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>

.github/workflows/build-apple.yml

@@ -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 \

.github/workflows/build-vulkan.yml

@@ -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)
 

.github/workflows/release.yml

@@ -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 \

SECURITY.md

@@ -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

build-xcframework.sh

@@ -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}
 )

common/arg.cpp

@@ -27,6 +27,7 @@
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
+#include <cmath>
 #include <cstdarg>
 #include <filesystem>
 #include <fstream>
@@ -60,6 +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) {
@@ -359,7 +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(),
@@ -412,11 +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) {
@@ -466,7 +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;
@@ -481,7 +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
@@ -509,14 +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;
@@ -630,11 +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);
     }
@@ -1360,6 +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();
@@ -2022,9 +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;
@@ -2036,7 +2022,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--repeat-penalty"}, "N",
         string_format("penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)", (double)params.sampling.penalty_repeat),
         [](common_params & params, const std::string & value) {
-            params.sampling.penalty_repeat = std::stof(value);
+            const float penalty_repeat = std::stof(value);
+            if (!std::isfinite(penalty_repeat) ||
+                penalty_repeat <= 0.0f ||
+                !std::isfinite(1.0f/penalty_repeat)) {
+                throw std::runtime_error("error: repeat-penalty must be finite and greater than 0\n");
+            }
+            params.sampling.penalty_repeat = penalty_repeat;
             params.sampling.user_sampling_config |= common_params_sampling_config::COMMON_PARAMS_SAMPLING_CONFIG_PENALTY_REPEAT;
         }
     ).set_sampling());
@@ -2044,14 +2036,22 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--presence-penalty"}, "N",
         string_format("repeat alpha presence penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_present),
         [](common_params & params, const std::string & value) {
-            params.sampling.penalty_present = std::stof(value);
+            const float penalty_present = std::stof(value);
+            if (!std::isfinite(penalty_present)) {
+                throw std::runtime_error("error: presence-penalty must be finite\n");
+            }
+            params.sampling.penalty_present = penalty_present;
         }
     ).set_sampling());
     add_opt(common_arg(
         {"--frequency-penalty"}, "N",
         string_format("repeat alpha frequency penalty (default: %.2f, 0.0 = disabled)", (double)params.sampling.penalty_freq),
         [](common_params & params, const std::string & value) {
-            params.sampling.penalty_freq = std::stof(value);
+            const float penalty_freq = std::stof(value);
+            if (!std::isfinite(penalty_freq)) {
+                throw std::runtime_error("error: frequency-penalty must be finite\n");
+            }
+            params.sampling.penalty_freq = penalty_freq;
         }
     ).set_sampling());
     add_opt(common_arg(
@@ -2081,9 +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;
@@ -2567,7 +2567,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.mtmd_batch_max_tokens = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MTMD_BATCH_MAX_TOKENS"));
-    if (llama_supports_rpc()) {
+    if (params.is_gen_docs || llama_supports_rpc()) {
         add_opt(common_arg(
             {"--rpc"}, "SERVERS",
             "comma-separated list of RPC servers (host:port)",
@@ -2968,20 +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)",
@@ -3316,7 +3302,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--tools"}, "TOOL1,TOOL2,...",
         "experimental: whether to enable built-in tools for AI agents - do not enable in untrusted environments (default: no tools)\n"
         "specify \"all\" to enable all tools\n"
-        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime\n"
+        "available tools: read_file, file_glob_search, grep_search, exec_shell_command, write_file, edit_file, get_datetime, get_info\n"
         "note: for security reasons, this will limit --cors-origins to localhost by default",
         [](common_params & params, const std::string & value) {
             params.server_tools = parse_csv_row(value);
@@ -4257,24 +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}));
 
@@ -4394,16 +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"},

common/arg.h

@@ -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;
 };
 

common/chat.cpp

@@ -2114,6 +2114,11 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
     auto extract_reasoning   = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
     auto include_grammar     = has_response_format || (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE);
 
+    std::optional<json> additional_context;
+    if (is_v4 && has_response_format) {
+        additional_context = json{ { "response_format", inputs.json_schema } };
+    }
+
     const std::string DSML         = "｜DSML｜";
     const std::string THINK_START  = "<think>";
     const std::string THINK_END    = "</think>";
@@ -2125,9 +2130,12 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
     const std::string PARAM_START  = "<" + DSML + "parameter";
     const std::string PARAM_END    = "</" + DSML + "parameter>";
     const std::string GEN_PROMPT   = "<｜Assistant｜>";
+    const std::string TC_SEPARATOR = "\n\n";
 
-    data.prompt             = common_chat_template_direct_apply_impl(tmpl, inputs, adjusted_messages);
-    data.generation_prompt  = common_chat_template_generation_prompt_impl(tmpl, inputs, adjusted_messages);
+    data.prompt = common_chat_template_direct_apply_impl(
+        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
+    data.generation_prompt = common_chat_template_generation_prompt_impl(
+        tmpl, inputs, adjusted_messages, std::nullopt, additional_context);
     data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.supports_thinking  = true;
     data.thinking_start_tag = THINK_START;
@@ -2141,9 +2149,16 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
     if (inputs.has_continuation()) {
         const auto & msg = inputs.continue_msg;
 
-        data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
-        if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
-            data.generation_prompt += THINK_END + msg.render_content();
+        if (is_v4 && msg.reasoning_content.empty()) {
+            data.generation_prompt = GEN_PROMPT + THINK_END;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += msg.render_content();
+            }
+        } else {
+            data.generation_prompt = GEN_PROMPT + THINK_START + msg.reasoning_content;
+            if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
+                data.generation_prompt += THINK_END + msg.render_content();
+            }
         }
 
         data.prompt += data.generation_prompt;
@@ -2242,7 +2257,9 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
 
         if (extract_reasoning && inputs.enable_thinking) {
             reasoning = p.optional(THINK_START + p.reasoning(p.until(THINK_END)) + THINK_END);
-            reasoning_with_tc = THINK_START + p.reasoning(p.until_one_of({ FC_START, THINK_END })) + obligatory_tool_calls;
+            reasoning_with_tc = THINK_START +
+                p.reasoning(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START, THINK_END })) +
+                p.space() + obligatory_tool_calls;
             allow_reasoning_with_tc = true;
         } else if (extract_reasoning) {
             // Thinking disabled but reasoning extraction requested: the generation prompt
@@ -2265,7 +2282,9 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
             return generation_prompt + reasoning + p.content(p.rest()) + end;
         }
 
-        auto content_before_tools = p.negate(p.literal(THINK_START)) + p.content(p.until(FC_START));
+        auto content_before_tools = p.negate(p.literal(THINK_START)) +
+            p.content(p.until_one_of({ TC_SEPARATOR + FC_START, FC_START })) +
+            p.space();
         return allow_reasoning_with_tc ? generation_prompt + (reasoning_with_tc | (reasoning + content_before_tools + tool_calls)) + end :
             generation_prompt + reasoning + content_before_tools + tool_calls + end;
     });

common/common.cpp

@@ -998,6 +998,23 @@ bool fs_is_directory(const std::string & path) {
     return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
 }
 
+std::string common_get_env(const std::string & name) {
+    const char * value = std::getenv(name.c_str());
+    return value == nullptr ? "" : value;
+}
+
+void common_set_env(const std::string & name, const std::string & value) {
+#if defined(_WIN32)
+    _putenv_s(name.c_str(), value.c_str());
+#else
+    if (value.empty()) {
+        unsetenv(name.c_str());
+    } else {
+        setenv(name.c_str(), value.c_str(), 1);
+    }
+#endif
+}
+
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
     auto ensure_trailing_slash = [](std::string p) {
@@ -1285,16 +1302,6 @@ 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);
@@ -1462,18 +1469,18 @@ common_init_result_ptr common_init_from_params(common_params & params, bool mode
 common_init_result::~common_init_result() = default;
 
 std::string common_get_model_endpoint() {
-    const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
-    // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
-    const char * hf_endpoint_env = getenv("HF_ENDPOINT");
-    const char * endpoint_env = model_endpoint_env ? model_endpoint_env : hf_endpoint_env;
-    std::string model_endpoint = "https://huggingface.co/";
-    if (endpoint_env) {
-        model_endpoint = endpoint_env;
-        if (model_endpoint.back() != '/') {
-            model_endpoint += '/';
-        }
+    std::string endpoint = common_get_env("MODEL_ENDPOINT");
+    if (endpoint.empty()) {
+        // the HF_ENDPOINT variable is respected for backward compatibility
+        endpoint = common_get_env("HF_ENDPOINT");
     }
-    return model_endpoint;
+    if (endpoint.empty()) {
+        return "https://huggingface.co/";
+    }
+    if (endpoint.back() != '/') {
+        endpoint += '/';
+    }
+    return endpoint;
 }
 
 char * common_get_model_or_exit(int argc, char * argv[]) {

common/common.h

@@ -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;
@@ -739,6 +730,12 @@ struct common_params {
     llama_progress_callback load_progress_callback = NULL;
     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
 };
 
 // call once at the start of a program if it uses libcommon
@@ -863,6 +860,15 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 
 bool glob_match(const std::string & pattern, const std::string & str);
 
+//
+// Environment utils
+//
+
+// portable environment access, an unset variable reads as an empty string
+// and setting an empty value unsets the variable
+std::string common_get_env(const std::string & name);
+void        common_set_env(const std::string & name, const std::string & value);
+
 //
 // Filesystem utils
 //

common/fit.cpp

@@ -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);
 

common/jinja/caps.cpp

@@ -482,6 +482,7 @@ caps caps_get(jinja::program & prog) {
             });
         },
         [&](context & ctx) {
+            ctx.set_val("enable_thinking", mk_val<value_bool>(true));
             caps_apply_preserve_reasoning(ctx, true);
         },
         nullptr, // tools_fn

common/sampling.cpp

@@ -184,9 +184,21 @@ std::string common_params_sampling::print() const {
     return std::string(result);
 }
 
-struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params) {
+struct common_sampler * common_sampler_init(
+        const struct llama_model * model,
+        struct common_params_sampling & params) {
+    if (!std::isfinite(params.penalty_repeat) ||
+        params.penalty_repeat <= 0.0f ||
+        !std::isfinite(1.0f/params.penalty_repeat)) {
+        throw std::invalid_argument("penalty_repeat must be finite and greater than 0");
+    }
+    if (!std::isfinite(params.penalty_freq)) {
+        throw std::invalid_argument("penalty_freq must be finite");
+    }
+    if (!std::isfinite(params.penalty_present)) {
+        throw std::invalid_argument("penalty_present must be finite");
+    }
     const llama_vocab * vocab = llama_model_get_vocab(model);
-
     llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
 
     lparams.no_perf = params.no_perf;
@@ -338,7 +350,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
                         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:
@@ -366,7 +378,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
                     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

common/sampling.h

@@ -37,7 +37,9 @@ struct common_sampler;
 // llama_sampler API overloads
 
 // note: can mutate params in some cases
-struct common_sampler * common_sampler_init(const struct llama_model * model, struct common_params_sampling & params);
+struct common_sampler * common_sampler_init(
+        const struct llama_model * model,
+        struct common_params_sampling & params);
 
 void common_sampler_free(struct common_sampler * gsmpl);
 

common/speculative.cpp

@@ -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 = {};

conversion/__init__.py

@@ -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",

conversion/chatglm.py

@@ -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')])
 

conversion/deepseek.py

@@ -447,12 +447,43 @@ class DeepseekV2Model(TextModel):
 class DeepseekV32Model(DeepseekV2Model):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK32
     skip_mtp = False
+    supports_mtp_export = True
+    _n_main_layers: int | None = None
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.block_count = self.hparams["num_hidden_layers"]
+        if not self.no_mtp:
+            self.block_count += self.hparams.get("num_nextn_predict_layers", 0)
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
 
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        # DeepSeek V3.2 appends the NextN/MTP block past num_hidden_layers
+        # (model.layers.61 -> blk.61 in the 62-block file).
+        assert cls._n_main_layers is not None
+        is_mtp = (m := re.match(r"model\.layers\.(\d+)\.", name)) is not None and int(m.group(1)) >= cls._n_main_layers
+
+        # --no-mtp: drop the appended NextN block entirely.
+        if is_mtp and cls.no_mtp:
+            return None
+        # --mtp: keep ONLY NextN-block tensors plus the shared embeddings/
+        # norm/lm_head (so the resulting GGUF carries just the draft head).
+        if cls.mtp_only and not is_mtp and name not in (
+            "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+        ):
+            return None
+
+        return name, gen
+
     def set_vocab(self):
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
@@ -463,7 +494,7 @@ class DeepseekV32Model(DeepseekV2Model):
         super().set_gguf_parameters()
 
         # NextN/MTP prediction layers
-        if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
+        if not self.no_mtp and (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
             self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
 
         # DSA indexer parameters
@@ -504,7 +535,10 @@ class DeepseekV4Model(TextModel):
             logger.info("Skipping %d DeepSeek-V4 MTP tensor(s) for conversion v0", type(self)._skipped_mtp_tensors)
 
         # add a default chat template; if the model has a built-in template, it will be overridden later
-        template_path = Path(__file__).parent.parent / "models" / "templates" / "deepseek-ai-DeepSeek-V4.jinja"
+        model_id_hint = self.remote_hf_model_id or self.dir_model.name
+        is_0731 = "0731" in model_id_hint
+        template_name = "deepseek-ai-DeepSeek-V4-Flash-0731.jinja" if is_0731 else "deepseek-ai-DeepSeek-V4.jinja"
+        template_path = Path(__file__).parent.parent / "models" / "templates" / template_name
         if template_path.is_file():
             with open(template_path, "r", encoding="utf-8") as f:
                 self.gguf_writer.add_chat_template(f.read())

conversion/glm.py

@@ -206,10 +206,70 @@ class Glm4MoeModel(TextModel):
 @ModelBase.register("Glm4MoeLiteForCausalLM")
 class Glm4MoeLiteModel(DeepseekV2Model):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
+    skip_mtp = False
+    supports_mtp_export = True
+    _n_main_layers: int | None = None
 
     def set_vocab(self):
         return self._set_vocab_glm()
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        num_hidden_layers = self.hparams["num_hidden_layers"]
+        self.num_nextn_predict_layers = self.hparams.get("num_nextn_predict_layers", 0)
+        self.skip_mtp = self.no_mtp or self.num_nextn_predict_layers == 0
+
+        if self.skip_mtp:
+            self.block_count = num_hidden_layers
+        else:
+            self.block_count = num_hidden_layers + self.num_nextn_predict_layers
+
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        if self.skip_mtp:
+            return
+
+        self.gguf_writer.add_nextn_predict_layers(self.num_nextn_predict_layers)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None):
+        type(self)._n_main_layers = self.hparams["num_hidden_layers"]
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
+
+    @classmethod
+    def filter_tensors(cls, item):
+        if (titem := super().filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+
+        if cls._n_main_layers is not None:
+            match = re.match(r"model\.layers\.(\d+)\.", name)
+            is_mtp = match is not None and int(match.group(1)) >= cls._n_main_layers
+            if is_mtp and cls.no_mtp:
+                return None
+            if cls.mtp_only and not is_mtp and name not in (
+                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+            ):
+                return None
+
+        return name, gen
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
 
 @ModelBase.register("GlmMoeDsaForCausalLM")
 class GlmMoeDsaModel(DeepseekV2Model):

conversion/llama.py

@@ -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():

conversion/qwen.py

@@ -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')])
 
@@ -268,8 +268,101 @@ class Qwen3MoeModel(Qwen2MoeModel):
         super().set_vocab()
 
 
+class _QwenMtpMixin:
+    """Shared MTP wiring for Qwen3-Next and Qwen3.5/3.6 text variants. The HF
+    config carries the MTP block under `mtp_num_hidden_layers` (computed from
+    the checkpoint when absent, e.g. Qwen3-Next) and the tensors under
+    `mtp.*`; we extend block_count, emit the nextn metadata key, and remap
+    `mtp.*` to the standard layer-indexed nextn naming so the existing
+    tensor_map handles them."""
+
+    supports_mtp_export = True
+    hparams: dict[str, Any]
+    model_arch: gguf.MODEL_ARCH
+    gguf_writer: gguf.GGUFWriter
+    block_count: int
+    tensor_map: gguf.TensorNameMap
+    no_mtp: bool
+    mtp_only: bool
+    _original_block_count: int | None = None
+    opt_num_mtp_layers: int = 0
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.block_count = self.hparams["num_hidden_layers"]
+        if not self.no_mtp:
+            n_mtp = self.hparams.get("mtp_num_hidden_layers", 0)
+            # Qwen-3-Next doesn't include `mtp_num_hidden_layers` in config.
+            if n_mtp == 0:
+                assert self.opt_num_mtp_layers != 0
+                n_mtp = self.opt_num_mtp_layers
+            self.block_count += n_mtp
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
+        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
+        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
+        type(self)._original_block_count = hparams.get(key)
+        type(self).opt_num_mtp_layers = 0
+        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)  # ty: ignore[unresolved-attribute]
+
+    @classmethod
+    def filter_tensors(cls, item):
+        assert cls._original_block_count is not None
+        # TODO: change TextModel to super()
+        if (titem := TextModel.filter_tensors(item)) is None:
+            return None
+        name, gen = titem
+        if name.startswith("model.mtp."):
+            name = name.replace("model.", "", 1)
+        if name.startswith("mtp."):
+            if cls.no_mtp:
+                return None
+            remapper = {
+                "fc":                    "eh_proj",
+                "pre_fc_norm_embedding": "enorm",
+                "pre_fc_norm_hidden":    "hnorm",
+                "norm":                  "shared_head.norm",
+            }
+            parts = name.split(".", 3)
+            if len(parts) == 4 and parts[1] == "layers" and parts[2].isdecimal():
+                mtp_idx = int(parts[2])
+                name = f"model.layers.{cls._original_block_count + mtp_idx}.{parts[3]}"
+                cls.opt_num_mtp_layers = max(cls.opt_num_mtp_layers, mtp_idx + 1)
+            elif len(parts) == 3 and parts[1] in remapper:
+                name = f"model.layers.{cls._original_block_count}.{remapper[parts[1]]}.{parts[2]}"
+        elif cls.mtp_only:
+            keep = name in (
+                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
+                "embed_tokens.weight", "norm.weight",
+            )
+            if not keep:
+                return None
+        return name, gen
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()  # ty: ignore[unresolved-attribute]
+        if self.no_mtp:
+            return
+        if (n := self.block_count - self.hparams["num_hidden_layers"]) > 0:
+            self.gguf_writer.add_nextn_predict_layers(n)
+
+    def prepare_metadata(self, vocab_only: bool):
+        from_dir = self.fname_out.is_dir()
+        super().prepare_metadata(vocab_only=vocab_only)  # ty: ignore[unresolved-attribute]
+
+        if not self.mtp_only or not from_dir:
+            return
+
+        output_type: str = self.ftype.name.partition("_")[2]  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        fname_default: str = gguf.naming_convention(
+            self.metadata.name, self.metadata.basename, self.metadata.finetune,                  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+            self.metadata.version, size_label=None, output_type=output_type, model_type=None)    # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
+        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
+
+
 @ModelBase.register("Qwen3NextForCausalLM")
-class Qwen3NextModel(Qwen2MoeModel):
+class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel):
     model_arch = gguf.MODEL_ARCH.QWEN3NEXT
 
     def set_gguf_parameters(self):
@@ -284,16 +377,6 @@ class Qwen3NextModel(Qwen2MoeModel):
             rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
         self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.rope_parameters.get("partial_rotary_factor", 0.25)))
 
-    @classmethod
-    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
-        name, gen = item
-
-        if name.startswith("mtp"):
-            # ignore MTP layers for now
-            return None
-
-        return super().filter_tensors(item)
-
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         if name.endswith(".A_log"):
             data_torch = -torch.exp(data_torch)
@@ -536,97 +619,13 @@ class _Qwen35MRopeMixin:
             self.gguf_writer.add_rope_dimension_sections(self._QWEN35_DEFAULT_MROPE_SECTION)
 
 
-class _Qwen35MtpMixin:
-    """Shared MTP wiring for Qwen3.5/3.6 text variants. The HF config carries
-    the MTP block under `mtp_num_hidden_layers` and the tensors under
-    `mtp.*`; we extend block_count, emit the nextn metadata key, and remap
-    `mtp.*` to the standard layer-indexed nextn naming so the existing
-    tensor_map handles them."""
-
-    supports_mtp_export = True
-    hparams: dict[str, Any]
-    model_arch: gguf.MODEL_ARCH
-    gguf_writer: gguf.GGUFWriter
-    block_count: int
-    tensor_map: gguf.TensorNameMap
-    no_mtp: bool
-    mtp_only: bool
-    _original_block_count: int | None = None
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.block_count = self.hparams["num_hidden_layers"]
-        if not self.no_mtp:
-            self.block_count += self.hparams.get("mtp_num_hidden_layers", 0)
-        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
-
-    def index_tensors(self, remote_hf_model_id: str | None = None) -> dict[str, Callable[[], Tensor]]:
-        hparams = {**self.hparams, **self.hparams.get("text_config", {})}
-        key = next((k for k in ["n_layers", "num_hidden_layers", "n_layer", "num_layers"] if k in hparams), None)
-        type(self)._original_block_count = hparams.get(key)
-        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)  # ty: ignore[unresolved-attribute]
-
-    @classmethod
-    def filter_tensors(cls, item):
-        assert cls._original_block_count is not None
-        # TODO: change TextModel to super()
-        if (titem := TextModel.filter_tensors(item)) is None:
-            return None
-        name, gen = titem
-        if name.startswith("model.mtp."):
-            name = name.replace("model.", "", 1)
-        if name.startswith("mtp."):
-            if cls.no_mtp:
-                return None
-            remapper = {
-                "fc":                    "eh_proj",
-                "pre_fc_norm_embedding": "enorm",
-                "pre_fc_norm_hidden":    "hnorm",
-                "norm":                  "shared_head.norm",
-            }
-            parts = name.split(".", 3)
-            if len(parts) == 4 and parts[1] == "layers" and parts[2].isdecimal():
-                mtp_idx = int(parts[2])
-                name = f"model.layers.{cls._original_block_count + mtp_idx}.{parts[3]}"
-            elif len(parts) == 3 and parts[1] in remapper:
-                name = f"model.layers.{cls._original_block_count}.{remapper[parts[1]]}.{parts[2]}"
-        elif cls.mtp_only:
-            keep = name in (
-                "model.embed_tokens.weight", "model.norm.weight", "lm_head.weight",
-                "embed_tokens.weight", "norm.weight",
-            )
-            if not keep:
-                return None
-        return name, gen
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()  # ty: ignore[unresolved-attribute]
-        if self.no_mtp:
-            return
-        if (n := self.hparams.get("mtp_num_hidden_layers", 0)) > 0:
-            self.gguf_writer.add_nextn_predict_layers(n)
-
-    def prepare_metadata(self, vocab_only: bool):
-        from_dir = self.fname_out.is_dir()
-        super().prepare_metadata(vocab_only=vocab_only)  # ty: ignore[unresolved-attribute]
-
-        if not self.mtp_only or not from_dir:
-            return
-
-        output_type: str = self.ftype.name.partition("_")[2]  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-        fname_default: str = gguf.naming_convention(
-            self.metadata.name, self.metadata.basename, self.metadata.finetune,                  # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-            self.metadata.version, size_label=None, output_type=output_type, model_type=None)    # pyright: ignore[reportAttributeAccessIssue] # ty: ignore[unresolved-attribute]
-        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
-
-
 @ModelBase.register("Qwen3_5ForConditionalGeneration", "Qwen3_5ForCausalLM")
-class Qwen3_5TextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase):
+class Qwen3_5TextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35
 
 
 @ModelBase.register("Qwen3_5MoeForConditionalGeneration", "Qwen3_5MoeForCausalLM")
-class Qwen3_5MoeTextModel(_Qwen35MtpMixin, _Qwen35MRopeMixin, _LinearAttentionVReorderBase):
+class Qwen3_5MoeTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35MOE
 
 

conversion/qwen3tts.py

@@ -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"))

docs/development/HOWTO-add-model.md

@@ -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
 

docs/ops.md

@@ -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 | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |

examples/gen-docs/gen-docs.cpp

@@ -70,6 +70,8 @@ static void write_table(std::ostringstream & ss, std::vector<common_arg *> & opt
 
 static void write_help(std::ostringstream & ss, const md_file & md) {
     common_params params;
+    params.is_gen_docs = true;
+
     auto ctx_arg = common_params_parser_init(params, md.ex);
 
     std::vector<common_arg *> common_options;

ggml/CMakeLists.txt

@@ -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/")

ggml/src/ggml-backend.cpp

@@ -765,8 +765,9 @@ struct ggml_backend_sched_split {
     int backend_id;
     int i_start;
     int i_end;
-    struct ggml_tensor * inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    struct ggml_tensor ** inputs;
     int n_inputs;
+    int inputs_capacity;
     // graph view of this split
     struct ggml_cgraph graph;
 };
@@ -805,8 +806,9 @@ struct ggml_backend_sched {
     int cur_copy;
     int next_copy;
     ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
-    struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    struct ggml_tensor ** graph_inputs;
     int n_graph_inputs;
+    int graph_inputs_capacity;
 
     struct ggml_context * ctx;
 
@@ -832,6 +834,36 @@ struct ggml_backend_sched {
 #define tensor_id_copy(id, backend_id, copy_id) sched->hv_tensor_copies[(id) * sched->n_backends * sched->n_copies + (backend_id) * sched->n_copies + (copy_id)]
 #define tensor_copy(tensor, backend_id, copy_id) tensor_id_copy(hash_id(tensor), backend_id, copy_id)
 
+static void ggml_backend_sched_split_inputs_grow(struct ggml_backend_sched_split * split) {
+    int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
+    if (split->inputs_capacity > 0) {
+        new_cap = 2*split->inputs_capacity;
+        GGML_LOG_WARN("%s: increasing split inputs capacity from %d to %d\n", __func__, split->inputs_capacity, new_cap);
+    }
+    auto * pnew = (struct ggml_tensor **) realloc((void *) split->inputs, new_cap * sizeof(struct ggml_tensor *));
+    if (pnew == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate %zu bytes\n", __func__, new_cap * sizeof(struct ggml_tensor *));
+        GGML_ABORT("failed to grow split inputs container");
+    }
+    split->inputs = pnew;
+    split->inputs_capacity = new_cap;
+}
+
+static void ggml_backend_sched_graph_inputs_grow(ggml_backend_sched_t sched) {
+    int new_cap = GGML_SCHED_MAX_SPLIT_INPUTS;
+    if (sched->graph_inputs_capacity > 0) {
+        new_cap = 2*sched->graph_inputs_capacity;
+        GGML_LOG_WARN("%s: increasing graph inputs capacity from %d to %d\n", __func__, sched->graph_inputs_capacity, new_cap);
+    }
+    auto * pnew = (struct ggml_tensor **) realloc((void *) sched->graph_inputs, new_cap * sizeof(struct ggml_tensor *));
+    if (pnew == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate %zu bytes\n", __func__, new_cap * sizeof(struct ggml_tensor *));
+        GGML_ABORT("failed to grow graph inputs container");
+    }
+    sched->graph_inputs = pnew;
+    sched->graph_inputs_capacity = new_cap;
+}
+
 // returns the priority of the backend, lower id is higher priority
 static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
     for (int i = 0; i < sched->n_backends; i++) {
@@ -1297,7 +1329,7 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                     }
                     // check if the split has too many inputs
                     // FIXME: count the number of inputs instead of only checking when full
-                    if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
+                    if (split->n_inputs >= split->inputs_capacity) {
                         const size_t id = hash_id(src);
                         int src_backend_id = sched->hv_tensor_backend_ids[id];
                         bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
@@ -1313,10 +1345,14 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                 split->i_end = i;
                 i_split++;
                 if (i_split >= sched->splits_capacity) {
+                    int old_cap = sched->splits_capacity;
                     sched->splits_capacity *= 2;
                     sched->splits = (ggml_backend_sched_split *)
                         realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
                     GGML_ASSERT(sched->splits != NULL);
+                    for (int k = old_cap; k < sched->splits_capacity; k++) {
+                        memset(&sched->splits[k], 0, sizeof(struct ggml_backend_sched_split));
+                    }
                 }
                 split = &sched->splits[i_split];
                 split->backend_id = node_backend_id;
@@ -1353,7 +1389,9 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
                         int n_graph_inputs = sched->n_graph_inputs++;
-                        GGML_ASSERT(n_graph_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        if (n_graph_inputs >= sched->graph_inputs_capacity) {
+                            ggml_backend_sched_graph_inputs_grow(sched);
+                        }
                         sched->graph_inputs[n_graph_inputs] = src;
                     }
                 }
@@ -1373,7 +1411,9 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
                         int n_inputs = split->n_inputs++;
-                        GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        if (n_inputs >= split->inputs_capacity) {
+                            ggml_backend_sched_split_inputs_grow(split);
+                        }
                         split->inputs[n_inputs] = src;
                     }
                     node->src[j] = tensor_id_copy(src_id, cur_backend_id, sched->cur_copy);
@@ -1399,7 +1439,11 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
         sched->prev_leaf_backend_ids = tmp;
     }
 
-    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sched->n_copies;
+    int total_inputs = sched->n_graph_inputs;
+    for (int i = 0; i < sched->n_splits; i++) {
+        total_inputs += sched->splits[i].n_inputs;
+    }
+    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + total_inputs * 2 * sched->n_copies;
 
     // remember the actual graph_size for performing reallocation checks later [GGML_SCHED_DEBUG_REALLOC]
     sched->debug_prev_graph_size = sched->debug_graph_size;
@@ -1782,6 +1826,9 @@ ggml_backend_sched_t ggml_backend_sched_new(
     sched->splits = (ggml_backend_sched_split *) calloc(initial_splits_capacity, sizeof(sched->splits[0]));
     sched->splits_capacity = initial_splits_capacity;
 
+    sched->graph_inputs_capacity = GGML_SCHED_MAX_SPLIT_INPUTS;
+    sched->graph_inputs = (struct ggml_tensor **) calloc(sched->graph_inputs_capacity, sizeof(struct ggml_tensor *));
+
     for (int b = 0; b < n_backends; b++) {
         sched->backends[b] = backends[b];
         sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
@@ -1814,7 +1861,11 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
     ggml_gallocr_free(sched->galloc);
     ggml_free(sched->ctx);
     ggml_hash_set_free(&sched->hash_set);
+    for (int i = 0; i < sched->splits_capacity; i++) {
+        free(sched->splits[i].inputs);
+    }
     free(sched->splits);
+    free(sched->graph_inputs);
     free(sched->hv_tensor_backend_ids);
     free(sched->hv_tensor_copies);
     free(sched->node_backend_ids);

ggml/src/ggml-cuda/common.cuh

@@ -627,7 +627,8 @@ template <typename T> struct block_reduce_policy<block_reduce_method::MAX, T> {
 };
 
 template <block_reduce_method reduce_method_t, const unsigned int block_size_template = 0, typename T>
-static __device__ T block_reduce(T val, T * shared_vals) {
+static __device__ T block_reduce(T val, [[maybe_unused]] T * shared_vals) {
+    // for multi-warp reductions, callers must not reuse shared_vals until all reads from this invocation have completed
     val                           = block_reduce_policy<reduce_method_t, T>::reduce(val);
     const unsigned int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
     if (block_size > WARP_SIZE) {

ggml/src/ggml-cuda/norm.cu

@@ -64,7 +64,7 @@ static __global__ void group_norm_f32(const float * x, float * dst, const int gr
         tmp += xi * xi;
     }
 
-    tmp = block_reduce<block_reduce_method::SUM, block_size>(tmp, s_sum);
+    tmp = block_reduce<block_reduce_method::SUM, block_size>(tmp, s_sum + 32);
 
     const float variance = tmp / group_size;
     const float scale = rsqrtf(variance + eps);
@@ -297,7 +297,7 @@ static void group_norm_f32_cuda(
         group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
     } else {
         const dim3 block_dims(1024, 1, 1);
-        group_norm_f32<1024><<<num_groups, block_dims, block_dims.x > WARP_SIZE ? 32 * sizeof(float): 0, stream>>>(x, dst, group_size, ne_elements, eps);
+        group_norm_f32<1024><<<num_groups, block_dims, block_dims.x > WARP_SIZE ? 2 * 32 * sizeof(float): 0, stream>>>(x, dst, group_size, ne_elements, eps);
     }
 }
 

ggml/src/ggml-cuda/softmax.cu

@@ -116,6 +116,11 @@ static __global__ void soft_max_f32(
         vals[col] = val;
     }
 
+    if (block_size > WARP_SIZE) {
+        // sync is needed as we reuse buf_iw across block_reduce invocations, see #26385
+        // for block_size <= WARP_SIZE, block_reduce does not access buf_iw
+        __syncthreads();
+    }
     // find the sum of exps in the block
     tmp = block_reduce<block_reduce_method::SUM, block_size_template>(tmp, buf_iw);
 
@@ -142,6 +147,8 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
                                                                 float * __restrict__ dst,
                                                                 float * __restrict__ tmp_maxs,
                                                                 float * __restrict__ tmp_sums,
+                                                                float * shared_vals_max,
+                                                                float * shared_vals_sum,
                                                                 const soft_max_params p) {
     namespace cg = cooperative_groups;
 
@@ -154,7 +161,6 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     float     local_vals[n_elem_per_thread] = { -INFINITY, -INFINITY, -INFINITY, -INFINITY };
     float     local_max                     = -INFINITY;
     const int step_size                     = gridDim.x * blockDim.x;
-    __shared__ float shared_vals[32];
 
     // Compute thread-local max
     for (int col = col_start; col < p.ncols;) {
@@ -171,7 +177,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     }
 
     // Compute CTA-level max
-    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals);
+    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals_max);
 
     // Store CTA-level max to GMEM
     if (tid == 0) {
@@ -186,7 +192,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     } else {
         local_max = -INFINITY;
     }
-    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals);
+    local_max = block_reduce<block_reduce_method::MAX>(local_max, shared_vals_max);
 
     // Compute softmax dividends, accumulate divisor
     float tmp_expf = 0.0f;
@@ -209,7 +215,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     }
 
     // Reduce divisor within CTA
-    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals);
+    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals_sum);
 
     // Store CTA-level sum to GMEM
     if (tid == 0) {
@@ -223,7 +229,7 @@ static __device__ void soft_max_f32_parallelize_cols_single_row(const float * __
     } else {
         tmp_expf = 0.0f;
     }
-    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals);
+    tmp_expf = block_reduce<block_reduce_method::SUM>(tmp_expf, shared_vals_sum);
 
     // Divide dividend by global sum + store data
     for (int col = col_start; col < p.ncols;) {
@@ -310,9 +316,11 @@ __launch_bounds__(8*WARP_SIZE, 1) static __global__ void soft_max_f32_paralleliz
 // https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/device-callable-apis.html#grid-synchronization
 // https://docs.nvidia.com/cuda/cuda-programming-guide/05-appendices/device-callable-apis.html#class-cluster-group
 {
+    __shared__ float shared_vals[2][32];
+
     for (int rowx = 0; rowx < p.ne01 * p.ne02 * p.ne03; rowx++) {
         soft_max_f32_parallelize_cols_single_row(x + int64_t(rowx) * p.ncols, dst + int64_t(rowx) * p.ncols, tmp_maxs,
-                                                 tmp_sums, p);
+                                                 tmp_sums, shared_vals[0], shared_vals[1], p);
     }
 }
 

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -477,6 +477,41 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_soft_max(ggml_me
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_lightning_indexer(
+        ggml_metal_library_t lib,
+        const ggml_tensor * op) {
+    GGML_ASSERT(op->op == GGML_OP_LIGHTNING_INDEXER);
+
+    char name[256];
+
+    snprintf(name, 256, "kernel_lightning_indexer_%s", ggml_type_name(op->src[1]->type));
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, name, name, nullptr);
+    }
+
+    return res;
+}
+
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_dsv4_hc(ggml_metal_library_t lib, ggml_op op) {
+    const char * name = nullptr;
+
+    switch (op) {
+        case GGML_OP_DSV4_HC_COMB: name = "kernel_dsv4_hc_comb_f32"; break;
+        case GGML_OP_DSV4_HC_PRE:  name = "kernel_dsv4_hc_pre_f32";  break;
+        case GGML_OP_DSV4_HC_POST: name = "kernel_dsv4_hc_post_f32"; break;
+        default: GGML_ABORT("fatal error");
+    }
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, name, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv(ggml_metal_library_t lib, const ggml_tensor * op) {
     GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
     GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
@@ -2117,6 +2152,23 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_opt_step_sgd(ggm
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_silu_back(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_SILU_BACK);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_silu_back_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_memset(ggml_metal_library_t lib, const ggml_tensor *  op) {
     GGML_ASSERT(op->type == GGML_TYPE_I64);
 

ggml/src/ggml-metal/ggml-metal-device.h

@@ -117,6 +117,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_diag
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_concat            (ggml_metal_library_t lib, enum ggml_type tsrc);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary             (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_silu_back         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum_rows          (ggml_metal_library_t lib, const struct ggml_tensor * op);
@@ -124,6 +125,8 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cumsum_bl
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cumsum_add        (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_tri               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_soft_max          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_lightning_indexer (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_dsv4_hc           (ggml_metal_library_t lib, enum ggml_op op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched  (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);

ggml/src/ggml-metal/ggml-metal-device.m

@@ -2,6 +2,7 @@
 
 #import "ggml-impl.h"
 #import "ggml-backend-impl.h"
+#import "ggml-metal-impl.h"
 
