step(ctx0, cdf_scaled);
+        ggml_tensor * idxf       = ggml_sum(ctx0, keep_mask);
+        idxf = ggml_clamp(ctx0, idxf, 0.0f, (float) keep_mask->ne[0] - 1);
+        ggml_tensor * ones = ggml_scale_bias(ctx0, idxf, 0.0f, 1.0f);
+
+        // top-p must include the crossing element, so force it to 1
+        ggml_tensor * keep_mask_2d = ggml_reshape_2d(ctx0, keep_mask, 1, keep_mask->ne[0]);
+        keep_mask_2d = ggml_set_rows(ctx0, keep_mask_2d, ones, ggml_cast(ctx0, idxf, GGML_TYPE_I32));
+        keep_mask    = ggml_reshape_1d(ctx0, keep_mask_2d, keep_mask->ne[0]);
+
+        // log(1) = 0 (keep), log(0) = -inf (drop)
+        ggml_tensor * bias = ggml_log(ctx0, keep_mask);
+        cur = ggml_add(ctx0, sorted_logits, bias);
+        cb(cur, "sample_top_p_logits", -1);
+    }
+
+    // draw one token: find where the cdf crosses inp_rand
+    ggml_tensor * probs  = ggml_soft_max(ctx0, cur);
+    ggml_tensor * cumsum = ggml_cumsum(ctx0, probs);
+
+    ggml_tensor * diff       = ggml_sub(ctx0, cumsum, inp_rand);
+    ggml_tensor * cross_mask = ggml_step(ctx0, diff);
+    ggml_tensor * idxf       = ggml_sum(ctx0, cross_mask);
+    ggml_tensor * idx        = ggml_cast(ctx0, ggml_scale_bias(ctx0, idxf, -1.0f, (float) cross_mask->ne[0]), GGML_TYPE_I32);
+
+    if (candidates) {
+        ggml_tensor * cand_2d = ggml_reshape_2d(ctx0, candidates, 1, candidates->ne[0]);
+        idx = ggml_get_rows(ctx0, cand_2d, idx);
+    }
+    cb(idx, "sample_token_id", -1);
+
+    return idx;
+}
+
+// returns a new cache with row row_idx set to value
+ggml_tensor * clip_graph_qwen3tts_gen::code_gen::cache_set(ggml_tensor * cache, int row_idx, ggml_tensor * value) const {
+    const int64_t n_embd  = cache->ne[0];
+    const int64_t n_cache = cache->ne[1];
+    GGML_ASSERT(row_idx >= 0 && row_idx < n_cache);
+
+    // append value as the last row, then gather it back into place
+    ggml_tensor * value_2d  = ggml_reshape_2d(ctx0, value, n_embd, 1);
+    ggml_tensor * cache_ext = ggml_concat(ctx0, cache, value_2d, 1); // [n_embd, n_cache + 1]
+
+    // gather indices [0..row_idx-1, n_cache, row_idx+1..n_cache-1]
+    // built via concat, since ggml_set_rows needs F32/F16 values, not an I32 index array
+    ggml_tensor * idx = const_i32(cache, (float) n_cache);
+    if (row_idx > 0) {
+        ggml_tensor * prefix = ggml_cast(ctx0, ggml_arange(ctx0, 0.0f, (float) row_idx, 1.0f), GGML_TYPE_I32);
+        idx = ggml_concat(ctx0, prefix, idx, 0);
+    }
+    if (row_idx < n_cache - 1) {
+        ggml_tensor * suffix = ggml_cast(ctx0, ggml_arange(ctx0, (float) (row_idx + 1), (float) n_cache, 1.0f), GGML_TYPE_I32);
+        idx = ggml_concat(ctx0, idx, suffix, 0);
+    }
+
+    ggml_tensor * result = ggml_get_rows(ctx0, cache_ext, idx);
+    cb(result, "cache_set_out", -1);
+    return result;
+}
+
+// builds a const i32 with no host upload: view a tensor, zero it via scale, add value, cast to i32
+ggml_tensor * clip_graph_qwen3tts_gen::code_gen::const_i32(ggml_tensor * anchor, float value) const {
+    ggml_tensor * v = ggml_view_1d(ctx0, anchor, 1, 0);
+    if (v->type != GGML_TYPE_F32) {
+        v = ggml_cast(ctx0, v, GGML_TYPE_F32);
+    }
+    return ggml_cast(ctx0, ggml_scale_bias(ctx0, v, 0.0f, value), GGML_TYPE_I32);
+}
+
+// causal keep-mask row for a query at position pos, window size n_kv_pad
+ggml_tensor * clip_graph_qwen3tts_gen::code_gen::causal_mask_row(int64_t n_kv_pad, int pos) const {
+    ggml_tensor * ones = ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv_pad, n_kv_pad), 1.0f);
+    ggml_tensor * keep = ggml_tri(ctx0, ones, GGML_TRI_TYPE_LOWER_DIAG);
+    ggml_tensor * row  = ggml_view_1d(ctx0, keep, n_kv_pad, (size_t) pos * keep->nb[1]);
+    ggml_tensor * mask = ggml_log(ctx0, row); // 0 = keep, -inf = masked
+    return ggml_reshape_4d(ctx0, mask, n_kv_pad, 1, 1, 1);
+}
+
+// talker hidden size -> predictor hidden size (small_to_mtp_projection)
+ggml_tensor * clip_graph_qwen3tts_gen::code_gen::project_in(ggml_tensor * cur) const {
+    if (!model.gen_code_proj_in_w) {
+        return cur;
+    }
+    cur = ggml_mul_mat(ctx0, model.gen_code_proj_in_w, cur);
+    if (model.gen_code_proj_in_b) {
+        cur = ggml_add(ctx0, cur, model.gen_code_proj_in_b);
+    }
+    return cur;
+}
+
+// one transformer layer at position pos; writes k/v into k_cache_layer/v_cache_layer at row pos
+ggml_tensor * clip_graph_qwen3tts_gen::code_gen::layer_forward(
+        ggml_tensor * cur,
+        const clip_layer & layer,
+        ggml_tensor * inp_pos,
+        ggml_tensor * kq_mask,
+        ggml_tensor *& k_cache_layer,
+        ggml_tensor *& v_cache_layer,
+        int64_t n_kv_pad,
+        int pos,
+        int il) const {
+    const int     n_head    = hparams.n_head;
+    const int     n_head_kv = hparams.n_head_kv;
+    const int64_t d_head    = layer.q_w->ne[1] / n_head; // real head_dim, not n_embd / n_head
+    const float   kq_scale  = 1.0f / sqrtf((float) d_head);
+
+    ggml_tensor * residual = cur;
+
+    ggml_tensor * h = ggml_rms_norm(ctx0, cur, hparams.eps);
+    h = ggml_mul(ctx0, h, layer.ln_1_w);
+
+    ggml_tensor * q = ggml_mul_mat(ctx0, layer.q_w, h);
+    ggml_tensor * k = ggml_mul_mat(ctx0, layer.k_w, h);
+    ggml_tensor * v = ggml_mul_mat(ctx0, layer.v_w, h);
+
+    q = ggml_reshape_3d(ctx0, q, d_head, n_head, 1);
+    k = ggml_reshape_3d(ctx0, k, d_head, n_head_kv, 1);
+
+    q = ggml_rms_norm(ctx0, q, hparams.eps);
+    q = ggml_mul(ctx0, q, layer.q_norm);
+    k = ggml_rms_norm(ctx0, k, hparams.eps);
+    k = ggml_mul(ctx0, k, layer.k_norm);
+
+    q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, (int) d_head, GGML_ROPE_TYPE_NEOX, 0,
+                      hparams.rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+    k = ggml_rope_ext(ctx0, k, inp_pos, nullptr, (int) d_head, GGML_ROPE_TYPE_NEOX, 0,
+                      hparams.rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+
+    // write k/v into the cache at row pos, flat layout
+    ggml_tensor * k_flat = ggml_reshape_1d(ctx0, k, d_head * n_head_kv);
+    k_cache_layer = cache_set(k_cache_layer, pos, k_flat);
+    v_cache_layer = cache_set(v_cache_layer, pos, v);
+
+    ggml_tensor * q_cur = ggml_reshape_4d(ctx0, q, d_head, n_head, 1, 1);
+    ggml_tensor * k_cur = ggml_reshape_4d(ctx0, k_cache_layer, d_head, n_head_kv, n_kv_pad, 1);
+    ggml_tensor * v_cur = ggml_reshape_4d(ctx0, v_cache_layer, d_head, n_head_kv, n_kv_pad, 1);
+
+    ggml_tensor * attn_out = build_attn(layer.o_w, layer.o_b, q_cur, k_cur, v_cur, kq_mask, kq_scale, il);
+
+    cur = ggml_add(ctx0, residual, attn_out);
+
+    ggml_tensor * h2 = ggml_rms_norm(ctx0, cur, hparams.eps);
+    h2 = ggml_mul(ctx0, h2, layer.ln_2_w);
+
+    ggml_tensor * gate = ggml_mul_mat(ctx0, layer.ff_gate_w, h2);
+    ggml_tensor * up   = ggml_mul_mat(ctx0, layer.ff_up_w, h2);
+    ggml_tensor * gu   = ggml_swiglu_split(ctx0, gate, up);
+    ggml_tensor * down = ggml_mul_mat(ctx0, layer.ff_down_w, gu);
+
+    return ggml_add(ctx0, cur, down);
+}
+
+// position 0: hidden bridge, seeds the k/v cache, no sampling
+// position 1: embed(code0), sample with lm_head[0], write out_code_cache[1]
+void clip_graph_qwen3tts_gen::code_gen::prefill(
+        std::vector<ggml_tensor *> & k_cache,
+        std::vector<ggml_tensor *> & v_cache,
+        ggml_tensor *& out_code_cache,
+        ggml_tensor * h_state,
+        ggml_tensor * code0_embd,
+        ggml_tensor * inp_rand) const {
+    const int64_t n_kv_pad = k_cache[0]->ne[1];
+
+    {
+        ggml_tensor * cur     = project_in(h_state);
+        ggml_tensor * kq_mask = causal_mask_row(n_kv_pad, 0);
+        ggml_tensor * inp_pos = const_i32(k_cache[0], 0.0f);
+        for (size_t il = 0; il < model.layers.size(); il++) {
+            cur = layer_forward(cur, model.layers[il], inp_pos, kq_mask, k_cache[il], v_cache[il], n_kv_pad, 0, (int) il);
+        }
+        // position 0's output is unused, it only seeded the cache
+    }
+
+    {
+        ggml_tensor * cur     = project_in(code0_embd);
+        ggml_tensor * kq_mask = causal_mask_row(n_kv_pad, 1);
+        ggml_tensor * inp_pos = const_i32(k_cache[0], 1.0f);
+        for (size_t il = 0; il < model.layers.size(); il++) {
+            cur = layer_forward(cur, model.layers[il], inp_pos, kq_mask, k_cache[il], v_cache[il], n_kv_pad, 1, (int) il);
+        }
+
+        cur = ggml_rms_norm(ctx0, cur, hparams.eps);
+        cur = ggml_mul(ctx0, cur, model.gen_code_norm_w);
+
+        ggml_tensor * head_w = model.gen_code_head_w;
+        ggml_tensor * head_g = ggml_view_2d(ctx0, head_w, head_w->ne[0], head_w->ne[1], head_w->nb[1], 0); // lm_head[0]
+        ggml_tensor * logits = ggml_mul_mat(ctx0, head_g, cur);
+
+        ggml_tensor * sampled = do_sampling(logits, inp_rand);
+        out_code_cache = cache_set(out_code_cache, 1, sampled);
+    }
+}
+
+// one decode step of code_predictor
+// at step_idx g:
+// - read code from out_code_cache[g], then embed it with codebook table g-1
+// - write new kv at cache row g+1, sample with lm_head[g]
+// - write result to out_code_cache[g+1]
+// step_idx must be in [1, n_acoustic - 1]
+ggml_tensor * clip_graph_qwen3tts_gen::code_gen::step(
+        std::vector<ggml_tensor *> & k_cache,
+        std::vector<ggml_tensor *> & v_cache,
+        ggml_tensor * out_code_cache,
+        ggml_tensor * inp_rand,
+        int step_idx) const {
+    const int64_t n_acoustic = model.gen_code_head_w->ne[2];
+    GGML_ASSERT(step_idx >= 1 && step_idx < n_acoustic);
+    GGML_ASSERT(k_cache.size() == model.layers.size());
+    GGML_ASSERT(v_cache.size() == model.layers.size());
+
+    const int64_t n_kv_pad = k_cache[0]->ne[1];
+    const int     pos      = step_idx + 1; // new cache row and RoPE position
+
+    // embed the previous code via this step's codebook table (rows are already scalars)
+    ggml_tensor * code_in = ggml_view_1d(ctx0, out_code_cache, 1, (size_t) step_idx * out_code_cache->nb[1]);
+
+    ggml_tensor * embd_w = model.gen_code_embd_w; // [n_embd_talker, vocab, n_acoustic]
+    ggml_tensor * embd_g = ggml_view_2d(ctx0, embd_w, embd_w->ne[0], embd_w->ne[1], embd_w->nb[1],
+                                        (size_t) (step_idx - 1) * embd_w->nb[2]);
+    ggml_tensor * cur = ggml_get_rows(ctx0, embd_g, code_in);
+    cur = ggml_reshape_1d(ctx0, cur, cur->ne[0]);
+    cb(cur, "step_embd_in", step_idx);
+
+    cur = project_in(cur);
+    cb(cur, "step_proj_in", step_idx);
+
+    ggml_tensor * kq_mask = causal_mask_row(n_kv_pad, pos);
+    ggml_tensor * inp_pos = const_i32(k_cache[0], (float) pos);
+
+    for (size_t il = 0; il < model.layers.size(); il++) {
+        cur = layer_forward(cur, model.layers[il], inp_pos, kq_mask, k_cache[il], v_cache[il], n_kv_pad, pos, (int) il);
+        cb(cur, "step_layer_out", (int) il);
+    }
+
+    // final norm, this step's lm_head, sample, write the result
+    cur = ggml_rms_norm(ctx0, cur, hparams.eps);
+    cur = ggml_mul(ctx0, cur, model.gen_code_norm_w);
+
+    ggml_tensor * head_w = model.gen_code_head_w; // [n_embd_pred, vocab, n_acoustic]
+    ggml_tensor * head_g = ggml_view_2d(ctx0, head_w, head_w->ne[0], head_w->ne[1], head_w->nb[1],
+                                        (size_t) step_idx * head_w->nb[2]);
+    ggml_tensor * logits = ggml_mul_mat(ctx0, head_g, cur);
+    cb(logits, "step_logits", step_idx);
+
+    ggml_tensor * sampled = do_sampling(logits, inp_rand);
+    cb(sampled, "step_sampled", step_idx);
+
+    return cache_set(out_code_cache, pos, sampled);
+}
+
+// causal conv1d, stride 1: prepend persisted left-context instead of zero-padding, then a plain conv
+// x: [T, IC] (T-first). w: [K, IC, OC]. state_name empty means K == 1 (no left-context). returns [T, OC]
+ggml_tensor * clip_graph_qwen3tts_gen::code2wav::causal_conv1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int dilation, const std::string & state_name) const {
+    const int K   = (int) w->ne[0];
+    const int pad = (K - 1) * dilation;
+
+    ggml_tensor * x_full = x;
+    if (pad > 0) {
+        ggml_tensor * left = state_in.at(state_name); // [pad, IC]
+        x_full = ggml_concat(ctx0, left, x, 0);
+    }
+    ggml_tensor * y = ggml_conv_1d(ctx0, w, x_full, 1, 0, dilation); // [T, OC, 1]
+    y = ggml_reshape_2d(ctx0, y, y->ne[0], y->ne[1]);
+    if (b) {
+        y = ggml_add(ctx0, y, ggml_reshape_2d(ctx0, b, 1, b->ne[0]));
+    }
+    if (pad > 0) {
+        ggml_tensor * new_left = ggml_cont(ctx0, ggml_view_2d(ctx0, x_full, pad, x_full->ne[1], x_full->nb[1],
+                                                              (size_t) (x_full->ne[0] - pad) * x_full->nb[0]));
+        state_out.push_back({state_name, new_left});
+    }
+    return y;
+}
+
+// causal depthwise conv1d, stride 1, dilation 1, kernel from w's shape.
+// x: [T, C]. w: [K, 1, C]. returns [T, C]. see causal_conv1d for the state contract.
+ggml_tensor * clip_graph_qwen3tts_gen::code2wav::causal_conv1d_dw(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, const std::string & state_name) const {
+    const int K   = (int) w->ne[0];
+    const int pad = K - 1;
+
+    ggml_tensor * x_full = x;
+    if (pad > 0) {
+        ggml_tensor * left = state_in.at(state_name); // [pad, C]
+        x_full = ggml_concat(ctx0, left, x, 0);
+    }
+    ggml_tensor * y = ggml_conv_1d_dw(ctx0, w, x_full, 1, 0, 1); // [T, C, 1]
+    y = ggml_reshape_2d(ctx0, y, y->ne[0], y->ne[1]);
+    if (b) {
+        y = ggml_add(ctx0, y, ggml_reshape_2d(ctx0, b, 1, b->ne[0]));
+    }
+    if (pad > 0) {
+        ggml_tensor * new_left = ggml_cont(ctx0, ggml_view_2d(ctx0, x_full, pad, x_full->ne[1], x_full->nb[1],
+                                                              (size_t) (x_full->ne[0] - pad) * x_full->nb[0]));
+        state_out.push_back({state_name, new_left});
+    }
+    return y;
+}
+
+// causal ConvTranspose1d, the (kernel - stride) overlap tail is kept as state for the next call
+// x: [T, IC], w: [K, OC, IC]. state_name empty means K == stride (no overlap). returns [T * stride, OC]
+ggml_tensor * clip_graph_qwen3tts_gen::code2wav::causal_conv_transpose1d(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int stride, const std::string & state_name) const {
+    const int     K        = (int) w->ne[0];
+    const int     OC       = (int) w->ne[1];
+    const int     trim     = K - stride;
+    const int64_t emit_len = x->ne[0] * stride;
+
+    // transposed conv as GEMM + col2im scatter-add, y: [emit_len + trim, OC]
+    ggml_tensor * w2  = ggml_reshape_2d(ctx0, w, (int64_t) K * OC, w->ne[2]);
+    w2                = ggml_cont(ctx0, ggml_transpose(ctx0, w2));
+    ggml_tensor * xt  = ggml_cont(ctx0, ggml_transpose(ctx0, x));
+    ggml_tensor * col = ggml_mul_mat(ctx0, w2, xt);
+    ggml_tensor * y   = ggml_col2im_1d(ctx0, col, stride, OC, 0);
+
+    ggml_tensor * out = y;
+    if (trim > 0) {
+        ggml_tensor * tail = state_in.at(state_name); // [trim, OC]
+        ggml_tensor * head = ggml_add(ctx0, ggml_view_2d(ctx0, y, trim, y->ne[1], y->nb[1], 0), tail);
+        if (emit_len > trim) {
+            ggml_tensor * middle = ggml_view_2d(ctx0, y, emit_len - trim, y->ne[1], y->nb[1], (size_t) trim * y->nb[0]);
+            out = ggml_concat(ctx0, head, middle, 0);
+        } else {
+            out = head;
+        }
+        ggml_tensor * new_tail = ggml_cont(ctx0, ggml_view_2d(ctx0, y, trim, y->ne[1], y->nb[1], (size_t) emit_len * y->nb[0]));
+        state_out.push_back({state_name, new_tail});
+    }
+    if (b) {
+        out = ggml_add(ctx0, out, ggml_reshape_2d(ctx0, b, 1, b->ne[0]));
+    }
+    return out;
+}
+
+// SnakeBeta activation: y = x + sin(alpha*x)^2 * inv_beta (alpha/inv_beta folded via exp/reciprocal at conversion time)
+// x: [T, C]. alpha/beta: [C], broadcasts over T
+ggml_tensor * clip_graph_qwen3tts_gen::code2wav::snake(ggml_tensor * x, ggml_tensor * alpha, ggml_tensor * beta) const {
+    ggml_tensor * a = ggml_reshape_2d(ctx0, alpha, 1, alpha->ne[0]);
+    ggml_tensor * b = ggml_reshape_2d(ctx0, beta,  1, beta->ne[0]);
+
+    // expand reshapes first so mul/sin/sqr/mul/add lands as consecutive nodes, letting backends fuse them
+    ggml_build_forward_expand(gf, a);
+    ggml_build_forward_expand(gf, b);
+
+    ggml_tensor * s = ggml_sin(ctx0, ggml_mul(ctx0, x, a));
+    s = ggml_sqr(ctx0, s);
+    s = ggml_mul(ctx0, s, b);
+    return ggml_add(ctx0, x, s);
+}
+
+// RVQ codebook decode: T frames of 16 codes -> 512-dim hidden (C-first, [512, T])
+// codebook 0 (semantic) and 1..15 (acoustic) sum within their group, project separately, then add
+ggml_tensor * clip_graph_qwen3tts_gen::code2wav::quant_decode(ggml_tensor * inp_codes) const {
+    const auto & c2w = model.c2w;
+    const int64_t T = inp_codes->ne[0];
+
+    // ids for codebook group g over all T frames, [T] I32
+    auto group_ids = [&](int g) {
+        return ggml_view_1d(ctx0, inp_codes, T, (size_t) g * inp_codes->nb[1]);
+    };
+
+    ggml_tensor * sem     = ggml_get_rows(ctx0, c2w.quant_first_cb_w, group_ids(0)); // [256, T]
+    ggml_tensor * sem_out = ggml_mul_mat(ctx0, c2w.quant_first_out_w, sem);          // [512, T]
+
+    ggml_tensor * acc = nullptr;
+    const int64_t n_acoustic = c2w.quant_rest_cb_w->ne[2];
+    for (int g = 1; g <= n_acoustic; g++) {
+        ggml_tensor * cb_g  = ggml_view_2d(ctx0, c2w.quant_rest_cb_w, c2w.quant_rest_cb_w->ne[0], c2w.quant_rest_cb_w->ne[1],
+                                           c2w.quant_rest_cb_w->nb[1], (size_t) (g - 1) * c2w.quant_rest_cb_w->nb[2]);
+        ggml_tensor * embd = ggml_get_rows(ctx0, cb_g, group_ids(g)); // [256, T]
+        acc = acc ? ggml_add(ctx0, acc, embd) : embd;
+    }
+    ggml_tensor * ac_out = ggml_mul_mat(ctx0, c2w.quant_rest_out_w, acc); // [512, T]
+
+    ggml_tensor * hidden = ggml_add(ctx0, sem_out, ac_out);
+    cb(hidden, "wav_quant_hidden", -1);
+    return hidden;
+}
+
+// one pre_transformer layer over a batch of N = sliding_window new frames
+// attention runs over [(W-1)-frame prefix from the last batch] + [N new frames]
+// RoPE positions come from a persisted counter, so phases line up across batches
+ggml_tensor * clip_graph_qwen3tts_gen::code2wav::tfm_layer_forward(ggml_tensor * cur, const clip_layer & layer, int il) const {
+    const int     n_head    = hparams.wav_tfm_n_head;
+    const int     n_head_kv = hparams.wav_tfm_n_head_kv;
+    const int64_t d_head    = layer.q_w->ne[1] / n_head;
+    const float   kq_scale  = 1.0f / sqrtf((float) d_head);
+    const int64_t W         = hparams.wav_tfm_swa; // == N, frames per batch
+    const int64_t N         = cur->ne[1];
+    const int64_t prefix    = W - 1;
+    const int64_t total_kv  = prefix + N;
+
+    ggml_tensor * residual = cur;
+    ggml_tensor * h = ggml_rms_norm(ctx0, cur, hparams.wav_tfm_eps);
+    h = ggml_mul(ctx0, h, layer.ln_1_w);
+
+    ggml_tensor * q = ggml_mul_mat(ctx0, layer.q_w, h); // [n_head*d_head, N]
+    ggml_tensor * k = ggml_mul_mat(ctx0, layer.k_w, h); // [n_head_kv*d_head, N]
+    ggml_tensor * v = ggml_mul_mat(ctx0, layer.v_w, h); // [n_head_kv*d_head, N]
+
+    q = ggml_reshape_3d(ctx0, q, d_head, n_head, N);
+    k = ggml_reshape_3d(ctx0, k, d_head, n_head_kv, N);
+
+    // real, ever-increasing positions: base (persisted) .. base+N-1
+    ggml_tensor * base   = ggml_reshape_1d(ctx0, state_in.at("tfm_pos"), 1);
+    ggml_tensor * offset = ggml_arange(ctx0, 0.0f, (float) N, 1.0f);
+    ggml_tensor * pos    = ggml_cast(ctx0, ggml_add(ctx0, offset, base), GGML_TYPE_I32);
+
+    q = ggml_rope_ext(ctx0, q, pos, nullptr, (int) d_head, GGML_ROPE_TYPE_NEOX, 0,
+                      hparams.wav_tfm_rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+    k = ggml_rope_ext(ctx0, k, pos, nullptr, (int) d_head, GGML_ROPE_TYPE_NEOX, 0,
+                      hparams.wav_tfm_rope_theta, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
+
+    // the position counter is the same for all layers, push it once from layer 0
+    if (il == 0) {
+        state_out.push_back({"tfm_pos", ggml_scale_bias(ctx0, state_in.at("tfm_pos"), 1.0f, (float) N)});
+    }
+
+    ggml_tensor * k_new = ggml_reshape_2d(ctx0, k, d_head * n_head_kv, N);
+    ggml_tensor * v_new = ggml_reshape_2d(ctx0, v, d_head * n_head_kv, N);
+
+    ggml_tensor * old_k = state_in.at("tfm_k_" + std::to_string(il)); // [d_head*n_head_kv, W-1]
+    ggml_tensor * old_v = state_in.at("tfm_v_" + std::to_string(il));
+
+    ggml_tensor * k_full = ggml_concat(ctx0, old_k, k_new, 1); // [.., prefix+N]
+    ggml_tensor * v_full = ggml_concat(ctx0, old_v, v_new, 1);
+
+    // next batch's prefix: the last (W-1) frames of this batch
+    state_out.push_back({"tfm_k_" + std::to_string(il),
+        ggml_cont(ctx0, ggml_view_2d(ctx0, k_full, k_full->ne[0], prefix, k_full->nb[1], (size_t) N * k_full->nb[1]))});
+    state_out.push_back({"tfm_v_" + std::to_string(il),
+        ggml_cont(ctx0, ggml_view_2d(ctx0, v_full, v_full->ne[0], prefix, v_full->nb[1], (size_t) N * v_full->nb[1]))});
+
+    // banded causal mask: key j is visible to query i iff 0 <= (prefix+i) - j < W
+    ggml_tensor * pos_k = ggml_reshape_2d(ctx0, ggml_arange(ctx0, 0.0f, (float) total_kv, 1.0f), total_kv, 1);
+    ggml_tensor * pos_q = ggml_reshape_2d(ctx0, ggml_arange(ctx0, (float) prefix, (float) (prefix + N), 1.0f), 1, N);
+    ggml_tensor * pos_q_grid = ggml_repeat_4d(ctx0, pos_q, total_kv, N, 1, 1);
+    ggml_tensor * diff = ggml_sub(ctx0, pos_q_grid, pos_k); // [total_kv, N]
+
+    ggml_tensor * causal_keep = ggml_step(ctx0, ggml_scale_bias(ctx0, diff, 1.0f, 0.5f));            // diff >= 0
+    ggml_tensor * in_window   = ggml_step(ctx0, ggml_scale_bias(ctx0, diff, -1.0f, (float) W - 0.5f)); // diff < W
+    ggml_tensor * keep = ggml_mul(ctx0, causal_keep, in_window);
+
+    // on a cold start, key j is real state only when j >= prefix - tfm_pos, mask out the rest
+    ggml_tensor * warm = ggml_step(ctx0, ggml_scale_bias(ctx0, ggml_add(ctx0, pos_k, base),
+                                                         1.0f, 0.5f - (float) prefix)); // j + pos > prefix - 0.5
+    keep = ggml_mul(ctx0, keep, warm);
+
+    ggml_tensor * mask = ggml_reshape_4d(ctx0, ggml_log(ctx0, keep), total_kv, N, 1, 1); // 0 = keep, -inf = masked
+
