hip: compile debug builds with -O2 on hip to avoid a compiler bug (#20392)

Co-authored-by: Mishusha <pmv26021975@gmail.com>

.github/workflows/build.yml

@@ -469,6 +469,7 @@ jobs:
           cd build
           export GGML_VK_VISIBLE_DEVICES=0
           export GGML_VK_DISABLE_F16=1
+          export GGML_VK_DISABLE_COOPMAT=1
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 4800
 

benches/nemotron/nemotron-dgx-spark.md

@@ -0,0 +1,72 @@
+# NVIDIA DGX Spark
+
+## System info
+
+```bash
+uname --all
+Linux spark-17ed 6.11.0-1016-nvidia #16-Ubuntu SMP PREEMPT_DYNAMIC Sun Sep 21 16:52:46 UTC 2025 aarch64 aarch64 aarch64 GNU/Linux
+
+g++ --version
+g++ (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0
+
+nvidia-smi
+Fri Mar  6 11:39:45 2026
++-----------------------------------------------------------------------------------------+
+| NVIDIA-SMI 580.95.05              Driver Version: 580.95.05      CUDA Version: 13.0     |
++-----------------------------------------+------------------------+----------------------+
+| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
+| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
+|                                         |                        |               MIG M. |
+|=========================================+========================+======================|
+|   0  NVIDIA GB10                    On  |   0000000F:01:00.0 Off |                  N/A |
+| N/A   52C    P0             13W /  N/A  | Not Supported          |      0%      Default |
+|                                         |                        |                  N/A |
++-----------------------------------------+------------------------+----------------------+
+```
+
+## ggml-org/nemotron-3-super-120b-GGUF
+
+Model: https://huggingface.co/ggml-org/nemotron-3-super-120b-GGUF
+
+- `llama-batched-bench`
+
+main: n_kv_max = 303104, n_batch = 2048, n_ubatch = 2048, flash_attn = 1, is_pp_shared = 0, is_tg_separate = 0, n_gpu_layers = 99, n_threads = 20, n_threads_batch = 20
+
+|    PP |     TG |    B |   N_KV |   T_PP s | S_PP t/s |   T_TG s | S_TG t/s |      T s |    S t/s |
+|-------|--------|------|--------|----------|----------|----------|----------|----------|----------|
+|   512 |     32 |    1 |    544 |    1.094 |   468.05 |    1.621 |    19.74 |    2.715 |   200.37 |
+|   512 |     32 |    2 |   1088 |    1.463 |   700.16 |    2.437 |    26.26 |    3.900 |   279.01 |
+|   512 |     32 |    4 |   2176 |    2.647 |   773.76 |    4.043 |    31.66 |    6.689 |   325.29 |
+|   512 |     32 |    8 |   4352 |    5.291 |   774.14 |    6.151 |    41.62 |   11.442 |   380.37 |
+|   512 |     32 |   16 |   8704 |   10.603 |   772.62 |   10.385 |    49.30 |   20.987 |   414.72 |
+|   512 |     32 |   32 |  17408 |   21.231 |   771.69 |   18.235 |    56.16 |   39.466 |   441.09 |
+|  4096 |     32 |    1 |   4128 |    5.340 |   767.05 |    1.616 |    19.81 |    6.956 |   593.47 |
+|  4096 |     32 |    2 |   8256 |   10.673 |   767.55 |    2.454 |    26.08 |   13.127 |   628.94 |
+|  4096 |     32 |    4 |  16512 |   21.348 |   767.46 |    4.072 |    31.44 |   25.420 |   649.57 |
+|  4096 |     32 |    8 |  33024 |   42.714 |   767.15 |    6.277 |    40.78 |   48.991 |   674.08 |
+|  4096 |     32 |   16 |  66048 |   85.385 |   767.54 |   10.596 |    48.32 |   95.981 |   688.14 |
+|  4096 |     32 |   32 | 132096 |  170.819 |   767.32 |   18.619 |    55.00 |  189.437 |   697.31 |
+|  8192 |     32 |    1 |   8224 |   10.690 |   766.32 |    1.619 |    19.76 |   12.310 |   668.10 |
+|  8192 |     32 |    2 |  16448 |   21.382 |   766.24 |    2.467 |    25.94 |   23.850 |   689.65 |
+|  8192 |     32 |    4 |  32896 |   42.782 |   765.92 |    4.098 |    31.23 |   46.881 |   701.69 |
+|  8192 |     32 |    8 |  65792 |   85.582 |   765.77 |    6.368 |    40.20 |   91.951 |   715.52 |
+|  8192 |     32 |   16 | 131584 |  171.066 |   766.21 |   10.774 |    47.52 |  181.840 |   723.62 |
+|  8192 |     32 |   32 | 263168 |  342.140 |   766.19 |   18.969 |    53.98 |  361.109 |   728.78 |
+
+
+- `llama-bench`
+
+| model                   |       size |     params | backend    | n_ubatch | fa |            test |                  t/s |
+| ----------------------- | ---------: | ---------: | ---------- | -------: | -: | --------------: | -------------------: |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |          pp2048 |        768.84 ± 0.90 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |            tg32 |         19.94 ± 0.16 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |  pp2048 @ d4096 |        764.51 ± 0.50 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |    tg32 @ d4096 |         19.95 ± 0.18 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |  pp2048 @ d8192 |        759.53 ± 0.71 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |    tg32 @ d8192 |         19.83 ± 0.18 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 | pp2048 @ d16384 |        747.98 ± 1.58 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |   tg32 @ d16384 |         19.84 ± 0.18 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 | pp2048 @ d32768 |        724.40 ± 2.70 |
+| nemotron 120B.A12B Q4_K |  65.10 GiB |   120.67 B | CUDA       |     2048 |  1 |   tg32 @ d32768 |         19.45 ± 0.18 |
+
+build: 04a65daab (8268)

common/CMakeLists.txt

@@ -81,6 +81,8 @@ add_library(${TARGET} STATIC
     preset.cpp
     preset.h
     regex-partial.cpp
+    reasoning-budget.cpp
+    reasoning-budget.h
     regex-partial.h
     sampling.cpp
     sampling.h

common/arg.cpp

@@ -2913,6 +2913,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             auto parsed = json::parse(value);
             for (const auto & item : parsed.items()) {
+                if (item.key() == "enable_thinking") {
+                    LOG_WRN("Setting 'enable_thinking' via --chat-template-kwargs is deprecated. "
+                            "Use --reasoning on / --reasoning off instead.\n");
+                }
                 params.default_template_kwargs[item.key()] = item.value().dump();
             }
         }
@@ -3048,14 +3052,39 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.reasoning_format = common_reasoning_format_from_name(value);
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_THINK"));
+    add_opt(common_arg(
+        {"-rea", "--reasoning"}, "[on|off|auto]",
+        "Use reasoning/thinking in the chat ('on', 'off', or 'auto', default: 'auto' (detect from template))",
+        [](common_params & params, const std::string & value) {
+            if (is_truthy(value)) {
+                params.enable_reasoning = 1;
+                params.default_template_kwargs["enable_thinking"] = "true";
+            } else if (is_falsey(value)) {
+                params.enable_reasoning = 0;
+                params.default_template_kwargs["enable_thinking"] = "false";
+            } else if (is_autoy(value)) {
+                params.enable_reasoning = -1;
+            } else {
+                throw std::invalid_argument(
+                    string_format("error: unknown value for --reasoning: '%s'\n", value.c_str()));
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_REASONING"));
     add_opt(common_arg(
         {"--reasoning-budget"}, "N",
-        "controls the amount of thinking allowed; currently only one of: -1 for unrestricted thinking budget, or 0 to disable thinking (default: -1)",
+        "token budget for thinking: -1 for unrestricted, 0 for immediate end, N>0 for token budget (default: -1)",
         [](common_params & params, int value) {
-            if (value != 0 && value != -1) { throw std::invalid_argument("invalid value"); }
+            if (value < -1) { throw std::invalid_argument("invalid value"); }
             params.reasoning_budget = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_THINK_BUDGET"));
+    add_opt(common_arg(
+        {"--reasoning-budget-message"}, "MESSAGE",
+        "message injected before the end-of-thinking tag when reasoning budget is exhausted (default: none)",
+        [](common_params & params, const std::string & value) {
+            params.reasoning_budget_message = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_THINK_BUDGET_MESSAGE"));
     add_opt(common_arg(
         {"--chat-template"}, "JINJA_TEMPLATE",
         string_format(

common/chat-auto-parser-generator.cpp

@@ -135,7 +135,9 @@ common_peg_parser analyze_reasoning::build_parser(parser_build_context & ctx) co
     if (thinking_forced_open || thinking_forced_closed) {
         // Thinking is forced open OR forced closed with enable_thinking=true
         // In both cases, expect only the closing tag (opening was in template)
-        return p.reasoning(p.until(end)) + end;
+        // However, since we might have incorrectly detected the open/close pattern,
+        // we admit an optional starting marker
+        return p.optional(p.literal(start)) + p.reasoning(p.until(end)) + end;
     }
     if (mode == reasoning_mode::TAG_BASED || mode == reasoning_mode::TOOLS_ONLY) {
         // Standard tag-based reasoning OR tools-only mode (reasoning appears with tools)

common/chat-peg-parser.cpp

@@ -6,7 +6,7 @@
 
 #include <nlohmann/json.hpp>
 
-using json = nlohmann::ordered_json;
+using ordered_json = nlohmann::ordered_json;
 
 static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
     int count = 0;
@@ -68,7 +68,7 @@ static int json_brace_depth(const std::string & s) {
 
 // JSON-escape a string and return the inner content (without surrounding quotes).
 static std::string escape_json_string_inner(const std::string & s) {
-    std::string escaped = json(s).dump();
+    std::string escaped = ordered_json(s).dump();
     if (escaped.size() >= 2 && escaped.front() == '"' && escaped.back() == '"') {
         return escaped.substr(1, escaped.size() - 2);
     }
@@ -309,7 +309,7 @@ void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
         if (arg_count > 0) {
             arg_entry = ",";
         }
-        arg_entry += json(trim(node.text)).dump() + ":";
+        arg_entry += ordered_json(trim(node.text)).dump() + ":";
         ++arg_count;
 
         auto & target = args_target();
@@ -343,7 +343,7 @@ void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
 
             // Try to parse as JSON value (number, bool, null, object, array)
             try {
-                json parsed = json::parse(value_content);
+                ordered_json parsed = ordered_json::parse(value_content);
                 if (parsed.is_string()) {
                     // Don't add closing quote yet (added by arg_close) for monotonic streaming
                     std::string escaped = parsed.dump();
@@ -408,7 +408,7 @@ void common_chat_peg_mapper::map(const common_peg_ast_node & node) {
 
 common_peg_parser common_chat_peg_builder::standard_constructed_tools(
     const std::map<std::string, std::string> & markers,
-    const nlohmann::json &                     tools,
+    const ordered_json &                       tools,
     bool                                       parallel_tool_calls,
     bool                                       force_tool_calls) {
     if (!tools.is_array() || tools.empty()) {
@@ -439,7 +439,7 @@ common_peg_parser common_chat_peg_builder::standard_constructed_tools(
         }
         const auto &   function = tool_def.at("function");
         std::string    name     = function.at("name");
-        nlohmann::json params = function.contains("parameters") ? function.at("parameters") : nlohmann::json::object();
+        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
 
         // Build argument parsers
         auto args = eps();
@@ -479,8 +479,8 @@ common_peg_parser common_chat_peg_builder::standard_constructed_tools(
 // Python-style tool calls: name(arg1="value1", arg2=123)
 // Used only by LFM2 for now, so we don't merge it into autoparser
 common_peg_parser common_chat_peg_builder::python_style_tool_calls(
-    const nlohmann::json & tools,
-    bool                   parallel_tool_calls) {
+    const ordered_json & tools,
+    bool                 parallel_tool_calls) {
     if (!tools.is_array() || tools.empty()) {
         return eps();
     }
@@ -493,7 +493,7 @@ common_peg_parser common_chat_peg_builder::python_style_tool_calls(
         }
         const auto &   function = tool_def.at("function");
         std::string    name     = function.at("name");
-        nlohmann::json params = function.contains("parameters") ? function.at("parameters") : nlohmann::json::object();
+        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
 
         auto args = eps();
         if (params.contains("properties") && !params["properties"].empty()) {
@@ -555,11 +555,11 @@ static std::pair<std::string, std::string> parse_key_spec(const std::string & ke
 
 // Mode 1: function_is_key — parse {"function_name": {...}}
 common_peg_parser common_chat_peg_builder::build_json_tools_function_is_key(
-    const nlohmann::json & tools,
-    const std::string &    args_key,
-    const std::string &    effective_args_key,
-    const std::string &    call_id_key,
-    const std::string &    gen_call_id_key) {
+    const ordered_json & tools,
+    const std::string &  args_key,
+    const std::string &  effective_args_key,
+    const std::string &  call_id_key,
+    const std::string &  gen_call_id_key) {
 
     auto tool_choices = choice();
 
@@ -569,7 +569,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_function_is_key(
         }
         const auto &   function = tool_def.at("function");
         std::string    name     = function.at("name");
-        nlohmann::json params = function.contains("parameters") ? function.at("parameters") : nlohmann::json::object();
+        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
 
         // Build inner object fields
         std::vector<common_peg_parser> inner_fields;
@@ -634,11 +634,11 @@ common_peg_parser common_chat_peg_builder::build_json_tools_function_is_key(
 
 // Mode 2: Nested keys (dot notation like "function.name")
 common_peg_parser common_chat_peg_builder::build_json_tools_nested_keys(
-    const nlohmann::json & tools,
-    const std::string &    effective_name_key,
-    const std::string &    effective_args_key,
-    const std::string &    call_id_key,
-    const std::string &    gen_call_id_key) {
+    const ordered_json & tools,
+    const std::string &  effective_name_key,
+    const std::string &  effective_args_key,
+    const std::string &  call_id_key,
+    const std::string &  gen_call_id_key) {
 
     auto tool_choices = choice();
 
@@ -655,7 +655,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_nested_keys(
         }
         const auto &   function = tool_def.at("function");
         std::string    name     = function.at("name");
-        nlohmann::json params = function.contains("parameters") ? function.at("parameters") : nlohmann::json::object();
+        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
 
         auto nested_name = literal("\"" + nested_name_field + "\"") + space() + literal(":") + space() +
                           literal("\"") + tool_name(literal(name)) + literal("\"");
@@ -706,7 +706,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_nested_keys(
 
 // Mode 3: Flat keys with optional ID fields and parameter ordering
 common_peg_parser common_chat_peg_builder::build_json_tools_flat_keys(
-    const nlohmann::json &           tools,
+    const ordered_json &             tools,
     const std::string &              effective_name_key,
     const std::string &              effective_args_key,
     const std::string &              call_id_key,
@@ -723,7 +723,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_flat_keys(
         }
         const auto &   function = tool_def.at("function");
         std::string    name     = function.at("name");
-        nlohmann::json params = function.contains("parameters") ? function.at("parameters") : nlohmann::json::object();
+        ordered_json   params   = function.contains("parameters") ? function.at("parameters") : ordered_json::object();
 
         auto tool_name_ = name_key_parser + space() + literal(":") + space() +
                          literal("\"") + tool_name(literal(name)) + literal("\"");
@@ -791,7 +791,7 @@ common_peg_parser common_chat_peg_builder::build_json_tools_flat_keys(
 common_peg_parser common_chat_peg_builder::standard_json_tools(
                                                        const std::string &              section_start,
                                                        const std::string &              section_end,
-                                                       const nlohmann::json &           tools,
+                                                       const ordered_json &             tools,
                                                        bool                             parallel_tool_calls,
                                                        bool                             force_tool_calls,
                                                        const std::string &              name_key,

common/chat-peg-parser.h

@@ -94,7 +94,7 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // parameters_order: order in which JSON fields should be parsed
     common_peg_parser standard_json_tools(const std::string &              section_start,
                                           const std::string &              section_end,
-                                          const nlohmann::json &           tools,
+                                          const nlohmann::ordered_json &   tools,
                                           bool                             parallel_tool_calls,
                                           bool                             force_tool_calls,
                                           const std::string &              name_key = "",
@@ -108,30 +108,30 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // Legacy-compatible helper for building XML/tagged style tool calls
     // Used by tests and manual parsers
     common_peg_parser standard_constructed_tools(const std::map<std::string, std::string> & markers,
-                                                 const nlohmann::json &                     tools,
+                                                 const nlohmann::ordered_json &             tools,
                                                  bool                                       parallel_tool_calls,
                                                  bool                                       force_tool_calls);
 
     // Helper for Python-style function call format: name(arg1="value1", arg2=123)
     // Used by LFM2 and similar templates
-    common_peg_parser python_style_tool_calls(const nlohmann::json & tools,
-                                              bool                   parallel_tool_calls);
+    common_peg_parser python_style_tool_calls(const nlohmann::ordered_json & tools,
+                                              bool                           parallel_tool_calls);
 
   private:
     // Implementation helpers for standard_json_tools — one per JSON tool call layout mode
-    common_peg_parser build_json_tools_function_is_key(const nlohmann::json & tools,
-                                                       const std::string &    args_key,
-                                                       const std::string &    effective_args_key,
-                                                       const std::string &    call_id_key,
-                                                       const std::string &    gen_call_id_key);
+    common_peg_parser build_json_tools_function_is_key(const nlohmann::ordered_json & tools,
+                                                       const std::string &            args_key,
+                                                       const std::string &            effective_args_key,
+                                                       const std::string &            call_id_key,
+                                                       const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_nested_keys(const nlohmann::json & tools,
-                                                   const std::string &    effective_name_key,
-                                                   const std::string &    effective_args_key,
-                                                   const std::string &    call_id_key,
-                                                   const std::string &    gen_call_id_key);
+    common_peg_parser build_json_tools_nested_keys(const nlohmann::ordered_json & tools,
+                                                   const std::string &            effective_name_key,
+                                                   const std::string &            effective_args_key,
+                                                   const std::string &            call_id_key,
+                                                   const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_flat_keys(const nlohmann::json &           tools,
+    common_peg_parser build_json_tools_flat_keys(const nlohmann::ordered_json &   tools,
                                                  const std::string &              effective_name_key,
                                                  const std::string &              effective_args_key,
                                                  const std::string &              call_id_key,

common/chat.cpp

@@ -857,7 +857,9 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
     auto extract_reasoning = inputs.reasoning_format != COMMON_REASONING_FORMAT_NONE;
     auto include_grammar   = true;
 
-    data.supports_thinking = true;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = "[THINK]";
+    data.thinking_end_tag   = "[/THINK]";
     data.prompt            = common_chat_template_direct_apply(tmpl, inputs, /* messages_override = */ adjusted_messages);
     data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
     data.preserved_tokens  = {
@@ -1165,9 +1167,11 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
                                                           const autoparser::templates_params & inputs) {
     common_chat_params data;
 
-    data.prompt            = common_chat_template_direct_apply(tmpl, inputs);
-    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
-    data.supports_thinking = true;
+    data.prompt             = common_chat_template_direct_apply(tmpl, inputs);
+    data.format             = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking  = true;
+    data.thinking_start_tag = "<think>";
+    data.thinking_end_tag   = "</think>";
     data.preserved_tokens  = {
         "<|tool_calls_section_begin|>",
         "<|tool_calls_section_end|>",
@@ -1350,6 +1354,77 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
     return data;
 }
 
+static common_chat_params common_chat_params_init_gigachat_v3(
+        const common_chat_template & tmpl,
+        const autoparser::templates_params & inputs) {
+
+    common_chat_params data;
+
+    data.prompt            = common_chat_template_direct_apply(tmpl, inputs);
+    data.format            = COMMON_CHAT_FORMAT_PEG_NATIVE;
+    data.supports_thinking = false;
+    data.preserved_tokens  = {
+        "<|message_sep|>\n\n",
+        "<|role_sep|>\n",
+    };
+
+    auto has_tools         = inputs.tools.is_array() && !inputs.tools.empty();
+    auto include_grammar   = has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE;
+    auto tool_call_start_prefix = "<|message_sep|>\n\nfunction call<|role_sep|>\n";
+
+    auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
+        if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
+            // Build a choice of all available tools
+            auto tool_choice = p.choice();
+            for (const auto & tool : inputs.tools) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                const auto & schema = function.at("parameters");
+
+                auto tool_name = p.json_member("name", "\"" + p.tool_name(p.literal(name)) + "\"");
+                auto tool_args = p.json_member("arguments", p.tool_args(p.schema(p.json(), "tool-" + name + "-schema", schema)));
+
+                auto tool_open = p.tool_open(p.literal("{") << tool_name);
+
+                tool_choice |= p.rule("tool-" + name, tool_open << "," << tool_args << "}");
+            }
+
+            // Define the tool call structure
+            auto min_calls = inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED ? 1 : 0;
+            auto max_calls = 1; // parallel toolcalls are not supported
+            auto tool_call = p.rule("tool-call", p.literal(tool_call_start_prefix) + tool_choice);
+            auto tool_calls = p.trigger_rule("tool-call-root", p.repeat(tool_call, /* min = */ min_calls, /* max = */ max_calls));
+
+            return p.content(p.until("<|message_sep|>\n\n")) << tool_calls;
+        }
+
+        // Content only parser
+        include_grammar = false;
+        return p.content(p.rest());
+
+    });
+
+    data.parser = parser.save();
+
+    if (include_grammar) {
+        data.grammar_lazy = has_tools && inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_AUTO;
+
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                auto schema = function.at("parameters");
+                builder.resolve_refs(schema);
+            });
+            parser.build_grammar(builder, data.grammar_lazy);
+        });
+
+        data.grammar_triggers = {
+            {COMMON_GRAMMAR_TRIGGER_TYPE_WORD, tool_call_start_prefix}
+        };
+    }
+    return data;
+}
+
 namespace workaround {
 
 static void map_developer_role_to_system(json & messages) {
@@ -1521,12 +1596,31 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
         return common_chat_params_init_lfm2(tmpl, params);
     }
 
+    // GigaChatV3 format detection
+    if (src.find("<|role_sep|>") != std::string::npos &&
+        src.find("<|message_sep|>") != std::string::npos &&
+        src.find("<|function_call|>") == std::string::npos
+    ) {
+        LOG_DBG("Using specialized template: GigaChatV3\n");
+        return common_chat_params_init_gigachat_v3(tmpl, params);
+    }
+
     try {
         LOG_DBG("Using differential autoparser\n");
         struct autoparser::autoparser autoparser;
         autoparser.analyze_template(tmpl);
         auto auto_params = autoparser::peg_generator::generate_parser(tmpl, params, autoparser);
         auto_params.supports_thinking = autoparser.reasoning.mode != autoparser::reasoning_mode::NONE;
+        if (auto_params.supports_thinking) {
+            auto_params.thinking_start_tag = autoparser.reasoning.start;
+            auto_params.thinking_end_tag   = autoparser.reasoning.end;
+            // FORCED_OPEN and FORCED_CLOSED both put <think> in the generation prompt
+            // (FORCED_CLOSED forces empty <think></think> when thinking is disabled,
+            //  but forces <think> open when thinking is enabled)
+            auto_params.thinking_forced_open =
+                autoparser.reasoning.mode == autoparser::reasoning_mode::FORCED_OPEN ||
+                autoparser.reasoning.mode == autoparser::reasoning_mode::FORCED_CLOSED;
+        }
         return auto_params;
     } catch (const std::exception & e) {
         throw std::invalid_argument(std::string("Unable to generate parser for this template. Automatic parser generation failed: ") + e.what());

common/chat.h

@@ -213,6 +213,8 @@ struct common_chat_params {
     bool                                grammar_lazy         = false;
     bool                                thinking_forced_open = false;
     bool                                supports_thinking    = false;
+    std::string                         thinking_start_tag;  // e.g., "<think>"
+    std::string                         thinking_end_tag;    // e.g., "</think>"
     std::vector<common_grammar_trigger> grammar_triggers;
     std::vector<std::string>            preserved_tokens;
     std::vector<std::string>            additional_stops;

common/common.h

@@ -235,6 +235,14 @@ struct common_params_sampling {
     std::vector<llama_logit_bias> logit_bias;     // logit biases to apply
     std::vector<llama_logit_bias> logit_bias_eog; // pre-calculated logit biases for EOG tokens
 
+    // reasoning budget sampler parameters
+    // these are populated by the server/CLI based on chat template params
+    int32_t                  reasoning_budget_tokens   = -1;   // -1 = disabled, >= 0 = token budget
+    bool                     reasoning_budget_activate_immediately = false;
+    std::vector<llama_token> reasoning_budget_start;           // start tag token sequence
+    std::vector<llama_token> reasoning_budget_end;             // end tag token sequence
+    std::vector<llama_token> reasoning_budget_forced;          // forced sequence (message + end tag)
+
     bool backend_sampling = false;
 
     bool has_logit_bias() const {
@@ -536,7 +544,9 @@ struct common_params {
     bool use_jinja = true;                                                                                  // NOLINT
     bool enable_chat_template = true;
     common_reasoning_format reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK;
+    int enable_reasoning = -1; // -1 = auto, 0 = disable, 1 = enable
     int reasoning_budget = -1;
+    std::string reasoning_budget_message; // message injected before end tag when budget exhausted
     bool prefill_assistant = true; // if true, any trailing assistant message will be prefilled into the response
     int sleep_idle_seconds = -1;   // if >0, server will sleep after this many seconds of idle time
 
@@ -916,7 +926,7 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 // MoE utils
 //
 