 #include <Foundation/Foundation.h>
 
@@ -1137,6 +1138,14 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 default:
                     return false;
             }
+        case GGML_OP_SILU_BACK:
+            return (op->src[0]->type == GGML_TYPE_F32) &&
+                (op->src[1]->type == GGML_TYPE_F32) &&
+                (op->type == GGML_TYPE_F32) &&
+                ggml_is_contiguous(op->src[0]) &&
+                ggml_is_contiguous(op->src[1]) &&
+                ggml_is_contiguous(op) &&
+                ggml_are_same_shape(op->src[0], op->src[1]);
         case GGML_OP_GLU:
             switch (ggml_get_glu_op(op)) {
                 case GGML_GLU_OP_REGLU:
@@ -1181,6 +1190,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_ADD_ID:
+            return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) && (op->src[0]->type == op->src[1]->type);
         case GGML_OP_ACC:
             return ggml_is_contiguous_rows(op->src[0]) && ggml_is_contiguous_rows(op->src[1]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_REPEAT:
@@ -1299,6 +1309,72 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                     return false;
             }
             return has_simdgroup_mm; // TODO: over-restricted for vec-kernels
+        case GGML_OP_LIGHTNING_INDEXER:
+            if (op->src[0]->ne[0] != OP_LIGHTNING_INDEXER_DK ||
+                op->src[0]->ne[1] != OP_LIGHTNING_INDEXER_NH) {
+                return false;
+            }
+            if (!has_simdgroup_mm ||
+                op->src[0]->type != GGML_TYPE_F32 ||
+                op->src[2]->type != GGML_TYPE_F32 ||
+                op->src[3]->type != GGML_TYPE_F16 ||
+                op->type         != GGML_TYPE_F32 ||
+                !ggml_is_contiguous_rows(op->src[0]) ||
+                !ggml_is_contiguous_rows(op->src[1]) ||
+                !ggml_is_contiguous_rows(op->src[2]) ||
+                !ggml_is_contiguous_rows(op->src[3])) {
+                return false;
+            }
+            switch (op->src[1]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                    return true;
+                case GGML_TYPE_BF16:
+                    return has_bfloat;
+                default:
+                    return false;
+            }
+        case GGML_OP_DSV4_HC_COMB:
+            return has_simdgroup_reduction &&
+                op->src[0]->type == GGML_TYPE_F32 &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->src[2]->type == GGML_TYPE_F32 &&
+                op->type         == GGML_TYPE_F32 &&
+                op->src[0]->ne[0] == 24 &&
+                op->src[1]->ne[0] >= 3 &&
+                op->src[2]->ne[0] == 24 &&
+                ggml_is_contiguous_rows(op->src[0]) &&
+                ggml_is_contiguous_rows(op->src[1]) &&
+                ggml_is_contiguous_rows(op->src[2]);
+        case GGML_OP_DSV4_HC_PRE:
+            return has_simdgroup_reduction &&
+                op->src[0]->type == GGML_TYPE_F32 &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->type         == GGML_TYPE_F32 &&
+                op->src[0]->ne[1] == 4 &&
+                op->src[1]->ne[0] == 4 &&
+                ggml_is_contiguous_rows(op->src[0]) &&
+                ggml_is_contiguous_rows(op->src[1]);
+        case GGML_OP_DSV4_HC_POST:
+            return has_simdgroup_reduction &&
+                op->src[0]->type == GGML_TYPE_F32 &&
+                op->src[1]->type == GGML_TYPE_F32 &&
+                op->src[2]->type == GGML_TYPE_F32 &&
+                op->src[3]->type == GGML_TYPE_F32 &&
+                op->type         == GGML_TYPE_F32 &&
+                op->src[1]->ne[1] == 4 &&
+                op->src[2]->ne[0] == 4 &&
+                op->src[3]->ne[0] == 4 &&
+                op->src[3]->ne[1] == 4 &&
+                ggml_is_contiguous_rows(op->src[0]) &&
+                ggml_is_contiguous_rows(op->src[1]) &&
+                ggml_is_contiguous_rows(op->src[2]) &&
+                ggml_is_contiguous_rows(op->src[3]);
         case GGML_OP_SSM_CONV:
         case GGML_OP_SSM_SCAN:
             return has_simdgroup_reduction;

ggml/src/ggml-metal/ggml-metal-impl.h

@@ -112,6 +112,13 @@
 #define OP_FLASH_ATTN_EXT_VEC_NQPSG 1
 #define OP_FLASH_ATTN_EXT_VEC_NCPSG 32
 
+#define OP_LIGHTNING_INDEXER_DK    128
+#define OP_LIGHTNING_INDEXER_NH     64
+#define OP_LIGHTNING_INDEXER_NHPTG   8
+#define OP_LIGHTNING_INDEXER_NKPSG   8
+#define OP_LIGHTNING_INDEXER_NSG     8
+#define OP_LIGHTNING_INDEXER_NBPTG   8
+
 #define OP_UNARY_NUM_SCALE      10
 #define OP_UNARY_NUM_FILL       11
 #define OP_UNARY_NUM_CLAMP      12
@@ -1171,6 +1178,66 @@ typedef struct {
     int64_t val;
 } ggml_metal_kargs_memset;
 
+typedef struct {
+    int32_t  n_kv;
+    int32_t  n_batch;
+    int32_t  mask_ne3;
+    uint64_t nb1;
+    uint64_t nb3;
+    uint64_t nbq1;
+    uint64_t nbq2;
+    uint64_t nbq3;
+    uint64_t nbk2;
+    uint64_t nbk3;
+    uint64_t nbw1;
+    uint64_t nbw3;
+    uint64_t nbm1;
+    uint64_t nbm3;
+} ggml_metal_kargs_lightning_indexer;
+
+typedef struct {
+    int32_t  n_tokens;
+    int32_t  n_iter;
+    uint64_t nb_m0;
+    uint64_t nb_m1;
+    uint64_t nb_s0;
+    uint64_t nb_b0;
+    uint64_t nb_d0;
+    uint64_t nb_d1;
+    uint64_t nb_d2;
+    float    eps;
+} ggml_metal_kargs_dsv4_hc_comb;
+
+typedef struct {
+    int32_t  n_embd;
+    int32_t  n_tokens;
+    uint64_t nb_x0;
+    uint64_t nb_x1;
+    uint64_t nb_x2;
+    uint64_t nb_w0;
+    uint64_t nb_w1;
+    uint64_t nb_d0;
+    uint64_t nb_d1;
+} ggml_metal_kargs_dsv4_hc_pre;
+
+typedef struct {
+    int32_t  n_embd;
+    int32_t  n_tokens;
+    uint64_t nb_x0;
+    uint64_t nb_x1;
+    uint64_t nb_r0;
+    uint64_t nb_r1;
+    uint64_t nb_r2;
+    uint64_t nb_p0;
+    uint64_t nb_p1;
+    uint64_t nb_c0;
+    uint64_t nb_c1;
+    uint64_t nb_c2;
+    uint64_t nb_d0;
+    uint64_t nb_d1;
+    uint64_t nb_d2;
+} ggml_metal_kargs_dsv4_hc_post;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -1222,4 +1289,8 @@ typedef struct {
     int64_t  np;
 } ggml_metal_kargs_opt_step_sgd;
 
+typedef struct {
+    int64_t ne;
+} ggml_metal_kargs_silu_back;
+
 #endif // GGML_METAL_IMPL

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -299,6 +299,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_unary(ctx, idx);
             } break;
+        case GGML_OP_SILU_BACK:
+            {
+                n_fuse = ggml_metal_op_silu_back(ctx, idx);
+            } break;
         case GGML_OP_GLU:
             {
                 n_fuse = ggml_metal_op_glu(ctx, idx);
@@ -316,6 +320,16 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_cumsum(ctx, idx);
             } break;
+        case GGML_OP_LIGHTNING_INDEXER:
+            {
+                n_fuse = ggml_metal_op_lightning_indexer(ctx, idx);
+            } break;
+        case GGML_OP_DSV4_HC_COMB:
+        case GGML_OP_DSV4_HC_PRE:
+        case GGML_OP_DSV4_HC_POST:
+            {
+                n_fuse = ggml_metal_op_dsv4_hc(ctx, idx);
+            } break;
         case GGML_OP_SOFT_MAX:
             {
                 n_fuse = ggml_metal_op_soft_max(ctx, idx);
@@ -1297,6 +1311,203 @@ int ggml_metal_op_diag(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_lightning_indexer(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_ASSERT(op->op == GGML_OP_LIGHTNING_INDEXER);
+
+    const ggml_tensor * q = op->src[0];
+    const ggml_tensor * k = op->src[1];
+    const ggml_tensor * w = op->src[2];
+    const ggml_tensor * m = op->src[3];
+
+    GGML_ASSERT(q->type == GGML_TYPE_F32);
+    GGML_ASSERT(k->type == GGML_TYPE_F32  ||
+                k->type == GGML_TYPE_F16  ||
+                k->type == GGML_TYPE_BF16 ||
+                k->type == GGML_TYPE_Q4_0 ||
+                k->type == GGML_TYPE_Q4_1 ||
+                k->type == GGML_TYPE_Q5_0 ||
+                k->type == GGML_TYPE_Q5_1 ||
+                k->type == GGML_TYPE_Q8_0);
+    GGML_ASSERT(w->type == GGML_TYPE_F32);
+    GGML_ASSERT(m->type == GGML_TYPE_F16);
+    GGML_ASSERT(op->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(q->ne[0] == OP_LIGHTNING_INDEXER_DK);
+    GGML_ASSERT(q->ne[1] == OP_LIGHTNING_INDEXER_NH);
+
+    ggml_metal_kargs_lightning_indexer args = {
+        /*.n_kv      =*/ (int32_t) k->ne[2],
+        /*.n_batch   =*/ (int32_t) q->ne[2],
+        /*.mask_ne3  =*/ (int32_t) m->ne[3],
+        /*.nb1       =*/ op->nb[1],
+        /*.nb3       =*/ op->nb[3],
+        /*.nbq1      =*/ q->nb[1],
+        /*.nbq2      =*/ q->nb[2],
+        /*.nbq3      =*/ q->nb[3],
+        /*.nbk2      =*/ k->nb[2],
+        /*.nbk3      =*/ k->nb[3],
+        /*.nbw1      =*/ w->nb[1],
+        /*.nbw3      =*/ w->nb[3],
+        /*.nbm1      =*/ m->nb[1],
+        /*.nbm3      =*/ m->nb[3],
+    };
+
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(q),  1);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(k),  2);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(w),  3);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(m),  4);
+    ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op), 5);
+
+    const int nsg   = OP_LIGHTNING_INDEXER_NSG;
+    const int nkptg = OP_LIGHTNING_INDEXER_NKPSG*nsg;
+    const int nbptg = OP_LIGHTNING_INDEXER_NBPTG;
+
+    auto pipeline = ggml_metal_library_get_pipeline_lightning_indexer(ctx->lib, op);
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes(enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_dispatch_threadgroups(enc,
+            (k->ne[2] + nkptg - 1)/nkptg,
+            (q->ne[2] + nbptg - 1)/nbptg,
+            q->ne[3], 32, nsg, 1);
+
+    return 1;
+}
+
+int ggml_metal_op_dsv4_hc(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_encoder_t enc = ctx->enc;
+    auto pipeline = ggml_metal_library_get_pipeline_dsv4_hc(ctx->lib, op->op);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+
+    switch (op->op) {
+        case GGML_OP_DSV4_HC_COMB:
+            {
+                const ggml_tensor * mixes = op->src[0];
+                const ggml_tensor * scale = op->src[1];
+                const ggml_tensor * base  = op->src[2];
+
+                GGML_ASSERT(mixes->type == GGML_TYPE_F32);
+                GGML_ASSERT(scale->type == GGML_TYPE_F32);
+                GGML_ASSERT(base->type  == GGML_TYPE_F32);
+                GGML_ASSERT(op->type    == GGML_TYPE_F32);
+                GGML_ASSERT(mixes->ne[0] == 24);
+                GGML_ASSERT(op->ne[0] == 4 && op->ne[1] == 4);
+
+                ggml_metal_kargs_dsv4_hc_comb args = {
+                    /*.n_tokens =*/ (int32_t) mixes->ne[1],
+                    /*.n_iter   =*/ ggml_get_op_params_i32(op, 1),
+                    /*.nb_m0    =*/ mixes->nb[0],
+                    /*.nb_m1    =*/ mixes->nb[1],
+                    /*.nb_s0    =*/ scale->nb[0],
+                    /*.nb_b0    =*/ base->nb[0],
+                    /*.nb_d0    =*/ op->nb[0],
+                    /*.nb_d1    =*/ op->nb[1],
+                    /*.nb_d2    =*/ op->nb[2],
+                    /*.eps      =*/ ggml_get_op_params_f32(op, 0),
+                };
+
+                ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(mixes), 1);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(scale), 2);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(base),  3);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),    4);
+
+                // One SIMDgroup owns one 4x4 Sinkhorn matrix. Packing up to four
+                // independent tokens per threadgroup keeps both decode and prompt
+                // dispatches compact without any threadgroup-memory synchronization.
+                const int nsg = std::min(4, args.n_tokens);
+                ggml_metal_encoder_dispatch_threadgroups(
+                        enc, (args.n_tokens + nsg - 1)/nsg, 1, 1, 32, nsg, 1);
+            } break;
+        case GGML_OP_DSV4_HC_PRE:
+            {
+                const ggml_tensor * x       = op->src[0];
+                const ggml_tensor * weights = op->src[1];
+
+                GGML_ASSERT(x->type       == GGML_TYPE_F32);
+                GGML_ASSERT(weights->type == GGML_TYPE_F32);
+                GGML_ASSERT(op->type      == GGML_TYPE_F32);
+                GGML_ASSERT(x->ne[1] == 4);
+
+                ggml_metal_kargs_dsv4_hc_pre args = {
+                    /*.n_embd   =*/ (int32_t) x->ne[0],
+                    /*.n_tokens =*/ (int32_t) x->ne[2],
+                    /*.nb_x0    =*/ x->nb[0],
+                    /*.nb_x1    =*/ x->nb[1],
+                    /*.nb_x2    =*/ x->nb[2],
+                    /*.nb_w0    =*/ weights->nb[0],
+                    /*.nb_w1    =*/ weights->nb[1],
+                    /*.nb_d0    =*/ op->nb[0],
+                    /*.nb_d1    =*/ op->nb[1],
+                };
+
+                ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(x),       1);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(weights), 2);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),      3);
+
+                const int n_tiles = (args.n_embd + 31)/32;
+                const int nsg = std::min(4, n_tiles);
+                ggml_metal_encoder_dispatch_threadgroups(
+                        enc, (n_tiles + nsg - 1)/nsg, args.n_tokens, 1, 32, nsg, 1);
+            } break;
+        case GGML_OP_DSV4_HC_POST:
+            {
+                const ggml_tensor * x        = op->src[0];
+                const ggml_tensor * residual = op->src[1];
+                const ggml_tensor * post     = op->src[2];
+                const ggml_tensor * comb     = op->src[3];
+
+                GGML_ASSERT(x->type        == GGML_TYPE_F32);
+                GGML_ASSERT(residual->type == GGML_TYPE_F32);
+                GGML_ASSERT(post->type     == GGML_TYPE_F32);
+                GGML_ASSERT(comb->type     == GGML_TYPE_F32);
+                GGML_ASSERT(op->type       == GGML_TYPE_F32);
+                GGML_ASSERT(residual->ne[1] == 4);
+
+                ggml_metal_kargs_dsv4_hc_post args = {
+                    /*.n_embd   =*/ (int32_t) x->ne[0],
+                    /*.n_tokens =*/ (int32_t) x->ne[1],
+                    /*.nb_x0    =*/ x->nb[0],
+                    /*.nb_x1    =*/ x->nb[1],
+                    /*.nb_r0    =*/ residual->nb[0],
+                    /*.nb_r1    =*/ residual->nb[1],
+                    /*.nb_r2    =*/ residual->nb[2],
+                    /*.nb_p0    =*/ post->nb[0],
+                    /*.nb_p1    =*/ post->nb[1],
+                    /*.nb_c0    =*/ comb->nb[0],
+                    /*.nb_c1    =*/ comb->nb[1],
+                    /*.nb_c2    =*/ comb->nb[2],
+                    /*.nb_d0    =*/ op->nb[0],
+                    /*.nb_d1    =*/ op->nb[1],
+                    /*.nb_d2    =*/ op->nb[2],
+                };
+
+                ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(x),        1);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(residual), 2);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(post),     3);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(comb),     4);
+                ggml_metal_encoder_set_buffer(enc, ggml_metal_get_buffer_id(op),       5);
+
+                const int n_tiles = (args.n_embd + 31)/32;
+                const int nsg = std::min(4, n_tiles);
+                ggml_metal_encoder_dispatch_threadgroups(
+                        enc, (n_tiles + nsg - 1)/nsg, args.n_tokens, 1, 32, nsg, 1);
+            } break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+
+    return 1;
+}
+
 int ggml_metal_op_soft_max(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 
@@ -3197,9 +3408,6 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
     GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
     GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
 
-    GGML_ASSERT(op->src[0]->type == GGML_TYPE_F32);
-    GGML_ASSERT(op->src[1]->type == GGML_TYPE_F32);
-
     GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
     GGML_ASSERT(ggml_is_contiguous_rows(op->src[1]));
 
@@ -3339,6 +3547,36 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
     return n_fuse;
 }
 
+int ggml_metal_op_silu_back(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    auto pipeline = ggml_metal_library_get_pipeline_silu_back(lib, op);
+
+    const int64_t ne = ggml_nelements(op);
+
+    ggml_metal_kargs_silu_back args = {
+        /*.ne =*/ ne,
+    };
+
+    int arg_idx{0};
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), arg_idx++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op), arg_idx++);
+
+    const int nth = std::min<int64_t>(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne);
+    const int64_t n = (ne + nth - 1) / nth;
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, nth, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_l2_norm(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 

ggml/src/ggml-metal/ggml-metal-ops.h

@@ -54,6 +54,8 @@ int ggml_metal_op_cumsum            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_get_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_set_rows          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_diag              (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_lightning_indexer (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_dsv4_hc           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_soft_max          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_conv          (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_ssm_scan          (ggml_metal_op_t ctx, int idx);
@@ -70,6 +72,7 @@ int ggml_metal_op_mul_mat_id        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_add_id            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_flash_attn_ext    (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_bin               (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_silu_back         (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_l2_norm           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_group_norm        (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_norm              (ggml_metal_op_t ctx, int idx);

ggml/src/ggml-metal/ggml-metal.metal

@@ -1255,6 +1255,20 @@ template [[host_name("kernel_unary_f32_f32_4")]] kernel kernel_unary_t kernel_un
 template [[host_name("kernel_unary_f16_f16")]]   kernel kernel_unary_t kernel_unary_impl<half,   half,   float>;
 template [[host_name("kernel_unary_f16_f16_4")]] kernel kernel_unary_t kernel_unary_impl<half4,  half4,  float4>;
 
+kernel void kernel_silu_back_f32(
+        constant ggml_metal_kargs_silu_back & args,
+        device const float * dy,
+        device const float * x,
+        device       float * dx,
+        uint gid [[thread_position_in_grid]]) {
+    if (gid >= args.ne) {
+        return;
+    }
+
+    const float s = 1.0f / (1.0f + exp(-x[gid]));
+    dx[gid] = dy[gid] * s * (1.0f + x[gid] * (1.0f - s));
+}
+
 // OP: 0 - add, 1 - sub, 2 - mul, 3 - div
 constant short FC_bin_op [[function_constant(FC_BIN + 0)]];
 constant short FC_bin_f  [[function_constant(FC_BIN + 1)]];
@@ -1418,6 +1432,8 @@ typedef decltype(kernel_bin_fuse_impl<float, float, float>) kernel_bin_fuse_t;
 
 template [[host_name("kernel_bin_fuse_f32_f32_f32")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float,  float,  float>;
 template [[host_name("kernel_bin_fuse_f32_f32_f32_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<float4, float4, float4>;
+template [[host_name("kernel_bin_fuse_f16_f16_f16")]]   kernel kernel_bin_fuse_t kernel_bin_fuse_impl<half,   half,   half>;
+template [[host_name("kernel_bin_fuse_f16_f16_f16_4")]] kernel kernel_bin_fuse_t kernel_bin_fuse_impl<half4,  half4,  half4>;
 
 kernel void kernel_add_id(
         constant ggml_metal_kargs_add_id & args,
@@ -11278,3 +11294,310 @@ kernel void kernel_count_equal(
 typedef decltype(kernel_count_equal<int32_t>) kernel_count_equal_t;
 
 template [[host_name("kernel_count_equal_i32")]] kernel kernel_count_equal_t kernel_count_equal<int32_t>;
+
+template<
+    typename kd4x4_t,
+    short nl_k,
+    void (*deq_k)(device const kd4x4_t *, short, thread half4x4 &)>
+kernel void kernel_lightning_indexer(
+        constant ggml_metal_kargs_lightning_indexer & args,
+        device const char * q,
+        device const char * k,
+        device const char * w,
+        device const char * m,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiitg[[thread_index_in_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    constexpr short DK    = OP_LIGHTNING_INDEXER_DK;
+    constexpr short NH    = OP_LIGHTNING_INDEXER_NH;
+    constexpr short NHPTG = OP_LIGHTNING_INDEXER_NHPTG;
+    constexpr short NKPSG = OP_LIGHTNING_INDEXER_NKPSG;
+    constexpr short NSG   = OP_LIGHTNING_INDEXER_NSG;
+    constexpr short NBPTG = OP_LIGHTNING_INDEXER_NBPTG;
+
+    constexpr short DK4  = DK/4;
+    constexpr short DK8  = DK/8;
+    constexpr short DK16 = DK/16;
+
+    constexpr short NK  = NKPSG*NSG; // keys    per threadgroup
+    constexpr short NTG = 32*NSG;    // threads per threadgroup
+
+    const int i_stream = tgpig.z;
+    const int i_kv_0   = tgpig.x*NK;            // first key of this threadgroup
+    const int i_kv     = i_kv_0 + sgitg*NKPSG;  // first key of this simdgroup
+
+    threadgroup half4x4   sk4x4[NK*DK16];
+    threadgroup half    * sk = (threadgroup half *) sk4x4;
+
+    for (short i = tiitg; i < NK*DK16; i += NTG) {
+        const short ik  = i/DK16;
+        const short i16 = i%DK16;
+
+        half4x4 tmp;
+
+        if (i_kv_0 + ik < args.n_kv) {
+            device const kd4x4_t * kr = (device const kd4x4_t *) (k + (i_kv_0 + ik)*args.nbk2 + i_stream*args.nbk3);
+
+            deq_k(kr + i16/nl_k, i16%nl_k, tmp);
+        } else {
+            FOR_UNROLL (short j = 0; j < 4; ++j) {
+                tmp[j] = half4(0.0h);
+            }
+        }
+
+        sk4x4[i] = tmp;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // K tile of this simdgroup, transposed to [DK, NKPSG]
+    simdgroup_half8x8 mk[DK8];
+
+    FOR_UNROLL (short i = 0; i < DK8; ++i) {
+        simdgroup_load(mk[i], sk + sgitg*NKPSG*DK + 8*i, DK, 0, true);
+    }
+
+    threadgroup half4   sq4[NHPTG*DK4];
+    threadgroup half  * sq = (threadgroup half *) sq4;
+
+    threadgroup float sw [NHPTG];
+    threadgroup float sqk[NSG*NHPTG*NKPSG];
+
+    const int i_batch_0 = tgpig.y*NBPTG;
+    const int n_batch   = min((int) NBPTG, args.n_batch - i_batch_0);
+
+    for (short ib = 0; ib < n_batch; ++ib) {
+        const int i_batch = i_batch_0 + ib;
+
+        device const char * pq = q + i_batch*args.nbq2 + i_stream*args.nbq3;
+        device const char * pw = w + i_batch*args.nbw1 + i_stream*args.nbw3;
+
+        float score = 0.0f;
+
+        FOR_UNROLL (short i_head = 0; i_head < NH; i_head += NHPTG) {
+            // stage the Q tile [DK, NHPTG] and the (prescaled) head weights
+            for (short i = tiitg; i < NHPTG*DK4; i += NTG) {
+                const short ih = i/DK4;
+                const short i4 = i%DK4;
+
+                device const float4 * q4 = (device const float4 *) (pq + (i_head + ih)*args.nbq1);
+
+                sq4[ih*DK4 + i4] = half4(q4[i4]);
+            }
+
+            if (tiitg < NHPTG) {
+                sw[tiitg] = ((device const float *) pw)[i_head + tiitg];
+            }
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+
+            simdgroup_float8x8 mqk = make_filled_simdgroup_matrix<float, 8>(0.0f);
+
+            FOR_UNROLL (short i = 0; i < DK8; ++i) {
+                simdgroup_half8x8 mq;
+
+                simdgroup_load(mq, sq + 8*i, DK, 0, false);
+                simdgroup_multiply_accumulate(mqk, mq, mk[i], mqk);
+            }
+
+            threadgroup float * pqk = sqk + sgitg*NHPTG*NKPSG;
+
+            simdgroup_store(mqk, pqk, NKPSG, 0, false);
+            simdgroup_barrier(mem_flags::mem_threadgroup);
+
+            // one lane per key: ReLU, apply the head weight and accumulate over the head tile
+            if (tiisg < NKPSG) {
+                FOR_UNROLL (short ih = 0; ih < NHPTG; ++ih) {
+                    score += max(pqk[ih*NKPSG + tiisg], 0.0f)*sw[ih];
+                }
+            }
+
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+        }
+
+        if (tiisg < NKPSG) {
+            const int ik = i_kv + tiisg;
+            if (ik < args.n_kv) {
+                device const half  * pm = (device const half  *) (m   + i_batch*args.nbm1 + (i_stream % args.mask_ne3)*args.nbm3);
+                device       float * pd = (device       float *) (dst + i_batch*args.nb1  + i_stream*args.nb3);
+
+                pd[ik] = score + (float) pm[ik];
+            }
+        }
+    }
+}
+
+typedef decltype(kernel_lightning_indexer<half4x4, 1, dequantize_f16>) kernel_lightning_indexer_t;
+
+template [[host_name("kernel_lightning_indexer_f32")]]  kernel kernel_lightning_indexer_t kernel_lightning_indexer<float4x4, 1, dequantize_f32>;
+template [[host_name("kernel_lightning_indexer_f16")]]  kernel kernel_lightning_indexer_t kernel_lightning_indexer<half4x4,  1, dequantize_f16>;
+
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_lightning_indexer_bf16")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<bfloat4x4, 1, dequantize_bf16>;
+#endif
+
+template [[host_name("kernel_lightning_indexer_q4_0")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<block_q4_0, 2, dequantize_q4_0>;
+template [[host_name("kernel_lightning_indexer_q4_1")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<block_q4_1, 2, dequantize_q4_1>;
+template [[host_name("kernel_lightning_indexer_q5_0")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<block_q5_0, 2, dequantize_q5_0>;
+template [[host_name("kernel_lightning_indexer_q5_1")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<block_q5_1, 2, dequantize_q5_1>;
+template [[host_name("kernel_lightning_indexer_q8_0")]] kernel kernel_lightning_indexer_t kernel_lightning_indexer<block_q8_0, 2, dequantize_q8_0>;
+
+kernel void kernel_dsv4_hc_comb_f32(
+        constant ggml_metal_kargs_dsv4_hc_comb & args,
+        device const char * mixes,
+        device const char * scale,
+        device const char * base,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+    constexpr ushort comb_offset = 2*hc;
+
+    const int it = tgpig.x*ntg.y + sgitg;
+    if (it >= args.n_tokens) {
+        return;
+    }
+
+    float scale_lane = 0.0f;
+    if (tiisg == 0) {
+        scale_lane = *(device const float *) (scale + 2*args.nb_s0);
+    }
+    const float scale_comb = simd_shuffle(scale_lane, 0);
+
+    float v = 0.0f;
+    if (tiisg < hc*hc) {
+        v = *(device const float *) (mixes + (comb_offset + tiisg)*args.nb_m0 + it*args.nb_m1)*scale_comb
+          + *(device const float *) (base   + (comb_offset + tiisg)*args.nb_b0);
+    }
+
+    // Softmax across destinations (the four contiguous lanes for each source).
+    float vmax = max(v, simd_shuffle_xor(v, 1));
+    vmax = max(vmax, simd_shuffle_xor(vmax, 2));
+    v = exp(v - vmax);
+
+    float sum = v + simd_shuffle_xor(v, 1);
+    sum += simd_shuffle_xor(sum, 2);
+    v = v/sum + args.eps;
+
+    // Normalize columns: equal destination indices are four lanes apart.
+    sum = v + simd_shuffle_xor(v, 4);
+    sum += simd_shuffle_xor(sum, 8);
+    v /= sum + args.eps;
+
+    for (int i = 1; i < args.n_iter; ++i) {
+        sum = v + simd_shuffle_xor(v, 1);
+        sum += simd_shuffle_xor(sum, 2);
+        v /= sum + args.eps;
+
+        sum = v + simd_shuffle_xor(v, 4);
+        sum += simd_shuffle_xor(sum, 8);
+        v /= sum + args.eps;
+    }
+
+    if (tiisg < hc*hc) {
+        const ushort idst = tiisg & 3;
+        const ushort isrc = tiisg >> 2;
+        *(device float *) (dst + idst*args.nb_d0 + isrc*args.nb_d1 + it*args.nb_d2) = v;
+    }
+}
+
+kernel void kernel_dsv4_hc_pre_f32(
+        constant ggml_metal_kargs_dsv4_hc_pre & args,
+        device const char * x,
+        device const char * weights,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+
+    const int it = tgpig.y;
+    const int i0 = ((int) tgpig.x*ntg.y + sgitg)*32 + tiisg;
+
+    float weight_lane = 0.0f;
+    if (tiisg < hc) {
+        weight_lane = *(device const float *) (weights + tiisg*args.nb_w0 + it*args.nb_w1);
+    }
+
+    float w[hc];
+    FOR_UNROLL (ushort ih = 0; ih < hc; ++ih) {
+        w[ih] = simd_shuffle(weight_lane, ih);
+    }
+
+    if (i0 >= args.n_embd) {
+        return;
+    }
+
+    device const char * xb = x + i0*args.nb_x0 + it*args.nb_x2;
+    float result = 0.0f;
+    FOR_UNROLL (ushort ih = 0; ih < hc; ++ih) {
+        result = fma(*(device const float *) (xb + ih*args.nb_x1), w[ih], result);
+    }
+
+    *(device float *) (dst + i0*args.nb_d0 + it*args.nb_d1) = result;
+}
+
+kernel void kernel_dsv4_hc_post_f32(
+        constant ggml_metal_kargs_dsv4_hc_post & args,
+        device const char * x,
+        device const char * residual,
+        device const char * post,
+        device const char * comb,
+        device       char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    constexpr ushort hc = 4;
+
+    const int it = tgpig.y;
+    const int i0 = ((int) tgpig.x*ntg.y + sgitg)*32 + tiisg;
+
+    float coeff_lane = 0.0f;
+    if (tiisg < hc) {
+        coeff_lane = *(device const float *) (post + tiisg*args.nb_p0 + it*args.nb_p1);
+    } else if (tiisg < hc + hc*hc) {
+        const ushort idx  = tiisg - hc;
+        const ushort idst = idx & 3;
+        const ushort isrc = idx >> 2;
+        coeff_lane = *(device const float *) (comb + idst*args.nb_c0 + isrc*args.nb_c1 + it*args.nb_c2);
+    }
+
+    float post_reg[hc];
+    float comb_reg[hc][hc];
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        post_reg[idst] = simd_shuffle(coeff_lane, idst);
+    }
+    FOR_UNROLL (ushort isrc = 0; isrc < hc; ++isrc) {
+        FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+            comb_reg[isrc][idst] = simd_shuffle(coeff_lane, hc + idst + hc*isrc);
+        }
+    }
+
+    if (i0 >= args.n_embd) {
+        return;
+    }
+
+    const float xv = *(device const float *) (x + i0*args.nb_x0 + it*args.nb_x1);
+    float result[hc];
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        result[idst] = xv*post_reg[idst];
+    }
+
+    device const char * rb = residual + i0*args.nb_r0 + it*args.nb_r2;
+    FOR_UNROLL (ushort isrc = 0; isrc < hc; ++isrc) {
+        const float rv = *(device const float *) (rb + isrc*args.nb_r1);
+        FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+            result[idst] = fma(rv, comb_reg[isrc][idst], result[idst]);
+        }
+    }
+
+    FOR_UNROLL (ushort idst = 0; idst < hc; ++idst) {
+        *(device float *) (dst + i0*args.nb_d0 + idst*args.nb_d1 + it*args.nb_d2) = result[idst];
+    }
+}

ggml/src/ggml-opencl/ggml-opencl.cpp

@@ -7065,7 +7065,7 @@ static inline bool use_flat_gemv_for_large_m_q4_K(const ggml_tensor *tensor) {
     return tensor->ne[1] >= 32768 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }
 
-static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_tensor *tensor) {
+static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
     // gemv_noshuffle variant perf drops for large M, use flat variant for large M.
     // threshold is well above typical hidden/FFN dims, but below typical vocab sizes.
     // q6_K flat gemv is worse for smaller K; 2048 seems to be a reasonable threshold.
@@ -7083,7 +7083,15 @@ static inline bool use_flat_gemv_for_large_m_q6_K(const ggml_tensor *tensor) {
     if ((tensor->ne[1] % 128 != 0) && tensor->ne[2] == 1 && tensor->ne[3] == 1) {
         return true;
     }
-    return tensor->ne[1] >= 32768 && tensor->ne[0] >= 2048 && tensor->ne[2] == 1 && tensor->ne[3] == 1;
+
+    // The gemv_noshuffle slowdown tracks TOTAL weight size, not ne0 alone; ne0 >= 2048 is a
+    // proxy for "large weight" that misses a narrow-hidden vocab-scale lm_head.
+    // Add a direct size escape so such weights also take the flat path, without changing
+    // which weights ne0 >= 2048 already routes there.
+    // The size escape is not taken on the A7X since its compiler miscompiles the flat K-quant GEMV
+    return tensor->ne[1] >= 32768
+        && (tensor->ne[0] >= 2048 || (backend_ctx->adreno_gen != ADRENO_GPU_GEN::A7X && ggml_nbytes(tensor) >= (256ull << 20)))
+        && tensor->ne[2] == 1 && tensor->ne[3] == 1;
 }
 
 static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
@@ -9403,7 +9411,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         cl_kernel kernel;
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
         kernel = backend_ctx->kernel_convert_block_q6_K;
-        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(backend_ctx, tensor)) {
             kernel = backend_ctx->kernel_convert_block_q6_K_noshuffle;
         }
 #else
@@ -9436,7 +9444,7 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         tensor->extra  = extra;
 
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
-        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(backend_ctx, tensor)) {
             cl_int M = tensor->ne[1];   // ne01
             cl_int K = tensor->ne[0];   // ne00
 
@@ -10473,7 +10481,7 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             CL_CHECK(clReleaseMemObject(data_device));
             return;
         }
-        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(tensor)) {
+        if (use_adreno_kernels(backend_ctx, tensor) && !use_flat_gemv_for_large_m_q6_K(backend_ctx, tensor)) {
             static ggml_cl_buffer buf_trans_ql;
             static ggml_cl_buffer buf_trans_qh;
             static ggml_cl_buffer buf_trans_s;
@@ -18895,7 +18903,7 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
         }
 