+    ggml_tensor * q_cur = ggml_reshape_4d(ctx0, q, d_head, n_head, N, 1);
+    ggml_tensor * k_cur = ggml_reshape_4d(ctx0, k_full, d_head, n_head_kv, total_kv, 1);
+    ggml_tensor * v_cur = ggml_reshape_4d(ctx0, v_full, d_head, n_head_kv, total_kv, 1);
+
+    ggml_tensor * attn_out = build_attn(layer.o_w, layer.o_b, q_cur, k_cur, v_cur, mask, kq_scale, il);
+    if (layer.ls_1_w) {
+        attn_out = ggml_mul(ctx0, attn_out, layer.ls_1_w);
+    }
+    cur = ggml_add(ctx0, residual, attn_out);
+
+    ggml_tensor * residual2 = cur;
+    ggml_tensor * h2 = ggml_rms_norm(ctx0, cur, hparams.wav_tfm_eps);
+    h2 = ggml_mul(ctx0, h2, layer.ln_2_w);
+
+    ggml_tensor * gate = ggml_mul_mat(ctx0, layer.ff_gate_w, h2);
+    ggml_tensor * up   = ggml_mul_mat(ctx0, layer.ff_up_w, h2);
+    ggml_tensor * gu   = ggml_swiglu_split(ctx0, gate, up);
+    ggml_tensor * down = ggml_mul_mat(ctx0, layer.ff_down_w, gu);
+    if (layer.ls_2_w) {
+        down = ggml_mul(ctx0, down, layer.ls_2_w);
+    }
+    return ggml_add(ctx0, residual2, down);
+}
+
+// dwconv -> LayerNorm -> pwconv1 -> GELU -> pwconv2 -> layer scale -> residual
+// x: [T, C] T-first; LayerNorm/pwconv need C on ne0, so this transposes in and back out
+ggml_tensor * clip_graph_qwen3tts_gen::code2wav::convnext_block(ggml_tensor * x, const clip_code2wav::upsample_block & blk, const std::string & state_prefix) const {
+    ggml_tensor * residual = x;
+
+    ggml_tensor * h = causal_conv1d_dw(x, blk.dwconv_w, blk.dwconv_b, state_prefix + "_dwconv"); // [T, C]
+    ggml_tensor * hc = ggml_cont(ctx0, ggml_transpose(ctx0, h)); // [C, T]
+
+    hc = ggml_norm(ctx0, hc, 1e-6f);
+    hc = ggml_mul(ctx0, hc, blk.norm_w);
+    hc = ggml_add(ctx0, hc, blk.norm_b);
+
+    ggml_tensor * g = ggml_mul_mat(ctx0, blk.pw1_w, hc);
+    g = ggml_add(ctx0, g, blk.pw1_b);
+    g = ggml_gelu(ctx0, g);
+    g = ggml_mul_mat(ctx0, blk.pw2_w, g);
+    g = ggml_add(ctx0, g, blk.pw2_b);
+    g = ggml_mul(ctx0, g, blk.gamma);
+
+    ggml_tensor * g_t = ggml_cont(ctx0, ggml_transpose(ctx0, g)); // back to [T, C]
+    return ggml_add(ctx0, residual, g_t);
+}
+
+// SnakeBeta -> dilated causal conv (k=7) -> SnakeBeta -> pointwise causal conv (k=1) -> residual.
+// x: [T, C]. returns [T, C].
+ggml_tensor * clip_graph_qwen3tts_gen::code2wav::dac_res_unit(ggml_tensor * x, const clip_code2wav::dac_res & res, int dilation, const std::string & state_name) const {
+    ggml_tensor * residual = x;
+    ggml_tensor * h = snake(x, res.act1_alpha, res.act1_beta);
+    h = causal_conv1d(h, res.conv1_w, res.conv1_b, dilation, state_name);
+    h = snake(h, res.act2_alpha, res.act2_beta);
+    h = causal_conv1d(h, res.conv2_w, res.conv2_b, 1, ""); // k=1, no left-context needed
+    return ggml_add(ctx0, residual, h);
+}
+
+// RVQ codes -> raw PCM for a batch of N = sliding_window frames
+ggml_tensor * clip_graph_qwen3tts_gen::code2wav::decode(ggml_tensor * inp_codes) const {
+    const auto & c2w = model.c2w;
+
+    // 1. quantizer decode: N frames of 16 codes -> [512, N] (C-first)
+    ggml_tensor * hidden = quant_decode(inp_codes);
+
+    // 2. pre_conv: [512, N] -> T-first [N, 512] -> causal conv k=3 -> [N, 1024]
+    ggml_tensor * x = ggml_cont(ctx0, ggml_transpose(ctx0, hidden)); // [N, 512]
+    x = causal_conv1d(x, c2w.pre_conv_w, c2w.pre_conv_b, 1, "pre_conv"); // [N, 1024]
+    cb(x, "wav_pre_conv_out", -1);
+
+    // 3. pre_transformer: back to C-first [1024, N], project down, run the layers, project back up
+    ggml_tensor * cur = ggml_cont(ctx0, ggml_transpose(ctx0, x)); // [1024, N]
+    cur = ggml_mul_mat(ctx0, c2w.tfm_in_proj_w, cur);
+    cur = ggml_add(ctx0, cur, c2w.tfm_in_proj_b); // [512 (tfm hidden), N]
+
+    for (int il = 0; il < hparams.wav_tfm_n_layer; il++) {
+        cur = tfm_layer_forward(cur, c2w.tfm_layers[il], il);
+    }
+
+    cur = ggml_rms_norm(ctx0, cur, hparams.wav_tfm_eps);
+    cur = ggml_mul(ctx0, cur, c2w.tfm_output_norm_w);
+    cur = ggml_mul_mat(ctx0, c2w.tfm_out_proj_w, cur);
+    cur = ggml_add(ctx0, cur, c2w.tfm_out_proj_b); // [1024, N]
+    cb(cur, "wav_tfm_out", -1);
+
+    // 4. upsample: 2x (causal ConvTranspose1d, stride 2 + ConvNeXt block), back to T-first
+    // kernel == stride here, so there is no overlap tail to persist
+    x = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); // [N, 1024]
+    for (size_t il = 0; il < c2w.upsample.size(); il++) {
+        const auto & up = c2w.upsample[il];
+        x = causal_conv_transpose1d(x, up.conv_w, up.conv_b, 2, "");
+        x = convnext_block(x, up, "up" + std::to_string(il));
+        cb(x, "wav_upsample_out", (int) il);
+    }
+
+    // 5. DAC decoder: conv_pre -> n blocks (SnakeBeta -> ConvTranspose1d -> 3 res units) -> conv_post
+    static constexpr int DAC_DILATIONS[3] = { 1, 3, 9 };
+
+    x = causal_conv1d(x, c2w.dac_entry_w, c2w.dac_entry_b, 1, "dac_entry");
+    cb(x, "wav_dac_entry_out", -1);
+
+    for (size_t il = 0; il < c2w.dac.size(); il++) {
+        const auto & blk = c2w.dac[il];
+        const int stride = (int) (blk.conv_w->ne[0] / 2); // kernel == 2*stride for all 4 blocks
+        const std::string blk_name = "dac" + std::to_string(il);
+        x = snake(x, blk.snake_alpha, blk.snake_beta);
+        x = causal_conv_transpose1d(x, blk.conv_w, blk.conv_b, stride, blk_name + "_tail");
+        for (size_t ir = 0; ir < blk.res.size(); ir++) {
+            x = dac_res_unit(x, blk.res[ir], DAC_DILATIONS[ir], blk_name + "_res" + std::to_string(ir));
+        }
+        cb(x, "wav_dac_block_out", (int) il);
+    }
+
+    x = snake(x, c2w.dac_post_snake_alpha, c2w.dac_post_snake_beta);
+    x = causal_conv1d(x, c2w.dac_post_conv_w, c2w.dac_post_conv_b, 1, "dac_post_conv"); // [n_samples, 1]
+
+    x = ggml_clamp(ctx0, x, -1.0f, 1.0f);
+    x = ggml_reshape_1d(ctx0, x, x->ne[0]);
+    cb(x, "wav_audio_out", -1);
+    return x;
+}
+
+// code2wav's persisted state buffers: RoPE position counter, K/V per pre_transformer layer,
+// left-context/tail per stateful conv. shape lookup only, no graph needed
+std::vector<c2w_state_slot> list_c2w_state_slots(const clip_hparams & hparams, const clip_model & model) {
+    const auto & c2w = model.c2w;
+    std::vector<c2w_state_slot> slots;
+
+    slots.push_back({"tfm_pos", 1, 1});
+
+    // prefix is (W-1) frames, the batch itself gives the other N=W frames (see tfm_layer_forward)
+    const int64_t d_head = c2w.tfm_layers[0].q_w->ne[1] / hparams.wav_tfm_n_head;
+    const int64_t kv_ch  = d_head * hparams.wav_tfm_n_head_kv;
+    const int64_t prefix = hparams.wav_tfm_swa - 1;
+    for (int il = 0; il < hparams.wav_tfm_n_layer; il++) {
+        slots.push_back({"tfm_k_" + std::to_string(il), kv_ch, prefix});
+        slots.push_back({"tfm_v_" + std::to_string(il), kv_ch, prefix});
+    }
+
+    slots.push_back({"pre_conv", c2w.pre_conv_w->ne[0] - 1, c2w.pre_conv_w->ne[1]});
+
+    for (size_t il = 0; il < c2w.upsample.size(); il++) {
+        const auto & up = c2w.upsample[il];
+        slots.push_back({"up" + std::to_string(il) + "_dwconv", up.dwconv_w->ne[0] - 1, up.dwconv_w->ne[2]});
+    }
+
+    slots.push_back({"dac_entry", c2w.dac_entry_w->ne[0] - 1, c2w.dac_entry_w->ne[1]});
+
+    static constexpr int DAC_DILATIONS[3] = { 1, 3, 9 };
+    for (size_t il = 0; il < c2w.dac.size(); il++) {
+        const auto & blk = c2w.dac[il];
+        const int64_t stride = blk.conv_w->ne[0] / 2; // kernel == 2*stride for all 4 blocks
+        const std::string blk_name = "dac" + std::to_string(il);
+        slots.push_back({blk_name + "_tail", stride, blk.conv_w->ne[1]});
+        for (size_t ir = 0; ir < blk.res.size(); ir++) {
+            const auto & res = blk.res[ir];
+            slots.push_back({blk_name + "_res" + std::to_string(ir),
+                              (res.conv1_w->ne[0] - 1) * DAC_DILATIONS[ir], res.conv1_w->ne[1]});
+        }
+    }
+
+    slots.push_back({"dac_post_conv", c2w.dac_post_conv_w->ne[0] - 1, c2w.dac_post_conv_w->ne[1]});
+
+    return slots;
+}
+
+// both sub-graphs are always built, so the topology stays constant
+// ggml_build_forward_select() then picks the one that actually runs
+ggml_cgraph * clip_graph_qwen3tts_gen::build() {
+    GGML_ASSERT(n_batch == 1); // this module only ever processes one frame at a time
+
+    int idx;
+    switch (gen_process) {
+        case CLIP_GEN_PROCESS_GEN_CODE: idx = 0; break;
+        case CLIP_GEN_PROCESS_GEN_WAV:  idx = 1; break;
+        default: GGML_ABORT("unknown gen_process");
+    }
+
+    // ---- CLIP_GEN_PROCESS_GEN_CODE: backbone hidden state -> 16 RVQ codes + next-step embd ----
+    // not build_inp_raw(), a GEN_WAV call's `img` has no hidden-state data
+    ggml_tensor * h_state = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, n_mmproj_embd);
+    ggml_set_name(h_state, "inp_raw"); // must keep this exact name, clip_encode() sets it by name
+    ggml_set_input(h_state);
+
+    ggml_tensor * code0 = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, 1);
+    ggml_set_name(code0, "inp_code0");
+    ggml_set_input(code0);
+
+    ggml_tensor * code0_embd = ggml_get_rows(ctx0, model.gen_code_out_embd_w, code0);
+    code0_embd = ggml_reshape_1d(ctx0, code0_embd, code0_embd->ne[0]);
+    cb(code0_embd, "code0_embd", -1);
+
+    const int64_t n_acoustic = model.gen_code_head_w->ne[2]; // 15
+    const int     n_codes    = (int) n_acoustic + 1;         // 16
+    const int64_t n_kv_pad   = n_codes;
+    const int     n_layer    = (int) model.layers.size();
+    const int     n_head     = hparams.n_head;
+    const int     n_head_kv  = hparams.n_head_kv;
+    const int64_t d_head     = model.layers[0].q_w->ne[1] / n_head;
+
+    // zero-filled per layer k/v caches, so masked-out rows can't hold garbage
+    std::vector<ggml_tensor *> k_cache(n_layer), v_cache(n_layer);
+    for (int il = 0; il < n_layer; il++) {
+        k_cache[il] = ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, d_head * n_head_kv, n_kv_pad), 0.0f);
+        v_cache[il] = ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, d_head * n_head_kv, n_kv_pad), 0.0f);
+    }
+
+    code_gen cg(*this, top_k, top_p);
+
+    ggml_tensor * out_code_cache = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, 1, n_codes);
+    out_code_cache = cg.cache_set(out_code_cache, 0, code0);
+
+    ggml_tensor * inp_rand0 = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+    ggml_set_name(inp_rand0, "inp_rand_0");
+    ggml_set_input(inp_rand0);
+
+    cg.prefill(k_cache, v_cache, out_code_cache, h_state, code0_embd, inp_rand0);
+
+    for (int g = 1; g < n_acoustic; g++) {
+        ggml_tensor * inp_rand = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+        ggml_set_name(inp_rand, ("inp_rand_" + std::to_string(g)).c_str());
+        ggml_set_input(inp_rand);
+        out_code_cache = cg.step(k_cache, v_cache, out_code_cache, inp_rand, g);
+    }
+
+    // output 1: this frame's 16 sampled codes, for the caller's code2wav window
+    ggml_tensor * out_codes = ggml_cont(ctx0, out_code_cache);
+    ggml_set_name(out_codes, "out_codes");
+    ggml_set_output(out_codes);
+
+    // output 2: sum of all 16 codebook embeddings, fed back to the talker for the next frame
+    ggml_tensor * out_embd = code0_embd;
+    for (int g = 1; g <= n_acoustic; g++) {
+        ggml_tensor * code_g = ggml_view_1d(ctx0, out_code_cache, 1, (size_t) g * out_code_cache->nb[1]);
+
+        ggml_tensor * embd_g = ggml_view_2d(ctx0, model.gen_code_embd_w, model.gen_code_embd_w->ne[0], model.gen_code_embd_w->ne[1],
+                                            model.gen_code_embd_w->nb[1], (size_t) (g - 1) * model.gen_code_embd_w->nb[2]);
+        ggml_tensor * e = ggml_get_rows(ctx0, embd_g, code_g);
+        e = ggml_reshape_1d(ctx0, e, e->ne[0]);
+
+        out_embd = ggml_add(ctx0, out_embd, e);
+    }
+    out_embd = ggml_reshape_2d(ctx0, out_embd, out_embd->ne[0], 1);
+    cb(out_embd, "gen_audio_out", -1);
+
+    // ---- CLIP_GEN_PROCESS_GEN_WAV: 16 RVQ codes -> raw PCM ----
+    const int n_frames = hparams.wav_tfm_swa; // frames per batch, == the attention window
+
+    ggml_tensor * inp_codes = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_frames, n_codes);
+    ggml_set_name(inp_codes, "inp_codes");
+    ggml_set_input(inp_codes);
+
+    code2wav c2w(*this);
+    for (const auto & slot : list_c2w_state_slots(hparams, model)) {
+        ggml_tensor * t = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, slot.ne0, slot.ne1);
+        ggml_set_name(t, ("state_in_" + slot.name).c_str());
+        ggml_set_input(t);
+        c2w.state_in[slot.name] = t;
+    }
+
+    ggml_tensor * out_audio = c2w.decode(inp_codes);
+    ggml_set_name(out_audio, "out_audio");
+    ggml_set_output(out_audio);
+
+    for (auto & slot : c2w.state_out) {
+        ggml_set_name(slot.second, ("state_out_" + slot.first).c_str());
+        ggml_set_output(slot.second);
+    }
+
+    // out_embd goes last, clip_encode() reads it back via ggml_graph_node(gf, -1)
+    ggml_tensor * outs[2];
+    outs[0] = out_codes; outs[1] = out_audio;
+    ggml_build_forward_select(gf, outs, 2, idx);
+    for (auto & slot : c2w.state_out) {
+        outs[0] = out_codes; outs[1] = slot.second;
+        ggml_build_forward_select(gf, outs, 2, idx);
+    }
+    outs[0] = out_embd; outs[1] = out_audio;
+    ggml_build_forward_select(gf, outs, 2, idx);
+
+    return gf;
+}
@@ -0,0 +1,197 @@
+#include "models.h"
+
+static constexpr int SPK_RES2NET_SCALE = 8; // enc_res2net_scale
+static constexpr int SPK_DILATIONS[3]  = { 2, 3, 4 }; // enc_dilations[1..3]
+
+// conv1d, kernel K, padding "same" (reflect), dilation d
+// x: [C, T] (ne[0]=C, ne[1]=T) -> [out_c, T]
+ggml_tensor * clip_graph_qwen3tts_spkenc::conv1d_same(ggml_tensor * x, ggml_tensor * w, ggml_tensor * b, int dilation) const {
+    const int K   = (int) w->ne[0];
+    const int IC  = (int) w->ne[1];
+    const int OC  = (int) w->ne[2];
+    const int pad = ((K - 1) * dilation) / 2;
+
+    // ggml_pad_reflect_1d pads ne[0], so bring T onto ne[0] first, same layout as im2col wants
+    ggml_tensor * x_t = ggml_cont(ctx0, ggml_transpose(ctx0, x)); // [T, IC]
+    if (pad > 0) {
+        x_t = ggml_pad_reflect_1d(ctx0, x_t, pad, pad); // [T + 2*pad, IC]
+    }
+    ggml_tensor * x4d = ggml_reshape_4d(ctx0, x_t, x_t->ne[0], IC, 1, 1);
+
+    // dummy F32 kernel, im2col only reads its shape, so a quantized w does not assert
+    ggml_tensor * dummy = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, K, IC, 1, 1);
+
+    ggml_tensor * col = ggml_im2col(ctx0, dummy, x4d, 1, 1, 0, 0, dilation, 1, false, GGML_TYPE_F32);
+    const int64_t T_out = col->ne[1];
+    col = ggml_reshape_2d(ctx0, col, (int64_t) K * IC, T_out);
+
+    ggml_tensor * w2d = ggml_reshape_2d(ctx0, w, (int64_t) K * IC, OC);
+    ggml_tensor * y   = ggml_mul_mat(ctx0, w2d, col); // [OC, T_out]
+    ggml_mul_mat_set_prec(y, GGML_PREC_F32);
+
+    ggml_tensor * b2d = ggml_reshape_2d(ctx0, b, OC, 1);
+    y = ggml_add(ctx0, y, b2d);
+    return y;
+}
+
+// Res2Net: split channel axis into `scale` chunks, chain dilated conv1d branches
+// x: [C, T] -> [C, T]
+ggml_tensor * clip_graph_qwen3tts_spkenc::res2net(ggml_tensor * x, const clip_layer & layer, int dilation, int scale) const {
+    const int64_t C  = x->ne[0];
+    const int64_t T  = x->ne[1];
+    const int64_t Cs = C / scale;
+
+    std::vector<ggml_tensor *> outs;
+    outs.reserve(scale);
+
+    auto chunk = [&](int i) -> ggml_tensor * {
+        return ggml_view_2d(ctx0, x, Cs, T, x->nb[1], (size_t) i * Cs * x->nb[0]);
+    };
+
+    ggml_tensor * prev = nullptr;
+    for (int i = 0; i < scale; i++) {
+        ggml_tensor * c = ggml_cont(ctx0, chunk(i));
+        if (i == 0) {
+            outs.push_back(c);
+            continue;
+        }
+        ggml_tensor * inp = (i >= 2) ? ggml_add(ctx0, c, prev) : c;
+        ggml_tensor * y   = conv1d_same(inp, layer.res2_conv_w[i - 1], layer.res2_conv_b[i - 1], dilation);
+        y                 = ggml_relu(ctx0, y);
+        outs.push_back(y);
+        prev = y;
+    }
+
+    ggml_tensor * acc = outs[0];
+    for (int i = 1; i < scale; i++) {
+        acc = ggml_concat(ctx0, acc, outs[i], 0);
+    }
+    return acc;
+}
+
+// squeeze-and-excitation gate. x: [C, T] -> [C, T]
+ggml_tensor * clip_graph_qwen3tts_spkenc::se_block(ggml_tensor * x, const clip_layer & layer) const {
+    // temporal mean, keepdim: transpose so T is on ne[0], reduce, transpose back
+    ggml_tensor * x_t  = ggml_cont(ctx0, ggml_transpose(ctx0, x));    // [T, C]
+    ggml_tensor * mean = ggml_mean(ctx0, x_t);                        // [1, C]
+    mean               = ggml_cont(ctx0, ggml_transpose(ctx0, mean)); // [C, 1]
+
+    ggml_tensor * h = conv1d_same(mean, layer.se_conv1_w, layer.se_conv1_b, 1);
+    h = ggml_relu(ctx0, h);
+    h = conv1d_same(h, layer.se_conv2_w, layer.se_conv2_b, 1);
+    h = ggml_sigmoid(ctx0, h); // [C, 1]
+
+    return ggml_mul(ctx0, x, h); // broadcast gate over T
+}
+
+// tdnn1 -> res2net -> tdnn2 -> se, plus residual. x: [C, T] -> [C, T]
+ggml_tensor * clip_graph_qwen3tts_spkenc::se_res2net_block(ggml_tensor * x, const clip_layer & layer, int dilation, int scale) const {
+    ggml_tensor * residual = x;
+    ggml_tensor * h  = conv1d_same(x, layer.conv_pw1_w, layer.conv_pw1_b, 1); // tdnn1
+    h = ggml_relu(ctx0, h);
+    h = res2net(h, layer, dilation, scale);
+    h = conv1d_same(h, layer.conv_pw2_w, layer.conv_pw2_b, 1); // tdnn2
+    h = ggml_relu(ctx0, h);
+    h = se_block(h, layer);
+    return ggml_add(ctx0, h, residual);
+}
+
+// attentive statistics pooling. x: [C, T] -> [2*C, 1]
+ggml_tensor * clip_graph_qwen3tts_spkenc::attentive_stats_pool(ggml_tensor * x) const {
+    const int64_t T = x->ne[1];
+
+    // mean over T: [C, 1]
+    ggml_tensor * x_t  = ggml_cont(ctx0, ggml_transpose(ctx0, x));
+    ggml_tensor * mean = ggml_mean(ctx0, x_t);
+    mean               = ggml_cont(ctx0, ggml_transpose(ctx0, mean));
+
+    // std over T: sqrt(clamp(mean((x - mean)^2), eps))
+    ggml_tensor * mean_rep = ggml_repeat(ctx0, mean, x);
+    ggml_tensor * centered = ggml_sub(ctx0, x, mean_rep);
+    ggml_tensor * var_t    = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_sqr(ctx0, centered)));
+    ggml_tensor * var      = ggml_mean(ctx0, var_t);
+    var                    = ggml_cont(ctx0, ggml_transpose(ctx0, var));
+    var                    = ggml_scale_bias(ctx0, var, 1.0f, 1e-12f);
+    ggml_tensor * std      = ggml_sqrt(ctx0, var);
+
+    // attention input: cat([x, mean, std]) along channel axis -> [3C, T]
+    ggml_tensor * std_rep = ggml_repeat(ctx0, std, x);
+    ggml_tensor * cat     = ggml_concat(ctx0, x, mean_rep, 0);
+    cat                   = ggml_concat(ctx0, cat, std_rep, 0);
+
+    // attention TDNN (3C -> attn_c) + ReLU, tanh, then 1x1 conv (attn_c -> C)
+    ggml_tensor * a = conv1d_same(cat, model.spk_asp_tdnn_w, model.spk_asp_tdnn_b, 1);
+    a = ggml_relu(ctx0, a);
+    a = ggml_tanh(ctx0, a);
+    a = conv1d_same(a, model.spk_asp_attn_w, model.spk_asp_attn_b, 1);
+
+    // softmax over T
+    ggml_tensor * a_t = ggml_cont(ctx0, ggml_transpose(ctx0, a));   // [T, C]
+    ggml_tensor * w_t = ggml_soft_max(ctx0, a_t);
+    ggml_tensor * w   = ggml_cont(ctx0, ggml_transpose(ctx0, w_t)); // [C, T]
+
+    // weighted mean: sum(w * x) over T, multiply by T to undo ggml_mean's 1/T scaling
+    ggml_tensor * wx     = ggml_mul(ctx0, w, x);
+    ggml_tensor * wx_t   = ggml_cont(ctx0, ggml_transpose(ctx0, wx));
+    ggml_tensor * w_mean = ggml_mean(ctx0, wx_t);
+    w_mean = ggml_scale(ctx0, w_mean, (float) T);
+    w_mean = ggml_cont(ctx0, ggml_transpose(ctx0, w_mean)); // [C, 1]
+
+    // weighted std: sum(w * (x - w_mean)^2) over T
+    ggml_tensor * w_mean_rep = ggml_repeat(ctx0, w_mean, x);
+    ggml_tensor * dev        = ggml_sub(ctx0, x, w_mean_rep);
+    ggml_tensor * w_var_in   = ggml_mul(ctx0, w, ggml_sqr(ctx0, dev));
+    ggml_tensor * w_var_t    = ggml_cont(ctx0, ggml_transpose(ctx0, w_var_in));
+    ggml_tensor * w_var      = ggml_mean(ctx0, w_var_t);
+    w_var                    = ggml_scale(ctx0, w_var, (float) T);
+    w_var                    = ggml_cont(ctx0, ggml_transpose(ctx0, w_var));
+    w_var                    = ggml_scale_bias(ctx0, w_var, 1.0f, 1e-12f);
+    ggml_tensor * w_std      = ggml_sqrt(ctx0, w_var);
+
+    return ggml_concat(ctx0, w_mean, w_std, 0); // [2C, 1]
+}
+
+ggml_cgraph * clip_graph_qwen3tts_spkenc::build() {
+    // inp_raw: [T, n_mel, 1, 1], from mtmd_audio_preprocessor_qwen3tts_spk
+    ggml_tensor * inp = build_inp_raw(1);
+    inp = ggml_reshape_2d(ctx0, inp, inp->ne[0], inp->ne[1]);
+
+    // this file's convention is [C, T]; the preprocessor delivers [T, C]
+    ggml_tensor * mel = ggml_cont(ctx0, ggml_transpose(ctx0, inp)); // [n_mel, T]
+    cb(mel, "mel", -1);
+
+    // frontend conv0 TDNN k=5, dilation=1: 128 -> 512
+    ggml_tensor * cur = conv1d_same(mel, model.conv1d_1_w, model.conv1d_1_b, 1);
+    cur = ggml_relu(ctx0, cur);
+    cb(cur, "frontend", -1);
+
+    // 3 SE-Res2Net blocks at dilations 2, 3, 4
+    GGML_ASSERT((int) model.layers.size() == 3);
+    std::vector<ggml_tensor *> blk_out(3);
+    for (int il = 0; il < 3; il++) {
+        cur = se_res2net_block(cur, model.layers[il], SPK_DILATIONS[il], SPK_RES2NET_SCALE);
+        blk_out[il] = cur;
+        cb(cur, "block_out", il);
+    }
+
+    // multi-layer feature aggregation: cat blk[0..2] then TDNN k=1 + ReLU
+    ggml_tensor * cat = ggml_concat(ctx0, blk_out[0], blk_out[1], 0);
+    cat = ggml_concat(ctx0, cat, blk_out[2], 0); // [1536, T]
+    ggml_tensor * mfa = conv1d_same(cat, model.conv_out_w, model.conv_out_b, 1);
+    mfa = ggml_relu(ctx0, mfa);
+    cb(mfa, "mfa", -1);
+
+    // attentive statistics pooling: [1536, T] -> [3072, 1]
+    ggml_tensor * stats = attentive_stats_pool(mfa);
+    cb(stats, "asp", -1);
+
+    // final FC k=1: [3072, 1] -> [enc_dim, 1]
+    ggml_tensor * emb = conv1d_same(stats, model.mm_fc_w, model.mm_fc_b, 1);
+
+    emb = ggml_reshape_1d(ctx0, emb, emb->ne[0]);
+    emb = ggml_cont(ctx0, emb);
+    cb(emb, "spk_embedding", -1);
+
+    ggml_build_forward_expand(gf, emb);
+    return gf;
+}
@@ -556,10 +556,8 @@ bool mtmd_audio_preprocessor_whisper::preprocess(const float *                 s
    }