-const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate)_(ch|)exps";
+const char * const LLM_FFN_EXPS_REGEX = "\\.ffn_(up|down|gate|gate_up)_(ch|)exps";
 
 inline std::string llm_ffn_exps_block_regex(int idx) {
     return string_format("blk\\.%d%s", idx, LLM_FFN_EXPS_REGEX);

common/reasoning-budget.cpp

@@ -0,0 +1,219 @@
+#include "reasoning-budget.h"
+#include "common.h"
+#include "unicode.h"
+
+#include "log.h"
+
+#include <cmath>
+#include <cstdint>
+#include <string>
+#include <vector>
+
+struct token_matcher {
+    std::vector<llama_token> tokens;
+    size_t pos = 0;
+
+    bool advance(llama_token token) {
+        if (tokens.empty()) {
+            return false;
+        }
+
+        if (token == tokens[pos]) {
+            pos++;
+            if (pos >= tokens.size()) {
+                pos = 0;
+                return true;
+            }
+        } else {
+            pos = 0;
+            if (token == tokens[0]) {
+                pos = 1;
+            }
+        }
+        return false;
+    }
+
+    void reset() { pos = 0; }
+};
+
+struct common_reasoning_budget_ctx {
+    const llama_vocab * vocab;
+
+    token_matcher start_matcher;
+    token_matcher end_matcher;
+    std::vector<llama_token> forced_tokens;
+
+    int32_t budget;           // maximum tokens in reasoning block
+    int32_t remaining;        // tokens remaining in budget
+
+    common_reasoning_budget_state state;
+
+    // for forcing
+    size_t force_pos;         // next position in forced_tokens to force
+};
+
+static const char * common_reasoning_budget_name(const struct llama_sampler * /*smpl*/) {
+    return "reasoning-budget";
+}
+
+static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_token token) {
+    auto * ctx = (common_reasoning_budget_ctx *) smpl->ctx;
+
+    switch (ctx->state) {
+        case REASONING_BUDGET_IDLE:
+        {
+            if (ctx->start_matcher.advance(token)) {
+                ctx->state = REASONING_BUDGET_COUNTING;
+                ctx->remaining = ctx->budget;
+                LOG_INF("reasoning-budget: activated, budget=%d tokens\n", ctx->budget);
+
+                if (ctx->remaining <= 0) {
+                    ctx->state = REASONING_BUDGET_FORCING;
+                    ctx->force_pos = 0;
+                    LOG_INF("reasoning-budget: budget=0, forcing immediately\n");
+                }
+            }
+            break;
+        }
+        case REASONING_BUDGET_COUNTING:
+        case REASONING_BUDGET_WAITING_UTF8:
+        {
+            if (ctx->end_matcher.advance(token)) {
+                ctx->state = REASONING_BUDGET_DONE;
+                LOG_INF("reasoning-budget: deactivated (natural end)\n");
+                break;
+            }
+
+            bool utf8_complete = true;
+            if (ctx->vocab != nullptr) {
+                const std::string piece = common_token_to_piece(ctx->vocab, token, false);
+                utf8_complete = common_utf8_is_complete(piece);
+            }
+
+            if (ctx->state == REASONING_BUDGET_WAITING_UTF8) {
+                if (utf8_complete) {
+                    ctx->state = REASONING_BUDGET_FORCING;
+                    ctx->force_pos = 0;
+                    ctx->end_matcher.reset();
+                    LOG_INF("reasoning-budget: UTF-8 complete, now forcing end sequence\n");
+                }
+            } else if (ctx->state == REASONING_BUDGET_COUNTING) {
+                ctx->remaining--;
+                if (ctx->remaining <= 0) {
+                    if (utf8_complete) {
+                        ctx->state = REASONING_BUDGET_FORCING;
+                        ctx->force_pos = 0;
+                        ctx->end_matcher.reset();
+                        LOG_INF("reasoning-budget: budget exhausted, forcing end sequence\n");
+                    } else {
+                        ctx->state = REASONING_BUDGET_WAITING_UTF8;
+                        ctx->end_matcher.reset();
+                        LOG_INF("reasoning-budget: budget exhausted, waiting for UTF-8 completion\n");
+                    }
+                }
+            }
+            break;
+        }
+        case REASONING_BUDGET_FORCING:
+            // force_pos is advanced in apply(), not here.
+            // This ensures the first forced token isn't skipped when the sampler
+            // is initialized directly in FORCING state (e.g. COUNTING + budget=0)
+            break;
+        case REASONING_BUDGET_DONE:
+            break;
+    }
+}
+
+static void common_reasoning_budget_apply(struct llama_sampler * smpl, llama_token_data_array * cur_p) {
+    auto * ctx = (common_reasoning_budget_ctx *) smpl->ctx;
+
+    if (ctx->state != REASONING_BUDGET_FORCING) {
+        // passthrough — don't modify logits
+        return;
+    }
+
+    if (ctx->force_pos >= ctx->forced_tokens.size()) {
+        return;
+    }
+
+    const llama_token forced = ctx->forced_tokens[ctx->force_pos];
+
+    // set all logits to -inf except the forced token
+    for (size_t i = 0; i < cur_p->size; i++) {
+        if (cur_p->data[i].id != forced) {
+            cur_p->data[i].logit = -INFINITY;
+        }
+    }
+
+    // advance to next forced token (done here rather than in accept so that
+    // the first forced token isn't skipped when starting in FORCING state)
+    ctx->force_pos++;
+    if (ctx->force_pos >= ctx->forced_tokens.size()) {
+        ctx->state = REASONING_BUDGET_DONE;
+        LOG_INF("reasoning-budget: forced sequence complete, done\n");
+    }
+}
+
+static void common_reasoning_budget_reset(struct llama_sampler * smpl) {
+    auto * ctx = (common_reasoning_budget_ctx *) smpl->ctx;
+    ctx->state = REASONING_BUDGET_IDLE;
+    ctx->remaining = ctx->budget;
+    ctx->start_matcher.reset();
+    ctx->end_matcher.reset();
+    ctx->force_pos = 0;
+}
+
+static struct llama_sampler * common_reasoning_budget_clone(const struct llama_sampler * smpl) {
+    const auto * ctx = (const common_reasoning_budget_ctx *) smpl->ctx;
+    return common_reasoning_budget_init(
+        ctx->vocab,
+        ctx->start_matcher.tokens,
+        ctx->end_matcher.tokens,
+        ctx->forced_tokens,
+        ctx->budget,
+        ctx->state);
+}
+
+static void common_reasoning_budget_free(struct llama_sampler * smpl) {
+    delete (common_reasoning_budget_ctx *) smpl->ctx;
+}
+
+static struct llama_sampler_i common_reasoning_budget_i = {
+    /* .name              = */ common_reasoning_budget_name,
+    /* .accept            = */ common_reasoning_budget_accept,
+    /* .apply             = */ common_reasoning_budget_apply,
+    /* .reset             = */ common_reasoning_budget_reset,
+    /* .clone             = */ common_reasoning_budget_clone,
+    /* .free              = */ common_reasoning_budget_free,
+    /* .backend_init      = */ nullptr,
+    /* .backend_accept    = */ nullptr,
+    /* .backend_apply     = */ nullptr,
+    /* .backend_set_input = */ nullptr,
+};
+
+struct llama_sampler * common_reasoning_budget_init(
+        const struct llama_vocab       * vocab,
+        const std::vector<llama_token> & start_tokens,
+        const std::vector<llama_token> & end_tokens,
+        const std::vector<llama_token> & forced_tokens,
+        int32_t                          budget,
+        common_reasoning_budget_state    initial_state) {
+    // promote COUNTING with budget <= 0 to FORCING
+    if (initial_state == REASONING_BUDGET_COUNTING && budget <= 0) {
+        initial_state = REASONING_BUDGET_FORCING;
+    }
+
+    return llama_sampler_init(
+        /* .iface = */ &common_reasoning_budget_i,
+        /* .ctx   = */ new common_reasoning_budget_ctx {
+            /* .vocab         = */ vocab,
+            /* .start_matcher = */ { start_tokens, 0 },
+            /* .end_matcher   = */ { end_tokens, 0 },
+            /* .forced_tokens = */ forced_tokens,
+            /* .budget        = */ budget,
+            /* .remaining     = */ budget,
+            /* .state         = */ initial_state,
+            /* .force_pos     = */ 0,
+        }
+    );
+}

common/reasoning-budget.h

@@ -0,0 +1,41 @@
+#pragma once
+
+#include "llama.h"
+
+#include <cstdint>
+#include <vector>
+
+enum common_reasoning_budget_state {
+    REASONING_BUDGET_IDLE,         // waiting for start sequence
+    REASONING_BUDGET_COUNTING,     // counting down tokens
+    REASONING_BUDGET_FORCING,      // forcing budget message + end sequence
+    REASONING_BUDGET_WAITING_UTF8, // budget exhausted, waiting for UTF-8 completion
+    REASONING_BUDGET_DONE,         // passthrough forever
+};
+
+// Creates a reasoning budget sampler that limits token generation inside a
+// reasoning block (e.g. between <think> and </think>).
+//
+// State machine: IDLE -> COUNTING -> WAITING_UTF8 -> FORCING -> DONE
+//   IDLE:         passthrough, watching for start_tokens sequence
+//   COUNTING:     counting down remaining tokens, watching for natural end_tokens
+//   WAITING_UTF8: budget exhausted, allowing tokens to complete a UTF-8 sequence
+//   FORCING:      forces forced_tokens token-by-token (all other logits -> -inf)
+//   DONE:         passthrough forever
+//
+// Parameters:
+//   vocab         - vocabulary (used for UTF-8 boundary detection; can be nullptr)
+//   start_tokens  - token sequence that activates counting
+//   end_tokens    - token sequence for natural deactivation
+//   forced_tokens - token sequence forced when budget expires
+//   budget        - max tokens allowed in the reasoning block
+//   initial_state - initial state of the sampler (e.g. IDLE or COUNTING)
+//                   note: COUNTING with budget <= 0 is promoted to FORCING
+//
+struct llama_sampler * common_reasoning_budget_init(
+        const struct llama_vocab       * vocab,
+        const std::vector<llama_token> & start_tokens,
+        const std::vector<llama_token> & end_tokens,
+        const std::vector<llama_token> & forced_tokens,
+        int32_t                          budget,
+        common_reasoning_budget_state    initial_state);

common/sampling.cpp

@@ -2,6 +2,7 @@
 
 #include "common.h"
 #include "log.h"
+#include "reasoning-budget.h"
 
 #include <algorithm>
 #include <cmath>
@@ -250,6 +251,17 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
         }
     }
 
+    // reasoning budget sampler — added first so it can force tokens before other samplers
+    if (params.reasoning_budget_tokens >= 0 && !params.reasoning_budget_forced.empty()) {
+        samplers.push_back(common_reasoning_budget_init(
+            vocab,
+            params.reasoning_budget_start,
+            params.reasoning_budget_end,
+            params.reasoning_budget_forced,
+            params.reasoning_budget_tokens,
+            params.reasoning_budget_activate_immediately ? REASONING_BUDGET_COUNTING : REASONING_BUDGET_IDLE));
+    }
+
     if (params.has_logit_bias()) {
         samplers.push_back(llama_sampler_init_logit_bias(llama_vocab_n_tokens(vocab), params.logit_bias.size(), params.logit_bias.data()));
     }

common/unicode.cpp

@@ -1,8 +1,10 @@
 #include "unicode.h"
+
+#include <algorithm>
 #include <cassert>
 #include <stdexcept>
-#include <vector>
 #include <string>
+#include <vector>
 
 // implementation adopted from src/unicode.cpp
 
@@ -67,6 +69,20 @@ utf8_parse_result common_parse_utf8_codepoint(std::string_view input, size_t off
     return utf8_parse_result(utf8_parse_result::INVALID);
 }
 
+bool common_utf8_is_complete(const std::string & s) {
+    if (s.empty()) {
+        return true;
+    }
+    for (int i = 1; i <= std::min(4, (int)s.size()); i++) {
+        unsigned char c = s[s.size() - i];
+        if ((c & 0xC0) != 0x80) {
+            int expected = (c >= 0xF0) ? 4 : (c >= 0xE0) ? 3 : (c >= 0xC0) ? 2 : 1;
+            return i >= expected;
+        }
+    }
+    return false;
+}
+
 std::string common_unicode_cpts_to_utf8(const std::vector<uint32_t> & cps) {
     std::string result;
     for (size_t i = 0; i < cps.size(); ++i) {

common/unicode.h

@@ -20,6 +20,9 @@ struct utf8_parse_result {
 // Returns 0 for invalid first bytes
 size_t common_utf8_sequence_length(unsigned char first_byte);
 
+// Check if a string ends with a complete UTF-8 sequence.
+bool common_utf8_is_complete(const std::string & s);
+
 // Parse a single UTF-8 codepoint from input
 utf8_parse_result common_parse_utf8_codepoint(std::string_view input, size_t offset);
 

convert_hf_to_gguf.py

@@ -144,6 +144,7 @@ class ModelBase:
         self.metadata_override = metadata_override
         self.model_name = model_name
         self.dir_model_card = dir_model  # overridden in convert_lora_to_gguf.py
+        self._is_nvfp4 = False
 
         # Apply heuristics to figure out typical tensor encoding based on first tensor's dtype
         # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie.
@@ -271,6 +272,9 @@ class ModelBase:
         return tensors
 
     def dequant_model(self):
+        if self._is_nvfp4:
+            return  # NVFP4 weights are repacked in _generate_nvfp4_tensors
+
         tensors_to_remove: list[str] = []
         new_tensors: dict[str, Callable[[], Tensor]] = {}
 
@@ -516,6 +520,13 @@ class ModelBase:
         raise NotImplementedError("set_gguf_parameters() must be implemented in subclasses")
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # skip NVFP4 auxiliary tensors (handled in _generate_nvfp4_tensors)
+        if self._is_nvfp4:
+            if name.endswith((".weight_scale", ".weight_scale_2", ".input_scale", ".k_scale", ".v_scale")):
+                return []
+            if name.endswith(".weight") and name.replace(".weight", ".weight_scale") in self.model_tensors:
+                return []
+
         new_name = self.map_tensor_name(name)
 
         # Handle gate/up expert tensor fusion if enabled
@@ -551,9 +562,135 @@ class ModelBase:
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         return ()
 
+    @staticmethod
+    def _nvfp4_pack(weight: Tensor, scale: Tensor) -> tuple[np.ndarray, list[int]]:
+        """Repack NVFP4 ModelOpt tensors into ggml super-block layout.
+        Preserves original E4M3 scale bits as UE4M3 (strip sign bit).
+        The per-tensor scale2 factor is stored as a separate tensor and applied at inference time via ggml_mul().
+        Returns (raw_data, logical_shape)."""
+
+        out_features = weight.shape[0]
+        n_blocks = scale.shape[1]
+
+        # Unpack ModelOpt nibble-packed weights
+        w = weight.reshape(out_features, n_blocks, 8)
+        vals = torch.stack([w & 0x0F, w >> 4], dim=-1).reshape(out_features, n_blocks, 16)
+
+        # Preserve original E4M3 scale bits as UE4M3 (strip sign bit)
+        d_ue = scale.view(torch.uint8).numpy().reshape(out_features, n_blocks) & 0x7F
+        qs = (vals[:, :, :8] | (vals[:, :, 8:] << 4)).to(torch.uint8).numpy()
+
+        # Pack into super-blocks: [4 UE4M3 scales, 32 qs bytes] = 36 bytes per 64 elements
+        n_super = n_blocks // 4
+        d_grouped = d_ue.reshape(out_features, n_super, 4)
+        qs_grouped = qs.reshape(out_features, n_super, 4, 8).reshape(out_features, n_super, 32)
+        raw = np.concatenate([d_grouped, qs_grouped], axis=-1).reshape(out_features, n_super * 36)
+        return raw, [out_features, n_super * 64]
+
+    @staticmethod
+    def _nvfp4_scale2_is_trivial(scale2: Tensor) -> bool:
+        return scale2.numel() <= 1 and abs(float(scale2.float().sum()) - 1.0) < 1e-6
+
+    def _repack_nvfp4(self, new_name: str, weight: Tensor, scale: Tensor, scale2: Tensor):
+        raw, shape = self._nvfp4_pack(weight, scale)
+        logger.info(f"Repacked {new_name} with shape {shape} and quantization NVFP4")
+        self.gguf_writer.add_tensor(new_name, raw, raw_dtype=gguf.GGMLQuantizationType.NVFP4)
+
+        # Emit per-tensor scale2 as a separate F32 tensor when non-trivial
+        if not self._nvfp4_scale2_is_trivial(scale2):
+            scale2_f32 = scale2.float().numpy().flatten()
+            scale_name = new_name.replace(".weight", ".scale")
+            logger.info(f"  + {scale_name} (per-tensor NVFP4 scale2, shape [{scale2_f32.size}])")
+            self.gguf_writer.add_tensor(scale_name, scale2_f32)
+
+    def _generate_nvfp4_tensors(self):
+        # Per-layer expert merging to avoid holding all experts in memory
+        expert_blocks: dict[tuple[int, str], list[tuple[int, np.ndarray]]] = {}
+        expert_scales: dict[tuple[int, str], list[tuple[int, float]]] = {}
+        expert_shapes: dict[tuple[int, str], list[int]] = {}
+        n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=True) or 0
+
+        for name in list(self.model_tensors.keys()):
+            if not name.endswith(".weight"):
+                continue
+            scale_name = name.replace(".weight", ".weight_scale")
+            scale2_name = name.replace(".weight", ".weight_scale_2")
+            if scale_name not in self.model_tensors:
+                continue
+            # Force eager materialization of lazy tensors
+            weight = LazyTorchTensor.to_eager(self.model_tensors[name]())
+            scale = LazyTorchTensor.to_eager(self.model_tensors[scale_name]())
+            scale2 = LazyTorchTensor.to_eager(self.model_tensors.get(scale2_name, lambda: torch.tensor(1.0))())
+
+            # Check if this is a per-expert tensor
+            m = re.search(r'\.experts\.(\d+)\.(gate_proj|up_proj|down_proj)\.weight$', name)
+            if m:
+                expert_id = int(m.group(1))
+                proj_type = m.group(2)
+                bid_m = re.search(r'\.layers\.(\d+)\.', name)
+                bid = int(bid_m.group(1)) if bid_m else 0
+                key = (bid, proj_type)
+
+                raw, shape = self._nvfp4_pack(weight, scale)
+
+                if key not in expert_blocks:
+                    expert_blocks[key] = []
+                    expert_scales[key] = []
+                    expert_shapes[key] = shape
+                expert_blocks[key].append((expert_id, raw.copy()))
+                # Collect per-expert scale2 (scalar per expert)
+                expert_scales[key].append((expert_id, float(scale2.float().sum())))
+
+                # Flush when all experts for this (layer, proj) are collected
+                if n_experts > 0 and len(expert_blocks[key]) >= n_experts:
+                    self._flush_nvfp4_experts(key, expert_blocks, expert_scales, expert_shapes, bid, proj_type)
+            else:
+                new_name = self.map_tensor_name(name)
+                self._repack_nvfp4(new_name, weight, scale, scale2)
+
+        # Flush any remaining experts (fallback if n_experts was unknown)
+        for (bid, proj_type) in list(expert_blocks.keys()):
+            self._flush_nvfp4_experts((bid, proj_type), expert_blocks, expert_scales, expert_shapes, bid, proj_type)
+
+    def _flush_nvfp4_experts(self, key, expert_blocks, expert_scales, expert_shapes, bid, proj_type):
+        experts = expert_blocks.pop(key)
+        scales = expert_scales.pop(key)
+        shape = expert_shapes.pop(key)
+
+        experts.sort(key=lambda x: x[0])
+        merged = np.stack([e[1] for e in experts], axis=0)
+        merged_name = f"model.layers.{bid}.mlp.experts.{proj_type}.weight"
+        new_name = self.map_tensor_name(merged_name)
+        logger.info(f"Repacked {new_name} with shape [{len(experts)}, {shape[0]}, {shape[1]}] and quantization NVFP4")
+        self.gguf_writer.add_tensor(new_name, merged, raw_dtype=gguf.GGMLQuantizationType.NVFP4)
+
+        # Emit per-expert scale2 tensor if any expert has non-trivial scale2
+        scales.sort(key=lambda x: x[0])
+        scale_vals = np.array([s[1] for s in scales], dtype=np.float32)
+        if not np.allclose(scale_vals, 1.0, atol=1e-6):
+            scale_name = new_name.replace(".weight", ".scale")
+            logger.info(f"  + {scale_name} (per-expert NVFP4 scale2, shape [{len(scales)}])")
+            self.gguf_writer.add_tensor(scale_name, scale_vals)
+
+        del experts, merged
+
     def prepare_tensors(self):
+        # detect NVFP4 quantization (ModelOpt format)
+        quant_algo = (self.hparams.get("quantization_config") or {}).get("quant_algo")
+        quant_config_file = self.dir_model / "hf_quant_config.json"
+
+        if not quant_algo and quant_config_file.is_file():
+            with open(quant_config_file, "r", encoding="utf-8") as f:
+                quant_algo = (json.load(f).get("quantization") or {}).get("quant_algo")
+
+        self._is_nvfp4 = quant_algo == "NVFP4"
+
         self.dequant_model()
 
+        # NVFP4 weights are repacked and written directly to gguf_writer
+        if self._is_nvfp4:
+            self._generate_nvfp4_tensors()
+
         # Handle empty tensor_map for models with block_count=0 (like MobileNetV5)
         if self.tensor_map.mapping:
             max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
@@ -4303,6 +4440,14 @@ class Qwen2MoeModel(TextModel):
         # process the experts separately
         name = name.replace("language_model.", "") # InternVL
 
+        # NVFP4 expert weights are handled in _generate_nvfp4_tensors
+        if self._is_nvfp4 and "experts" in name:
+            if name.endswith((".weight", ".weight_scale", ".weight_scale_2", ".input_scale")):
+                if name.endswith(".weight") and name.replace(".weight", ".weight_scale") in self.model_tensors:
+                    return
+                if not name.endswith(".weight"):
+                    return
+
         # handle aggregated expert tensors
         # GGUF stores dimensions reversed from PyTorch, so:
         # PyTorch (A,B,C) -> GGUF writes [C,B,A] -> GGML reads ne={C,B,A}
@@ -4917,7 +5062,7 @@ class Phi2Model(TextModel):
         self.gguf_writer.add_add_bos_token(False)
 
 
-@ModelBase.register("Phi3ForCausalLM")
+@ModelBase.register("Phi3ForCausalLM", "Phi4ForCausalLMV")
 class Phi3MiniModel(TextModel):
     model_arch = gguf.MODEL_ARCH.PHI3
 
@@ -5092,6 +5237,129 @@ class Phi3MiniModel(TextModel):
         yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32))
         yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith(("model.vision_tower.", "vision_tower.", "model.mm_projector.", "mm_projector.")):
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("Phi4ForCausalLMV")
+class Phi4VisionMmprojModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+
+        self.vision_total_layers = int(self.find_vparam(self.n_block_keys))
+        if self.vision_total_layers < 2:
+            raise ValueError(
+                f"Phi-4 vision mmproj conversion requires at least 2 vision layers, got {self.vision_total_layers}"
+            )
+
+        # Phi-4 uses SigLIP2 hidden_states[-2], so export one fewer encoder block and
+        # drop post-layernorm/head weights. This makes the GGUF runtime output match
+        # the feature map consumed by the patched siglip.cpp Phi-4 projector path.
+        self.vision_export_layers = self.vision_total_layers - 1
+        self.vision_last_layer_idx = self.vision_total_layers - 1
+
+        for key in self.n_block_keys:
+            if key in self.hparams_vision:
+                self.hparams_vision[key] = self.vision_export_layers
+                break
+
+        self.block_count = self.vision_export_layers
+        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.MMPROJ, self.block_count)
+
+        patch_size = self.preprocessor_config.get("patch_size")
+        if patch_size is None:
+            raise KeyError("Phi-4 vision mmproj conversion requires patch_size in preprocessor_config.json")
+
+        self.hparams_vision["patch_size"] = patch_size
+
+        pos_emb_name = next(
+            (
+                name for name in self.model_tensors
+                if name.endswith("vision_model.embeddings.position_embedding.weight")
+            ),
+            None,
+        )
+        if pos_emb_name is None:
+            raise KeyError("Phi-4 vision mmproj conversion could not find position_embedding.weight")
+
+        pos_emb_shape = self.model_tensors[pos_emb_name]().shape
+        base_grid_tokens = int(pos_emb_shape[0])
+        grid_side = math.isqrt(base_grid_tokens)
+        if grid_side * grid_side != base_grid_tokens:
+            raise ValueError(f"Unexpected Phi-4 position embedding shape: {tuple(pos_emb_shape)}")
+
+        self.hparams_vision["image_size"] = grid_side * patch_size
+
+        min_num_patches = self.preprocessor_config.get("min_num_patches", self.global_config.get("min_num_patches"))
+        max_num_patches = self.preprocessor_config.get("max_num_patches", self.global_config.get("max_num_patches"))
+        if min_num_patches is None or max_num_patches is None:
+            raise KeyError("Phi-4 vision mmproj conversion requires min_num_patches and max_num_patches")
+
+        self.min_pixels = int(min_num_patches) * patch_size * patch_size
+        self.max_pixels = int(max_num_patches) * patch_size * patch_size
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.PHI4)
+        self.gguf_writer.add_vision_min_pixels(self.min_pixels)
+        self.gguf_writer.add_vision_max_pixels(self.max_pixels)
+        self.gguf_writer.add_vision_use_gelu(True)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams_vision.get("layer_norm_eps", 1e-6))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith(("model.vision_tower.vision_tower.", "vision_tower.")):
+            if ".vision_model.head." in name:
+                return
+
+            new_name = name.replace("model.vision_tower.vision_tower.", "vision_tower.")
+
+            if ".vision_model.post_layernorm." in new_name:
+                return
+
+            if bid is not None and bid == self.vision_last_layer_idx:
+                return
+
+            if new_name.endswith("vision_model.embeddings.patch_embedding.weight"):
+                assert self.hparams_vision is not None
+                if data_torch.ndim != 2:
+                    raise ValueError(f"Unexpected Phi-4 patch embedding shape: {tuple(data_torch.shape)}")
+
+                patch_area = self.hparams_vision["patch_size"] ** 2
+                in_features = data_torch.shape[1]
+                if in_features % patch_area != 0:
+                    raise ValueError(
+                        f"Phi-4 patch embedding input dim {in_features} is not divisible by patch area {patch_area}"
+                    )
+
+                num_channels = in_features // patch_area
+                patch_size = self.hparams_vision["patch_size"]
+                data_torch = data_torch.view(data_torch.shape[0], patch_size, patch_size, num_channels)
+                data_torch = data_torch.permute(0, 3, 1, 2)
+
+            yield from super().modify_tensors(data_torch, new_name, bid)
+            return
+
+        if name.startswith(("model.mm_projector.", "mm_projector.")):
+            local_name = name
+            local_name = local_name.replace("model.mm_projector.", "")
+            local_name = local_name.replace("mm_projector.", "")
+
+            if not (local_name.startswith("0.") or local_name.startswith("2.")):
+                return
+
+            suffix = ".bias" if local_name.endswith(".bias") else ".weight"
+            mm_idx = int(local_name.split(".", maxsplit=1)[0])
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.V_MMPROJ, mm_idx, suffix=suffix), data_torch)
+            return
+
+        return
+
 
 @ModelBase.register("PhiMoEForCausalLM")
 class PhiMoeModel(Phi3MiniModel):
@@ -9743,20 +10011,35 @@ class NemotronHModel(GraniteHybridModel):
         # M: Mamba2, *: Attention, -: MLP
         # MoE:
         # M: Mamba2, *: Attention, E: Expert
-        hybrid_override_pattern = self.hparams["hybrid_override_pattern"]
-        self._ssm_layers = [i for i, val in enumerate(hybrid_override_pattern) if val == "M"]
-        self._mlp_layers = [i for i, val in enumerate(hybrid_override_pattern) if val == ("E" if self.is_moe else "-")]
+        pattern = self.hparams.get("hybrid_override_pattern") or self.hparams.get("layers_block_type")
+        if pattern is None:
+            self._ssm_layers = []
+            self._mlp_layers = []
+        elif isinstance(pattern, str):
+            self._ssm_layers = [i for i, val in enumerate(pattern) if val == "M"]
+            self._mlp_layers = [i for i, val in enumerate(pattern) if val == ("E" if self.is_moe else "-")]
+        else:
+            self._ssm_layers = [i for i, val in enumerate(pattern) if val == "mamba"]
+            self._mlp_layers = [i for i, val in enumerate(pattern) if val == "moe"]
 
     def get_attn_layers(self):
-        hybrid_override_pattern = self.hparams["hybrid_override_pattern"]
-        assert len(hybrid_override_pattern) == self.block_count, "Mismatch between hybrid override and num_hidden_layers!"
-        return [i for i, val in enumerate(hybrid_override_pattern) if val == "*"]
+        pattern = self.hparams.get("hybrid_override_pattern") or self.hparams.get("layers_block_type")
+        if pattern is None:
+            return []
+        assert len(pattern) == self.block_count, f"Mismatch between pattern ({len(pattern)}) and block_count ({self.block_count})!"
+        if isinstance(pattern, str):
+            return [i for i, val in enumerate(pattern) if val == "*"]
+
+        return [i for i, val in enumerate(pattern) if val == "attention"]
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
 
-        self.gguf_writer.add_key_length(self.head_dim)
-        self.gguf_writer.add_value_length(self.head_dim)
+        head_dim = self.head_dim
+        if head_dim is None:
+            raise ValueError("Could not find the attention head dim in config")
+        self.gguf_writer.add_key_length(head_dim)
+        self.gguf_writer.add_value_length(head_dim)
 
         # Set feed_forward_length
         # NOTE: This will trigger an override warning. This is preferable to
@@ -9784,6 +10067,9 @@ class NemotronHModel(GraniteHybridModel):
             if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
                 self.gguf_writer.add_expert_used_count(n_experts_used)
 
+            if (latent_size := self.hparams.get("moe_latent_size")) is not None:
+                self.gguf_writer.add_moe_latent_size(latent_size)
+
     def set_vocab(self):
         super().set_vocab()
 
@@ -9803,6 +10089,13 @@ class NemotronHModel(GraniteHybridModel):
             name = name[len("language_model."):]
 
         if self.is_moe and bid is not None:
+            # Skip Multi-Token Prediction (MTP) tensors. These are used for
+            # for speculative decoding but we don't include them in this model
+            # conversion. See https://github.com/ggml-org/llama.cpp/pull/18886
+            if "mtp" in name:
+                logger.info(f"gguf: Skipping MTP (Speculative) layer: {name}")
+                return []
+
             if name.endswith("mixer.gate.e_score_correction_bias"):
                 new_name = name.replace("e_score_correction_bias", "e_score_correction.bias")
                 yield from ModelBase.modify_tensors(self, data_torch, new_name, bid)

docs/backend/SYCL.md

@@ -382,17 +382,27 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 ## Windows
 
-### I. Setup Environment
-
-1. Install GPU driver
+### Install GPU driver
 
 Intel GPU drivers instructions guide and download page can be found here: [Get Intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
 
-2. Install Visual Studio
+### Option 1: download the binary package directly
+
+Download the binary package for Windows from: https://github.com/ggml-org/llama.cpp/releases.
+
+Extract the package to local folder, run the llama tools directly. Refer to [Run the inference](#iii-run-the-inference-1).
+
+Note, the package includes the SYCL running time and all depended dll files, no need to install oneAPI package and activte them.
+
+### Option 2: build locally from the source code.
+
+#### I. Setup environment
+
+1. Install Visual Studio
 
 If you already have a recent version of Microsoft Visual Studio, you can skip this step. Otherwise, please refer to the official download page for [Microsoft Visual Studio](https://visualstudio.microsoft.com/).
 