         // q6_K x fp32
-        if (src0t == GGML_TYPE_Q6_K && src1t == GGML_TYPE_F32 && !use_flat_gemv_for_large_m_q6_K(src0)) {
+        if (src0t == GGML_TYPE_Q6_K && src1t == GGML_TYPE_F32 && !use_flat_gemv_for_large_m_q6_K(backend_ctx, src0)) {
             ggml_cl_mul_mat_q6_K_f32_adreno(backend, src0, src1, dst);
             return;
         }
@@ -24260,7 +24268,7 @@ static void ggml_cl_glu(ggml_backend_t backend, const ggml_tensor * src0, const
     }
 
     const size_t nrows = ggml_nrows(src0);
-    size_t nth = 512;
+    size_t nth = backend_ctx->max_workgroup_size < 512 ? backend_ctx->max_workgroup_size : 512;
     size_t global_work_size[] = {nrows*nth, 1, 1};
     size_t local_work_size[] = {nth, 1, 1};
 

ggml/src/ggml-sycl/concat.cpp

@@ -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);

ggml/src/ggml-sycl/fattn-onednn.cpp

@@ -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;

ggml/src/ggml-sycl/fattn.cpp

@@ -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) {

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -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],

ggml/src/ggml-vulkan/vulkan-shaders/gla.comp

@@ -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];
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -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"}}));

gguf-py/gguf/constants.py

@@ -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,
@@ -2329,7 +2518,13 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.SSM_NORM,
         MODEL_TENSOR.SSM_IN,
         MODEL_TENSOR.SSM_BETA_ALPHA,
-        MODEL_TENSOR.SSM_OUT
+        MODEL_TENSOR.SSM_OUT,
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
     ],
     MODEL_ARCH.QWEN3VL: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -3215,6 +3410,13 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
         MODEL_TENSOR.FFN_EXP_PROBS_B,
+        # NextN/MTP tensors
+        MODEL_TENSOR.NEXTN_EH_PROJ,
+        MODEL_TENSOR.NEXTN_EMBED_TOKENS,
+        MODEL_TENSOR.NEXTN_ENORM,
+        MODEL_TENSOR.NEXTN_HNORM,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD,
+        MODEL_TENSOR.NEXTN_SHARED_HEAD_NORM,
     ],
     MODEL_ARCH.DEEPSEEK2OCR: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -4634,6 +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
@@ -4908,6 +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"

gguf-py/gguf/gguf_reader.py

@@ -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

gguf-py/gguf/gguf_writer.py

@@ -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)
 

gguf-py/gguf/scripts/gguf_convert_endian.py

@@ -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,
 }

gguf-py/gguf/tensor_mapping.py

@@ -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: (

gguf-py/tests/test_gguf_reader_validation.py

@@ -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)

include/llama.h

@@ -1424,19 +1424,19 @@ 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)
-                               float   penalty_repeat,   // 1.0 = disabled
-                               float   penalty_freq,     // 0.0 = disabled
-                               float   penalty_present); // 0.0 = disabled
+                             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
 
     ///  @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);
 

models/templates/deepseek-ai-DeepSeek-V4-Flash-0731.jinja

@@ -0,0 +1,140 @@
+{%- if not add_generation_prompt is defined -%}
+  {%- set add_generation_prompt = false -%}
+{%- endif -%}
+{%- if not thinking is defined -%}
+  {%- if enable_thinking is defined -%}
+    {%- set thinking = enable_thinking -%}
+  {%- else -%}
+    {%- set thinking = false -%}
+  {%- endif -%}
+{%- endif -%}
+{%- if not drop_thinking is defined -%}
+  {%- set drop_thinking = true -%}
+{%- endif -%}
+{%- set dsml_token = '｜DSML｜' -%}
+{%- set thinking_start_token = '<think>' -%}
+{%- set thinking_end_token = '</think>' -%}
+{%- set reasoning_effort_high = 'Reasoning Effort: Absolute maximum with no shortcuts permitted.\nYou MUST be very thorough in your thinking and comprehensively decompose the problem to resolve the root cause, rigorously stress-testing your logic against all potential paths, edge cases, and adversarial scenarios.\nExplicitly write out your entire deliberation process, documenting every intermediate step, considered alternative, and rejected hypothesis to ensure absolutely no assumption is left unchecked.\n\n' -%}
+{%- set reasoning_effort_max = 'Reasoning Effort: Beyond maximum — exhaustive, relentless, and uncompromising.\nYou MUST reason with the utmost depth and rigor, leaving absolutely nothing to chance: exhaustively decompose the problem into its most fundamental components, trace every causal chain to its root, and resolve the underlying cause rather than any surface symptom.\nDo not stop reasoning until you have independently verified the solution from multiple angles and are certain that no assumption remains unchecked and no error remains undiscovered.\n\n' -%}
+{%- set response_format_template = '## Response Format:\n\nYou MUST strictly adhere to the following schema to reply:\n' -%}
+{%- set has_tools = false -%}
+{%- set tools_header = '## Tools\n\nYou have access to a set of tools to help answer the user\'s question. You can invoke tools by writing a "<' + dsml_token + 'tool_calls>" block like the following:\n\n<' + dsml_token + 'tool_calls>\n<' + dsml_token + 'invoke name="$TOOL_NAME">\n<' + dsml_token + 'parameter name="$PARAMETER_NAME" string="true|false">$PARAMETER_VALUE</' + dsml_token + 'parameter>\n...\n</' + dsml_token + 'invoke>\n<' + dsml_token + 'invoke name="$TOOL_NAME2">\n...\n</' + dsml_token + 'invoke>\n</' + dsml_token + 'tool_calls>\n\nString parameters should be specified as is and set `string="true"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string="false"`.\n\nIf thinking_mode is enabled (triggered by ' + thinking_start_token + '), you MUST output your complete reasoning inside ' + thinking_start_token + '...' + thinking_end_token + ' BEFORE any tool calls or final response.\n\nOtherwise, output directly after ' + thinking_end_token + ' with tool calls or final response.\n\n### Available Tool Schemas\n\n' -%}
+{%- set tools_footer = '\nYou MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.\n' -%}
+{%- set ns = namespace(system_prompt='', is_first_sp=true, has_tool_calls=false) -%}
+{%- for message in messages -%}
+  {%- if message['role'] == 'system' -%}
+    {%- if ns.is_first_sp -%}
+      {%- set ns.system_prompt = ns.system_prompt + (message['content'] or '') -%}
+      {%- set ns.is_first_sp = false -%}
+    {%- else -%}
+      {%- set ns.system_prompt = ns.system_prompt + '\n\n' + (message['content'] or '') -%}
+    {%- endif -%}
+  {%- endif -%}
+{%- endfor -%}
+{%- if tools is defined and tools -%}
+  {%- set has_tools = true -%}
+  {%- set ts = namespace(schemas='') -%}
+  {%- for tool in tools -%}
+    {%- if tool['type'] == 'function' -%}
+      {%- set ts.schemas = ts.schemas + (tool['function'] | tojson) + '\n' -%}
+    {%- endif -%}
+  {%- endfor -%}
+  {%- if ns.system_prompt -%}
+    {%- set ns.system_prompt = ns.system_prompt + '\n\n' + tools_header + ts.schemas + tools_footer -%}
+  {%- else -%}
+    {%- set ns.system_prompt = tools_header + ts.schemas + tools_footer -%}
+  {%- endif -%}
+{%- endif -%}
+{%- if response_format is defined -%}
+  {%- if ns.system_prompt -%}
+    {%- set ns.system_prompt = ns.system_prompt + '\n\n' -%}
+  {%- endif -%}
+  {%- set ns.system_prompt = ns.system_prompt + response_format_template + (response_format | tojson) -%}
+{%- endif -%}
+{{- bos_token -}}
+{%- if messages and thinking and reasoning_effort is defined and reasoning_effort == 'high' -%}
+  {{- reasoning_effort_high -}}
+{%- elif messages and thinking and reasoning_effort is defined and reasoning_effort == 'max' -%}
+  {{- reasoning_effort_max -}}
+{%- endif -%}
+{{- ns.system_prompt -}}
+{%- set last_user_idx = namespace(value=-1) -%}
+{%- for message in messages -%}
+  {%- if message['role'] == 'user' or message['role'] == 'developer' or message['role'] == 'tool' -%}
+    {%- set last_user_idx.value = loop.index0 -%}
+  {%- endif -%}
+{%- endfor -%}
+{%- set state = namespace(in_user=false) -%}
+{%- for message in messages -%}
+  {%- if message['role'] == 'tool' -%}
+    {%- set ns.has_tool_calls = true -%}
+  {%- endif -%}
+{%- endfor -%}
+{%- for message in messages -%}
+  {%- if message['role'] == 'user' or message['role'] == 'developer' -%}
+    {%- if state.in_user -%}
+      {{- '\n\n' -}}
+    {%- else -%}
+      {{- '<｜User｜>' -}}
+      {%- set state.in_user = true -%}
+    {%- endif -%}
+    {{- message['content'] or '' -}}
+  {%- elif message['role'] == 'tool' -%}
+    {%- if state.in_user -%}
+      {{- '\n\n' -}}
+    {%- else -%}
+      {{- '<｜User｜>' -}}
+      {%- set state.in_user = true -%}
+    {%- endif -%}
+    {{- '<tool_result>' + (message['content'] or '') + '</tool_result>' -}}
+  {%- elif message['role'] == 'assistant' -%}
+    {%- set state.in_user = false -%}
+    {{- '<｜Assistant｜>' -}}
+    {%- set is_after_last_user = loop.index0 > last_user_idx.value -%}
+    {%- set keep_reasoning = thinking and ((not drop_thinking) or has_tools or is_after_last_user or ns.has_tool_calls) -%}
+    {%- if keep_reasoning -%}
+      {{- thinking_start_token -}}
+      {%- if message['reasoning_content'] is defined and message['reasoning_content'] -%}
+        {{- message['reasoning_content'] -}}
+      {%- endif -%}
+      {{- thinking_end_token -}}
+    {%- else -%}
+      {{- thinking_end_token -}}
+    {%- endif -%}
+    {%- if message['content'] is defined and message['content'] -%}
+      {{- message['content'] -}}
+    {%- endif -%}
+    {%- if message['tool_calls'] -%}
+      {{- '\n\n<' + dsml_token + 'tool_calls>\n' -}}
+      {%- for tool in message['tool_calls'] -%}
+        {%- set func = tool['function'] -%}
+        {{- '<' + dsml_token + 'invoke name="' + func['name'] + '">\n' -}}
+        {%- set args = func['arguments'] -%}
+        {%- if args is string -%}
+          {%- set args = args | from_json -%}
+        {%- endif -%}
+        {%- for key, val in args.items() -%}
+          {%- if val is string -%}
+            {{- '<' + dsml_token + 'parameter name="' + key + '" string="true">' + val + '</' + dsml_token + 'parameter>\n' -}}
+          {%- else -%}
+            {{- '<' + dsml_token + 'parameter name="' + key + '" string="false">' + (val | tojson) + '</' + dsml_token + 'parameter>\n' -}}
+          {%- endif -%}
+        {%- endfor -%}
+        {%- if not args -%}
+          {{- '\n' -}}
+        {%- endif -%}
+        {{- '</' + dsml_token + 'invoke>\n' -}}
+      {%- endfor -%}
+      {{- '</' + dsml_token + 'tool_calls>' -}}
+    {%- endif -%}
+    {{- '<｜end▁of▁sentence｜>' -}}
+  {%- endif -%}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+  {{- '<｜Assistant｜>' -}}
+  {%- if thinking -%}
+    {{- thinking_start_token -}}
+  {%- else -%}
+    {{- thinking_end_token -}}
+  {%- endif -%}
+{%- endif -%}

models/templates/deepseek-ai-DeepSeek-V4.jinja

@@ -9,11 +9,14 @@
   {%- endif -%}
 {%- endif -%}
 {%- if not drop_thinking is defined -%}
-  {%- set drop_thinking = false -%}
+  {%- set drop_thinking = true -%}
 {%- endif -%}
 {%- set dsml_token = '｜DSML｜' -%}
 {%- set thinking_start_token = '<think>' -%}
 {%- set thinking_end_token = '</think>' -%}
+{%- set reasoning_effort_max = 'Reasoning Effort: Absolute maximum with no shortcuts permitted.\nYou MUST be very thorough in your thinking and comprehensively decompose the problem to resolve the root cause, rigorously stress-testing your logic against all potential paths, edge cases, and adversarial scenarios.\nExplicitly write out your entire deliberation process, documenting every intermediate step, considered alternative, and rejected hypothesis to ensure absolutely no assumption is left unchecked.\n\n' -%}
+{%- set response_format_template = '## Response Format:\n\nYou MUST strictly adhere to the following schema to reply:\n' -%}
+{%- set has_tools = false -%}
 {%- set tools_header = '## Tools\n\nYou have access to a set of tools to help answer the user\'s question. You can invoke tools by writing a "<' + dsml_token + 'tool_calls>" block like the following:\n\n<' + dsml_token + 'tool_calls>\n<' + dsml_token + 'invoke name="$TOOL_NAME">\n<' + dsml_token + 'parameter name="$PARAMETER_NAME" string="true|false">$PARAMETER_VALUE</' + dsml_token + 'parameter>\n...\n</' + dsml_token + 'invoke>\n<' + dsml_token + 'invoke name="$TOOL_NAME2">\n...\n</' + dsml_token + 'invoke>\n</' + dsml_token + 'tool_calls>\n\nString parameters should be specified as is and set `string="true"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string="false"`.\n\nIf thinking_mode is enabled (triggered by ' + thinking_start_token + '), you MUST output your complete reasoning inside ' + thinking_start_token + '...' + thinking_end_token + ' BEFORE any tool calls or final response.\n\nOtherwise, output directly after ' + thinking_end_token + ' with tool calls or final response.\n\n### Available Tool Schemas\n\n' -%}
 {%- set tools_footer = '\nYou MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.\n' -%}
 {%- set ns = namespace(system_prompt='', is_first_sp=true, has_tool_calls=false) -%}
@@ -28,6 +31,7 @@
   {%- endif -%}
 {%- endfor -%}
 {%- if tools is defined and tools -%}
+  {%- set has_tools = true -%}
   {%- set ts = namespace(schemas='') -%}
   {%- for tool in tools -%}
     {%- if tool['type'] == 'function' -%}
@@ -40,7 +44,16 @@
     {%- set ns.system_prompt = tools_header + ts.schemas + tools_footer -%}
   {%- endif -%}
 {%- endif -%}
+{%- if response_format is defined -%}
+  {%- if ns.system_prompt -%}
+    {%- set ns.system_prompt = ns.system_prompt + '\n\n' -%}
+  {%- endif -%}
+  {%- set ns.system_prompt = ns.system_prompt + response_format_template + (response_format | tojson) -%}
+{%- endif -%}
 {{- bos_token -}}
+{%- if messages and thinking and reasoning_effort is defined and reasoning_effort == 'max' -%}
+  {{- reasoning_effort_max -}}
+{%- endif -%}
 {{- ns.system_prompt -}}
 {%- set last_user_idx = namespace(value=-1) -%}
 {%- for message in messages -%}
@@ -75,8 +88,8 @@
     {%- set state.in_user = false -%}
     {{- '<｜Assistant｜>' -}}
     {%- set is_after_last_user = loop.index0 > last_user_idx.value -%}
-    {%- set retain_reasoning = (not drop_thinking) or (is_after_last_user or ns.has_tool_calls) -%}
-    {%- if retain_reasoning and thinking -%}
+    {%- set keep_reasoning = thinking and ((not drop_thinking) or has_tools or is_after_last_user or ns.has_tool_calls) -%}
+    {%- if keep_reasoning -%}
       {{- thinking_start_token -}}
       {%- if message['reasoning_content'] is defined and message['reasoning_content'] -%}
         {{- message['reasoning_content'] -}}
@@ -104,6 +117,9 @@
             {{- '<' + dsml_token + 'parameter name="' + key + '" string="false">' + (val | tojson) + '</' + dsml_token + 'parameter>\n' -}}
           {%- endif -%}
         {%- endfor -%}
+        {%- if not args -%}
+          {{- '\n' -}}
+        {%- endif -%}
         {{- '</' + dsml_token + 'invoke>\n' -}}
       {%- endfor -%}
       {{- '</' + dsml_token + 'tool_calls>' -}}
@@ -118,4 +134,4 @@
   {%- else -%}
     {{- thinking_end_token -}}
   {%- endif -%}
-{%- endif -%}
+{%- endif -%}

scripts/sync-ggml.last

@@ -1 +1 @@
-06ca97616793248fadb410ea8d69c7511b2005e4
+90951f99af1fbebef3fbdd58ff5b8715b0bb9c43

scripts/sync_vendor.py

@@ -5,7 +5,7 @@ import os
 import sys
 import subprocess
 
-HTTPLIB_VERSION = "refs/tags/v0.51.0"
+HTTPLIB_VERSION = "refs/tags/v0.52.0"
 
 vendor = {
     "https://github.com/nlohmann/json/releases/latest/download/json.hpp":     "vendor/nlohmann/json.hpp",
@@ -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([

scripts/ui-assets.cmake

@@ -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()

skills/code-review/SKILL.md

@@ -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)
 

src/CMakeLists.txt

@@ -25,6 +25,7 @@ add_library(llama
             llama-kv-cache.cpp
             llama-kv-cache-iswa.cpp
             llama-kv-cache-dsa.cpp
+            llama-kv-cache-msa.cpp
             llama-kv-cache-dsv4.cpp
             llama-memory.cpp
             llama-memory-hybrid.cpp

src/llama-arch.cpp

@@ -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;

src/llama-arch.h

@@ -149,6 +149,7 @@ enum llm_arch {
     LLM_ARCH_MINIMAX_M3,
     LLM_ARCH_DFLASH,
     LLM_ARCH_NANBEIGE,
+    LLM_ARCH_QWEN3TTS,
     LLM_ARCH_UNKNOWN,
 };
 

src/llama-ext.h

@@ -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);

src/llama-graph.cpp

@@ -8,6 +8,7 @@
 #include "llama-kv-cache.h"
 #include "llama-kv-cache-iswa.h"
 #include "llama-kv-cache-dsa.h"
+#include "llama-kv-cache-msa.h"
 #include "llama-kv-cache-dsv4.h"
 #include "llama-memory-hybrid.h"
 #include "llama-memory-hybrid-iswa.h"
@@ -518,6 +519,40 @@ bool llm_graph_input_attn_k::can_reuse(const llm_graph_params & params) {
     return res;
 }
 
+llm_graph_input_attn_kv_msa::llm_graph_input_attn_kv_msa(
+        const llama_hparams & hparams,
+        const llama_cparams & cparams,
+        const llama_kv_cache_msa_context * mctx) :
+    llm_graph_input_attn_kv(hparams, cparams, mctx->get_base()),
+    mctx_msa(mctx) {
+}
+
+void llm_graph_input_attn_kv_msa::set_input(const llama_ubatch * ubatch) {
+    llm_graph_input_attn_kv::set_input(ubatch);
+
+    if (self_k_idxs_idx) {
+        mctx_msa->get_idx()->set_input_k_idxs(self_k_idxs_idx, ubatch);
+    }
+}
+
+bool llm_graph_input_attn_kv_msa::can_reuse(const llm_graph_params & params) {
+    mctx_msa = static_cast<const llama_kv_cache_msa_context *>(params.mctx);
+
+    // the parent class operates on the base cache context
+    this->mctx = mctx_msa->get_base();
+
+    bool res = true;
+
+    res &= self_k_idxs->ne[0] == params.ubatch.n_tokens;
+    if (self_k_idxs_idx) {
+        res &= self_k_idxs_idx->ne[0] == params.ubatch.n_tokens;
+    }
+
+    res &= can_reuse_kq_mask(self_kq_mask, this->mctx, params.ubatch, params.cparams);
+
+    return res;
+}
+
 void llm_graph_input_attn_k_dsa::set_input(const llama_ubatch * ubatch) {
     mctx->get_mla()->set_input_k_idxs(self_k_idxs_mla, ubatch);
 
@@ -3187,6 +3222,34 @@ llm_graph_input_attn_k_dsa * llm_graph_context::build_attn_inp_k_dsa() const {
     return (llm_graph_input_attn_k_dsa *) res->add_input(std::move(inp));
 }
 
+llm_graph_input_attn_kv_msa * llm_graph_context::build_attn_inp_kv_msa(bool msa_enabled) const {
+    const auto * mctx_cur = static_cast<const llama_kv_cache_msa_context *>(mctx);
+
+    auto inp = std::make_unique<llm_graph_input_attn_kv_msa>(hparams, cparams, mctx_cur);
+
+    const auto * mctx_base = mctx_cur->get_base();
+    const auto * mctx_idx  = mctx_cur->get_idx();
+
+    {
+        GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA");
+
+        inp->self_k_idxs = mctx_base->build_input_k_idxs(ctx0, ubatch);
+        inp->self_v_idxs = mctx_base->build_input_v_idxs(ctx0, ubatch);
+
+        inp->self_kq_mask = build_attn_inp_kq_mask(ctx0, mctx_base, ubatch, cparams);
+        inp->self_kq_mask_cnv = inp->self_kq_mask;
+    }
+
+    inp->self_k_rot = mctx_base->build_input_k_rot(ctx0);
+    inp->self_v_rot = mctx_base->build_input_v_rot(ctx0);
+
+    if (msa_enabled) {
+        inp->self_k_idxs_idx = mctx_idx->build_input_k_idxs(ctx0, ubatch);
+    }
+
+    return (llm_graph_input_attn_kv_msa *) res->add_input(std::move(inp));
+}
+
 // TODO: maybe separate the inner implementation into a separate function
 //       like with the non-sliding window equivalent
 //       once sliding-window hybrid caches are a thing.

src/llama-graph.h

@@ -23,6 +23,7 @@ struct llama_memory_context_i;
 
 class llama_kv_cache_context;
 class llama_kv_cache_dsa_context;
+class llama_kv_cache_msa_context;
 class llama_kv_cache_dsv4_raw_context;
 class llama_kv_cache_dsv4_context;
 class llama_kv_cache_iswa_context;
@@ -425,6 +426,26 @@ public:
     const llama_kv_cache_dsa_context * mctx;
 };
 
+// standard K/V attention input against the base cache, plus destination indices for the indexer key cache
+class llm_graph_input_attn_kv_msa : public llm_graph_input_attn_kv {
+public:
+    llm_graph_input_attn_kv_msa(
+            const llama_hparams & hparams,
+            const llama_cparams & cparams,
+            const llama_kv_cache_msa_context * mctx);
+    ~llm_graph_input_attn_kv_msa() = default;
+
+    void set_input(const llama_ubatch * ubatch) override;
+
+    bool can_reuse(const llm_graph_params & params) override;
+
+    ggml_tensor * get_k_idxs_idx() const { return self_k_idxs_idx; }
+
+    ggml_tensor * self_k_idxs_idx = nullptr; // I64 [n_batch]
+
+    const llama_kv_cache_msa_context * mctx_msa;
+};
+
 class llm_graph_input_attn_kv_iswa : public llm_graph_input_i {
 public:
     llm_graph_input_attn_kv_iswa(
@@ -1169,6 +1190,8 @@ struct llm_graph_context {
 
     llm_graph_input_attn_k_dsa * build_attn_inp_k_dsa() const;
 
+    llm_graph_input_attn_kv_msa * build_attn_inp_kv_msa(bool msa_enabled) const;
+
     ggml_tensor * build_attn(
             llm_graph_input_attn_k_dsa * inp,
             ggml_tensor * wo,

src/llama-hparams.cpp

@@ -180,16 +180,6 @@ uint32_t llama_hparams::n_embd_v_gqa_max() const {
     return val;
 }
 
-uint32_t llama_hparams::n_embd_k_idx(uint32_t il) const {
-    if (!indexer_kv || indexer_head_size == 0) {
-        return 0; // arch without a MSA indexer
-    }
-    if (il < n_layer_dense_lead) {
-        return 0; // leading dense layers carry no indexer
-    }
-    return indexer_head_size; // 128
-}
-
 uint32_t llama_hparams::n_embd_r() const {
     if (wkv_head_size != 0) {
         // for RWKV models

src/llama-hparams.h

@@ -230,8 +230,6 @@ struct llama_hparams {
     // MSA
     uint32_t indexer_block_size  = 0;
     uint32_t indexer_local_blocks = 0;
-    // MSA stores its indexer keys in the main KV cache (k_idx tensors);
-    bool indexer_kv = false;
 
     // Indexer is "full" (1) or "shared" (0)
     // Shared indexers reuse top-k from previous full layer
@@ -356,9 +354,6 @@ struct llama_hparams {
     uint32_t n_embd_k_gqa_max() const;
     uint32_t n_embd_v_gqa_max() const;
 
-    // dimension of the single-head MSA indexer key stream
-    uint32_t n_embd_k_idx(uint32_t il = 0) const;
-
     // dimension of the rolling state embeddings
     // corresponds to Mamba's conv_states size or RWKV's token_shift states size
     uint32_t n_embd_r() const;

src/llama-kv-cache-dsa.cpp

@@ -23,7 +23,8 @@ llama_kv_cache_dsa::llama_kv_cache_dsa(
                  uint32_t   n_pad,
                  uint32_t   n_swa,
            llama_swa_type   swa_type,
-    const layer_filter_cb & filter,
+    const layer_filter_cb & filter_mla,
+    const layer_filter_cb & filter_lid,
     const  layer_reuse_cb & reuse) :
     hparams_lid(model.hparams), n_stream(unified ? 1 : n_seq_max) {
 
@@ -32,7 +33,7 @@ llama_kv_cache_dsa::llama_kv_cache_dsa(
     kv_mla = std::make_unique<llama_kv_cache>(
             model, model.hparams, type_k, type_v,
             v_trans, offload, unified, kv_size, n_seq_max, n_pad,
-            n_swa, swa_type, nullptr, filter, reuse, nullptr);
+            n_swa, swa_type, nullptr, filter_mla, reuse, nullptr);
 
     // we use llama_kv_cache for caching indexer keys
     // by hand-tweaking some hparams we fool it to create
@@ -49,7 +50,7 @@ llama_kv_cache_dsa::llama_kv_cache_dsa(
     kv_lid = std::make_unique<llama_kv_cache>(
             model, hparams_lid, type_k, type_v,
             v_trans, offload, unified, kv_size, n_seq_max, n_pad,
-            n_swa, swa_type, nullptr, filter, reuse, nullptr);
+            n_swa, swa_type, nullptr, filter_lid, reuse, nullptr);
 }
 
 void llama_kv_cache_dsa::clear(bool data) {

src/llama-kv-cache-dsa.h

@@ -26,7 +26,8 @@ public:
                      uint32_t   n_pad,
                      uint32_t   n_swa,
                llama_swa_type   swa_type,
-        const layer_filter_cb & filter,
+        const layer_filter_cb & filter_mla,
+        const layer_filter_cb & filter_lid,
         const  layer_reuse_cb & reuse);
 
     ~llama_kv_cache_dsa() = default;

src/llama-kv-cache-msa.cpp

@@ -0,0 +1,395 @@
+#include "llama-kv-cache-msa.h"
+
+#include "llama-impl.h"
+#include "llama-batch.h"
+#include "llama-model.h"
+
+#include <algorithm>
+#include <cassert>
+#include <cmath>
+
+// llama_kv_cache_msa
+
+llama_kv_cache_msa::llama_kv_cache_msa(
+        const llama_model & model,
+                ggml_type   type_k,
+                ggml_type   type_v,
+                     bool   v_trans,
+                     bool   offload,
+                     bool   unified,
+                 uint32_t   kv_size,
+                 uint32_t   n_seq_max,
+                 uint32_t   n_pad,
+                 uint32_t   n_swa,
+           llama_swa_type   swa_type,
+    const layer_filter_cb & filter,
+    const layer_filter_cb & filter_idx,
+    const  layer_reuse_cb & reuse) :
+    hparams_idx(model.hparams),
+    n_stream(unified ? 1 : n_seq_max), n_seq_max(n_seq_max), n_pad(n_pad),
+    n_swa(n_swa), swa_type(swa_type) {
+
+    LLAMA_LOG_INFO("%s: creating main KV cache, size = %u cells\n", __func__, kv_size);
+
+    kv_base = std::make_unique<llama_kv_cache>(
+            model, model.hparams, type_k, type_v,
+            v_trans, offload, unified, kv_size, n_seq_max, n_pad,
+            n_swa, swa_type, nullptr, filter, reuse, nullptr);
+
+    // the MSA indexer uses a single key head per layer
+    std::fill(hparams_idx.n_head_kv_arr.begin(), hparams_idx.n_head_kv_arr.end(), 1);
+    hparams_idx.n_embd_head_k_full = model.hparams.indexer_head_size;
+    // the rope parameters are kept identical to the main cache
+
+    LLAMA_LOG_INFO("%s: creating indexer KV cache, size = %u cells\n", __func__, kv_size);
+
+    kv_idx = std::make_unique<llama_kv_cache>(
+            model, hparams_idx, type_k, type_v,
+            v_trans, offload, unified, kv_size, n_seq_max, n_pad,
+            n_swa, swa_type, nullptr, filter_idx, reuse, nullptr);
+}
+
+void llama_kv_cache_msa::clear(bool data) {
+    kv_base->clear(data);
+    kv_idx ->clear(data);
+}
+
+bool llama_kv_cache_msa::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
+    bool res = true;
+
+    res = res & kv_base->seq_rm(seq_id, p0, p1);
+    res = res & kv_idx ->seq_rm(seq_id, p0, p1);
+
+    return res;
+}
+
+void llama_kv_cache_msa::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) {
+    kv_base->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+    kv_idx ->seq_cp(seq_id_src, seq_id_dst, p0, p1);
+}
+
+void llama_kv_cache_msa::seq_keep(llama_seq_id seq_id) {
+    kv_base->seq_keep(seq_id);
+    kv_idx ->seq_keep(seq_id);
+}
+
+void llama_kv_cache_msa::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) {
+    kv_base->seq_add(seq_id, p0, p1, shift);
+    kv_idx ->seq_add(seq_id, p0, p1, shift);
+}
+
+void llama_kv_cache_msa::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    kv_base->seq_div(seq_id, p0, p1, d);
+    kv_idx ->seq_div(seq_id, p0, p1, d);
+}
+
+llama_pos llama_kv_cache_msa::seq_pos_min(llama_seq_id seq_id) const {
+    return kv_base->seq_pos_min(seq_id);
+}
+
+llama_pos llama_kv_cache_msa::seq_pos_max(llama_seq_id seq_id) const {
+    return kv_base->seq_pos_max(seq_id);
+}
+
+std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache_msa::memory_breakdown() const {
+    std::map<ggml_backend_buffer_type_t, size_t> mb = kv_base->memory_breakdown();
+    for (const auto & buft_size : kv_idx->memory_breakdown()) {
+        mb[buft_size.first] += buft_size.second;
+    }
+    return mb;
+}
+
+llama_memory_context_ptr llama_kv_cache_msa::init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) {
+    GGML_UNUSED(embd_all);
+
+    do {
+        balloc.split_reset();
+
+        std::vector<llama_ubatch> ubatches;
+        while (true) {
+            auto ubatch = n_stream == 1 ? balloc.split_simple(n_ubatch) : balloc.split_equal(n_ubatch, true, 0);
+
+            if (ubatch.n_tokens == 0) {
+                break;
+            }
+
+            ubatches.push_back(std::move(ubatch));
+        }
+
+        if (balloc.get_n_used() < balloc.get_n_tokens()) {
+            // failed to find a suitable split
+            break;
+        }
+
+        auto sinfos_base = kv_base->prepare(ubatches);
+        if (sinfos_base.empty()) {
+            break;
+        }
+
+        auto sinfos_idx = kv_idx->prepare(ubatches);
+        if (sinfos_idx.empty()) {
+            break;
+        }
+
+        assert(sinfos_base.size() == sinfos_idx.size());
+
+        return std::make_unique<llama_kv_cache_msa_context>(
+                this, std::move(sinfos_base), std::move(sinfos_idx), std::move(ubatches));
+    } while (false);
+
+    return std::make_unique<llama_kv_cache_msa_context>(LLAMA_MEMORY_STATUS_FAILED_PREPARE);
+}
+
+llama_memory_context_ptr llama_kv_cache_msa::init_full() {
+    return std::make_unique<llama_kv_cache_msa_context>(this);
+}
+
+llama_memory_context_ptr llama_kv_cache_msa::init_update(llama_context * lctx, bool optimize) {
+    return std::make_unique<llama_kv_cache_msa_context>(this, lctx, optimize);
+}
+
+bool llama_kv_cache_msa::get_can_shift() const {
+    return kv_base->get_can_shift() &&
+           kv_idx ->get_can_shift() &&
+           kv_base->get_size() == kv_idx->get_size();
+}
+
+void llama_kv_cache_msa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    kv_base->state_write(io, seq_id, flags);
+    kv_idx ->state_write(io, seq_id, flags);
+}
+
+void llama_kv_cache_msa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    kv_base->state_read(io, seq_id, flags);
+    kv_idx ->state_read(io, seq_id, flags);
+}
+
+llama_kv_cache * llama_kv_cache_msa::get_base() const {
+    return kv_base.get();
+}
+
+llama_kv_cache * llama_kv_cache_msa::get_idx() const {
+    return kv_idx.get();
+}
+
+// llama_kv_cache_msa_context
+
+llama_kv_cache_msa_context::llama_kv_cache_msa_context(llama_memory_status status) :
+    kv(nullptr), status(status) {}
+
+llama_kv_cache_msa_context::llama_kv_cache_msa_context(
+        llama_kv_cache_msa * kv) :
+    kv(kv),
+    ctx_base(kv->get_base()->init_full()),
+    ctx_idx (kv->get_idx ()->init_full()),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_idx->get_status())) {
+}
+
+llama_kv_cache_msa_context::llama_kv_cache_msa_context(
+        llama_kv_cache_msa * kv,
+        llama_context * lctx,
+        bool optimize) :
+    kv(kv),
+    ctx_base(kv->get_base()->init_update(lctx, optimize)),
+    ctx_idx (kv->get_idx ()->init_update(lctx, optimize)),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_idx->get_status())) {
+}
+
+llama_kv_cache_msa_context::llama_kv_cache_msa_context(
+        llama_kv_cache_msa * kv,
+        slot_info_vec_t sinfos_base,
+        slot_info_vec_t sinfos_idx,
+        std::vector<llama_ubatch> ubatches) :
+    kv(kv),
+    ubatches(std::move(ubatches)),
+    // here we copy the ubatches. not sure if this is ideal
+    ctx_base(new llama_kv_cache_context(kv->get_base(), std::move(sinfos_base), this->ubatches)),
+    ctx_idx (new llama_kv_cache_context(kv->get_idx (), std::move(sinfos_idx),  this->ubatches)),
+    status(llama_memory_status_combine(ctx_base->get_status(), ctx_idx->get_status())) {
+}
+
+llama_kv_cache_msa_context::~llama_kv_cache_msa_context() = default;
+
+bool llama_kv_cache_msa_context::next() {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    ctx_base->next();
+    ctx_idx ->next();
+
+    if (++i_next >= ubatches.size()) {
+        return false;
+    }
+
+    return true;
+}
+
+bool llama_kv_cache_msa_context::apply() {
+    assert(!llama_memory_status_is_fail(status));
+
+    bool res = true;
+
+    res = res & ctx_base->apply();
+    res = res & ctx_idx ->apply();
+
+    return res;
+}
+
+llama_memory_status llama_kv_cache_msa_context::get_status() const {
+    return status;
+}
+
+const llama_ubatch & llama_kv_cache_msa_context::get_ubatch() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return ubatches[i_next];
+}
+
+const llama_kv_cache_context * llama_kv_cache_msa_context::get_base() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return static_cast<const llama_kv_cache_context *>(ctx_base.get());
+}
+
+const llama_kv_cache_context * llama_kv_cache_msa_context::get_idx() const {
+    assert(status == LLAMA_MEMORY_STATUS_SUCCESS);
+
+    return static_cast<const llama_kv_cache_context *>(ctx_idx.get());
+}
+
+uint32_t llama_kv_cache_msa_context::get_n_pos() const {
+    // pad the value so that the graph remains constant across batches and can be reused
+    const uint32_t n_pad_cur = std::max(kv->get_n_pad(), 256u);
+
+    llama_pos pos_max = -1;
+
+    for (llama_seq_id seq_id = 0; seq_id < (llama_seq_id) kv->get_n_seq_max(); ++seq_id) {
+        pos_max = std::max(pos_max, kv->seq_pos_max(seq_id));
+    }
+
+    return std::max(n_pad_cur, GGML_PAD((uint32_t) (pos_max + 1), n_pad_cur));
+}
+
+void llama_kv_cache_msa_context::set_input_cell_pos(ggml_tensor * dst, const llama_ubatch * ubatch, int32_t div) const {
+    GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
+    GGML_ASSERT(dst->type == GGML_TYPE_I32);
+    GGML_ASSERT(div > 0);
+
+    const int64_t n_tokens    = ubatch->n_tokens;
+    const int64_t n_kv        = dst->ne[0];
+    const int64_t n_stream_ub = dst->ne[1];
+
+    GGML_ASSERT(n_tokens % n_stream_ub == 0);
+    const int64_t n_tps = n_tokens/n_stream_ub;
+
+    int32_t * data = (int32_t *) dst->data;
+
+    for (int64_t s = 0; s < n_stream_ub; ++s) {
+        const llama_seq_id seq_id = ubatch->seq_id[s*n_tps][0];
+
+        const auto & cells = kv->get_base()->get_cells(seq_id);
+
+        for (int64_t j = 0; j < n_kv; ++j) {
+            // the value for empty or other-sequence cells is irrelevant as consumers mask them
+            data[s*n_kv + j] =
+                cells.is_empty(j) || !cells.seq_has(j, seq_id)
+                    ? 0
+                    : (int32_t) (cells.pos_get(j)/div);
+        }
+    }
+}
+
+void llama_kv_cache_msa_context::set_input_pos_slot(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+    GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
+    GGML_ASSERT(dst->type == GGML_TYPE_I32 || dst->type == GGML_TYPE_F32);
+
+    const int64_t n_tokens    = ubatch->n_tokens;
+    const int64_t n_pos       = dst->ne[0];
+    const int64_t n_stream_ub = dst->ne[1];
+
+    GGML_ASSERT(n_tokens % n_stream_ub == 0);
+    const int64_t n_tps = n_tokens/n_stream_ub;
+
+    for (int64_t s = 0; s < n_stream_ub; ++s) {
+        const llama_seq_id seq_id = ubatch->seq_id[s*n_tps][0];
+
+        const auto & cells = kv->get_base()->get_cells(seq_id);
+
+        std::vector<int32_t> map(n_pos, 0);
+
+        for (uint32_t j = 0; j < cells.size(); ++j) {
+            if (cells.is_empty(j) || !cells.seq_has(j, seq_id)) {
+                continue;
+            }
+
+            const llama_pos p0 = cells.pos_get(j);
+
+            if (p0 < 0 || p0 >= n_pos) {
+                continue;
+            }
+
+            map[p0] = (int32_t) j;
+        }
+
+        if (dst->type == GGML_TYPE_I32) {
+            int32_t * data = (int32_t *) dst->data + s*n_pos;
+            std::copy(map.begin(), map.end(), data);
+        } else {
+            float * data = (float *) dst->data + s*n_pos;
+            for (int64_t p = 0; p < n_pos; ++p) {
+                data[p] = (float) map[p];
+            }
+        }
+    }
+}
+
+void llama_kv_cache_msa_context::set_input_pos_mask(ggml_tensor * dst, const llama_ubatch * ubatch) const {
+    GGML_ASSERT(ggml_backend_buffer_is_host(dst->buffer));
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int64_t n_tokens = ubatch->n_tokens;
+    const int64_t n_pos    = dst->ne[0];
+
+    GGML_ASSERT(dst->ne[1] == n_tokens);
+
+    const uint32_t       n_swa    = kv->get_n_swa();
+    const llama_swa_type swa_type = kv->get_swa_type();
+
+    float * data = (float *) dst->data;
+
+    std::fill(data, data + n_pos*n_tokens, -INFINITY);
+
+    for (int64_t i = 0; i < n_tokens; ++i) {
+        const llama_seq_id seq_id = ubatch->seq_id[i][0];
+
+        const auto & cells = kv->get_base()->get_cells(seq_id);
+
+        const llama_pos p1 = ubatch->pos[i];
+
+        for (uint32_t j = 0; j < cells.size(); ++j) {
+            if (cells.is_empty(j) || !cells.seq_has(j, seq_id)) {
+                continue;
+            }
+
+            const llama_pos p0 = cells.pos_get(j);
+
+            if (p0 < 0 || p0 >= n_pos) {
+                continue;
+            }
+
+            // causal mask
+            if (p0 > p1) {
+                continue;
+            }
+
+            // apply SWA if any
+            if (llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1)) {
+                continue;
+            }
+
+            data[i*n_pos + p0] = 0.0f;
+        }
+    }
+}