    std::vector<float> smpl;
-    // if input is too short, pad with zeros
-    // this is to avoid potential issues with stage1/2 padding in log_mel_spectrogram
-    // TODO: maybe handle this better
-    size_t min_samples = (size_t) hparams.audio_sample_rate * (hparams.audio_chunk_len + 1);  // +1 second margin
+    // reflection padding needs one sample plus half an FFT window
+    size_t min_samples = (size_t) hparams.audio_n_fft / 2 + 1;
    if (n_samples < min_samples) {
        smpl.resize(min_samples, 0.0f);
        std::memcpy(smpl.data(), samples, n_samples * sizeof(float));
@@ -791,6 +789,66 @@ bool mtmd_audio_preprocessor_mimo_audio::preprocess(const float *
    return true;
 }

+//
+// mtmd_audio_preprocessor_qwen3tts_spk
+//
+// same as mel_spectrogram() in modeling_qwen3_tts.py
+// ECAPA-TDNN takes the whole clip in one pass, so no Whisper-style chunking or normalization
+//
+
+void mtmd_audio_preprocessor_qwen3tts_spk::initialize() {
+    cache.fill_sin_cos_table(hparams.audio_n_fft);
+    cache.fill_hann_window(hparams.audio_window_len, true);
+    cache.fill_mel_filterbank_matrix(hparams.n_mel_bins, hparams.audio_n_fft, hparams.audio_sample_rate);
+}
+
+bool mtmd_audio_preprocessor_qwen3tts_spk::preprocess(const float *                 samples,
+                                                      size_t                        n_samples,
+                                                      std::vector<mtmd_audio_mel> & output) {
+    if (n_samples == 0) {
+        return false;
+    }
+
+    GGML_ASSERT(!cache.sin_vals.empty());
+    GGML_ASSERT(!cache.cos_vals.empty());
+    GGML_ASSERT(!cache.filters.data.empty());
+
+    // reflect pad by (n_fft - hop) / 2 = 384, matching center=False STFT framing
+    const int pad = (hparams.audio_n_fft - hparams.audio_hop_len) / 2;
+    if (n_samples < (size_t) pad + 1) {
+        return false;
+    }
+
+    std::vector<float> padded(n_samples + 2 * pad, 0.0f);
+    for (int i = 0; i < pad; i++) {
+        padded[i] = samples[pad - i];
+    }
+    std::copy(samples, samples + n_samples, padded.begin() + pad);
+    for (int i = 0; i < pad; i++) {
+        padded[n_samples + pad + i] = samples[n_samples - 2 - i];
+    }
+
+    filter_params params;
+    params.n_mel            = hparams.n_mel_bins;
+    params.n_fft_bins       = 1 + (hparams.audio_n_fft / 2);
+    params.hann_window_size = hparams.audio_window_len;
+    params.hop_length       = hparams.audio_hop_len;
+    params.sample_rate      = hparams.audio_sample_rate;
+    params.no_padding       = true; // reflect padding already applied above
+    params.use_natural_log  = true;
+    params.use_magnitude    = true;
+    params.mel_floor        = 1e-5f;
+
+    mtmd_audio_mel out;
+    bool ok = log_mel_spectrogram(padded.data(), (int) padded.size(), 4, params, cache, out);
+    if (!ok) {
+        return false;
+    }
+
+    output.push_back(std::move(out));
+    return true;
+}
+
 //
 // mtmd_audio_preprocessor_conformer
 //
@@ -120,6 +120,15 @@ struct mtmd_audio_preprocessor_mimo_audio : mtmd_audio_preprocessor {
    mtmd_audio_cache cache;
 };

+struct mtmd_audio_preprocessor_qwen3tts_spk : mtmd_audio_preprocessor {
+    mtmd_audio_preprocessor_qwen3tts_spk(const clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) {}
+    void initialize() override;
+    bool preprocess(const float * samples, size_t n_samples, std::vector<mtmd_audio_mel> & output) override;
+
+  private:
+    mtmd_audio_cache cache;
+};
+
 struct mtmd_audio_preprocessor_parakeet : mtmd_audio_preprocessor {
    mtmd_audio_preprocessor_parakeet(clip_ctx * ctx) : mtmd_audio_preprocessor(ctx) { }
    void initialize() override;
@@ -116,6 +116,14 @@ struct mtmd_cli_context {
            exit(1);
        }

+        init_vision_context(params);
+
+        if (!mtmd_helper_model_can_chat(lctx, ctx_vision.get())) {
+            LOG_ERR("Model does not support chat mode\n");
+            LOG_ERR("Hint: for TTS models, please use llama-tts\n");
+            exit(1);
+        }
+
        if (!llama_model_chat_template(model, nullptr) && params.chat_template.empty()) {
            LOG_ERR("Model does not have chat template.\n");
            LOG_ERR("  For old llava models, you may need to use '--chat-template vicuna'\n");
@@ -129,8 +137,6 @@ struct mtmd_cli_context {
        chat_history.clear();
        LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(tmpls.get(), params.use_jinja, params.default_template_kwargs).c_str());

-        init_vision_context(params);
-
        // load antiprompt tokens for legacy templates
        if (params.chat_template == "vicuna") {
            antiprompt_tokens = common_tokenize(lctx, "ASSISTANT:", false, true);
@@ -0,0 +1,180 @@
+#pragma once
+
+// shared internal utilities for the mtmd-helper-*.cpp translation units
+// (mtmd-helper.cpp, mtmd-helper-gen.cpp)
+// NOT part of the public mtmd-helper.h API
+
+#include "ggml.h"
+#include "llama.h"
+#include "mtmd.h"
+
+#include <cstdarg>
+#include <cstdio>
+#include <cstdlib>
+#include <vector>
+
+//
+// logging
+//
+
+struct mtmd_helper_logger {
+    ggml_log_callback default_callback = [](ggml_log_level level, const char * text, void * user_data) {
+        (void) level;
+        (void) user_data;
+        fputs(text, stderr);
+        fflush(stderr);
+    };
+
+    ggml_log_callback log_callback = default_callback;
+    void * log_callback_user_data;
+
+    void log_v(enum ggml_log_level level, const char * format, va_list args) {
+        if (format == NULL) {
+            return;
+        }
+        va_list args_copy;
+        va_copy(args_copy, args);
+        char buffer[128];
+        int len = vsnprintf(buffer, 128, format, args);
+        if (len < 128) {
+            log_callback(level, buffer, log_callback_user_data);
+        } else {
+            char * buffer2 = (char *) calloc(len + 1, sizeof(char));
+            vsnprintf(buffer2, len + 1, format, args_copy);
+            buffer2[len] = 0;
+            log_callback(level, buffer2, log_callback_user_data);
+            free(buffer2);
+        }
+        va_end(args_copy);
+    }
+
+    void log(enum ggml_log_level level, const char * format, ...) {
+        va_list args;
+        va_start(args, format);
+        log_v(level, format, args);
+        va_end(args);
+    }
+};
+
+// inline, so all TUs including this header share one instance
+inline mtmd_helper_logger g_logger;
+
+#define LOG_DBG(...) g_logger.log(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
+#define LOG_INF(...) g_logger.log(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
+#define LOG_WRN(...) g_logger.log(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)
+#define LOG_ERR(...) g_logger.log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+
+//
+// embd batch
+//
+
+// helper struct to make working with embd batch easier
+// note: this will be removed after llama_batch_ext refactoring
+struct decode_embd_batch {
+    int n_pos_per_embd;
+    int n_mmproj_embd;
+    std::vector<llama_pos>      pos;
+    std::vector<llama_pos>      pos_view; // used by mrope
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id>   seq_id_0;
+    std::vector<llama_seq_id *> seq_ids;
+    std::vector<int8_t>         logits;
+    llama_batch batch;
+    decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
+        GGML_ASSERT(n_tokens > 0 && n_pos_per_embd > 0 && n_mmproj_embd > 0);
+        pos     .resize(n_tokens * n_pos_per_embd);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        seq_ids [n_tokens] = nullptr;
+        batch = {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ embd,
+            /*pos            =*/ pos.data(),
+            /*n_seq_id       =*/ n_seq_id.data(),
+            /*seq_id         =*/ seq_ids.data(),
+            /*logits         =*/ logits.data(),
+        };
+    }
+
+    void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) {
+        seq_id_0[0] = seq_id;
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.pos     [i] = pos_0 + i;
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    // M-RoPE for image
+    void set_position_mrope_2d(const std::vector<mtmd_decoder_pos> & rel_pos, llama_seq_id seq_id) {
+        GGML_ASSERT(n_pos_per_embd == 4);
+        GGML_ASSERT(!rel_pos.empty() && (int32_t)rel_pos.size() == batch.n_tokens);
+        seq_id_0[0] = seq_id;
+        for (int32_t i = 0; i < batch.n_tokens; i++) {
+            pos[i                     ] = rel_pos[i].t;
+            pos[i + batch.n_tokens    ] = rel_pos[i].y;
+            pos[i + batch.n_tokens * 2] = rel_pos[i].x;
+            pos[i + batch.n_tokens * 3] = rel_pos[i].z;
+        }
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    // M-RoPE for audio
+    void set_position_mrope_1d(llama_pos pos_0, llama_seq_id seq_id) {
+        GGML_ASSERT(n_pos_per_embd == 4);
+        seq_id_0[0] = seq_id;
+        for (int i = 0; i < batch.n_tokens; i++) {
+            pos[i                     ] = pos_0 + i;
+            pos[i + batch.n_tokens    ] = pos_0 + i;
+            pos[i + batch.n_tokens * 2] = pos_0 + i;
+            pos[i + batch.n_tokens * 3] = pos_0 + i;
+        }
+        for (int i = 0; i < batch.n_tokens; i++) {
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+
+    llama_batch get_view(int offset, int n_tokens) {
+        GGML_ASSERT(offset >= 0 && n_tokens > 0 && offset + n_tokens <= batch.n_tokens);
+        llama_pos * pos_ptr;
+        pos_view.clear();
+        pos_view.reserve(n_tokens * n_pos_per_embd);
+        if (n_pos_per_embd > 1) {
+            // mrope
+            // for example, with layout of src: 1234...1234...1234...1234...
+            //       offset 2 will give us dst: 34...34...34...34...
+            for (int i = 0; i < n_pos_per_embd; i++) {
+                // assume n_tokens is less than or equal to batch.n_tokens
+                // batch.n_tokens is number of **total** tokens
+                // n_tokens is number of viewed token
+                size_t src_idx = i * batch.n_tokens + offset;
+                pos_view.insert(pos_view.end(),
+                    pos.data() + src_idx,
+                    pos.data() + src_idx + n_tokens);
+            }
+            pos_ptr = pos_view.data();
+        } else {
+            // normal
+            pos_ptr = pos.data() + offset;
+        }
+        return {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ batch.embd     + offset * n_mmproj_embd,
+            /*pos            =*/ pos_ptr,
+            /*n_seq_id       =*/ batch.n_seq_id + offset,
+            /*seq_id         =*/ batch.seq_id   + offset,
+            /*logits         =*/ batch.logits   + offset,
+        };
+    }
+};
@@ -0,0 +1,505 @@
+#include "mtmd.h"
+#include "mtmd-helper.h"
+#include "mtmd-helper-common.h"
+#include "llama.h"
+#include "../src/llama-ext.h"
+
+#include <algorithm>
+#include <cstring>
+#include <memory>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#ifdef MTMD_INTERNAL_HEADER
+#error "mtmd-helper is a public library outside of mtmd. it must not include internal headers"
+#endif
+
+//
+// Audio generation helpers
+//
+
+// --tts-lang codes -> language names used by the codec_language special tokens
+static const std::unordered_map<std::string, std::string> tts_lang_codes = {
+    { "zh", "chinese"    },
+    { "en", "english"    },
+    { "de", "german"     },
+    { "it", "italian"    },
+    { "pt", "portuguese" },
+    { "es", "spanish"    },
+    { "ja", "japanese"   },
+    { "ko", "korean"     },
+    { "fr", "french"     },
+    { "ru", "russian"    },
+};
+
+static std::string tts_resolve_lang(const std::string & lang) {
+    auto it = tts_lang_codes.find(lang);
+    return it != tts_lang_codes.end() ? it->second : lang;
+}
+
+static llama_token find_special_token(const llama_vocab * vocab, const std::string & piece) {
+    const int32_t n = llama_vocab_n_tokens(vocab);
+    for (llama_token t = 0; t < n; t++) {
+        if (piece == llama_vocab_get_text(vocab, t)) {
+            return t;
+        }
+    }
+    return LLAMA_TOKEN_NULL;
+}
+
+static bool write_wav16(std::vector<char> & buf, const std::vector<float> & pcm, int32_t rate) {
+    // RIFF chunk sizes are 32-bit; refuse to emit a file with a truncated header
+    if (pcm.size() > ((size_t) UINT32_MAX - 36) / 2) {
+        return false;
+    }
+    const uint32_t data_sz   = (uint32_t) (pcm.size() * 2);
+    const uint32_t riff_sz   = 36 + data_sz;
+    const uint32_t fmt_sz    = 16, byte_rate = (uint32_t) rate * 2;
+    const uint16_t fmt = 1, ch = 1, align = 2, bits = 16;
+    const uint32_t rate32    = (uint32_t) rate;
+    auto put = [&](const void * p, size_t n) {
+        const char * c = (const char *) p;
+        buf.insert(buf.end(), c, c + n);
+    };
+    put("RIFF", 4); put(&riff_sz, 4); put("WAVE", 4);
+    put("fmt ", 4); put(&fmt_sz, 4);
+    put(&fmt, 2); put(&ch, 2); put(&rate32, 4);
+    put(&byte_rate, 4); put(&align, 2); put(&bits, 2);
+    put("data", 4); put(&data_sz, 4);
+    for (float v : pcm) {
+        int16_t s = (int16_t) (std::max(-1.0f, std::min(1.0f, v)) * 32767.0f);
+        put(&s, 2);
+    }
+    return true;
+}
+
+class mtmd_gen_audio_pipeline {
+public:
+    mtmd_gen_audio_pipeline(llama_context * lctx, mtmd_context * mctx)
+        : lctx(lctx), mctx(mctx), model(llama_get_model(lctx)), vocab(llama_model_get_vocab(model)),
+          n_embd(llama_model_n_embd(model)), info(mtmd_gen_audio_get_info(mctx)) {}
+    virtual ~mtmd_gen_audio_pipeline() = default;
+
+    virtual void reset() = 0;
+    virtual int32_t set_input(const mtmd_helper_gen_audio_inp * inp) = 0;
+    // decodes at most n_batch prompt tokens; returns remaining count (0 = done), <0 on error
+    virtual int32_t step_prompt(int32_t n_batch) = 0;
+    // sampled can be LLAMA_TOKEN_NULL for pipelines with no discrete backbone token,
+    // those read what they need from h_state_in instead
+    virtual int32_t step_gen(llama_token sampled, const float * h_state_in, const float ** h_state_out) = 0;
+    virtual int32_t get_output(int32_t * out_sample_rate, const char ** out_data, size_t * out_data_len, int64_t * out_n_samples) = 0;
+
+protected:
+    llama_context * lctx;
+    mtmd_context  * mctx;
+    const llama_model * model;
+    const llama_vocab  * vocab;
+    int n_embd;
+    mtmd_gen_audio_info info;
+};
+
+// Qwen3-TTS: backbone samples codec_0, code_predictor gives the other 15 codebooks,
+// then code2wav decodes them to PCM
+class qwen3tts_gen_audio_pipeline : public mtmd_gen_audio_pipeline {
+public:
+    using mtmd_gen_audio_pipeline::mtmd_gen_audio_pipeline;
+
+    void reset() override {
+        seq_id = 0;
+        pos = 0;
+        codes_buf.clear();
+        c2w_state.clear();
+        audio_pcm.clear();
+        overlay.clear();
+        overlay_idx = 0;
+        h_state_buf.clear();
+        out_buf.clear();
+        prompt_embd_buf.clear();
+        prompt_batch.reset();
+        n_prompt = 0;
+        prompt_pos = 0;
+    }
+
+    int32_t set_input(const mtmd_helper_gen_audio_inp * inp) override {
+        reset();
+        seq_id = inp->seq_id;
+
+        if (!ensure_cache()) {
+            return 1;
+        }
+
+        const std::string lang   = tts_resolve_lang((inp->lang && inp->lang[0]) ? inp->lang : "english");
+        const llama_token c_lang = find_special_token(vocab, ("<|codec_language_" + lang + "|>").c_str());
+        if (c_lang == LLAMA_TOKEN_NULL) {
+            LOG_ERR("mtmd_helper_gen_audio: unknown language '%s'\n", lang.c_str());
+            return 1;
+        }
+
+        std::vector<float> speaker_embd;
+        if (inp->speaker_ref) {
+            if (!encode_speaker(inp->speaker_ref, speaker_embd)) {
+                return 1;
+            }
+        }
+
+        const int n_e = n_embd;
+        auto row = [&](llama_token t) {
+            return std::vector<float>(tok_embd.begin() + (size_t) t * n_e,
+                                       tok_embd.begin() + (size_t) (t + 1) * n_e);
+        };
+        auto sum_row = [&](llama_token a, llama_token b) {
+            std::vector<float> va = row(a), vb = row(b);
+            for (int i = 0; i < n_e; i++) va[(size_t) i] += vb[(size_t) i];
+            return va;
+        };
+        auto sum_vec = [&](llama_token a, const std::vector<float> & vb) {
+            std::vector<float> va = row(a);
+            for (int i = 0; i < n_e; i++) va[(size_t) i] += vb[(size_t) i];
+            return va;
+        };
+
+        // upstream chat wrap, then slices: [0:3] role, [3:-5] utterance body
+        const std::string full = "<|im_start|>assistant\n" + std::string(inp->prompt, inp->prompt_len) +
+                                  "<|im_end|>\n<|im_start|>assistant\n";
+        std::vector<llama_token> ids(full.size() + 16);
+        int n_ids = llama_tokenize(vocab, full.c_str(), (int32_t) full.size(), ids.data(), (int32_t) ids.size(),
+                                   false, true);
+        if (n_ids < 8) {
+            LOG_ERR("mtmd_helper_gen_audio: tokenization failed\n");
+            return 1;
+        }
+        ids.resize((size_t) n_ids);
+
+        std::vector<std::vector<float>> prompt;
+        for (int i = 0; i < 3; i++) prompt.push_back(row(ids[(size_t) i]));
+        prompt.push_back(sum_row(tts_pad, c_think));
+        prompt.push_back(sum_row(tts_pad, c_think_b));
+        prompt.push_back(sum_row(tts_pad, c_lang));
+        prompt.push_back(sum_row(tts_pad, c_think_e));
+        if (!speaker_embd.empty()) prompt.push_back(sum_vec(tts_pad, speaker_embd));
+        prompt.push_back(sum_row(tts_bos, codec_pad));
+        for (int i = 3; i < n_ids - 5; i++) prompt.push_back(sum_row(ids[(size_t) i], codec_pad));
+        prompt.push_back(sum_row(tts_eos, codec_pad));
+        prompt.push_back(sum_row(tts_pad, codec_bos));
+
+        n_prompt = (int) prompt.size();
+
+        // the talker uses the qwen3vl interleaved mrope, all sections are equal for a text/codec stream
+        mrope = llama_model_rope_type(model) == LLAMA_ROPE_TYPE_MROPE ||
+                llama_model_rope_type(model) == LLAMA_ROPE_TYPE_IMROPE;
+        const int n_pos_per_embd = mrope ? 4 : 1;
+
+        prompt_embd_buf.resize((size_t) n_prompt * (size_t) n_e);
+        for (int i = 0; i < n_prompt; i++) {
+            memcpy(prompt_embd_buf.data() + (size_t) i * n_e, prompt[(size_t) i].data(), (size_t) n_e * sizeof(float));
+        }
+
+        prompt_batch.reset(new decode_embd_batch(prompt_embd_buf.data(), n_prompt, n_pos_per_embd, n_e));
+        if (mrope) prompt_batch->set_position_mrope_1d(0, seq_id);
+        else       prompt_batch->set_position_normal  (0, seq_id);
+        prompt_pos = 0;
+
+        pos = 0;
+        top_k = inp->top_k > 0 ? inp->top_k : 50;
+        top_p = inp->top_p > 0 ? inp->top_p : 1.0f;
+        out_type = inp->out_type;
+
+        // the text stream keeps flowing during generation: after frame k, the input adds
+        // trailing text row k on top of the codes embedding, then tts_eos, then tts_pad
+        for (int i = 3; i < n_ids - 5; i++) overlay.push_back(row(ids[(size_t) i]));
+        overlay.push_back(row(tts_eos));
+        overlay.push_back(row(tts_pad));
+
+        return 0;
+    }
+
+    int32_t step_prompt(int32_t n_batch) override {
+        GGML_ASSERT(n_batch > 0);
+        if (prompt_pos >= n_prompt) {
+            return 0;
+        }
+        const int32_t n_tokens_batch = std::min(n_batch, n_prompt - prompt_pos);
+        llama_batch batch_view = prompt_batch->get_view(prompt_pos, n_tokens_batch);
+
+        const bool is_last_batch = (prompt_pos + n_tokens_batch) == n_prompt;
+        if (is_last_batch) {
+            batch_view.logits[n_tokens_batch - 1] = 1;
+        }
+
+        if (llama_decode(lctx, batch_view) != 0) {
+            LOG_ERR("mtmd_helper_gen_audio: prompt decode failed\n");
+            return -1;
+        }
+
+        pos        += n_tokens_batch;
+        prompt_pos += n_tokens_batch;
+
+        if (prompt_pos >= n_prompt) {
+            // prompt fully processed, its embedding buffer is no longer needed
+            prompt_batch.reset();
+            prompt_embd_buf.clear();
+            return 0;
+        }
+        return n_prompt - prompt_pos;
+    }
+
+    int32_t step_gen(llama_token sampled, const float * h_state_in, const float ** h_state_out) override {
+        mtmd_gen_inp inp{};
+        inp.type  = MTMD_GEN_PROCESS_TYPE_GEN_CODE;
+        inp.code0 = sampled - codec_0;
+        inp.embd  = const_cast<float *>(h_state_in);
+        inp.top_k = top_k;
+        inp.top_p = top_p;
+        mtmd_gen_out out{};
+        if (mtmd_gen_audio_process(mctx, &inp, &out) != 0) {
+            LOG_ERR("mtmd_helper_gen_audio: gen_code process failed\n");
+            return 1;
+        }
+
+        codes_buf.insert(codes_buf.end(), out.codes, out.codes + out.n_codes);
+        if (out.n_codes > 0 && codes_buf.size() / out.n_codes >= window_frames) {
+            if (!flush_gen_wav()) {
+                return 1;
+            }
+        }
+
+        std::vector<float> fb(out.embd, out.embd + n_embd);
+        const auto & ov = overlay[std::min(overlay_idx, overlay.size() - 1)];
+        for (int i = 0; i < n_embd; i++) fb[(size_t) i] += ov[(size_t) i];
+        overlay_idx++;
+
+        const int n_pos_per_embd = mrope ? 4 : 1;
+        decode_embd_batch batch_embd(fb.data(), 1, n_pos_per_embd, n_embd);
+        if (mrope) batch_embd.set_position_mrope_1d(pos, seq_id);
+        else       batch_embd.set_position_normal  (pos, seq_id);
+        batch_embd.batch.logits[0] = 1;
+        pos++;
+
+        if (llama_decode(lctx, batch_embd.batch) != 0) {
+            LOG_ERR("mtmd_helper_gen_audio: decode failed\n");
+            return 1;
+        }
+
+        const float * he = llama_get_embeddings_ith(lctx, -1);
+        h_state_buf.assign(he, he + n_embd);
+        *h_state_out = h_state_buf.data();
+
+        return 0;
+    }
+
+    int32_t get_output(int32_t * out_sample_rate, const char ** out_data, size_t * out_data_len, int64_t * out_n_samples) override {
+        if (!flush_gen_wav()) {
+            return 1;
+        }
+
+        *out_sample_rate = info.sample_rate;
+        if (out_n_samples) {
+            *out_n_samples = (int64_t) audio_pcm.size();
+        }
+
+        if (out_type == MTMD_HELPER_GEN_AUDIO_OUTTYPE_PCM) {
+            *out_data     = (const char *) audio_pcm.data();
+            *out_data_len = audio_pcm.size() * sizeof(float);
+            return 0;
+        }
+
+        out_buf.clear();
+        if (!write_wav16(out_buf, audio_pcm, info.sample_rate)) {
+            LOG_ERR("mtmd_helper_gen_audio: output too large for WAV\n");
+            return 1;
+        }
+        *out_data     = out_buf.data();
+        *out_data_len = out_buf.size();
+        return 0;
+    }
+
+private:
+    bool ensure_cache() {
+        if (specials_ok) {
+            return true;
+        }
+        codec_0   = find_special_token(vocab, "<|codec_0|>");
+        codec_bos = find_special_token(vocab, "<|codec_bos|>");
+        codec_eos = find_special_token(vocab, "<|codec_eos_token|>");
+        codec_pad = find_special_token(vocab, "<|codec_pad|>");
+        c_think   = find_special_token(vocab, "<|codec_think|>");
+        c_think_b = find_special_token(vocab, "<|codec_think_bos|>");
+        c_think_e = find_special_token(vocab, "<|codec_think_eos|>");
+        tts_pad   = find_special_token(vocab, "<tts_pad>");
+        tts_bos   = find_special_token(vocab, "<tts_text_bos>");
+        tts_eos   = find_special_token(vocab, "<tts_text_eod>");
+        for (llama_token t : { codec_0, codec_bos, codec_eos, codec_pad,
+                               c_think, c_think_b, c_think_e,
+                               tts_pad, tts_bos, tts_eos }) {
+            if (t == LLAMA_TOKEN_NULL) {
+                LOG_ERR("mtmd_helper_gen_audio: missing a required special token in vocab\n");
+                return false;
+            }
+        }
+        const uint32_t n_tok_embd = llama_model_get_tok_embd(model, nullptr);
+        if (n_tok_embd == 0) {
+            LOG_ERR("mtmd_helper_gen_audio: model has no token embeddings\n");
+            return false;
+        }
+        tok_embd.resize(n_tok_embd);
+        if (llama_model_get_tok_embd(model, tok_embd.data()) != n_tok_embd) {
+            LOG_ERR("mtmd_helper_gen_audio: token embedding copy failed\n");
+            return false;
+        }
+        specials_ok = true;
+        return true;
+    }
+
+    // runs the reference wav through the speaker encoder, returns one x-vector embedding row
+    bool encode_speaker(mtmd_bitmap * bitmap, std::vector<float> & out) {
+        if (!mtmd_support_audio(mctx)) {
+            LOG_ERR("mtmd_helper_gen_audio: mmproj has no speaker/audio encoder\n");
+            return false;
+        }
+        const std::string  marker = mtmd_default_marker();
+        mtmd_input_text     text{ marker.c_str(), marker.size(), false, true };
+        mtmd_input_chunks * chunks = mtmd_input_chunks_init();
+        const mtmd_bitmap * bptr = bitmap;
+        bool ok = mtmd_tokenize(mctx, chunks, &text, &bptr, 1) == 0;
+        if (ok) {
+            ok = false;
+            for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
+                const mtmd_input_chunk * chunk = mtmd_input_chunks_get(chunks, i);
+                if (mtmd_input_chunk_get_type(chunk) != MTMD_INPUT_CHUNK_TYPE_AUDIO) {
+                    continue;
+                }
+                if (mtmd_encode_chunk(mctx, chunk) != 0) {
+                    LOG_ERR("mtmd_helper_gen_audio: speaker encode failed\n");
+                    break;
+                }
+                const float * embd = mtmd_get_output_embd(mctx);
+                const size_t  n = (size_t) llama_model_n_embd_inp(model) * mtmd_input_chunk_get_n_tokens(chunk);
+                out.assign(embd, embd + n);
+                ok = true;
+                break;
+            }
+        }
+        mtmd_input_chunks_free(chunks);
+        return ok;
+    }
+
+    // one GEN_WAV process() call over the buffered codes, state is carried across batches
+    bool flush_gen_wav() {
+        if (codes_buf.empty()) {
+            return true;
+        }
+        mtmd_gen_inp inp{};
+        inp.type       = MTMD_GEN_PROCESS_TYPE_GEN_WAV;
+        inp.codes      = codes_buf.data();
+        inp.n_codes    = codes_buf.size();
+        inp.state_data = c2w_state.empty() ? nullptr : (const char *) c2w_state.data();
+        inp.state_size = c2w_state.size();
+        mtmd_gen_out out{};
+        if (mtmd_gen_audio_process(mctx, &inp, &out) != 0) {
+            LOG_ERR("mtmd_helper_gen_audio: gen_wav process failed\n");
+            return false;
+        }
+        audio_pcm.insert(audio_pcm.end(), out.audio, out.audio + out.n_samples);
+        c2w_state.assign(out.state_data, out.state_data + out.state_size);
+        codes_buf.clear();
+        return true;
+    }
+
+    // vocab specials fixed across the whole session, looked up once
+    bool specials_ok = false;
+    llama_token codec_0    = LLAMA_TOKEN_NULL;
+    llama_token codec_bos  = LLAMA_TOKEN_NULL;
+    llama_token codec_eos  = LLAMA_TOKEN_NULL;
+    llama_token codec_pad  = LLAMA_TOKEN_NULL;
+    llama_token c_think    = LLAMA_TOKEN_NULL;
+    llama_token c_think_b  = LLAMA_TOKEN_NULL;
+    llama_token c_think_e  = LLAMA_TOKEN_NULL;
+    llama_token tts_pad    = LLAMA_TOKEN_NULL;
+    llama_token tts_bos    = LLAMA_TOKEN_NULL;
+    llama_token tts_eos    = LLAMA_TOKEN_NULL;
+    std::vector<float> tok_embd; // whole token embedding matrix, n_vocab * n_embd
+
+    // must match hparams.wav_tfm_swa hardcoded in clip.cpp
+    size_t window_frames = 72;
+
+    // per-generation state, cleared by reset()
+    llama_seq_id seq_id = 0;
+    bool mrope = false;
+    int pos = 0;
+    // prompt decode state, consumed batch-by-batch by step_prompt()
+    std::vector<float> prompt_embd_buf;
+    std::unique_ptr<decode_embd_batch> prompt_batch;
+    int n_prompt = 0;
+    int prompt_pos = 0;
+    int32_t top_k = 50;
+    float   top_p = 1.0f;
+    std::vector<int32_t> codes_buf;
+    std::vector<uint8_t> c2w_state;
+    std::vector<float>   audio_pcm;
+    std::vector<std::vector<float>> overlay;
+    size_t overlay_idx = 0;
+    std::vector<float> h_state_buf;
+    mtmd_helper_gen_audio_outtype out_type = MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV;
+    std::vector<char> out_buf;
+};
+
+static std::unique_ptr<mtmd_gen_audio_pipeline> make_pipeline(llama_context * lctx, mtmd_context * mctx) {
+    switch (mtmd_gen_audio_get_info(mctx).type) {
+        case MTMD_GEN_AUDIO_TYPE_QWEN3TTS:
+            return std::unique_ptr<mtmd_gen_audio_pipeline>(new qwen3tts_gen_audio_pipeline(lctx, mctx));
+        default:
+            return nullptr;
+    }
+}
+
+struct mtmd_helper_gen_audio {
+    std::unique_ptr<mtmd_gen_audio_pipeline> pipeline;
+};
+
+mtmd_helper_gen_audio * mtmd_helper_gen_audio_init(struct llama_context * lctx, struct mtmd_context * mctx) {
+    auto * ctx = new mtmd_helper_gen_audio();
+    ctx->pipeline = make_pipeline(lctx, mctx);
+    return ctx;
+}
+
+void mtmd_helper_gen_audio_free(mtmd_helper_gen_audio * ctx) {
+    delete ctx;
+}
+
+void mtmd_helper_gen_audio_reset(mtmd_helper_gen_audio * ctx) {
+    if (ctx->pipeline) {
+        ctx->pipeline->reset();
+    }
+}
+
+int32_t mtmd_helper_gen_audio_set_input(mtmd_helper_gen_audio * ctx, const mtmd_helper_gen_audio_inp * inp) {
+    if (!ctx->pipeline) {
+        LOG_ERR("mtmd_helper_gen_audio: unsupported or missing gen-audio pipeline\n");
+        return 1;
+    }
+    return ctx->pipeline->set_input(inp);
+}
+
+int32_t mtmd_helper_gen_audio_step_prompt(mtmd_helper_gen_audio * ctx, int32_t n_batch) {
+    if (!ctx->pipeline) {
+        return -1;
+    }
+    return ctx->pipeline->step_prompt(n_batch);
+}
+
+int32_t mtmd_helper_gen_audio_step_gen(mtmd_helper_gen_audio * ctx, llama_token sampled,
+                                       const float * h_state_in, const float ** h_state_out) {
+    if (!ctx->pipeline) {
+        return 1;
+    }
+    return ctx->pipeline->step_gen(sampled, h_state_in, h_state_out);
+}
+
+int32_t mtmd_helper_gen_audio_get_output(mtmd_helper_gen_audio * ctx, int32_t * out_sample_rate,
+                                         const char ** out_data, size_t * out_data_len, int64_t * out_n_samples) {
+    if (!ctx->pipeline) {
+        return 1;
+    }
+    return ctx->pipeline->get_output(out_sample_rate, out_data, out_data_len, out_n_samples);
+}
@@ -9,6 +9,7 @@