-3. Install Intel® oneAPI Base toolkit
+2. Install Intel® oneAPI Base toolkit
 
 SYCL backend depends on:
   - Intel® oneAPI DPC++/C++ compiler/running-time.
@@ -443,25 +453,25 @@ Output (example):
 [ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Iris(R) Xe Graphics 1.3 [1.3.28044]
 ```
 
-4. Install build tools
+3. Install build tools
 
 a. Download & install cmake for Windows: https://cmake.org/download/ (CMake can also be installed from Visual Studio Installer)
 b. The new Visual Studio will install Ninja as default. (If not, please install it manually: https://ninja-build.org/)
 
 
-### II. Build llama.cpp
+#### II. Build llama.cpp
 
 You could download the release package for Windows directly, which including binary files and depended oneAPI dll files.
 
 Choose one of following methods to build from source code.
 
-#### 1. Script
+##### Option 1: Script
 
 ```sh
 .\examples\sycl\win-build-sycl.bat
 ```
 
-#### 2. CMake
+##### Option 2: CMake
 
 On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
@@ -490,7 +500,7 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-completion
 ```
 
-#### 3. Visual Studio
+##### Option 3: Visual Studio
 
 You have two options to use Visual Studio to build llama.cpp:
 - As CMake Project using CMake presets.
@@ -500,7 +510,7 @@ You have two options to use Visual Studio to build llama.cpp:
 
 All following commands are executed in PowerShell.
 
-##### - Open as a CMake Project
+###### - Open as a CMake Project
 
 You can use Visual Studio to open the `llama.cpp` folder directly as a CMake project. Before compiling, select one of the SYCL CMake presets:
 
@@ -515,7 +525,7 @@ You can use Visual Studio to open the `llama.cpp` folder directly as a CMake pro
     cmake --build build --config Release -j --target llama-completion
     ```
 
-##### - Generating a Visual Studio Solution
+###### - Generating a Visual Studio Solution
 
 You can use Visual Studio solution to build and work on llama.cpp on Windows. You need to convert the CMake Project into a `.sln` file.
 
@@ -603,7 +613,7 @@ found 2 SYCL devices:
 
 ```
 
-#### Choose level-zero devices
+##### Choose level-zero devices
 
 |Chosen Device ID|Setting|
 |-|-|
@@ -611,7 +621,7 @@ found 2 SYCL devices:
 |1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
 |0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"` or `set ONEAPI_DEVICE_SELECTOR="level_zero:*"`|
 
-#### Execute
+##### Execute
 
 Choose one of following methods to run.
 
@@ -669,7 +679,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 ## Environment Variable
 
-#### Build
+### Build
 
 | Name               | Value                                 | Function                                    |
 |--------------------|---------------------------------------|---------------------------------------------|
@@ -684,7 +694,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 1. FP32 or FP16 have different performance impact to LLM. Recommended to test them for better prompt processing performance on your models. You need to rebuild the code after change `GGML_SYCL_F16=OFF/ON`.
 
-#### Runtime
+### Runtime
 
 | Name              | Value            | Function                                                                                                                  |
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
@@ -777,7 +787,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
   ```
 
 ### **GitHub contribution**:
-Please add the `SYCL :` prefix/tag in issues/PRs titles to help the SYCL contributors to check/address them without delay.
+Please add the `[SYCL]` prefix/tag in issues/PRs titles to help the SYCL contributors to check/address them without delay.
 
 ## TODO
 

docs/ops.md

@@ -80,7 +80,7 @@ Legend:
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ | ❌ | ❌ |
+|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |

docs/ops/WebGPU.csv

@@ -5023,20 +5023,20 @@
 "WebGPU: WebGPU","ARGMAX","type=f32,ne=[1024,12,1,1]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ARGMAX","type=f32,ne=[2000,10,1,1]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","ARGMAX","type=f32,ne=[5438,3,1,1]","support","1","yes","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[2,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,2,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,2,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,1,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=i32,ne=[10,5,4,1],nr=[2,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=i16,ne=[10,5,4,1],nr=[1,1,1,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[2,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,2,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,2,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,1,2]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=i32,ne=[10,5,4,3],nr=[2,1,1,1]","support","0","no","WebGPU"
-"WebGPU: WebGPU","REPEAT","type=i16,ne=[10,5,4,3],nr=[1,1,1,2]","support","0","no","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[2,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,2,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,2,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,1],nr=[1,1,1,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=i32,ne=[10,5,4,1],nr=[2,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=i16,ne=[10,5,4,1],nr=[1,1,1,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[2,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,2,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,2,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=f32,ne=[10,5,4,3],nr=[1,1,1,2]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=i32,ne=[10,5,4,3],nr=[2,1,1,1]","support","1","yes","WebGPU"
+"WebGPU: WebGPU","REPEAT","type=i16,ne=[10,5,4,3],nr=[1,1,1,2]","support","1","yes","WebGPU"
 "WebGPU: WebGPU","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,1,1,1],v=0","support","0","no","WebGPU"
 "WebGPU: WebGPU","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[2,1,1,1],v=0","support","0","no","WebGPU"
 "WebGPU: WebGPU","REPEAT_BACK","type=f32,ne=[8,6,4,2],nr=[1,2,1,1],v=0","support","0","no","WebGPU"

ggml/include/ggml.h

@@ -427,7 +427,8 @@ extern "C" {
         // GGML_TYPE_IQ4_NL_4_8 = 37,
         // GGML_TYPE_IQ4_NL_8_8 = 38,
         GGML_TYPE_MXFP4   = 39, // MXFP4 (1 block)
-        GGML_TYPE_COUNT   = 40,
+        GGML_TYPE_NVFP4   = 40, // NVFP4 (4 blocks, E4M3 scale)
+        GGML_TYPE_COUNT   = 41,
     };
 
     // precision
@@ -463,6 +464,7 @@ extern "C" {
         GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
         GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
         GGML_FTYPE_MOSTLY_MXFP4   = 25, // except 1d tensors
+        GGML_FTYPE_MOSTLY_NVFP4   = 26, // except 1d tensors
     };
 
     // available tensor operations:
@@ -2464,6 +2466,8 @@ extern "C" {
         bool                  lower,
         bool                  uni);
 
+    // TODO: add ggml_gated_delta_net_set_bcast() to be able to configure Q, K broadcast type: tiled vs interleaved [TAG_GGML_GDN_BCAST]
+    // ref: https://github.com/ggml-org/llama.cpp/pull/19468#discussion_r2786394306
     GGML_API struct ggml_tensor * ggml_gated_delta_net(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,

ggml/src/ggml-common.h

@@ -102,6 +102,9 @@ typedef sycl::half2 ggml_half2;
 #define QI_MXFP4 (QK_MXFP4 / (4 * QR_MXFP4))
 #define QR_MXFP4 2
 
+#define QI_NVFP4 (QK_NVFP4 / (4 * QR_NVFP4))
+#define QR_NVFP4 2
+
 #define QI5_0 (QK5_0 / (4 * QR5_0))
 #define QR5_0 2
 
@@ -194,6 +197,14 @@ typedef struct {
 } block_mxfp4;
 static_assert(sizeof(block_mxfp4) == sizeof(uint8_t) + QK_MXFP4/2, "wrong mxfp4 block size/padding");
 
+#define QK_NVFP4 64
+#define QK_NVFP4_SUB 16  // sub-block size for per-group scales
+typedef struct {
+    uint8_t d[QK_NVFP4/QK_NVFP4_SUB]; // UE4M3 scales (4 bytes, one per 16-element sub-block)
+    uint8_t qs[QK_NVFP4/2];           // packed 4-bit E2M1 values (32 bytes)
+} block_nvfp4;
+static_assert(sizeof(block_nvfp4) == sizeof(uint8_t)*(QK_NVFP4/QK_NVFP4_SUB) + QK_NVFP4/2, "wrong nvfp4 block size/padding");
+
 #define QK5_0 32
 typedef struct {
     ggml_half d;           // delta

ggml/src/ggml-cpu/arch-fallback.h

@@ -15,6 +15,7 @@
 #define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
 #define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@@ -79,6 +80,8 @@
 #define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
+// quants.c
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
@@ -108,6 +111,7 @@
 // ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
@@ -155,6 +159,7 @@
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
@@ -201,6 +206,7 @@
 #define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
 #define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x1_generic ggml_quantize_mat_q8_0_4x1
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
@@ -240,6 +246,7 @@
 #elif defined(__s390x__)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 #define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
 #define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
 #define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
@@ -302,6 +309,7 @@
 #define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
 #define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
 #define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8

ggml/src/ggml-cpu/arch/arm/quants.c

@@ -650,6 +650,90 @@ void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
     *s = sumf;
 }
 
+void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_NVFP4 == 0);
+
+    const block_nvfp4 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    // Each NVFP4 super-block (64 elements) spans 2 q8_0 blocks
+    const int nb = n / QK_NVFP4;
+
+    float sumf = 0;
+
+#if defined __ARM_NEON
+    const int8x16_t values = vld1q_s8(kvalues_mxfp4);
+    const uint8x16_t m4b = vdupq_n_u8(0x0f);
+    float32x4_t acc = vdupq_n_f32(0.0f);
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const uint8x16_t q4bits_0 = vld1q_u8(x[ib].qs);
+        const uint8x16_t q4bits_1 = vld1q_u8(x[ib].qs + 16);
+
+        const int8x16_t q4_lo_0 = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits_0, m4b));
+        const int8x16_t q4_hi_0 = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits_0, 4));
+        const int8x16_t q4_lo_1 = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits_1, m4b));
+        const int8x16_t q4_hi_1 = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits_1, 4));
+
+        const int8x16_t q8_0a = vld1q_s8(y[2*ib].qs);
+        const int8x16_t q8_0b = vld1q_s8(y[2*ib].qs + 16);
+        const int8x16_t q8_lo_0 = vcombine_s8(vget_low_s8(q8_0a), vget_low_s8(q8_0b));
+        const int8x16_t q8_hi_0 = vcombine_s8(vget_high_s8(q8_0a), vget_high_s8(q8_0b));
+
+        const int8x16_t q8_1a = vld1q_s8(y[2*ib+1].qs);
+        const int8x16_t q8_1b = vld1q_s8(y[2*ib+1].qs + 16);
+        const int8x16_t q8_lo_1 = vcombine_s8(vget_low_s8(q8_1a), vget_low_s8(q8_1b));
+        const int8x16_t q8_hi_1 = vcombine_s8(vget_high_s8(q8_1a), vget_high_s8(q8_1b));
+
+        const int32x4_t p0 = vaddq_s32(
+            ggml_vdotq_s32(vdupq_n_s32(0), q4_lo_0, q8_lo_0),
+            ggml_vdotq_s32(vdupq_n_s32(0), q4_hi_0, q8_hi_0));
+        const int32x4_t p1 = vaddq_s32(
+            ggml_vdotq_s32(vdupq_n_s32(0), q4_lo_1, q8_lo_1),
+            ggml_vdotq_s32(vdupq_n_s32(0), q4_hi_1, q8_hi_1));
+
+        const int32x4_t sums = vpaddq_s32(p0, p1);
+
+        // Decode 4 UE4M3 scales to f32 and multiply with q8 scales
+        const float dy0 = GGML_CPU_FP16_TO_FP32(y[2*ib].d);
+        const float dy1 = GGML_CPU_FP16_TO_FP32(y[2*ib+1].d);
+        const float32x4_t nvsc = {
+            ggml_ue4m3_to_fp32(x[ib].d[0]),
+            ggml_ue4m3_to_fp32(x[ib].d[1]),
+            ggml_ue4m3_to_fp32(x[ib].d[2]),
+            ggml_ue4m3_to_fp32(x[ib].d[3])
+        };
+        const float32x4_t scales = vmulq_f32(nvsc, (float32x4_t){dy0, dy0, dy1, dy1});
+
+        acc = vfmaq_f32(acc, vcvtq_f32_s32(sums), scales);
+    }
+    sumf = vaddvq_f32(acc);
+#else
+    for (int ib = 0; ib < nb; ++ib) {
+        for (int si = 0; si < 4; ++si) {
+            const float d = ggml_ue4m3_to_fp32(x[ib].d[si]);
+            const int q8b = si / 2;
+            const int q8o = (si % 2) * QK_NVFP4_SUB;
+            const float dy = GGML_CPU_FP16_TO_FP32(y[2*ib + q8b].d);
+
+            int sumi_lo = 0, sumi_hi = 0;
+            for (int j = 0; j < QK_NVFP4_SUB/2; ++j) {
+                const uint8_t qv = x[ib].qs[si*(QK_NVFP4_SUB/2) + j];
+                sumi_lo += y[2*ib + q8b].qs[q8o + j +               0] * kvalues_mxfp4[qv & 0xf];
+                sumi_hi += y[2*ib + q8b].qs[q8o + j + QK_NVFP4_SUB/2] * kvalues_mxfp4[qv >>  4];
+            }
+            sumf += dy * d * (sumi_lo + sumi_hi);
+        }
+    }
+#endif
+    *s = sumf;
+}
+
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;

ggml/src/ggml-cpu/ggml-cpu.c

@@ -270,6 +270,12 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_Q8_0,
         .nrows                    = 1,
     },
+    [GGML_TYPE_NVFP4] = {
+        .from_float               = quantize_row_nvfp4,
+        .vec_dot                  = ggml_vec_dot_nvfp4_q8_0,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+    },
     [GGML_TYPE_Q2_K] = {
         .from_float               = quantize_row_q2_K,
         .vec_dot                  = ggml_vec_dot_q2_K_q8_K,

ggml/src/ggml-cpu/ops.cpp

@@ -670,6 +670,7 @@ void ggml_compute_forward_add(
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -1119,6 +1120,7 @@ void ggml_compute_forward_add1(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -1247,6 +1249,7 @@ void ggml_compute_forward_acc(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -4334,6 +4337,7 @@ void ggml_compute_forward_out_prod(
         case GGML_TYPE_Q5_1:
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -4609,6 +4613,7 @@ void ggml_compute_forward_set(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -4831,6 +4836,7 @@ void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -5555,6 +5561,7 @@ void ggml_compute_forward_clamp(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
+        case GGML_TYPE_NVFP4:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -10436,8 +10443,8 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
 
     const float * state_in_base = (const float *)src_state->data;
 
-    const int64_t rq1 = nev1 / neq1;
-    const int64_t rk1 = nev1 / nek1;
+  //const int64_t rq1 = nev1 / neq1;
+  //const int64_t rk1 = nev1 / nek1;
     const int64_t rq3 = nev3 / neq3;
     const int64_t rk3 = nev3 / nek3;
 
@@ -10447,8 +10454,8 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
         const int64_t iv1 = ir % H; // head_index
         const int64_t iv3 = ir / H; // sequence
 
-        const int64_t iq1 = iv1 / rq1;
-        const int64_t ik1 = iv1 / rk1;
+        const int64_t iq1 = iv1 % neq1;
+        const int64_t ik1 = iv1 % nek1;
 
         const int64_t iq3 = iv3 / rq3;
         const int64_t ik3 = iv3 / rk3;
@@ -10468,7 +10475,7 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
             const float * v_d = (const float *)((const char *)src_v->data + iv3 * nbv3 + t * nbv2 + iv1 * nbv1);
 
             const float beta_val = *(const float *)((const char *)src_beta->data + iv3 * nbb3 + t * nbb2 + iv1 * nbb1);
-            const float * g_d   =  (const float *)((const char *)src_g->data    + iv3 * nbg3 + t * nbg2 + iv1 * nbg1);
+            const float * g_d    =  (const float *)((const char *)src_g->data    + iv3 * nbg3 + t * nbg2 + iv1 * nbg1);
 
             if (kda) {
                 for (int64_t i = 0; i < S_v; ++i) {
@@ -10501,7 +10508,6 @@ static void ggml_compute_forward_gated_delta_net_one_chunk(
 
             attn_data += S_v * H; // advance to next token
         }
-
     }
 }
 

ggml/src/ggml-cpu/quants.c

@@ -50,6 +50,10 @@ void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, i
     quantize_row_mxfp4_ref(x, y, k);
 }
 
+void quantize_row_nvfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_nvfp4_ref(x, y, k);
+}
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -216,6 +220,42 @@ void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     *s = sumf;
 }
 
+// NVFP4: super-block of 64 elements = 4 sub-blocks of 16 = 2 q8_0 blocks
+void ggml_vec_dot_nvfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_NVFP4 == 0);
+
+    const block_nvfp4 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+    const int nb = n / QK_NVFP4;
+
+    float sumf = 0;
+
+    for (int ib = 0; ib < nb; ++ib) {
+        for (int s_idx = 0; s_idx < 4; ++s_idx) {
+            const float d = ggml_ue4m3_to_fp32(x[ib].d[s_idx]);
+            const int q8_block = s_idx / 2;
+            const int q8_off   = (s_idx % 2) * QK_NVFP4_SUB;
+            const float dy = GGML_CPU_FP16_TO_FP32(y[2*ib + q8_block].d);
+
+            int sumi_lo = 0, sumi_hi = 0;
+            for (int j = 0; j < QK_NVFP4_SUB/2; ++j) {
+                const uint8_t qv = x[ib].qs[s_idx*(QK_NVFP4_SUB/2) + j];
+                sumi_lo += y[2*ib + q8_block].qs[q8_off + j +               0] * kvalues_mxfp4[qv & 0xf];
+                sumi_hi += y[2*ib + q8_block].qs[q8_off + j + QK_NVFP4_SUB/2] * kvalues_mxfp4[qv >>  4];
+            }
+
+            sumf += dy * d * (sumi_lo + sumi_hi);
+        }
+    }
+    *s = sumf;
+}
+
 void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;

ggml/src/ggml-cpu/quants.h

@@ -20,6 +20,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
 void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
 void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_nvfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
 void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
@@ -42,6 +43,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
 void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
 void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -73,6 +75,7 @@ void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
 void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
 void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_nvfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
 void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);

ggml/src/ggml-cuda/gated_delta_net.cu

@@ -1,36 +1,36 @@
 #include "gated_delta_net.cuh"
-#include "ggml-cuda/common.cuh"
 
 template <int S_v, bool KDA>
-__global__ void __launch_bounds__(S_v, 1)
-gated_delta_net_cuda(const float * q,
-                     const float * k,
-                     const float * v,
-                     const float * g,
-                     const float * beta,
-                     const float * curr_state,
-                     float *       dst,
-                     const int64_t H,
-                     const int64_t n_tokens,
-                     const int64_t n_seqs,
-                     const int64_t sq1,
-                     const int64_t sq2,
-                     const int64_t sq3,
-                     const int64_t sv1,
-                     const int64_t sv2,
-                     const int64_t sv3,
-                     const int64_t sb1,
-                     const int64_t sb2,
-                     const int64_t sb3,
-                     const int64_t rq1,
-                     const int64_t rq3,
-                     const float   scale) {
-    const int64_t h_idx    = blockIdx.x;
-    const int64_t sequence = blockIdx.y;
-    const int     col      = threadIdx.x;  // each thread owns one column
+__global__ void gated_delta_net_cuda(const float * q,
+                                     const float * k,
+                                     const float * v,
+                                     const float * g,
+                                     const float * beta,
+                                     const float * curr_state,
+                                     float *       dst,
+                                     int64_t       H,
+                                     int64_t       n_tokens,
+                                     int64_t       n_seqs,
+                                     int64_t       sq1,
+                                     int64_t       sq2,
+                                     int64_t       sq3,
+                                     int64_t       sv1,
+                                     int64_t       sv2,
+                                     int64_t       sv3,
+                                     int64_t       sb1,
+                                     int64_t       sb2,
+                                     int64_t       sb3,
+                                     const uint3   neqk1_magic,
+                                     const uint3   rq3_magic,
+                                     float         scale) {
+    const uint32_t h_idx    = blockIdx.x;
+    const uint32_t sequence = blockIdx.y;
+    // each warp owns one column, using warp-level primitives to reduce across rows
+    const int      lane     = threadIdx.x;
+    const int      col      = blockIdx.z * blockDim.y + threadIdx.y;
 
-    const int64_t iq1 = h_idx / rq1;
-    const int64_t iq3 = sequence / rq3;
+    const uint32_t iq1 = fastmodulo(h_idx, neqk1_magic);
+    const uint32_t iq3 = fastdiv(sequence, rq3_magic);
 
     const int64_t attn_score_elems = S_v * H * n_tokens * n_seqs;
     float *       attn_data        = dst;
@@ -41,17 +41,14 @@ gated_delta_net_cuda(const float * q,
     curr_state += state_offset;
     attn_data += (sequence * n_tokens * H + h_idx) * S_v;
 
-    // GCN and CDNA devices spill registers, we use shared mem for them. See https://github.com/ggml-org/llama.cpp/pull/20282#issuecomment-4025770229
-    // TODO: check optimal path for RDNA1 and RDNA2 devices.
-#if (defined(GGML_USE_HIP) && !defined(RDNA3) && !defined(RDNA4)) || defined(GGML_USE_MUSA)
-    extern __shared__ float s_shared[];
-    float * s = s_shared + col * S_v;
-#else
-    float s[S_v];
-#endif
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size() < S_v ? ggml_cuda_get_physical_warp_size() : S_v;
+    static_assert(S_v % warp_size == 0, "S_v must be a multiple of warp_size");
+    constexpr int rows_per_lane = (S_v + warp_size - 1) / warp_size;
+    float         s_shard[rows_per_lane];
 #pragma unroll
-    for (int i = 0; i < S_v; i++) {
-        s[i] = curr_state[i * S_v + col];
+    for (int r = 0; r < rows_per_lane; r++) {
+        const int i = r * warp_size + lane;
+        s_shard[r]  = curr_state[i * S_v + col];
     }
 
     for (int t = 0; t < n_tokens; t++) {
@@ -69,46 +66,61 @@ gated_delta_net_cuda(const float * q,
             const float g_val = expf(*g_t);
 
             // kv[col] = (S^T @ k)[col] = sum_i S[i][col] * k[i]
-            float kv_col = 0.0f;
+            float kv_shard = 0.0f;
 #pragma unroll
-            for (int i = 0; i < S_v; i++) {
-                kv_col += s[i] * k_t[i];
+            for (int r = 0; r < rows_per_lane; r++) {
+                const int i = r * warp_size + lane;
+                kv_shard += s_shard[r] * k_t[i];
             }
+            float kv_col = warp_reduce_sum<warp_size>(kv_shard);
 
             // delta[col] = (v[col] - g * kv[col]) * beta
             float delta_col = (v_t[col] - g_val * kv_col) * beta_val;
 
             // fused: S[i][col] = g * S[i][col] + k[i] * delta[col]
             // attn[col] = (S^T @ q)[col] = sum_i S[i][col] * q[i]
-            float attn_col = 0.0f;
+            float attn_partial = 0.0f;
 #pragma unroll
-            for (int i = 0; i < S_v; i++) {
-                s[i] = g_val * s[i] + k_t[i] * delta_col;
-                attn_col += s[i] * q_t[i];
+            for (int r = 0; r < rows_per_lane; r++) {
+                const int i = r * warp_size + lane;
+                s_shard[r]  = g_val * s_shard[r] + k_t[i] * delta_col;
+                attn_partial += s_shard[r] * q_t[i];
             }
 
-            attn_data[col] = attn_col * scale;
+            float attn_col = warp_reduce_sum<warp_size>(attn_partial);
+
+            if (lane == 0) {
+                attn_data[col] = attn_col * scale;
+            }
         } else {
             // kv[col] = sum_i g[i] * S[i][col] * k[i]
-            float kv_col = 0.0f;
+            float kv_shard = 0.0f;
 #pragma unroll
-            for (int i = 0; i < S_v; i++) {
-                kv_col += expf(g_t[i]) * s[i] * k_t[i];
+            for (int r = 0; r < rows_per_lane; r++) {
+                const int i = r * warp_size + lane;
+                kv_shard += expf(g_t[i]) * s_shard[r] * k_t[i];
             }
 
+            float kv_col = warp_reduce_sum<warp_size>(kv_shard);
+
             // delta[col] = (v[col] - kv[col]) * beta
             float delta_col = (v_t[col] - kv_col) * beta_val;
 
             // fused: S[i][col] = g[i] * S[i][col] + k[i] * delta[col]
             // attn[col] = (S^T @ q)[col] = sum_i S[i][col] * q[i]
-            float attn_col = 0.0f;
+            float attn_partial = 0.0f;
 #pragma unroll
-            for (int i = 0; i < S_v; i++) {
-                s[i] = expf(g_t[i]) * s[i] + k_t[i] * delta_col;
-                attn_col += s[i] * q_t[i];
+            for (int r = 0; r < rows_per_lane; r++) {
+                const int i = r * warp_size + lane;
+                s_shard[r]  = expf(g_t[i]) * s_shard[r] + k_t[i] * delta_col;
+                attn_partial += s_shard[r] * q_t[i];
             }
 
-            attn_data[col] = attn_col * scale;
+            float attn_col = warp_reduce_sum<warp_size>(attn_partial);
+
+            if (lane == 0) {
+                attn_data[col] = attn_col * scale;
+            }
         }
 
         attn_data += S_v * H;
@@ -116,8 +128,9 @@ gated_delta_net_cuda(const float * q,
 
     // Write state back to global memory
 #pragma unroll
-    for (int i = 0; i < S_v; i++) {
-        state[i * S_v + col] = s[i];
+    for (int r = 0; r < rows_per_lane; r++) {
+        const int i          = r * warp_size + lane;
+        state[i * S_v + col] = s_shard[r];
     }
 }
 
@@ -135,35 +148,43 @@ static void launch_gated_delta_net(
         const float * q_d, const float * k_d, const float * v_d,
         const float * g_d, const float * b_d, const float * s_d,
         float * dst_d,
-        int64_t S_v, int64_t H, int64_t n_tokens, int64_t n_seqs,
-        int64_t sq1, int64_t sq2, int64_t sq3,
-        int64_t sv1, int64_t sv2, int64_t sv3,
-        int64_t sb1, int64_t sb2, int64_t sb3,
-        int64_t rq1, int64_t rq3,
+        int64_t S_v,   int64_t H, int64_t n_tokens, int64_t n_seqs,
+        int64_t sq1,   int64_t sq2, int64_t sq3,
+        int64_t sv1,   int64_t sv2, int64_t sv3,
+        int64_t sb1,   int64_t sb2, int64_t sb3,
+        int64_t neqk1, int64_t rq3,
         float scale, cudaStream_t stream) {
+    //TODO: Add chunked kernel for even faster pre-fill
+    const int warp_size = ggml_cuda_info().devices[ggml_cuda_get_device()].warp_size;
+    const int num_warps = 4;
+    dim3      grid_dims(H, n_seqs, (S_v + num_warps - 1) / num_warps);
+    dim3      block_dims(warp_size <= S_v ? warp_size : S_v, num_warps, 1);
 