src/llama-kv-cache-msa.h

@@ -0,0 +1,153 @@
+#pragma once
+
+#include "llama-kv-cache.h"
+
+#include <vector>
+
+// llama_kv_cache_msa
+
+// uses two instances of llama_kv_cache, one for K/V tensors, and one for the MSA indexer tensors
+// both receive identical sequence operations and identical ubatches, so their cell layouts stay in synced.
+// the context also exposes per-ubatch pos - cell translation maps populated from llama_kv_cells via
+// llama_kv_cache::get_cells(), which the model graph uses to run MSA block selection in position space
+
+class llama_kv_cache_msa : public llama_memory_i {
+public:
+    llama_kv_cache_msa(
+            const llama_model & model,
+                    ggml_type   type_k,
+                    ggml_type   type_v,
+                         bool   v_trans,
+                         bool   offload,
+                         bool   unified,
+                     uint32_t   kv_size,
+                     uint32_t   n_seq_max,
+                     uint32_t   n_pad,
+                     uint32_t   n_swa,
+               llama_swa_type   swa_type,
+        const layer_filter_cb & filter,
+        const layer_filter_cb & filter_idx,
+        const  layer_reuse_cb & reuse);
+
+    ~llama_kv_cache_msa() = default;
+
+    // llama_memory_i
+
+    llama_memory_context_ptr init_batch(
+            llama_batch_allocr & balloc,
+            uint32_t n_ubatch,
+            bool embd_all) override;
+
+    llama_memory_context_ptr init_full() override;
+
+    llama_memory_context_ptr init_update(llama_context * lctx, bool optimize) override;
+
+    bool get_can_shift() const override;
+
+    void clear(bool data) override;
+
+    bool seq_rm  (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1) override;
+    void seq_cp  (llama_seq_id seq_id_src, llama_seq_id seq_id_dst, llama_pos p0, llama_pos p1) override;
+    void seq_keep(llama_seq_id seq_id)                                                          override;
+    void seq_add (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, llama_pos shift) override;
+    void seq_div (llama_seq_id seq_id,                              llama_pos p0, llama_pos p1, int d) override;
+
+    llama_pos seq_pos_min(llama_seq_id seq_id) const override;
+    llama_pos seq_pos_max(llama_seq_id seq_id) const override;
+
+    std::map<ggml_backend_buffer_type_t, size_t> memory_breakdown() const override;
+
+    // state write/load
+
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
+
+    // llama_kv_cache_msa specific API
+
+    llama_kv_cache * get_base() const;
+    llama_kv_cache * get_idx () const;
+
+    uint32_t       get_n_pad()    const { return n_pad; }
+    uint32_t       get_n_seq_max() const { return n_seq_max; }
+    uint32_t       get_n_swa()    const { return n_swa; }
+    llama_swa_type get_swa_type() const { return swa_type; }
+
+private:
+    // keep the indexer KV cache hparams instance here as llama_kv_cache stores only a reference
+    llama_hparams hparams_idx;
+
+    const uint32_t n_stream  = 1;
+    const uint32_t n_seq_max = 1;
+    const uint32_t n_pad     = 1;
+
+    const uint32_t       n_swa    = 0;
+    const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
+
+    std::unique_ptr<llama_kv_cache> kv_base;
+    std::unique_ptr<llama_kv_cache> kv_idx;
+};
+
+class llama_kv_cache_msa_context : public llama_memory_context_i {
+public:
+    using slot_info_vec_t = llama_kv_cache::slot_info_vec_t;
+
+    // used for errors
+    llama_kv_cache_msa_context(llama_memory_status status);
+
+    // used to create a full-cache context
+    llama_kv_cache_msa_context(
+            llama_kv_cache_msa * kv);
+
+    // used to create an update context
+    llama_kv_cache_msa_context(
+            llama_kv_cache_msa * kv,
+            llama_context * lctx,
+            bool optimize);
+
+    // used to create a batch processing context from a batch
+    llama_kv_cache_msa_context(
+            llama_kv_cache_msa * kv,
+            slot_info_vec_t sinfos_base,
+            slot_info_vec_t sinfos_idx,
+            std::vector<llama_ubatch> ubatches);
+
+    virtual ~llama_kv_cache_msa_context();
+
+    // llama_memory_context_i
+
+    bool next()  override;
+    bool apply() override;
+
+    llama_memory_status  get_status() const override;
+    const llama_ubatch & get_ubatch() const override;
+
+    // llama_kv_cache_msa_context specific API
+
+    const llama_kv_cache_context * get_base() const;
+    const llama_kv_cache_context * get_idx () const;
+
+    // max position currently present in the cache plus one, padded MSA blocks are defined over token positions
+    // so the block-selection tensors are sized by this value rather than by the number of cells
+    uint32_t get_n_pos() const;
+
+    // position <-> cell translation maps, populated from the base cache cells
+    // the model graph relates cache contents to token positions only through these per ubatch inputs
+    // value for empty or other-sequence cells is 0 so consumers must mask them
+    void set_input_cell_pos(ggml_tensor * dst, const llama_ubatch * ubatch, int32_t div) const;
+    // positions without a cell map to cell 0, consumers must mask them assumes one sequence per stream
+    void set_input_pos_slot(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+    void set_input_pos_mask(ggml_tensor * dst, const llama_ubatch * ubatch) const;
+
+private:
+    llama_kv_cache_msa * kv;
+
+    // the index of the next ubatch to process
+    size_t i_next = 0;
+
+    std::vector<llama_ubatch> ubatches;
+
+    const llama_memory_context_ptr ctx_base;
+    const llama_memory_context_ptr ctx_idx;
+
+    const llama_memory_status status;
+};

src/llama-kv-cache.cpp

@@ -112,7 +112,7 @@ llama_kv_cache::llama_kv_cache(
         auto it = ctx_map.find(buft);
         if (it == ctx_map.end()) {
             ggml_init_params params = {
-                /*.mem_size   =*/ size_t(3u*(1 + n_stream)*n_layer*ggml_tensor_overhead()), //Reserve tensor metadata for up to 3 tensors per layer (K, V, and optional K_idx), plus one view per tensor per stream.
+                /*.mem_size   =*/ size_t(2u*(1 + n_stream)*n_layer*ggml_tensor_overhead()),
                 /*.mem_buffer =*/ NULL,
                 /*.no_alloc   =*/ true,
             };
@@ -242,25 +242,9 @@ llama_kv_cache::llama_kv_cache(
             v_stream.push_back(has_v ? ggml_view_2d(ctx, v, n_embd_v_gqa, kv_size, v->nb[1], s*v->nb[2]) : nullptr);
         }
 
-        const uint32_t n_embd_k_idx = hparams.n_embd_k_idx(il);
-        ggml_tensor * k_idx = n_embd_k_idx > 0
-            ? ggml_new_tensor_3d(ctx, GGML_TYPE_F32, n_embd_k_idx, kv_size, n_stream)
-            : nullptr;
-        if (k_idx) {
-            ggml_format_name(k_idx, "cache_k_idx_l%d", il);
-            msa_strict_slots = (n_stream == n_seq_max);
-        }
-
-        std::vector<ggml_tensor *> k_idx_stream;
-        for (uint32_t s = 0; s < n_stream; ++s) {
-            k_idx_stream.push_back(k_idx
-                ? ggml_view_2d(ctx, k_idx, n_embd_k_idx, kv_size, k_idx->nb[1], s*k_idx->nb[2])
-                : nullptr);
-        }
-
         map_layer_ids[il] = layers.size();
 
-        layers.push_back({ il, k, v, k_idx, k_stream, v_stream, k_idx_stream });
+        layers.push_back({ il, k, v, k_stream, v_stream, });
     }
 
     if (reuse) {
@@ -309,24 +293,13 @@ llama_kv_cache::llama_kv_cache(
     }
 
     {
-        const size_t memory_size_k     = size_k_bytes();
-        const size_t memory_size_v     = size_v_bytes();
-        const size_t memory_size_k_idx = size_k_idx_bytes();
-        const size_t memory_size_total = memory_size_k + memory_size_v + memory_size_k_idx;
+        const size_t memory_size_k = size_k_bytes();
+        const size_t memory_size_v = size_v_bytes();
 
-        constexpr float mib = 1024.0f * 1024.0f;
-
-        const std::string k_log = format(", K (%s): %7.2f MiB", ggml_type_name(type_k), (float) memory_size_k / mib);
-        const std::string v_log = format(", V (%s): %7.2f MiB", ggml_type_name(type_v), (float) memory_size_v / mib);
-
-        std::string k_idx_log;
-        if (memory_size_k_idx > 0) {
-            k_idx_log = format(", K_idx (%s): %7.2f MiB", ggml_type_name(GGML_TYPE_F32), (float) memory_size_k_idx / mib);
-        }
-
-        LLAMA_LOG_INFO("%s: size = %7.2f MiB (%6u cells, %3d layers, %2u/%u seqs)%s%s%s\n", __func__,
-                (float) memory_size_total / mib, kv_size, (int) layers.size(), n_seq_max, n_stream,
-                k_log.c_str(), v_log.c_str(), k_idx_log.c_str());
+        LLAMA_LOG_INFO("%s: size = %7.2f MiB (%6u cells, %3d layers, %2u/%u seqs), K (%s): %7.2f MiB, V (%s): %7.2f MiB\n", __func__,
+                (float)(memory_size_k + memory_size_v) / (1024.0f * 1024.0f), kv_size, (int) layers.size(), n_seq_max, n_stream,
+                ggml_type_name(type_k), (float)memory_size_k / (1024.0f * 1024.0f),
+                ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
     }
 
     // TODO: refactor [TAG_KV_CACHE_SHARE_CELLS]
@@ -419,39 +392,6 @@ bool llama_kv_cache::seq_rm(llama_seq_id seq_id, llama_pos p0, llama_pos p1) {
         p1 = std::numeric_limits<llama_pos>::max();
     }
 
-    // empty range - nothing to remove
-    if (p0 >= p1) {
-        return true;
-    }
-
-    // MSA anchors block selection to absolute cache slots (slot == position). Tail trim and full removal preserve this invariant, but removing a prefix
-    // or middle range would free slots while later cells survive, desynchronizing the indexer cache. Reject such removals before modifying the cache.
-    if (msa_strict_slots) {
-        for (llama_seq_id sid = 0; sid < (llama_seq_id) seq_to_stream.size(); ++sid) {
-            if (seq_id >= 0 && sid != seq_id) {
-                continue;
-            }
-
-            const auto & cells = v_cells[seq_to_stream[sid]];
-
-            const llama_pos pmin = cells.seq_pos_min(sid);
-            const llama_pos pmax = cells.seq_pos_max(sid);
-
-            if (pmin < 0) {
-                continue;   // empty sequence
-            }
-
-            const bool overlaps    = p0 <= pmax && p1 > pmin;   // the range removes something
-            const bool leaves_tail = p1 <= pmax;                // cells beyond the range survive
-
-            if (overlaps && leaves_tail) {
-                LLAMA_LOG_WARN("%s: MSA: partial (non-suffix) removal [%d, %d) for seq %d is not supported "
-                        "(block selection is anchored to cache slots) - rejected\n", __func__, p0, p1, sid);
-                return false;
-            }
-        }
-    }
-
     if (seq_id >= 0) {
         auto & cells = v_cells[seq_to_stream[seq_id]];
         auto & head  = v_heads[seq_to_stream[seq_id]];
@@ -906,10 +846,6 @@ bool llama_kv_cache::update(llama_context * lctx, bool do_shift, const stream_co
                 if (layer.v_stream[ssrc]) {
                     ggml_backend_tensor_copy(layer.v_stream[ssrc], layer.v_stream[sdst]);
                 }
-                if (layer.k_idx_stream[ssrc]) {
-                    GGML_ASSERT(layer.k_idx_stream[sdst]);
-                    ggml_backend_tensor_copy(layer.k_idx_stream[ssrc], layer.k_idx_stream[sdst]);
-                }
             }
         }
     }
@@ -1058,44 +994,6 @@ llama_kv_cache::slot_info llama_kv_cache::find_slot(const llama_ubatch & ubatch,
 
         const auto & cells = v_cells[seq_to_stream[seq_id]];
 
-        if (n_tokens > cells.size()) {
-            LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %u\n", __func__, n_tokens, cells.size());
-            return { };
-        }
-
-        // MSA block selection assumes slot == logical position (append-only streams).
-        if (msa_strict_slots) {
-            for (uint32_t ii = 0; ii < n_tokens; ++ii) {
-                const llama_pos pos = ubatch.pos[s*n_tokens + ii];
-
-                if (pos < 0 || (uint64_t) pos >= cells.size()) {
-                    LLAMA_LOG_WARN("%s: MSA: position %d is outside the cache range [0, %u)\n",
-                            __func__, pos, cells.size());
-                    return { };
-                }
-
-                const uint32_t idx = (uint32_t) pos;
-
-                if (!cells.is_empty(idx)) {
-                    LLAMA_LOG_WARN("%s: MSA: required slot %u is already occupied (stream %u)\n",
-                            __func__, idx, seq_to_stream[seq_id]);
-                    return { };
-                }
-
-                // strictly increasing positions, rules out duplicates and, for contiguous requests, is tightened to exact adjacency
-                if (!res.idxs[s].empty() && (cont ? idx != res.idxs[s].back() + 1
-                                                  : idx <= res.idxs[s].back())) {
-                    LLAMA_LOG_WARN("%s: MSA: token positions are not %s within the ubatch\n",
-                            __func__, cont ? "contiguous" : "strictly increasing");
-                    return { };
-                }
-
-                res.idxs[s].push_back(idx);
-            }
-
-            continue;
-        }
-
         uint32_t head_cur = v_heads[seq_to_stream[seq_id]];
 
         // if we have enough unused cells before the current head ->
@@ -1104,6 +1002,11 @@ llama_kv_cache::slot_info llama_kv_cache::find_slot(const llama_ubatch & ubatch,
             head_cur = 0;
         }
 
+        if (n_tokens > cells.size()) {
+            LLAMA_LOG_ERROR("%s: n_tokens = %d > size = %u\n", __func__, n_tokens, cells.size());
+            return { };
+        }
+
         uint32_t n_tested = 0;
 
         // for continuous slots, we test that all tokens in the ubatch fit, starting from the current head
@@ -1210,15 +1113,6 @@ void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch &
 
             const auto idx = sinfo.idxs[s][ii];
 
-            if (msa_strict_slots && (llama_pos) idx != ubatch.pos[i]) {
-                LLAMA_LOG_ERROR("%s: MSA slot/position invariant violated: "
-                        "writing pos %d into cell %u (stream %u). The indexer cache "
-                        "would desync and block selection would silently corrupt. "
-                        "This is a bug, please report it with reproduction steps.\n",
-                        __func__, ubatch.pos[i], idx, sinfo.strm[s]);
-                GGML_ABORT("MSA: slot != pos");
-            }
-
             if (!cells.is_empty(idx)) {
                 assert(cells.seq_count(idx) == 1);
 
@@ -1262,8 +1156,7 @@ void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch &
             LLAMA_LOG_DEBUG("%s: purging positions [%d, %d] of sequence %d from KV cache\n",
                     __func__, cells.seq_pos_min(s), seq_pos_max_rm[s], s);
 
-            // under MSA strict slots this path should be unreachable, since strict MSA placement never selects occupied cells
-            GGML_ASSERT(seq_rm(s, cells.seq_pos_min(s), seq_pos_max_rm[s] + 1));
+            seq_rm(s, cells.seq_pos_min(s), seq_pos_max_rm[s] + 1);
         }
     }
 
@@ -1283,12 +1176,6 @@ bool llama_kv_cache::get_can_shift() const {
     if (hparams.n_pos_per_embd() > 1) {
         return false;
     }
-    // shifting would leave k_idx stale
-    for (const auto & layer : layers) {
-        if (layer.k_idx) {
-            return false;
-        }
-    }
     return true;
 }
 
@@ -1337,6 +1224,12 @@ ggml_tensor * llama_kv_cache::get_k_storage(int32_t il) const {
     return layers[ikv].k;
 }
 
+const llama_kv_cells & llama_kv_cache::get_cells(llama_seq_id seq_id) const {
+    GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size());
+
+    return v_cells[seq_to_stream[seq_id]];
+}
+
 uint32_t llama_kv_cache::get_n_kv(const slot_info & sinfo) const {
     uint32_t result = 0;
 
@@ -1405,23 +1298,6 @@ ggml_tensor * llama_kv_cache::get_v(ggml_context * ctx, int32_t il, uint32_t n_k
             ggml_row_size(v->type, kv_size*n_embd_v_gqa)*sinfo.s0);
 }
 
-ggml_tensor * llama_kv_cache::get_k_idx(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const {
-    const int32_t ikv = map_layer_ids.at(il);
-    auto * k_idx = layers[ikv].k_idx;
-    GGML_ASSERT(k_idx);
-
-    const uint64_t kv_size = get_size();
-    const int64_t  n_idx   = k_idx->ne[0];                 // 128
-    const uint32_t ns      = sinfo.s1 - sinfo.s0 + 1;
-
-    return ggml_view_4d(ctx, k_idx,
-            n_idx, 1, n_kv, ns,
-            ggml_row_size(k_idx->type, n_idx),             // nb1 (single head)
-            ggml_row_size(k_idx->type, n_idx),             // nb2 (per cell)
-            ggml_row_size(k_idx->type, n_idx*kv_size),     // nb3 (per stream)
-            ggml_row_size(k_idx->type, n_idx*kv_size)*sinfo.s0);
-}
-
 ggml_tensor * llama_kv_cache::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const {
     GGML_UNUSED(sinfo);
 
@@ -1523,28 +1399,6 @@ ggml_tensor * llama_kv_cache::build_input_k_idxs(ggml_context * ctx, const llama
     return k_idxs;
 }
 
-ggml_tensor * llama_kv_cache::cpy_k_idx(ggml_context * ctx, ggml_tensor * k_idx_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const {
-    GGML_UNUSED(sinfo);
-    const int32_t ikv = map_layer_ids.at(il);
-    ggml_tensor * k_idx = layers[ikv].k_idx;
-    GGML_ASSERT(k_idx && "cpy_k_idx on a layer with no indexer cache");
-
-    const int64_t n_embd_head = k_idx_cur->ne[0];          // 128
-    const int64_t n_head      = k_idx_cur->ne[1];          // 1
-    const int64_t n_tokens    = k_idx_cur->ne[2];
-    const int64_t n_embd_gqa  = n_embd_head*n_head;        // 128
-
-    GGML_ASSERT(ggml_row_size(k_idx_cur->type, n_embd_head) == k_idx_cur->nb[1]);
-    k_idx_cur = ggml_view_2d(ctx, k_idx_cur, n_embd_gqa, n_tokens, k_idx_cur->nb[2], 0);
-
-    const int64_t n_stream = k_idx->ne[2];
-    if (n_stream > 1) {
-        const int64_t kv_size = get_size();
-        k_idx = ggml_reshape_2d(ctx, k_idx, n_embd_gqa, kv_size*n_stream);
-    }
-    return ggml_set_rows(ctx, k_idx, k_idx_cur, k_idxs);   // same k_idxs as the K store
-}
-
 ggml_tensor * llama_kv_cache::build_input_v_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
     const uint32_t n_tokens = ubatch.n_tokens;
 
@@ -1979,18 +1833,6 @@ size_t llama_kv_cache::size_v_bytes() const {
     return size_v_bytes;
 }
 
-size_t llama_kv_cache::size_k_idx_bytes() const {
-    size_t size_k_idx_bytes = 0;
-
-    for (const auto & layer : layers) {
-        if (layer.k_idx) {
-            size_k_idx_bytes += ggml_nbytes(layer.k_idx);
-        }
-    }
-
-    return size_k_idx_bytes;
-}
-
 ggml_tensor * llama_kv_cache::build_rope_shift(
         const llama_cparams & cparams,
                ggml_context * ctx,
@@ -2303,36 +2145,6 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
         }
     }
 
-    if (size_k_idx_bytes() > 0) {
-        const uint32_t has_k_idx_u32 = 1;
-        io.write(&has_k_idx_u32, sizeof(has_k_idx_u32));
-
-        for (const auto & layer : layers) {
-            const uint32_t layer_has_k_idx = layer.k_idx ? 1 : 0;
-            io.write(&layer_has_k_idx, sizeof(layer_has_k_idx));
-
-            if (!layer_has_k_idx) {
-                continue;
-            }
-
-            GGML_ASSERT(layer.k_idx_stream[cr.strm]);
-
-            const int32_t k_idx_type_i = (int32_t) layer.k_idx->type;
-            io.write(&k_idx_type_i, sizeof(k_idx_type_i));
-
-            const uint64_t k_idx_size_row = ggml_row_size(layer.k_idx->type, layer.k_idx->ne[0]);
-            io.write(&k_idx_size_row, sizeof(k_idx_size_row));
-
-            for (const auto & range : cr.data) {
-                const size_t range_size = range.second - range.first;
-                const size_t buf_size   = range_size * k_idx_size_row;
-                const size_t offset     = range.first * k_idx_size_row;
-
-                io.write_tensor(layer.k_idx_stream[cr.strm], offset, buf_size);
-            }
-        }
-    }
-
     if (!v_trans) {
         for (const auto & layer : layers) {
             const uint32_t il = layer.il;
@@ -2581,68 +2393,6 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
         }
     }
 
-    if (size_k_idx_bytes() > 0) {
-        uint32_t has_k_idx_u32 = 0;
-        io.read(&has_k_idx_u32, sizeof(has_k_idx_u32));
-
-        if (has_k_idx_u32 != 1) {
-            LLAMA_LOG_ERROR("%s: missing k_idx data in KV cache state\n", __func__);
-            return false;
-        }
-
-        for (const auto & layer : layers) {
-            uint32_t layer_has_k_idx = 0;
-            io.read(&layer_has_k_idx, sizeof(layer_has_k_idx));
-
-            const uint32_t expected_layer_has_k_idx = layer.k_idx ? 1 : 0;
-
-            if (layer_has_k_idx != expected_layer_has_k_idx) {
-                LLAMA_LOG_ERROR(
-                    "%s: mismatched k_idx state for layer: got %u, expected %u\n",
-                    __func__, layer_has_k_idx, expected_layer_has_k_idx);
-                return false;
-            }
-
-            if (!layer_has_k_idx) {
-                continue;
-            }
-
-            GGML_ASSERT(layer.k_idx_stream[strm]);
-
-            int32_t k_idx_type_i = -1;
-            io.read(&k_idx_type_i, sizeof(k_idx_type_i));
-
-            if (k_idx_type_i != (int32_t) layer.k_idx->type) {
-                LLAMA_LOG_ERROR(
-                    "%s: mismatched k_idx type: got %d, expected %d\n",
-                    __func__, k_idx_type_i, (int32_t) layer.k_idx->type);
-                return false;
-            }
-
-            uint64_t k_idx_size_row = 0;
-            io.read(&k_idx_size_row, sizeof(k_idx_size_row));
-
-            const uint64_t expected_k_idx_size_row = ggml_row_size(layer.k_idx->type, layer.k_idx->ne[0]);
-
-            if (k_idx_size_row != expected_k_idx_size_row) {
-                LLAMA_LOG_ERROR(
-                    "%s: mismatched k_idx row size: got %zu, expected %zu\n",
-                    __func__, (size_t) k_idx_size_row, (size_t) expected_k_idx_size_row);
-                return false;
-            }
-
-            if (cell_count) {
-                if (sinfo.is_contiguous()) {
-                    io.read_tensor(layer.k_idx_stream[strm], sinfo.head() * k_idx_size_row, cell_count * k_idx_size_row);
-                } else {
-                    for (uint32_t i = 0; i < cell_count; ++i) {
-                        io.read_tensor(layer.k_idx_stream[strm], sinfo.idxs[0][i] * k_idx_size_row, k_idx_size_row);
-                    }
-                }
-            }
-        }
-    }
-
     if (!this->v_trans) {
         for (const auto & layer : layers) {
             const uint32_t il = layer.il;
@@ -2844,10 +2594,6 @@ ggml_tensor * llama_kv_cache_context::get_v(ggml_context * ctx, int32_t il) cons
     return kv->get_v(ctx, il, n_kv, sinfos[i_cur]);
 }
 
-ggml_tensor * llama_kv_cache_context::get_k_idx(ggml_context * ctx, int32_t il) const {
-    return kv->get_k_idx(ctx, il, n_kv, sinfos[i_cur]);
-}
-
 ggml_tensor * llama_kv_cache_context::cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const {
     return kv->cpy_k(ctx, k_cur, k_idxs, il, sinfos[i_cur]);
 }
@@ -2856,10 +2602,6 @@ ggml_tensor * llama_kv_cache_context::cpy_v(ggml_context * ctx, ggml_tensor * v_
     return kv->cpy_v(ctx, v_cur, v_idxs, il, sinfos[i_cur]);
 }
 
-ggml_tensor * llama_kv_cache_context::cpy_k_idx(ggml_context * ctx, ggml_tensor * k_idx_cur, ggml_tensor * k_idxs, int32_t il) const {
-    return kv->cpy_k_idx(ctx, k_idx_cur, k_idxs, il, sinfos[i_cur]);
-}
-
 ggml_tensor * llama_kv_cache_context::build_input_k_idxs(ggml_context * ctx, const llama_ubatch & ubatch) const {
     return kv->build_input_k_idxs(ctx, ubatch);
 }

src/llama-kv-cache.h

@@ -164,6 +164,8 @@ public:
     std::vector<uint32_t> get_layer_ids() const;
     ggml_tensor * get_k_storage(int32_t il) const;
 
+    const llama_kv_cells & get_cells(llama_seq_id seq_id) const;
+
     //
     // graph_build API
     //
@@ -173,12 +175,10 @@ public:
     // get views of the current state of the cache
     ggml_tensor * get_k(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
     ggml_tensor * get_v(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
-    ggml_tensor * get_k_idx(ggml_context * ctx, int32_t il, uint32_t n_kv, const slot_info & sinfo) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
     ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const;
     ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il, const slot_info & sinfo) const;
-    ggml_tensor * cpy_k_idx(ggml_context * ctx, ggml_tensor * k_idx_cur, ggml_tensor * k_idxs, int32_t il, const slot_info & sinfo) const;
 
     //
     // preparation API
@@ -230,11 +230,9 @@ private:
 
         ggml_tensor * k;
         ggml_tensor * v;
-        ggml_tensor * k_idx;   // MSA single-head indexer keys, F32
 
         std::vector<ggml_tensor *> k_stream;
         std::vector<ggml_tensor *> v_stream;
-        std::vector<ggml_tensor *> k_idx_stream;
     };
 
     bool v_trans = true;  // the value tensor is transposed
@@ -263,9 +261,6 @@ private:
     // env: LLAMA_KV_CACHE_DEBUG
     int debug = 0;
 
-    // set when a k_idx (indexer) cache exists and the stream layout supports MSA (single seq, or one stream per seq)
-    bool msa_strict_slots = false;
-
     // this is the SWA type of the cache - not to be confused with the model SWA type
     const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE;
 
@@ -298,7 +293,6 @@ private:
 
     size_t size_k_bytes() const;
     size_t size_v_bytes() const;
-    size_t size_k_idx_bytes() const;
 
     ggml_tensor * build_rope_shift(
             const llama_cparams & cparams,
@@ -378,7 +372,6 @@ public:
     // get views of the current state of the cache
     ggml_tensor * get_k(ggml_context * ctx, int32_t il) const;
     ggml_tensor * get_v(ggml_context * ctx, int32_t il) const;
-    ggml_tensor * get_k_idx(ggml_context * ctx, int32_t il) const;
 
     // store k_cur and v_cur in the cache based on the provided head location
     // note: the heads in k_cur and v_cur should be laid out contiguously in memory
@@ -388,7 +381,6 @@ public:
     //   - v_idxs [n_tokens] or [n_tokens*n_embd_v_gqa] depending if V cache is transposed
     ggml_tensor * cpy_k(ggml_context * ctx, ggml_tensor * k_cur, ggml_tensor * k_idxs, int32_t il) const;
     ggml_tensor * cpy_v(ggml_context * ctx, ggml_tensor * v_cur, ggml_tensor * v_idxs, int32_t il) const;
-    ggml_tensor * cpy_k_idx(ggml_context * ctx, ggml_tensor * k_idx_cur, ggml_tensor * k_idxs, int32_t il) const;
 
     // create destination indices for each head of the current batch for where it would be written in the KV cache
     // the indices address the global KV cache (not per stream) - this is not relevant for the user of this API, but

src/llama-model-loader.cpp

@@ -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));

src/llama-model-loader.h

@@ -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);

src/llama-model.cpp

@@ -11,6 +11,7 @@
 #include "llama-kv-cache.h"
 #include "llama-kv-cache-iswa.h"
 #include "llama-kv-cache-dsa.h"
+#include "llama-kv-cache-msa.h"
 #include "llama-kv-cache-dsv4.h"
 #include "llama-memory-hybrid.h"
 #include "llama-memory-hybrid-iswa.h"
@@ -111,6 +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:
@@ -2071,9 +2074,13 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
             {
                 res = nullptr;
             } break;
-        case LLM_ARCH_DEEPSEEK32:
+        case LLM_ARCH_MINIMAX_M3:
             {
-                res = new llama_kv_cache_dsa(
+                // sparse (MSA) layers carry an indexer key cache, but leading dense layers do not
+                llama_kv_cache::layer_filter_cb filter_idx =
+                    [&](int32_t il) { return (uint32_t) il >= hparams.n_layer_dense_lead; };
+
+                res = new llama_kv_cache_msa(
                         *this,
                         params.type_k,
                         params.type_v,
@@ -2086,9 +2093,11 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         hparams.n_swa,
                         hparams.swa_type,
                         nullptr,
+                        filter_idx,
                         nullptr);
             } break;
         case LLM_ARCH_GLM_DSA:
+        case LLM_ARCH_DEEPSEEK32:
             {
                 if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP && hparams.n_layer_nextn > 0) {
                     // The NextN/MTP draft head runs dense MLA (no DSA indexer), so the
@@ -2117,10 +2126,11 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                 } else {
                     // Main context: DSA cache for the trunk layers only - the nextn
                     // layer(s) are never attended by the trunk graph.
-                    llama_kv_cache::layer_filter_cb filter = nullptr;
+                    llama_kv_cache::layer_filter_cb filter_mla = nullptr;
                     if (hparams.n_layer_nextn > 0) {
-                        filter = [&](uint32_t il) { return il < hparams.n_layer(); };
+                        filter_mla = [&](uint32_t il) { return il < hparams.n_layer(); };
                     }
+                    llama_kv_cache::layer_filter_cb filter_lid = [&](uint32_t il) { return il < hparams.n_layer() && (arch != LLM_ARCH_GLM_DSA || hparams.is_indexer_full(il)); };
 
                     res = new llama_kv_cache_dsa(
                             *this,
@@ -2134,7 +2144,8 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                             1,
                             hparams.n_swa,
                             hparams.swa_type,
-                            filter,
+                            filter_mla,
+                            filter_lid,
                             nullptr);
                 }
             } break;
@@ -2211,11 +2222,11 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
         // checks
         default:
             {
-                // The MTP head is dense-attention only on hybrid Qwen3.5/3.6, so use a plain
+                // The MTP head is dense-attention only on hybrid Qwen3-Next/3.5/3.6, so use a plain
                 // attention KV cache for the MTP context instead of the hybrid wrapper.
-                const bool mtp_on_hybrid_qwen35 =
+                const bool mtp_on_hybrid_qwen =
                     params.ctx_type == LLAMA_CONTEXT_TYPE_MTP &&
-                    (arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE);
+                    (arch == LLM_ARCH_QWEN3NEXT || arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE);
 
                 if (llm_arch_is_recurrent(arch)) {
                     res = new llama_memory_recurrent(
@@ -2227,7 +2238,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                             cparams.n_seq_max,
                             cparams.n_rs_seq,
                             nullptr);
-                } else if (llm_arch_is_hybrid(arch) && !mtp_on_hybrid_qwen35) {
+                } else if (llm_arch_is_hybrid(arch) && !mtp_on_hybrid_qwen) {
                     // The main difference between hybrid architectures is the
                     // layer filters, so pick the right one here
                     llama_memory_hybrid::layer_filter_cb filter_attn = nullptr;
@@ -2242,7 +2253,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         filter_recr = [&](uint32_t il) {
                             return hparams.is_recr(il) && hparams.n_ff(il) == 0;
                         };
-                    } else if (arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE) {
+                    } else if (arch == LLM_ARCH_QWEN3NEXT || arch == LLM_ARCH_QWEN35 || arch == LLM_ARCH_QWEN35MOE) {
                         filter_attn = [&](uint32_t il) {
                             return il < hparams.n_layer() && !hparams.is_recr(il);
                         };
@@ -2308,12 +2319,12 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params,
                         };
                     }
 
-                    if (mtp_on_hybrid_qwen35) {
+                    if (mtp_on_hybrid_qwen) {
                         filter = [&](uint32_t il) { return il >= hparams.n_layer(); };
                     }
 
                     if ((arch == LLM_ARCH_STEP35 || arch == LLM_ARCH_HY_V3 || arch == LLM_ARCH_GLM_DSA ||
-                            arch == LLM_ARCH_MIMO2) &&
+                            arch == LLM_ARCH_MIMO2 || arch == LLM_ARCH_DEEPSEEK32) &&
                             hparams.n_layer_nextn > 0) {
                         if (params.ctx_type == LLAMA_CONTEXT_TYPE_MTP) {
                             filter = [&](uint32_t il) { return il >= hparams.n_layer(); };
@@ -2684,6 +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:
@@ -2858,7 +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);
@@ -2877,6 +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);
@@ -2898,3 +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;
+}

src/llama-model.h

@@ -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;

src/llama-sampler.cpp

@@ -2638,7 +2638,8 @@ struct llama_sampler * llama_sampler_init_grammar_lazy_patterns(
 
 // penalties
 
-struct llama_sampler_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;
@@ -2648,10 +2649,50 @@ struct llama_sampler_penalties {
 