 #include "mtmd.h"
 #include "mtmd-helper.h"
+#include "mtmd-helper-common.h"
 #include "llama.h"

 #include <algorithm>
@@ -45,50 +46,6 @@
 // internal logging functions
 //

-struct mtmd_helper_logger {
-    ggml_log_callback default_callback = [](ggml_log_level level, const char * text, void * user_data) {
-        (void) level;
-        (void) user_data;
-        fputs(text, stderr);
-        fflush(stderr);
-    };
-
-    ggml_log_callback log_callback = default_callback;
-    void * log_callback_user_data;
-
-    void log_v(enum ggml_log_level level, const char * format, va_list args) {
-        if (format == NULL) {
-            return;
-        }
-        va_list args_copy;
-        va_copy(args_copy, args);
-        char buffer[128];
-        int len = vsnprintf(buffer, 128, format, args);
-        if (len < 128) {
-            log_callback(level, buffer, log_callback_user_data);
-        } else {
-            char * buffer2 = (char *) calloc(len + 1, sizeof(char));
-            vsnprintf(buffer2, len + 1, format, args_copy);
-            buffer2[len] = 0;
-            log_callback(level, buffer2, log_callback_user_data);
-            free(buffer2);
-        }
-        va_end(args_copy);
-    }
-
-    void log(enum ggml_log_level level, const char * format, ...) {
-        va_list args;
-        va_start(args, format);
-        log_v(level, format, args);
-        va_end(args);
-    }
-} g_logger;
-
-#define LOG_DBG(...) g_logger.log(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
-#define LOG_INF(...) g_logger.log(GGML_LOG_LEVEL_INFO,  __VA_ARGS__)
-#define LOG_WRN(...) g_logger.log(GGML_LOG_LEVEL_WARN,  __VA_ARGS__)
-#define LOG_ERR(...) g_logger.log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
-
 void mtmd_helper_log_set(ggml_log_callback log_callback, void * user_data) {
    if (log_callback == nullptr) {
        log_callback = g_logger.default_callback;
@@ -127,117 +84,6 @@ void mtmd_helper_image_get_decoder_pos(const mtmd_image_tokens * chunks, llama_p
    }
 }

-// helper struct to make working with embd batch easier
-// note: this will be removed after llama_batch_ext refactoring
-struct decode_embd_batch {
-    int n_pos_per_embd;
-    int n_mmproj_embd;
-    std::vector<llama_pos>      pos;
-    std::vector<llama_pos>      pos_view; // used by mrope
-    std::vector<int32_t>        n_seq_id;
-    std::vector<llama_seq_id>   seq_id_0;
-    std::vector<llama_seq_id *> seq_ids;
-    std::vector<int8_t>         logits;
-    llama_batch batch;
-    decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
-        GGML_ASSERT(n_tokens > 0 && n_pos_per_embd > 0 && n_mmproj_embd > 0);
-        pos     .resize(n_tokens * n_pos_per_embd);
-        n_seq_id.resize(n_tokens);
-        seq_ids .resize(n_tokens + 1);
-        logits  .resize(n_tokens);
-        seq_id_0.resize(1);
-        seq_ids [n_tokens] = nullptr;
-        batch = {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ embd,
-            /*pos            =*/ pos.data(),
-            /*n_seq_id       =*/ n_seq_id.data(),
-            /*seq_id         =*/ seq_ids.data(),
-            /*logits         =*/ logits.data(),
-        };
-    }
-
-    void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) {
-        seq_id_0[0] = seq_id;
-        for (int i = 0; i < batch.n_tokens; i++) {
-            batch.pos     [i] = pos_0 + i;
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
-        }
-    }
-
-    // M-RoPE for image
-    void set_position_mrope_2d(const std::vector<mtmd_decoder_pos> & rel_pos, llama_seq_id seq_id) {
-        GGML_ASSERT(n_pos_per_embd == 4);
-        GGML_ASSERT(!rel_pos.empty() && (int32_t)rel_pos.size() == batch.n_tokens);
-        seq_id_0[0] = seq_id;
-        for (int32_t i = 0; i < batch.n_tokens; i++) {
-            pos[i                     ] = rel_pos[i].t;
-            pos[i + batch.n_tokens    ] = rel_pos[i].y;
-            pos[i + batch.n_tokens * 2] = rel_pos[i].x;
-            pos[i + batch.n_tokens * 3] = rel_pos[i].z;
-        }
-        for (int i = 0; i < batch.n_tokens; i++) {
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
-        }
-    }
-
-    // M-RoPE for audio
-    void set_position_mrope_1d(llama_pos pos_0, llama_seq_id seq_id) {
-        GGML_ASSERT(n_pos_per_embd == 4);
-        seq_id_0[0] = seq_id;
-        for (int i = 0; i < batch.n_tokens; i++) {
-            pos[i                     ] = pos_0 + i;
-            pos[i + batch.n_tokens    ] = pos_0 + i;
-            pos[i + batch.n_tokens * 2] = pos_0 + i;
-            pos[i + batch.n_tokens * 3] = pos_0 + i;
-        }
-        for (int i = 0; i < batch.n_tokens; i++) {
-            batch.n_seq_id[i] = 1;
-            batch.seq_id  [i] = seq_id_0.data();
-            batch.logits  [i] = false;
-        }
-    }
-
-    llama_batch get_view(int offset, int n_tokens) {
-        GGML_ASSERT(offset >= 0 && n_tokens > 0 && offset + n_tokens <= batch.n_tokens);
-        llama_pos * pos_ptr;
-        pos_view.clear();
-        pos_view.reserve(n_tokens * n_pos_per_embd);
-        if (n_pos_per_embd > 1) {
-            // mrope
-            // for example, with layout of src: 1234...1234...1234...1234...
-            //       offset 2 will give us dst: 34...34...34...34...
-            for (int i = 0; i < n_pos_per_embd; i++) {
-                // assume n_tokens is less than or equal to batch.n_tokens
-                // batch.n_tokens is number of **total** tokens
-                // n_tokens is number of viewed token
-                size_t src_idx = i * batch.n_tokens + offset;
-                pos_view.insert(pos_view.end(),
-                    pos.data() + src_idx,
-                    pos.data() + src_idx + n_tokens);
-            }
-            pos_ptr = pos_view.data();
-        } else {
-            // normal
-            pos_ptr = pos.data() + offset;
-        }
-        return {
-            /*n_tokens       =*/ n_tokens,
-            /*tokens         =*/ nullptr,
-            /*embd           =*/ batch.embd     + offset * n_mmproj_embd,
-            /*pos            =*/ pos_ptr,
-            /*n_seq_id       =*/ batch.n_seq_id + offset,
-            /*seq_id         =*/ batch.seq_id   + offset,
-            /*logits         =*/ batch.logits   + offset,
-        };
-    }
-};
-
 // Helper class to set non-causal attention via RAII
 class scope_non_causal {
 public:
@@ -1084,3 +930,18 @@ int32_t mtmd_helper_video_read_next(mtmd_helper_video * ctx,
    GGML_ASSERT(false && "video is not supported in this build (MTMD_VIDEO is set to OFF)");
 #endif
 }
+
+bool mtmd_helper_model_can_chat(llama_context * lctx, mtmd_context * mctx) {
+    if (!mctx) {
+        return true;
+    }
+
+    auto * model = llama_get_model(lctx);
+    auto * tmpl = llama_model_chat_template(model, nullptr);
+    auto info = mtmd_gen_audio_get_info(mctx);
+
+    // tts-only model cannot be used for chat (no chat template)
+    bool is_tts_only = info.type != MTMD_GEN_AUDIO_TYPE_NONE && tmpl == nullptr;
+
+    return !is_tts_only;
+}
@@ -157,6 +157,73 @@ MTMD_API int32_t mtmd_helper_video_read_next(mtmd_helper_video * ctx,
            mtmd_bitmap ** out_bitmap,
            char ** out_text);

+// return true if model can be used for chat
+MTMD_API bool mtmd_helper_model_can_chat(struct llama_context * lctx, struct mtmd_context * mctx);
+
+//
+// Audio generation helpers
+// (early-stage experimental, subjected to breaking changes)
+//
+
+// audio generation helper context
+// contains accumulator for generated audio features and PCM audio
+struct mtmd_helper_gen_audio;
+typedef struct mtmd_helper_gen_audio mtmd_helper_gen_audio;
+
+enum mtmd_helper_gen_audio_outtype {
+    MTMD_HELPER_GEN_AUDIO_OUTTYPE_PCM, // raw PCM
+    MTMD_HELPER_GEN_AUDIO_OUTTYPE_WAV, // WAV PCM 16-bit LE, mono
+};
+struct mtmd_helper_gen_audio_inp {
+    llama_seq_id seq_id;
+
+    const char * prompt;
+    size_t       prompt_len;
+
+    mtmd_bitmap * speaker_ref; // optional, can be NULL
+    const char * lang; // optional, can be NULL
+
+    int32_t top_k;
+    float   top_p;
+
+    enum mtmd_helper_gen_audio_outtype out_type;
+};
+
+MTMD_API mtmd_helper_gen_audio * mtmd_helper_gen_audio_init(
+                                    struct llama_context * lctx,
+                                    struct mtmd_context * mctx);
+
+MTMD_API void mtmd_helper_gen_audio_free(mtmd_helper_gen_audio * ctx);
+
+MTMD_API void mtmd_helper_gen_audio_reset(mtmd_helper_gen_audio * ctx);
+
+MTMD_API int32_t mtmd_helper_gen_audio_set_input(
+                        mtmd_helper_gen_audio * ctx,
+                        const struct mtmd_helper_gen_audio_inp * inp);
+
+// processes at most n_batch prompt tokens per call
+// returns: >0 = number of prompt tokens remaining, 0 = done, <0 = error
+MTMD_API int32_t mtmd_helper_gen_audio_step_prompt(
+                        mtmd_helper_gen_audio * ctx,
+                        int32_t n_batch);
+
+// generates one frame; must only be called after step_prompt() has returned 0
+// h_state_out is valid until next step_gen() or reset() call
+MTMD_API int32_t mtmd_helper_gen_audio_step_gen(
+                        mtmd_helper_gen_audio * ctx,
+                        llama_token sampled,
+                        const float *  h_state_in,
+                        const float ** h_state_out);
+
+// out_data valid until next get_output() or reset() call
+// out_n_samples (optional, can be NULL) receives the number of generated PCM samples
+MTMD_API int32_t mtmd_helper_gen_audio_get_output(
+                        mtmd_helper_gen_audio * ctx,
+                        int32_t * out_sample_rate,
+                        const char ** out_data,
+                        size_t * out_data_len,
+                        int64_t * out_n_samples);
+
 #ifdef __cplusplus
 } // extern "C"
 #endif
@@ -177,6 +244,31 @@ struct mtmd_helper_video_deleter {
 };
 using video_ptr = std::unique_ptr<mtmd_helper_video, mtmd_helper_video_deleter>;