-    dim3 grid_dims(H, n_seqs, 1);
-    dim3 block_dims(S_v, 1, 1);
+    const uint3 neqk1_magic = init_fastdiv_values(neqk1);
+    const uint3 rq3_magic   = init_fastdiv_values(rq3);
 
     int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
 
     switch (S_v) {
-        case 32: {
-            constexpr int sv = 32;
-            size_t smem = calculate_smem(sv, cc);
-            gated_delta_net_cuda<sv, KDA><<<grid_dims, block_dims, smem, stream>>>(
+        case 16:
+            gated_delta_net_cuda<16, KDA><<<grid_dims, block_dims, 0, stream>>>(
                 q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
                 n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
-                sb1, sb2, sb3, rq1, rq3, scale);
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
+            break;
+        case 32:
+            gated_delta_net_cuda<32, KDA><<<grid_dims, block_dims, 0, stream>>>(
+                q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
+                n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
             break;
-        }
         case 64: {
             constexpr int sv = 64;
             size_t smem = calculate_smem(sv, cc);
             gated_delta_net_cuda<sv, KDA><<<grid_dims, block_dims, smem, stream>>>(
                 q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
                 n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
-                sb1, sb2, sb3, rq1, rq3, scale);
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
             break;
         }
         case 128: {
@@ -172,7 +193,7 @@ static void launch_gated_delta_net(
             gated_delta_net_cuda<sv, KDA><<<grid_dims, block_dims, smem, stream>>>(
                 q_d, k_d, v_d, g_d, b_d, s_d, dst_d, H,
                 n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
-                sb1, sb2, sb3, rq1, rq3, scale);
+                sb1, sb2, sb3, neqk1_magic, rq3_magic, scale);
             break;
         }
         default:
@@ -190,10 +211,12 @@ void ggml_cuda_op_gated_delta_net(ggml_backend_cuda_context & ctx, ggml_tensor *
     ggml_tensor * src_state = dst->src[5];
 
     GGML_TENSOR_LOCALS(int64_t, neq, src_q, ne);
-    GGML_TENSOR_LOCALS(size_t, nbq, src_q, nb);
+    GGML_TENSOR_LOCALS(size_t , nbq, src_q, nb);
+    GGML_TENSOR_LOCALS(int64_t, nek, src_k, ne);
+    GGML_TENSOR_LOCALS(size_t , nbk, src_k, nb);
     GGML_TENSOR_LOCALS(int64_t, nev, src_v, ne);
-    GGML_TENSOR_LOCALS(size_t, nbv, src_v, nb);
-    GGML_TENSOR_LOCALS(size_t, nbb, src_beta, nb);
+    GGML_TENSOR_LOCALS(size_t,  nbv, src_v, nb);
+    GGML_TENSOR_LOCALS(size_t,  nbb, src_beta, nb);
 
     const int64_t S_v      = nev0;
     const int64_t H        = nev1;
@@ -202,7 +225,9 @@ void ggml_cuda_op_gated_delta_net(ggml_backend_cuda_context & ctx, ggml_tensor *
 
     const bool kda = (src_g->ne[0] == S_v);
 
-    const int64_t rq1 = nev1 / neq1;
+    GGML_ASSERT(neq1 == nek1);
+    const int64_t neqk1 = neq1;
+
     const int64_t rq3 = nev3 / neq3;
 
     const float * q_d = (const float *) src_q->data;
@@ -241,10 +266,10 @@ void ggml_cuda_op_gated_delta_net(ggml_backend_cuda_context & ctx, ggml_tensor *
     if (kda) {
         launch_gated_delta_net<true>(q_d, k_d, v_d, g_d, b_d, s_d, dst_d,
             S_v, H, n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
-            sb1, sb2, sb3, rq1, rq3, scale, stream);
+            sb1, sb2, sb3, neqk1, rq3, scale, stream);
     } else {
         launch_gated_delta_net<false>(q_d, k_d, v_d, g_d, b_d, s_d, dst_d,
             S_v, H, n_tokens, n_seqs, sq1, sq2, sq3, sv1, sv2, sv3,
-            sb1, sb2, sb3, rq1, rq3, scale, stream);
+            sb1, sb2, sb3, neqk1, rq3, scale, stream);
     }
 }

ggml/src/ggml-hip/CMakeLists.txt

@@ -11,6 +11,10 @@ endif()
 list(APPEND CMAKE_PREFIX_PATH  ${ROCM_PATH})
 list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
 
+if (NOT DEFINED CMAKE_HIP_FLAGS_DEBUG)
+    set(CMAKE_HIP_FLAGS_DEBUG "-g -O2")
+endif()
+
 # CMake on Windows doesn't support the HIP language yet
 if (WIN32)
     set(CXX_IS_HIPCC TRUE)

ggml/src/ggml-impl.h

@@ -491,6 +491,61 @@ static inline float ggml_e8m0_to_fp32_half(uint8_t x) {
 #define GGML_E8M0_TO_FP32(x) ggml_e8m0_to_fp32(x)
 #define GGML_E8M0_TO_FP32_HALF(x) ggml_e8m0_to_fp32_half(x)
 
+// UE4M3: unsigned, 4 exp bits (bias=7), 3 mantissa bits
+// Returns value * 0.5 to match kvalues_mxfp4 convention (kvalues = 2 * E2M1_float)
+static inline float ggml_ue4m3_to_fp32(uint8_t x) {
+    if (x == 0 || x == 0x7F) {
+        return 0.0f;
+    }
+    int   exp = (x >> 3) & 0xF;
+    int   man = x & 0x7;
+    float raw;
+    if (exp == 0) {
+        raw = ldexpf((float) man, -9);
+    } else {
+        raw = ldexpf(1.0f + (float) man / 8.0f, exp - 7);
+    }
+    return raw * 0.5f;
+}
+
+static inline uint8_t ggml_fp32_to_ue4m3(float x) {
+    if (!(x > 0.0f)) {
+        return 0;
+    }
+    if (x > 448.0f) {
+        x = 448.0f;
+    }
+    uint32_t bits;
+    memcpy(&bits, &x, 4);
+    int fp32_exp  = ((bits >> 23) & 0xFF) - 127;
+    int fp32_man  = (bits >> 20) & 0x7;
+    int ue4m3_exp = fp32_exp + 7;
+    if (ue4m3_exp <= 0) {
+        // subnormal: value = man * 2^-9, man = round(x * 2^9)
+        int man = (int) (x * 512.0f + 0.5f);
+        if (man > 7) {
+            man = 7;
+        }
+        if (man < 1) {
+            return 0;
+        }
+        return (uint8_t) man;
+    }
+    if (ue4m3_exp >= 15) {
+        return 0x7E;
+    }
+    int round_bit = (bits >> 19) & 1;
+    int ue4m3_man = fp32_man + round_bit;
+    if (ue4m3_man > 7) {
+        ue4m3_man = 0;
+        ue4m3_exp++;
+        if (ue4m3_exp >= 15) {
+            return 0x7E;
+        }
+    }
+    return (uint8_t) ((ue4m3_exp << 3) | ue4m3_man);
+}
+
 /**
  * Converts brain16 to float32.
  *

ggml/src/ggml-metal/ggml-metal-context.m

@@ -47,7 +47,7 @@ struct ggml_metal {
     uint64_t fuse_cnt[GGML_OP_COUNT];
 
     // capture state
-    bool capture_next_compute;
+    int capture_compute;
     bool capture_started;
 
     id<MTLCaptureScope> capture_scope;
@@ -158,10 +158,17 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
     GGML_LOG_INFO("%s: use concurrency    = %s\n", __func__, res->use_concurrency    ? "true" : "false");
     GGML_LOG_INFO("%s: use graph optimize = %s\n", __func__, res->use_graph_optimize ? "true" : "false");
 
-    res->capture_next_compute = false;
+    res->capture_compute = 0;
     res->capture_started = false;
     res->capture_scope = nil;
 
+    {
+        const char * val = getenv("GGML_METAL_CAPTURE_COMPUTE");
+        if (val) {
+            res->capture_compute = atoi(val);
+        }
+    }
+
     res->has_error = false;
 
     res->gf = nil;
@@ -458,9 +465,13 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
 
         ctx->n_nodes_per_cb = (ctx->n_nodes_1 + ctx->n_cb - 1) / ctx->n_cb;
 
-        const bool use_capture = ctx->capture_next_compute;
+        if (ctx->capture_compute >= 0) {
+            ctx->capture_compute--;
+        }
+
+        const bool use_capture = ctx->capture_compute == 0;
         if (use_capture) {
-            ctx->capture_next_compute = false;
+            ctx->capture_compute = -1;
 
             // make sure all previous computations have finished before starting the capture
             if (ctx->cmd_buf_last) {
@@ -469,6 +480,10 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
             }
 
             if (!ctx->capture_started) {
+                NSString * path = [NSString stringWithFormat:@"/tmp/perf-metal-%d.gputrace", getpid()];
+
+                GGML_LOG_WARN("%s: capturing graph in %s\n", __func__, [path UTF8String]);
+
                 // create capture scope
                 id<MTLDevice> device = ggml_metal_device_get_obj(ctx->dev);
                 ctx->capture_scope = [[MTLCaptureManager sharedCaptureManager] newCaptureScopeWithDevice:device];
@@ -476,7 +491,7 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
                 MTLCaptureDescriptor * descriptor = [MTLCaptureDescriptor new];
                 descriptor.captureObject = ctx->capture_scope;
                 descriptor.destination = MTLCaptureDestinationGPUTraceDocument;
-                descriptor.outputURL = [NSURL fileURLWithPath:[NSString stringWithFormat:@"/tmp/perf-metal.gputrace"]];
+                descriptor.outputURL = [NSURL fileURLWithPath:path];
 
                 NSError * error = nil;
                 if (![[MTLCaptureManager sharedCaptureManager] startCaptureWithDescriptor:descriptor error:&error]) {
@@ -683,7 +698,7 @@ void ggml_metal_set_n_cb(ggml_metal_t ctx, int n_cb) {
             idx_end,
             ctx->use_fusion,
             ctx->use_concurrency,
-            ctx->capture_next_compute,
+            ctx->capture_compute,
             ctx->debug_graph,
             ctx->debug_fusion);
 
@@ -718,5 +733,5 @@ bool ggml_metal_supports_family(ggml_metal_t ctx, int family) {
 }
 
 void ggml_metal_capture_next_compute(ggml_metal_t ctx) {
-    ctx->capture_next_compute = true;
+    ctx->capture_compute = 1;
 }

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

@@ -577,6 +577,41 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv(ggml_metal_
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_gated_delta_net(ggml_metal_library_t lib, const ggml_tensor * op) {
+    char base[256];
+    char name[256];
+
+    // v is src[2], dimensions: S_v = ne[0], H = ne[1]
+    const int ne20 = op->src[2]->ne[0]; // S_v
+    const int ne21 = op->src[2]->ne[1]; // H
+    const int ne30 = op->src[3]->ne[0]; // G
+
+    const int nsg = op->src[2]->ne[0]/32;
+
+    GGML_ASSERT(op->src[5]->type == GGML_TYPE_F32);
+    GGML_ASSERT(op->ne[0] == ne20 * ne21);
+    GGML_ASSERT(ne20 % 32 == 0);
+
+    snprintf(base, 256, "kernel_gated_delta_net_%s_%d", ggml_type_name(op->src[0]->type), nsg);
+    snprintf(name, 256, "%s_ne20=%d_ne30=%d", base, ne20, ne30);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        ggml_metal_cv_t cv = ggml_metal_cv_init();
+
+        ggml_metal_cv_set_int16(cv, ne20, FC_GATED_DELTA_NET + 0);
+        ggml_metal_cv_set_int16(cv, ne30, FC_GATED_DELTA_NET + 1);
+
+        res = ggml_metal_library_compile_pipeline(lib, base, name, cv);
+
+        ggml_metal_cv_free(cv);
+    }
+
+    res.nsg = nsg;
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri(ggml_metal_library_t lib, const ggml_tensor * op) {
     char base[256];
     char name[256];

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

@@ -125,6 +125,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv
 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);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rwkv              (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_gated_delta_net   (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_solve_tri         (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_ext        (ggml_metal_library_t lib, enum ggml_type tsrc0, enum ggml_type tsrc1, int nsg, int nxpsg, int r1ptg);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm            (ggml_metal_library_t lib, const struct ggml_tensor * op);

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

@@ -1155,10 +1155,12 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:
             return true;
+        case GGML_OP_GATED_DELTA_NET:
+            return op->src[2]->ne[0] % 32 == 0;
         case GGML_OP_SOLVE_TRI:
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
-            return has_simdgroup_reduction;
+            return has_simdgroup_reduction && op->src[0]->type != GGML_TYPE_NVFP4;
         case GGML_OP_SET:
         case GGML_OP_CPY:
         case GGML_OP_DUP:
@@ -1216,7 +1218,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 };
             }
         case GGML_OP_GET_ROWS:
-            return true;
+            return op->src[0]->type != GGML_TYPE_NVFP4;
         case GGML_OP_SET_ROWS:
             {
                 if (op->src[0]->type != GGML_TYPE_F32) {

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

@@ -35,7 +35,7 @@
 #define N_R0_Q4_K 2
 #define N_SG_Q4_K 2
 
-#define N_R0_Q5_K 2
+#define N_R0_Q5_K 1
 #define N_SG_Q5_K 2
 
 #define N_R0_Q6_K 2
@@ -84,6 +84,7 @@
 #define FC_BIN                         1300
 #define FC_SUM_ROWS                    1400
 #define FC_UPSCALE                     1500
+#define FC_GATED_DELTA_NET             1600
 
 // op-specific constants
 #define OP_FLASH_ATTN_EXT_NQPSG 8
@@ -793,6 +794,44 @@ typedef struct {
     uint64_t nb0;
 } ggml_metal_kargs_ssm_scan;
 
+typedef struct {
+    int32_t  ne00;
+    int32_t  ne01;
+    int32_t  ne02;
+    int32_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne10;
+    int32_t  ne11;
+    int32_t  ne12;
+    int32_t  ne13;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb13;
+    int32_t  ne20;
+    int32_t  ne21;
+    int32_t  ne22;
+    int32_t  ne23;
+    uint64_t nb20;
+    uint64_t nb21;
+    uint64_t nb22;
+    uint64_t nb23;
+    int32_t  ns02;
+    int32_t  ns12;
+    int32_t  ns22;
+    int32_t  ne0;
+    int32_t  ne1;
+    int32_t  ne2;
+    int32_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_gated_delta_net;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;

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

@@ -333,6 +333,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_rwkv(ctx, idx);
             } break;
+        case GGML_OP_GATED_DELTA_NET:
+            {
+                n_fuse = ggml_metal_op_gated_delta_net(ctx, idx);
+            } break;
         case GGML_OP_SOLVE_TRI:
             {
                 n_fuse = ggml_metal_op_solve_tri(ctx, idx);
@@ -1562,6 +1566,81 @@ int ggml_metal_op_rwkv(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_gated_delta_net(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;
+
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne1, op->src[1], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb1, op->src[1], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne2, op->src[2], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb2, op->src[2], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    auto pipeline = ggml_metal_library_get_pipeline_gated_delta_net(lib, op);
+
+    int ida = 0;
+
+    ggml_metal_kargs_gated_delta_net args = {
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.ne03 =*/ ne03,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.ne10 =*/ ne10,
+        /*.ne11 =*/ ne11,
+        /*.ne12 =*/ ne12,
+        /*.ne13 =*/ ne13,
+        /*.nb10 =*/ nb10,
+        /*.nb11 =*/ nb11,
+        /*.nb12 =*/ nb12,
+        /*.nb13 =*/ nb13,
+        /*.ne20 =*/ ne20,
+        /*.ne21 =*/ ne21,
+        /*.ne22 =*/ ne22,
+        /*.ne23 =*/ ne23,
+        /*.nb20 =*/ nb20,
+        /*.nb21 =*/ nb21,
+        /*.nb22 =*/ nb22,
+        /*.nb23 =*/ nb23,
+        /*.ns02 =*/ (int32_t) (nb02/sizeof(float)),
+        /*.ns12 =*/ (int32_t) (nb12/sizeof(float)),
+        /*.ns22 =*/ (int32_t) (nb22/sizeof(float)),
+        /*.ne0  =*/ ne0,
+        /*.ne1  =*/ ne1,
+        /*.ne2  =*/ ne2,
+        /*.ne3  =*/ ne3,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+    };
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args),                  ida++);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), ida++); // q
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[1]), ida++); // k
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[2]), ida++); // v
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[3]), ida++); // gate
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[4]), ida++); // beta
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[5]), ida++); // state
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         ida++); // dst
+
+    const int nsg = pipeline.nsg;
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, op->src[2]->ne[0]/nsg, op->src[2]->ne[1], op->src[2]->ne[3], 32, nsg, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_solve_tri(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 

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

@@ -58,6 +58,7 @@ 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);
 int ggml_metal_op_rwkv              (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_gated_delta_net   (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_solve_tri         (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_set               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_cpy               (ggml_metal_op_t ctx, int idx);

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

@@ -2434,6 +2434,227 @@ kernel void kernel_rwkv_wkv7_f32(
     }
 }
 
+constant short FC_gated_delta_net_ne20 [[function_constant(FC_GATED_DELTA_NET + 0)]];
+constant short FC_gated_delta_net_ne30 [[function_constant(FC_GATED_DELTA_NET + 1)]];
+
+#if 1
+template<short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B
+    const uint i21 = tgpig.y; // H
+    const uint i20 = tgpig.x*NSG + ty;
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    device const float * s_ptr = (device const float *) (s) + (i23*args.ne21 + i21)*S_v*S_v + i20;
+
+    float ls[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        ls[j] = s_ptr[is*S_v];
+    }
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    for (short t = 0; t < args.ne22; t++) {
+        float s_k = 0.0f;
+
+        if (G == 1) {
+            const float g_exp = exp(g_ptr[0]);
+
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= g_exp;
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        } else {
+            // KDA
+            FOR_UNROLL (short j = 0; j < NSG; j++) {
+                const short is = tx*NSG + j;
+                ls[j] *= exp(g_ptr[is]);
+
+                s_k += ls[j]*k_ptr[is];
+            }
+        }
+
+        s_k = simd_sum(s_k);
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        float y = 0.0f;
+
+        FOR_UNROLL (short j = 0; j < NSG; j++) {
+            const short is = tx*NSG + j;
+            ls[j] += k_ptr[is]*d;
+
+            y += ls[j]*q_ptr[is];
+        }
+
+        y = simd_sum(y);
+
+        if (tx == 0) {
+            dst_attn[t*args.ne21*S_v] = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+    }
+
+    device float * dst_state = (device float *) (dst) + args.ne23*args.ne22*args.ne21*S_v + (i23*args.ne21 + i21)*S_v*S_v + i20;
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        dst_state[is*S_v] = ls[j];
+    }
+
+#undef S_v
+#undef G
+}
+
+typedef decltype(kernel_gated_delta_net_impl<4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<4>;
+
+#else
+// a simplified version of the above
+// no performance improvement, so keep the above version for now
+
+template<typename T, short NSG>
+kernel void kernel_gated_delta_net_impl(
+        constant ggml_metal_kargs_gated_delta_net & args,
+        device const char * q,
+        device const char * k,
+        device const char * v,
+        device const char * g,
+        device const char * b,
+        device const char * s,
+        device       char * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]])  {
+#define S_v FC_gated_delta_net_ne20
+#define G   FC_gated_delta_net_ne30
+
+    const uint tx = tpitg.x;
+    const uint ty = tpitg.y;
+
+    const uint i23 = tgpig.z; // B
+    const uint i21 = tgpig.y; // H
+    const uint i20 = tgpig.x*NSG + ty;
+
+    const uint i01 = i21 % args.ne01;
+    const uint i11 = i21 % args.ne11;
+
+    const float scale = 1.0f / sqrt((float)S_v);
+
+    device const float * s_ptr = (device const float *) (s) + (i23*args.ne21 + i21)*S_v*S_v + i20;
+
+    float lsf[NSG];
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        lsf[j] = s_ptr[is*S_v];
+    }
+
+    thread T * ls = (thread T *) (lsf);
+
+    device float * dst_attn = (device float *) (dst) + (i23*args.ne22*args.ne21 + i21)*S_v + i20;
+
+    device const float * q_ptr = (device const float *) (q + i23*args.nb03 + i01*args.nb01);
+    device const float * k_ptr = (device const float *) (k + i23*args.nb13 + i11*args.nb11);
+    device const float * v_ptr = (device const float *) (v + i23*args.nb23 + i21*args.nb21);
+
+    device const float * b_ptr  = (device const float *) (b) + (i23*args.ne22*args.ne21 + i21);
+    device const float * g_ptr  = (device const float *) (g) + (i23*args.ne22*args.ne21 + i21)*G;
+
+    for (short t = 0; t < args.ne22; t++) {
+        device const T * qt_ptr = (device const T *) (q_ptr);
+        device const T * kt_ptr = (device const T *) (k_ptr);
+        device const T * gt_ptr = (device const T *) (g_ptr);
+
+        if (G == 1) {
+            *ls *= exp(g_ptr[0]);
+        } else {
+            // KDA
+            *ls *= exp(gt_ptr[tx]);
+        }
+
+        const float s_k = simd_sum(dot(*ls, kt_ptr[tx]));
+
+        const float d = (v_ptr[i20] - s_k)*b_ptr[0];
+
+        *ls += kt_ptr[tx]*d;
+
+        const float y = simd_sum(dot(*ls, qt_ptr[tx]));
+
+        if (tx == 0) {
+            *dst_attn = y*scale;
+        }
+
+        q_ptr += args.ns02;
+        k_ptr += args.ns12;
+        v_ptr += args.ns22;
+
+        b_ptr += args.ne21;
+        g_ptr += args.ne21*G;
+
+        dst_attn += args.ne21*S_v;
+    }
+
+    device float * dst_state  = (device float *) (dst) + args.ne23*args.ne22*args.ne21*S_v + (i23*args.ne21 + i21)*S_v*S_v + i20;
+    device T     * dstt_state = (device T     *) (dst_state);
+
+    FOR_UNROLL (short j = 0; j < NSG; j++) {
+        const short is = tx*NSG + j;
+        dst_state[is*S_v] = lsf[j];
+    }
+
+#undef S_v
+#undef G
+}
+
+typedef decltype(kernel_gated_delta_net_impl<float4, 4>) kernel_gated_delta_net_t;
+
+template [[host_name("kernel_gated_delta_net_f32_1")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float,  1>;
+template [[host_name("kernel_gated_delta_net_f32_2")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float2, 2>;
+template [[host_name("kernel_gated_delta_net_f32_4")]] kernel kernel_gated_delta_net_t kernel_gated_delta_net_impl<float4, 4>;
+#endif
+
 constant short FC_solve_tri_nsg [[function_constant(FC_SOLVE_TRI + 0)]];
 constant short FC_solve_tri_n   [[function_constant(FC_SOLVE_TRI + 1)]];
 constant short FC_solve_tri_k   [[function_constant(FC_SOLVE_TRI + 2)]];
@@ -9081,6 +9302,7 @@ template [[host_name("kernel_mul_mm_id_map0_ne20_6" )]] kernel kernel_mul_mm_id_
 template [[host_name("kernel_mul_mm_id_map0_ne20_8" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<8>;
 template [[host_name("kernel_mul_mm_id_map0_ne20_10")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<10>;
 template [[host_name("kernel_mul_mm_id_map0_ne20_16")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<16>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_22")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<22>;
 
 template<typename S0, typename S0_4x4, typename S0_8x8, typename S1, typename S1_2x4, typename S1_8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread S0_4x4 &), typename T0, typename T0_4x4, typename T1, typename T1_2x4>
 kernel void kernel_mul_mm_id(

ggml/src/ggml-quants.c

@@ -304,6 +304,41 @@ void quantize_row_mxfp4_ref(const float * GGML_RESTRICT x, block_mxfp4 * GGML_RE
     }
 }
 
+void quantize_row_nvfp4_ref(const float * GGML_RESTRICT x, block_nvfp4 * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK_NVFP4;
+    static const int qk_sub = QK_NVFP4_SUB;
+    static const int n_sub = QK_NVFP4 / QK_NVFP4_SUB;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        for (int s = 0; s < n_sub; s++) {
+            const float * xb = x + i*qk + s*qk_sub;
+
+            float amax = 0.0f;
+            for (int j = 0; j < qk_sub; j++) {
+                if (amax < fabsf(xb[j])) {
+                    amax = fabsf(xb[j]);
+                }
+            }
+
+            // UE4M3 scale: amax / 6.0 maps the max E2M1 value (6.0) to amax
+            const uint8_t ue = ggml_fp32_to_ue4m3(amax / 6.0f);
+            y[i].d[s] = ue;
+            const float d = ggml_ue4m3_to_fp32(ue);
+
+            for (int j = 0; j < qk_sub/2; ++j) {
+                const uint8_t x0 = best_index_mxfp4(xb[0        + j], d);
+                const uint8_t x1 = best_index_mxfp4(xb[qk_sub/2 + j], d);
+
+                y[i].qs[s*(qk_sub/2) + j] = x0 | (x1 << 4);
+            }
+        }
+    }
+}
+
 void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
     static const int qk = QK4_0;
 
@@ -434,6 +469,31 @@ void dequantize_row_mxfp4(const block_mxfp4 * GGML_RESTRICT x, float * GGML_REST
     }
 }
 
+void dequantize_row_nvfp4(const block_nvfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+    static const int qk = QK_NVFP4;
+    static const int qk_sub = QK_NVFP4_SUB;
+    static const int n_sub = QK_NVFP4 / QK_NVFP4_SUB;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        for (int s = 0; s < n_sub; s++) {
+            const float d = ggml_ue4m3_to_fp32(x[i].d[s]);
+            float * yb = y + i*qk + s*qk_sub;
+
+            for (int j = 0; j < qk_sub/2; ++j) {
+                const int8_t v0 = kvalues_mxfp4[x[i].qs[s*(qk_sub/2) + j] & 0x0F];
+                const int8_t v1 = kvalues_mxfp4[x[i].qs[s*(qk_sub/2) + j] >>   4];
+
+                yb[j + 0       ] = v0*d;
+                yb[j + qk_sub/2] = v1*d;
+            }
+        }
+    }
+}
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -2098,6 +2158,12 @@ size_t quantize_mxfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst,
     return nrow * ggml_row_size(GGML_TYPE_MXFP4, n_per_row);
 }
 
+size_t quantize_nvfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_UNUSED(quant_weights);
+    quantize_row_nvfp4_ref(src, dst, (int64_t)nrow*n_per_row);
+    return nrow * ggml_row_size(GGML_TYPE_NVFP4, n_per_row);
+}
+
 // ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs)
 
 void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k) {
@@ -5244,6 +5310,12 @@ bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbyte
             {
                 VALIDATE_ROW_DATA_E_E8M0_IMPL(block_mxfp4, data, nb);
             } break;
+        case GGML_TYPE_NVFP4:
+            {
+                // UE4M3 scales are uint8_t — all byte values are valid
+                GGML_UNUSED(data);
+                GGML_UNUSED(nb);
+            } break;
         case GGML_TYPE_Q2_K:
             {
                 VALIDATE_ROW_DATA_DM_F16_IMPL(block_q2_K, data, nb, d, dmin);

ggml/src/ggml-quants.h

@@ -22,6 +22,7 @@ GGML_API void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 *
 GGML_API void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
 
 GGML_API void quantize_row_mxfp4_ref(const float * GGML_RESTRICT x, block_mxfp4 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_nvfp4_ref(const float * GGML_RESTRICT x, block_nvfp4 * GGML_RESTRICT y, int64_t k);
 