     // a frequency map to count token occurrences
     std::unordered_map<llama_token, int> token_count;
+
+    // backend graph inputs
+    ggml_tensor * inp_token_ids = nullptr;
+    ggml_tensor * inp_counts    = nullptr;
+
+    // backend helpers
+    int32_t n_max   = 0;
+    bool has_candidates = false;
+
+    std::vector<int32_t> host_token_ids;
+    std::vector<int32_t> host_counts;
+
+    static bool is_disabled(
+            int32_t penalty_last_n,
+            float   penalty_repeat,
+            float   penalty_freq,
+            float   penalty_present) {
+        return penalty_last_n == 0 ||
+            (penalty_repeat == 1.0f && penalty_freq == 0.0f && penalty_present == 0.0f);
+    }
+
+    bool is_disabled() const {
+        return is_disabled(penalty_last_n, penalty_repeat, penalty_freq, penalty_present);
+    }
+
+    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)
+        , penalty_present (penalty_present)
+        , prev            (penalty_last_n) {
+    }
 };
 
-static const char * llama_sampler_penalties_name(const struct llama_sampler * /*smpl*/) {
-    return "penalties";
+static const char * llama_sampler_penalties_name(const struct llama_sampler * smpl) {
+    auto * ctx = (llama_sampler_penalties *) smpl->ctx;
+    return ctx->get_name();
 }
 
 static void llama_sampler_penalties_accept(struct llama_sampler * smpl, llama_token token) {
@@ -2688,8 +2729,7 @@ static void llama_sampler_penalties_accept(struct llama_sampler * smpl, llama_to
 static void llama_sampler_penalties_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
     auto * ctx = (llama_sampler_penalties *) smpl->ctx;
 
-    if ((ctx->penalty_last_n == 0) ||
-        (ctx->penalty_repeat == 1.0f && ctx->penalty_freq == 0.0f && ctx->penalty_present == 0.0f)) {
+    if (ctx->is_disabled()) {
         return;
     }
 
@@ -2727,6 +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,
@@ -2736,7 +2777,8 @@ static struct llama_sampler * llama_sampler_penalties_clone(const struct llama_s
     {
         auto * result_ctx = (llama_sampler_penalties *) result->ctx;
 
-        result_ctx->prev = ctx->prev;
+        result_ctx->prev        = ctx->prev;
+        result_ctx->token_count = ctx->token_count;
     }
 
     return result;
@@ -2746,6 +2788,170 @@ static void llama_sampler_penalties_free(struct llama_sampler * smpl) {
     delete (llama_sampler_penalties *) smpl->ctx;
 }
 
+static bool llama_sampler_penalties_backend_init(
+        struct llama_sampler       * smpl,
+        ggml_backend_buffer_type_t   buft) {
+    auto * sctx = (llama_sampler_penalties *) smpl->ctx;
+
+    const bool res = llama_sampler_backend_support(smpl, buft);
+
+    sctx->init(res);
+
+    return res;
+}
+
+static void llama_sampler_penalties_backend_apply(
+        struct llama_sampler      * smpl,
+        struct ggml_context       * ctx,
+        struct ggml_cgraph        * gf,
+        struct llama_sampler_data * data) {
+    GGML_UNUSED(gf);
+
+    auto * sctx = (llama_sampler_penalties *) smpl->ctx;
+
+    if (sctx->is_disabled()) {
+        return;
+    }
+
+    GGML_ASSERT(sctx->n_vocab > 0);
+
+    sctx->has_candidates = data->candidates != nullptr;
+    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);
+    ggml_set_name(sctx->inp_token_ids, "penalties_token_ids");
+    ggml_set_input(sctx->inp_token_ids);
+
+    sctx->inp_counts = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sctx->n_max);
+    ggml_set_name(sctx->inp_counts, "penalties_counts");
+    ggml_set_input(sctx->inp_counts);
+
+    if ((int32_t) sctx->host_token_ids.size() != sctx->n_max) {
+        sctx->host_token_ids.assign(sctx->n_max, 0);
+        sctx->host_counts.assign(sctx->n_max, 0);
+    }
+
+    // flatten
+    ggml_tensor * logits = ggml_reshape_1d(ctx, data->logits, ggml_nelements(data->logits));
+    ggml_tensor * gathered = logits;
+    ggml_tensor * counts_f32 = ggml_cast(ctx, sctx->inp_counts, GGML_TYPE_F32);
+
+    if (sctx->has_candidates) {
+        ggml_tensor * candidates = ggml_reshape_1d(
+                ctx, data->candidates, ggml_nelements(data->candidates));
+        const int64_t n_candidates = candidates->ne[0];
+        GGML_ASSERT(n_candidates == ggml_nelements(logits));
+
+        ggml_tensor * counts_rows = ggml_fill(
+                ctx, ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, sctx->n_vocab), 0.0f);
+        ggml_tensor * scatter_rows = ggml_reshape_2d(ctx, counts_f32, 1, sctx->n_max);
+        counts_rows = ggml_set_rows(ctx, counts_rows, scatter_rows, sctx->inp_token_ids);
+        counts_f32 = ggml_get_rows(ctx, counts_rows, candidates);
+        counts_f32 = ggml_reshape_1d(ctx, counts_f32, n_candidates);
+    } else {
+        ggml_tensor * logits_rows = ggml_reshape_2d(ctx, logits, 1, ggml_nelements(logits));
+        gathered = ggml_get_rows(ctx, logits_rows, sctx->inp_token_ids);
+        gathered = ggml_reshape_1d(ctx, gathered, sctx->n_max);
+    }
+
+    ggml_tensor * active_mask = ggml_step(ctx, counts_f32);
+    ggml_tensor * inactive_mask = ggml_sub(ctx, ggml_fill(ctx, active_mask, 1.0f), active_mask);
+
+    ggml_tensor * penalized = gathered;
+
+    if (sctx->penalty_repeat != 1.0f) {
+        ggml_tensor * pos_mask = ggml_step(ctx, penalized);
+        ggml_tensor * neg_mask = ggml_sub(ctx, ggml_fill(ctx, pos_mask, 1.0f), pos_mask);
+
+        ggml_tensor * pos_scale = ggml_scale(ctx, pos_mask, 1.0f/sctx->penalty_repeat);
+        ggml_tensor * neg_scale = ggml_scale(ctx, neg_mask, sctx->penalty_repeat);
+        ggml_tensor * repeat_scale = ggml_add(ctx, pos_scale, neg_scale);
+
+        // scale inactive entries with 1 to avoid -INF * 0 = NaN for values masked by top-p
+        repeat_scale = ggml_mul(ctx, repeat_scale, active_mask);
+        repeat_scale = ggml_add(ctx, repeat_scale, inactive_mask);
+        penalized = ggml_mul(ctx, gathered, repeat_scale);
+    }
+
+    if (sctx->penalty_freq != 0.0f) {
+        ggml_tensor * penalty_freq = ggml_scale(ctx, counts_f32, sctx->penalty_freq);
+        penalized = ggml_sub(ctx, penalized, penalty_freq);
+    }
+
+    if (sctx->penalty_present != 0.0f) {
+        ggml_tensor * penalty_present = ggml_scale(ctx, active_mask, sctx->penalty_present);
+        penalized = ggml_sub(ctx, penalized, penalty_present);
+    }
+
+    if (sctx->has_candidates) {
+        data->logits = penalized;
+    } else {
+        ggml_tensor * logits_rows = ggml_reshape_2d(ctx, logits, 1, ggml_nelements(logits));
+        ggml_tensor * scatter_rows = ggml_reshape_2d(ctx, penalized, 1, sctx->n_max);
+        logits_rows = ggml_set_rows(ctx, logits_rows, scatter_rows, sctx->inp_token_ids);
+        data->logits = ggml_reshape_1d(ctx, logits_rows, ggml_nelements(logits));
+    }
+}
+
+static void llama_sampler_penalties_backend_set_input(struct llama_sampler * smpl) {
+    auto * sctx = (llama_sampler_penalties *) smpl->ctx;
+
+    if (!sctx->inp_token_ids || !sctx->inp_counts || sctx->n_max <= 0 || sctx->n_vocab <= 0) {
+        return;
+    }
+
+    if (sctx->is_disabled()) {
+        return;
+    }
+
+    // fill active entries from the map
+    int32_t n_active = 0;
+
+    for (const auto & it : sctx->token_count) {
+        GGML_ASSERT(n_active < sctx->n_max);
+        sctx->host_token_ids[n_active] = it.first;
+        sctx->host_counts   [n_active] = it.second;
+        ++n_active;
+    }
+
+    // Sorting is required because backend_apply uses ggml_set_rows (a scatter-back operation)
+    std::vector<std::pair<int32_t, int32_t>> entries;
+    entries.reserve(n_active);
+    for (int32_t i = 0; i < n_active; ++i) {
+        entries.emplace_back(sctx->host_token_ids[i], sctx->host_counts[i]);
+    }
+    std::sort(entries.begin(), entries.end(), [](const auto & a, const auto & b) {
+        return a.first < b.first;
+    });
+    for (int32_t i = 0; i < n_active; ++i) {
+        sctx->host_token_ids[i] = entries[i].first;
+        sctx->host_counts   [i] = entries[i].second;
+    }
+
+    // Padding: Finds a filler token id that is not present in token_count.
+    // Use it to do padding for the arrays, it avoids resizing every time.
+    // The arrays must always have exactly n_max entries (the GPU tensor is a fixed size).
+    int32_t filler = 0;
+    if (n_active < sctx->n_max) {
+        while (sctx->token_count.find(filler) != sctx->token_count.end()) {
+            ++filler;
+        }
+        GGML_ASSERT(filler < sctx->n_vocab);
+    }
+
+    // Fill the rest of the arrays with the filler token id and count 0.
+    // Inactive slots are padded with a unique dummy token ID (count = 0).
+    // The uniqueness matters because ggml_set_rows with duplicate indices can produce non-deterministic or incorrect results.
+    // Using a filler token with count 0 that isn't in the active set is safe, because the active_mask step in backend_apply filters them out via ggml_step(counts_f32)
+    for (int32_t i = n_active; i < sctx->n_max; ++i) {
+        sctx->host_token_ids[i] = filler;
+        sctx->host_counts   [i] = 0;
+    }
+
+    ggml_backend_tensor_set(sctx->inp_token_ids, sctx->host_token_ids.data(), 0, sctx->n_max * sizeof(int32_t));
+    ggml_backend_tensor_set(sctx->inp_counts,    sctx->host_counts.data(),    0, sctx->n_max * sizeof(int32_t));
+}
+
 static struct llama_sampler_i llama_sampler_penalties_i = {
     /* .name              = */ llama_sampler_penalties_name,
     /* .accept            = */ llama_sampler_penalties_accept,
@@ -2753,35 +2959,33 @@ static struct llama_sampler_i llama_sampler_penalties_i = {
     /* .reset             = */ llama_sampler_penalties_reset,
     /* .clone             = */ llama_sampler_penalties_clone,
     /* .free              = */ llama_sampler_penalties_free,
-    /* .backend_init      = */ nullptr,
+    /* .backend_init      = */ llama_sampler_penalties_backend_init,
     /* .backend_accept    = */ nullptr,
-    /* .backend_apply     = */ nullptr,
-    /* .backend_set_input = */ nullptr,
+    /* .backend_apply     = */ llama_sampler_penalties_backend_apply,
+    /* .backend_set_input = */ llama_sampler_penalties_backend_set_input,
 };
 
 struct llama_sampler * llama_sampler_init_penalties(
+        int32_t n_vocab,
         int32_t penalty_last_n,
         float penalty_repeat,
         float penalty_freq,
         float penalty_present) {
     penalty_last_n = std::max(penalty_last_n, 0);
 
-    const bool is_empty = (penalty_last_n == 0 || (penalty_repeat == 1.0f && penalty_freq == 0.0f && penalty_present == 0.0f));
-
-    if (is_empty) {
+    if (llama_sampler_penalties::is_disabled(
+                penalty_last_n, penalty_repeat, penalty_freq, penalty_present)) {
         return llama_sampler_init_empty("?penalties");
     }
 
     return llama_sampler_init(
         /* .iface = */ &llama_sampler_penalties_i,
-        /* .ctx   = */ new llama_sampler_penalties {
-            /* .penalty_last_n  = */ penalty_last_n,
-            /* .penalty_repeat  = */ penalty_repeat,
-            /* .penalty_freq    = */ penalty_freq,
-            /* .penalty_present = */ penalty_present,
-            /* .prev            = */ ring_buffer<llama_token>(penalty_last_n),
-            /* .token_count     = */ {},
-        }
+        /* .ctx   = */ new llama_sampler_penalties(
+            n_vocab,
+            penalty_last_n,
+            penalty_repeat,
+            penalty_freq,
+            penalty_present)
     );
 }
 
@@ -2874,8 +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;
@@ -2951,8 +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;
@@ -3165,7 +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
     {
@@ -3196,8 +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;
@@ -3234,23 +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

src/llama-sampler.h

@@ -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,

src/llama-vocab.cpp

@@ -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;
             }
@@ -2532,6 +2534,12 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
             const std::string & key = kv(std::get<0>(it));
             int32_t & id = std::get<1>(it);
 
+            if (id >= 0 && static_cast<size_t>(id) >= id_to_token.size()) {
+                LLAMA_LOG_WARN("%s: default special token '%s' = %d out of vocab range, disabling\n",
+                    __func__, key.c_str(), id);
+                id = LLAMA_TOKEN_NULL;
+            }
+
             uint32_t new_id;
             if (!ml.get_key(std::get<0>(it), new_id, false)) {
                 continue;
@@ -3619,12 +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;
                     }
                 }
 

src/models/deepseek2.cpp

@@ -37,6 +37,11 @@ void llama_model_deepseek2::load_arch_hparams(llama_model_loader & ml) {
         hparams.rope_yarn_log_mul /= 0.1f;
     }
 
+    // NextN/MTP
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);
+    GGML_ASSERT(hparams.n_layer_nextn == 0 ||
+        hparams.n_layer() + hparams.n_layer_nextn == hparams.n_layer_all);
+
     // (optional) temperature tuning - used by mistral-large
     ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE,  hparams.f_attn_temp_scale,       false);
     ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.n_attn_temp_floor_scale, false); // FIXME why not use temperature_length?
@@ -52,10 +57,20 @@ void llama_model_deepseek2::load_arch_hparams(llama_model_loader & ml) {
     }
 }
 
-void llama_model_deepseek2::load_arch_tensors(llama_model_loader &) {
+void llama_model_deepseek2::load_arch_tensors(llama_model_loader & ml) {
     LLAMA_LOAD_LOCALS;
     const int64_t n_expert_shared = hparams.n_expert_shared;
 
+    const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);
+    const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
+    const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
+    const int trunk_flags = mtp_only  ? TENSOR_NOT_REQUIRED : 0;
+    int       mtp_flags   = trunk_only ? TENSOR_NOT_REQUIRED : 0;
+
+    if (!ml.load_mtp) {
+        mtp_flags |= TENSOR_SKIP;
+    }
+
     const bool is_mla = hparams.is_mla();
 
     // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
@@ -81,44 +96,45 @@ void llama_model_deepseek2::load_arch_tensors(llama_model_loader &) {
         output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
     }
 
-    for (int i = 0; i < n_layer; ++i) {
+    for (int i = 0; i < n_layer_all; ++i) {
         auto & layer = layers[i];
+        const int flags = i < n_layer ? trunk_flags : mtp_flags;
 
-        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
         if (q_lora_rank > 0) {
-            layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, 0);
+            layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, flags);
         }
 
-        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
+        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, flags);
 
         if (q_lora_rank > 0) {
-            layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
-            layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
+            layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, flags);
+            layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, flags);
         } else {
-            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_embd_head_k_mla}, 0);
+            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_embd_head_k_mla}, flags);
         }
 
-        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, 0);
+        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, flags);
 
         // note: only old legacy GGUF files will have the unsplit wkv_b tensor in
         if (is_mla) {
-            layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, 0);
-            layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, 0);
+            layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, flags);
+            layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, flags);
         } else {
-            layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v_mla)}, 0);
+            layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v_mla)}, flags);
         }
 
-        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, flags);
 
-        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
 
         if (i < (int) hparams.n_layer_dense_lead) {
-            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
-            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
-            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);
         } else {
-            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
-            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | flags);
 
             if (n_expert == 0) {
                 throw std::runtime_error("n_expert must be > 0");
@@ -128,21 +144,281 @@ void llama_model_deepseek2::load_arch_tensors(llama_model_loader &) {
             }
 
             // MoE branch
-            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
-            create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, flags);
+            create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, flags);
 
             // Shared expert branch
-            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
-            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
-            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, flags);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+        }
+
+        // NextN/MTP tensors
+        if (i >= n_layer) {
+            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, mtp_flags);
+            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, mtp_flags);
+            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, mtp_flags);
+            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED | flags);
         }
     }
 }
 
 std::unique_ptr<llm_graph_context> llama_model_deepseek2::build_arch_graph(const llm_graph_params & params) const {
+    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {
+        return std::make_unique<graph_mtp>(*this, params);
+    }
     return std::make_unique<graph>(*this, params);
 }
 
+llama_model_deepseek2::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params) :
+    llm_graph_context(params) {
+    GGML_ASSERT(hparams.n_layer_nextn > 0 && "GLM4 MTP requires n_layer_nextn > 0");
+    GGML_ASSERT(hparams.n_layer_nextn == 1 && "GLM4 MTP currently only supports a single MTP block");
+    GGML_ASSERT(hparams.is_mla() && "GLM4 MTP requires MLA");
+    GGML_ASSERT(hparams.f_attn_temp_scale == 0.0f && "GLM4 MTP does not support attention temperature scaling");
+
+    // The appended MTP block is stored immediately after the main decoder layers.
+    const int il = hparams.n_layer();
+    const auto & layer = model.layers[il];
+
+    GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");
+    GGML_ASSERT(layer.nextn.enorm   && "MTP block missing nextn.enorm");
+    GGML_ASSERT(layer.nextn.hnorm   && "MTP block missing nextn.hnorm");
+
+    GGML_ASSERT((uint32_t) il >= hparams.n_layer_dense_lead && "GLM4 MTP block expected to use MoE FFN");
+
+    const int64_t n_embd_head_k_mla   = hparams.n_embd_head_k_mla();
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+    const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+    const int64_t kv_lora_rank        = hparams.n_lora_kv;
+
+    GGML_ASSERT(n_embd_head_qk_nope >= 1);
+    GGML_ASSERT(hparams.n_lora_q > 0);
+    GGML_ASSERT(layer.wq_a);
+    GGML_ASSERT(layer.attn_q_a_norm);
+    GGML_ASSERT(layer.wq_b);
+    GGML_ASSERT(layer.wkv_a_mqa);
+    GGML_ASSERT(layer.attn_kv_a_norm);
+    GGML_ASSERT(layer.wk_b);
+
+    const bool has_split_exps =
+            layer.ffn_up_exps   != nullptr &&
+            layer.ffn_gate_exps != nullptr;
+
+    const bool has_fused_exps = layer.ffn_gate_up_exps != nullptr;
+
+    GGML_ASSERT(has_split_exps || has_fused_exps);
+    GGML_ASSERT(layer.ffn_norm);
+    GGML_ASSERT(layer.ffn_gate_inp);
+    GGML_ASSERT(layer.ffn_down_exps);
+    GGML_ASSERT(layer.ffn_gate_shexp);
+    GGML_ASSERT(layer.ffn_down_shexp);
+    GGML_ASSERT(layer.ffn_up_shexp);
+
+    auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);
+
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    ggml_set_input(inp->tokens);
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);
+    ggml_set_input(inp->embd);
+
+    ggml_tensor * tok_embd;
+    if (ubatch.token) {
+        ggml_tensor * tok_embd_w = layer.nextn.embed_tokens
+                ? layer.nextn.embed_tokens
+                : model.tok_embd;
+
+        tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
+    } else {
+        tok_embd = inp->embd;
+    }
+    cb(tok_embd, "mtp_tok_embd", il);
+
+    inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
+    ggml_set_input(inp->h);
+    ggml_set_name(inp->h, "mtp_h_input");
+
+    ggml_tensor * h_embd = inp->h;
+
+    res->add_input(std::move(inp));
+
+    ggml_tensor * inp_pos     = build_inp_pos();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    auto * inp_attn_k = build_attn_inp_k();
+
+    ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
+    cb(h_norm, "mtp_hnorm", il);
+
+    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
+    cb(e_norm, "mtp_enorm", il);
+
+    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, 0);
+    cb(concat, "mtp_concat", il);
+
+    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);
+    cb(cur, "mtp_eh_proj", il);
+
+    ggml_tensor * inpSA = cur;
+
+    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_attn_norm", il);
+
+    ggml_tensor * q = ggml_mul_mat(ctx0, layer.wq_a, cur);
+    cb(q, "mtp_q_a", il);
+
+    q = build_norm(q, layer.attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+    cb(q, "mtp_q_a_norm", il);
+
+    q = ggml_mul_mat(ctx0, layer.wq_b, q);
+    cb(q, "mtp_q_b", il);
+
+    ggml_tensor * q_nope =
+        ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
+                ggml_row_size(q->type, n_embd_head_k_mla),
+                ggml_row_size(q->type, n_embd_head_k_mla) * n_head, 0);
+    cb(q_nope, "mtp_q_nope", il);
+
+    ggml_tensor * q_pe =
+        ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
+                ggml_row_size(q->type, n_embd_head_k_mla),
+                ggml_row_size(q->type, n_embd_head_k_mla) * n_head,
+                ggml_row_size(q->type, n_embd_head_qk_nope));
+    cb(q_pe, "mtp_q_pe", il);
+
+    ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, layer.wkv_a_mqa, cur);
+    cb(kv_cmpr_pe, "mtp_kv_cmpr_pe", il);
+
+    ggml_tensor * kv_cmpr =
+        ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+                ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+    cb(kv_cmpr, "mtp_kv_cmpr", il);
+
+    ggml_tensor * k_pe =
+        ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+                ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+    cb(k_pe, "mtp_k_pe", il);
+
+    kv_cmpr = build_norm(kv_cmpr, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+    cb(kv_cmpr, "mtp_kv_cmpr_norm", il);
+
+    GGML_ASSERT(ext_factor >= 0.0f);
+
+    const float attn_factor_org =
+            attn_factor * (1.0f + 0.1f * logf(1.0f / freq_scale));
+
+    const float mscale =
+            attn_factor_org * (1.0f + 0.1f * hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+
+    const float kq_scale =
+            1.0f * mscale * mscale / sqrtf(float(n_embd_head_k_mla));
+
+    q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr,
+            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+    cb(q_pe, "mtp_q_pe_rope", il);
+
+    k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr,
+            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+    cb(k_pe, "mtp_k_pe_rope", il);
+
+    q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+    cb(q_nope, "mtp_q_nope_perm", il);
+
+    ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, layer.wk_b, q_nope);
+    cb(q_nope_absorbed, "mtp_q_nope_absorbed", il);
+
+    q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+    cb(q_nope_absorbed, "mtp_q_nope_absorbed_perm", il);
+
+    ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
+    cb(Qcur, "mtp_Qcur", il);
+
+    kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, hparams.n_lora_kv, 1, n_tokens);
+    cb(kv_cmpr, "mtp_kv_cmpr_reshape", il);
+
+    ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
+    cb(Kcur, "mtp_Kcur", il);
+
+    ggml_tensor * Vcur = kv_cmpr;
+    cb(Vcur, "mtp_Vcur", il);
+
+    cur = build_attn(inp_attn_k,
+            layer.wo, nullptr, layer.wo_s,
+            Qcur, Kcur, Vcur, nullptr, nullptr, layer.wv_b, kq_scale, il);
+    cb(cur, "mtp_attn_out", il);
+
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "mtp_ffn_inp", il);
+
+    cur = build_norm(ffn_inp, layer.ffn_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_ffn_norm", il);
+
+    ggml_tensor * moe_out = build_moe_ffn(cur,
+            layer.ffn_gate_inp,
+            layer.ffn_up_exps,
+            layer.ffn_gate_exps,
+            layer.ffn_down_exps,
+            layer.ffn_exp_probs_b,
+            n_expert, n_expert_used,
+            LLM_FFN_SILU, hparams.expert_weights_norm,
+            hparams.expert_weights_scale,
+            (llama_expert_gating_func_type) hparams.expert_gating_func,
+            il,
+            nullptr,
+            layer.ffn_gate_up_exps);
+    cb(moe_out, "mtp_ffn_moe_out", il);
+
+    ggml_tensor * ffn_shexp = build_ffn(cur,
+            layer.ffn_up_shexp, nullptr, nullptr,
+            layer.ffn_gate_shexp, nullptr, nullptr,
+            layer.ffn_down_shexp, nullptr, nullptr,
+            nullptr, LLM_FFN_SILU, LLM_FFN_PAR, il);
+    cb(ffn_shexp, "mtp_ffn_shexp", il);
+
+    cur = ggml_add(ctx0, moe_out, ffn_shexp);
+    cb(cur, "mtp_ffn_out", il);
+
+    cur = ggml_add(ctx0, cur, ffn_inp);
+    cb(cur, "mtp_post_ffn", il);
+
+    ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
+            ? layer.nextn.shared_head_norm
+            : model.output_norm;
+    GGML_ASSERT(head_norm_w && "GLM4 MTP: missing both nextn.shared_head_norm and output_norm");
+
+    cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+    cb(cur, "mtp_shared_head_norm", -1);
+
+    ggml_tensor * head_w = layer.nextn.shared_head_head
+            ? layer.nextn.shared_head_head
+            : model.output;
+
+    ggml_tensor * head_s = layer.nextn.shared_head_head
+            ? layer.nextn.shared_head_head_s
+            : model.output_s;
+
+    GGML_ASSERT(head_w && "GLM4 MTP: missing LM head (nextn.shared_head_head or model.output)");
+
+    cur = build_lora_mm(head_w, cur, head_s);
+    cb(cur, "result_output", -1);
+
+    res->t_logits = cur;
+    ggml_build_forward_expand(gf, cur);
+}
+
 llama_model_deepseek2::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
@@ -365,7 +641,7 @@ llama_model_deepseek2::graph::graph(const llama_model & model, const llm_graph_p
                             Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
             }
         }
-        if (il == n_layer - 1 && inp_out_ids) {
+        if (il == n_layer - 1 && inp_out_ids && (!cparams.embeddings_nextn || cparams.embeddings_nextn_masked)) {
             cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }
@@ -425,6 +701,13 @@ llama_model_deepseek2::graph::graph(const llama_model & model, const llm_graph_p
 
     cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
 
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (cparams.embeddings_nextn && !cparams.embeddings_nextn_masked && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
     cb(cur, "result_norm", -1);
     res->t_embd = cur;
 

src/models/deepseek32.cpp

@@ -44,13 +44,24 @@ void llama_model_deepseek32::load_arch_hparams(llama_model_loader & ml) {
     GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer");
 
     switch (hparams.n_layer()) {
-        case 62: type = LLM_TYPE_685B_A37B; break;
+        case 61: type = LLM_TYPE_685B_A37B; break;
         default: type = LLM_TYPE_UNKNOWN;
     }
 }
 
-void llama_model_deepseek32::load_arch_tensors(llama_model_loader &) {
+void llama_model_deepseek32::load_arch_tensors(llama_model_loader & ml) {
     LLAMA_LOAD_LOCALS;
+
+    const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);
+    const std::string mtp_probe = "blk." + std::to_string(n_layer) + ".nextn.eh_proj.weight";
+    const bool trunk_only = (hparams.n_layer_nextn > 0) && (ml.get_weight(mtp_probe.c_str()) == nullptr);
+    const int trunk_flags = mtp_only   ? TENSOR_NOT_REQUIRED : 0;
+    int       mtp_flags   = trunk_only ? TENSOR_NOT_REQUIRED : 0;
+
+    if (!ml.load_mtp) {
+        mtp_flags |= TENSOR_SKIP;
+    }
+
     const bool is_mla = hparams.is_mla();
     if (!is_mla) {
         throw std::runtime_error("DEEPSEEK32 architecture requires MLA");
@@ -80,12 +91,7 @@ void llama_model_deepseek32::load_arch_tensors(llama_model_loader &) {
     }
 
     for (int i = 0; i < n_layer_all; ++i) {
-        int flags = 0;
-        if (i >= n_layer) {
-            // skip all tensors in the NextN layers
-            // TODO @ngxson : TENSOR_NOT_REQUIRED was a hack, need to remove it later
-            flags |= TENSOR_SKIP | TENSOR_NOT_REQUIRED;
-        }
+        const int flags = (i >= n_layer) ? mtp_flags : trunk_flags;
 
         auto & layer = layers[i];
 
@@ -138,7 +144,7 @@ void llama_model_deepseek32::load_arch_tensors(llama_model_loader &) {
             layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
         }
 
-        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+        // NextN/MTP tensors - conditionally load for last nextn_predict_layers
         if (i >= n_layer) {
             layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
             layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
@@ -153,6 +159,9 @@ void llama_model_deepseek32::load_arch_tensors(llama_model_loader &) {
 }
 
 std::unique_ptr<llm_graph_context> llama_model_deepseek32::build_arch_graph(const llm_graph_params & params) const {
+    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {
+        return std::make_unique<graph_mtp>(*this, params);
+    }
     return std::make_unique<graph>(*this, params);
 }
 
@@ -430,7 +439,9 @@ llama_model_deepseek32::graph::graph(const llama_model & model, const llm_graph_
                         Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, top_k, kq_scale, il);
             }
         }
-        if (il == n_layer - 1 && inp_out_ids) {
+        // when unmasked nextn embeddings are requested, t_h_nextn must keep all rows,
+        // so the early output masking has to be skipped (it is applied after the final norm instead)
+        if (il == n_layer - 1 && inp_out_ids && (!cparams.embeddings_nextn || cparams.embeddings_nextn_masked)) {
             cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }
@@ -493,6 +504,14 @@ llama_model_deepseek32::graph::graph(const llama_model & model, const llm_graph_
 
     cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
 
+    // post-norm hidden state feeds the NextN/MTP draft head
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (cparams.embeddings_nextn && !cparams.embeddings_nextn_masked && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
     cb(cur, "result_norm", -1);
     res->t_embd = cur;
 
@@ -504,3 +523,243 @@ llama_model_deepseek32::graph::graph(const llama_model & model, const llm_graph_
 
     ggml_build_forward_expand(gf, cur);
 }
+
+// LLM_GRAPH_TYPE_DECODER_MTP draft head for DeepSeek V3.2 (DEEPSEEK32).
+// Semantics mirror the deepseek-family NextN/MTP layer:
+//   enorm(embed) + hnorm(prev_hidden) -> concat(e, h) -> eh_proj ->
+//   full deepseek32 decoder block (dense MLA attention + sigmoid-gated MoE FFN
+//   with shared expert, exactly as the trunk deepseek2 graph builds it) ->
+//   shared_head_norm (fallback output_norm) -> shared LM head.
+// The DSA indexer is not used at runtime.
+llama_model_deepseek32::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)
+    : llm_graph_context(params) {
+    GGML_ASSERT(hparams.n_layer_nextn > 0 && "DEEPSEEK32 MTP requires n_layer_nextn > 0");
+    GGML_ASSERT(hparams.n_layer_nextn == 1 && "DEEPSEEK32 MTP currently only supports a single MTP block");
+    GGML_ASSERT(hparams.is_mla() && "DEEPSEEK32 MTP requires MLA");
+
+    const int il = hparams.n_layer() + cparams.nextn_layer_offset;
+    GGML_ASSERT(cparams.nextn_layer_offset >= 0 &&
+                cparams.nextn_layer_offset < (int) hparams.n_layer_nextn &&
+                "nextn_layer_offset out of range [0, n_layer_nextn)");
+    const auto & layer = model.layers[il];
+
+    GGML_ASSERT(layer.nextn.eh_proj && "MTP block missing nextn.eh_proj");
+    GGML_ASSERT(layer.nextn.enorm   && "MTP block missing nextn.enorm");
+    GGML_ASSERT(layer.nextn.hnorm   && "MTP block missing nextn.hnorm");
+    GGML_ASSERT(layer.ffn_gate_inp  && "MTP block missing ffn_gate_inp");
+
+    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+    const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+    const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
+
+    const uint32_t kv_lora_rank = hparams.n_lora_kv;
+
+    // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
+    // See the deepseek2 trunk graph for the detailed explanation - this must match it EXACTLY.
+    GGML_ASSERT(ext_factor >= 0.0f);
+    const float attn_factor_org = attn_factor * (1.0f + 0.1f * logf(1.0f / freq_scale));
+
+    const float mscale   = attn_factor_org * (1.0f + 0.1f * hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
+    const float kq_scale = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k));
+
+    // TODO: extract in a common llm_graph_context::build_inp_embd_h()
+    auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);
+
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    ggml_set_input(inp->tokens);
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);
+    ggml_set_input(inp->embd);
+
+    ggml_tensor * tok_embd;
+    if (ubatch.token) {
+        ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
+
+        tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
+    } else {
+        tok_embd = inp->embd;
+    }
+    cb(tok_embd, "mtp_tok_embd", il);
+
+    inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
+    ggml_set_input(inp->h);
+    ggml_set_name(inp->h, "mtp_h_input");
+
+    ggml_tensor * h_embd = inp->h;
+
+    res->add_input(std::move(inp));
+
+    ggml_tensor * inp_pos     = build_inp_pos();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    // MLA with the absorption optimization uses a K-only cache (V is a view of K)
+    auto * inp_attn = build_attn_inp_k();
+
+    ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
+    cb(h_norm, "mtp_hnorm", il);
+
+    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
+    cb(e_norm, "mtp_enorm", il);
+
+    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0);
+    cb(concat, "mtp_concat", il);
+
+    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);
+    cb(cur, "mtp_eh_proj", il);
+
+    ggml_tensor * inpSA = cur;
+
+    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_attn_norm", il);
+
+    // self-attention: dense MLA, same construction as the deepseek2 trunk graph
+    {
+        ggml_tensor * q = ggml_mul_mat(ctx0, layer.wq_a, cur);
+        cb(q, "mtp_q", il);
+
+        q = build_norm(q, layer.attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
+        cb(q, "mtp_q", il);
+
+        q = ggml_mul_mat(ctx0, layer.wq_b, q);
+        cb(q, "mtp_q", il);
+
+        // split into {n_embd_head_qk_nope, n_head, n_tokens}
+        ggml_tensor * q_nope =
+            ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+                         ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
+        cb(q_nope, "mtp_q_nope", il);
+
+        // and {n_embd_head_qk_rope, n_head, n_tokens}
+        ggml_tensor * q_pe = ggml_view_3d(
+            ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
+            ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope));
+        cb(q_pe, "mtp_q_pe", il);
+
+        ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, layer.wkv_a_mqa, cur);
+        cb(kv_cmpr_pe, "mtp_kv_cmpr_pe", il);
+
+        // split into {kv_lora_rank, n_tokens}
+        ggml_tensor * kv_cmpr =
+            ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
+                         ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
+        cb(kv_cmpr, "mtp_kv_cmpr", il);
+
+        // and {n_embd_head_qk_rope, 1, n_tokens}
+        ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
+                                          ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                          ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
+                                          ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
+        cb(k_pe, "mtp_k_pe", il);
+
+        q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+        cb(q_pe, "mtp_q_pe", il);
+
+        k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                             ext_factor, attn_factor, beta_fast, beta_slow);
+        cb(k_pe, "mtp_k_pe", il);
+
+        kv_cmpr = build_norm(kv_cmpr, layer.attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
+        cb(kv_cmpr, "mtp_kv_cmpr", il);
+
+        // {n_embd_head_qk_nope, n_tokens, n_head}
+        q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
+        cb(q_nope, "mtp_q_nope_perm", il);
+
+        // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
+        ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, layer.wk_b, q_nope);
+        cb(q_nope_absorbed, "mtp_q_nope_absorbed", il);
+
+        // {kv_lora_rank, n_head, n_tokens}
+        q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
+        cb(q_nope_absorbed, "mtp_q_nope_absorbed_perm", il);
+
+        // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
+        // note: rope must go first for in-place context shifting in build_rope_shift()
+        ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
+        cb(Qcur, "mtp_Qcur", il);
+
+        kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
+        cb(kv_cmpr, "mtp_kv_cmpr_reshape", il);
+
+        // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
+        ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
+        cb(Kcur, "mtp_Kcur", il);
+
+        // {kv_lora_rank, 1, n_tokens}
+        ggml_tensor * Vcur = kv_cmpr;
+        cb(Vcur, "mtp_Vcur", il);
+
+        // note: MLA with the absorption optimization converts into MQA (ie: GQA with 1 group)
+        cur = build_attn(inp_attn,
+                layer.wo, NULL, layer.wo_s,
+                Qcur, Kcur, Vcur, nullptr, nullptr, layer.wv_b, kq_scale, il);
+        cb(cur, "mtp_attn_out", il);
+    }
+
+    ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+    cb(ffn_inp, "mtp_ffn_inp", il);
+
+    cur = build_norm(ffn_inp, layer.ffn_norm, NULL, LLM_NORM_RMS, il);
+    cb(cur, "mtp_ffn_norm", il);
+
+    // MoE FFN with shared expert - same construction as the deepseek2 trunk graph
+    ggml_tensor * moe_out = build_moe_ffn(cur,
+        layer.ffn_gate_inp,
+        layer.ffn_up_exps,
+        layer.ffn_gate_exps,
+        layer.ffn_down_exps,
+        layer.ffn_exp_probs_b,
+        n_expert, n_expert_used,
+        LLM_FFN_SILU, hparams.expert_weights_norm,
+        hparams.expert_weights_scale,
+        (llama_expert_gating_func_type) hparams.expert_gating_func,
+        il,
+        nullptr,
+        layer.ffn_gate_up_exps,
+        layer.ffn_up_exps_s,
+        layer.ffn_gate_exps_s,
+        layer.ffn_down_exps_s);
+    cb(moe_out, "mtp_ffn_moe_out", il);
+
+    // FFN shared expert
+    ggml_tensor * ffn_shexp =
+        build_ffn(cur,
+            layer.ffn_up_shexp, NULL, layer.ffn_up_shexp_s,
+            layer.ffn_gate_shexp, NULL, layer.ffn_gate_shexp_s,
+            layer.ffn_down_shexp, NULL, layer.ffn_down_shexp_s,
+            NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
+    cb(ffn_shexp, "mtp_ffn_shexp", il);
+
+    cur = ggml_add(ctx0, moe_out, ffn_shexp);
+    cb(cur, "mtp_ffn_out", il);
+
+    cur = ggml_add(ctx0, cur, ffn_inp);
+    cb(cur, "mtp_post_ffn", il);
+
+    // shared_head_norm applied after the decoder block, before the shared LM head.
+    // The post-norm hidden state seeds the next MTP step.
+    ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
+            ? layer.nextn.shared_head_norm
+            : model.output_norm;
+    GGML_ASSERT(head_norm_w && "DEEPSEEK32 MTP: missing both nextn.shared_head_norm and output_norm");
+    cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);
+
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    cb(cur, "mtp_shared_head_norm", -1);
+
+    ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;
+    ggml_tensor * head_s = layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : model.output_s;
+    GGML_ASSERT(head_w && "DEEPSEEK32 MTP: missing LM head (nextn.shared_head_head or model.output)");
+    cur = build_lora_mm(head_w, cur, head_s);
+    cb(cur, "result_output", -1);
+
+    res->t_logits = cur;
+    ggml_build_forward_expand(gf, cur);
+}
+

src/models/deepseek4.cpp

@@ -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);

src/models/dflash.cpp

@@ -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);

src/models/minimax-m3.cpp

@@ -1,5 +1,5 @@
 #include "models.h"
-#include "llama-kv-cache.h"
+#include "llama-kv-cache-msa.h"
 #include <cmath>
 #include <vector>
 #include <cstdint>
@@ -7,7 +7,8 @@
 // MiniMax-M3: MiniMax-M2 style GQA (per-head QK-norm, partial rotary) with
 // DeepSeek-V3 leading-dense + routed/shared experts (sigmoid gating, routed scaling),
 // swigluoai activation, and MiniMax Sparse Attention (MSA). MTP is not in released model weights.
-// Notes: Blocks are anchored to absolute KV cache slots.
+// MSA blocks are defined over token positions. The graph translates between position space (block
+// selection) and cell space (K/V/indexer storage) via per-ubatch pos<->cell maps populated from llama_kv_cells
 
 void llama_model_minimax_m3::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -23,7 +24,6 @@ void llama_model_minimax_m3::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_ATTENTION_INDEXER_BLOCK_SIZE,    hparams.indexer_block_size);
     ml.get_key(LLM_KV_ATTENTION_INDEXER_LOCAL_BLOCKS,  hparams.indexer_local_blocks);
     msa_p = { (int) hparams.indexer_block_size, (int) hparams.indexer_top_k, (int) hparams.indexer_local_blocks };
-    hparams.indexer_kv = true;
 
     switch (hparams.n_layer()) {
         case 60: type = LLM_TYPE_428B_A23B; break;
@@ -86,43 +86,83 @@ std::unique_ptr<llm_graph_context> llama_model_minimax_m3::build_arch_graph(cons
     return std::make_unique<graph>(*this, params);
 }
 