+// audio generation-related C++ wrappers
+struct mtmd_helper_gen_audio_deleter {
+    void operator()(mtmd_helper_gen_audio * val) { mtmd_helper_gen_audio_free(val); }
+};
+using gen_audio_ptr = std::unique_ptr<mtmd_helper_gen_audio, mtmd_helper_gen_audio_deleter>;
+struct gen_audio {
+    gen_audio_ptr ctx;
+    gen_audio(struct llama_context * lctx, struct mtmd_context * mctx) : ctx(mtmd_helper_gen_audio_init(lctx, mctx)) {}
+    void reset() {
+        mtmd_helper_gen_audio_reset(ctx.get());
+    }
+    int32_t set_input(const struct mtmd_helper_gen_audio_inp * inp) {
+        return mtmd_helper_gen_audio_set_input(ctx.get(), inp);
+    }
+    int32_t step_prompt(int32_t n_batch) {
+        return mtmd_helper_gen_audio_step_prompt(ctx.get(), n_batch);
+    }
+    int32_t step_gen(llama_token sampled, const float * h_state, const float ** h_state_out) {
+        return mtmd_helper_gen_audio_step_gen(ctx.get(), sampled, h_state, h_state_out);
+    }
+    int32_t get_output(int32_t * out_sample_rate, const char ** out_data, size_t * out_data_len, int64_t * out_n_samples = nullptr) {
+        return mtmd_helper_gen_audio_get_output(ctx.get(), out_sample_rate, out_data, out_data_len, out_n_samples);
+    }
+};
+
 } // namespace mtmd_helper
 #endif

@@ -262,6 +262,13 @@ struct mtmd_context {
    struct clip_ctx * ctx_a; // audio
    std::vector<float> out_embd; // image embedding vector

+    // generation context
+    struct clip_ctx * ctx_gen_a; // audio
+    std::vector<int32_t> gen_out_codes; // this frame's 16 sampled codes (GEN_CODE)
+    std::vector<float>   gen_out_embd;  // next-step hidden state fed back to backbone (GEN_CODE)
+    std::vector<float>   gen_out_audio; // decoded PCM samples for the current frame (GEN_WAV)
+    std::vector<uint8_t> gen_out_state; // state to feed into the next GEN_WAV call
+
    bool print_timings;
    int n_threads;
    std::string media_marker;
@@ -354,6 +361,7 @@ struct mtmd_context {
        auto res = clip_init(mmproj_fname, ctx_clip_params);
        ctx_v = res.ctx_v;
        ctx_a = res.ctx_a;
+        ctx_gen_a = res.ctx_gen_a;
        if (!ctx_v && !ctx_a) {
            throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
        }
@@ -378,6 +386,15 @@ struct mtmd_context {
                "hint: you may be using wrong mmproj\n",
                n_embd_text, n_embd_clip));
        }
+        if (ctx_gen_a) {
+            int n_embd_gen = clip_n_mmproj_embd(ctx_gen_a);
+            if (n_embd_text > 0 && n_embd_text != n_embd_gen) {
+                throw std::runtime_error(string_format(
+                    "mismatch between text model (n_embd = %d) and gen-audio mmproj (n_embd = %d)\n"
+                    "hint: you may be using wrong mmproj\n",
+                    n_embd_text, n_embd_gen));
+            }
+        }
        if (ctx_v) {
            init_vision();
        }
@@ -740,6 +757,10 @@ struct mtmd_context {
                    aud_end = "<|mimo_audio_end|>";
                    audio_preproc = std::make_unique<mtmd_audio_preprocessor_mimo_audio>(ctx_a);
                } break;
+            case PROJECTOR_TYPE_QWEN3TTS_SPKENC:
+                {
+                    audio_preproc = std::make_unique<mtmd_audio_preprocessor_qwen3tts_spk>(ctx_a);
+                } break;
            default:
                throw std::runtime_error(string_format("%s: unexpected audio projector type %d\n", __func__, proj));
        }
@@ -780,6 +801,7 @@ struct mtmd_context {
    ~mtmd_context() {
        clip_free(ctx_a);
        clip_free(ctx_v);
+        clip_free(ctx_gen_a);
    }

 private:
@@ -1553,6 +1575,125 @@ float * mtmd_get_output_embd(mtmd_context * ctx) {
    return ctx->out_embd.data();
 }

+//
+// audio generation
+//
+
+mtmd_gen_audio_info mtmd_gen_audio_get_info(const mtmd_context * ctx) {
+    mtmd_gen_audio_info info;
+    if (!ctx->ctx_gen_a) {
+        info.type = MTMD_GEN_AUDIO_TYPE_NONE;
+        return info;
+    }
+    switch (clip_get_projector_type(ctx->ctx_gen_a)) {
+        case PROJECTOR_TYPE_QWEN3TTS_GEN:
+            info.type = MTMD_GEN_AUDIO_TYPE_QWEN3TTS;
+            info.sample_rate = 24000;
+            break;
+        default:
+            info.type = MTMD_GEN_AUDIO_TYPE_NONE;
+            break;
+    }
+    return info;
+}
+
+static int32_t mtmd_gen_audio_process_impl(mtmd_context * ctx, const mtmd_gen_inp * inp, mtmd_gen_out * out) {
+    clip_ctx * ctx_clip = ctx->ctx_gen_a;
+    if (!ctx_clip) {
+        LOG_ERR("%s: model does not support audio generation\n", __func__);
+        return 1;
+    }
+
+    if (inp->type == MTMD_GEN_PROCESS_TYPE_GEN_CODE) {
+        const size_t n_embd = (size_t) clip_n_mmproj_embd(ctx_clip);
+
+        clip_image_f32 hidden_state;
+        hidden_state.set_size({(int) n_embd, 1}, false, true);
+        hidden_state.cpy_buf(std::vector<float>(inp->embd, inp->embd + n_embd));
+
+        clip_image_f32_batch batch;
+        batch.is_audio = true;
+        batch.entries.push_back(std::move(hidden_state));
+
+        std::vector<float>   out_embd(n_embd);
+        std::vector<int32_t> out_codes;
+
+        clip_encode_params params;
+        params.imgs        = &batch;
+        params.n_threads   = ctx->n_threads;
+        params.gen_process = CLIP_GEN_PROCESS_GEN_CODE;
+        params.out_embd    = &out_embd;
+        params.out_codes   = &out_codes;
+        params.code0       = inp->code0;
+        params.top_k       = inp->top_k;
+        params.top_p       = inp->top_p;
+
+        if (!clip_encode(ctx_clip, &params)) {
+            LOG_ERR("%s: clip_encode failed (gen_code)\n", __func__);
+            return 1;
+        }
+
+        ctx->gen_out_embd  = std::move(out_embd);
+        ctx->gen_out_codes = std::move(out_codes);
+
+        out->embd    = ctx->gen_out_embd.data();
+        out->codes   = ctx->gen_out_codes.data();
+        out->n_codes = ctx->gen_out_codes.size();
+        return 0;
+    }
+
+    // MTMD_GEN_PROCESS_TYPE_GEN_WAV
+    if (!inp->codes || inp->n_codes == 0) {
+        LOG_ERR("%s: codes required for gen_wav\n", __func__);
+        return 1;
+    }
+    std::vector<int32_t> in_codes(inp->codes, inp->codes + inp->n_codes);
+    std::vector<uint8_t> in_state;
+    if (inp->state_data) {
+        in_state.assign(inp->state_data, inp->state_data + inp->state_size);
+    }
+
+    // gen_wav has no hidden-state input, the batch entry is an unused placeholder
+    // TODO @ngxson : some models in the future may require hidden-state input, need to update this code later
+    clip_image_f32 dummy;
+    dummy.set_size({1, 1}, false, true);
+    dummy.cpy_buf(std::vector<float>(1, 0.0f));
+
+    clip_image_f32_batch batch;
+    batch.is_audio = true;
+    batch.entries.push_back(std::move(dummy));
+
+    clip_encode_params params;
+    params.imgs        = &batch;
+    params.n_threads   = ctx->n_threads;
+    params.gen_process = CLIP_GEN_PROCESS_GEN_WAV;
+    params.codes       = &in_codes;
+    params.out_audio   = &ctx->gen_out_audio;
+    params.state_in    = inp->state_data ? &in_state : nullptr;
+    params.state_out   = &ctx->gen_out_state;
+
+    if (!clip_encode(ctx_clip, &params)) {
+        LOG_ERR("%s: clip_encode failed (code2wav)\n", __func__);
+        return 1;
+    }
+
+    out->audio      = ctx->gen_out_audio.data();
+    out->n_samples  = ctx->gen_out_audio.size();
+    out->state_data = (const char *) ctx->gen_out_state.data();
+    out->state_size = ctx->gen_out_state.size();
+
+    return 0;
+}
+
+int32_t mtmd_gen_audio_process(mtmd_context * ctx, const struct mtmd_gen_inp * inp, struct mtmd_gen_out * out) {
+    try {
+        return mtmd_gen_audio_process_impl(ctx, inp, out);
+    } catch (const std::exception & e) {
+        LOG_ERR("%s: error: %s\n", __func__, e.what());
+        return 1;
+    }
+}
+
 mtmd_batch * mtmd_batch_init(mtmd_context * ctx) {
    return new mtmd_batch(ctx);
 }
@@ -327,6 +327,60 @@ struct mtmd_caps {
 };
 MTMD_API struct mtmd_caps mtmd_get_cap_from_file(const char * mmproj_fname);

+/////////////////////////////////////////
+// EXPERIMENTAL API for audio generation, subjected to breaking changes
+
+// represent the pipeline type
+enum mtmd_gen_audio_type {
+    MTMD_GEN_AUDIO_TYPE_NONE, // not supported
+    MTMD_GEN_AUDIO_TYPE_QWEN3TTS,
+};
+struct mtmd_gen_audio_info {
+    enum mtmd_gen_audio_type type;
+    int32_t sample_rate; // in Hz, for example 24000 for qwen3tts
+};
+MTMD_API struct mtmd_gen_audio_info mtmd_gen_audio_get_info(const mtmd_context * ctx);
+
+enum mtmd_gen_process_type {
+    MTMD_GEN_PROCESS_TYPE_GEN_CODE, // h_state to semantic (codes, mel-spectrogram, etc.)
+    MTMD_GEN_PROCESS_TYPE_GEN_WAV,  // convert semantic to PCM audio
+                                    // for qwen3tts, this is code2wav
+};
+struct mtmd_gen_inp {
+    enum mtmd_gen_process_type type;
+
+    // for MTMD_GEN_PROCESS_TYPE_GEN_CODE
+    int32_t code0;  // the sampled codebook 0 entry from backbone
+    float * embd;   // the hidden state from backbone, must have n_text_embd elements
+    int32_t top_k;
+    float   top_p;
+
+    // for MTMD_GEN_PROCESS_TYPE_GEN_WAV
+    int32_t * codes;
+    size_t    n_codes;
+    const char * state_data;
+    size_t       state_size;
+};
+struct mtmd_gen_out {
+    // note: output memory is allocated by the context, valid until next process() call
+
+    // for MTMD_GEN_PROCESS_TYPE_GEN_CODE
+    const int32_t * codes;
+    size_t n_codes;
+    const float * embd; // the generated hidden state, to be fed back to backbone
+                        // it must have n_text_embd elements
+
+    // for MTMD_GEN_PROCESS_TYPE_GEN_WAV
+    const float * audio;
+    size_t        n_samples;
+    const char * state_data;
+    size_t       state_size;
+};
+// note: this API is stateless, caller must handle state management and audio frame accumulation
+MTMD_API int32_t mtmd_gen_audio_process(mtmd_context * ctx,
+                                const struct mtmd_gen_inp * inp,
+                                struct mtmd_gen_out * out);
+
 /////////////////////////////////////////

 // test function, to be used in test-mtmd-c-api.c
@@ -215,7 +215,7 @@ def run_mtmd_cli(spec: "ModelSpec", model_path, mmproj_path, image_path, bin_pat
            "--dry-multiplier", "0.8",
            "--dry-base", "1.75",
            "--dry-allowed-length", "2",
-            "--dry-penalty-last-n", "-1",
+            "--dry-penalty-last-n", "64",
            "--dry-sequence-breaker", "none",
        ]
    if spec.n_ctx is not None:
@@ -133,14 +133,14 @@ For the full list of features, please refer to [server's changelog](https://gith
 | `--xtc-probability N` | xtc probability (default: 0.00, 0.0 = disabled) |
 | `--xtc-threshold N` | xtc threshold (default: 0.10, 1.0 = disabled) |
 | `--typical, --typical-p N` | locally typical sampling, parameter p (default: 1.00, 1.0 = disabled) |
-| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
+| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled) |
 | `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.00, 1.0 = disabled) |
 | `--presence-penalty N` | repeat alpha presence penalty (default: 0.00, 0.0 = disabled) |
 | `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.00, 0.0 = disabled) |
 | `--dry-multiplier N` | set DRY sampling multiplier (default: 0.00, 0.0 = disabled) |
 | `--dry-base N` | set DRY sampling base value (default: 1.75) |
 | `--dry-allowed-length N` | set allowed length for DRY sampling (default: 2) |
-| `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
+| `--dry-penalty-last-n N` | set DRY penalty for the last n tokens (default: 64, 0 = disable) |
 | `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers ('\n', ':', '"', '*') in the process; use "none" to not use any sequence breakers |
 | `--adaptive-target N` | adaptive-p: select tokens near this probability (valid range 0.0 to 1.0; negative = disabled) (default: -1.00)<br/>[(more info)](https://github.com/ggml-org/llama.cpp/pull/17927) |
 | `--adaptive-decay N` | adaptive-p: decay rate for target adaptation over time. lower values are more reactive, higher values are more stable.<br/>(valid range 0.0 to 0.99) (default: 0.90) |
@@ -476,7 +476,7 @@ These words will not be included in the completion, so make sure to add them to

 `repeat_penalty`: Control the repetition of token sequences in the generated text. Default: `1.1`

-`repeat_last_n`: Last n tokens to consider for penalizing repetition. Default: `64`, where `0` is disabled and `-1` is ctx-size.
+`repeat_last_n`: Last n tokens to consider for penalizing repetition. Default: `64`, where `0` is disabled.

 `presence_penalty`: Repeat alpha presence penalty. Default: `0.0`, which is disabled.

@@ -488,7 +488,7 @@ These words will not be included in the completion, so make sure to add them to

 `dry_allowed_length`: Tokens that extend repetition beyond this receive exponentially increasing penalty: multiplier * base ^ (length of repeating sequence before token - allowed length). Default: `2`

-`dry_penalty_last_n`: How many tokens to scan for repetitions. Default: `-1`, where `0` is disabled and `-1` is context size.
+`dry_penalty_last_n`: How many tokens to scan for repetitions. Default: `64`, where `0` is disabled.

 `dry_sequence_breakers`: Specify an array of sequence breakers for DRY sampling. Only a JSON array of strings is accepted. Default: `['\n', ':', '"', '*']`

@@ -796,7 +796,7 @@ By default, it is read-only. To make POST request to change global properties, y
      "dry_multiplier": 0.0,
      "dry_base": 1.75,
      "dry_allowed_length": 2,
-      "dry_penalty_last_n": -1,
+      "dry_penalty_last_n": 64,
      "dry_sequence_breakers": [
        "\n",
        ":",
@@ -1076,6 +1076,10 @@ In *router mode* the query param `?model={model_id}` has to be set. This endpoin
 | `llamacpp:n_tokens_max` | Counter | High watermark of the context size observed. |
 | `llamacpp:n_decode_total` | Counter | Total Number of llama_decode() calls. |
 | `llamacpp:n_busy_slots_per_decode` | Gauge | Average number of busy slots per llama_decode() call. |
+| `llamacpp:spec_decode_num_draft_tokens_total` | Counter | Total draft tokens generated (0 when spec-decode is off). |
+| `llamacpp:spec_decode_num_accepted_tokens_total` | Counter | Total draft tokens accepted by the target model (0 when spec-decode is off). |
+| `llamacpp:spec_decode_num_drafts_total` | Counter | Total speculative decoding verification steps (0 when spec-decode is off). |
+| `llamacpp:spec_decode_num_accepted_tokens_per_pos_total` | Counter | Accepted tokens per draft position (labeled `position="N"`; absent when spec-decode is off or before the first completed speculative request). |

 ### POST `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.

@@ -845,6 +845,11 @@ struct server_metrics {
    uint64_t n_decode_total     = 0;
    uint64_t n_busy_slots_total = 0;

+    uint64_t n_draft_tokens_total      = 0;
+    uint64_t n_draft_accepted_total    = 0;
+    uint64_t n_draft_verif_steps_total = 0;
+    std::vector<uint64_t> n_accepted_per_pos_total;
+
    void init() {
        t_start = ggml_time_us();
    }
@@ -863,6 +868,17 @@ struct server_metrics {
        n_tokens_predicted         += slot.n_decoded;
        t_tokens_generation        += slot.t_token_generation;
        t_tokens_generation_total  += slot.t_token_generation;
+
+        n_draft_tokens_total      += slot.n_draft_total;
+        n_draft_accepted_total    += slot.n_draft_accepted;
+        n_draft_verif_steps_total += slot.n_draft_verif_steps;
+
+        if (n_accepted_per_pos_total.size() < slot.n_accepted_per_pos.size()) {
+            n_accepted_per_pos_total.resize(slot.n_accepted_per_pos.size(), 0);
+        }
+        for (size_t i = 0; i < slot.n_accepted_per_pos.size(); i++) {
+            n_accepted_per_pos_total[i] += slot.n_accepted_per_pos[i];
+        }
    }

    void on_decoded(const std::vector<server_slot> & slots) {
@@ -1807,8 +1823,7 @@ private:
        // initialize samplers
        if (task.need_sampling()) {
            try {
-                slot.smpl.reset(common_sampler_init(
-                        model_tgt, task.params.sampling, (int32_t) llama_n_ctx(ctx_tgt)));
+                slot.smpl.reset(common_sampler_init(model_tgt, task.params.sampling));
            } catch (std::exception & e) {
                std::string err_msg = std::string("Failed to initialize samplers: ") + e.what();
                send_error(task, err_msg, ERROR_TYPE_INVALID_REQUEST);
@@ -2553,6 +2568,11 @@ private:
                    res->n_decode_total          = metrics.n_decode_total;
                    res->n_busy_slots_total      = metrics.n_busy_slots_total;

+                    res->n_draft_tokens_total      = metrics.n_draft_tokens_total;
+                    res->n_draft_accepted_total    = metrics.n_draft_accepted_total;
+                    res->n_draft_verif_steps_total = metrics.n_draft_verif_steps_total;
+                    res->n_accepted_per_pos_total  = metrics.n_accepted_per_pos_total;
+
                    if (task.metrics_reset_bucket) {
                        metrics.reset_bucket();
                    }
@@ -4148,7 +4168,6 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
            task.params = server_schema::eval_llama_cmpl_schema(
                    ctx_server.vocab,
                    params,
-                    meta->slot_n_ctx,
                    meta->logit_bias_eog,
                    data);

@@ -4442,6 +4461,18 @@ void server_routes::init_routes() {
                    {"name",  "n_tokens_max"},
                    {"help",  "Largest observed n_tokens."},
                    {"value",  res_task->n_tokens_max}
+            }, {
+                    {"name",  "spec_decode_num_draft_tokens_total"},
+                    {"help",  "Total draft tokens generated"},
+                    {"value",  res_task->n_draft_tokens_total}
+            }, {
+                    {"name",  "spec_decode_num_accepted_tokens_total"},
+                    {"help",  "Total draft tokens accepted by the target model"},
+                    {"value",  res_task->n_draft_accepted_total}
+            }, {
+                    {"name",  "spec_decode_num_drafts_total"},
+                    {"help",  "Total speculative decoding verification steps"},
+                    {"value",  res_task->n_draft_verif_steps_total}
            }}},
            {"gauge", {{
                    {"name",  "prompt_tokens_seconds"},
@@ -4483,6 +4514,17 @@ void server_routes::init_routes() {
            }
        }

+        // labeled counter: one time series per draft position
+        if (!res_task->n_accepted_per_pos_total.empty()) {
+            prometheus << "# HELP llamacpp:spec_decode_num_accepted_tokens_per_pos_total"
+                          " Accepted tokens per draft position\n"
+                       << "# TYPE llamacpp:spec_decode_num_accepted_tokens_per_pos_total counter\n";
+            for (size_t i = 0; i < res_task->n_accepted_per_pos_total.size(); i++) {
+                prometheus << "llamacpp:spec_decode_num_accepted_tokens_per_pos_total{position=\""
+                           << i << "\"} " << res_task->n_accepted_per_pos_total[i] << "\n";
+            }
+        }
+
        res->headers["Process-Start-Time-Unix"] = std::to_string(res_task->t_start);
        res->content_type = "text/plain; version=0.0.4";
        res->status = 200;
@@ -124,8 +124,8 @@ std::vector<std::unique_ptr<field>> make_llama_cmpl_schema(const common_params &
        ->set_desc("Dynamic temperature exponent, controls how entropy maps to temperature"));

    add((new field_num("repeat_last_n", params.sampling.penalty_last_n))
-        ->set_hard_limits(-1, INT32_MAX)
-        ->set_desc("Last n tokens to consider for penalizing repetition (0 = disabled, -1 = ctx-size)"));
+        ->set_hard_limits(0, INT32_MAX)
+        ->set_desc("Last n tokens to consider for penalizing repetition (0 = disabled)"));

    add((new field_num("repeat_penalty", params.sampling.penalty_repeat))
        ->set_desc("Control the repetition of token sequences in the generated text (1.0 = disabled)"));
@@ -151,8 +151,8 @@ std::vector<std::unique_ptr<field>> make_llama_cmpl_schema(const common_params &
        ->set_desc("Tokens that extend repetition beyond this length receive exponentially increasing penalty: multiplier * base ^ (sequence_length - allowed_length)"));

    add((new field_num("dry_penalty_last_n", params.sampling.dry_penalty_last_n))
-        ->set_hard_limits(-1, INT32_MAX)
-        ->set_desc("How many tokens to scan for repetitions (0 = disabled, -1 = context size)"));
+        ->set_hard_limits(0, INT32_MAX)
+        ->set_desc("How many tokens to scan for repetitions (0 = disabled)"));

    add((new field_num("mirostat", params.sampling.mirostat))
        ->set_limits(0, 2)
@@ -515,7 +515,6 @@ std::vector<std::unique_ptr<field>> make_llama_cmpl_schema(const common_params &
 task_params eval_llama_cmpl_schema(
                const llama_vocab * vocab,
                const common_params & params_base,
-                const int n_ctx_slot,
                const std::vector<llama_logit_bias> & logit_bias_eog,
                const json & data) {
    task_params params;
@@ -549,15 +548,6 @@ task_params eval_llama_cmpl_schema(

    // post-processing
    {
-        if (params.sampling.penalty_last_n == -1) {
-            // note: should be the slot's context and not the full context, but it's ok
-            params.sampling.penalty_last_n = n_ctx_slot;
-        }
-
-        if (params.sampling.dry_penalty_last_n == -1) {
-            params.sampling.dry_penalty_last_n = n_ctx_slot;
-        }
-
        // if "reasoning_format" is not provided, its handler will not be called, we will need to handle it here
        auto reasoning_format = params.chat_parser_params.reasoning_format;
        params.chat_parser_params.reasoning_in_content = params.stream && (reasoning_format == COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY);
@@ -98,7 +98,6 @@ std::vector<std::unique_ptr<field>> make_llama_cmpl_schema(
 task_params eval_llama_cmpl_schema(
                    const llama_vocab * vocab,
                    const common_params & params_base,
-                    const int n_ctx_slot,
                    const std::vector<llama_logit_bias> & logit_bias_eog,
                    const json & data);

@@ -1560,6 +1560,11 @@ json server_task_result_metrics::to_json() {
        { "n_decode_total",                  n_decode_total },
        { "n_busy_slots_total",              n_busy_slots_total },

+        { "n_draft_tokens_total",            n_draft_tokens_total },
+        { "n_draft_accepted_total",          n_draft_accepted_total },
+        { "n_draft_verif_steps_total",       n_draft_verif_steps_total },
+        { "n_accepted_per_pos_total",        n_accepted_per_pos_total },
+
        { "slots",                           slots_data },
    };
 }
@@ -532,6 +532,11 @@ struct server_task_result_metrics : server_task_result {
    uint64_t n_decode_total     = 0;
    uint64_t n_busy_slots_total = 0;

+    uint64_t n_draft_tokens_total      = 0;
+    uint64_t n_draft_accepted_total    = 0;
+    uint64_t n_draft_verif_steps_total = 0;
+    std::vector<uint64_t> n_accepted_per_pos_total;
+
    // while we can also use std::vector<server_slot> this requires copying the slot object which can be quite messy
    // therefore, we use json to temporarily store the slot.to_json() result
    json slots_data = json::array();
@@ -10,17 +10,67 @@
 #include <ctime>
 #include <atomic>
 #include <cstring>
+#include <cstdlib>
 #include <algorithm>
 #include <unordered_set>
+#include <tuple>
 #include <functional>
 #include <memory>

+#if defined(_WIN32)
+#   ifndef NOMINMAX
+#       define NOMINMAX
+#   endif
+#   include <windows.h>
+#endif
+
 namespace fs = std::filesystem;