 GGML_API void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
 GGML_API void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
@@ -48,6 +49,7 @@ GGML_API void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GG
 //GGML_API void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
 GGML_API void dequantize_row_mxfp4(const block_mxfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_nvfp4(const block_nvfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 
 GGML_API void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
 GGML_API void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
@@ -95,6 +97,7 @@ GGML_API size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTR
 GGML_API size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
 GGML_API size_t quantize_mxfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_nvfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
 
 GGML_API void iq2xs_init_impl(enum ggml_type type);
 GGML_API void iq2xs_free_impl(enum ggml_type type);

ggml/src/ggml-webgpu/ggml-webgpu-shader-lib.hpp

@@ -198,6 +198,22 @@ struct ggml_webgpu_concat_pipeline_key_hash {
     }
 };
 
+/** Repeat **/
+
+struct ggml_webgpu_repeat_pipeline_key {
+    int type;
+
+    bool operator==(const ggml_webgpu_repeat_pipeline_key & other) const { return type == other.type; }
+};
+
+struct ggml_webgpu_repeat_pipeline_key_hash {
+    size_t operator()(const ggml_webgpu_repeat_pipeline_key & key) const {
+        size_t seed = 0;
+        ggml_webgpu_hash_combine(seed, key.type);
+        return seed;
+    }
+};
+
 /** Binary **/
 
 struct ggml_webgpu_binary_pipeline_key {
@@ -431,6 +447,8 @@ class ggml_webgpu_shader_lib {
         binary_pipelines;                                              // type/op/inplace/overlap
     std::unordered_map<ggml_webgpu_concat_pipeline_key, webgpu_pipeline, ggml_webgpu_concat_pipeline_key_hash>
         concat_pipelines;                                              // type
+    std::unordered_map<ggml_webgpu_repeat_pipeline_key, webgpu_pipeline, ggml_webgpu_repeat_pipeline_key_hash>
+        repeat_pipelines;                                              // type
     std::unordered_map<ggml_webgpu_flash_attn_pipeline_key, webgpu_pipeline, ggml_webgpu_flash_attn_pipeline_key_hash>
         flash_attn_pipelines;
     std::unordered_map<ggml_webgpu_legacy_mul_mat_pipeline_key,
@@ -1147,7 +1165,7 @@ class ggml_webgpu_shader_lib {
         }
 
         std::vector<std::string> defines;
-        std::string variant = "concat";
+        std::string              variant = "concat";
 
         switch (key.type) {
             case GGML_TYPE_F32:
@@ -1164,15 +1182,56 @@ class ggml_webgpu_shader_lib {
 
         defines.push_back(std::string("WG_SIZE=") + std::to_string(context.max_wg_size));
 
-        auto processed = preprocessor.preprocess(wgsl_concat, defines);
-        auto decisions = std::make_shared<ggml_webgpu_generic_shader_decisions>();
-        decisions->wg_size = context.max_wg_size;
+        auto processed           = preprocessor.preprocess(wgsl_concat, defines);
+        auto decisions           = std::make_shared<ggml_webgpu_generic_shader_decisions>();
+        decisions->wg_size       = context.max_wg_size;
         webgpu_pipeline pipeline = ggml_webgpu_create_pipeline(device, processed, variant);
-        pipeline.context = decisions;
-        concat_pipelines[key] = pipeline;
+        pipeline.context         = decisions;
+        concat_pipelines[key]    = pipeline;
         return concat_pipelines[key];
     }
 
+    webgpu_pipeline get_repeat_pipeline(const ggml_webgpu_shader_lib_context & context) {
+        ggml_webgpu_repeat_pipeline_key key = {
+            .type = context.dst->type,
+        };
+
+        auto it = repeat_pipelines.find(key);
+        if (it != repeat_pipelines.end()) {
+            return it->second;
+        }
+
+        std::vector<std::string> defines;
+        std::string              variant = "repeat";
+
+        switch (key.type) {
+            case GGML_TYPE_F32:
+                defines.push_back("TYPE_F32");
+                variant += "_f32";
+                break;
+            case GGML_TYPE_I32:
+                defines.push_back("TYPE_I32");
+                variant += "_i32";
+                break;
+            case GGML_TYPE_I16:
+                defines.push_back("TYPE_I16");
+                variant += "_i16";
+                break;
+            default:
+                GGML_ABORT("Unsupported type for repeat shader");
+        }
+
+        defines.push_back(std::string("WG_SIZE=") + std::to_string(context.max_wg_size));
+
+        auto processed           = preprocessor.preprocess(wgsl_repeat, defines);
+        auto decisions           = std::make_shared<ggml_webgpu_generic_shader_decisions>();
+        decisions->wg_size       = context.max_wg_size;
+        webgpu_pipeline pipeline = ggml_webgpu_create_pipeline(device, processed, variant);
+        pipeline.context         = decisions;
+        repeat_pipelines[key]    = pipeline;
+        return repeat_pipelines[key];
+    }
+
     webgpu_pipeline get_flash_attn_pipeline(const ggml_webgpu_shader_lib_context & context) {
         const bool has_mask  = context.src3 != nullptr;
         const bool has_sinks = context.src4 != nullptr;

ggml/src/ggml-webgpu/ggml-webgpu.cpp

@@ -1567,6 +1567,48 @@ static webgpu_command ggml_webgpu_concat(webgpu_context & ctx,
     return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
 }
 
+static webgpu_command ggml_webgpu_repeat(webgpu_context & ctx, ggml_tensor * src0, ggml_tensor * dst) {
+    uint32_t ne = (uint32_t) ggml_nelements(dst);
+
+    std::vector<uint32_t> params = { ne,
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, src0) /
+                                                 ggml_type_size(src0->type)),
+                                     (uint32_t) (ggml_webgpu_tensor_misalignment(ctx, dst) / ggml_type_size(dst->type)),
+                                     (uint32_t) (src0->nb[0] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src0->nb[1] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src0->nb[2] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src0->nb[3] / ggml_type_size(src0->type)),
+                                     (uint32_t) (src0->ne[0]),
+                                     (uint32_t) (src0->ne[1]),
+                                     (uint32_t) (src0->ne[2]),
+                                     (uint32_t) (src0->ne[3]),
+                                     (uint32_t) (dst->ne[0]),
+                                     (uint32_t) (dst->ne[1]),
+                                     (uint32_t) (dst->ne[2]) };
+
+    std::vector<wgpu::BindGroupEntry> entries = {
+        { .binding = 0,
+         .buffer  = ggml_webgpu_tensor_buf(src0),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, src0),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, src0) },
+        { .binding = 1,
+         .buffer  = ggml_webgpu_tensor_buf(dst),
+         .offset  = ggml_webgpu_tensor_align_offset(ctx, dst),
+         .size    = ggml_webgpu_tensor_binding_size(ctx, dst)  }
+    };
+
+    ggml_webgpu_shader_lib_context shader_lib_ctx = {
+        .src0        = src0,
+        .dst         = dst,
+        .max_wg_size = ctx->global_ctx->capabilities.limits.maxComputeInvocationsPerWorkgroup,
+    };
+
+    webgpu_pipeline pipeline  = ctx->shader_lib->get_repeat_pipeline(shader_lib_ctx);
+    auto *          decisions = static_cast<ggml_webgpu_generic_shader_decisions *>(pipeline.context.get());
+    uint32_t        wg_x      = CEIL_DIV(ne, decisions->wg_size);
+    return ggml_backend_webgpu_build(ctx->global_ctx, ctx->param_buf_pool, pipeline, params, entries, wg_x);
+}
+
 static webgpu_command ggml_webgpu_rms_norm(webgpu_context & ctx, ggml_tensor * src, ggml_tensor * dst) {
     int inplace = ggml_webgpu_tensor_equal(src, dst);
 
@@ -2158,6 +2200,8 @@ static std::optional<webgpu_command> ggml_webgpu_encode_node(webgpu_context ctx,
             return ggml_webgpu_binary_op(ctx, src0, src1, node);
         case GGML_OP_CONCAT:
             return ggml_webgpu_concat(ctx, src0, src1, node);
+        case GGML_OP_REPEAT:
+            return ggml_webgpu_repeat(ctx, src0, node);
         case GGML_OP_RMS_NORM:
             return ggml_webgpu_rms_norm(ctx, src0, node);
         case GGML_OP_ROPE:
@@ -2919,10 +2963,10 @@ static ggml_backend_buffer_type_t ggml_backend_webgpu_device_get_buffer_type(ggm
         /* .iface = */ {
                         /* .get_name         = */ ggml_backend_webgpu_buffer_type_get_name,
                         /* .alloc_buffer     = */
-            ggml_backend_webgpu_buffer_type_alloc_buffer,  /* .get_alignment    = */
-            ggml_backend_webgpu_buffer_type_get_alignment, /* .get_max_size     = */
-            ggml_backend_webgpu_buffer_type_get_max_size,  /* .get_alloc_size   = */
-            ggml_backend_webgpu_buffer_type_get_alloc_size, /* .is_host          = */ NULL,                // defaults to false
+            ggml_backend_webgpu_buffer_type_alloc_buffer,                                    /* .get_alignment    = */
+            ggml_backend_webgpu_buffer_type_get_alignment,                                   /* .get_max_size     = */
+            ggml_backend_webgpu_buffer_type_get_max_size,                                    /* .get_alloc_size   = */
+            ggml_backend_webgpu_buffer_type_get_alloc_size, /* .is_host          = */ NULL,  // defaults to false
         },
         /* .device  = */
         dev,
@@ -3000,6 +3044,9 @@ static bool ggml_backend_webgpu_device_supports_op(ggml_backend_dev_t dev, const
         case GGML_OP_CONCAT:
             supports_op = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32);
             break;
+        case GGML_OP_REPEAT:
+            supports_op = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_I32 || src0->type == GGML_TYPE_I16);
+            break;
         case GGML_OP_CPY:
         case GGML_OP_CONT:
             supports_op = ((op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16) &&

ggml/src/ggml-webgpu/wgsl-shaders/repeat.wgsl

@@ -0,0 +1,67 @@
+enable f16;
+
+struct Params {
+    ne: u32,
+
+    offset_src0: u32,
+    offset_dst: u32,
+
+    stride_src0_0: u32,
+    stride_src0_1: u32,
+    stride_src0_2: u32,
+    stride_src0_3: u32,
+
+    a_ne0: u32,
+    a_ne1: u32,
+    a_ne2: u32,
+    a_ne3: u32,
+
+    ne0: u32,
+    ne1: u32,
+    ne2: u32,
+};
+
+#ifdef TYPE_F32
+#define DataType f32
+#endif
+#ifdef TYPE_I32
+#define DataType i32
+#endif
+#ifdef TYPE_I16
+// same size (16-bit) is sufficient for repeat
+#define DataType f16
+#endif
+
+@group(0) @binding(0)
+var<storage, read_write> src0: array<DataType>;
+
+@group(0) @binding(1)
+var<storage, read_write> dst: array<DataType>;
+
+@group(0) @binding(2)
+var<uniform> params: Params;
+
+@compute @workgroup_size(WG_SIZE)
+fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
+    if (gid.x < params.ne) {
+        var i = gid.x;
+        let i3 = i / (params.ne2 * params.ne1 * params.ne0);
+        i = i % (params.ne2 * params.ne1 * params.ne0);
+        let i2 = i / (params.ne1 * params.ne0);
+        i = i % (params.ne1 * params.ne0);
+        let i1 = i / params.ne0;
+        let i0 = i % params.ne0;
+
+        let a_i0 = i0 % params.a_ne0;
+        let a_i1 = i1 % params.a_ne1;
+        let a_i2 = i2 % params.a_ne2;
+        let a_i3 = i3 % params.a_ne3;
+
+        let a_index = a_i0 * params.stride_src0_0 +
+                           a_i1 * params.stride_src0_1 +
+                           a_i2 * params.stride_src0_2 +
+                           a_i3 * params.stride_src0_3;
+
+        dst[params.offset_dst + gid.x] = src0[params.offset_src0 + a_index];
+    }
+}

ggml/src/ggml.c

@@ -718,6 +718,14 @@ static const struct ggml_type_traits type_traits[GGML_TYPE_COUNT] = {
         .to_float                 = (ggml_to_float_t) dequantize_row_mxfp4,
         .from_float_ref           = (ggml_from_float_t)quantize_row_mxfp4_ref,
     },
+    [GGML_TYPE_NVFP4] = {
+        .type_name                = "nvfp4",
+        .blck_size                = QK_NVFP4,
+        .type_size                = sizeof(block_nvfp4),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_nvfp4,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_nvfp4_ref,
+    },
     [GGML_TYPE_Q2_K] = {
         .type_name                = "q2_K",
         .blck_size                = QK_K,
@@ -1374,6 +1382,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_Q5_1:          wtype = GGML_TYPE_Q5_1;  break;
         case GGML_FTYPE_MOSTLY_Q8_0:          wtype = GGML_TYPE_Q8_0;  break;
         case GGML_FTYPE_MOSTLY_MXFP4:         wtype = GGML_TYPE_MXFP4; break;
+        case GGML_FTYPE_MOSTLY_NVFP4:         wtype = GGML_TYPE_NVFP4; break;
         case GGML_FTYPE_MOSTLY_Q2_K:          wtype = GGML_TYPE_Q2_K;  break;
         case GGML_FTYPE_MOSTLY_Q3_K:          wtype = GGML_TYPE_Q3_K;  break;
         case GGML_FTYPE_MOSTLY_Q4_K:          wtype = GGML_TYPE_Q4_K;  break;
@@ -7641,6 +7650,7 @@ size_t ggml_quantize_chunk(
         case GGML_TYPE_Q5_1:    result = quantize_q5_1(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q8_0:    result = quantize_q8_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_MXFP4:   result = quantize_mxfp4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_NVFP4:   result = quantize_nvfp4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q2_K:    result = quantize_q2_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q3_K:    result = quantize_q3_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_Q4_K:    result = quantize_q4_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;

gguf-py/gguf/constants.py

@@ -125,6 +125,7 @@ class Keys:
         EXPERT_GROUP_SCALE                = "{arch}.expert_group_scale"
         EXPERTS_PER_GROUP                 = "{arch}.experts_per_group"
         MOE_EVERY_N_LAYERS                = "{arch}.moe_every_n_layers"
+        MOE_LATENT_SIZE                   = "{arch}.moe_latent_size"
         NEXTN_PREDICT_LAYERS              = "{arch}.nextn_predict_layers"
         NUM_DEEPSTACK_LAYERS              = "{arch}.n_deepstack_layers"
         POOLING_TYPE                      = "{arch}.pooling_type"
@@ -543,6 +544,8 @@ class MODEL_TENSOR(IntEnum):
     FFN_DOWN_CHEXP       = auto()
     FFN_UP_CHEXP         = auto()
     FFN_EXP_PROBS_B      = auto()
+    MOE_LATENT_DOWN      = auto() # nemotron 3 super
+    MOE_LATENT_UP        = auto() # nemotron 3 super
     ATTN_Q_NORM          = auto()
     ATTN_K_NORM          = auto()
     LAYER_OUT_NORM       = auto()
@@ -986,6 +989,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.FFN_UP_EXP:                "blk.{bid}.ffn_up_exps",
     MODEL_TENSOR.FFN_GATE_UP_EXP:           "blk.{bid}.ffn_gate_up_exps",
     MODEL_TENSOR.FFN_EXP_PROBS_B:           "blk.{bid}.exp_probs_b",
+    MODEL_TENSOR.MOE_LATENT_DOWN:           "blk.{bid}.ffn_latent_down",      # nemotron 3 super
+    MODEL_TENSOR.MOE_LATENT_UP:             "blk.{bid}.ffn_latent_up",        # nemotron 3 super
     MODEL_TENSOR.LAYER_OUT_NORM:            "blk.{bid}.layer_output_norm",
     MODEL_TENSOR.PER_LAYER_TOKEN_EMBD:      "per_layer_token_embd",           # gemma3n
     MODEL_TENSOR.PER_LAYER_MODEL_PROJ:      "per_layer_model_proj",           # gemma3n
@@ -2913,6 +2918,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_GATE_INP,
         MODEL_TENSOR.FFN_UP_EXP,
         MODEL_TENSOR.FFN_DOWN_EXP,
+        # expert latent
+        MODEL_TENSOR.MOE_LATENT_DOWN,
+        MODEL_TENSOR.MOE_LATENT_UP,
         # shared expert
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
@@ -3776,6 +3784,7 @@ class GGMLQuantizationType(IntEnum):
     TQ1_0   = 34
     TQ2_0   = 35
     MXFP4   = 39
+    NVFP4   = 40
 
 
 class ExpertGatingFuncType(IntEnum):
@@ -3872,6 +3881,7 @@ class VisionProjectorType:
     GEMMA3 = "gemma3"
     GEMMA3NV = "gemma3nv"
     GEMMA3NA = "gemma3na"
+    PHI4 = "phi4"
     IDEFICS3 = "idefics3"
     PIXTRAL = "pixtral"
     LLAMA4 = "llama4"
@@ -3933,6 +3943,7 @@ GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
     GGMLQuantizationType.TQ1_0:   (256, 2 + 4 * 13),
     GGMLQuantizationType.TQ2_0:   (256, 2 + 64),
     GGMLQuantizationType.MXFP4:   (32, 1 + 16),
+    GGMLQuantizationType.NVFP4:   (64, 4 + 32),
 }
 
 

gguf-py/gguf/gguf_writer.py

@@ -139,10 +139,13 @@ class GGUFWriter:
                 size = prod(shape)
 
                 if "_exps." in name:
-                    expert_count = shape[-2 if ".bias" in name else -3]
-                    expert_params += (size // expert_count)
-                    expert_sum += expert_count
-                    n_expert_tensors += 1
+                    if len(shape) >= 3:
+                        expert_count = shape[-2 if ".bias" in name else -3]
+                        expert_params += (size // expert_count)
+                        expert_sum += expert_count
+                        n_expert_tensors += 1
+                    else:
+                        shared_params += size
                 else:
                     shared_params += size
 
@@ -859,6 +862,9 @@ class GGUFWriter:
     def add_moe_every_n_layers(self, value: int) -> None:
         self.add_uint32(Keys.LLM.MOE_EVERY_N_LAYERS.format(arch=self.arch), value)
 
+    def add_moe_latent_size(self, value: int) -> None:
+        self.add_uint32(Keys.LLM.MOE_LATENT_SIZE.format(arch=self.arch), value)
+
     def add_nextn_predict_layers(self, count: int) -> None:
         self.add_uint32(Keys.LLM.NEXTN_PREDICT_LAYERS.format(arch=self.arch), count)
 

gguf-py/gguf/quants.py

@@ -704,6 +704,65 @@ class MXFP4(__Quant, qtype=GGMLQuantizationType.MXFP4):
         return (d * qs.astype(np.float32))
 
 
+class NVFP4(__Quant, qtype=GGMLQuantizationType.NVFP4):
+    # E2M1 values doubled (kvalues_mxfp4 convention)
+    kvalues = (0, 1, 2, 3, 4, 6, 8, 12, 0, -1, -2, -3, -4, -6, -8, -12)
+
+    @staticmethod
+    def ue4m3_to_fp32(x: np.ndarray) -> np.ndarray:
+        """Decode unsigned E4M3 (bias=7) to float, with 0.5 factor for kvalues convention."""
+        exp = (x >> 3).astype(np.int32) & 0xF
+        man = (x & 0x7).astype(np.float32)
+        raw = np.where(
+            exp == 0,
+            man * 2**-9,
+            (1.0 + man / 8.0) * (2.0 ** (exp.astype(np.float32) - 7)))
+        return np.where((x == 0) | (x == 0x7F), 0.0, raw * 0.5)
+
+    @staticmethod
+    def fp32_to_ue4m3(x: np.ndarray) -> np.ndarray:
+        """Vectorized float32 to unsigned E4M3, matching ggml_fp32_to_ue4m3 in C."""
+        x = np.clip(x, 0.0, 448.0).astype(np.float32)
+        bits = x.view(np.uint32)
+        fp32_exp = ((bits >> 23) & 0xFF).astype(np.int32) - 127
+        fp32_man = ((bits >> 20) & 0x7).astype(np.int32)
+        ue4m3_exp = fp32_exp + 7
+
+        # Subnormal
+        sub_man = np.clip((x * 512.0 + 0.5).astype(np.int32), 0, 7)
+        sub_result = np.where(sub_man >= 1, sub_man, 0).astype(np.uint8)
+
+        # Normal with rounding
+        round_bit = ((bits >> 19) & 1).astype(np.int32)
+        man = fp32_man + round_bit
+        exp = ue4m3_exp.copy()
+        overflow = man > 7
+        man = np.where(overflow, 0, man)
+        exp = np.where(overflow, exp + 1, exp)
+        normal_result = np.where(exp >= 15, np.uint8(0x7E), ((exp << 3) | man).astype(np.uint8))
+
+        return np.where(x <= 0.0, np.uint8(0),
+                        np.where(ue4m3_exp <= 0, sub_result,
+                        np.where(ue4m3_exp >= 15, np.uint8(0x7E), normal_result)))
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_super = blocks.shape[0]
+
+        d_bytes, qs = np.hsplit(blocks, [4])
+        d = cls.ue4m3_to_fp32(d_bytes).reshape(n_super, 4, 1)  # (n_super, 4, 1)
+
+        qs = qs.reshape(n_super, 4, 8)
+        lo = (qs & np.uint8(0x0F)).view(np.int8)
+        hi = (qs >> np.uint8(4)).view(np.int8)
+        vals = np.concatenate([lo, hi], axis=-1)  # (n_super, 4, 16)
+
+        kvalues = np.array(cls.kvalues, dtype=np.int8).reshape(1, 1, 16)
+        vals = np.take_along_axis(kvalues, vals, axis=-1)
+
+        return (d * vals.astype(np.float32)).reshape(n_super, 64)
+
+
 class IQ2_XXS(__Quant, qtype=GGMLQuantizationType.IQ2_XXS):
     ksigns: bytes = (
         b"\x00\x81\x82\x03\x84\x05\x06\x87\x88\x09\x0a\x8b\x0c\x8d\x8e\x0f"

gguf-py/gguf/scripts/gguf_convert_endian.py

@@ -65,6 +65,7 @@ byteswap_tensors = {
     gguf.GGMLQuantizationType.Q4_K:  byteswap_q4_k,
     gguf.GGMLQuantizationType.Q6_K:  byteswap_q6_k,
     gguf.GGMLQuantizationType.MXFP4: byteswap_noop,
+    gguf.GGMLQuantizationType.NVFP4: byteswap_noop,
 }
 
 

gguf-py/gguf/tensor_mapping.py

@@ -571,6 +571,14 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.experts.gate_up_proj",
         ),
 
+        MODEL_TENSOR.MOE_LATENT_DOWN: (
+            "backbone.layers.{bid}.mixer.fc1_latent_proj",                 # nemotron 3 super
+        ),
+
+        MODEL_TENSOR.MOE_LATENT_UP: (
+            "backbone.layers.{bid}.mixer.fc2_latent_proj",                 # nemotron 3 super
+        ),
+
         # Feed-forward down
         MODEL_TENSOR.FFN_DOWN: (
             "gpt_neox.layers.{bid}.mlp.dense_4h_to_h",                # gptneox

gguf-py/tests/test_quants.py

@@ -68,6 +68,7 @@ class GGMLQuants:
             "q2_K", "q3_K", "q4_K", "q5_K", "q6_K",
             "tq1_0", "tq2_0",
             "mxfp4",
+            "nvfp4",
             "iq2_xxs", "iq2_xs", "iq2_s", "iq3_xxs", "iq3_s", "iq1_s", "iq1_m",
             "iq4_nl", "iq4_xs",
         ):

include/llama.h

@@ -153,6 +153,7 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_TQ1_0         = 36, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_TQ2_0         = 37, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_MXFP4_MOE     = 38, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_NVFP4         = 39, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };

models/templates/GigaChat3-10B-A1.8B.jinja

@@ -0,0 +1,355 @@
+{#--------TOOL RENDERING FUNCTIONS---------#}
+
+{#---------------------------------------------------------------
+  Converts JSON Schema (dict) to a TypeScript type definition
+----------------------------------------------------------------#}
+{%- macro json_schema_to_typescript(schema, indent="") -%}
+	{%- set ADDITIONAL_JSON_KEYS = ['format', 'maxItems', 'maximum', 'minItems', 'minimum', 'pattern'] -%}
+	{%- set ty = schema.get("type") -%}
+
+	{# ---------------- OBJECT ---------------- #}
+	{%- if ty == "object" -%}
+		{{- "{\n" -}}
+
+		{# Start building property list #}
+		{%- set props = schema.get("properties", {}) -%}
+		{%- set required = schema.get("required", []) -%}
+		{%- set has_additional_props = schema.get("additionalProperties") is defined -%}
+		{%- set additional_props_type = none -%}
+		{%- if has_additional_props -%}
+			{%- if schema.additionalProperties == true -%}
+				{%- set additional_props_type = {'type': 'any'} -%}
+			{%- elif schema.additionalProperties is mapping -%}
+				{%- set additional_props_type = schema.additionalProperties -%}
+			{%- endif -%}
+		{%- endif -%}
+
+		{%- for key, val in props.items() -%}
+			{# ---------- Description Comments ---------- #}
+			{%- if "description" in val -%}
+				{%- for line in val['description'].split('\n') -%}
+					{%- if line.strip() -%}
+						{{- indent + '// ' + line + '\n' -}}
+					{%- endif -%}
+				{%- endfor -%}
+			{%- endif -%}
+
+			{# ---------- Additional JSON Keys ---------- #}
+			{%- for add_key, add_val in val.items() -%}
+				{%- if add_key in ADDITIONAL_JSON_KEYS -%}
+					{%- if add_val is string -%}
+						{{- indent + '// ' + add_key + ': "' + add_val + '"' + '\n'  -}}
+					{%- else -%}
+						{{- indent + '// ' + add_key + ': ' ~ add_val ~ '\n' -}}
+					{%- endif -%}
+				{%- endif -%}
+			{%- endfor -%}
+
+			{# ---------- Property Definition ---------- #}
+            {%- set type_str = json_schema_to_typescript(
+                val, 
+                indent + "  "
+            ) -%}
+
+			{{- indent + key + ('' if key in required else '?') + ': ' + type_str + ',' -}}
+
+			{%- if "default" in val or "defalut_value" in val -%}
+				{%- set default = val.get("default", val.get("defalut_value")) -%}
+				{%- if default is string -%}
+					{{- ' // default: "' + default + '"' -}}
+				{%- else -%}
+					{{- ' // default: ' ~ default -}}
+				{%- endif -%}
+			{%- endif -%}
+
+			{{- "\n" -}}
+		{%- endfor -%}
+
+		{# Handle additionalProperties as index signature #}
+		{%- if has_additional_props and additional_props_type is not none -%}
+			{%- set additional_type_str = json_schema_to_typescript(
+				additional_props_type,
+				indent + "  "
+			) -%}
+			{{- indent + '[key: string]: ' + additional_type_str + '\n' -}}
+		{%- endif -%}
+
+		{{- indent[: (indent|length - "  "|length) ] + '}' -}}
+
+	{# ---------------- STRING ---------------- #}
+	{%- elif ty == "string" -%}
+		{%- if schema.get("enum") -%}
+            {%- set ns = namespace(enum = []) -%}
+            {%- for en in schema['enum'] -%}
+                {%- set ns.enum = ns.enum + ['"' ~ en ~ '"'] -%}
+            {%- endfor -%}
+			{{- ns.enum | join(' | ') -}}
+		{%- elif schema.get("format", "none") in ['date-time', 'date'] -%}
+			{{- 'Date' -}}
+		{%- else -%}
+			{{- 'string' -}}
+		{%- endif -%}
+
+	{# ---------------- NUMBER / INTEGER ---------------- #}
+	{%- elif ty in ["number", "integer"] -%}
+		{%- if schema.get("enum") -%}
+			{{- schema.enum | join(' | ') -}}
+		{%- else -%}
+			{{- 'number' -}}
+		{%- endif -%}
+
+	{# ---------------- BOOLEAN ---------------- #}
+	{%- elif ty == "boolean" -%}
+		{{- 'boolean' -}}
+
+	{# ---------------- ARRAY ---------------- #}
+	{%- elif ty == "array" -%}
+		{%- if "items" in schema -%}
+			{{- json_schema_to_typescript(schema['items'], indent) + '[]' -}}
+		{%- else -%}
+			{{- 'Array<any>' -}}
+		{%- endif -%}
+
+	{# ---------------- FALLBACK ---------------- #}
+	{%- else -%}
+		{{- 'any' -}}
+	{%- endif -%}
+{%- endmacro -%}
+
+{#---------------------------------------------------------------
+  Renders a namespace and its tool definitions in TypeScript style
+----------------------------------------------------------------#}
+
+{%- macro render_tool_namespace(namespace_name, tools) -%}
+	{%- set ns = namespace(sections = ['namespace ' ~ namespace_name ~ ' {']) -%}
+
+	{%- for tool in tools -%}
+		{%- if tool.function -%}
+			{%- set tool = tool.function -%}
+		{%- endif -%}
+
+		{%- set ns_tool = namespace(content_lines=[]) -%}
+
+		{# ---------- TOOL DESCRIPTION ---------- #}
+		{%- if tool.get('description') -%}
+			{%- for line in tool['description'].split('\n') -%}
+				{%- if line.strip() -%}
+					{%- set ns_tool.content_lines = ns_tool.content_lines + ['// ' ~ line] -%}
+				{%- endif -%}
+			{%- endfor -%}
+		{%- endif -%}
+
+		{# ---------- TOOL SIGNATURE ---------- #}
+		{%- set main_body = "" -%}
+		{%- set params = tool.get("parameters") -%}
+		{%- if params and params.get("properties") -%}
+			{%- set param_type = json_schema_to_typescript(params, "  ") -%}
+			{%- set main_body = 'type ' ~ tool.name ~ ' = (_: ' ~ param_type ~ ') => ' -%}
+		{%- else -%}
+			{%- set main_body = 'type ' ~ tool.name ~ ' = () => ' -%}
+		{%- endif -%}
+
+		{# ---------- RETURN TYPE ---------- #}
+		{%- set return_params = tool.get("return_parameters") -%}
+		{%- if return_params and return_params.get("properties") -%}
+			{%- set return_type = json_schema_to_typescript(return_params, "  ") -%}
+			{%- set main_body = main_body ~ return_type -%}
+		{%- else -%}
+			{%- set main_body = main_body ~ 'any' -%}
+		{%- endif -%}
+
+		{%- set main_body = main_body ~ ';\n' -%}
+
+		{%- set ns_tool.content_lines = ns_tool.content_lines + [main_body] -%}
+
+		{# ---------- ADD TOOL TO SECTIONS ---------- #}
+		{%- set ns.sections = ns.sections + [ns_tool.content_lines | join('\n')] -%}
+	{%- endfor -%}
+
+	{%- set ns.sections = ns.sections + ['} // namespace ' ~ namespace_name] -%}
+
+	{{- ns.sections | join('\n') -}}
+{%- endmacro -%}
+
+
+{# ----------- MESSAGE RENDERING HELPER FUNCTIONS ------------ #}
+
+{%- macro render_role_message(message, role=None) -%}
+    {%- if not role -%}
+        {%- set role = message["role"] -%}
+    {%- endif -%}
+
+    {%- set message_content = message['content'] or '' -%}
+    {%- if message_content is not string -%}
+        {%- set message_content = message_content | tojson(ensure_ascii=False) -%}
+    {%- endif -%}
+
+    {{- role + add_tokens.role_sep + message_content + add_tokens.message_sep -}}
+
+{%- endmacro -%}
+
+
+{%- macro render_function_call(message) -%}
+    {%- set call = message['content'] -%}
+    {%- if call.function -%}
+        {%- set call = call.function -%}
+    {%- endif -%}
+
+    {%- set arguments = call['arguments'] -%}
+	{%- if arguments is not string -%}
+		{%- set arguments = arguments| tojson(ensure_ascii=False) -%}
+	{%- endif -%}
+
+    {{- render_role_message(
+        {
+            'role': 'function call',
+            'content': '{"name": "' ~ call['name'] ~ '", "arguments": ' ~ arguments ~ '}'
+        }
+    ) -}}
+{%- endmacro -%}
+
+{# ----- SPECIAL TOKENS ----- #}
+
+{%- set add_tokens = namespace(
+    role_sep="<|role_sep|>\n", 
+    message_sep="<|message_sep|>\n\n"
+) -%}
+
+{# ----- DEFAULT DEVSYSTEM ----- #}
+
+{%- set DEVSYSTEM -%}
+<role_description>
+Description of the roles available in the dialog.
+
+`developer system`
+A message added by Sber before the main dialog. It has the highest priority and sets global, non-overridable conditions (for example, conversation rules, the safety policy, the assistant's overall response style, etc.).
+
+`system`
+A system instruction added by developers or by the user, but with a lower priority than `developer system`. It usually describes the assistant's instructions, a specific response style, and other conditions for this particular dialog.
+
+`user`
+A message or request from the user. The assistant follows it if it does not conflict with higher-priority instructions (see <instruction_priority>).
+
+`user memory`
+A sequence of the most up-to-date long-term facts about the user at the time of their request, presented as a JSON list of strings. Facts are listed in chronological order, meaning newer facts are appended to the end of the sequence. When facts are changed or deleted, records of previous facts remain in the sequence. The assistant saves facts using a function and uses them in accordance with the <memory_guidelines> block below.
+
+`added files`
+Metadata about files available for use in the dialog, presented in JSON format. It contains the following keys: id (a unique file identifier), name (file name), type (file type).
+
+`assistant`
+The assistant's reply to the user's request. If the system instruction or the user does not set additional rules for `assistant`, this reply must comply with the instructions in the <assistant_guidelines> block below. The list of functions available to call is contained in `function descriptions`. The name of the required function and its arguments will be generated next by the `function call` role. In its replies, the assistant follows the instructions in accordance with <instruction_priority>.
+
+`function descriptions`
+Function descriptions in TypeScript format. A function is a special tool (or a set of instructions) that the assistant can call to perform specific actions, computations, or obtain data needed to solve the user's task. Each function description contains blocks with the name, description, and arguments. Sometimes the description contains separate blocks with return parameters and usage examples that illustrate the correct call and arguments.
+
+`function call`
+The function that `assistant` calls based on the dialog context, and its arguments. The function is invoked in strict accordance with the instructions in the <function_usage> block.
+
+`function result`
+The result of the last function call.
+</role_description>
+
+<available_modalities>
+The assistant can work with the following modalities: text, available functions.
+</available_modalities>
+
+<instruction_priority>
+If instructions from different roles conflict within the dialog context, observe the following priorities:  
+`developer system` > `system` > `user` > `function descriptions` > `function result` > `user memory`
+</instruction_priority>
+
+<function_usage>
+Basic instructions for working with functions.
+
+Only call those functions that are described in `function descriptions`.
+
+Call available functions when, according to their description, such a call will help provide a more complete and/or accurate answer to the user's request. Fill in function arguments using information from the dialog context. If a function could help answer the request but a required argument is missing from the context, ask the user for the missing data before calling the function. If a necessary function is unavailable or an error occurs, briefly inform the user and, if possible, suggest an alternative.
+</function_usage>
+
+<memory_guidelines>
+Rules for using facts in long-term memory:
+
+If there is no message under the `user memory` role in the dialog, this is equivalent to the absence of long-term facts about the user in memory. In that case, information about the user is limited to the current dialog, and no new facts should be saved.
+</memory_guidelines>
+
+<assistant_guidelines>
+You are a helpful assistant.
+
+# Instructions
+- Strictly follow the instruction priority.
+- Maintain a logical chain of reasoning when answering the user's question.
+- For complex questions (for example, STEM), try to answer in detail unless the system message or dialog context limits the response length.
+- Be helpful, truthful, and avoid unsafe or prohibited content in your responses.
+- Try to reply in the language in which the user asked their question.
+</assistant_guidelines>
+
+A dialog will follow below.
+The dialog may include various roles described in the <role_description> block.
+Each turn begins with the role name and a special token that marks the end of the role's full name, and ends with a special end-of-turn token.
+Your task is to continue the dialog from the last specified role in accordance with the dialog context.
+{%- endset -%}
+
+
+{#- ---------------------- RENDERING STARTS HERE ---------------------- -#}
+
+
+{# ----- RENDER BOS TOKEN ----- #}
+{{- bos_token -}}
+
+
+{# ----- RENDER DEVSYSTEM ----- #}
+{{- render_role_message({"role": "developer system", "content": DEVSYSTEM}) -}}
+
+{# ----- RENDER SYSTEM IF PRESENT ----- #}
+{%- if messages and messages[0]['role'] == 'system' -%}
+    {{- render_role_message(messages[0]) -}}
+    {%- set messages = messages[1:] -%}
+{%- endif -%}
+
+{# ----- RENDER TOOLS ----- #}
+{%- if tools -%}
+    {%- set tools_content = (
+        render_tool_namespace('functions', tools) 
+        + "\n\n"
+    ) -%}
+    {{- render_role_message({'role': 'function descriptions', 'content': tools_content}) -}}
+{%- endif -%}
+
+{# ----- MAIN MESSAGE LOOP ----- #}
+{%- for message in messages -%}
+
+    {# ----- TOOL MESSAGE -------#}
+    {%- if message['role'] == 'tool' -%}
+        {{- render_role_message(message, 'function result') -}}
+
+
+    {# ----- ASSISTANT MESSAGE ----- #}
+    {%- elif message['role'] == 'assistant' -%}
+
+        {# ----- FUNCTION CALL PART CHECKING: SINGLE CALL SETUP ----- #}
+        {%- if message.tool_calls is defined and message.tool_calls -%}
+            {%- set function_call = message.tool_calls[0] -%}
+        {%- else -%}
+            {%- set function_call = None -%}
+        {%- endif -%}
+
+        {# ----- MAIN ASSISTANT RENDERING ----- #}
+
+        {{- render_role_message({'role': 'assistant', 'content': message.content}) -}}
+        {%- if function_call -%}
+            {{- render_function_call({'role': 'function call', 'content': function_call}) -}}
+        {%- endif -%}
+
+        
+    {# ----- OTHER MESSAGES ----- #}
+    {%- else -%}
+        {{- render_role_message(message) -}}
+    {%- endif -%}
+
+    {# ----- ADDING GENERATION PROMPT ----- #}
+
+    {%- if loop.last and add_generation_prompt and message['role'] != 'assistant' -%}
+        {{- 'assistant' + add_tokens.role_sep -}}
+    {%- endif -%}
+
+{%- endfor -%}

models/templates/GigaChat3.1-10B-A1.8B.jinja

@@ -0,0 +1,339 @@
+{#--------TOOL RENDERING FUNCTIONS---------#}
+
+{#---------------------------------------------------------------
+  Converts JSON Schema (dict) to a TypeScript type definition
+----------------------------------------------------------------#}
+{%- macro json_schema_to_typescript(schema, indent="") -%}
+	{%- set ADDITIONAL_JSON_KEYS = ['format', 'maxItems', 'maximum', 'minItems', 'minimum', 'pattern'] -%}
+	{%- set ty = schema.get("type") -%}
+
+	{# ---------------- OBJECT ---------------- #}
+	{%- if ty == "object" -%}
+		{{- "{\n" -}}
+
+		{# Start building property list #}
+		{%- set props = schema.get("properties", {}) -%}
+		{%- set required = schema.get("required", []) -%}
+		{%- set has_additional_props = schema.get("additionalProperties") is defined -%}
+		{%- set additional_props_type = none -%}
+		{%- if has_additional_props -%}
+			{%- if schema.additionalProperties == true -%}
+				{%- set additional_props_type = {'type': 'any'} -%}
+			{%- elif schema.additionalProperties is mapping -%}
+				{%- set additional_props_type = schema.additionalProperties -%}
+			{%- endif -%}
+		{%- endif -%}
+
+		{%- for key, val in props.items() -%}
+			{# ---------- Description Comments ---------- #}
+			{%- if "description" in val -%}
+				{%- for line in val['description'].split('\n') -%}
+					{%- if line.strip() -%}
+						{{- indent + '// ' + line + '\n' -}}
+					{%- endif -%}
+				{%- endfor -%}
+			{%- endif -%}
+
+			{# ---------- Additional JSON Keys ---------- #}
+			{%- for add_key, add_val in val.items() -%}
+				{%- if add_key in ADDITIONAL_JSON_KEYS -%}
+					{%- if add_val is string -%}
+						{{- indent + '// ' + add_key + ': "' + add_val + '"' + '\n'  -}}
+					{%- else -%}
+						{{- indent + '// ' + add_key + ': ' ~ add_val ~ '\n' -}}
+					{%- endif -%}
+				{%- endif -%}
+			{%- endfor -%}
+
+			{# ---------- Property Definition ---------- #}
+            {%- set type_str = json_schema_to_typescript(
+                val, 
+                indent + "  "
+            ) -%}
+
+			{{- indent + key + ('' if key in required else '?') + ': ' + type_str + ',' -}}
+
+			{%- if "default" in val or "defalut_value" in val -%}
+				{%- set default = val.get("default", val.get("defalut_value")) -%}
+				{%- if default is string -%}
+					{{- ' // default: "' + default + '"' -}}
+				{%- else -%}
+					{{- ' // default: ' ~ default -}}
+				{%- endif -%}
+			{%- endif -%}
+
+			{{- "\n" -}}
+		{%- endfor -%}
+
+		{# Handle additionalProperties as index signature #}
+		{%- if has_additional_props and additional_props_type is not none -%}
+			{%- set additional_type_str = json_schema_to_typescript(
+				additional_props_type,
+				indent + "  "
+			) -%}
+			{{- indent + '[key: string]: ' + additional_type_str + '\n' -}}
+		{%- endif -%}
+
+		{{- indent[: (indent|length - "  "|length) ] + '}' -}}
+
+	{# ---------------- STRING ---------------- #}
+	{%- elif ty == "string" -%}
+		{%- if schema.get("enum") -%}
+            {%- set ns = namespace(enum = []) -%}
+            {%- for en in schema['enum'] -%}
+                {%- set ns.enum = ns.enum + ['"' ~ en ~ '"'] -%}
+            {%- endfor -%}
+			{{- ns.enum | join(' | ') -}}
+		{%- elif schema.get("format", "none") in ['date-time', 'date'] -%}
+			{{- 'Date' -}}
+		{%- else -%}
+			{{- 'string' -}}
+		{%- endif -%}
+
+	{# ---------------- NUMBER / INTEGER ---------------- #}
+	{%- elif ty in ["number", "integer"] -%}
+		{%- if schema.get("enum") -%}
+			{{- schema.enum | join(' | ') -}}
+		{%- else -%}
+			{{- 'number' -}}
+		{%- endif -%}
+
+	{# ---------------- BOOLEAN ---------------- #}
+	{%- elif ty == "boolean" -%}
+		{{- 'boolean' -}}
+
+	{# ---------------- ARRAY ---------------- #}
+	{%- elif ty == "array" -%}
+		{%- if "items" in schema -%}
+			{{- json_schema_to_typescript(schema['items'], indent) + '[]' -}}
+		{%- else -%}
+			{{- 'Array<any>' -}}
+		{%- endif -%}
+
+	{# ---------------- FALLBACK ---------------- #}
+	{%- else -%}
+		{{- 'any' -}}
+	{%- endif -%}
+{%- endmacro -%}
+
+{#---------------------------------------------------------------
+  Renders a namespace and its tool definitions in TypeScript style
+----------------------------------------------------------------#}
+
+{%- macro render_tool_namespace(namespace_name, tools) -%}
+	{%- set ns = namespace(sections = ['namespace ' ~ namespace_name ~ ' {']) -%}
+
+	{%- for tool in tools -%}
+		{%- if tool.function -%}
+			{%- set tool = tool.function -%}
+		{%- endif -%}
+
+		{%- set ns_tool = namespace(content_lines=[]) -%}
+
+		{# ---------- TOOL DESCRIPTION ---------- #}
+		{%- if tool.get('description') -%}
+			{%- for line in tool['description'].split('\n') -%}
+				{%- if line.strip() -%}
+					{%- set ns_tool.content_lines = ns_tool.content_lines + ['// ' ~ line] -%}
+				{%- endif -%}
+			{%- endfor -%}
+		{%- endif -%}
+
+		{# ---------- TOOL SIGNATURE ---------- #}
+		{%- set main_body = "" -%}
+		{%- set params = tool.get("parameters") -%}
+		{%- if params and params.get("properties") -%}
+			{%- set param_type = json_schema_to_typescript(params, "  ") -%}
+			{%- set main_body = 'type ' ~ tool.name ~ ' = (_: ' ~ param_type ~ ') => ' -%}
+		{%- else -%}
+			{%- set main_body = 'type ' ~ tool.name ~ ' = () => ' -%}
+		{%- endif -%}
+
+		{# ---------- RETURN TYPE ---------- #}
+		{%- set return_params = tool.get("return_parameters") -%}
+		{%- if return_params and return_params.get("properties") -%}
+			{%- set return_type = json_schema_to_typescript(return_params, "  ") -%}
+			{%- set main_body = main_body ~ return_type -%}
+		{%- else -%}
+			{%- set main_body = main_body ~ 'any' -%}
+		{%- endif -%}
+
+		{%- set main_body = main_body ~ ';\n' -%}
+
+		{%- set ns_tool.content_lines = ns_tool.content_lines + [main_body] -%}
+
+		{# ---------- ADD TOOL TO SECTIONS ---------- #}
+		{%- set ns.sections = ns.sections + [ns_tool.content_lines | join('\n')] -%}
+	{%- endfor -%}
+
+	{%- set ns.sections = ns.sections + ['} // namespace ' ~ namespace_name] -%}
+
+	{{- ns.sections | join('\n') -}}
+{%- endmacro -%}
+
+
+{# ----------- MESSAGE RENDERING HELPER FUNCTIONS ------------ #}
+
+{%- macro render_function_call(call) -%}
+    {%- if call.function -%}
+        {%- set call = call.function -%}
+    {%- endif -%}
+
+    {%- set arguments = call['arguments'] -%}
+	{%- if arguments is not string -%}
+		{%- set arguments = arguments| tojson(ensure_ascii=False) -%}
+	{%- endif -%}
+
+    {{- '{"name": "' ~ call['name'] ~ '", "arguments": ' ~ arguments ~ '}' -}}
+{%- endmacro -%}
+
+
+{%- macro render_role_message(message, role=None) -%}
+    {%- if not role -%}
+        {%- set role = message["role"] -%}
+    {%- endif -%}
+
+    {%- set message_content = message['content'] or '' -%}
+    {%- if message_content is not string -%}
+        {%- set message_content = message_content | tojson(ensure_ascii=False) -%}
+    {%- endif -%}
+
+    {{- role + add_tokens.role_sep + message_content -}}
+
+    {%- if message.tool_calls is defined and message.tool_calls -%}
+        {{- add_tokens.function_call + render_function_call(message.tool_calls[0]) -}}
+    {%- endif -%}
+
+    {{- add_tokens.message_sep -}}
+
+{%- endmacro -%}
+
+
+
+{# ----- SPECIAL TOKENS ----- #}
+
+{%- set add_tokens = namespace(
+    role_sep="<|role_sep|>\n", 
+    message_sep="<|message_sep|>\n\n",
+    function_call="<|function_call|>"
+) -%}
+
+{# ----- DEFAULT DEVSYSTEM ----- #}
+
+{%- set DEVSYSTEM -%}
+<role_description>
+Description of the roles available in the dialog.
+
+`developer system`
+A message added by Sber before the main dialog. It has the highest priority and sets global, non-overridable conditions (for example, conversation rules, the safety policy, the assistant's overall response style, etc.).
+
+`system`
+A system instruction added by developers or by the user, but with a lower priority than `developer system`. It usually describes the assistant's instructions, a specific response style, and other conditions for this particular dialog.
+
+`user`
+A message or request from the user. The assistant follows it if it does not conflict with higher-priority instructions (see <instruction_priority>).
+
+`user memory`
+A sequence of the most up-to-date long-term facts about the user at the time of their request, presented as a JSON list of strings. Facts are listed in chronological order, meaning newer facts are appended to the end of the sequence. When facts are changed or deleted, records of previous facts remain in the sequence. The assistant saves facts using a function and uses them in accordance with the <memory_guidelines> block below.
+
+`added files`
+Metadata about files available for use in the dialog, presented in JSON format. It contains the following keys: id (a unique file identifier), name (file name), type (file type).
+
+`assistant`
+The assistant's reply to the user's request. If the system instruction or the user does not set additional rules for `assistant`, this reply must comply with the instructions in the <assistant_guidelines> block below. The list of functions available to call is contained in `function descriptions`. The name of the required function and its arguments will be generated next by the `function call` role. In its replies, the assistant follows the instructions in accordance with <instruction_priority>.
+
+`function descriptions`
+Function descriptions in TypeScript format. A function is a special tool (or a set of instructions) that the assistant can call to perform specific actions, computations, or obtain data needed to solve the user's task. Each function description contains blocks with the name, description, and arguments. Sometimes the description contains separate blocks with return parameters and usage examples that illustrate the correct call and arguments.
+
+`function call`
+The function that `assistant` calls based on the dialog context, and its arguments. The function is invoked in strict accordance with the instructions in the <function_usage> block.
+
+`function result`
+The result of the last function call.
+</role_description>
+
+<available_modalities>
+The assistant can work with the following modalities: text, available functions.
+</available_modalities>
+
+<instruction_priority>
+If instructions from different roles conflict within the dialog context, observe the following priorities:  
+`developer system` > `system` > `user` > `function descriptions` > `function result` > `user memory`
+</instruction_priority>
+
+<function_usage>
+Basic instructions for working with functions.
+
+Only call those functions that are described in `function descriptions`.
+
+Call available functions when, according to their description, such a call will help provide a more complete and/or accurate answer to the user's request. Fill in function arguments using information from the dialog context. If a function could help answer the request but a required argument is missing from the context, ask the user for the missing data before calling the function. If a necessary function is unavailable or an error occurs, briefly inform the user and, if possible, suggest an alternative.
+</function_usage>
+
+<memory_guidelines>
+Rules for using facts in long-term memory:
+
+If there is no message under the `user memory` role in the dialog, this is equivalent to the absence of long-term facts about the user in memory. In that case, information about the user is limited to the current dialog, and no new facts should be saved.
+</memory_guidelines>
+
+<assistant_guidelines>
+You are a helpful assistant.
+
+# Instructions
+- Strictly follow the instruction priority.
+- Maintain a logical chain of reasoning when answering the user's question.
+- For complex questions (for example, STEM), try to answer in detail unless the system message or dialog context limits the response length.
+- Be helpful, truthful, and avoid unsafe or prohibited content in your responses.
+- Try to reply in the language in which the user asked their question.
+</assistant_guidelines>
+
+A dialog will follow below.
+The dialog may include various roles described in the <role_description> block.
+Each turn begins with the role name and a special token that marks the end of the role's full name, and ends with a special end-of-turn token.
+Your task is to continue the dialog from the last specified role in accordance with the dialog context.
+{%- endset -%}
+
+
+{#- ---------------------- RENDERING STARTS HERE ---------------------- -#}
+
+
+{# ----- RENDER BOS TOKEN ----- #}
+{{- bos_token -}}
+
+
+{# ----- RENDER DEVSYSTEM ----- #}
+{{- render_role_message({"role": "developer system", "content": DEVSYSTEM}) -}}
+
+{# ----- RENDER SYSTEM IF PRESENT ----- #}
+{%- if messages and messages[0]['role'] == 'system' -%}
+    {{- render_role_message(messages[0]) -}}
+    {%- set messages = messages[1:] -%}
+{%- endif -%}
+
+{# ----- RENDER TOOLS ----- #}
+{%- if tools -%}
+    {%- set tools_content = (
+        render_tool_namespace('functions', tools) 
+        + "\n\n"
+    ) -%}
+    {{- render_role_message({'role': 'function descriptions', 'content': tools_content}) -}}
+{%- endif -%}
+
+{# ----- MAIN MESSAGE LOOP ----- #}
+{%- for message in messages -%}
+
+    {# ----- TOOL MESSAGE -------#}
+    {%- if message['role'] == 'tool' -%}
+        {{- render_role_message(message, 'function result') -}}
+        
+    {# ----- OTHER MESSAGES ----- #}
+    {%- else -%}
+        {{- render_role_message(message) -}}
+    {%- endif -%}
+
+    {# ----- ADDING GENERATION PROMPT ----- #}
+
+    {%- if loop.last and add_generation_prompt and message['role'] != 'assistant' -%}
+        {{- 'assistant' + add_tokens.role_sep -}}
+    {%- endif -%}
+
+{%- endfor -%}

scripts/compare-llama-bench.py

@@ -293,6 +293,10 @@ class LlamaBenchData:
         for t in self.repo.tags:
             if t.name == name:
                 return t.commit.hexsha[:self.build_len]
+        for remote in self.repo.remotes:
+            for ref in remote.refs:
+                if ref.name == name or ref.remote_head == name:
+                    return ref.commit.hexsha[:self.build_len]
         for c in self.repo.iter_commits("--all"):
             if c.hexsha[:self.build_len] == name[:self.build_len]:
                 return c.hexsha[:self.build_len]

src/llama-arch.cpp

@@ -185,6 +185,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_EXPERT_GROUP_SCALE,                "%s.expert_group_scale"                },
     { LLM_KV_EXPERTS_PER_GROUP,                 "%s.experts_per_group"                 },
     { LLM_KV_MOE_EVERY_N_LAYERS,                "%s.moe_every_n_layers"                },
+    { LLM_KV_MOE_LATENT_SIZE,                   "%s.moe_latent_size"                   },
     { LLM_KV_NEXTN_PREDICT_LAYERS,              "%s.nextn_predict_layers"              },
     { LLM_KV_NUM_DEEPSTACK_LAYERS,              "%s.n_deepstack_layers"                },
     { LLM_KV_POOLING_TYPE,                      "%s.pooling_type"                      },
@@ -365,6 +366,8 @@ static const std::map<llm_tensor, const char *> LLM_TENSOR_NAMES = {
     { LLM_TENSOR_FFN_UP_SHEXP,                           "blk.%d.ffn_up_shexp" },
     { LLM_TENSOR_FFN_DOWN_SHEXP,                         "blk.%d.ffn_down_shexp" },
     { LLM_TENSOR_FFN_EXP_PROBS_B,                        "blk.%d.exp_probs_b" },
+    { LLM_TENSOR_FFN_LATENT_DOWN,                        "blk.%d.ffn_latent_down" },
+    { LLM_TENSOR_FFN_LATENT_UP,                          "blk.%d.ffn_latent_up" },
     { LLM_TENSOR_ATTN_NORM_2,                            "blk.%d.attn_norm_2" },
     { LLM_TENSOR_ATTN_QKV,                               "blk.%d.attn_qkv" },
     { LLM_TENSOR_LAYER_OUT_NORM,                         "blk.%d.layer_output_norm" },
@@ -1879,6 +1882,8 @@ static std::set<llm_tensor> llm_get_tensor_names(llm_arch arch) {
                 LLM_TENSOR_FFN_UP_EXPS,
                 LLM_TENSOR_FFN_DOWN_EXPS,
                 LLM_TENSOR_FFN_EXP_PROBS_B,
+                LLM_TENSOR_FFN_LATENT_DOWN,
+                LLM_TENSOR_FFN_LATENT_UP,
                 // MoE shared expert layer
                 LLM_TENSOR_FFN_DOWN_SHEXP,
                 LLM_TENSOR_FFN_UP_SHEXP,
@@ -2754,6 +2759,9 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_NEXTN_HNORM,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
     {LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD,     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_NEXTN_SHARED_HEAD_NORM,     {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
+    // Nemotron 3 Super
+    {LLM_TENSOR_FFN_LATENT_DOWN,            {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
+    {LLM_TENSOR_FFN_LATENT_UP,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 };
 