-// per-query local-force bias for MSA selection
-// local window always wins a slot
-class llm_graph_input_msa_local : public llm_graph_input_i {
+class llm_graph_input_msa : public llm_graph_input_i {
 public:
-    llm_graph_input_msa_local(int blk, int local, int64_t nblk) : blk(blk), local(local), nblk(nblk) {}
+    llm_graph_input_msa(const llama_kv_cache_msa_context * mctx, int blk, int local) :
+        mctx(mctx), blk(blk), local(local) {}
 
     void set_input(const llama_ubatch * ubatch) override {
-        if (!bias || !ubatch->pos) {
-            return;
-        }
-        const int64_t n_tokens = ubatch->n_tokens;
-        std::vector<float> data((size_t) nblk * n_tokens, 0.0f);
-        for (int64_t i = 0; i < n_tokens; ++i) {
-            const int64_t L = ubatch->pos[i] / blk;
-            for (int l = 0; l < local && L - l >= 0; ++l) {
-                if (L - l < nblk) {
-                    data[(size_t) i * nblk + (L - l)] = 1e30f;
+        if (pos_slot_i) { mctx->set_input_pos_slot(pos_slot_i, ubatch); }
+        if (pos_slot_f) { mctx->set_input_pos_slot(pos_slot_f, ubatch); }
+        if (cell_blk)   { mctx->set_input_cell_pos(cell_blk, ubatch, blk); }
+        if (pos_mask)   { mctx->set_input_pos_mask(pos_mask, ubatch); }
+
+        // local-force bias over position blocks
+        if (bias && ubatch->pos) {
+            const int64_t n_tokens = ubatch->n_tokens;
+            const int64_t nblk     = bias->ne[0];
+            std::vector<float> data((size_t) nblk * n_tokens, 0.0f);
+            for (int64_t i = 0; i < n_tokens; ++i) {
+                const int64_t L = ubatch->pos[i] / blk;
+                for (int l = 0; l < local && L - l >= 0; ++l) {
+                    if (L - l < nblk) {
+                        data[(size_t) i * nblk + (L - l)] = 1e30f;
+                    }
                 }
             }
+            ggml_backend_tensor_set(bias, data.data(), 0, data.size() * sizeof(float));
         }
-        ggml_backend_tensor_set(bias, data.data(), 0, data.size() * sizeof(float));
     }
 
-    // valid as long as the bias tensor dims still match the new ubatch/cache window
+    // valid as long as the tensor dims still match the new ubatch/cache window and the
+    // ubatch is in the same regime (decode graphs have pos_slot_f, batch graphs cell_blk)
     bool can_reuse(const llm_graph_params & params) override {
-        const auto * mctx = static_cast<const llama_kv_cache_context *>(params.mctx);
+        const auto * mctx_new = static_cast<const llama_kv_cache_msa_context *>(params.mctx);
+
+        this->mctx = mctx_new;
+
+        const int64_t n_ps = GGML_PAD((int64_t) mctx_new->get_n_pos(), blk);
+        const int64_t ns   = params.cparams.kv_unified ? 1 : params.ubatch.n_seqs_unq;
+
+        const bool decode = params.ubatch.n_tokens == ns;   // one token per stream
 
         bool res = true;
-        res &= bias->ne[1] == params.ubatch.n_tokens;
-        res &= bias->ne[0] * blk == (int64_t) mctx->get_n_kv();
+
+        res &= bias->ne[0] * blk == n_ps;
+        res &= bias->ne[1]       == params.ubatch.n_tokens;
+
+        res &= pos_mask->ne[0] == n_ps;
+        res &= pos_mask->ne[1] == params.ubatch.n_tokens;
+
+        res &= pos_slot_i->ne[0] == n_ps;
+        res &= pos_slot_i->ne[1] == ns;
+
+        res &= decode == (pos_slot_f != nullptr);
+        res &= decode == (cell_blk   == nullptr);
+
+        if (pos_slot_f) {
+            res &= pos_slot_f->ne[0] == n_ps;
+            res &= pos_slot_f->ne[1] == ns;
+        }
+
+        if (cell_blk) {
+            res &= cell_blk->ne[0] == (int64_t) mctx_new->get_base()->get_n_kv();
+            res &= cell_blk->ne[1] == ns;
+        }
+
         return res;
     }
 
-    ggml_tensor * bias = nullptr;
-    int     blk;
-    int     local;
-    int64_t nblk;
+    ggml_tensor * bias       = nullptr; // F32 [nblk, n_tokens] local-force bias (position blocks)
+    ggml_tensor * pos_mask   = nullptr; // F32 [n_ps, n_tokens] 0/-inf visibility, by position
+    ggml_tensor * pos_slot_i = nullptr; // I32 [n_ps, ns]       pos -> cell (get_rows index)
+    ggml_tensor * pos_slot_f = nullptr; // F32 [n_ps, ns]       pos -> cell (gatherable values, decode)
+    ggml_tensor * cell_blk   = nullptr; // I32 [n_kv, ns]       cell -> position block (batch)
+
+    const llama_kv_cache_msa_context * mctx;
+
+    int blk;
+    int local;
 };
 
 // One FA call for all GQA groups (and at multi-stream decode, all streams) by mapping them onto the FA sequence dim (ne[3])
@@ -173,7 +213,9 @@ llama_model_minimax_m3::graph::graph(const llama_model & model, const llm_graph_
     inpL = build_inp_embd(model.tok_embd);
 
     ggml_tensor * inp_pos = build_inp_pos();
-    auto inp_attn = build_attn_inp_kv();
+
+    // ==========================================
+    // TODO: avoid such kind of complexity in the model graphs
 
     // MSA calls ggml_flash_attn_ext directly and assumes the non-transposed V layout that
     // llama.cpp only provides when flash attention is enabled. Block selection is anchored
@@ -185,6 +227,8 @@ llama_model_minimax_m3::graph::graph(const llama_model & model, const llm_graph_
     const bool streams_ok  = cparams.n_seq_max == 1 || !cparams.kv_unified;
     const bool msa_enabled = fa_on && streams_ok;
 
+    auto * inp_attn = build_attn_inp_kv_msa(msa_enabled);
+
     static bool warned_no_fa = false;
     if (!fa_on && !warned_no_fa) {
         LLAMA_LOG_WARN("%s: flash attention disabled; MSA requires it -> running DENSE attention "
@@ -197,36 +241,54 @@ llama_model_minimax_m3::graph::graph(const llama_model & model, const llm_graph_
                        "-> running DENSE attention. Output may be degraded. Drop --kv-unified to enable MSA.\n", __func__);
         warned_unified = true;
     }
+    // ==========================================
 
     // hoisted per-graph MSA state (shared by every sparse layer)
-    llm_graph_input_msa_local * msa_loc = nullptr;
+    llm_graph_input_msa * msa = nullptr;
     ggml_tensor * msa_kqm = nullptr;
-    ggml_tensor * msa_mf  = nullptr;
-    int64_t n_kv = 0, nblk = 0, ns = 1, n_tps = 0;
+    ggml_tensor * msa_mf  = nullptr;   // F32 copy of the FA mask for the final mask add
+    int64_t n_kv = 0, n_ps = 0, nblk = 0, ns = 1, n_tps = 0;
     bool msa_decode = false;           // gather (1 token per stream) vs mask
     const int     blk = mm.msa_p.blk;
     const int64_t Hd  = hparams.indexer_n_head;   // one indexer head per GQA group
 
     if (msa_enabled) {
+        const auto * mctx_msa = static_cast<const llama_kv_cache_msa_context *>(mctx);
+
         msa_kqm = inp_attn->get_kq_mask();
         n_kv  = msa_kqm->ne[0];
         n_tps = msa_kqm->ne[1];        // tokens per stream
         ns    = msa_kqm->ne[3];        // streams in this ubatch
         GGML_ASSERT(msa_kqm->type == GGML_TYPE_F16 && "MSA requires the FA (f16) mask");
         GGML_ASSERT(n_tps*ns == n_tokens);
-        GGML_ASSERT(n_kv % blk == 0 &&
-            "MSA: KV/mask n_kv must be a multiple of indexer.block_size (128); "
-            "the flash-attention KV padding must be a multiple of the block size. "
-            "A non-multiple would silently drop the partial tail block.");
-        nblk = n_kv / blk;
+
+        // the position axis covers every position currently in the cache and is padded to whole blocks
+        n_ps = GGML_PAD((int64_t) mctx_msa->get_n_pos(), blk);
+        nblk = n_ps / blk;
         msa_decode = n_tps == 1;
 
-        msa_mf = ggml_cast(ctx0, msa_kqm, GGML_TYPE_F32);
+        auto inp = std::make_unique<llm_graph_input_msa>(mctx_msa, blk, mm.msa_p.local);
 
-        auto loc = std::make_unique<llm_graph_input_msa_local>(blk, mm.msa_p.local, nblk);
-        loc->bias = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, nblk, n_tokens);  // stream-grouped tokens
-        ggml_set_input(loc->bias);
-        msa_loc = (llm_graph_input_msa_local *) res->add_input(std::move(loc));
+        inp->bias = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, nblk, n_tokens);  // stream-grouped tokens
+        ggml_set_input(inp->bias);
+
+        inp->pos_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_ps, n_tokens);
+        ggml_set_input(inp->pos_mask);
+
+        inp->pos_slot_i = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_ps, ns);
+        ggml_set_input(inp->pos_slot_i);
+
+        if (msa_decode) {
+            inp->pos_slot_f = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_ps, ns);
+            ggml_set_input(inp->pos_slot_f);
+        } else {
+            inp->cell_blk = ggml_new_tensor_2d(ctx0, GGML_TYPE_I32, n_kv, ns);
+            ggml_set_input(inp->cell_blk);
+
+            msa_mf = ggml_cast(ctx0, msa_kqm, GGML_TYPE_F32);
+        }
+
+        msa = (llm_graph_input_msa *) res->add_input(std::move(inp));
     }
 
     ggml_tensor * inp_out_ids = build_inp_out_ids();
@@ -283,9 +345,11 @@ llama_model_minimax_m3::graph::graph(const llama_model & model, const llm_graph_
                 ik = ggml_rope_ext(ctx0, ik, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig,
                                    freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
 
-                const auto * mctx_cur = inp_attn->mctx;
-                ggml_build_forward_expand(gf, mctx_cur->cpy_k_idx(ctx0, ik, inp_attn->get_k_idxs(), il));
-                ggml_tensor * ik_kv = mctx_cur->get_k_idx(ctx0, il);
+                const auto * mctx_msa_l = static_cast<const llama_kv_cache_msa_context *>(mctx);
+                const auto * mctx_cur = mctx_msa_l->get_base();
+                const auto * mctx_idx = mctx_msa_l->get_idx();
+                ggml_build_forward_expand(gf, mctx_idx->cpy_k(ctx0, ik, inp_attn->get_k_idxs_idx(), il));
+                ggml_tensor * ik_kv = mctx_idx->get_k(ctx0, il);
 
                 if (inp_attn->self_k_rot) {
                     Qcur = llama_mul_mat_hadamard(ctx0, Qcur, inp_attn->self_k_rot);
@@ -316,42 +380,52 @@ llama_model_minimax_m3::graph::graph(const llama_model & model, const llm_graph_
 
                 if (msa_decode) {
                     // decode: batched over streams top-k + gather, one grouped FA
-                    // scores: per-stream batched matmul over the stream dim (ne[3]).
-                    // the cache views are not contiguous across streams (stride = kv_size, not n_kv)
-                    ggml_tensor * ikv4 = ggml_view_4d(ctx0, ik_kv, n_idx_dim, n_kv, 1, ns,
-                            ik_kv->nb[2], ik_kv->nb[3], ik_kv->nb[3], 0);
+                    // gather the indexer keys through the pos -> cell map
+                    ggml_tensor * ik3 = ggml_view_3d(ctx0, ik_kv, n_idx_dim, n_kv, ns,
+                            ik_kv->nb[2], ik_kv->nb[3], 0);
+                    ggml_tensor * ikp = ggml_get_rows(ctx0, ik3, msa->pos_slot_i);   // [n_idx_dim, n_ps, ns]
                     ggml_tensor * iq4 = ggml_reshape_4d(ctx0, iq, n_idx_dim, Hd, 1, ns);
-                    ggml_tensor * sc  = ggml_mul_mat(ctx0, ikv4, iq4);
+                    ggml_tensor * sc  = ggml_mul_mat(ctx0,
+                            ggml_reshape_4d(ctx0, ikp, n_idx_dim, n_ps, 1, ns), iq4);
                     ggml_mul_mat_set_prec(sc, GGML_PREC_F32);
-                    sc = ggml_add_inplace(ctx0, sc, msa_mf);
+                    // unmapped positions come out -inf, so they can never rank into the top-k
+                    sc = ggml_add_inplace(ctx0, sc,
+                            ggml_reshape_4d(ctx0, msa->pos_mask, n_ps, 1, 1, ns));
                     ggml_tensor * bs = ggml_pool_2d(ctx0, sc, GGML_OP_POOL_MAX, blk, 1, blk, 1, 0, 0);
                     cb(bs, "msa_bs", il);
 
                     ggml_tensor * bsf = ggml_add(ctx0, bs,
-                            ggml_reshape_4d(ctx0, msa_loc->bias, nblk, 1, 1, ns));
-                    ggml_tensor * idx = ggml_top_k(ctx0, bsf, K);
+                            ggml_reshape_4d(ctx0, msa->bias, nblk, 1, 1, ns));
+                    ggml_tensor * idx = ggml_top_k(ctx0, bsf, K);   // position blocks
 
-                    // token idx: tj[t,k,h,s] = blk*idx[k,h,s] + t   (for the mask gather)
-                    // row   idx: tr[t,k,h,s] = tj*HKV + h           (for the per-stream K/V gather)
+                    // pos idx:  tj[t,k,h,s] = blk*idx[k,h,s] + t   (positions - mask gather)
+                    // cell idx: cs[t,k,h,s] = pos_slot[tj]         (pos -> cell translation)
+                    // row idx:  tr[t,k,h,s] = cs*HKV + h           (per-stream K/V gather)
                     ggml_tensor * a = ggml_scale(ctx0, ggml_cast(ctx0, idx, GGML_TYPE_F32), (float) blk);
                     a = ggml_reshape_4d(ctx0, a, 1, K, Hd, ns);
                     ggml_tensor * tj = ggml_add(ctx0,
                             ggml_repeat_4d(ctx0, a, blk, K, Hd, ns),
                             ggml_reshape_3d(ctx0, ggml_arange(ctx0, 0.0f, (float) blk, 1.0f), blk, 1, 1));
-                    ggml_tensor * tr = ggml_add(ctx0,
-                            ggml_scale(ctx0, tj, (float) HKV),
-                            ggml_reshape_3d(ctx0, ggml_arange(ctx0, 0.0f, (float) HKV, 1.0f), 1, 1, Hd));
 
                     ggml_tensor * tokj = ggml_cast(ctx0, ggml_reshape_2d(ctx0, tj, (int64_t) blk*K*Hd, ns), GGML_TYPE_I32);
+
+                    ggml_tensor * cs = ggml_get_rows(ctx0,
+                            ggml_reshape_3d(ctx0, msa->pos_slot_f, 1, n_ps, ns), tokj);   // [1, blk*K*Hd, ns]
+                    cs = ggml_reshape_4d(ctx0, cs, blk, K, Hd, ns);
+
+                    ggml_tensor * tr = ggml_add(ctx0,
+                            ggml_scale(ctx0, cs, (float) HKV),
+                            ggml_reshape_3d(ctx0, ggml_arange(ctx0, 0.0f, (float) HKV, 1.0f), 1, 1, Hd));
+
                     ggml_tensor * tokr = ggml_cast(ctx0, ggml_reshape_2d(ctx0, tr, (int64_t) blk*K*Hd, ns), GGML_TYPE_I32);
 
                     ggml_tensor * k3 = ggml_view_3d(ctx0, k, D, HKV*n_kv, ns, k->nb[1], k->nb[3], 0);
                     ggml_tensor * v3 = ggml_view_3d(ctx0, v, D, HKV*n_kv, ns, v->nb[1], v->nb[3], 0);
-                    ggml_tensor * m3 = ggml_reshape_3d(ctx0, msa_kqm, 1, n_kv, ns);
+                    ggml_tensor * mp = ggml_reshape_3d(ctx0, msa->pos_mask, 1, n_ps, ns);
 
                     ggml_tensor * kg = ggml_get_rows(ctx0, k3, tokr);
                     ggml_tensor * vg = ggml_get_rows(ctx0, v3, tokr);
-                    ggml_tensor * mg = ggml_get_rows(ctx0, m3, tokj);
+                    ggml_tensor * mg = ggml_get_rows(ctx0, mp, tokj);
 
                     // fold (group, stream) onto the FA channel dim
                     const ggml_type kt = ggml_is_quantized(k->type) ? GGML_TYPE_F16 : k->type;
@@ -372,12 +446,16 @@ llama_model_minimax_m3::graph::graph(const llama_model & model, const llm_graph_
                                 iq->nb[1], iq->nb[2], st*n_tps*iq->nb[2]);
                         ggml_tensor * ik_s = ggml_view_2d(ctx0, ik_kv, n_idx_dim, n_kv,
                                 ik_kv->nb[2], st*ik_kv->nb[3]);
-                        ggml_tensor * mf_s = ggml_view_3d(ctx0, msa_mf, n_kv, 1, n_tps,
-                                msa_mf->nb[1], msa_mf->nb[1], st*msa_mf->nb[3]);
-                        ggml_tensor * km_s = ggml_view_3d(ctx0, msa_kqm, n_kv, n_tps, 1,
-                                msa_kqm->nb[1], msa_kqm->nb[3], st*msa_kqm->nb[3]);
-                        ggml_tensor * bias_s = ggml_view_3d(ctx0, msa_loc->bias, nblk, 1, n_tps,
-                                msa_loc->bias->nb[1], msa_loc->bias->nb[1], st*n_tps*msa_loc->bias->nb[1]);
+                        ggml_tensor * psl_s = ggml_view_1d(ctx0, msa->pos_slot_i, n_ps,
+                                st*msa->pos_slot_i->nb[1]);
+                        ggml_tensor * pm_s = ggml_view_3d(ctx0, msa->pos_mask, n_ps, 1, n_tps,
+                                msa->pos_mask->nb[1], msa->pos_mask->nb[1], st*n_tps*msa->pos_mask->nb[1]);
+                        ggml_tensor * cb_s = ggml_view_1d(ctx0, msa->cell_blk, n_kv,
+                                st*msa->cell_blk->nb[1]);
+                        ggml_tensor * mf_s = ggml_view_3d(ctx0, msa_mf, n_kv, n_tps, 1,
+                                msa_mf->nb[1], msa_mf->nb[3], st*msa_mf->nb[3]);
+                        ggml_tensor * bias_s = ggml_view_3d(ctx0, msa->bias, nblk, 1, n_tps,
+                                msa->bias->nb[1], msa->bias->nb[1], st*n_tps*msa->bias->nb[1]);
                         ggml_tensor * q_s = ggml_view_3d(ctx0, Qcur, D, n_head, n_tps,
                                 Qcur->nb[1], Qcur->nb[2], st*n_tps*Qcur->nb[2]);
                         ggml_tensor * k_s = ggml_view_4d(ctx0, k, D, HKV, n_kv, 1,
@@ -385,14 +463,16 @@ llama_model_minimax_m3::graph::graph(const llama_model & model, const llm_graph_
                         ggml_tensor * v_s = ggml_view_4d(ctx0, v, D, HKV, n_kv, 1,
                                 v->nb[1], v->nb[2], v->nb[3], st*v->nb[3]);
 
-                        // block scores: bs = maxpool_blk(idx_q * idx_k^T + causal mask)
+                        // block scores: the indexer keys are gathered through the pos -> cell map first
                         // scores are unscaled, only the top-k ordering matters
-                        ggml_tensor * sc = ggml_mul_mat(ctx0, ik_s,
+                        ggml_tensor * ikp = ggml_get_rows(ctx0, ik_s, psl_s);   // [n_idx_dim, n_ps]
+                        ggml_tensor * sc = ggml_mul_mat(ctx0, ikp,
                                 ggml_reshape_2d(ctx0, iq_s, n_idx_dim, Hd*n_tps));
                         // indexer scores run in F32
                         ggml_mul_mat_set_prec(sc, GGML_PREC_F32);
-                        sc = ggml_reshape_3d(ctx0, sc, n_kv, Hd, n_tps);
-                        sc = ggml_add_inplace(ctx0, sc, mf_s);
+                        sc = ggml_reshape_3d(ctx0, sc, n_ps, Hd, n_tps);
+                        // unmapped positions (holes, padding, empty cells) come out -inf
+                        sc = ggml_add_inplace(ctx0, sc, pm_s);
                         ggml_tensor * bs = ggml_pool_2d(ctx0, sc, GGML_OP_POOL_MAX, blk, 1, blk, 1, 0, 0);
                         cb(bs, "msa_bs", il);
 
@@ -416,14 +496,16 @@ llama_model_minimax_m3::graph::graph(const llama_model & model, const llm_graph_
                         bm = ggml_cont(ctx0, ggml_permute(ctx0, bm, 0, 2, 1, 3)); // [nblk, n_tps, Hd]
                         cb(bm, "msa_block_mask", il);
 
-                        // expand block -> token granularity (j = bk*blk + t),
-                        // then combine with the causal mask in place
-                        ggml_tensor * bmx = ggml_repeat_4d(ctx0,
-                                ggml_reshape_3d(ctx0, bm, 1, nblk, n_tps*Hd),
-                                blk, nblk, n_tps*Hd, 1);
+                        // expand block -> cell granularity through the cell -> position block
+                        // map, then combine with the causal mask. empty cells are masked by the causal mask.
+                        ggml_tensor * bm2 = ggml_cont(ctx0, ggml_transpose(ctx0,
+                                ggml_reshape_2d(ctx0, bm, nblk, n_tps*Hd)));       // [n_tps*Hd, nblk]
+                        ggml_tensor * bmc = ggml_get_rows(ctx0, bm2, cb_s);        // [n_tps*Hd, n_kv] F32
+                        ggml_tensor * bmx = ggml_cont(ctx0, ggml_transpose(ctx0, bmc));
                         bmx = ggml_reshape_3d(ctx0, bmx, n_kv, n_tps, Hd);
-                        ggml_tensor * mask4 = ggml_add_inplace(ctx0, bmx, km_s);
-                        mask4 = ggml_reshape_4d(ctx0, mask4, n_kv, n_tps, 1, Hd);
+                        ggml_tensor * mask4 = ggml_add_inplace(ctx0, bmx, mf_s);
+                        mask4 = ggml_cast(ctx0,
+                                ggml_reshape_4d(ctx0, mask4, n_kv, n_tps, 1, Hd), GGML_TYPE_F16);
                         cb(mask4, "msa_mask4", il);
 
                         // cache views with groups on ne[3];

src/models/models.h

@@ -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;
@@ -1084,6 +1089,10 @@ struct llama_model_deepseek2 : public llama_model_base {
         graph(const llama_model & model, const llm_graph_params & params);
     };
 
+    struct graph_mtp : public llm_graph_context {
+        graph_mtp(const llama_model & model, const llm_graph_params & params);
+    };
+
     std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
 };
 
@@ -1097,6 +1106,10 @@ struct llama_model_deepseek32 : public llama_model_base {
         graph(const llama_model & model, const llm_graph_params & params);
     };
 
+    struct graph_mtp : public llm_graph_context {
+        graph_mtp(const llama_model & model, const llm_graph_params & params);
+    };
+
     std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
 };
 
@@ -2033,6 +2046,10 @@ struct llama_model_qwen3next : public llama_model_base {
         const llama_model & model;
     };
 
+    struct graph_mtp : public llm_graph_context {
+        graph_mtp(const llama_model & model, const llm_graph_params & params);
+    };
+
     std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
 };
 

src/models/qwen3next.cpp

@@ -13,7 +13,11 @@ void llama_model_qwen3next::load_arch_hparams(llama_model_loader & ml) {
     ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
     ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
 
-    // Mark recurrent layers (linear attention layers)
+    // NextN/MTP: extra decoder block appended beyond the main stack
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.n_layer_nextn, false);
+    GGML_ASSERT(hparams.n_layer_nextn < hparams.n_layer_all && "n_layer_nextn must be < n_layer_all");
+
+    // Mark recurrent layers (linear attention layers).
     if (!ml.get_key_or_arr(LLM_KV_ATTENTION_RECURRENT_LAYERS, hparams.is_recr_impl, hparams.n_layer_all, false)) {
         uint32_t full_attn_interval = 4;
         ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false);
@@ -28,13 +32,17 @@ void llama_model_qwen3next::load_arch_hparams(llama_model_loader & ml) {
     }
 }
 
-void llama_model_qwen3next::load_arch_tensors(llama_model_loader &) {
+void llama_model_qwen3next::load_arch_tensors(llama_model_loader & ml) {
     LLAMA_LOAD_LOCALS;
 
     if (n_expert == 0) {
         throw std::runtime_error(arch_name() + " model cannot have zero experts");
     }
 
+    const bool mtp_only = (hparams.n_layer_nextn > 0) && (ml.get_weight("blk.0.attn_norm.weight") == nullptr);
+    const int trunk_flags = mtp_only ? TENSOR_NOT_REQUIRED : 0;
+    int mtp_flags = !ml.load_mtp ? TENSOR_SKIP : 0;
+
     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
 
     // output
@@ -61,49 +69,73 @@ void llama_model_qwen3next::load_arch_tensors(llama_model_loader &) {
     const int64_t qkvz_dim = key_dim * 2 + value_dim * 2;
     const int64_t ba_dim   = n_v_heads * 2;
 
-    for (int i = 0; i < n_layer; ++i) {
-        auto & layer = layers[i];
-        const uint32_t n_ff_shexp = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : hparams.n_ff(i);
+    auto load_block_trunk = [&](int il, int flags) {
+        auto & layer = layers[il];
+        const uint32_t n_ff_shexp = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : hparams.n_ff(il);
 
-        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), { n_embd }, 0);
-        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0);
+        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", il), { n_embd }, flags);
+        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", il), { n_embd }, flags);
 
-        if (!hparams.is_recr(i)) {
+        if (!hparams.is_recr(il)) {
             // Attention layers
-            create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, 0);
-            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
-
+            create_tensor_qkv(layer, il, n_embd, n_embd_head_k * n_head * 2, n_embd_k_gqa, n_embd_v_gqa, flags);
+            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", il), { n_embd_head_k * n_head, n_embd }, flags);
             // Q/K normalization for attention layers
-            layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0);
-            layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0);
+            layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", il), { n_embd_head_k }, flags);
+            layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", il), { n_embd_head_k }, flags);
         } else {
             // Linear attention (gated delta net) specific tensors
             // Create tensors with calculated dimensions
             // note: ssm_in is used by legacy GGUF
-            layer.ssm_in         = create_tensor(tn(LLM_TENSOR_SSM_IN,         "weight", i), { n_embd, qkvz_dim }, TENSOR_NOT_REQUIRED);
-            layer.wqkv           = create_tensor(tn(LLM_TENSOR_ATTN_QKV,       "weight", i), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED);
-            layer.wqkv_gate      = create_tensor(tn(LLM_TENSOR_ATTN_GATE,      "weight", i), { n_embd, value_dim }, TENSOR_NOT_REQUIRED);
-            layer.ssm_conv1d     = create_tensor(tn(LLM_TENSOR_SSM_CONV1D,     "weight", i), { hparams.ssm_d_conv, conv_dim }, 0);
-            layer.ssm_dt         = create_tensor(tn(LLM_TENSOR_SSM_DT,         "bias",   i), { hparams.ssm_dt_rank }, 0);
-            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN,             i), { hparams.ssm_dt_rank }, 0);
-            layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", i), { n_embd, ba_dim }, 0);
-            layer.ssm_norm       = create_tensor(tn(LLM_TENSOR_SSM_NORM,       "weight", i), { head_v_dim }, 0);
-            layer.ssm_out        = create_tensor(tn(LLM_TENSOR_SSM_OUT,        "weight", i), { value_dim, n_embd }, 0);
+            layer.ssm_in         = create_tensor(tn(LLM_TENSOR_SSM_IN,         "weight", il), { n_embd, qkvz_dim }, TENSOR_NOT_REQUIRED | flags);
+            layer.wqkv           = create_tensor(tn(LLM_TENSOR_ATTN_QKV,       "weight", il), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED | flags);
+            layer.wqkv_gate      = create_tensor(tn(LLM_TENSOR_ATTN_GATE,      "weight", il), { n_embd, value_dim }, TENSOR_NOT_REQUIRED | flags);
+            layer.ssm_conv1d     = create_tensor(tn(LLM_TENSOR_SSM_CONV1D,     "weight", il), { hparams.ssm_d_conv, conv_dim }, flags);
+            layer.ssm_dt         = create_tensor(tn(LLM_TENSOR_SSM_DT,         "bias",   il), { hparams.ssm_dt_rank }, flags);
+            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN,             il), { hparams.ssm_dt_rank }, flags);
+            layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", il), { n_embd, ba_dim }, flags);
+            layer.ssm_norm       = create_tensor(tn(LLM_TENSOR_SSM_NORM,       "weight", il), { head_v_dim }, flags);
+            layer.ssm_out        = create_tensor(tn(LLM_TENSOR_SSM_OUT,        "weight", il), { value_dim, n_embd }, flags);
         }
 
-        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert }, 0);
-        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0);
-        create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0);
+        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", il), { n_embd, n_expert }, flags);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", il), { n_ff_exp, n_embd, n_expert }, flags);
+        create_tensor_gate_up_exps(layer, il, n_embd, n_ff_exp, n_expert, flags);
 
         // Shared experts
-        layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), { n_embd }, 0);
-        layer.ffn_gate_shexp     = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP,     "weight", i), { n_embd, n_ff_shexp }, 0);
-        layer.ffn_up_shexp       = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,       "weight", i), { n_embd, n_ff_shexp }, 0);
-        layer.ffn_down_shexp     = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP,     "weight", i), { n_ff_shexp, n_embd }, 0);
+        layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", il), { n_embd }, flags);
+        layer.ffn_gate_shexp     = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP,     "weight", il), { n_embd, n_ff_shexp }, flags);
+        layer.ffn_up_shexp       = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,       "weight", il), { n_embd, n_ff_shexp }, flags);
+        layer.ffn_down_shexp     = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP,     "weight", il), { n_ff_shexp, n_embd }, flags);
+    };
+
+    auto load_block_mtp = [&](int il) {
+        // MTP head is identical to the trunk block (full attention + FFN)
+        load_block_trunk(il, mtp_flags);
+
+        auto & layer = layers[il];
+
+        // NextN-specific tensors that define the MTP block.
+        layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ,          "weight", il), { 2 * n_embd, n_embd }, mtp_flags);
+        layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,            "weight", il), { n_embd },             mtp_flags);
+        layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,            "weight", il), { n_embd },             mtp_flags);
+        layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", il), { n_embd, n_vocab },    mtp_flags | TENSOR_NOT_REQUIRED);
+        layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", il), { n_embd, n_vocab },    mtp_flags | TENSOR_NOT_REQUIRED);
+        layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", il), { n_embd },             mtp_flags | TENSOR_NOT_REQUIRED);
+    };
+
+    for (int i = 0; i < n_layer; i++) {
+        load_block_trunk(i, trunk_flags);
+    }
+    for (int i = n_layer; i < n_layer_all; i++) {
+        load_block_mtp(i);
     }
 }
 
 std::unique_ptr<llm_graph_context> llama_model_qwen3next::build_arch_graph(const llm_graph_params & params) const {
+    if (params.gtype == LLM_GRAPH_TYPE_DECODER_MTP) {
+        return std::make_unique<graph_mtp>(*this, params);
+    }
     return std::make_unique<graph>(*this, params);
 }
 
@@ -120,6 +152,7 @@ llama_model_qwen3next::graph::graph(const llama_model & model, const llm_graph_p
     ggml_tensor * inp_pos     = build_inp_pos();
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
+    // MTP/NextN layers are loaded as extra decoder blocks but not executed in the main pass.
     for (int il = 0; il < n_layer; ++il) {
         res->t_layer_inp[il] = inpL;
 
@@ -139,7 +172,7 @@ llama_model_qwen3next::graph::graph(const llama_model & model, const llm_graph_p
             cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);
         }
 
-        if (il == n_layer - 1 && inp_out_ids) {
+        if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
             cur   = ggml_get_rows(ctx0, cur, inp_out_ids);
             inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
         }
@@ -171,9 +204,16 @@ llama_model_qwen3next::graph::graph(const llama_model & model, const llm_graph_p
     }
     cur = inpL;
 
-    // Final norm
+    // post-norm hidden state is input to both the LM head and the MTP head
     cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1);
 
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    if (!cparams.embeddings_nextn_masked && inp_out_ids) {
+        cur = ggml_get_rows(ctx0, cur, inp_out_ids);
+    }
+
     cb(cur, "result_norm", -1);
     res->t_embd = cur;
 
@@ -186,14 +226,6 @@ llama_model_qwen3next::graph::graph(const llama_model & model, const llm_graph_p
     ggml_build_forward_expand(gf, cur);
 }
 