 //
 // internal helpers
 //

+#if defined(_WIN32)
+// A chunk can end in the middle of a multi-byte sequence, so the incomplete
+// tail is dropped before validating what precedes it.
+static bool is_utf8_text(const std::string & text) {
+    return is_valid_utf8(text.substr(0, validate_utf8(text)));
+}
+
+// A child process writes its output in the OEM code page, which is not UTF-8
+// on a western Windows install, so accented text reaches the JSON layer as
+// invalid bytes and is replaced there. Text that already decodes as UTF-8 is
+// returned untouched, so a child that emits UTF-8 is never decoded twice.
+// run() spawns without a console, so the console code page does not apply.
+static std::string console_output_to_utf8(const std::string & text) {
+    if (text.empty() || is_utf8_text(text)) {
+        return text;
+    }
+
+    const UINT cp = GetOEMCP();
+
+    // fail rather than emit replacement characters when the code page is wrong
+    const int wide_len = MultiByteToWideChar(cp, MB_ERR_INVALID_CHARS, text.data(), (int) text.size(), nullptr, 0);
+    if (wide_len <= 0) {
+        return text;
+    }
+    std::wstring wide(wide_len, L'\0');
+    MultiByteToWideChar(cp, MB_ERR_INVALID_CHARS, text.data(), (int) text.size(), wide.data(), wide_len);
+
+    const int utf8_len = WideCharToMultiByte(CP_UTF8, 0, wide.data(), wide_len, nullptr, 0, nullptr, nullptr);
+    if (utf8_len <= 0) {
+        return text;
+    }
+    std::string utf8(utf8_len, '\0');
+    WideCharToMultiByte(CP_UTF8, 0, wide.data(), wide_len, utf8.data(), utf8_len, nullptr, nullptr);
+    return utf8;
+}
+#else
+static std::string console_output_to_utf8(const std::string & text) {
+    return text;
+}
+#endif
+
 json server_tool::to_json() const {
    return {
        {"display_name", display_name},
@@ -34,7 +84,40 @@ json server_tool::to_json() const {
 }

 static constexpr size_t SERVER_TOOL_GIT_LS_FILES_MAX_OUTPUT = 8 * 1024 * 1024; // 8 MB
-static constexpr int SERVER_TOOL_GIT_LS_FILES_TIMEOUT = 15; // seconds
+// budget for one listing call, shared by the git and walker paths
+static constexpr int SERVER_TOOL_LIST_ENTRIES_TIMEOUT = 15; // seconds
+
+// entry kinds a directory listing may return
+enum class list_kind {
+    files, // regular files only
+    dirs,  // directories only
+    all,   // both
+};
+
+// home directory, read once at first use (getenv is not thread safe against setenv)
+static const std::string & home_dir() {
+    static const std::string home = [] {
+        const char * h = getenv("HOME");
+#ifdef _WIN32
+        if (h == nullptr) h = getenv("USERPROFILE");
+#endif
+        return h ? std::string(h) : std::string();
+    }();
+    return home;
+}
+
+static std::string expand_home(const std::string & path) {
+    if (path.empty() || path[0] != '~') return path;
+    if (path.size() > 1 && path[1] != '/' && path[1] != '\\') return path;
+    const std::string & home = home_dir();
+    if (home.empty()) return path;
+    return home + path.substr(1);
+}
+
+// depth of a '/'-separated relative path: "a/b/c" is 3
+static int entry_depth(const std::string & rel) {
+    return 1 + (int) std::count(rel.begin(), rel.end(), '/');
+}

 class tools_io {
 public:
@@ -51,8 +134,17 @@ public:
    virtual bool file_size(const std::string & path, uintmax_t & out_size) const = 0;
    virtual bool read_file(const std::string & path, std::string & out) const = 0;
    virtual bool write_file(const std::string & path, const std::string & content) const = 0;
-    // paths relative to `base`, '/'-separated; sets `err` if `base` isn't a directory
-    virtual std::vector<std::string> list_files(const std::string & base, std::string & err) const = 0;
+    // resolve `path` against the IO's working directory; absolute paths are returned unchanged
+    virtual std::string resolve(const std::string & path) const = 0;
+    struct list_entry {
+        std::string rel; // '/'-separated, relative to `base`
+        bool is_dir = false;
+    };
+    // entries relative to `base`; sets `err` if `base` isn't a directory
+    // max_depth == 0 means unlimited, 1 means direct children of `base` only
+    // `base` must already be resolved (absolute); `caller_path` is the path the
+    // caller passed, used only for error messages
+    virtual std::vector<list_entry> list_entries(const std::string & base, const std::string & caller_path, int max_depth, list_kind kind, std::string & err, bool & truncated) const = 0;
    // on_chunk, if set, is called with each chunk of output as it is read (before truncation cuts in);
    // returning false terminates the process early (e.g. the client disconnected)
    virtual exec_result run(
@@ -67,6 +159,22 @@ public:
    // cwd, if non-empty, is used to resolve relative paths and as the working directory for run()
    explicit tools_io_basic(std::string cwd = "") : cwd(std::move(cwd)) {}

+    // expands a leading `~`, then resolves `path` against `cwd` (or the server
+    // working directory when `cwd` is unset); the result is always absolute
+    std::string resolve(const std::string & path) const override {
+        std::string p = expand_home(path);
+        if (fs::path(p).is_absolute()) {
+            return p;
+        }
+        if (cwd.empty()) {
+            std::error_code ec;
+            fs::path cur = fs::current_path(ec);
+            if (ec) return p;
+            return (cur / p).string();
+        }
+        return (fs::path(cwd) / p).string();
+    }
+
    bool is_directory(const std::string & path) const override {
        std::error_code ec;
        return fs::is_directory(resolve(path), ec) && !ec;
@@ -105,34 +213,41 @@ public:
        return (bool) f;
    }

-    std::vector<std::string> list_files(const std::string & base, std::string & err) const override {
+    std::vector<list_entry> list_entries(const std::string & base, const std::string & caller_path, int max_depth, list_kind kind, std::string & err, bool & truncated) const override {
        err.clear();
-        std::string abs_base = resolve(base);
-        if (!is_directory(base)) {
-            err = "path does not exist or is not a directory: " + base;
+        truncated = false;
+        std::error_code ec;
+        if (!fs::is_directory(base, ec) || ec) {
+            err = "path does not exist or is not a directory: " + caller_path;
            return {};
        }

-        auto res = run(
-            {"git", "-C", abs_base, "ls-files", "--cached", "--others", "--exclude-standard"},
-            SERVER_TOOL_GIT_LS_FILES_MAX_OUTPUT, SERVER_TOOL_GIT_LS_FILES_TIMEOUT);
+        const auto deadline = std::chrono::steady_clock::now() + std::chrono::seconds(SERVER_TOOL_LIST_ENTRIES_TIMEOUT);

-        if (res.exit_code == 0 && !res.timed_out) {
-            std::vector<std::string> result;
-            std::istringstream iss(res.output);
-            std::string line;
-            while (std::getline(iss, line)) {
-                if (!line.empty() && line.back() == '\r') line.pop_back();
-                if (line.empty()) continue;
-                std::replace(line.begin(), line.end(), '\\', '/');
-                if (is_regular_file((fs::path(base) / line).string())) {
-                    result.push_back(line);
+        // git ls-files cannot list directories; use the walker when they are requested
+        if (kind == list_kind::files) {
+            auto res = run(
+                {"git", "-C", base, "ls-files", "--cached", "--others", "--exclude-standard"},
+                SERVER_TOOL_GIT_LS_FILES_MAX_OUTPUT, SERVER_TOOL_LIST_ENTRIES_TIMEOUT);
+
+            if (res.exit_code == 0 && !res.timed_out) {
+                std::vector<list_entry> result;
+                std::istringstream iss(res.output);
+                std::string line;
+                while (std::getline(iss, line)) {
+                    if (!line.empty() && line.back() == '\r') line.pop_back();
+                    if (line.empty()) continue;
+                    std::replace(line.begin(), line.end(), '\\', '/');
+                    if (max_depth > 0 && entry_depth(line) > max_depth) continue;
+                    if (is_regular_file((fs::path(base) / line).string())) {
+                        result.push_back({line, false});
+                    }
                }
+                return result;
            }
-            return result;
        }

-        return list_files_fallback(abs_base);
+        return list_entries_fallback(base, max_depth, kind, deadline, truncated);
    }

    exec_result run(
@@ -179,14 +294,14 @@ public:
                    size_t len = strlen(buf);
                    if (output.size() + len <= max_output) {
                        output.append(buf, len);
-                        if (on_chunk && !on_chunk(std::string(buf, len))) {
+                        if (on_chunk && !on_chunk(console_output_to_utf8(std::string(buf, len)))) {
                            proc.terminate();
                            break;
                        }
                    } else {
                        size_t remaining = max_output - output.size();
                        output.append(buf, remaining);
-                        if (on_chunk && remaining > 0) on_chunk(std::string(buf, remaining));
+                        if (on_chunk && remaining > 0) on_chunk(console_output_to_utf8(std::string(buf, remaining)));
                        truncated = true;
                    }
                }
@@ -200,7 +315,7 @@ public:

        res.exit_code = proc.join();

-        res.output    = output;
+        res.output    = console_output_to_utf8(output);
        res.timed_out = timed_out.load();
        if (truncated) {
            res.output += "\n[output truncated]";
@@ -211,14 +326,6 @@ public:
 private:
    std::string cwd;

-    // resolves `path` against `cwd` if `path` is relative and `cwd` is set; otherwise returns `path` unchanged
-    std::string resolve(const std::string & path) const {
-        if (cwd.empty() || fs::path(path).is_absolute()) {
-            return path;
-        }
-        return (fs::path(cwd) / path).string();
-    }
-
    static const std::unordered_set<std::string> & junk_dir_names() {
        static const std::unordered_set<std::string> names = {
            ".git", ".svn", ".hg", "node_modules", "__pycache__",
@@ -227,28 +334,50 @@ private:
        return names;
    }

-    std::vector<std::string> list_files_fallback(const std::string & base) const {
-        std::vector<std::string> result;
+    std::vector<list_entry> list_entries_fallback(const std::string & base, int max_depth, list_kind kind,
+                                                  std::chrono::steady_clock::time_point deadline, bool & truncated) const {
+        std::vector<list_entry> result;
        std::error_code ec;

-        std::vector<std::pair<fs::path, fs::path>> stack;
-        stack.emplace_back(fs::path(base), fs::path());
+        std::vector<std::tuple<fs::path, fs::path, int>> stack;
+        stack.emplace_back(fs::path(base), fs::path(), 0);

        while (!stack.empty()) {
-            auto [dir, rel_dir] = stack.back();
+            auto [dir, rel_dir, depth] = stack.back();
            stack.pop_back();

-            for (const auto & entry : fs::directory_iterator(dir, fs::directory_options::skip_permission_denied, ec)) {
+            // the throwing increment would escape the tool on a directory that
+            // goes away mid walk, so step the iterator explicitly
+            fs::directory_iterator it(dir, fs::directory_options::skip_permission_denied, ec);
+            for (const fs::directory_iterator end; it != end; it.increment(ec)) {
                if (ec) break;
+                if (std::chrono::steady_clock::now() >= deadline) {
+                    truncated = true;
+                    return result;
+                }
+                const fs::directory_entry & entry = *it;
                std::string fname = entry.path().filename().string();
                std::error_code tec;
                if (entry.is_directory(tec)) {
+                    std::string rel = (rel_dir / fname).string();
+                    std::replace(rel.begin(), rel.end(), '\\', '/');
+                    if (kind == list_kind::dirs || kind == list_kind::all) {
+                        result.push_back({rel, true});
+                    }
+                    // junk directories stay selectable but are never walked: they
+                    // hold nothing worth searching and can be enormous
                    if (junk_dir_names().count(fname) > 0) continue;
-                    stack.emplace_back(entry.path(), rel_dir / fname);
+                    // do not descend into symlinks: a link can point back to an
+                    // ancestor and loop forever
+                    if (!entry.is_symlink(tec) && (max_depth == 0 || depth + 1 < max_depth)) {
+                        stack.emplace_back(entry.path(), rel_dir / fname, depth + 1);
+                    }
                } else if (entry.is_regular_file(tec)) {
                    std::string rel = (rel_dir / fname).string();
                    std::replace(rel.begin(), rel.end(), '\\', '/');
-                    result.push_back(rel);
+                    if (kind == list_kind::files || kind == list_kind::all) {
+                        result.push_back({rel, false});
+                    }
                }
            }
        }
@@ -363,6 +492,9 @@ struct server_tool_read_file : server_tool {
 //

 static constexpr size_t SERVER_TOOL_FILE_SEARCH_MAX_RESULTS = 100;
+static constexpr const char * SERVER_TOOL_FILE_SEARCH_TYPE_FILE = "file";
+static constexpr const char * SERVER_TOOL_FILE_SEARCH_TYPE_DIR  = "dir";
+static constexpr const char * SERVER_TOOL_FILE_SEARCH_TYPE_ALL  = "all";

 struct server_tool_file_glob_search : server_tool {
    server_tool_file_glob_search() {
@@ -382,13 +514,18 @@ struct server_tool_file_glob_search : server_tool {
                    "and common junk directories (.git, node_modules, build, dist, etc.) otherwise. "
                    "A pattern with no '/' (e.g. \"*.cpp\") matches the file's basename at any depth. "
                    "A pattern containing '/' matches the full relative path; unless already anchored with "
-                    "\"**/\" or a leading '/', it is automatically prefixed with \"**/\"."},
+                    "\"**/\" or a leading '/', it is automatically prefixed with \"**/\". "
+                    "Use type=\"dir\" or \"all\" to also list directories; directory entries are suffixed with '/' in the output. "
+                    "Note: directory listings do not apply .gitignore filtering."},
                {"parameters", {
                    {"type", "object"},
                    {"properties", {
-                        {"path",    {{"type", "string"}, {"description", "Base directory to search in"}}},
-                        {"include", {{"type", "string"}, {"description", "Glob pattern for files to include (e.g. \"*.cpp\" or \"src/**/*.cpp\"). Default: **"}}},
-                        {"exclude", {{"type", "string"}, {"description", "Glob pattern for files to exclude"}}},
+                        {"path",      {{"type", "string"},  {"description", "Base directory to search in"}}},
+                        {"include",   {{"type", "string"},  {"description", "Glob pattern for files to include (e.g. \"*.cpp\" or \"src/**/*.cpp\"). Default: **"}}},
+                        {"exclude",   {{"type", "string"},  {"description", "Glob pattern for files to exclude"}}},
+                        {"type",      {{"type", "string"},  {"description", "Entry type to return: \"file\" (default), \"dir\" or \"all\""}}},
+                        {"max_depth", {{"type", "integer"}, {"description", "Maximum depth to descend into subdirectories (default: 0 = unlimited; 1 = direct children only)"}}},
+                        {"limit",     {{"type", "integer"}, {"description", string_format("Maximum number of results to return (default %zu; values below 1 fall back to the default)", SERVER_TOOL_FILE_SEARCH_MAX_RESULTS)}}},
                    }},
                    {"required", json::array({"path"})},
                }},
@@ -397,30 +534,56 @@ struct server_tool_file_glob_search : server_tool {
    }

    json invoke(json params, server_tool::stream *) const override {
-        std::string base    = params.at("path").get<std::string>();
-        std::string include = json_value(params, "include", std::string("**"));
-        std::string exclude = json_value(params, "exclude", std::string(""));
-
        auto io = make_tools_io(params);
+
+        std::string base      = io->resolve(params.at("path").get<std::string>());
+        // normalize to forward slashes so the web UI (which assumes '/') can
+        // join the relative entries into absolute paths on Windows too
+        std::replace(base.begin(), base.end(), '\\', '/');
+        std::string include   = json_value(params, "include", std::string("**"));
+        std::string exclude   = json_value(params, "exclude", std::string(""));
+        std::string type      = json_value(params, "type",    std::string("file"));
+        int         max_depth = std::max(0, json_value(params, "max_depth", 0));
+        int         limit     = json_value(params, "limit", (int) SERVER_TOOL_FILE_SEARCH_MAX_RESULTS);
+        if (limit < 1) limit = SERVER_TOOL_FILE_SEARCH_MAX_RESULTS;
+        limit = std::min(limit, (int) SERVER_TOOL_FILE_SEARCH_MAX_RESULTS);
+
+        list_kind kind;
+        if (type == SERVER_TOOL_FILE_SEARCH_TYPE_FILE) {
+            kind = list_kind::files;
+        } else if (type == SERVER_TOOL_FILE_SEARCH_TYPE_DIR) {
+            kind = list_kind::dirs;
+        } else if (type == SERVER_TOOL_FILE_SEARCH_TYPE_ALL) {
+            kind = list_kind::all;
+        } else {
+            return {{"error", "invalid type: " + type + " (expected \"file\", \"dir\" or \"all\")"}};
+        }
+
        std::string err;
-        auto files = io->list_files(base, err);
+        bool truncated = false;
+        auto entries = io->list_entries(base, params.at("path").get<std::string>(), max_depth, kind, err, truncated);
        if (!err.empty()) {
            return {{"error", err}};
        }

-        std::vector<std::string> matches;
-        for (const auto & rel : files) {
-            if (!path_glob_match(include, rel)) continue;
-            if (!exclude.empty() && path_glob_match(exclude, rel)) continue;
-            matches.push_back(rel);
+        std::vector<tools_io::list_entry> matches;
+        for (const auto & entry : entries) {
+            if (!path_glob_match(include, entry.rel)) continue;
+            if (!exclude.empty() && path_glob_match(exclude, entry.rel)) continue;
+            matches.push_back(entry);
        }

        size_t total = matches.size();
-        size_t shown = std::min(total, SERVER_TOOL_FILE_SEARCH_MAX_RESULTS);
+        size_t shown = std::min(total, (size_t) limit);

        std::ostringstream output_text;
+        json entries_json = json::array();
        for (size_t i = 0; i < shown; i++) {
-            output_text << matches[i] << "\n";
+            output_text << matches[i].rel << (matches[i].is_dir ? "/" : "") << "\n";
+            entries_json.push_back({
+                {"path", matches[i].rel},
+                {"type", matches[i].is_dir ? "dir" : "file"},
+            });
        }

        output_text << "\n---\nTotal matches: " << total << "\n";
@@ -429,8 +592,16 @@ struct server_tool_file_glob_search : server_tool {
                "[%zu results limit reached (%zu total matches). Refine the glob pattern to narrow the search.]\n",
                shown, total);
        }
+        if (truncated) {
+            output_text << "[search timed out, results truncated]\n";
+        }

-        return {{"plain_text_response", output_text.str()}};
+        // `base` is always absolute (resolve falls back to the server cwd), so
+        // API clients (e.g. the web UI picker) can join the relative entries
+        // into absolute paths. `plain_text_response` is what the model sees;
+        // `entries` is the same data as structured JSON for the UI picker,
+        // which reads `entries`/`base` instead of re-parsing the text.
+        return {{"plain_text_response", output_text.str()}, {"entries", entries_json}, {"base", base}};
    }
 };

@@ -513,18 +684,20 @@ struct server_tool_grep_search : server_tool {
        // collect (absolute_path, display_path) pairs to search
        std::vector<std::pair<std::string, std::string>> files;

-        if (io->is_regular_file(path)) {
-            files.emplace_back(path, path);
-        } else if (io->is_directory(path)) {
+        const std::string abs_path = io->resolve(path);
+        if (io->is_regular_file(abs_path)) {
+            files.emplace_back(abs_path, path);
+        } else if (io->is_directory(abs_path)) {
            std::string err;
-            auto candidates = io->list_files(path, err);
+            bool truncated = false;
+            auto candidates = io->list_entries(abs_path, path, 0, list_kind::files, err, truncated);
            if (!err.empty()) {
                return {{"error", err}};
            }
-            for (const auto & rel : candidates) {
-                if (!path_glob_match(include, rel)) continue;
-                if (!exclude.empty() && path_glob_match(exclude, rel)) continue;
-                files.emplace_back((fs::path(path) / rel).string(), rel);
+            for (const auto & entry : candidates) {
+                if (!path_glob_match(include, entry.rel)) continue;
+                if (!exclude.empty() && path_glob_match(exclude, entry.rel)) continue;
+                files.emplace_back((fs::path(abs_path) / entry.rel).string(), entry.rel);
            }
        } else {
            return {{"error", "path does not exist: " + path}};
@@ -1094,6 +1267,9 @@ struct server_tool_get_datetime : server_tool {
 // get_info: returns runtime info (OS name/version and cwd)
 //

+static constexpr size_t SERVER_TOOL_GET_INFO_MAX_OUTPUT = 4096;
+static constexpr int    SERVER_TOOL_GET_INFO_TIMEOUT    = 5; // seconds
+
 struct server_tool_get_info : server_tool {
    server_tool_get_info() {
        name = "get_info";
@@ -1119,9 +1295,9 @@ struct server_tool_get_info : server_tool {
        auto io = make_tools_io(params);

 #ifdef _WIN32
-        auto res = io->run({"cmd", "/c", "ver"}, 4096, 5);
+        auto res = io->run({"cmd", "/c", "ver"}, SERVER_TOOL_GET_INFO_MAX_OUTPUT, SERVER_TOOL_GET_INFO_TIMEOUT);
 #else
-        auto res = io->run({"uname", "-a"}, 4096, 5);
+        auto res = io->run({"uname", "-a"}, SERVER_TOOL_GET_INFO_MAX_OUTPUT, SERVER_TOOL_GET_INFO_TIMEOUT);
 #endif
        // "ver" prints a blank line before the version, so the output is stripped on both ends;
        // a failed spawn or a timeout leaves a diagnostic in res.output, which is not an OS name
@@ -164,3 +164,101 @@ def test_tools_builtin_edit_file_rejects_overlapping_edits():
    finally:
        if os.path.exists(log_path):
            os.remove(log_path)
+
+
+def test_tools_builtin_file_glob_search_type_dir(tmp_path):
+    global server
+    server.start()
+
+    (tmp_path / "project-alpha" / "src").mkdir(parents=True)
+    (tmp_path / "project-alpha" / "README.md").write_text("alpha")
+    (tmp_path / "project-alpha" / "src" / "main.cpp").write_text("int main() {}")
+    (tmp_path / "project-beta").mkdir()
+    (tmp_path / "project-beta" / "notes.txt").write_text("beta")
+
+    res = call_tool("file_glob_search", {"path": str(tmp_path), "type": "dir"})
+    text = res["plain_text_response"]
+    assert "project-alpha/" in text
+    assert "project-beta/" in text
+    assert "project-alpha/src/" in text
+    assert "README.md" not in text
+    types = {e["path"]: e["type"] for e in res["entries"]}
+    assert types["project-alpha"] == "dir"
+    assert types["project-alpha/src"] == "dir"
+
+    res_all = call_tool("file_glob_search", {"path": str(tmp_path), "type": "all", "include": "*proj*"})
+    paths = [e["path"] for e in res_all["entries"]]
+    assert "project-alpha" in paths
+    assert "project-beta" in paths
+
+
+def test_tools_builtin_file_glob_search_max_depth_and_limit(tmp_path):
+    global server
+    server.start()
+
+    (tmp_path / "a" / "b" / "c").mkdir(parents=True)
+    (tmp_path / "top.txt").write_text("top")
+    (tmp_path / "a" / "mid.txt").write_text("mid")
+    (tmp_path / "a" / "b" / "deep.txt").write_text("deep")
+
+    res = call_tool("file_glob_search", {"path": str(tmp_path), "max_depth": 1})
+    assert "top.txt" in res["plain_text_response"]
+    assert "mid.txt" not in res["plain_text_response"]
+
+    res = call_tool("file_glob_search", {"path": str(tmp_path), "max_depth": 2})
+    assert "mid.txt" in res["plain_text_response"]
+    assert "deep.txt" not in res["plain_text_response"]
+
+    res = call_tool("file_glob_search", {"path": str(tmp_path), "limit": 1})
+    assert len(res["entries"]) == 1
+    assert "Total matches: 3" in res["plain_text_response"]
+
+
+def test_tools_builtin_file_glob_search_rejects_invalid_type(tmp_path):
+    global server
+    server.start()
+
+    err = call_tool_expect_error("file_glob_search", {"path": str(tmp_path), "type": "bogus"})
+    assert "invalid type" in err
+
+
+def test_tools_builtin_cwd_header_overrides_model_param(tmp_path):
+    global server
+    server.start()
+
+    workdir = tmp_path / "workdir"
+    workdir.mkdir()
+    (workdir / "marker.txt").write_text("marker")
+
+    # a model-provided "cwd" in the params is overridden by the x-tool-cwd header
+    res = call_tool("read_file", {"path": "marker.txt", "cwd": "/definitely/not/a/real/path"},
+                    headers={"x-tool-cwd": str(workdir)})
+    assert "marker" in res["plain_text_response"]
+
+
+def test_tools_builtin_cwd_relative_paths(tmp_path):
+    global server
+    server.start()
+
+    workdir = tmp_path / "workdir"
+    workdir.mkdir()
+    (workdir / "rel.txt").write_text("relative-content")
+
+    headers = {"x-tool-cwd": str(workdir)}
+
+    # relative paths in file tools resolve against the header cwd
+    res = call_tool("read_file", {"path": "rel.txt"}, headers=headers)
+    assert "relative-content" in res["plain_text_response"]
+
+    res = call_tool("write_file", {"path": "sub/out.txt", "content": "written"}, headers=headers)
+    assert (workdir / "sub" / "out.txt").read_text() == "written"
+
+    res = call_tool("file_glob_search", {"path": ".", "include": "*.txt"}, headers=headers)
+    assert "rel.txt" in res["plain_text_response"]
+
+    # absolute paths are unaffected by the cwd
+    other = tmp_path / "other"
+    other.mkdir()
+    (other / "abs.txt").write_text("absolute-content")
+    res = call_tool("read_file", {"path": str(other / "abs.txt")}, headers=headers)
+    assert "absolute-content" in res["plain_text_response"]
@@ -1,6 +1,6 @@
 set(TARGET llama-tts)
 add_executable(${TARGET} tts.cpp)
-target_link_libraries(${TARGET} PRIVATE llama llama-common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama llama-common mtmd ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

 if(LLAMA_TOOLS_INSTALL)
@@ -1,117 +1,34 @@
-# llama.cpp/example/tts
-This example demonstrates the Text To Speech feature. It uses a
-[model](https://www.outeai.com/blog/outetts-0.2-500m) from
-[outeai](https://www.outeai.com/).
+# llama.cpp TTS

-## Quickstart
-If you have built llama.cpp with SSL support you can simply run the
-following command and the required models will be downloaded automatically:
-```console
-$ build/bin/llama-tts --tts-oute-default -p "Hello world" && aplay output.wav
-```
-For details about the models and how to convert them to the required format
-see the following sections.
+This is a tool to demonstrate audio generation capability in llama.cpp via `libmtmd`. It was added via PR [#26254](https://github.com/ggml-org/llama.cpp/pull/26254)

-### Model conversion
-Checkout or download the model that contains the LLM model:
-```console
-$ pushd models
-$ git clone --branch main --single-branch --depth 1 https://huggingface.co/OuteAI/OuteTTS-0.2-500M
-$ cd OuteTTS-0.2-500M && git lfs install && git lfs pull
-$ popd
-```
-Convert the model to .gguf format:
-```console
-(venv) python convert_hf_to_gguf.py models/OuteTTS-0.2-500M \
-    --outfile models/outetts-0.2-0.5B-f16.gguf --outtype f16
-```
-The generated model will be `models/outetts-0.2-0.5B-f16.gguf`.
+Note: this tool used to serve as a demo for OuteTTS, but it was converted to a more model-agnostic tool.