 LLM_KV::LLM_KV(llm_arch arch, const char * suffix) : arch(arch), suffix(suffix) {}

src/llama-arch.h

@@ -189,6 +189,7 @@ enum llm_kv {
     LLM_KV_EXPERT_GROUP_SCALE,
     LLM_KV_EXPERTS_PER_GROUP,
     LLM_KV_MOE_EVERY_N_LAYERS,
+    LLM_KV_MOE_LATENT_SIZE,
     LLM_KV_NEXTN_PREDICT_LAYERS,
     LLM_KV_NUM_DEEPSTACK_LAYERS,
     LLM_KV_POOLING_TYPE,
@@ -385,6 +386,8 @@ enum llm_tensor {
     LLM_TENSOR_FFN_GATE_CHEXPS,
     LLM_TENSOR_FFN_UP_CHEXPS,
     LLM_TENSOR_FFN_EXP_PROBS_B,
+    LLM_TENSOR_FFN_LATENT_DOWN,
+    LLM_TENSOR_FFN_LATENT_UP,
     LLM_TENSOR_ATTN_Q_NORM,
     LLM_TENSOR_ATTN_K_NORM,
     LLM_TENSOR_LAYER_OUT_NORM,

src/llama-context.cpp

@@ -151,7 +151,8 @@ llama_context::llama_context(
     cparams.auto_fa    = params.flash_attn_type == LLAMA_FLASH_ATTN_TYPE_AUTO;
 
     cparams.fused_gdn_ar = true;
-    cparams.fused_gdn_ch = false; // TODO: implement
+    cparams.fused_gdn_ch = true;
+    cparams.auto_fgdn    = true;
 
     // with causal attention, the batch size is limited by the context size
     cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
@@ -462,37 +463,81 @@ void llama_context::sched_reserve() {
         cparams.auto_fa = false;
     }
 
-    if (cparams.fused_gdn_ar) {
-        auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
-        if (!gf) {
-            throw std::runtime_error("failed to reserve graph for fused Gated Delta Net check");
-        }
+    if (cparams.auto_fgdn) {
+        LLAMA_LOG_INFO("%s: resolving fused Gated Delta Net support:\n", __func__);
 
-        const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FGDNAR) + 1;
-        bool gdn_device_mismatch = false;
-        for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
-            ggml_tensor * n = ggml_graph_node(gf, i);
-            if (n->op != GGML_OP_GATED_DELTA_NET) {
-                continue;
+        if (cparams.fused_gdn_ar) {
+            auto * gf = graph_reserve(1, n_seqs, n_outputs, mctx.get(), true);
+            if (!gf) {
+                throw std::runtime_error("failed to reserve graph for fused Gated Delta Net check (autoregressive)");
             }
-            ggml_backend_dev_t device_gdn = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
 
-            GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FGDNAR "-", prefix_len) == 0);
-            const int il = std::stoi(n->name + prefix_len);
-            ggml_backend_dev_t device_kv = model.dev_layer(il);
-            if (device_gdn != device_kv) {
-                LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the fused Gated Delta Net tensor "
-                        "is assigned to device %s (usually due to missing support)\n",
-                        __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_gdn));
-                gdn_device_mismatch = true;
-                break;
+            const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FGDN_AR) + 1;
+            bool gdn_device_mismatch = false;
+            for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
+                ggml_tensor * n = ggml_graph_node(gf, i);
+                if (n->op != GGML_OP_GATED_DELTA_NET) {
+                    continue;
+                }
+                ggml_backend_dev_t device_gdn = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
+
+                GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FGDN_AR "-", prefix_len) == 0);
+                const int il = std::stoi(n->name + prefix_len);
+                ggml_backend_dev_t device_kv = model.dev_layer(il);
+                if (device_gdn != device_kv) {
+                    LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the fused Gated Delta Net tensor "
+                            "is assigned to device %s (usually due to missing support)\n",
+                            __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_gdn));
+                    gdn_device_mismatch = true;
+                    break;
+                }
+            }
+
+            if (gdn_device_mismatch) {
+                cparams.fused_gdn_ar = false;
+                LLAMA_LOG_WARN("%s: fused Gated Delta Net (autoregressive) not supported, set to disabled\n", __func__);
+            } else {
+                LLAMA_LOG_INFO("%s: fused Gated Delta Net (autoregressive) enabled\n", __func__);
             }
         }
 
-        if (gdn_device_mismatch) {
-            cparams.fused_gdn_ar = false;
-            LLAMA_LOG_WARN("%s: fused Gated Delta Net not supported, set to disabled\n", __func__);
+        if (cparams.fused_gdn_ch) {
+            // more than one token in the batch per sequence in order to take the chunked path
+            auto * gf = graph_reserve(16*n_seqs, n_seqs, n_outputs, mctx.get(), true);
+            if (!gf) {
+                throw std::runtime_error("failed to reserve graph for fused Gated Delta Net check (chunked)");
+            }
+
+            const size_t prefix_len = strlen(LLAMA_TENSOR_NAME_FGDN_CH) + 1;
+            bool gdn_device_mismatch = false;
+            for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
+                ggml_tensor * n = ggml_graph_node(gf, i);
+                if (n->op != GGML_OP_GATED_DELTA_NET) {
+                    continue;
+                }
+                ggml_backend_dev_t device_gdn = ggml_backend_get_device(ggml_backend_sched_get_tensor_backend(sched.get(), n));
+
+                GGML_ASSERT(strncmp(n->name, LLAMA_TENSOR_NAME_FGDN_CH "-", prefix_len) == 0);
+                const int il = std::stoi(n->name + prefix_len);
+                ggml_backend_dev_t device_kv = model.dev_layer(il);
+                if (device_gdn != device_kv) {
+                    LLAMA_LOG_WARN("%s: layer %d is assigned to device %s but the fused Gated Delta Net tensor "
+                            "is assigned to device %s (usually due to missing support)\n",
+                            __func__, il, ggml_backend_dev_name(device_kv), ggml_backend_dev_name(device_gdn));
+                    gdn_device_mismatch = true;
+                    break;
+                }
+            }
+
+            if (gdn_device_mismatch) {
+                cparams.fused_gdn_ch = false;
+                LLAMA_LOG_WARN("%s: fused Gated Delta Net (chunked) not supported, set to disabled\n", __func__);
+            } else {
+                LLAMA_LOG_INFO("%s: fused Gated Delta Net (chunked) enabled\n", __func__);
+            }
         }
+
+        cparams.auto_fgdn = false;
     }
 
     // reserve worst-case graph

src/llama-cparams.h

@@ -33,6 +33,7 @@ struct llama_cparams {
     bool auto_fa;
     bool fused_gdn_ar;       // use fused gated delta net (autoregressive)
     bool fused_gdn_ch;       // use fused gated delta net (chunked)
+    bool auto_fgdn;
     bool no_perf;
     bool warmup;
     bool op_offload;

src/llama-graph.cpp

@@ -900,7 +900,8 @@ ggml_tensor * llm_graph_context::build_cvec(
 
 ggml_tensor * llm_graph_context::build_lora_mm(
           ggml_tensor * w,
-          ggml_tensor * cur) const {
+          ggml_tensor * cur,
+          ggml_tensor * w_s) const {
     ggml_tensor * res = ggml_mul_mat(ctx0, w, cur);
 
     for (const auto & lora : *loras) {
@@ -921,6 +922,10 @@ ggml_tensor * llm_graph_context::build_lora_mm(
         res = ggml_add(ctx0, res, ab_cur);
     }
 
+    if (w_s) {
+        res = ggml_mul(ctx0, res, w_s);
+    }
+
     return res;
 }
 
@@ -1166,7 +1171,10 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
          llama_expert_gating_func_type gating_op,
                  int   il,
          ggml_tensor * probs_in,
-         ggml_tensor * gate_up_exps) const {
+         ggml_tensor * gate_up_exps,
+         ggml_tensor * up_exps_s,
+         ggml_tensor * gate_exps_s,
+         ggml_tensor * down_exps_s) const {
     return build_moe_ffn(
         cur,
         gate_inp,  /* gate_inp_b  */ nullptr,
@@ -1182,7 +1190,11 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         gating_op,
         il,
         probs_in,
-        gate_up_exps
+        gate_up_exps,
+        /* gate_up_exps_b */ nullptr,
+        up_exps_s,
+        gate_exps_s,
+        down_exps_s
     );
 }
 
@@ -1206,7 +1218,10 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
                  int   il,
          ggml_tensor * probs_in,
          ggml_tensor * gate_up_exps,
-         ggml_tensor * gate_up_exps_b) const {
+         ggml_tensor * gate_up_exps_b,
+         ggml_tensor * up_exps_s,
+         ggml_tensor * gate_exps_s,
+         ggml_tensor * down_exps_s) const {
     const int64_t n_embd   = cur->ne[0];
     const int64_t n_tokens = cur->ne[1];
     const bool weight_before_ffn = arch == LLM_ARCH_LLAMA4; // for llama4, we apply the sigmoid-ed weights before the FFN
@@ -1358,6 +1373,15 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             cb(gate_up, "ffn_moe_gate_up_biased", il);
         }
 
+        // apply per-expert scale2 to merged gate_up (use up_exps_s since gate and up are fused)
+        if (up_exps_s) {
+            ggml_tensor * s = ggml_reshape_3d(ctx0, up_exps_s, 1, n_expert, 1);
+            s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
+            s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
+            gate_up = ggml_mul(ctx0, gate_up, s);
+            cb(gate_up, "ffn_moe_gate_up_scaled", il);
+        }
+
         const int64_t n_ff = gate_up->ne[0] / 2;
         cur = ggml_view_3d(ctx0, gate_up, n_ff, gate_up->ne[1], gate_up->ne[2], gate_up->nb[1], gate_up->nb[2], 0);
         cb(cur, "ffn_moe_gate", il);
@@ -1373,6 +1397,15 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             cb(up, "ffn_moe_up_biased", il);
         }
 
+        // apply per-expert scale2 to up
+        if (up_exps_s) {
+            ggml_tensor * s = ggml_reshape_3d(ctx0, up_exps_s, 1, n_expert, 1);
+            s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
+            s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
+            up = ggml_mul(ctx0, up, s);
+            cb(up, "ffn_moe_up_scaled", il);
+        }
+
         if (gate_exps) {
             cur = build_lora_mm_id(gate_exps, cur, selected_experts); // [n_ff, n_expert_used, n_tokens]
             cb(cur, "ffn_moe_gate", il);
@@ -1384,6 +1417,15 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             cur = ggml_add_id(ctx0, cur, gate_exps_b, selected_experts);
             cb(cur, "ffn_moe_gate_biased", il);
         }
+
+        // apply per-expert scale2 to gate
+        if (gate_exps_s) {
+            ggml_tensor * s = ggml_reshape_3d(ctx0, gate_exps_s, 1, n_expert, 1);
+            s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
+            s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
+            cur = ggml_mul(ctx0, cur, s);
+            cb(cur, "ffn_moe_gate_scaled", il);
+        }
     }
 
     const bool has_gate = gate_exps || gate_up_exps;
@@ -1463,6 +1505,15 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
         cb(experts, "ffn_moe_down_biased", il);
     }
 
+    // apply per-expert scale2 to down
+    if (down_exps_s) {
+        ggml_tensor * s = ggml_reshape_3d(ctx0, down_exps_s, 1, n_expert, 1);
+        s = ggml_repeat_4d(ctx0, s, 1, n_expert, n_tokens, 1);
+        s = ggml_get_rows(ctx0, s, selected_experts); // [1, n_expert_used, n_tokens]
+        experts = ggml_mul(ctx0, experts, s);
+        cb(experts, "ffn_moe_down_scaled", il);
+    }
+
     if (!weight_before_ffn) {
         experts = ggml_mul(ctx0, experts, weights);
         cb(cur, "ffn_moe_weighted", il);

src/llama-graph.h

@@ -764,10 +764,11 @@ struct llm_graph_context {
              ggml_tensor * cur,
                      int   il) const;
 
-    // do mat_mul, while optionally apply lora
+    // do mat_mul, while optionally apply lora and per-tensor scale
     ggml_tensor * build_lora_mm(
               ggml_tensor * w,
-              ggml_tensor * cur) const;
+              ggml_tensor * cur,
+              ggml_tensor * w_s = nullptr) const;
 
     // do mat_mul_id, while optionally apply lora
     ggml_tensor * build_lora_mm_id(
@@ -814,7 +815,10 @@ struct llm_graph_context {
             llama_expert_gating_func_type gating_op,
                      int   il,
              ggml_tensor * probs_in = nullptr,
-             ggml_tensor * gate_up_exps = nullptr) const;
+             ggml_tensor * gate_up_exps = nullptr,
+             ggml_tensor * up_exps_s = nullptr,
+             ggml_tensor * gate_exps_s = nullptr,
+             ggml_tensor * down_exps_s = nullptr) const;
 
     ggml_tensor * build_moe_ffn(
              ggml_tensor * cur,
@@ -836,7 +840,10 @@ struct llm_graph_context {
                      int   il,
              ggml_tensor * probs_in = nullptr,
              ggml_tensor * gate_up_exps = nullptr,
-             ggml_tensor * gate_up_exps_b = nullptr) const;
+             ggml_tensor * gate_up_exps_b = nullptr,
+             ggml_tensor * up_exps_s = nullptr,
+             ggml_tensor * gate_exps_s = nullptr,
+             ggml_tensor * down_exps_s = nullptr) const;
 
     //
     // inputs

src/llama-hparams.h

@@ -89,6 +89,7 @@ struct llama_hparams {
     bool     expert_weights_norm  = false;
     uint32_t expert_gating_func   = LLAMA_EXPERT_GATING_FUNC_TYPE_NONE;
     uint32_t moe_every_n_layers   = 0;
+    uint32_t moe_latent_size      = 0;
     uint32_t nextn_predict_layers = 0;
 
     float f_norm_eps;

src/llama-impl.h

@@ -70,6 +70,6 @@ std::string llama_format_tensor_shape(const struct ggml_tensor * t);
 
 std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i);
 
-#define LLAMA_TENSOR_NAME_FATTN  "__fattn__"
-#define LLAMA_TENSOR_NAME_FGDNAR "__fgdnar__"
-#define LLAMA_TENSOR_NAME_FGDNCH "__fgdnch__"
+#define LLAMA_TENSOR_NAME_FATTN   "__fattn__"
+#define LLAMA_TENSOR_NAME_FGDN_AR "__fgdn_ar__"
+#define LLAMA_TENSOR_NAME_FGDN_CH "__fgdn_ch__"

src/llama-model-loader.cpp

@@ -42,6 +42,7 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q5_1:     return "Q5_1";
         case LLAMA_FTYPE_MOSTLY_Q8_0:     return "Q8_0";
         case LLAMA_FTYPE_MOSTLY_MXFP4_MOE: return "MXFP4 MoE";
+        case LLAMA_FTYPE_MOSTLY_NVFP4:    return "NVFP4";
         case LLAMA_FTYPE_MOSTLY_Q2_K:     return "Q2_K - Medium";
         case LLAMA_FTYPE_MOSTLY_Q2_K_S:   return "Q2_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:   return "Q3_K - Small";
@@ -724,6 +725,7 @@ llama_model_loader::llama_model_loader(
             case GGML_TYPE_IQ4_NL:  ftype = LLAMA_FTYPE_MOSTLY_IQ4_NL;  break;
             case GGML_TYPE_IQ4_XS:  ftype = LLAMA_FTYPE_MOSTLY_IQ4_XS;  break;
             case GGML_TYPE_IQ3_S:   ftype = LLAMA_FTYPE_MOSTLY_IQ3_S;   break;
+            case GGML_TYPE_NVFP4:   ftype = LLAMA_FTYPE_MOSTLY_NVFP4;   break;
             default:
                 {
                     LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));

src/llama-model.cpp

@@ -135,6 +135,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_100B_A6B:      return "100B.A6B";
         case LLM_TYPE_102B_A12B:     return "102B.A12B";
         case LLM_TYPE_106B_A12B:     return "106B.A12B";
+        case LLM_TYPE_120B_A12B:     return "120B.A12B";
         case LLM_TYPE_122B_A10B:     return "122B.A10B";
         case LLM_TYPE_196B_A11B:     return "196B.A11B";
         case LLM_TYPE_230B_A10B:     return "230B.A10B";
@@ -1861,10 +1862,12 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
                 ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
                 ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
+                ml.get_key(LLM_KV_MOE_LATENT_SIZE,                   hparams.moe_latent_size, false);
 
                 switch (hparams.n_layer) {
                     case 52: type = LLM_TYPE_31B_A3_5B; break; // Nemotron-H_MOE 31B
                     case 56: type = LLM_TYPE_9B; break;
+                    case 88: type = LLM_TYPE_120B_A12B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -5007,23 +5010,23 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.attn_sub_norm = create_tensor(tn(LLM_TENSOR_ATTN_SUB_NORM, "weight", i), {n_embd}, 0);
 
                         layer.wq       = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd}, 0);
-                        layer.wq_scale = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+                        layer.wq_s     = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
                         layer.wk       = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
-                        layer.wk_scale = create_tensor(tn(LLM_TENSOR_ATTN_K,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+                        layer.wk_s     = create_tensor(tn(LLM_TENSOR_ATTN_K,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
                         layer.wv       = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
-                        layer.wv_scale = create_tensor(tn(LLM_TENSOR_ATTN_V,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+                        layer.wv_s     = create_tensor(tn(LLM_TENSOR_ATTN_V,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
                         layer.wo       = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
-                        layer.wo_scale = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+                        layer.wo_s     = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
 
                         layer.ffn_norm     = create_tensor(tn(LLM_TENSOR_FFN_NORM,     "weight", i), {n_embd}, 0);
                         layer.ffn_sub_norm = create_tensor(tn(LLM_TENSOR_FFN_SUB_NORM, "weight", i), {n_ff}, 0);
 
                         layer.ffn_gate       = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
-                        layer.ffn_gate_scale = create_tensor(tn(LLM_TENSOR_FFN_GATE, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+                        layer.ffn_gate_s = create_tensor(tn(LLM_TENSOR_FFN_GATE, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
                         layer.ffn_down       = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
-                        layer.ffn_down_scale = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+                        layer.ffn_down_s = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
                         layer.ffn_up         = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
-                        layer.ffn_up_scale   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+                        layer.ffn_up_s   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
                     }
                 } break;
             case LLM_ARCH_T5:
@@ -5544,6 +5547,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                     const int64_t n_ssm_head = hparams.ssm_dt_rank;
                     const int64_t n_group    = hparams.ssm_n_group;
                     const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
+                    const int64_t moe_n_embd = hparams.moe_latent_size > 0 ? hparams.moe_latent_size : n_embd;
 
                     // embeddings
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -5603,8 +5607,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                                 layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, 0);
 
                                 // MoE branch
-                                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
-                                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+                                layer.ffn_latent_down = create_tensor(tn(LLM_TENSOR_FFN_LATENT_DOWN, "weight", i), {n_embd, moe_n_embd}, TENSOR_NOT_REQUIRED);
+                                layer.ffn_latent_up   = create_tensor(tn(LLM_TENSOR_FFN_LATENT_UP,   "weight", i), {moe_n_embd, n_embd}, TENSOR_NOT_REQUIRED);
+
+                                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   moe_n_embd, n_expert}, 0);
+                                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {moe_n_embd, n_ff_exp, n_expert}, 0);
 
                                 // Shared expert branch
                                 layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
@@ -7436,6 +7443,48 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             default:
                 throw std::runtime_error("unknown architecture");
         }
+
+        // generic pass: load optional per-tensor/per-expert ".scale" tensors (e.g. NVFP4 scale2)
+        // this avoids having to add scale loading to every architecture
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = layers[i];
+
+            // attention weight scales (per-tensor, shape {1})
+            if (!layer.wq_s && layer.wq) {
+                layer.wq_s = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.wk_s && layer.wk) {
+                layer.wk_s = create_tensor(tn(LLM_TENSOR_ATTN_K,   "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.wv_s && layer.wv) {
+                layer.wv_s = create_tensor(tn(LLM_TENSOR_ATTN_V,   "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.wo_s && layer.wo) {
+                layer.wo_s = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+
+            // dense FFN weight scales (per-tensor, shape {1})
+            if (!layer.ffn_gate_s && layer.ffn_gate) {
+                layer.ffn_gate_s = create_tensor(tn(LLM_TENSOR_FFN_GATE, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.ffn_down_s && layer.ffn_down) {
+                layer.ffn_down_s = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.ffn_up_s && layer.ffn_up) {
+                layer.ffn_up_s = create_tensor(tn(LLM_TENSOR_FFN_UP, "scale", i), {1}, TENSOR_NOT_REQUIRED);
+            }
+
+            // MoE expert weight scales (per-expert, shape {n_expert})
+            if (!layer.ffn_gate_exps_s && layer.ffn_gate_exps) {
+                layer.ffn_gate_exps_s = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "scale", i), {n_expert}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.ffn_down_exps_s && layer.ffn_down_exps) {
+                layer.ffn_down_exps_s = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "scale", i), {n_expert}, TENSOR_NOT_REQUIRED);
+            }
+            if (!layer.ffn_up_exps_s && layer.ffn_up_exps) {
+                layer.ffn_up_exps_s = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "scale", i), {n_expert}, TENSOR_NOT_REQUIRED);
+            }
+        }
     }
 
     ml.done_getting_tensors();

src/llama-model.h

@@ -126,6 +126,7 @@ enum llm_type {
     LLM_TYPE_100B_A6B,
     LLM_TYPE_102B_A12B, // Solar-Open
     LLM_TYPE_106B_A12B, // GLM-4.5-Air
+    LLM_TYPE_120B_A12B, // Nemotron 3 Super
     LLM_TYPE_122B_A10B, // Qwen3.5
     LLM_TYPE_196B_A11B, // Step3.5-Flash
     LLM_TYPE_230B_A10B, // Minimax M2
@@ -294,6 +295,15 @@ struct llama_layer {
     struct ggml_tensor * ffn_up_exps_b     = nullptr;
     struct ggml_tensor * ffn_gate_up_exps_b = nullptr;
 
+    // ff MoE per-expert scales (NVFP4 per-tensor scale2)
+    struct ggml_tensor * ffn_gate_exps_s   = nullptr;
+    struct ggml_tensor * ffn_down_exps_s   = nullptr;
+    struct ggml_tensor * ffn_up_exps_s     = nullptr;
+
+    // ff MoE latent proj
+    struct ggml_tensor * ffn_latent_down = nullptr;
+    struct ggml_tensor * ffn_latent_up   = nullptr;
+
     // ff shared expert (shexp)
     struct ggml_tensor * ffn_gate_inp_shexp = nullptr;
     struct ggml_tensor * ffn_gate_shexp     = nullptr;
@@ -387,13 +397,13 @@ struct llama_layer {
     struct ggml_tensor * rope_freqs = nullptr;
 
     // bitnet scale
-    struct ggml_tensor * wq_scale       = nullptr;
-    struct ggml_tensor * wk_scale       = nullptr;
-    struct ggml_tensor * wv_scale       = nullptr;
-    struct ggml_tensor * wo_scale       = nullptr;
-    struct ggml_tensor * ffn_gate_scale = nullptr;
-    struct ggml_tensor * ffn_up_scale   = nullptr;
-    struct ggml_tensor * ffn_down_scale = nullptr;
+    struct ggml_tensor * wq_s       = nullptr;
+    struct ggml_tensor * wk_s       = nullptr;
+    struct ggml_tensor * wv_s       = nullptr;
+    struct ggml_tensor * wo_s       = nullptr;
+    struct ggml_tensor * ffn_gate_s = nullptr;
+    struct ggml_tensor * ffn_up_s   = nullptr;
+    struct ggml_tensor * ffn_down_s = nullptr;
 
     // altup & laurel
     struct ggml_tensor * per_layer_inp_gate   = nullptr;

src/models/bitnet.cpp

@@ -29,10 +29,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
         // self-attention
         {
             // compute Q and K and RoPE them
-            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-            if (model.layers[il].wq_scale) {
-                Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale);
-            }
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
             cb(Qcur, "Qcur", il);
             if (model.layers[il].bq) {
                 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
@@ -40,10 +37,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
             }
 
             // B1.K
-            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-            if (model.layers[il].wk_scale) {
-                Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale);
-            }
+            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur, model.layers[il].wk_s);
             cb(Kcur, "Kcur", il);
             if (model.layers[il].bk) {
                 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
@@ -51,10 +45,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
             }
 
             // B1.V
-            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-            if (model.layers[il].wv_scale) {
-                Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale);
-            }
+            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur, model.layers[il].wv_s);
             cb(Vcur, "Vcur", il);
             if (model.layers[il].bv) {
                 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
@@ -90,10 +81,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
                     LLM_NORM_RMS, il);
             cb(cur, "attn_sub_norm", il);
 
-            cur = build_lora_mm(model.layers[il].wo, cur);
-            if (model.layers[il].wo_scale) {
-                cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale);
-            }
+            cur = build_lora_mm(model.layers[il].wo, cur, model.layers[il].wo_s);
             if (model.layers[il].bo) {
                 cur = ggml_add(ctx0, cur, model.layers[il].bo);
             }
@@ -115,8 +103,8 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
         cb(cur, "ffn_norm", il);
 
         cur = build_ffn(cur,
-                model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_scale,
-                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale,
+                model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_s,
+                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
                 NULL,                      NULL, NULL,
                 NULL,
                 LLM_FFN_SILU, LLM_FFN_PAR, il);
@@ -127,10 +115,7 @@ llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_pa
                 LLM_NORM_RMS, il);
         cb(cur, "ffn_sub_norm", il);
 
-        cur = build_lora_mm(model.layers[il].ffn_down, cur);
-        if (model.layers[il].ffn_down_scale) {
-            cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale);
-        }
+        cur = build_lora_mm(model.layers[il].ffn_down, cur, model.layers[il].ffn_down_s);
         cb(cur, "ffn_down", il);
 
         cur = ggml_add(ctx0, cur, ffn_inp);

src/models/delta-net-base.cpp

@@ -41,13 +41,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     GGML_ASSERT(b->ne[0] == 1   && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
     GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v      && s->ne[3] == n_seqs);
 
-    if (cparams.fused_gdn_ch) {
-        //ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
-        //cb(result, LLAMA_TENSOR_NAME_FGDNCH, il);
-
-        GGML_ABORT("not implemented yet");
-    }
-
     const float scale = 1.0f / sqrtf(S_k);
 
     q = ggml_scale(ctx0, q, scale);
@@ -325,26 +318,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
     GGML_ASSERT(b->ne[0] == 1   && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
     GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v      && s->ne[3] == n_seqs);
 