-// utility to get one slice from the third dimension
-// input dim:  [x, y, c, b]
-// output dim: [x, y, 1, b]
-static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) {
-    return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3],
-        t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c);
-}
-
 ggml_tensor * llama_model_qwen3next::graph::build_norm_gated(
         ggml_tensor * input,
         ggml_tensor * weights,
@@ -216,7 +248,7 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_attn(
     // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention
 
     // Qwen3Next uses a single Q projection that outputs query + gate
-    ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur);
+    ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
     cb(Qcur_full, "Qcur_full", il);
 
     Qcur_full = ggml_reshape_4d(ctx0, Qcur_full, n_embd_head * 2, n_head, n_tokens, 1);
@@ -232,10 +264,10 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_attn(
                      Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full));
     cb(gate, "gate", il);
 
-    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
+    ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
     cb(Kcur, "Kcur", il);
 
-    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
+    ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
     cb(Vcur, "Vcur", il);
 
     Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
@@ -274,8 +306,6 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_attn(
     gate = ggml_sigmoid(ctx0, gate);
     cb(gate, "gate_sigmoid", il);
 
-    gate = ggml_reshape_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
-
     cur = ggml_mul(ctx0, cur, gate);
     cb(cur, "attn_gated", il);
 
@@ -550,16 +580,19 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_ffn(ggml_tensor * cur, c
                 LLM_FFN_SILU, true,
                 hparams.expert_weights_scale,
                 LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,
-                nullptr, model.layers[il].ffn_gate_up_exps);
+                nullptr, model.layers[il].ffn_gate_up_exps,
+                model.layers[il].ffn_up_exps_s,
+                model.layers[il].ffn_gate_exps_s,
+                model.layers[il].ffn_down_exps_s);
         cb(moe_out, "ffn_moe_out", il);
 
         // Add shared experts if present - following Qwen3Next reference implementation
         if (model.layers[il].ffn_up_shexp != nullptr) {
             ggml_tensor * ffn_shexp =
                 build_ffn(cur,
-                    model.layers[il].ffn_up_shexp,   NULL, NULL,
-                    model.layers[il].ffn_gate_shexp, NULL, NULL,
-                    model.layers[il].ffn_down_shexp, NULL, NULL,
+                    model.layers[il].ffn_up_shexp,   NULL, model.layers[il].ffn_up_shexp_s,
+                    model.layers[il].ffn_gate_shexp, NULL, model.layers[il].ffn_gate_shexp_s,
+                    model.layers[il].ffn_down_shexp, NULL, model.layers[il].ffn_down_shexp_s,
                     NULL,
                     LLM_FFN_SILU, LLM_FFN_PAR, il);
             cb(ffn_shexp, "ffn_shexp", il);
@@ -593,3 +626,198 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_ffn(ggml_tensor * cur, c
     }
     return cur;
 }
+
+// LLM_GRAPH_TYPE_DECODER_MTP draft head for Qwen3-Next
+llama_model_qwen3next::graph_mtp::graph_mtp(const llama_model & model, const llm_graph_params & params)
+    : llm_graph_context(params) {
+    GGML_ASSERT(hparams.n_layer_nextn > 0 && "QWEN3NEXT MTP requires n_layer_nextn > 0");
+    GGML_ASSERT(hparams.n_layer_nextn == 1 && "QWEN3NEXT MTP currently only supports a single MTP block");
+
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+
+    const int il = hparams.n_layer();
+    const auto & layer = model.layers[il];
+
+    GGML_ASSERT(layer.nextn.eh_proj    && "MTP block missing nextn.eh_proj");
+    GGML_ASSERT(layer.nextn.enorm      && "MTP block missing nextn.enorm");
+    GGML_ASSERT(layer.nextn.hnorm      && "MTP block missing nextn.hnorm");
+    GGML_ASSERT(layer.ffn_gate_inp     && "MTP block missing ffn_gate_inp");
+
+    // TODO: extract in a common llm_graph_context::build_inp_embd_h()
+    auto inp = std::make_unique<llm_graph_input_embd_h>(hparams.n_embd);
+
+    inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+    ggml_set_input(inp->tokens);
+
+    inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd_inp(), n_tokens);
+    ggml_set_input(inp->embd);
+
+    // TODO: make static using `ggml_build_forward_select()`
+    //       see llm_graph_context::build_inp_embd() for reference
+    ggml_tensor * tok_embd;
+    if (ubatch.token) {
+        ggml_tensor * tok_embd_w = layer.nextn.embed_tokens ? layer.nextn.embed_tokens : model.tok_embd;
+
+        tok_embd = ggml_get_rows(ctx0, tok_embd_w, inp->tokens);
+    } else {
+        tok_embd = inp->embd;
+    }
+    cb(tok_embd, "mtp_tok_embd", il);
+
+    inp->h = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, hparams.n_embd, n_tokens);
+    ggml_set_input(inp->h);
+    ggml_set_name(inp->h, "mtp_h_input");
+
+    ggml_tensor * h_embd = inp->h;
+
+    res->add_input(std::move(inp));
+
+    ggml_tensor * inp_pos     = build_inp_pos();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * h_norm = build_norm(h_embd, layer.nextn.hnorm, nullptr, LLM_NORM_RMS, il);
+    cb(h_norm, "mtp_hnorm", il);
+
+    ggml_tensor * e_norm = build_norm(tok_embd, layer.nextn.enorm, nullptr, LLM_NORM_RMS, il);
+    cb(e_norm, "mtp_enorm", il);
+
+    ggml_tensor * concat = ggml_concat(ctx0, e_norm, h_norm, /*dim=*/ 0);
+    cb(concat, "mtp_concat", il);
+
+    ggml_tensor * cur = build_lora_mm(layer.nextn.eh_proj, concat, layer.nextn.eh_proj_s);
+    cb(cur, "mtp_eh_proj", il);
+
+    ggml_tensor * inpSA = cur;
+
+    cur = build_norm(cur, layer.attn_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_attn_norm", il);
+
+    ggml_tensor * Qcur_full = build_lora_mm(layer.wq, cur, layer.wq_s);
+    cb(Qcur_full, "mtp_Qcur_full", il);
+
+    ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full,
+            n_embd_head, n_head, n_tokens,
+            ggml_element_size(Qcur_full) * n_embd_head * 2,
+            ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head,
+            0);
+    Qcur = build_norm(Qcur, layer.attn_q_norm, nullptr, LLM_NORM_RMS, il);
+    cb(Qcur, "mtp_Qcur_normed", il);
+
+    ggml_tensor * Kcur = build_lora_mm(layer.wk, cur, layer.wk_s);
+    Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+    Kcur = build_norm(Kcur, layer.attn_k_norm, nullptr, LLM_NORM_RMS, il);
+    cb(Kcur, "mtp_Kcur_normed", il);
+
+    ggml_tensor * Vcur = build_lora_mm(layer.wv, cur, layer.wv_s);
+    Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+    Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr,
+            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+    Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr,
+            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+            ext_factor, attn_factor, beta_fast, beta_slow);
+
+    cb(Qcur, "mtp_Qcur", il);
+    cb(Kcur, "mtp_Kcur", il);
+    cb(Vcur, "mtp_Vcur", il);
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f
+            ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    cur = build_attn(inp_attn,
+            nullptr, nullptr, nullptr,
+            Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+    cb(cur, "mtp_attn_pregate", il);
+
+    ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full,
+            n_embd_head, n_head, n_tokens,
+            ggml_element_size(Qcur_full) * n_embd_head * 2,
+            ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head,
+            ggml_element_size(Qcur_full) * n_embd_head);
+
+    // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont
+    gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens);
+    cb(gate, "mtp_gate", il);
+
+    cur = ggml_mul(ctx0, cur, ggml_sigmoid(ctx0, gate));
+    cur = build_lora_mm(layer.wo, cur, layer.wo_s);
+    cb(cur, "mtp_attn_out", il);
+
+    if (inp_out_ids) {
+        cur   = ggml_get_rows(ctx0, cur,   inp_out_ids);
+        inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+    }
+
+    cur = ggml_add(ctx0, cur, inpSA);
+    cb(cur, "mtp_attn_residual", il);
+
+    ggml_tensor * ffn_residual = cur;
+    cur = build_norm(cur, layer.attn_post_norm, nullptr, LLM_NORM_RMS, il);
+    cb(cur, "mtp_attn_post_norm", il);
+
+    // MoE FFN — routed experts plus gated shared expert (mirrors the trunk).
+    ggml_tensor * moe_out =
+        build_moe_ffn(cur,
+            layer.ffn_gate_inp,
+            layer.ffn_up_exps,
+            layer.ffn_gate_exps,
+            layer.ffn_down_exps,
+            nullptr,
+            n_expert, n_expert_used,
+            LLM_FFN_SILU, true,
+            hparams.expert_weights_scale,
+            LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il,
+            nullptr, layer.ffn_gate_up_exps,
+            layer.ffn_up_exps_s,
+            layer.ffn_gate_exps_s,
+            layer.ffn_down_exps_s);
+    cb(moe_out, "mtp_ffn_moe_out", il);
+
+    if (layer.ffn_up_shexp != nullptr) {
+        ggml_tensor * ffn_shexp =
+            build_ffn(cur,
+                layer.ffn_up_shexp,   nullptr, layer.ffn_up_shexp_s,
+                layer.ffn_gate_shexp, nullptr, layer.ffn_gate_shexp_s,
+                layer.ffn_down_shexp, nullptr, layer.ffn_down_shexp_s,
+                nullptr,
+                LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(ffn_shexp, "mtp_ffn_shexp", il);
+
+        ggml_tensor * shared_gate = build_lora_mm(layer.ffn_gate_inp_shexp, cur);
+        shared_gate = ggml_sigmoid(ctx0, shared_gate);
+        cb(shared_gate, "mtp_shared_expert_gate_sigmoid", il);
+
+        ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate);
+        cb(ffn_shexp, "mtp_ffn_shexp_gated", il);
+
+        cur = ggml_add(ctx0, moe_out, ffn_shexp);
+    } else {
+        cur = moe_out;
+    }
+    cb(cur, "mtp_ffn_out", il);
+
+    cur = ggml_add(ctx0, cur, ffn_residual);
+    cb(cur, "mtp_post_ffn", il);
+
+    ggml_tensor * head_norm_w = layer.nextn.shared_head_norm
+            ? layer.nextn.shared_head_norm
+            : model.output_norm;
+    GGML_ASSERT(head_norm_w && "QWEN3NEXT MTP: missing both nextn.shared_head_norm and output_norm");
+    cur = build_norm(cur, head_norm_w, nullptr, LLM_NORM_RMS, -1);
+
+    cb(cur, "h_nextn", -1);
+    res->t_h_nextn = cur;
+
+    ggml_tensor * head_w = layer.nextn.shared_head_head ? layer.nextn.shared_head_head : model.output;
+    ggml_tensor * head_s = layer.nextn.shared_head_head ? layer.nextn.shared_head_head_s : model.output_s;
+    GGML_ASSERT(head_w && "QWEN3NEXT MTP: missing LM head (nextn.shared_head_head or model.output)");
+    cur = build_lora_mm(head_w, cur, head_s);
+    cb(cur, "result_output", -1);
+
+    res->t_logits = cur;
+    ggml_build_forward_expand(gf, cur);
+}

src/models/qwen3tts.cpp

@@ -0,0 +1,3 @@
+#include "models.h"
+
+// llama_model_qwen3tts reuses llama_model_qwen3vl's hparams/tensors/graph logic

src/models/qwen3vl.cpp

@@ -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;
 

tests/CMakeLists.txt

@@ -258,6 +258,9 @@ llama_build_and_test(test-thread-safety.cpp ARGS -m "${MODEL_DEST}" -ngl 99 -p "
 set_tests_properties(test-thread-safety PROPERTIES FIXTURES_REQUIRED test-download-model)
 
 llama_build_and_test(test-arg-parser.cpp)
+llama_build_and_test(test-model-resolution.cpp)
+# the test serves its repos from an httplib server, and the library links it privately
+target_link_libraries(test-model-resolution PRIVATE cpp-httplib)
 
 if (NOT LLAMA_SANITIZE_ADDRESS AND NOT GGML_SCHED_NO_REALLOC)
   # TODO: repair known memory leaks

tests/test-arg-parser.cpp

@@ -99,6 +99,42 @@ static void test(void) {
     argv = {"binary_name", "-sm", "hello"};
     assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_COMMON));
 
+    {
+        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));
+
+        argv = {"binary_name", "--repeat-penalty", "-1"};
+        assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), penalty_params, LLAMA_EXAMPLE_COMMON));
+
+        argv = {"binary_name", "--repeat-penalty", "nan"};
+        assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), penalty_params, LLAMA_EXAMPLE_COMMON));
+
+        argv = {"binary_name", "--repeat-penalty", "inf"};
+        assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), penalty_params, LLAMA_EXAMPLE_COMMON));
+
+        argv = {"binary_name", "--repeat-penalty", "-inf"};
+        assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), penalty_params, LLAMA_EXAMPLE_COMMON));
+
+        const char * penalty_options[] = {"--frequency-penalty", "--presence-penalty"};
+        const char * nonfinite_values[] = {"nan", "inf", "-inf"};
+        for (const char * option : penalty_options) {
+            for (const char * value : nonfinite_values) {
+                argv = {"binary_name", option, value};
+                assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), penalty_params, LLAMA_EXAMPLE_COMMON));
+            }
+        }
+    }
+
     // non-existence arg in specific example (--draft cannot be used outside llama-speculative)
     argv = {"binary_name", "--draft", "123"};
     assert(false == common_params_parse(argv.size(), list_str_to_char(argv).data(), params, LLAMA_EXAMPLE_EMBEDDING));

tests/test-backend-ops.cpp

@@ -8069,6 +8069,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_dsv4_hc_comb(1, 1));
     test_cases.emplace_back(new test_dsv4_hc_comb(17, 4));
     test_cases.emplace_back(new test_dsv4_hc_comb(257, 8));
+    test_cases.emplace_back(new test_dsv4_hc_comb(17, 20));
 
     test_cases.emplace_back(new test_dsv4_hc_pre(1, 1));
     test_cases.emplace_back(new test_dsv4_hc_pre(31, 17));
@@ -8078,6 +8079,7 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_dsv4_hc_post(1, 1));
     test_cases.emplace_back(new test_dsv4_hc_post(31, 17));
     test_cases.emplace_back(new test_dsv4_hc_post(128, 257));
+    test_cases.emplace_back(new test_dsv4_hc_post(4096, 21));
 
     // glu ops
     for (ggml_type type : {GGML_TYPE_F16, GGML_TYPE_F32}) {
@@ -9729,6 +9731,12 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
         }
     }
 
+    for (int kv : { 1, 7, 8, 63, 64, 65 }) {
+        for (ggml_type type_K : {GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0}) {
+            test_cases.emplace_back(new test_lightning_indexer(128, 64, kv, 32, 4, 1, type_K));
+        }
+    }
+
     return test_cases;
 }
 #ifdef _MSC_VER

tests/test-backend-sampler.cpp

@@ -8,12 +8,15 @@
 #endif
 
 #include <algorithm>
+#include <cmath>
 #include <cstdlib>
 #include <cstring>
 #include <fstream>
+#include <functional>
 #include <map>
 #include <string>
 #include <unordered_map>
+#include <unordered_set>
 #include <vector>
 
 struct test_args {
@@ -761,6 +764,564 @@ static void test_backend_logit_bias_sampling(const test_params & params) {
     printf("backend logit bias sampling test PASSED\n");
 }
 
+static void accept_prompt(llama_sampler * smpl, const llama_vocab * vocab, const std::string & prompt) {
+    const llama_token bos = llama_vocab_bos(vocab);
+    if (bos != LLAMA_TOKEN_NULL) {
+        llama_sampler_accept(smpl, bos);
+    }
+
+    std::vector<llama_token> tokens(64);
+    int32_t n_tokens = llama_tokenize(vocab, prompt.c_str(), (int32_t) prompt.size(),
+            tokens.data(), (int32_t) tokens.size(), false, false);
+    if (n_tokens < 0) {
+        tokens.resize(-n_tokens);
+        n_tokens = llama_tokenize(vocab, prompt.c_str(), (int32_t) prompt.size(),
+                tokens.data(), (int32_t) tokens.size(), false, false);
+    }
+
+    for (int32_t i = 0; i < n_tokens; ++i) {
+        llama_sampler_accept(smpl, tokens[i]);
+    }
+}
+
+static std::vector<float> decode_raw_logits(const test_params & params, const std::string & prompt) {
+    const int seq_id = 0;
+    const int n_vocab = llama_vocab_n_tokens(llama_model_get_vocab(params.model.get()));
+    std::vector<llama_sampler_seq_config> empty_configs;
+    test_context ctx(params, empty_configs);
+
+    GGML_ASSERT(ctx.decode({{ seq_id, prompt }}));
+
+    float * logits = llama_get_logits_ith(ctx.ctx.get(), ctx.idx_for_seq(seq_id));
+    GGML_ASSERT(logits != nullptr);
+    return std::vector<float>(logits, logits + n_vocab);
+}
+
+static std::vector<llama_token_data> apply_cpu_sampler(
+        const std::vector<float> & raw_logits,
+        llama_sampler * sampler) {
+    std::vector<llama_token_data> data;
+    data.reserve(raw_logits.size());
+    for (llama_token token = 0; token < (llama_token) raw_logits.size(); ++token) {
+        data.push_back({ token, raw_logits[token], 0.0f });
+    }
+
+    llama_token_data_array cur_p = { data.data(), data.size(), -1, false };
+    llama_sampler_apply(sampler, &cur_p);
+    data.resize(cur_p.size);
+    return data;
+}
+
+using sampler_setup_fn = std::function<void(llama_sampler *)>;
+using sampler_init_fn = std::function<llama_sampler *()>;
+
+enum class penalties_position {
+    before_filter,
+    after_filter,
+};
+
+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,
+        float penalty_present,
+        penalties_position position) {
+    const auto add_penalties = [&]() {
+        llama_sampler_chain_add(chain, llama_sampler_init_penalties(
+                    n_vocab, penalty_last_n, penalty_repeat, penalty_freq, penalty_present));
+    };
+
+    if (position == penalties_position::before_filter) {
+        add_penalties();
+        llama_sampler_chain_add(chain, init_filter());
+    } else {
+        llama_sampler_chain_add(chain, init_filter());
+        add_penalties();
+    }
+}
+
+static llama_sampler_ptr make_sampler_chain(
+        const sampler_setup_fn & add_samplers,
+        const sampler_setup_fn & accept_history) {
+    llama_sampler_ptr chain(llama_sampler_chain_init(llama_sampler_chain_default_params()));
+    add_samplers(chain.get());
+    accept_history(chain.get());
+    return chain;
+}
+
+struct backend_sampler_output {
+    std::vector<float> logits;
+    std::vector<llama_token> candidates;
+};
+
+static backend_sampler_output run_backend_sampler(
+        const test_params & params,
+        const std::string & prompt,
+        llama_sampler * sampler) {
+    const int seq_id = 0;
+    std::vector<llama_sampler_seq_config> configs = {{ seq_id, sampler }};
+    test_context ctx(params, configs);
+
+    GGML_ASSERT(ctx.decode({{ seq_id, prompt }}));
+    llama_synchronize(ctx.ctx.get());
+
+    const int32_t idx = ctx.idx_for_seq(seq_id);
+    const uint32_t n_logits = llama_get_sampled_logits_count_ith(ctx.ctx.get(), idx);
+    const uint32_t n_candidates = llama_get_sampled_candidates_count_ith(ctx.ctx.get(), idx);
+    float * logits = llama_get_sampled_logits_ith(ctx.ctx.get(), idx);
+    llama_token * candidates = llama_get_sampled_candidates_ith(ctx.ctx.get(), idx);
+    GGML_ASSERT(logits != nullptr);
+
+    backend_sampler_output result;
+    result.logits.assign(logits, logits + n_logits);
+    result.candidates.resize(n_logits);
+
+    if (n_candidates == 0) {
+        for (uint32_t i = 0; i < n_logits; ++i) {
+            result.candidates[i] = (llama_token) i;
+        }
+    } else {
+        GGML_ASSERT(candidates != nullptr);
+        GGML_ASSERT(n_candidates == n_logits);
+        std::memcpy(result.candidates.data(), candidates, n_candidates * sizeof(llama_token));
+    }
+
+    return result;
+}
+
+struct sampler_comparison_output {
+    std::vector<llama_token_data> expected;
+    backend_sampler_output actual;
+};
+
+static sampler_comparison_output run_sampler_comparison(
+        const test_params & params,
+        const std::string & prompt,
+        const std::vector<float> & raw_logits,
+        const sampler_setup_fn & add_samplers,
+        const sampler_setup_fn & accept_history) {
+    llama_sampler_ptr cpu_chain = make_sampler_chain(add_samplers, accept_history);
+    llama_sampler_ptr backend_chain = make_sampler_chain(add_samplers, accept_history);
+    return {
+        apply_cpu_sampler(raw_logits, cpu_chain.get()),
+        run_backend_sampler(params, prompt, backend_chain.get()),
+    };
+}
+
+static std::unordered_map<llama_token, float> map_logits(const std::vector<llama_token_data> & data) {
+    std::unordered_map<llama_token, float> result;
+    result.reserve(data.size());
+    for (const auto & item : data) {
+        result[item.id] = item.logit;
+    }
+    return result;
+}
+
+struct sampler_comparison_stats {
+    int n_mismatch = 0;
+    int n_masked = 0;
+    float max_diff = 0.0f;
+};
+
+static sampler_comparison_stats compare_sampler_outputs(
+        const char * name,
+        const std::unordered_map<llama_token, float> & expected,
+        const backend_sampler_output & actual,
+        bool allow_extra_candidates = false) {
+    GGML_ASSERT(actual.logits.size() == actual.candidates.size());
+
+    sampler_comparison_stats result;
+    std::unordered_set<llama_token> seen;
+    seen.reserve(actual.candidates.size());
+
+    for (size_t i = 0; i < actual.logits.size(); ++i) {
+        const llama_token token = actual.candidates[i];
+        const float logit = actual.logits[i];
+        if (!seen.insert(token).second || std::isnan(logit)) {
+            if (result.n_mismatch < 5) {
+                printf("%s token %d has invalid backend output\n", name, token);
+            }
+            ++result.n_mismatch;
+            continue;
+        }
+
+        const auto it = expected.find(token);
+        if (it == expected.end()) {
+            if (std::isinf(logit) && logit < 0.0f) {
+                ++result.n_masked;
+            } else if (!allow_extra_candidates) {
+                if (result.n_mismatch < 5) {
+                    printf("%s token %d was not masked\n", name, token);
+                }
+                ++result.n_mismatch;
+            }
+            continue;
+        }
+
+        const float diff = fabsf(it->second - logit);
+        result.max_diff = std::max(result.max_diff, diff);
+        if (!std::isfinite(logit) || diff > 1e-3f) {
+            if (result.n_mismatch < 5) {
+                printf("%s mismatch token %d: cpu=%.6f backend=%.6f diff=%.6f\n",
+                        name, token, it->second, logit, diff);
+            }
+            ++result.n_mismatch;
+        }
+    }
+
+    for (const auto & item : expected) {
+        if (seen.find(item.first) == seen.end()) {
+            if (result.n_mismatch < 5) {
+                printf("%s missing backend token %d\n", name, item.first);
+            }
+            ++result.n_mismatch;
+        }
+    }
+
+    printf("%s logits: max_diff=%.6f n_masked=%d n_mismatch=%d\n",
+            name, result.max_diff, result.n_masked, result.n_mismatch);
+    return result;
+}
+
+static float find_backend_logit(const backend_sampler_output & output, llama_token token) {
+    for (size_t i = 0; i < output.candidates.size(); ++i) {
+        if (output.candidates[i] == token) {
+            return output.logits[i];
+        }
+    }
+    GGML_ABORT("backend token not found");
+}
+
+static sampler_comparison_output run_penalties_comparison(
+        const test_params & params,
+        int32_t penalty_last_n,
+        float penalty_repeat,
+        float penalty_freq,
+        float penalty_present,
+        const std::string & prompt,
+        const std::function<void(llama_sampler *)> & extra_accept = {}) {
+    const auto * vocab = llama_model_get_vocab(params.model.get());
+    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(
+                    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);
+        if (extra_accept) {
+            extra_accept(chain);
+        }
+    };
+
+    return run_sampler_comparison(
+            params, prompt, raw_logits, add_samplers, accept_history);
+}
+
+static void compare_penalties_logits(
+        const test_params & params,
+        int32_t penalty_last_n,
+        float penalty_repeat,
+        float penalty_freq,
+        float penalty_present,
+        const std::string & prompt,
+        const std::function<void(llama_sampler *)> & extra_accept = {}) {
+    const sampler_comparison_output output = run_penalties_comparison(
+            params, penalty_last_n, penalty_repeat, penalty_freq, penalty_present, prompt, extra_accept);
+
+    GGML_ASSERT(output.expected.size() == output.actual.logits.size());
+
+    const sampler_comparison_stats stats = compare_sampler_outputs(
+            "penalties", map_logits(output.expected), output.actual);
+    GGML_ASSERT(stats.n_masked == 0);
+    GGML_ASSERT(stats.n_mismatch == 0);
+}
+
+static void test_penalty_parameter_values(const test_params & params) {
+    struct penalty_test_case {
+        const char * name;
+        float repeat;
+        float frequency;
+        float presence;
+    };
+
+    const penalty_test_case cases[] = {
+        { "frequency -1",   1.0f, -1.0f,     0.0f },
+        { "frequency 0",    1.0f,  0.0f,     0.0f },
+        { "frequency 1",    1.0f,  1.0f,     0.0f },
+        { "presence -1",    1.0f,  0.0f,    -1.0f },
+        { "presence 0",     1.0f,  0.0f,     0.0f },
+        { "presence 1",     1.0f,  0.0f,     1.0f },
+        { "repeat 1",       1.0f,  0.0f,     0.0f },
+    };
+
+    int n_failed = 0;
+    for (const auto & test : cases) {
+        const sampler_comparison_output output = run_penalties_comparison(
+                params, 64, test.repeat, test.frequency, test.presence, "Hello Hello world");
+        GGML_ASSERT(output.expected.size() == output.actual.logits.size());
+        const sampler_comparison_stats stats = compare_sampler_outputs(
+                test.name, map_logits(output.expected), output.actual);
+        n_failed += stats.n_mismatch != 0;
+    }
+
+    GGML_ASSERT(n_failed == 0);
+}
+
+static void compare_top_k_penalties_logits(
+        const test_params & params,
+        int32_t k,
+        int32_t penalty_last_n,
+        float penalty_repeat,
+        float penalty_freq,
+        float penalty_present,
+        const std::string & prompt,
+        penalties_position position) {
+    const auto * vocab = llama_model_get_vocab(params.model.get());
+    const std::vector<float> raw_logits = decode_raw_logits(params, prompt);
+    const int n_vocab = (int) raw_logits.size();
+
+    GGML_ASSERT(n_vocab > k);
+
+    const sampler_init_fn init_top_k = [k]() {
+        return llama_sampler_init_top_k(k);
+    };
+    llama_sampler_ptr top_k(init_top_k());
+    const std::vector<llama_token_data> top_k_data = apply_cpu_sampler(raw_logits, top_k.get());
+    GGML_ASSERT(top_k_data.size() == (size_t) k);
+    const llama_token retained_history_token = top_k_data[0].id;
+
+    llama_token excluded_history_token = LLAMA_TOKEN_NULL;
+    for (llama_token token = 0; token < n_vocab; ++token) {
+        const auto it = std::find_if(top_k_data.begin(), top_k_data.end(), [token](const llama_token_data & data) {
+            return data.id == token;
+        });
+        if (it == top_k_data.end()) {
+            excluded_history_token = token;
+            break;
+        }
+    }
+    GGML_ASSERT(excluded_history_token != LLAMA_TOKEN_NULL);
+
+    const auto add_samplers = [&](llama_sampler * chain) {
+        add_filter_and_penalties(chain, init_top_k, n_vocab,
+                penalty_last_n, penalty_repeat, penalty_freq, penalty_present, position);
+    };
+
+    auto accept_history = [&](llama_sampler * smpl) {
+        accept_prompt(smpl, vocab, prompt);
+        llama_sampler_accept(smpl, excluded_history_token);
+        llama_sampler_accept(smpl, excluded_history_token);
+        llama_sampler_accept(smpl, retained_history_token);
+        llama_sampler_accept(smpl, retained_history_token);
+    };
+
+    const sampler_comparison_output output = run_sampler_comparison(
+            params, prompt, raw_logits, add_samplers, accept_history);
+
+    GGML_ASSERT(output.expected.size() == (size_t) k);
+    GGML_ASSERT(output.actual.logits.size() == (size_t) k);
+
+    const std::unordered_map<llama_token, float> expected_logits = map_logits(output.expected);
+
+    if (position == penalties_position::after_filter) {
+        GGML_ASSERT(expected_logits.find(retained_history_token) != expected_logits.end());
+        GGML_ASSERT(fabsf(expected_logits.at(retained_history_token) - raw_logits[retained_history_token]) > 1e-6f);
+        GGML_ASSERT(expected_logits.find(excluded_history_token) == expected_logits.end());
+        GGML_ASSERT(std::find(output.actual.candidates.begin(), output.actual.candidates.end(),
+                    excluded_history_token) == output.actual.candidates.end());
+    } else {
+        const std::unordered_map<llama_token, float> unpenalized_logits = map_logits(top_k_data);
+        bool changed = false;
+        for (const auto & item : expected_logits) {
+            const auto it = unpenalized_logits.find(item.first);
+            if (it == unpenalized_logits.end() || fabsf(it->second - item.second) > 1e-6f) {
+                changed = true;
+                break;
+            }
+        }
+        GGML_ASSERT(changed);
+    }
+
+    const char * name = position == penalties_position::before_filter
+        ? "penalties top-k"
+        : "top-k penalties";
+    const sampler_comparison_stats stats = compare_sampler_outputs(
+            name, expected_logits, output.actual);
+    GGML_ASSERT(stats.n_masked == 0);
+    GGML_ASSERT(stats.n_mismatch == 0);
+}
+
+static void compare_masking_penalties_logits(
+        const test_params & params,
+        const char * filter_name,
+        const sampler_init_fn & init_filter,
+        int32_t penalty_last_n,
+        float penalty_repeat,
+        float penalty_freq,
+        float penalty_present,
+        const std::string & prompt,
+        penalties_position position,
+        bool allow_extra_candidates,
+        bool add_history = true) {
+    const auto * vocab = llama_model_get_vocab(params.model.get());
+    const std::vector<float> raw_logits = decode_raw_logits(params, prompt);
+    const int n_vocab = (int) raw_logits.size();
+    llama_sampler_ptr filter(init_filter());
+    const std::vector<llama_token_data> filtered_data = apply_cpu_sampler(raw_logits, filter.get());
+    GGML_ASSERT(!filtered_data.empty());
+    GGML_ASSERT(filtered_data.size() < (size_t) n_vocab);
+
+    const llama_token penalized_token = filtered_data[0].id;
+    std::unordered_set<llama_token> retained_tokens;
+    retained_tokens.reserve(filtered_data.size());
+    for (const auto & data : filtered_data) {
+        retained_tokens.insert(data.id);
+    }
+
+    llama_token masked_token = LLAMA_TOKEN_NULL;
+    for (llama_token token = 0; token < n_vocab; ++token) {
+        if (retained_tokens.find(token) == retained_tokens.end()) {
+            masked_token = token;
+            break;
+        }
+    }
+    GGML_ASSERT(masked_token != LLAMA_TOKEN_NULL);
+
+    const auto add_samplers = [&](llama_sampler * chain) {
+        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) {
+        if (!add_history) {
+            return;
+        }
+        accept_prompt(smpl, vocab, prompt);
+        llama_sampler_accept(smpl, penalized_token);
+        llama_sampler_accept(smpl, penalized_token);
+        llama_sampler_accept(smpl, masked_token);
+        llama_sampler_accept(smpl, masked_token);
+    };
+
+    const sampler_comparison_output output = run_sampler_comparison(
+            params, prompt, raw_logits, add_samplers, accept_history);
+
+    GGML_ASSERT(output.actual.logits.size() == (size_t) n_vocab);
+
+    const std::unordered_map<llama_token, float> expected_logits = map_logits(output.expected);
+
+    GGML_ASSERT(expected_logits.find(masked_token) == expected_logits.end());
+    if (add_history) {
+        if (position == penalties_position::after_filter) {
+            GGML_ASSERT(expected_logits.find(penalized_token) != expected_logits.end());
+            GGML_ASSERT(fabsf(expected_logits.at(penalized_token) - raw_logits[penalized_token]) > 1e-6f);
+        } else {
+            llama_sampler_ptr penalties(llama_sampler_init_penalties(
+                        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()));
+            GGML_ASSERT(fabsf(penalized_logits.at(penalized_token) - raw_logits[penalized_token]) > 1e-6f);
+        }
+    }
+
+    const std::string name = position == penalties_position::before_filter
+        ? "penalties " + std::string(filter_name)
+        : std::string(filter_name) + " penalties";
+    const sampler_comparison_stats stats = compare_sampler_outputs(
+            name.c_str(), expected_logits, output.actual, allow_extra_candidates);
+    const float masked_logit = find_backend_logit(output.actual, masked_token);
+    GGML_ASSERT(stats.n_masked > 0);
+    GGML_ASSERT(std::isinf(masked_logit) && masked_logit < 0.0f);
+    GGML_ASSERT(stats.n_mismatch == 0);
+}
+
+static void test_backend_penalties_sampling(const test_params & params) {
+    printf("Testing backend penalties (repeat + freq + presence)\n");
+    compare_penalties_logits(params, 64, 1.1f, 0.5f, 0.25f, "Hello Hello world");
+
+    printf("Testing backend penalties with penalty_last_n > 64\n");
+    const auto * vocab = llama_model_get_vocab(params.model.get());
+    std::vector<llama_token> tokens(8);
+    int32_t n_tok = llama_tokenize(vocab, "a", 1, tokens.data(), (int32_t) tokens.size(), false, false);
+    if (n_tok < 0) {
+        tokens.resize(-n_tok);
+        n_tok = llama_tokenize(vocab, "a", 1, tokens.data(), (int32_t) tokens.size(), false, false);
+    }
+    GGML_ASSERT(n_tok > 0);
+    const llama_token tok = tokens[0];
+
+    compare_penalties_logits(params, 80, 1.15f, 0.1f, 0.05f, "a", [tok](llama_sampler * smpl) {
+        // accept_prompt already accepted BOS + one 'a'; fill the ring to n=80
+        for (int i = 0; i < 78; ++i) {
+            llama_sampler_accept(smpl, tok);
+        }
+    });
+
+    printf("Testing backend penalties without filler entries\n");
+    compare_penalties_logits(params, 64, 1.1f, 0.5f, 0.25f, "Hello", [](llama_sampler * smpl) {
+        for (llama_token token = 0; token < 64; ++token) {
+            llama_sampler_accept(smpl, token);
+        }
+    });
+
+    printf("Testing backend top-k followed by penalties\n");
+    compare_top_k_penalties_logits(params, 8, 64, 1.1f, 0.5f, 0.25f, "Hello",
+            penalties_position::after_filter);
+
+    printf("Testing backend penalties followed by top-k\n");
+    compare_top_k_penalties_logits(params, 8, 64, 1.1f, 0.5f, 0.25f, "Hello",
+            penalties_position::before_filter);
+
+    printf("Testing backend top-p followed by penalties\n");
+    compare_masking_penalties_logits(params, "top-p", []() {
+        return llama_sampler_init_top_p(0.9f, 0);
+    }, 64, 1.1f, 0.5f, 0.25f, "Hello", penalties_position::after_filter, true);
+
+    printf("Testing backend top-p followed by penalties with a large history window\n");
+    compare_masking_penalties_logits(params, "top-p large-window", []() {
+        return llama_sampler_init_top_p(0.9f, 0);
+    }, 4096, 1.1f, 0.5f, 0.25f, "Hello", penalties_position::after_filter, true);
+
+    printf("Testing backend penalties followed by top-p\n");
+    compare_masking_penalties_logits(params, "top-p", []() {
+        return llama_sampler_init_top_p(0.9f, 0);
+    }, 64, 1.1f, 0.5f, 0.25f, "Hello", penalties_position::before_filter, true);
+
+    printf("Testing backend min-p followed by penalties\n");
+    compare_masking_penalties_logits(params, "min-p", []() {
+        return llama_sampler_init_min_p(0.1f, 0);
+    }, 64, 1.1f, 0.5f, 0.25f, "Hello", penalties_position::after_filter, false);
+
+    printf("Testing backend penalties followed by min-p\n");
+    compare_masking_penalties_logits(params, "min-p", []() {
+        return llama_sampler_init_min_p(0.1f, 0);
+    }, 64, 1.1f, 0.5f, 0.25f, "Hello", penalties_position::before_filter, false);
+
+    printf("Testing backend top-p followed by penalties with empty history\n");
+    compare_masking_penalties_logits(params, "top-p empty", []() {
+        return llama_sampler_init_top_p(0.9f, 0);
+    }, 64, 1.1f, 0.5f, 0.25f, "Hello", penalties_position::after_filter, true, false);
+
+    printf("Testing backend top-p followed by individual penalties\n");
+    compare_masking_penalties_logits(params, "top-p repeat", []() {
+        return llama_sampler_init_top_p(0.9f, 0);
+    }, 64, 1.1f, 0.0f, 0.0f, "Hello", penalties_position::after_filter, true);
+    compare_masking_penalties_logits(params, "top-p frequency", []() {
+        return llama_sampler_init_top_p(0.9f, 0);
+    }, 64, 1.0f, 0.5f, 0.0f, "Hello", penalties_position::after_filter, true);
+    compare_masking_penalties_logits(params, "top-p presence", []() {
+        return llama_sampler_init_top_p(0.9f, 0);
+    }, 64, 1.0f, 0.0f, 0.25f, "Hello", penalties_position::after_filter, true);
+
+    printf("Testing backend penalty parameter values\n");
+    test_penalty_parameter_values(params);
+
+    printf("backend penalties sampling test PASSED\n");
+}
+
 // This test verifies that it is possible to have two different backend samplers,
 // one that uses the backend dist sampler, and another that uses CPU dist sampler.
 static void test_backend_mixed_sampling(const test_params & params) {
@@ -1014,6 +1575,7 @@ struct backend_test_case {
 static const backend_test_case BACKEND_TESTS[] = {
     { "greedy",          test_backend_greedy_sampling,         true  },
     { "logit_bias",      test_backend_logit_bias_sampling,     true  },
+    { "penalties",       test_backend_penalties_sampling,      true  },
     { "temp",            test_backend_temp_sampling,           true  },
     { "temp_ext",        test_backend_temp_ext_sampling,       true  },
     { "top_k",           test_backend_top_k_sampling,          true  },

tests/test-chat.cpp

@@ -3987,6 +3987,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect_tool_calls({
                 { "special_function", R"({"arg1": 1})", {} },
             })
+            .expect_reconstruction()
             .run();
 