-We can optionally quantize this to Q8_0 using the following command:
-```console
-$ build/bin/llama-quantize models/outetts-0.2-0.5B-f16.gguf \
-    models/outetts-0.2-0.5B-q8_0.gguf q8_0
-```
-The quantized model will be `models/outetts-0.2-0.5B-q8_0.gguf`.
+## Common usage

-Next we do something similar for the audio decoder. First download or checkout
-the model for the voice decoder:
-```console
-$ pushd models
-$ git clone --branch main --single-branch --depth 1 https://huggingface.co/novateur/WavTokenizer-large-speech-75token
-$ cd WavTokenizer-large-speech-75token && git lfs install && git lfs pull
-$ popd
-```
-This model file is a PyTorch checkpoint (.ckpt) and we first need to convert it to
-huggingface format:
-```console
-(venv) python tools/tts/convert_pt_to_hf.py \
-    models/WavTokenizer-large-speech-75token/wavtokenizer_large_speech_320_24k.ckpt
-...
-Model has been successfully converted and saved to models/WavTokenizer-large-speech-75token/model.safetensors
-Metadata has been saved to models/WavTokenizer-large-speech-75token/index.json
-Config has been saved to models/WavTokenizer-large-speech-75tokenconfig.json
-```
-Then we can convert the huggingface format to gguf:
-```console
-(venv) python convert_hf_to_gguf.py models/WavTokenizer-large-speech-75token \
-    --outfile models/wavtokenizer-large-75-f16.gguf --outtype f16
-...
-INFO:hf-to-gguf:Model successfully exported to models/wavtokenizer-large-75-f16.gguf
+Simple usage:
+
+```sh
+llama-tts -hf ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF -p "Hello world" --output out.wav
 ```

-### Running the example
+Common params:
+- Sampling params such as `--top-k`, `--top-p`, `--temp`, etc.
+- `-n <number_of_frames>` limits the output length, e.g. `-n 500`. Note that how many milliseconds each frame represents varies by model
+- Core inference params such as `-ngl`, `-b`, `-ub`, etc.

-With both of the models generated, the LLM model and the voice decoder model,
-we can run the example:
-```console
-$ build/bin/llama-tts -m  ./models/outetts-0.2-0.5B-q8_0.gguf \
-    -mv ./models/wavtokenizer-large-75-f16.gguf \
-    -p "Hello world"
-...
-main: audio written to file 'output.wav'
-```
-The output.wav file will contain the audio of the prompt. This can be heard
-by playing the file with a media player. On Linux the following command will
-play the audio:
-```console
-$ aplay output.wav
-```
+## Qwen3-TTS

-### Running the example with llama-server
-Running this example with `llama-server` is also possible and requires two
-server instances to be started. One will serve the LLM model and the other
-will serve the voice decoder model.
+Available params:
+- `--tts-lang` can be `zh`, `en`, `de`, `it`, `pt`, `es`, `ja`, `ko`, `fr`, `ru` (default: `en`)
+- `--tts-speaker-file` should point to a speaker reference audio file (wav, mp3)

-The LLM model server can be started with the following command:
-```console
-$ ./build/bin/llama-server -m ./models/outetts-0.2-0.5B-q8_0.gguf --port 8020
-```
+Example usage:

-And the voice decoder model server can be started using:
-```console
-./build/bin/llama-server -m ./models/wavtokenizer-large-75-f16.gguf --port 8021 --embeddings --pooling none
-```
-
-Then we can run [tts-outetts.py](tts-outetts.py) to generate the audio.
-
-First create a virtual environment for python and install the required
-dependencies (this in only required to be done once):
-```console
-$ python3 -m venv venv
-$ source venv/bin/activate
-(venv) pip install requests numpy
-```
-
-And then run the python script using:
-```conole
-(venv) python ./tools/tts/tts-outetts.py http://localhost:8020 http://localhost:8021 "Hello world"
-spectrogram generated: n_codes: 90, n_embd: 1282
-converting to audio ...
-audio generated: 28800 samples
-audio written to file "output.wav"
-```
-And to play the audio we can again use aplay or any other media player:
-```console
-$ aplay output.wav
+```sh
+llama-tts -hf ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF \
+    -p "Hello world" \
+    --tts-lang english \
+    --tts-speaker-file speaker.mp3 \
+    --output out.wav
 ```
@@ -1,180 +0,0 @@
-# convert the https://huggingface.co/novateur/WavTokenizer-large-speech-75token to HF format
-# the goal is to be able to reuse the convert_hf_to_gguf.py after that to create a GGUF file with the WavTokenizer decoder
-#
-# TODO: this script is LLM-generated and probably very inefficient and should be rewritten
-
-import torch
-import json
-import os
-import sys
-import re
-
-from safetensors.torch import save_file
-
-# default
-model_path = './model.pt'
-
-# read from CLI
-if len(sys.argv) > 1:
-    model_path = sys.argv[1]
-
-# get the directory of the input model
-path_dst = os.path.dirname(model_path)
-
-print(f"Loading model from {model_path}")
-
-model = torch.load(model_path, map_location='cpu')
-
-#print(model)
-
-# print all keys
-for key in model.keys():
-    print(key)
-    if key == 'hyper_parameters':
-        #print(model[key])
-        # dump as json pretty
-        print(json.dumps(model[key], indent=4))
-    #if key != 'state_dict' and key != 'optimizer_states':
-    #    print(model[key])
-
-# Check if the loaded model is a state_dict or a model instance
-if isinstance(model, torch.nn.Module):
-    state_dict = model.state_dict()
-else:
-    state_dict = model
-
-# Print the structure of the state_dict to understand its format
-print("State dictionary keys:")
-for key in state_dict.keys():
-    print(key)
-
-# Ensure the state_dict is flat and contains only torch.Tensor objects
-def flatten_state_dict(state_dict, parent_key='', sep='.'):
-    items = []
-    items_new = []
-
-    for k, v in state_dict.items():
-        new_key = f"{parent_key}{sep}{k}" if parent_key else k
-        if isinstance(v, torch.Tensor):
-            items.append((new_key, v))
-        elif isinstance(v, dict):
-            items.extend(flatten_state_dict(v, new_key, sep=sep).items())
-            return dict(items)
-
-    size_total_mb = 0
-
-    for key, value in list(items):
-        # keep only what we need for inference
-        if not key.startswith('state_dict.feature_extractor.encodec.quantizer.') and \
-           not key.startswith('state_dict.backbone.') and \
-           not key.startswith('state_dict.head.out'):
-               print('Skipping key: ', key)
-               continue
-
-        new_key = key
-
-        new_key = new_key.replace('state_dict.', '')
-        new_key = new_key.replace('pos_net', 'posnet')
-
-        # check if matches "backbone.posnet.%d.bias" or "backbone.posnet.%d.weight"
-        if new_key.startswith("backbone.posnet."):
-            match = re.match(r"backbone\.posnet\.(\d+)\.(bias|weight)", new_key)
-            if match:
-               new_key = f"backbone.posnet.{match.group(1)}.norm.{match.group(2)}"
-
-        # "feature_extractor.encodec.quantizer.vq.layers.0._codebook.embed" -> "backbone.embedding.weight"
-        if new_key == "feature_extractor.encodec.quantizer.vq.layers.0._codebook.embed":
-            new_key = "backbone.embedding.weight"
-
-        # these are the only rows used
-        # ref: https://github.com/edwko/OuteTTS/blob/a613e79c489d8256dd657ea9168d78de75895d82/outetts/wav_tokenizer/audio_codec.py#L100
-        if new_key.endswith("norm.scale.weight"):
-            new_key = new_key.replace("norm.scale.weight", "norm.weight")
-            value = value[0]
-
-        if new_key.endswith("norm.shift.weight"):
-            new_key = new_key.replace("norm.shift.weight", "norm.bias")
-            value = value[0]
-
-        if new_key.endswith("gamma"):
-            new_key = new_key.replace("gamma", "gamma.weight")
-
-        # convert from 1D [768] to 2D [768, 1] so that ggml_add can broadcast the bias
-        if (new_key.endswith("norm.weight") or new_key.endswith("norm1.weight") or new_key.endswith("norm2.weight") or new_key.endswith(".bias")) and (new_key.startswith("backbone.posnet") or new_key.startswith("backbone.embed.bias")):
-            value = value.unsqueeze(1)
-
-        if new_key.endswith("dwconv.bias"):
-            value = value.unsqueeze(1)
-
-        size_mb = value.element_size() * value.nelement() / (1024 * 1024)
-        print(f"{size_mb:8.2f} MB - {new_key}: {value.shape}")
-
-        size_total_mb += size_mb
-
-        #print(key, '->', new_key, ': ', value)
-        #print(key, '->', new_key)
-
-        items_new.append((new_key, value))
-
-    print(f"Total size: {size_total_mb:8.2f} MB")
-
-    return dict(items_new)
-
-flattened_state_dict = flatten_state_dict(state_dict)
-
-
-# Convert the model to the safetensors format
-output_path = path_dst + '/model.safetensors'
-save_file(flattened_state_dict, output_path)
-
-print(f"Model has been successfully converted and saved to {output_path}")
-
-# Calculate the total size of the .safetensors file
-total_size = os.path.getsize(output_path)
-
-# Create the weight map
-weight_map = {
-    "model.safetensors": ["*"]  # Assuming all weights are in one file
-}
-
-# Create metadata for the index.json file
-metadata = {
-    "total_size": total_size,
-    "weight_map": weight_map
-}
-
-# Save the metadata to index.json
-index_path = path_dst + '/index.json'
-with open(index_path, 'w') as f:
-    json.dump(metadata, f, indent=4)
-
-print(f"Metadata has been saved to {index_path}")
-
-config = {
-    "architectures": [
-        "WavTokenizerDec"
-    ],
-    "hidden_size": 1282,
-    "n_embd_features": 512,
-    "n_ff": 2304,
-    "vocab_size": 4096,
-    "n_head": 1,
-    "layer_norm_epsilon": 1e-6,
-    "group_norm_epsilon": 1e-6,
-    "group_norm_groups": 32,
-    "max_position_embeddings": 8192, # ?
-    "n_layer": 12,
-    "posnet": {
-        "n_embd": 768,
-        "n_layer": 6
-    },
-    "convnext": {
-        "n_embd": 768,
-        "n_layer": 12
-    },
-}
-
-with open(path_dst + '/config.json', 'w') as f:
-    json.dump(config, f, indent=4)
-
-print(f"Config has been saved to {path_dst + 'config.json'}")
@@ -1,299 +0,0 @@
-import sys
-#import json
-#import struct
-import requests
-import re
-import struct
-import numpy as np
-from concurrent.futures import ThreadPoolExecutor
-
-
-def fill_hann_window(size, periodic=True):
-    if periodic:
-        return np.hanning(size + 1)[:-1]
-    return np.hanning(size)
-
-
-def irfft(n_fft, complex_input):
-    return np.fft.irfft(complex_input, n=n_fft)
-
-
-def fold(buffer, n_out, n_win, n_hop, n_pad):
-    result = np.zeros(n_out)
-    n_frames = len(buffer) // n_win
-
-    for i in range(n_frames):
-        start = i * n_hop
-        end = start + n_win
-        result[start:end] += buffer[i * n_win:(i + 1) * n_win]
-
-    return result[n_pad:-n_pad] if n_pad > 0 else result
-
-
-def process_frame(args):
-    l, n_fft, ST, hann = args
-    frame = irfft(n_fft, ST[l])
-    frame = frame * hann
-    hann2 = hann * hann
-    return frame, hann2
-
-
-def embd_to_audio(embd, n_codes, n_embd, n_thread=4):
-    embd = np.asarray(embd, dtype=np.float32).reshape(n_codes, n_embd)
-
-    n_fft = 1280
-    n_hop = 320
-    n_win = 1280
-    n_pad = (n_win - n_hop) // 2
-    n_out = (n_codes - 1) * n_hop + n_win
-
-    hann = fill_hann_window(n_fft, True)
-
-    E = np.zeros((n_embd, n_codes), dtype=np.float32)
-    for l in range(n_codes):
-        for k in range(n_embd):
-            E[k, l] = embd[l, k]
-
-    half_embd = n_embd // 2
-    S = np.zeros((n_codes, half_embd + 1), dtype=np.complex64)
-
-    for k in range(half_embd):
-        for l in range(n_codes):
-            mag = E[k, l]
-            phi = E[k + half_embd, l]
-
-            mag = np.clip(np.exp(mag), 0, 1e2)
-            S[l, k] = mag * np.exp(1j * phi)
-
-    res = np.zeros(n_codes * n_fft)
-    hann2_buffer = np.zeros(n_codes * n_fft)
-
-    with ThreadPoolExecutor(max_workers=n_thread) as executor:
-        args = [(l, n_fft, S, hann) for l in range(n_codes)]
-        results = list(executor.map(process_frame, args))
-
-        for l, (frame, hann2) in enumerate(results):
-            res[l*n_fft:(l+1)*n_fft] = frame
-            hann2_buffer[l*n_fft:(l+1)*n_fft] = hann2
-
-    audio = fold(res, n_out, n_win, n_hop, n_pad)
-    env = fold(hann2_buffer, n_out, n_win, n_hop, n_pad)
-
-    mask = env > 1e-10
-    audio[mask] /= env[mask]
-
-    return audio
-
-
-def save_wav(filename, audio_data, sample_rate):
-    num_channels = 1
-    bits_per_sample = 16
-    bytes_per_sample = bits_per_sample // 8
-    data_size = len(audio_data) * bytes_per_sample
-    byte_rate = sample_rate * num_channels * bytes_per_sample
-    block_align = num_channels * bytes_per_sample
-    chunk_size = 36 + data_size  # 36 = size of header minus first 8 bytes
-
-    header = struct.pack(
-        '<4sI4s4sIHHIIHH4sI',
-        b'RIFF',
-        chunk_size,
-        b'WAVE',
-        b'fmt ',
-        16,                # fmt chunk size
-        1,                 # audio format (PCM)
-        num_channels,
-        sample_rate,
-        byte_rate,
-        block_align,
-        bits_per_sample,
-        b'data',
-        data_size
-    )
-
-    audio_data = np.clip(audio_data * 32767, -32768, 32767)
-    pcm_data = audio_data.astype(np.int16)
-
-    with open(filename, 'wb') as f:
-        f.write(header)
-        f.write(pcm_data.tobytes())
-
-
-def process_text(text: str):
-    text = re.sub(r'\d+(\.\d+)?', lambda x: x.group(), text.lower()) # TODO this needs to be fixed
-    text = re.sub(r'[-_/,\.\\]', ' ', text)
-    text = re.sub(r'[^a-z\s]', '', text)
-    text = re.sub(r'\s+', ' ', text).strip()
-    return text.split()
-
-# usage:
-# python tts-outetts.py http://server-llm:port http://server-dec:port "text"
-
-if len(sys.argv) <= 3:
-    print("usage: python tts-outetts.py http://server-llm:port http://server-dec:port \"text\"")
-    exit(1)
-
-host_llm = sys.argv[1]
-host_dec = sys.argv[2]
-text = sys.argv[3]
-
-prefix = """<|im_start|>
-<|text_start|>the<|text_sep|>overall<|text_sep|>package<|text_sep|>from<|text_sep|>just<|text_sep|>two<|text_sep|>people<|text_sep|>is<|text_sep|>pretty<|text_sep|>remarkable<|text_sep|>sure<|text_sep|>i<|text_sep|>have<|text_sep|>some<|text_sep|>critiques<|text_sep|>about<|text_sep|>some<|text_sep|>of<|text_sep|>the<|text_sep|>gameplay<|text_sep|>aspects<|text_sep|>but<|text_sep|>its<|text_sep|>still<|text_sep|>really<|text_sep|>enjoyable<|text_sep|>and<|text_sep|>it<|text_sep|>looks<|text_sep|>lovely<|text_sep|>"""
-
-words = process_text(text)
-words = "<|text_sep|>".join([i.strip() for i in words])
-words += "<|text_end|>\n"
-
-# voice data
-# TODO: load from json
-#suffix = """<|audio_start|>
-#the<|t_0.08|><|code_start|><|257|><|740|><|636|><|913|><|788|><|1703|><|code_end|>
-#overall<|t_0.36|><|code_start|><|127|><|201|><|191|><|774|><|700|><|532|><|1056|><|557|><|798|><|298|><|1741|><|747|><|1662|><|1617|><|1702|><|1527|><|368|><|1588|><|1049|><|1008|><|1625|><|747|><|1576|><|728|><|1019|><|1696|><|1765|><|code_end|>
-#package<|t_0.56|><|code_start|><|935|><|584|><|1319|><|627|><|1016|><|1491|><|1344|><|1117|><|1526|><|1040|><|239|><|1435|><|951|><|498|><|723|><|1180|><|535|><|789|><|1649|><|1637|><|78|><|465|><|1668|><|901|><|595|><|1675|><|117|><|1009|><|1667|><|320|><|840|><|79|><|507|><|1762|><|1508|><|1228|><|1768|><|802|><|1450|><|1457|><|232|><|639|><|code_end|>
-#from<|t_0.19|><|code_start|><|604|><|782|><|1682|><|872|><|1532|><|1600|><|1036|><|1761|><|647|><|1554|><|1371|><|653|><|1595|><|950|><|code_end|>
-#just<|t_0.25|><|code_start|><|1782|><|1670|><|317|><|786|><|1748|><|631|><|599|><|1155|><|1364|><|1524|><|36|><|1591|><|889|><|1535|><|541|><|440|><|1532|><|50|><|870|><|code_end|>
-#two<|t_0.24|><|code_start|><|1681|><|1510|><|673|><|799|><|805|><|1342|><|330|><|519|><|62|><|640|><|1138|><|565|><|1552|><|1497|><|1552|><|572|><|1715|><|1732|><|code_end|>
-#people<|t_0.39|><|code_start|><|593|><|274|><|136|><|740|><|691|><|633|><|1484|><|1061|><|1138|><|1485|><|344|><|428|><|397|><|1562|><|645|><|917|><|1035|><|1449|><|1669|><|487|><|442|><|1484|><|1329|><|1832|><|1704|><|600|><|761|><|653|><|269|><|code_end|>
-#is<|t_0.16|><|code_start|><|566|><|583|><|1755|><|646|><|1337|><|709|><|802|><|1008|><|485|><|1583|><|652|><|10|><|code_end|>
-#pretty<|t_0.32|><|code_start|><|1818|><|1747|><|692|><|733|><|1010|><|534|><|406|><|1697|><|1053|><|1521|><|1355|><|1274|><|816|><|1398|><|211|><|1218|><|817|><|1472|><|1703|><|686|><|13|><|822|><|445|><|1068|><|code_end|>
-#remarkable<|t_0.68|><|code_start|><|230|><|1048|><|1705|><|355|><|706|><|1149|><|1535|><|1787|><|1356|><|1396|><|835|><|1583|><|486|><|1249|><|286|><|937|><|1076|><|1150|><|614|><|42|><|1058|><|705|><|681|><|798|><|934|><|490|><|514|><|1399|><|572|><|1446|><|1703|><|1346|><|1040|><|1426|><|1304|><|664|><|171|><|1530|><|625|><|64|><|1708|><|1830|><|1030|><|443|><|1509|><|1063|><|1605|><|1785|><|721|><|1440|><|923|><|code_end|>
-#sure<|t_0.36|><|code_start|><|792|><|1780|><|923|><|1640|><|265|><|261|><|1525|><|567|><|1491|><|1250|><|1730|><|362|><|919|><|1766|><|543|><|1|><|333|><|113|><|970|><|252|><|1606|><|133|><|302|><|1810|><|1046|><|1190|><|1675|><|code_end|>
-#i<|t_0.08|><|code_start|><|123|><|439|><|1074|><|705|><|1799|><|637|><|code_end|>
-#have<|t_0.16|><|code_start|><|1509|><|599|><|518|><|1170|><|552|><|1029|><|1267|><|864|><|419|><|143|><|1061|><|0|><|code_end|>
-#some<|t_0.16|><|code_start|><|619|><|400|><|1270|><|62|><|1370|><|1832|><|917|><|1661|><|167|><|269|><|1366|><|1508|><|code_end|>
-#critiques<|t_0.60|><|code_start|><|559|><|584|><|1163|><|1129|><|1313|><|1728|><|721|><|1146|><|1093|><|577|><|928|><|27|><|630|><|1080|><|1346|><|1337|><|320|><|1382|><|1175|><|1682|><|1556|><|990|><|1683|><|860|><|1721|><|110|><|786|><|376|><|1085|><|756|><|1523|><|234|><|1334|><|1506|><|1578|><|659|><|612|><|1108|><|1466|><|1647|><|308|><|1470|><|746|><|556|><|1061|><|code_end|>
-#about<|t_0.29|><|code_start|><|26|><|1649|><|545|><|1367|><|1263|><|1728|><|450|><|859|><|1434|><|497|><|1220|><|1285|><|179|><|755|><|1154|><|779|><|179|><|1229|><|1213|><|922|><|1774|><|1408|><|code_end|>
-#some<|t_0.23|><|code_start|><|986|><|28|><|1649|><|778|><|858|><|1519|><|1|><|18|><|26|><|1042|><|1174|><|1309|><|1499|><|1712|><|1692|><|1516|><|1574|><|code_end|>
-#of<|t_0.07|><|code_start|><|197|><|716|><|1039|><|1662|><|64|><|code_end|>
-#the<|t_0.08|><|code_start|><|1811|><|1568|><|569|><|886|><|1025|><|1374|><|code_end|>
-#gameplay<|t_0.48|><|code_start|><|1269|><|1092|><|933|><|1362|><|1762|><|1700|><|1675|><|215|><|781|><|1086|><|461|><|838|><|1022|><|759|><|649|><|1416|><|1004|><|551|><|909|><|787|><|343|><|830|><|1391|><|1040|><|1622|><|1779|><|1360|><|1231|><|1187|><|1317|><|76|><|997|><|989|><|978|><|737|><|189|><|code_end|>
-#aspects<|t_0.56|><|code_start|><|1423|><|797|><|1316|><|1222|><|147|><|719|><|1347|><|386|><|1390|><|1558|><|154|><|440|><|634|><|592|><|1097|><|1718|><|712|><|763|><|1118|><|1721|><|1311|><|868|><|580|><|362|><|1435|><|868|><|247|><|221|><|886|><|1145|><|1274|><|1284|><|457|><|1043|><|1459|><|1818|><|62|><|599|><|1035|><|62|><|1649|><|778|><|code_end|>
-#but<|t_0.20|><|code_start|><|780|><|1825|><|1681|><|1007|><|861|><|710|><|702|><|939|><|1669|><|1491|><|613|><|1739|><|823|><|1469|><|648|><|code_end|>
-#its<|t_0.09|><|code_start|><|92|><|688|><|1623|><|962|><|1670|><|527|><|599|><|code_end|>
-#still<|t_0.27|><|code_start|><|636|><|10|><|1217|><|344|><|713|><|957|><|823|><|154|><|1649|><|1286|><|508|><|214|><|1760|><|1250|><|456|><|1352|><|1368|><|921|><|615|><|5|><|code_end|>
-#really<|t_0.36|><|code_start|><|55|><|420|><|1008|><|1659|><|27|><|644|><|1266|><|617|><|761|><|1712|><|109|><|1465|><|1587|><|503|><|1541|><|619|><|197|><|1019|><|817|><|269|><|377|><|362|><|1381|><|507|><|1488|><|4|><|1695|><|code_end|>
-#enjoyable<|t_0.49|><|code_start|><|678|><|501|><|864|><|319|><|288|><|1472|><|1341|><|686|><|562|><|1463|><|619|><|1563|><|471|><|911|><|730|><|1811|><|1006|><|520|><|861|><|1274|><|125|><|1431|><|638|><|621|><|153|><|876|><|1770|><|437|><|987|><|1653|><|1109|><|898|><|1285|><|80|><|593|><|1709|><|843|><|code_end|>
-#and<|t_0.15|><|code_start|><|1285|><|987|><|303|><|1037|><|730|><|1164|><|502|><|120|><|1737|><|1655|><|1318|><|code_end|>
-#it<|t_0.09|><|code_start|><|848|><|1366|><|395|><|1601|><|1513|><|593|><|1302|><|code_end|>
-#looks<|t_0.27|><|code_start|><|1281|><|1266|><|1755|><|572|><|248|><|1751|><|1257|><|695|><|1380|><|457|><|659|><|585|><|1315|><|1105|><|1776|><|736|><|24|><|736|><|654|><|1027|><|code_end|>
-#lovely<|t_0.56|><|code_start|><|634|><|596|><|1766|><|1556|><|1306|><|1285|><|1481|><|1721|><|1123|><|438|><|1246|><|1251|><|795|><|659|><|1381|><|1658|><|217|><|1772|><|562|><|952|><|107|><|1129|><|1112|><|467|><|550|><|1079|><|840|><|1615|><|1469|><|1380|><|168|><|917|><|836|><|1827|><|437|><|583|><|67|><|595|><|1087|><|1646|><|1493|><|1677|><|code_end|>"""
-
-# TODO: tokenization is slow for some reason - here is pre-tokenized input
-suffix = [ 151667, 198, 1782, 155780, 151669, 151929, 152412, 152308, 152585, 152460, 153375, 151670, 198, 74455,
-          155808, 151669, 151799, 151873, 151863, 152446, 152372, 152204, 152728, 152229, 152470, 151970, 153413,
-          152419, 153334, 153289, 153374, 153199, 152040, 153260, 152721, 152680, 153297, 152419, 153248, 152400,
-          152691, 153368, 153437, 151670, 198, 1722, 155828, 151669, 152607, 152256, 152991, 152299, 152688, 153163,
-          153016, 152789, 153198, 152712, 151911, 153107, 152623, 152170, 152395, 152852, 152207, 152461, 153321,
-          153309, 151750, 152137, 153340, 152573, 152267, 153347, 151789, 152681, 153339, 151992, 152512, 151751,
-          152179, 153434, 153180, 152900, 153440, 152474, 153122, 153129, 151904, 152311, 151670, 198, 1499, 155791,
-          151669, 152276, 152454, 153354, 152544, 153204, 153272, 152708, 153433, 152319, 153226, 153043, 152325,
-          153267, 152622, 151670, 198, 4250, 155797, 151669, 153454, 153342, 151989, 152458, 153420, 152303, 152271,
-          152827, 153036, 153196, 151708, 153263, 152561, 153207, 152213, 152112, 153204, 151722, 152542, 151670, 198,
-          19789, 155796, 151669, 153353, 153182, 152345, 152471, 152477, 153014, 152002, 152191, 151734, 152312, 152810,
-          152237, 153224, 153169, 153224, 152244, 153387, 153404, 151670, 198, 16069, 155811, 151669, 152265, 151946,
-          151808, 152412, 152363, 152305, 153156, 152733, 152810, 153157, 152016, 152100, 152069, 153234, 152317,
-          152589, 152707, 153121, 153341, 152159, 152114, 153156, 153001, 153504, 153376, 152272, 152433, 152325,
-          151941, 151670, 198, 285, 155788, 151669, 152238, 152255, 153427, 152318, 153009, 152381, 152474, 152680,
-          152157, 153255, 152324, 151682, 151670, 198, 32955, 155804, 151669, 153490, 153419, 152364, 152405, 152682,
-          152206, 152078, 153369, 152725, 153193, 153027, 152946, 152488, 153070, 151883, 152890, 152489, 153144,
-          153375, 152358, 151685, 152494, 152117, 152740, 151670, 198, 37448, 480, 155840, 151669, 151902, 152720,
-          153377, 152027, 152378, 152821, 153207, 153459, 153028, 153068, 152507, 153255, 152158, 152921, 151958,
-          152609, 152748, 152822, 152286, 151714, 152730, 152377, 152353, 152470, 152606, 152162, 152186, 153071,
-          152244, 153118, 153375, 153018, 152712, 153098, 152976, 152336, 151843, 153202, 152297, 151736, 153380,
-          153502, 152702, 152115, 153181, 152735, 153277, 153457, 152393, 153112, 152595, 151670, 198, 19098, 155808,
-          151669, 152464, 153452, 152595, 153312, 151937, 151933, 153197, 152239, 153163, 152922, 153402, 152034,
-          152591, 153438, 152215, 151673, 152005, 151785, 152642, 151924, 153278, 151805, 151974, 153482, 152718,
-          152862, 153347, 151670, 198, 72, 155780, 151669, 151795, 152111, 152746, 152377, 153471, 152309, 151670, 198,
-          19016, 155788, 151669, 153181, 152271, 152190, 152842, 152224, 152701, 152939, 152536, 152091, 151815, 152733,
-          151672, 151670, 198, 14689, 155788, 151669, 152291, 152072, 152942, 151734, 153042, 153504, 152589, 153333,
-          151839, 151941, 153038, 153180, 151670, 198, 36996, 8303, 155832, 151669, 152231, 152256, 152835, 152801,
-          152985, 153400, 152393, 152818, 152765, 152249, 152600, 151699, 152302, 152752, 153018, 153009, 151992,
-          153054, 152847, 153354, 153228, 152662, 153355, 152532, 153393, 151782, 152458, 152048, 152757, 152428,
-          153195, 151906, 153006, 153178, 153250, 152331, 152284, 152780, 153138, 153319, 151980, 153142, 152418,
-          152228, 152733, 151670, 198, 9096, 155801, 151669, 151698, 153321, 152217, 153039, 152935, 153400, 152122,
-          152531, 153106, 152169, 152892, 152957, 151851, 152427, 152826, 152451, 151851, 152901, 152885, 152594,
-          153446, 153080, 151670, 198, 14689, 155795, 151669, 152658, 151700, 153321, 152450, 152530, 153191, 151673,
-          151690, 151698, 152714, 152846, 152981, 153171, 153384, 153364, 153188, 153246, 151670, 198, 1055, 155779,
-          151669, 151869, 152388, 152711, 153334, 151736, 151670, 198, 1782, 155780, 151669, 153483, 153240, 152241,
-          152558, 152697, 153046, 151670, 198, 5804, 1363, 155820, 151669, 152941, 152764, 152605, 153034, 153434,
-          153372, 153347, 151887, 152453, 152758, 152133, 152510, 152694, 152431, 152321, 153088, 152676, 152223,
-          152581, 152459, 152015, 152502, 153063, 152712, 153294, 153451, 153032, 152903, 152859, 152989, 151748,
-          152669, 152661, 152650, 152409, 151861, 151670, 198, 300, 7973, 155828, 151669, 153095, 152469, 152988,
-          152894, 151819, 152391, 153019, 152058, 153062, 153230, 151826, 152112, 152306, 152264, 152769, 153390,
-          152384, 152435, 152790, 153393, 152983, 152540, 152252, 152034, 153107, 152540, 151919, 151893, 152558,
-          152817, 152946, 152956, 152129, 152715, 153131, 153490, 151734, 152271, 152707, 151734, 153321, 152450,
-          151670, 198, 8088, 155792, 151669, 152452, 153497, 153353, 152679, 152533, 152382, 152374, 152611, 153341,
-          153163, 152285, 153411, 152495, 153141, 152320, 151670, 198, 1199, 155781, 151669, 151764, 152360, 153295,
-          152634, 153342, 152199, 152271, 151670, 198, 43366, 155799, 151669, 152308, 151682, 152889, 152016, 152385,
-          152629, 152495, 151826, 153321, 152958, 152180, 151886, 153432, 152922, 152128, 153024, 153040, 152593,
-          152287, 151677, 151670, 198, 53660, 155808, 151669, 151727, 152092, 152680, 153331, 151699, 152316, 152938,
-          152289, 152433, 153384, 151781, 153137, 153259, 152175, 153213, 152291, 151869, 152691, 152489, 151941,
-          152049, 152034, 153053, 152179, 153160, 151676, 153367, 151670, 198, 268, 4123, 480, 155821, 151669, 152350,
-          152173, 152536, 151991, 151960, 153144, 153013, 152358, 152234, 153135, 152291, 153235, 152143, 152583,
-          152402, 153483, 152678, 152192, 152533, 152946, 151797, 153103, 152310, 152293, 151825, 152548, 153442,
-          152109, 152659, 153325, 152781, 152570, 152957, 151752, 152265, 153381, 152515, 151670, 198, 437, 155787,
-          151669, 152957, 152659, 151975, 152709, 152402, 152836, 152174, 151792, 153409, 153327, 152990, 151670, 198,
-          275, 155781, 151669, 152520, 153038, 152067, 153273, 153185, 152265, 152974, 151670, 198, 94273, 155799,
-          151669, 152953, 152938, 153427, 152244, 151920, 153423, 152929, 152367, 153052, 152129, 152331, 152257,
-          152987, 152777, 153448, 152408, 151696, 152408, 152326, 152699, 151670, 198, 385, 16239, 155828, 151669,
-          152306, 152268, 153438, 153228, 152978, 152957, 153153, 153393, 152795, 152110, 152918, 152923, 152467,
-          152331, 153053, 153330, 151889, 153444, 152234, 152624, 151779, 152801, 152784, 152139, 152222, 152751,
-          152512, 153287, 153141, 153052, 151840, 152589, 152508, 153499, 152109, 152255, 151739, 152267, 152759,
-          153318, 153165, 153349, 151670, ]
-
-response = requests.post(
-    host_llm + "/completion",
-    json={
-        "prompt": [prefix + words, *suffix],
-        "n_predict": 1024,
-        "cache_prompt": True,
-        "return_tokens": True,
-        "samplers": ["top_k"],
-        "top_k": 16,
-        "seed": 1003,
-    }
-)
-
-response_json = response.json()
-
-#print(json.dumps(response_json, indent=4))
-#print(json.dumps(response_json["prompt"], indent=4).replace("\\n", "\n"))
-#print(json.dumps(response_json["timings"], indent=4))
-#print(json.dumps(response_json["tokens"], indent=4))
-
-codes = response_json["tokens"]
-
-codes = [t - 151672 for t in codes if t >= 151672 and t <= 155772]
-
-response = requests.post(
-    host_dec + "/embeddings",
-    json={
-        "input": [*codes],
-    }
-)
-
-response_json = response.json()
-
-#print(json.dumps(response_json, indent=4))
-
-# spectrogram
-embd = response_json[0]["embedding"]
-
-n_codes = len(embd)
-n_embd = len(embd[0])
-
-print('spectrogram generated: n_codes: %d, n_embd: %d' % (n_codes, n_embd))
-
-# post-process the spectrogram to convert to audio
-print('converting to audio ...')
-audio = embd_to_audio(embd, n_codes, n_embd)
-print('audio generated: %d samples' % len(audio))
-
-filename = "output.wav"
-sample_rate = 24000 # sampling rate
-
-# zero out first 0.25 seconds
-audio[:24000 // 4] = 0.0
-
-save_wav(filename, audio, sample_rate)
-print('audio written to file "%s"' % filename)
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Niklas WenzelandGitHub	360e1349f0	tests: re-enable MiniMax M3 in `test-llama-archs` (#26633 )	2026-08-05 17:58:34 +02:00
Saba FallahandGitHub	b06aa774c0	mtmd: Unlimited-OCR fix max_tiles, setting in converter (#25614 )	2026-08-05 15:30:14 +02:00
Aldehir RojasandGitHub	cd0fa6051a	grammar : degrade max repetition >= 2000 to unbounded (#26613 )	2026-08-05 07:39:10 -05:00
Xuan-Son Nguyen GitHub Saba Fallah	717dad5c8e	mtmd: support multi-row batching for deepseek-ocr (#26154 ) * mtmd: support multi-row batching for deepseek-ocr * mtmd: weave deepseek-ocr rows in one shot instead of per row (#26615) --------- Co-authored-by: Saba Fallah <sabafallah@gmail.com>	2026-08-05 13:34:52 +02:00
Sergey MalininandGitHub	9a688e51e6	fit: Fix memory allocation for MTP layers (#26605 )	2026-08-05 13:29:45 +02:00
Xuan-Son NguyenandGitHub	9303cdd8d3	security : clarify about AI-generated reports (#26579 ) * security : clarify about AI-generated reports * nits * nits 2	2026-08-05 13:27:06 +02:00
Bhavik ShardaandGitHub	a035a88878	server: Adding spec-decode counters to /metrics endpoint (#26389 ) * * server: add spec-decode counters to /metrics endpoint * server: fixed review comments and now aligned param names exactly with vLLM.	2026-08-05 12:36:01 +02:00
Andreas Krebbel GitHub Sigbjørn Skjæret	020760adfc	convert: Add endianness conversion for Q1 and TQ2 quantizations (#26618 ) * Add endianness conversion for Q1 and TQ2 quantizations * lint --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>	2026-08-05 18:06:09 +08:00
Xuan-Son NguyenandGitHub	61881b1f7f	vendor : apply patches for subprocess.h (#26606 )	2026-08-05 11:26:20 +02:00
Aleksander GrygierandGitHub	3e3a7a416d	ui: show generation statistics by default in chat settings (#26624 )	2026-08-05 11:03:23 +02:00
Niklas WenzelandGitHub	d52ec04a66	build : remove GGML_METAL_USE_BF16 from all build scripts (#26604 )	2026-08-05 10:44:34 +02:00
Aleksander GrygierandGitHub	e031d95679	ui: Update vulnerable packages + cleanup Storybook config (#26607 ) * chore: Upgrade Storybook * chore: Bump package-lock * chore: bump vitest to 4.1.10 * ui: bump fast-uri to 3.1.5 * ui: bump ip-address to 10.4.0 * ui: bump js-yaml to 4.3.1 * ui: bump immutable to 5.1.9 * ui: bump postcss to 8.5.25 * ui: bump brace-expansion to safe versions * ui: bump sharp to 0.35.3 via override * ui: bump body-parser to 2.3.0 * ui: bump vite to 7.3.6 and esbuild to 0.28.1 Assisted-by: Claude Sonnet * ui: bump hono to 4.13.0 * ui: bump dompurify to 3.4.13 * ui: bump @sveltejs/kit to 2.70.2 * ui: bump @modelcontextprotocol/sdk to 1.30.0 * ui: bump valibot to 1.4.2 via override * chore: Remove legacy setup file * refactor: Nits cleanup	2026-08-05 08:06:37 +02:00
Evan HuusandGitHub	6ea215d171	Prefer npm ci over install for security (#26601 )	2026-08-05 00:14:22 +02:00
PascalandGitHub	4308a4f035	server: decode Windows OEM output to UTF-8 in built-in tools (#26597 ) a child process writes in the OEM code page, which is not UTF-8 on a western Windows install, so accented output reaches the JSON layer as invalid bytes and gets replaced there, silently losing the characters run() spawns without a console, so the child never inherits the console code page and GetOEMCP is the one that applies decode with MB_ERR_INVALID_CHARS so a wrong code page returns the text untouched instead of emitting replacement characters, and pass text that already decodes as UTF-8 through so a child emitting UTF-8 is never decoded twice the check drops an incomplete trailing sequence before validating, since a streamed chunk can end in the middle of a multi-byte character	2026-08-04 22:24:55 +02:00
Abhinay KrishnaandGitHub	474c92e722	mtmd: correcting duplicate empty audio chunks for short inputs (#26536 ) * correcting duplicate empty audio chunks for short inputs * tests.sh code restored	2026-08-04 22:05:56 +02:00
Oliver SimonsandGitHub	a6aa6f5450	sampler : remove "full-context windows" from history-based samplers (#26524 ) * Resolve -1 to 1024 instead of ctx-len for samplers Because of backend-sampling we initialize samplers before the complete llama_context is there. Therefore, we cannot infer the resolved context length yet at the time we construct the samplers. * Shared default of 64 for history-based samplers, remove context_size	2026-08-04 21:28:55 +03:00
Guilherme Quintino GitHub Sigbjørn Skjæret	76c956c137	gguf-split: Add option to delete split parts during merge (#26538 ) * Add delete-files option to split parameters Added a new option to delete split files during execution to free up disk space. * Add test for delete files on merge option * Fix tests * Update tools/gguf-split/gguf-split.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> * Update tools/gguf-split/gguf-split.cpp Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co> * Uncomment tests * Improvements to address PR comments * Fix formatting * Fix formatting * Rename --delete-files to --delete-splits * Comment tests * Move delete inside loop * style cleanup --------- Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>	2026-08-04 21:27:47 +03:00
Aleksander Grygier GitHub Pascal	2f56fc3431	ui: CWD for agent (#26518 ) * server : extend file_glob_search for UI pickers * ui : add per-conversation working directory with picker * ui : add path navigation and search scope to cwd picker Treat path-like queries (starting with / or ~) as directory navigation instead of glob-matching the whole query: search the parent for the last segment, and descend into an exactly-typed directory by listing its children. Show the effective search scope in the footer and auto-search on open so the current directory and its siblings appear immediately. Assisted-by: Claude * db : persist per-call tool cwd on tool result messages * ui : abbreviate tool paths under home with a tilde * ui : show the per-call cwd on exec shell rows * ui : clarify the synthetic cwd message for the model * ui : reuse the trailing cwd row on a repeated pick * ui : don't jump when a cwd row is injected mid-chat * chore: Formatting * refactor: Cleanup comments * ui : unify working directory naming and add a synthetic-message flag * ui : render synthetic cwd rows without a scroll jump * ui : decouple the working directory picker into utils and sub-components * ui : add get_info tool call block * chore: Formatting * refactor: Cleanup * refactor: Cleanup * refactor: Cleanup * fix: UI * server : harden file_glob_search listing (kind enum, timeout, symlink guard, absolute base) * ui : use persisted isSynthetic flag for cwd rows, drop legacy formats * ui : cache picker search, fail visibly on native resolve * ui : escape glob metacharacters in picker search glob * ui : simplify auto-scroll pin * chore: Format * fix: Use `SvelteMap` * refactor: Post-review fixes * ui: accept Windows roots in the working directory picker recognize a drive root (C:) and a UNC share (//host/share) as path navigation, alongside the POSIX root and ~, so a query like D:\repos lists that directory instead of glob-matching it under the home dir split below the root, so a bare drive resolves to its root rather than to a drive-relative prefix rewrite backslashes into forward slashes only when the query carries a Windows root, since a backslash is a legal POSIX filename character paths keep travelling with forward slashes, which is what the server returns and what Windows accepts --------- Co-authored-by: Pascal <admin@serveurperso.com>	2026-08-04 19:05:48 +02:00
Xuan-Son Nguyen GitHub Pascal	0713275082	mtmd: support Qwen3-TTS (note: breaking change to llama-tts binary) (#26254 ) * convert text model * main model load ok * convert encoder ok * speaker encoder loading ok * speaker enc graph * adapt vocab for backbone (with some tricks) * add suppress_tokens * poc new mtmd gen api * convert code_predictor to gguf * load gen_code model ok * add clip_encode * wire up * code gen cgraph init version Co-authored-by: Pascal <admin@serveurperso.com> * code2wav convert to gguf * code2wav graph ok * wire up in/out * (wip) subgraph * wire up * wip, correct code2wav * demo (to be removed) * code2wav preserve kv between calls * demo voice clone * llama: add llama_model_get_tok_embd * mtmd_helper_gen_audio API * fix clamp cold prefix Co-authored-by: Pascal <admin@serveurperso.com> * fuse snake op Co-authored-by: Pascal <admin@serveurperso.com> * demo: use proper sampling * update dev docs * polymorphism helper * revamp llama-tts binary * update docs * fix compile * fix lint * nits * add guide + docs * more timings info * clean up code comments * security fixes * update docs * use ggml_build_forward_select, clean up comments * fix ci * use ISO 639-1 language code * rename CODE2WAV --> GEN_WAV, update docs * clean up * clean up tts.cpp * add seq_id * add step_prompt() * mtmd_helper_model_can_chat * clean up comments --------- Co-authored-by: Pascal <admin@serveurperso.com>	2026-08-04 17:26:15 +02:00
Georgi GerganovandGitHub	1c3c9674de	models : fix dflash wo_a reshape on load (#26577 )	2026-08-04 16:56:49 +03:00
Niklas WenzelandGitHub	6b5224cfcc	ci: fix pre-built binaries no longer working on macOS 15 and below (#26375 ) * ci: fix pre-built binaries no longer working on macOS 15 and below * ci: add macOS deployment target to disabled KleidiAI build	2026-08-04 15:03:38 +02:00
Daniel BeveniusandGitHub	7bd8282c37	speculative : refactor enabled configs common_speculative_init (#26510 ) This commit contains a suggestion to reduce some code duplication in common_speculative_init when adding the enabled speculative decoding configurations. No tests were added but the existing server tests still passes with this change: ```console $ ./tests.sh unit/test_speculative.py -v -x ```	2026-08-04 13:17:15 +02:00
hclandGitHub	5788b510a1	gguf-py: validate n_dims and guard against uint64 overflow in reader (#25401 ) The Python GGUF reader lacked two guards the C++ loader has: - n_dims read as uint32 with no GGML_MAX_DIMS bound -> crafted file with huge n_dims triggers oversized memmap read / OOM. - np.prod(dims) on uint64 wraps silently -> a crafted dims triple can overflow to a tiny element count, passing an undersized read through. Add a GGML_MAX_DIMS check and compute the element count with Python ints. Fixes #25378	2026-08-04 12:12:48 +03:00
Georgi Gerganov	2e17f69ef4	sync : ggml	2026-08-04 11:54:54 +03:00
Georgi Gerganov	15831f579a	ggml : bump version to 0.18.1 (ggml/1578)	2026-08-04 11:54:54 +03:00
Angel GalindoandGitHub	b5746d28ce	convert : add missing return after setting tekken vocab (#25947 )	2026-08-04 11:41:18 +03:00
Pranav UttarkarandGitHub	f26efa02a7	vulkan backend ops: implemented GATED_LINEAR_ATTN (#25601 ) * vulkan : add GATED_LINEAR_ATTN op * docs : update Vulkan ops * vulkan : remove unused GLA spec constant * Updated ops.md * ops.md update	2026-08-04 11:40:54 +03:00
Sigbjørn SkjæretandGitHub	cf06ad7dfe	vocab : validate plamo2 byte tokens (#26511 ) * validate plamo2 byte tokens * --typo	2026-08-04 11:40:02 +03:00
Caleb DeLeeuwandGitHub	b06fbc968b	convert : import bytes_to_unicode from convert_slow_tokenizer (#26217 ) bytes_to_unicode was removed from transformers.models.gpt2.tokenization_gpt2 in huggingface/transformers#40936, but it had already been copied into transformers.convert_slow_tokenizer in huggingface/transformers#30334 (transformers 4.54.1), so import it directly from there. Applies the same fix to chatglm.py.	2026-08-04 10:34:30 +03:00
Georgi GerganovandGitHub	1269cb1ff1	model : allow reshape of tensors during load (#26531 )	2026-08-04 09:06:44 +03:00
Oliver SimonsandGitHub	935cad6497	llama : move n_vocab from llama_sampler_data to penalty_sampler (#26520 ) This matches how it is done for logit_bias and mirostat samplers, see https://github.com/ggml-org/llama.cpp/pull/25262#discussion_r3703951151	2026-08-04 09:02:49 +03:00
EveandGitHub	22dc605c4e	ci: fix vulkan llvmpipe runs (#26533 )	2026-08-04 03:28:57 +00:00
TitaniumtownandGitHub	6c8dcaa7ae	sycl: parallelize the non-contiguous concat kernel (#25852 ) * sycl: parallelize the non-contiguous concat kernel Launch geometry only: the non-contiguous concat kernel launched a single-lane work-group (1, 1, 1), now it will launch a (1, 1, SYCL_CONCAT_BLOCK_SIZE) one. SYCL_CONCAT_BLOCK_SIZE is defined in `ggml/src/ggml-sycl/presets.hpp`. llama-bench (Arc Pro B70, Qwen3.6-27B-UD-Q4_K_XL, -fa on, q8_0 KV), on top of upstream master: pp2048 920 -> 1006 t/s (+9.4%) * sycl: cap non-contiguous concat block at ne0 * sycl: make non-contiguous concat block width env-tunable (GGML_SYCL_CONCAT_BLOCK_SIZE) * Revert "sycl: make non-contiguous concat block width env-tunable (GGML_SYCL_CONCAT_BLOCK_SIZE)" This reverts commit `2709909e79`.	2026-08-04 10:08:05 +08:00
Ozymandias_EBON GitHub Claude	66fa168a56	Extended SYCL oneDNN SDPA to non-FP16 KV caches (Q4_0–Q8_0 and FP32) (#25874 ) * sycl: extend oneDNN SDPA to Q4_0-Q8_0 and F32 KV caches Extends the oneDNN SDPA path (PR #25222) to handle non-F16 KV caches by dequantizing or converting K/V to dense FP16 on-device before feeding them into the SDPA graph. The fused systolic kernel then runs identically to the native FP16 path. Supported KV types: - Q4_0, Q4_1, Q5_0, Q5_1, Q8_0: to_fp16_sycl / to_fp16_nc_sycl - F32: cont_to_f16_sycl<float> - BF16 and IQ types are excluded (no conversion kernel available) Gate: non-F16 requires K >= 1024 and Q >= 32 (prefill only). F16 KV runs at any length (existing behavior). Also includes the stream sync fix (stream->wait_and_throw() unconditional, PR #25741 by @malsbat) and removal of V_is_K_view aliasing (K and V are always dequantized to separate buffers). Co-Authored-By: Claude <noreply@anthropic.com> * docs: drop GGML_SYCL_FA_DEBUG from SYCL.md (not shipped in this PR) Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>	2026-08-04 10:07:23 +08:00