         // Tool call with negative number
@@ -4212,6 +4213,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .expect_tool_calls({
                 { "special_function", R"({"arg1": 1})", {} },
             })
+            .expect_reconstruction()
             .run();
 
         // Tool call with multiple params (mixed types)
@@ -4268,6 +4270,24 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .run();
     }
 
+    {
+        // The DSML separator belongs to the tool call block, not assistant content.
+        auto tst = peg_tester("models/templates/deepseek-ai-DeepSeek-V4-Flash-0731.jinja", detailed_debug);
+        tst.test(
+               "\n\n"
+               "<｜DSML｜tool_calls>\n"
+               "<｜DSML｜invoke name=\"special_function\">\n"
+               "<｜DSML｜parameter name=\"arg1\" string=\"false\">1</｜DSML｜parameter>\n"
+               "</｜DSML｜invoke>\n"
+               "</｜DSML｜tool_calls>")
+            .enable_thinking(false)
+            .reasoning_format(COMMON_REASONING_FORMAT_DEEPSEEK)
+            .tools({ special_function_tool })
+            .expect(message_assist_call)
+            .expect_reconstruction()
+            .run();
+    }
+
     // GLM-4.6 tests - format: <tool_call>function_name\n<arg_key>...</arg_key>\n<arg_value>...</arg_value>\n</tool_call>
     {
         auto tst = peg_tester("models/templates/GLM-4.6.jinja", detailed_debug);
@@ -6359,6 +6379,7 @@ static void test_template_generation_prompt() {
         std::vector<common_chat_msg> messages;
         bool                         add_generation_prompt  = true;
         common_chat_continuation     continue_final_message = COMMON_CHAT_CONTINUATION_NONE;
+        bool                         enable_thinking        = true;
     };
 
     auto basic = [&]() {
@@ -6390,6 +6411,7 @@ static void test_template_generation_prompt() {
         inputs.messages               = opts.messages;
         inputs.add_generation_prompt  = opts.add_generation_prompt;
         inputs.continue_final_message = opts.continue_final_message;
+        inputs.enable_thinking        = opts.enable_thinking;
 
         auto params = common_chat_templates_apply(tmpls.get(), inputs);
 
@@ -6488,6 +6510,156 @@ static void test_template_generation_prompt() {
         check(tmpls, continuation_reasoning(), "<｜Assistant｜><think>I'm");
     }
 
+    const std::string deepseek_v4_reasoning_effort_max = "Reasoning Effort: Absolute maximum";
+    const std::string deepseek_v4_flash_0731_reasoning_effort_max = "Reasoning Effort: Beyond maximum";
+
+    {
+        auto tmpls = read_templates("models/templates/deepseek-ai-DeepSeek-V4.jinja");
+        check(tmpls, basic(),                  "<｜Assistant｜><think>");
+        check(tmpls, continuation_content(),   "<｜Assistant｜><think>I'm thinking</think>Hello, ");
+        check(tmpls, continuation_reasoning(), "<｜Assistant｜><think>I'm");
+
+        auto continuation_content_no_thinking = continuation_content();
+        continuation_content_no_thinking.messages = { system_msg, message_user, simple_assist_msg("Hello, ") };
+        continuation_content_no_thinking.enable_thinking = false;
+        check(tmpls, continuation_content_no_thinking, "<｜Assistant｜></think>Hello, ");
+
+        common_chat_templates_inputs max_inputs;
+        max_inputs.messages = { system_msg, message_user };
+        max_inputs.chat_template_kwargs["reasoning_effort"] = R"("max")";
+        auto max_params = common_chat_templates_apply(tmpls.get(), max_inputs);
+        assert_contains(max_params.prompt, deepseek_v4_reasoning_effort_max);
+
+        auto high_inputs = max_inputs;
+        high_inputs.chat_template_kwargs["reasoning_effort"] = R"("high")";
+        auto high_params = common_chat_templates_apply(tmpls.get(), high_inputs);
+        assert_not_contains(high_params.prompt, deepseek_v4_reasoning_effort_max);
+
+        auto low_inputs = max_inputs;
+        low_inputs.chat_template_kwargs["reasoning_effort"] = R"("low")";
+        auto low_params = common_chat_templates_apply(tmpls.get(), low_inputs);
+        assert_not_contains(low_params.prompt, deepseek_v4_reasoning_effort_max);
+
+        common_chat_templates_inputs default_effort_inputs;
+        default_effort_inputs.messages = { system_msg, message_user };
+        auto default_effort_params = common_chat_templates_apply(tmpls.get(), default_effort_inputs);
+        assert_not_contains(default_effort_params.prompt, deepseek_v4_reasoning_effort_max);
+
+        auto non_thinking_max_inputs = max_inputs;
+        non_thinking_max_inputs.enable_thinking = false;
+        auto non_thinking_max_params = common_chat_templates_apply(tmpls.get(), non_thinking_max_inputs);
+        assert_not_contains(non_thinking_max_params.prompt, deepseek_v4_reasoning_effort_max);
+
+        common_chat_templates_inputs response_format_inputs;
+        response_format_inputs.messages = { system_msg, message_user };
+        response_format_inputs.tools = { get_time_tool };
+        response_format_inputs.json_schema =
+            R"({"type":"object","properties":{"answer":{"type":"string"}}})";
+        auto response_format_params = common_chat_templates_apply(tmpls.get(), response_format_inputs);
+        const auto tools_pos = response_format_params.prompt.find("## Tools");
+        const auto response_format_pos = response_format_params.prompt.find(
+            "## Response Format:\n\nYou MUST strictly adhere to the following schema to reply:\n");
+        if (tools_pos == std::string::npos || response_format_pos == std::string::npos || tools_pos > response_format_pos) {
+            LOG_ERR("Expected response format after tools\nActual: %s\n", response_format_params.prompt.c_str());
+            common_log_flush(common_log_main());
+            throw std::runtime_error("Test failed");
+        }
+        assert_contains(response_format_params.prompt, R"("answer": {"type": "string"})");
+
+        response_format_inputs.json_schema = "{}";
+        auto json_object_params = common_chat_templates_apply(tmpls.get(), response_format_inputs);
+        assert_contains(json_object_params.prompt,
+                        "## Response Format:\n\nYou MUST strictly adhere to the following schema to reply:\n{}");
+
+        common_chat_msg assistant_history;
+        assistant_history.role              = "assistant";
+        assistant_history.content           = "Previous answer";
+        assistant_history.reasoning_content = "Previous reasoning";
+
+        common_chat_msg user_followup;
+        user_followup.role    = "user";
+        user_followup.content = "Follow up";
+
+        common_chat_templates_inputs default_history_inputs;
+        default_history_inputs.messages = { message_user, assistant_history, user_followup };
+        auto default_history_params = common_chat_templates_apply(tmpls.get(), default_history_inputs);
+        assert_contains(default_history_params.prompt, "<｜Assistant｜></think>Previous answer");
+
+        auto drop_thinking_inputs = default_history_inputs;
+        drop_thinking_inputs.chat_template_kwargs["drop_thinking"] = "false";
+        auto drop_thinking_params = common_chat_templates_apply(tmpls.get(), drop_thinking_inputs);
+        assert_contains(drop_thinking_params.prompt, "<｜Assistant｜><think>Previous reasoning</think>Previous answer");
+
+        auto preserve_reasoning_inputs = default_history_inputs;
+        preserve_reasoning_inputs.chat_template_kwargs["preserve_reasoning"] = "true";
+        auto preserve_reasoning_params = common_chat_templates_apply(tmpls.get(), preserve_reasoning_inputs);
+        assert_contains(preserve_reasoning_params.prompt, "<｜Assistant｜><think>Previous reasoning</think>Previous answer");
+        assert_equals(true, common_chat_templates_get_caps(tmpls.get()).at("supports_preserve_reasoning"));
+
+        auto no_preserve_reasoning_inputs = default_history_inputs;
+        no_preserve_reasoning_inputs.chat_template_kwargs["preserve_reasoning"] = "false";
+        auto no_preserve_reasoning_params = common_chat_templates_apply(tmpls.get(), no_preserve_reasoning_inputs);
+        assert_contains(no_preserve_reasoning_params.prompt, "<｜Assistant｜></think>Previous answer");
+
+        common_chat_msg empty_tool_call = simple_assist_msg("", "", "empty_args", "{}");
+        common_chat_templates_inputs empty_tool_inputs;
+        empty_tool_inputs.messages = { message_user, empty_tool_call };
+        empty_tool_inputs.tools    = { empty_args_tool };
+        auto empty_tool_params = common_chat_templates_apply(tmpls.get(), empty_tool_inputs);
+        assert_contains(empty_tool_params.prompt,
+                        "<｜DSML｜invoke name=\"empty_args\">\n\n</｜DSML｜invoke>");
+    }
+
+    {
+        auto tmpls = read_templates("models/templates/deepseek-ai-DeepSeek-V4-Flash-0731.jinja");
+        check(tmpls, basic(),                  "<｜Assistant｜><think>");
+        check(tmpls, continuation_content(),   "<｜Assistant｜><think>I'm thinking</think>Hello, ");
+        check(tmpls, continuation_reasoning(), "<｜Assistant｜><think>I'm");
+
+        auto continuation_content_no_thinking = continuation_content();
+        continuation_content_no_thinking.messages = { system_msg, message_user, simple_assist_msg("Hello, ") };
+        continuation_content_no_thinking.enable_thinking = false;
+        check(tmpls, continuation_content_no_thinking, "<｜Assistant｜></think>Hello, ");
+
+        common_chat_templates_inputs high_inputs;
+        high_inputs.messages = { system_msg, message_user };
+        high_inputs.chat_template_kwargs["reasoning_effort"] = R"("high")";
+        auto high_params = common_chat_templates_apply(tmpls.get(), high_inputs);
+        assert_contains(high_params.prompt, deepseek_v4_reasoning_effort_max);
+
+        auto max_inputs = high_inputs;
+        max_inputs.chat_template_kwargs["reasoning_effort"] = R"("max")";
+        auto max_params = common_chat_templates_apply(tmpls.get(), max_inputs);
+        assert_contains(max_params.prompt, deepseek_v4_flash_0731_reasoning_effort_max);
+
+        auto low_inputs = high_inputs;
+        low_inputs.chat_template_kwargs["reasoning_effort"] = R"("low")";
+        auto low_params = common_chat_templates_apply(tmpls.get(), low_inputs);
+        assert_not_contains(low_params.prompt, deepseek_v4_reasoning_effort_max);
+        assert_not_contains(low_params.prompt, deepseek_v4_flash_0731_reasoning_effort_max);
+
+        common_chat_templates_inputs default_effort_inputs;
+        default_effort_inputs.messages = { system_msg, message_user };
+        auto default_effort_params = common_chat_templates_apply(tmpls.get(), default_effort_inputs);
+        assert_not_contains(default_effort_params.prompt, deepseek_v4_reasoning_effort_max);
+        assert_not_contains(default_effort_params.prompt, deepseek_v4_flash_0731_reasoning_effort_max);
+
+        auto non_thinking_max_inputs = max_inputs;
+        non_thinking_max_inputs.enable_thinking = false;
+        auto non_thinking_max_params = common_chat_templates_apply(tmpls.get(), non_thinking_max_inputs);
+        assert_not_contains(non_thinking_max_params.prompt, deepseek_v4_flash_0731_reasoning_effort_max);
+
+        common_chat_templates_inputs response_format_inputs;
+        response_format_inputs.messages = { system_msg, message_user };
+        response_format_inputs.tools = { get_time_tool };
+        response_format_inputs.json_schema =
+            R"({"type":"object","properties":{"answer":{"type":"string"}}})";
+        auto response_format_params = common_chat_templates_apply(tmpls.get(), response_format_inputs);
+        assert_contains(response_format_params.prompt,
+                        "## Response Format:\n\nYou MUST strictly adhere to the following schema to reply:\n");
+        assert_contains(response_format_params.prompt, R"("answer": {"type": "string"})");
+    }
+
     {
         auto tmpls = read_templates("models/templates/openbmb-MiniCPM5-1B.jinja");
         check(tmpls, basic(),                  "<|im_start|>assistant\n<think>\n");

tests/test-llama-archs.cpp

@@ -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;

tests/test-model-resolution.cpp

@@ -0,0 +1,506 @@
+// tests the HF model resolution and the model handler assembly end-to-end on
+// synthetic repo listings: a local httplib server bound to the loopback
+// serves hardcoded HF API responses, so the real client, hf_cache, resolution
+// and CLI parsing run against them without external network access
+
+#include "arg.h"
+#include "common.h"
+#include "download.h"
+#include "http.h"
+#include "log.h"
+
+#include <nlohmann/json.hpp>
+
+#include <algorithm>
+#include <cstdio>
+#include <cstdlib>
+#include <filesystem>
+#include <map>
+#include <thread>
+#include <string>
+#include <vector>
+
+// the case and reordering being checked, printed with every failure
+static std::string g_context;
+
+// independent of NDEBUG, so the checks stay alive in Release builds
+#define REQUIRE(x) do {                                                         \
+    if (!(x)) {                                                                 \
+        fprintf(stderr, "%s:%d: [%s] REQUIRE(%s) failed\n",                     \
+                __FILE__, __LINE__, g_context.c_str(), #x);                     \
+        std::abort();                                                           \
+    }                                                                           \
+} while (0)
+
+#define REQUIRE_EQ(actual, expected) do {                                       \
+    if (!((actual) == (expected))) {                                            \
+        fprintf(stderr, "%s:%d: [%s] REQUIRE_EQ(%s, %s) failed\n  actual:   '%s'\n  expected: '%s'\n", \
+                __FILE__, __LINE__, g_context.c_str(), #actual, #expected,      \
+                std::string(actual).c_str(), std::string(expected).c_str());    \
+        std::abort();                                                           \
+    }                                                                           \
+} while (0)
+
+//
+// synthetic repos keyed by repo id, served over the loopback by a real
+// httplib server, so the tested code runs its own client and transport
+//
+
+static std::map<std::string, std::vector<std::string>> g_repos;
+
+static const char * COMMIT = "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa";
+
+// the server lives in main, so its destructor runs before the static teardown
+// tears down the winsock state httplib brings in
+static void serve_repos(httplib::Server & server) {
+    server.Get(R"(/api/models/(.+)/refs)", [](const httplib::Request & req, httplib::Response & res) {
+        if (g_repos.count(req.matches[1])) {
+            res.set_content(nlohmann::json{{"branches", {{{"name", "main"}, {"targetCommit", COMMIT}}}}}.dump(),
+                            "application/json");
+        } else {
+            res.status = 404;
+        }
+    });
+    server.Get(R"(/api/models/(.+)/tree/.+)", [](const httplib::Request & req, httplib::Response & res) {
+        if (!g_repos.count(req.matches[1])) {
+            res.status = 404;
+            return;
+        }
+        auto files = nlohmann::json::array();
+        size_t i = 0;
+        for (const auto & p : g_repos[req.matches[1]]) {
+            char oid[41];
+            snprintf(oid, sizeof(oid), "%040lx", (unsigned long) ++i);
+            files.push_back({{"type", "file"}, {"path", p}, {"size", 1}, {"oid", oid}});
+        }
+        res.set_content(files.dump(), "application/json");
+    });
+}
+
+static common_params_model model_ref(const std::string & hf_repo, const std::string & hf_file = "") {
+    common_params_model m;
+    m.hf_repo = hf_repo;
+    m.hf_file = hf_file;
+    return m;
+}
+
+// the model cache is isolated under a temporary directory named after the
+// loopback port, so concurrent runs on a shared machine keep their own, and
+// the local path the handler wires for a file is snapshots/<commit>/<path>
+static std::filesystem::path cache_dir;
+
+static std::string cached(std::string repo_id, const std::string & path) {
+    string_replace_all(repo_id, "/", "--");
+    return (cache_dir / ("models--" + repo_id) / "snapshots" / COMMIT / path).string();
+}
+
+//
+// fixtures mimicking real repo layouts
+//
+
+// flat layout in the style of ggml-org/gemma-4-31B-it-GGUF
+static const std::vector<std::string> flat = {
+    "README.md",
+    "model-BF16.gguf",
+    "model-Q4_K_M.gguf",
+    "model-Q8_0.gguf",
+    "mmproj-model-BF16.gguf",
+    "mmproj-model-Q8_0.gguf",
+    "mtp-model-BF16.gguf",
+    "mtp-model-Q4_0.gguf",
+    "mtp-model-Q8_0.gguf",
+    "dflash-model-BF16.gguf",
+    "dflash-model-Q8_0.gguf",
+};
+
+// quants in subdirectories with sharded files and root sidecars,
+// in the style of stepfun-ai/Step-3.7-Flash-GGUF
+static const std::vector<std::string> subdir = {
+    "mmproj-model-f16.gguf",
+    "model-mtp-BF16.gguf",
+    "model-mtp-Q8_0.gguf",
+    "Q3_K_M/model-Q3_K_M-00001-of-00003.gguf",
+    "Q3_K_M/model-Q3_K_M-00002-of-00003.gguf",
+    "Q3_K_M/model-Q3_K_M-00003-of-00003.gguf",
+    "Q8_0/model-Q8_0-00001-of-00002.gguf",
+    "Q8_0/model-Q8_0-00002-of-00002.gguf",
+};
+
+// sidecar quants exist where the full model quant does not,
+// in the style of ggml-org/Qwen3.6-27B-GGUF
+static const std::vector<std::string> hole = {
+    "model-BF16.gguf",
+    "model-Q4_K_M.gguf",
+    "model-Q8_0.gguf",
+    "mtp-model-BF16.gguf",
+    "mtp-model-Q4_0.gguf",
+    "mtp-model-Q8_0.gguf",
+    "dflash-model-BF16.gguf",
+    "dflash-model-Q8_0.gguf",
+};
+
+// unsloth-style naming with UD quants and a suffix MTP file
+static const std::vector<std::string> unsloth = {
+    "model-UD-Q8_K_XL.gguf",
+    "mmproj-BF16.gguf",
+    "model-MTP-BF16.gguf",
+};
+
+// bartowski-style vendor prefix and mradermacher-style dot quant
+static const std::vector<std::string> vendors = {
+    "TheDrummer_Model-24B-v4.1-Q8_0.gguf",
+    "BlackSheep-24B.Q8_0.gguf",
+};
+
+// every speculative sidecar type at the same quant
+static const std::vector<std::string> quad = {
+    "model-Q8_0.gguf",
+    "mtp-model-Q8_0.gguf",
+    "dflash-model-Q8_0.gguf",
+    "eagle3-model-Q8_0.gguf",
+    "dspark-model-Q8_0.gguf",
+};
+
+static const std::vector<std::string> dflash_only = {
+    "model-Q8_0.gguf",
+    "dflash-model-Q8_0.gguf",
+};
+
+static const std::vector<std::string> eagle3_only = {
+    "model-Q8_0.gguf",
+    "eagle3-model-Q8_0.gguf",
+};
+
+// a single full quant with dspark sidecars at other quants,
+// in the style of ggml-org/DeepSeek-V4-Flash-0731-GGUF
+static const std::vector<std::string> spark = {
+    "README.md",
+    "model-MXFP4.gguf",
+    "dspark-model-BF16.gguf",
+    "dspark-model-MXFP4.gguf",
+};
+
+// dspark outranks dflash in the type auto-selection
+static const std::vector<std::string> dspark_dflash = {
+    "model-Q8_0.gguf",
+    "dflash-model-Q8_0.gguf",
+    "dspark-model-Q8_0.gguf",
+};
+
+//
+// table-driven plan resolution through the real entry point,
+// each case replayed on multiple deterministic reorderings of the listing,
+// except the cases whose pick legitimately depends on the listing order
+//
+
+struct plan_case {
+    const char * name;
+    const std::vector<std::string> & files;
+    const char * hf_repo;
+    const char * hf_file;
+    bool sidecars;        // request mmproj + mtp + dflash + eagle3 + dspark
+    bool order_dependent; // the expected pick depends on the listing order
+    const char * primary;
+    std::vector<std::string> model_files;
+    const char * mmproj;
+    const char * mtp;
+    const char * dflash;
+    const char * eagle3;
+    const char * dspark;
+};
+
+static const plan_case plan_cases[] = {
+    // exact tag picks the matching primary, sidecars follow the tag
+    {"flat exact tag", flat, "test/repo:Q8_0", "", true, false,
+     "model-Q8_0.gguf", {"model-Q8_0.gguf"},
+     "mmproj-model-Q8_0.gguf", "mtp-model-Q8_0.gguf", "dflash-model-Q8_0.gguf", "", ""},
+
+    // no tag falls back to the default quant preference
+    {"flat default", flat, "test/repo", "", false, false,
+     "model-Q4_K_M.gguf", {"model-Q4_K_M.gguf"},
+     "", "", "", "", ""},
+
+    // no tag and no default match falls back to the first model in the listing
+    {"unsloth fallback", unsloth, "test/repo", "", true, true,
+     "model-UD-Q8_K_XL.gguf", {"model-UD-Q8_K_XL.gguf"},
+     "mmproj-BF16.gguf", "", "", "", ""},
+
+    // explicit hf_file picks that exact file
+    {"flat hf_file", flat, "test/repo", "model-BF16.gguf", false, false,
+     "model-BF16.gguf", {"model-BF16.gguf"},
+     "", "", "", "", ""},
+
+    // missing hf_file resolves nothing
+    {"flat missing hf_file", flat, "test/repo", "nope.gguf", false, false,
+     "", {},
+     "", "", "", "", ""},
+
+    // a sharded primary brings all its parts, a subdir primary finds the root sidecar
+    {"subdir shards", subdir, "test/repo:Q3_K_M", "", true, false,
+     "Q3_K_M/model-Q3_K_M-00001-of-00003.gguf",
+     {"Q3_K_M/model-Q3_K_M-00001-of-00003.gguf",
+      "Q3_K_M/model-Q3_K_M-00002-of-00003.gguf",
+      "Q3_K_M/model-Q3_K_M-00003-of-00003.gguf"},
+     "mmproj-model-f16.gguf", "model-mtp-Q8_0.gguf", "", "", ""},
+
+    // a tag with no matching full model still resolves the requested sidecars
+    {"hole tag sidecar", hole, "test/repo:Q4_0", "", true, false,
+     "", {},
+     "", "mtp-model-Q4_0.gguf", "dflash-model-Q8_0.gguf", "", ""},
+
+    // the same tag without a requested sidecar resolves nothing
+    {"hole tag alone", hole, "test/repo:Q4_0", "", false, false,
+     "", {},
+     "", "", "", "", ""},
+
+    // no tag anchors the sidecars on the primary quant
+    {"hole default anchor", hole, "test/repo", "", true, false,
+     "model-Q4_K_M.gguf", {"model-Q4_K_M.gguf"},
+     "", "mtp-model-Q4_0.gguf", "dflash-model-Q8_0.gguf", "", ""},
+
+    // the mtp- keyword is case sensitive, a suffix -MTP file is not discovered
+    {"unsloth suffix mtp", unsloth, "test/repo:Q8_K_XL", "", true, false,
+     "model-UD-Q8_K_XL.gguf", {"model-UD-Q8_K_XL.gguf"},
+     "mmproj-BF16.gguf", "", "", "", ""},
+
+    // vendor prefixes and the dot quant convention both match the tag,
+    // first match wins between two files at the same quant
+    {"vendor prefix", vendors, "test/repo:Q8_0", "", false, true,
+     "TheDrummer_Model-24B-v4.1-Q8_0.gguf", {"TheDrummer_Model-24B-v4.1-Q8_0.gguf"},
+     "", "", "", "", ""},
+
+    // every sidecar type resolves at the tag
+    {"quad exact tag", quad, "test/repo:Q8_0", "", true, false,
+     "model-Q8_0.gguf", {"model-Q8_0.gguf"},
+     "", "mtp-model-Q8_0.gguf", "dflash-model-Q8_0.gguf", "eagle3-model-Q8_0.gguf", "dspark-model-Q8_0.gguf"},
+
+    // no tag anchors the dspark sidecar on the only full quant
+    {"spark default anchor", spark, "test/repo", "", true, false,
+     "model-MXFP4.gguf", {"model-MXFP4.gguf"},
+     "", "", "", "", "dspark-model-MXFP4.gguf"},
+
+    // a tag with no matching full model still resolves the exact dspark sidecar
+    {"spark tag sidecar", spark, "test/repo:BF16", "", true, false,
+     "", {},
+     "", "", "", "", "dspark-model-BF16.gguf"},
+};
+
+static void check_plan(const plan_case & c) {
+    common_download_opts opts;
+    opts.download_mmproj = c.sidecars;
+    opts.download_mtp    = c.sidecars;
+    opts.download_dflash = c.sidecars;
+    opts.download_eagle3 = c.sidecars;
+    opts.download_dspark = c.sidecars;
+
+    auto plan = common_download_get_hf_plan(model_ref(c.hf_repo, c.hf_file), opts);
+
+    REQUIRE_EQ(plan.primary.path, c.primary);
+    REQUIRE_EQ(plan.mmproj.path,  c.mmproj);
+    REQUIRE_EQ(plan.mtp.path,     c.mtp);
+    REQUIRE_EQ(plan.dflash.path,  c.dflash);
+    REQUIRE_EQ(plan.eagle3.path,  c.eagle3);
+    REQUIRE_EQ(plan.dspark.path,  c.dspark);
+
+    // exact shard set, order insensitive; the primary must be the first split
+    std::vector<std::string> actual;
+    for (const auto & f : plan.model_files) {
+        actual.push_back(f.path);
+    }
+    std::sort(actual.begin(), actual.end());
+    auto expected = c.model_files;
+    std::sort(expected.begin(), expected.end());
+    REQUIRE(actual == expected);
+    if (!expected.empty()) {
+        REQUIRE(plan.primary.path == expected.front());
+    }
+}
+
+static void test_plan_resolution() {
+    printf("test-model-resolution: plan resolution on %zu cases\n", sizeof(plan_cases) / sizeof(plan_cases[0]));
+
+    for (const auto & c : plan_cases) {
+        printf("  %s\n", c.name);
+        // invariant: the resolution is insensitive to the listing order
+        for (size_t rot = 0; rot < c.files.size(); ++rot) {
+            if (c.order_dependent && rot > 0) {
+                continue;
+            }
+            g_context = std::string(c.name) + ", reordering " + std::to_string(rot);
+            auto files = c.files;
+            std::rotate(files.begin(), files.begin() + rot, files.end());
+            if (rot % 2 == 1) {
+                std::reverse(files.begin(), files.end());
+            }
+            g_repos["test/repo"] = files;
+            check_plan(c);
+        }
+    }
+    g_repos.clear();
+}
+
+//
+// end-to-end assembly: real CLI parsing, real handler init resolving over the
+// loopback, downloads skipped by flipping offline before apply
+//
+
+static void assemble(std::vector<std::string> argv, common_params & params) {
+    std::vector<char *> cargv;
+    g_context.clear();
+    for (auto & a : argv) {
+        g_context += g_context.empty() ? a : " " + a;
+        cargv.push_back(a.data());
+    }
+    bool ok = common_params_parse((int) cargv.size(), cargv.data(), params, LLAMA_EXAMPLE_SERVER);
+    REQUIRE(ok);
+
+    auto handler = common_models_handler_init(params, LLAMA_EXAMPLE_SERVER);
+
+    // skip the network execution, on_done still wires the params
+    params.offline = true;
+    common_models_handler_apply(handler, params);
+}
+
+static void test_task_assembly() {
+    printf("test-model-resolution: end-to-end assembly\n");
+
+    g_repos["test/main"]   = flat;
+    g_repos["test/hole"]   = hole;
+    g_repos["test/quad"]   = quad;
+    g_repos["test/dflash"] = dflash_only;
+    g_repos["test/eagle3"] = eagle3_only;
+    g_repos["test/spark"]  = spark;
+    g_repos["test/pair"]   = dspark_dflash;
+    g_repos["test/small"]  = {"draft-model-Q4_K_M.gguf"};
+    g_repos["test/preset"] = {"preset.ini", "model-Q8_0.gguf"};
+
+    {
+        // plain -hf wires the model and its mmproj, nothing speculative
+        common_params params;
+        assemble({"server", "-hf", "test/main:Q8_0"}, params);
+        REQUIRE_EQ(params.model.path,  cached("test/main", "model-Q8_0.gguf"));
+        REQUIRE_EQ(params.mmproj.path, cached("test/main", "mmproj-model-Q8_0.gguf"));
+        REQUIRE(params.speculative.draft.mparams.path.empty());
+    }
+    {
+        // --no-mmproj disables the mmproj discovery
+        common_params params;
+        assemble({"server", "-hf", "test/main:Q8_0", "--no-mmproj"}, params);
+        REQUIRE(params.mmproj.path.empty());
+    }
+    {
+        // an explicit --mmproj wins over the discovery
+        common_params params;
+        assemble({"server", "-hf", "test/main:Q8_0", "--mmproj", "/local/mmproj.gguf"}, params);
+        REQUIRE(params.mmproj.path == "/local/mmproj.gguf");
+    }
+    {
+        // -hf with a spec type wires the sidecar of the main repo as fallback draft
+        common_params params;
+        assemble({"server", "-hf", "test/main:Q8_0", "--spec-type", "draft-mtp"}, params);
+        REQUIRE_EQ(params.speculative.draft.mparams.path, cached("test/main", "mtp-model-Q8_0.gguf"));
+    }
+    {
+        // -hfd with a spec type wires the draft repo sidecar at its tag,
+        // not its full model, and suppresses the main repo fallback
+        common_params params;
+        assemble({"server", "-hf", "test/hole:Q8_0", "-hfd", "test/hole:Q4_0", "--spec-type", "draft-mtp"}, params);
+        REQUIRE_EQ(params.speculative.draft.mparams.path, cached("test/hole", "mtp-model-Q4_0.gguf"));
+    }
+    {
+        // an explicit -md file wins over the sidecar resolution
+        common_params params;
+        assemble({"server", "-hf", "test/main:Q8_0", "-hfd", "test/main", "-md", "mtp-model-BF16.gguf", "--spec-type", "draft-mtp"}, params);
+        REQUIRE_EQ(params.speculative.draft.mparams.path, cached("test/main", "mtp-model-BF16.gguf"));
+    }
+    {
+        // -hfd without a spec type auto-selects the type, mtp first when all ship
+        common_params params;
+        assemble({"server", "-hf", "test/main:Q8_0", "-hfd", "test/quad:Q8_0"}, params);
+        REQUIRE(params.speculative.types == std::vector<enum common_speculative_type>{COMMON_SPECULATIVE_TYPE_DRAFT_MTP});
+        REQUIRE_EQ(params.speculative.draft.mparams.path, cached("test/quad", "mtp-model-Q8_0.gguf"));
+    }
+    {
+        // auto-selection with only a dflash sidecar
+        common_params params;
+        assemble({"server", "-hf", "test/main:Q8_0", "-hfd", "test/dflash:Q8_0"}, params);
+        REQUIRE(params.speculative.types == std::vector<enum common_speculative_type>{COMMON_SPECULATIVE_TYPE_DRAFT_DFLASH});
+        REQUIRE_EQ(params.speculative.draft.mparams.path, cached("test/dflash", "dflash-model-Q8_0.gguf"));
+    }
+    {
+        // auto-selection with only an eagle3 sidecar
+        common_params params;
+        assemble({"server", "-hf", "test/main:Q8_0", "-hfd", "test/eagle3:Q8_0"}, params);
+        REQUIRE(params.speculative.types == std::vector<enum common_speculative_type>{COMMON_SPECULATIVE_TYPE_DRAFT_EAGLE3});
+        REQUIRE_EQ(params.speculative.draft.mparams.path, cached("test/eagle3", "eagle3-model-Q8_0.gguf"));
+    }
+    {
+        // auto-selection prefers dspark over dflash when both ship
+        common_params params;
+        assemble({"server", "-hf", "test/main:Q8_0", "-hfd", "test/pair:Q8_0"}, params);
+        REQUIRE(params.speculative.types == std::vector<enum common_speculative_type>{COMMON_SPECULATIVE_TYPE_DRAFT_DSPARK});
+        REQUIRE_EQ(params.speculative.draft.mparams.path, cached("test/pair", "dspark-model-Q8_0.gguf"));
+    }
+    {
+        // -hf with the dspark spec type wires the sidecar of the main repo,
+        // anchored on the only full quant
+        common_params params;
+        assemble({"server", "-hf", "test/spark", "--spec-type", "draft-dspark"}, params);
+        REQUIRE_EQ(params.model.path, cached("test/spark", "model-MXFP4.gguf"));
+        REQUIRE_EQ(params.speculative.draft.mparams.path, cached("test/spark", "dspark-model-MXFP4.gguf"));
+    }
+    {
+        // -hfd on a repo without sidecars keeps resolving a full model as draft
+        common_params params;
+        assemble({"server", "-hf", "test/main:Q8_0", "-hfd", "test/small"}, params);
+        REQUIRE(params.speculative.types == std::vector<enum common_speculative_type>{COMMON_SPECULATIVE_TYPE_NONE});
+        REQUIRE_EQ(params.speculative.draft.mparams.path, cached("test/small", "draft-model-Q4_K_M.gguf"));
+    }
+    {
+        // a preset repo wires the preset and clears the model for router mode
+        common_params params;
+        assemble({"server", "-hf", "test/preset"}, params);
+        REQUIRE_EQ(params.models_preset, cached("test/preset", "preset.ini"));
+        REQUIRE(params.model.path.empty());
+        REQUIRE(params.model.hf_repo.empty());
+    }
+
+    g_repos.clear();
+}
+
+int main(void) {
+    // unbuffered, so a crash cannot swallow the reports already printed
+    setvbuf(stdout, nullptr, _IONBF, 0);
+    setvbuf(stderr, nullptr, _IONBF, 0);
+
+    // the negative cases legitimately log errors on every reordering,
+    // keep the output down to the reports
+    common_log_pause(common_log_main());
+
+    // the loopback endpoint also keeps the client init from rejecting
+    // https on the builds without TLS support
+    httplib::Server server;
+    serve_repos(server);
+    int port = server.bind_to_any_port("127.0.0.1");
+
+    // isolate the cache, its location is read once so it is set
+    // before anything else
+    cache_dir = std::filesystem::temp_directory_path() /
+                ("test-model-resolution-cache-" + std::to_string(port));
+    std::filesystem::remove_all(cache_dir);
+    common_set_env("LLAMA_CACHE", cache_dir.string());
+
+    std::thread server_thread([&server] { server.listen_after_bind(); });
+    server.wait_until_ready();
+    common_set_env("MODEL_ENDPOINT", "http://127.0.0.1:" + std::to_string(port) + "/");
+
+    test_plan_resolution();
+    test_task_assembly();
+
+    server.stop();
+    server_thread.join();
+
+    std::filesystem::remove_all(cache_dir);
+    printf("test-model-resolution: all tests OK\n");
+    return 0;
+}

tests/test-sampling.cpp

@@ -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]);

tools/cli/README.md

@@ -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) |

tools/completion/README.md

@@ -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
 

tools/gguf-split/gguf-split.cpp

@@ -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) {

tools/gguf-split/tests.sh

@@ -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