-    if (cparams.fused_gdn_ar) {
-        ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
-        cb(result, LLAMA_TENSOR_NAME_FGDNAR, il);
-
-        ggml_tensor * output = ggml_view_4d(ctx0, result,
-            S_v, H_v, n_tokens, n_seqs,
-            ggml_row_size(result->type, S_v),
-            ggml_row_size(result->type, S_v * H_v),
-            ggml_row_size(result->type, S_v * H_v * n_tokens), 0);
-
-        ggml_tensor * new_state = ggml_view_4d(ctx0, result,
-            S_v, S_v, H_v, n_seqs,
-            ggml_row_size(result->type, S_v),
-            ggml_row_size(result->type, S_v * S_v),
-            ggml_row_size(result->type, S_v * S_v * H_v),
-            ggml_row_size(result->type, S_v * H_v * n_tokens * n_seqs));
-
-        return {output, new_state};
-    }
-
     const float scale = 1.0f / sqrtf(S_k);
 
     q = ggml_scale(ctx0, q, scale);
@@ -401,3 +374,78 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
 
     return {o, s};
 }
+
+std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net_fused(
+        ggml_tensor * q,
+        ggml_tensor * k,
+        ggml_tensor * v,
+        ggml_tensor * g,
+        ggml_tensor * b,
+        ggml_tensor * s,
+        int           il) {
+    const int64_t S_k      = q->ne[0];
+    const int64_t H_k      = q->ne[1];
+    const int64_t n_tokens = q->ne[2];
+    const int64_t n_seqs   = q->ne[3];
+
+    const int64_t S_v = v->ne[0];
+    const int64_t H_v = v->ne[1];
+
+    GGML_ASSERT(S_k == S_v);
+    GGML_ASSERT(H_v % H_k == 0);
+
+    GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
+    GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
+    GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
+
+    GGML_ASSERT(g->ne[0] == 1   || g->ne[0] == S_v);
+    GGML_ASSERT(                   g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
+    GGML_ASSERT(b->ne[0] == 1   && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
+    GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v      && s->ne[3] == n_seqs);
+
+    ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
+    if (n_tokens == 1) {
+        cb(result, LLAMA_TENSOR_NAME_FGDN_AR, il);
+    } else {
+        cb(result, LLAMA_TENSOR_NAME_FGDN_CH, il);
+    }
+
+    ggml_tensor * output = ggml_view_4d(ctx0, result,
+            S_v, H_v, n_tokens, n_seqs,
+            ggml_row_size(result->type, S_v),
+            ggml_row_size(result->type, S_v * H_v),
+            ggml_row_size(result->type, S_v * H_v * n_tokens), 0);
+
+    ggml_tensor * new_state = ggml_view_4d(ctx0, result,
+            S_v, S_v, H_v, n_seqs,
+            ggml_row_size(result->type, S_v),
+            ggml_row_size(result->type, S_v * S_v),
+            ggml_row_size(result->type, S_v * S_v * H_v),
+            ggml_row_size(result->type, S_v * H_v * n_tokens * n_seqs));
+
+    return {output, new_state};
+}
+
+std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net(
+        ggml_tensor * q,
+        ggml_tensor * k,
+        ggml_tensor * v,
+        ggml_tensor * g,
+        ggml_tensor * b,
+        ggml_tensor * s,
+        int           il) {
+    const int64_t n_seq_tokens = q->ne[2];
+
+    if (n_seq_tokens == 1) {
+        if (cparams.fused_gdn_ar) {
+            return build_delta_net_fused(q, k, v, g, b, s, il);
+        }
+        return build_delta_net_autoregressive(q, k, v, g, b, s, il);
+    }
+
+    if (cparams.fused_gdn_ch) {
+        return build_delta_net_fused(q, k, v, g, b, s, il);
+    }
+
+    return build_delta_net_chunking(q, k, v, g, b, s, il);
+}

src/models/kimi-linear.cpp

@@ -169,9 +169,7 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
             Kcur = ggml_l2_norm(ctx0, Kcur, eps_norm);
 
             // Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens
-            std::pair<ggml_tensor *, ggml_tensor *> attn_out = n_seq_tokens == 1 ?
-                build_delta_net_autoregressive(Qcur, Kcur, Vcur, g1, beta, state, il) :
-                build_delta_net_chunking(Qcur, Kcur, Vcur, g1, beta, state, il);
+            auto attn_out = build_delta_net(Qcur, Kcur, Vcur, g1, beta, state, il);
 
             ggml_tensor * output = ggml_cont(ctx0, attn_out.first);
             ggml_tensor * new_state = attn_out.second;

src/models/llama.cpp

@@ -43,19 +43,19 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
             ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
 
             // compute Q and K and RoPE them
-            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
             cb(Qcur, "Qcur", il);
             if (model.layers[il].bq) {
                 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
                 cb(Qcur, "Qcur", 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);
             if (model.layers[il].bk) {
                 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
                 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);
             if (model.layers[il].bv) {
                 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
@@ -91,6 +91,9 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
             cur = build_attn(inp_attn,
                     model.layers[il].wo, model.layers[il].bo,
                     Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            if (model.layers[il].wo_s) {
+                cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);
+            }
             cb(cur, "attn_out", il);
         }
         if (il == n_layer - 1 && inp_out_ids) {
@@ -109,9 +112,9 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
             cb(cur, "ffn_norm", il);
 
             cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   model.layers[il].ffn_up_s,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, model.layers[il].ffn_gate_s,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, model.layers[il].ffn_down_s,
                     NULL,
                     LLM_FFN_SILU, LLM_FFN_PAR, il);
             cb(cur, "ffn_out", il);
@@ -132,7 +135,11 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
                     LLM_FFN_SILU, true,
                     hparams.expert_weights_scale,
                     LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    il);
+                    il,
+                    nullptr, nullptr,
+                    model.layers[il].ffn_up_exps_s,
+                    model.layers[il].ffn_gate_exps_s,
+                    model.layers[il].ffn_down_exps_s);
             cb(cur, "ffn_moe_out", il);
         }
         cur = ggml_add(ctx0, cur, ffn_inp);

src/models/models.h

@@ -44,6 +44,26 @@ struct llm_build_delta_net_base : public llm_graph_context {
                 ggml_tensor * b,
                 ggml_tensor * s,
                 int           il);
+
+    // use the ggml_gated_delta_net fused operator
+    std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_fused(
+                ggml_tensor * q,
+                ggml_tensor * k,
+                ggml_tensor * v,
+                ggml_tensor * g,
+                ggml_tensor * b,
+                ggml_tensor * s,
+                        int   il);
+
+    // choose one of two implementations above based on the number of tokens
+    std::pair<ggml_tensor *, ggml_tensor *> build_delta_net(
+                ggml_tensor * q,
+                ggml_tensor * k,
+                ggml_tensor * v,
+                ggml_tensor * g,
+                ggml_tensor * b,
+                ggml_tensor * s,
+                        int   il);
 };
 
 struct llm_build_rwkv6_base : public llm_graph_context {

src/models/nemotron-h.cpp

@@ -114,9 +114,18 @@ ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const lla
                 LLM_FFN_RELU_SQR, LLM_FFN_PAR, il);
         cb(cur, "ffn_out", il);
     } else {
-        ggml_tensor * ffn_inp = cur;
+        ggml_tensor * inp_emb    = cur;
+        ggml_tensor * inp_latent = cur;
+
+        if (model.layers[il].ffn_latent_down) {
+            inp_latent = ggml_mul_mat(ctx0, model.layers[il].ffn_latent_down, cur);
+        }
+
+        ggml_tensor * router_logits = build_lora_mm(model.layers[il].ffn_gate_inp, cur);
+        cb(router_logits, "ffn_moe_logits", il);
+
         ggml_tensor * moe_out =
-            build_moe_ffn(ffn_inp,
+            build_moe_ffn(inp_latent,
                     model.layers[il].ffn_gate_inp,
                     model.layers[il].ffn_up_exps,
                     nullptr, // no gate
@@ -126,10 +135,15 @@ ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const lla
                     LLM_FFN_RELU_SQR, hparams.expert_weights_norm,
                     hparams.expert_weights_scale,
                     LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
-                    il);
+                    il,
+                    router_logits);
         cb(moe_out, "ffn_moe_out", il);
 
-        ggml_tensor * ffn_shexp = build_ffn(ffn_inp,
+        if (model.layers[il].ffn_latent_up) {
+            moe_out = ggml_mul_mat(ctx0, model.layers[il].ffn_latent_up, moe_out);
+        }
+
+        ggml_tensor * ffn_shexp = build_ffn(inp_emb,
                     model.layers[il].ffn_up_shexp,  NULL, NULL,
                     NULL /* no gate */           ,  NULL, NULL,
                     model.layers[il].ffn_down_shexp, NULL, NULL,

src/models/qwen3.cpp

@@ -30,13 +30,13 @@ llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_para
         // self-attention
         {
             // compute Q and K and RoPE them
-            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
             cb(Qcur, "Qcur", 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);
 
             Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
@@ -68,6 +68,9 @@ llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_para
             cur = build_attn(inp_attn,
                     model.layers[il].wo, model.layers[il].bo,
                     Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            if (model.layers[il].wo_s) {
+                cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);
+            }
         }
         if (il == n_layer - 1 && inp_out_ids) {
             cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
@@ -83,9 +86,9 @@ llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_para
         cb(cur, "ffn_norm", il);
 
         cur = build_ffn(cur,
-                model.layers[il].ffn_up,   NULL, NULL,
-                model.layers[il].ffn_gate, NULL, NULL,
-                model.layers[il].ffn_down, NULL, NULL,
+                model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_s,
+                model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
+                model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,
                 NULL,
                 LLM_FFN_SILU, LLM_FFN_PAR, il);
         cb(cur, "ffn_out", il);

src/models/qwen35.cpp

@@ -321,9 +321,9 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
 
     // if head keys and value keys are different, repeat to force tensors into matching shapes
-    if (num_k_heads != num_v_heads) {
+    // note: need explicit repeat only if we are not using the fused GDN
+    if (num_k_heads != num_v_heads && (!cparams.fused_gdn_ar || !cparams.fused_gdn_ch)) {
         GGML_ASSERT(num_v_heads % num_k_heads == 0);
-        // TODO: try to avoid these explicit repeats by utilizing op broadcast
         q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
         k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
     }
@@ -332,12 +332,8 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
     cb(k_conv, "k_conv_predelta", il);
     cb(v_conv, "v_conv_predelta", il);
 
-    std::pair<ggml_tensor *, ggml_tensor *> attn_out;
-    if (n_seq_tokens == 1) {
-        attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
-    } else {
-        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
-    }
+    auto attn_out = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state, il);
+
     ggml_tensor * output    = attn_out.first;
     ggml_tensor * new_state = attn_out.second;
     cb(output, "attn_output", il);

src/models/qwen35moe.cpp

@@ -321,9 +321,9 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
 
     // if head keys and value keys are different, repeat to force tensors into matching shapes
-    if (num_k_heads != num_v_heads) {
+    // note: need explicit repeat only if we are not using the fused GDN
+    if (num_k_heads != num_v_heads && (!cparams.fused_gdn_ar || !cparams.fused_gdn_ch)) {
         GGML_ASSERT(num_v_heads % num_k_heads == 0);
-        // TODO: try to avoid these explicit repeats by utilizing op broadcast
         q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
         k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
     }
@@ -332,12 +332,8 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
     cb(k_conv, "k_conv_predelta", il);
     cb(v_conv, "v_conv_predelta", il);
 
-    std::pair<ggml_tensor *, ggml_tensor *> attn_out;
-    if (n_seq_tokens == 1) {
-        attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
-    } else {
-        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
-    }
+    auto attn_out = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state, il);
+
     ggml_tensor * output    = attn_out.first;
     ggml_tensor * new_state = attn_out.second;
     cb(output, "attn_output", il);

src/models/qwen3moe.cpp

@@ -30,13 +30,13 @@ llm_build_qwen3moe::llm_build_qwen3moe(const llama_model & model, const llm_grap
         // self_attention
         {
             // compute Q and K and RoPE them
-            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
+            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur, model.layers[il].wq_s);
             cb(Qcur, "Qcur", 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);
 
             Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
@@ -68,6 +68,9 @@ llm_build_qwen3moe::llm_build_qwen3moe(const llama_model & model, const llm_grap
             cur = build_attn(inp_attn,
                     model.layers[il].wo, model.layers[il].bo,
                     Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+            if (model.layers[il].wo_s) {
+                cur = ggml_mul(ctx0, cur, model.layers[il].wo_s);
+            }
         }
         if (il == n_layer - 1 && inp_out_ids) {
             cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
@@ -93,7 +96,11 @@ llm_build_qwen3moe::llm_build_qwen3moe(const llama_model & model, const llm_grap
                     LLM_FFN_SILU, true,
                     hparams.expert_weights_scale,
                     LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
-                    il);
+                    il,
+                    nullptr, nullptr,
+                    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);
         cur = moe_out;
 

src/models/qwen3next.cpp

@@ -406,6 +406,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
 
     // if head keys and value keys are different, repeat to force tensors into matching shapes
+    // TODO: avoid repeats for fused GDN, needs broadcast configuration for GDN op [TAG_GGML_GDN_BCAST]
     if (num_k_heads != num_v_heads) {
         GGML_ASSERT(num_v_heads % num_k_heads == 0);
         int64_t repeat_factor = num_v_heads / num_k_heads;
@@ -431,13 +432,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
     cb(k_conv, "k_conv_predelta", il);
     cb(v_conv, "v_conv_predelta", il);
 
-    // Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens
-    std::pair<ggml_tensor *, ggml_tensor *> attn_out; // pair of (output, new_state)
-    if (n_seq_tokens == 1) {
-        attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
-    } else {
-        attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
-    }
+    auto attn_out = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state, il);
+
     ggml_tensor * output    = attn_out.first;
     ggml_tensor * new_state = attn_out.second;
     cb(output, "attn_output", il);

tests/CMakeLists.txt

@@ -149,6 +149,7 @@ endif ()
 if (NOT WIN32 OR NOT BUILD_SHARED_LIBS)
     # these tests are disabled on Windows because they use internal functions not exported with LLAMA_API (when building with shared libraries)
     llama_build_and_test(test-sampling.cpp)
+    llama_build_and_test(test-reasoning-budget.cpp)
     llama_build_and_test(test-grammar-parser.cpp)
     llama_build_and_test(test-grammar-integration.cpp)
     llama_build_and_test(test-llama-grammar.cpp)

tests/test-backend-ops.cpp

@@ -7854,10 +7854,6 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 1056, 1, 67,  {1,  1}, {4, 1}, {0, 2, 1, 3}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 16, 32, 32, { 1,  1}, {1, 1}, {0, 1, 2, 3}, 64, 3));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F32, GGML_TYPE_F32, 64, 77, 77, {12,1}, {1,1}));
-    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 2, 1, 3, {128, 1024}, {1, 1}));
-    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 2, 3, 4, {128, 1024}, {1, 1}));
-    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 2, 1, 3, {128*1024, 1}, {1, 1}, {0, 2, 1, 3}));
-    test_cases.emplace_back(new test_mul_mat(GGML_TYPE_F16, GGML_TYPE_F32, 2, 1, 3, {128*1024, 1}, {1, 1}, {0, 1, 2, 3}, 64));
 
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 576, 512, 576, {1,1}, {1,1}));
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 1, 2048, 8192, {1,  1}, {1, 1}));
@@ -8451,6 +8447,9 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     }
 
     test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 32, 128, 1, 1));
+    test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 32, 16, 1, 1));
+    test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 32, 16, 1, 1, 1, true, true));
+    test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 32, 16, 1, 1, 1, false, true));
     test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 16, 64, 1, 2));
     test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 4, 64, 4, 1));
     test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 4, 64, 4, 2));
@@ -8460,10 +8459,12 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     // KDA (vector gate)
     test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 4, 64, 1, 1, 1, false, true));
     test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 4, 64, 1, 2, 1, false, true));
+    test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 4, 16, 1, 2, 1, false, true));
     test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 4, 32, 4, 1, 1, false, true));
     test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 4, 64, 4, 2, 1, false, true));
     test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 8, 32, 4, 2, 2, false, true));
     test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 4, 64, 4, 2, 1, true,  true));
+    test_cases.emplace_back(new test_gated_delta_net(GGML_TYPE_F32, 4, 16, 4, 2, 1, true,  true));
 
 #if 0
     // these tests are disabled to save execution time, sbut they can be handy for debugging

tests/test-chat.cpp

@@ -2765,6 +2765,42 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
             .run();
 
     }
+
+    // GigaChat V3
+    {
+        auto tst = peg_tester("models/templates/GigaChat3-10B-A1.8B.jinja", detailed_debug);
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("<|message_sep|>\n\nfunction call<|role_sep|>\n{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}")
+            .tools({ special_function_tool })
+            .expect(message_assist_call)
+            .run();
+
+        tst.test(
+            "Hello, world!\nWhat's up?"
+            "<|message_sep|>\n\nfunction call<|role_sep|>\n{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}"
+        )
+            .tools({ special_function_tool })
+            .expect(message_assist_call_content)
+            .run();
+    }
+
+    // GigaChat V3.1
+    {
+        auto tst = peg_tester("models/templates/GigaChat3.1-10B-A1.8B.jinja", detailed_debug);
+        tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
+        tst.test("<|function_call|>{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}")
+            .tools({ special_function_tool })
+            .expect(message_assist_call)
+            .run();
+
+        tst.test(
+            "Hello, world!\nWhat's up?"
+            "<|function_call|>{\"name\": \"special_function\", \"arguments\": {\"arg1\": 1}}"
+        )
+            .tools({ special_function_tool })
+            .expect(message_assist_call_content)
+            .run();
+    }
 }
 
 // Test the developer role to system workaround with a simple mock template

tests/test-quantize-fns.cpp

@@ -20,8 +20,10 @@ constexpr float MAX_QUANTIZATION_TOTAL_ERROR_TERNARY = 0.01f;
 constexpr float MAX_QUANTIZATION_TOTAL_ERROR_2BITS = 0.0075f;
 constexpr float MAX_QUANTIZATION_TOTAL_ERROR_3BITS = 0.0040f;
 constexpr float MAX_QUANTIZATION_TOTAL_ERROR_3BITS_XXS = 0.0050f;
+constexpr float MAX_QUANTIZATION_TOTAL_ERROR_FP4 = 0.0030f;
 constexpr float MAX_DOT_PRODUCT_ERROR = 0.02f;
 constexpr float MAX_DOT_PRODUCT_ERROR_LOWBIT = 0.04f;
+constexpr float MAX_DOT_PRODUCT_ERROR_FP4 = 0.03f;
 constexpr float MAX_DOT_PRODUCT_ERROR_TERNARY = 0.15f;
 
 static const char* RESULT_STR[] = {"ok", "FAILED"};
@@ -149,7 +151,8 @@ int main(int argc, char * argv[]) {
                 type == GGML_TYPE_IQ2_S   ? MAX_QUANTIZATION_TOTAL_ERROR_2BITS :
                 type == GGML_TYPE_Q3_K    ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS :
                 type == GGML_TYPE_IQ3_S   ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS :
-                type == GGML_TYPE_IQ3_XXS ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS_XXS : MAX_QUANTIZATION_TOTAL_ERROR;
+                type == GGML_TYPE_IQ3_XXS ? MAX_QUANTIZATION_TOTAL_ERROR_3BITS_XXS :
+                type == GGML_TYPE_NVFP4   ? MAX_QUANTIZATION_TOTAL_ERROR_FP4 : MAX_QUANTIZATION_TOTAL_ERROR;
             failed = !(total_error < max_quantization_error);
             num_failed += failed;
             if (failed || verbose) {
@@ -169,6 +172,8 @@ int main(int argc, char * argv[]) {
                                           ? MAX_DOT_PRODUCT_ERROR_LOWBIT
                                           : type == GGML_TYPE_TQ1_0 || type == GGML_TYPE_TQ2_0
                                           ? MAX_DOT_PRODUCT_ERROR_TERNARY
+                                          : type == GGML_TYPE_NVFP4
+                                          ? MAX_DOT_PRODUCT_ERROR_FP4
                                           : MAX_DOT_PRODUCT_ERROR;
             failed = !(vec_dot_error < max_allowed_error);
             num_failed += failed;

tests/test-reasoning-budget.cpp

@@ -0,0 +1,238 @@
+#include "reasoning-budget.h"
+#include "unicode.h"
+
+#include "llama.h"
+#include "ggml.h"
+
+#ifdef NDEBUG
+#undef NDEBUG
+#endif
+
+#include <cmath>
+#include <cstddef>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+// Reasoning budget sampler test helper
+// These tests use nullptr vocab which safely falls back to treating all tokens as complete
+// (The UTF-8 boundary detection logic is tested separately in test_utf8_boundary_detection)
+static void test_reasoning_budget(
+    const char * test_name,
+    const std::vector<llama_token> & sequence,
+    const std::vector<llama_token> & start_tokens,
+    const std::vector<llama_token> & end_tokens,
+    const std::vector<llama_token> & forced_tokens,
+    int32_t budget,
+    common_reasoning_budget_state initial_state,
+    size_t expected_force_start,   // token index where forcing should start (SIZE_MAX = never)
+    size_t expected_force_end      // token index where forcing should end (after this, no more forcing)
+) {
+    // Find the maximum token ID to ensure our vocab covers all tokens
+    llama_token max_token = 0;
+    for (auto t : sequence) max_token = std::max(max_token, t);
+    for (auto t : start_tokens) max_token = std::max(max_token, t);
+    for (auto t : end_tokens) max_token = std::max(max_token, t);
+    for (auto t : forced_tokens) max_token = std::max(max_token, t);
+
+    // Create a minimal sampler with mock vocabulary
+    // For this test, we use nullptr as vocab since we're testing state transitions
+    // The UTF-8 boundary check will treat all tokens as complete (safe fallback)
+    auto * sampler = common_reasoning_budget_init(
+        nullptr,  // vocab - not used for basic state machine tests
+        start_tokens,
+        end_tokens,
+        forced_tokens,
+        budget,
+        initial_state
+    );
+
+    // Create a test token data array for checking forcing behavior
+    // Vocab size must be large enough to include all tokens (start, end, forced, sequence)
+    std::vector<llama_token_data> cur;
+    const size_t n_vocab = (size_t)max_token + 1;
+    for (size_t i = 0; i < n_vocab; i++) {
+        cur.emplace_back(llama_token_data{(llama_token)i, logf((float)(i+1)), 0.0f});
+    }
+    llama_token_data_array cur_p = { cur.data(), cur.size(), -1, false };
+
+    size_t actual_force_start = SIZE_MAX;
+    size_t actual_force_end = SIZE_MAX;
+
+    // Feed the sequence and track when forcing occurs
+    for (size_t i = 0; i < sequence.size(); i++) {
+        llama_sampler_accept(sampler, sequence[i]);
+
+        // Check if we're in forcing state by applying and seeing if logits are modified
+        cur_p.selected = -1;
+        for (size_t j = 0; j < cur.size(); j++) {
+            cur[j].logit = logf((float)(j+1));  // reset logits
+        }
+
+        llama_sampler_apply(sampler, &cur_p);
+
+        // Check if forcing is active (all logits except one should be -INFINITY)
+        size_t finite_count = 0;
+        llama_token finite_token = -1;
+        for (size_t j = 0; j < cur.size(); j++) {
+            if (std::isfinite(cur[j].logit)) {
+                finite_count++;
+                finite_token = cur[j].id;
+            }
+        }
+
+        fprintf(stderr, "    i=%zu: token=%d, finite_count=%zu, finite_token=%d\n", i, (int)sequence[i], finite_count, (int)finite_token);
+
+        if (finite_count == 1) {
+            if (actual_force_start == SIZE_MAX) {
+                actual_force_start = i;
+            }
+            actual_force_end = i;
+        } else if (actual_force_start != SIZE_MAX && actual_force_end != SIZE_MAX) {
+            // Forcing stopped
+            break;
+        }
+    }
+
+    llama_sampler_free(sampler);
+
+    // Verify forcing occurred at expected positions
+    if (expected_force_start == SIZE_MAX) {
+        if (actual_force_start != SIZE_MAX) {
+            fprintf(stderr, "Test '%s' FAILED: Expected no forcing, but forcing occurred at %zu\n", test_name, actual_force_start);
+            GGML_ASSERT(false && "Expected no forcing, but forcing occurred");
+        }
+    } else {
+        if (actual_force_start == SIZE_MAX) {
+            fprintf(stderr, "Test '%s' FAILED: Expected forcing but none occurred\n", test_name);
+            GGML_ASSERT(false && "Expected forcing but none occurred");
+        }
+        if (actual_force_start != expected_force_start) {
+            fprintf(stderr, "Test '%s' FAILED: Forcing started at %zu, expected %zu\n", test_name, actual_force_start, expected_force_start);
+            GGML_ASSERT(false && "Forcing started at wrong position");
+        }
+    }
+
+    if (expected_force_end != SIZE_MAX) {
+        if (actual_force_end < expected_force_end) {
+            fprintf(stderr, "Test '%s' FAILED: Forcing ended at %zu, expected >= %zu\n", test_name, actual_force_end, expected_force_end);
+            GGML_ASSERT(false && "Forcing ended too early");
+        }
+    }
+
+    fprintf(stderr, "  Test '%s' passed (force_start=%zu, force_end=%zu)\n", test_name, actual_force_start, actual_force_end);
+    (void)sequence;
+}
+
+// UTF-8 boundary detection unit test
+// Tests common_utf8_is_complete() from reasoning-budget.h
+static void test_utf8_boundary_detection() {
+    // Complete sequences
+    GGML_ASSERT(common_utf8_is_complete("hello"));
+    GGML_ASSERT(common_utf8_is_complete(""));
+    GGML_ASSERT(common_utf8_is_complete("\xC2\xA0"));            // complete 2-byte UTF-8 (U+00A0)
+    GGML_ASSERT(common_utf8_is_complete("\xE2\x80\x9C"));        // complete 3-byte UTF-8 (left double quote)
+    GGML_ASSERT(common_utf8_is_complete("\xF0\x9F\x98\x80"));    // complete 4-byte UTF-8 (emoji)
+    GGML_ASSERT(common_utf8_is_complete("abc\xC3\xA9"));         // ASCII + complete 2-byte
+
+    // Incomplete sequences
+    GGML_ASSERT(!common_utf8_is_complete(std::string("\xC2", 1)));            // 2-byte start, missing continuation
+    GGML_ASSERT(!common_utf8_is_complete(std::string("\xE2\x80", 2)));        // 3-byte start + 1 cont, missing 1
+    GGML_ASSERT(!common_utf8_is_complete(std::string("\xE2", 1)));            // 3-byte start, missing 2
+    GGML_ASSERT(!common_utf8_is_complete(std::string("\xF0\x9F\x98", 3)));    // 4-byte start + 2 cont, missing 1
+    GGML_ASSERT(!common_utf8_is_complete(std::string("\xF0\x9F", 2)));        // 4-byte start + 1 cont, missing 2
+    GGML_ASSERT(!common_utf8_is_complete(std::string("\xF0", 1)));            // 4-byte start, missing 3
+    GGML_ASSERT(!common_utf8_is_complete(std::string("\x80", 1)));            // orphan continuation byte
+
+    // Mixed: ASCII followed by start of multi-byte
+    GGML_ASSERT(!common_utf8_is_complete(std::string("hello\xC3", 6)));       // ASCII + incomplete 2-byte
+    GGML_ASSERT(common_utf8_is_complete(std::string("hello\xC3\xA9", 7)));    // ASCII + complete 2-byte
+}
+
+int main(void) {
+    // Reasoning budget sampler tests
+    printf("Testing reasoning budget sampler... ");
+
+    // Test 1: Basic budget with start/end tokens - no forcing (natural end before budget exhausted)
+    {
+        const std::vector<llama_token> start = {100};  // start token
+        const std::vector<llama_token> end = {101};    // end token
+        const std::vector<llama_token> forced = {102}; // forced token (not used in this test)
+        const std::vector<llama_token> sequence = {100, 50, 51, 101, 52}; // start, two tokens, end, one more
+
+        test_reasoning_budget("natural end before budget exhausted", sequence, start, end, forced,
+            5,      // budget of 5 tokens
+            REASONING_BUDGET_IDLE,
+            SIZE_MAX, SIZE_MAX); // no forcing expected (natural end)
+    }
+
+    // Test 2: Budget exhausted, forcing should occur
+    // Flow: i=0 accept(100)->COUNTING, i=1 accept(50)->remaining=1, i=2 accept(51)->remaining=0->FORCING
+    // Forcing is active at i=2 and i=3 (when apply() is called while in FORCING state)
+    // At i=4, force_pos becomes 2 which equals forced_tokens.size(), so state becomes DONE
+    {
+        const std::vector<llama_token> start = {100};
+        const std::vector<llama_token> end = {101};
+        const std::vector<llama_token> forced = {102, 101}; // forced message + end
+        const std::vector<llama_token> sequence = {100, 50, 51, 52, 53}; // start + 4 tokens (budget=2)
+
+        test_reasoning_budget("budget exhausted forcing", sequence, start, end, forced,
+            2,      // budget of 2 tokens
+            REASONING_BUDGET_IDLE,
+            2,      // forcing starts at i=2 (after accept(51) depletes budget, apply() forces)
+            3);     // forcing continues through i=3 (at i=4 state becomes DONE)
+    }
+
+    // Test 3: Activate immediately with budget=0, forcing should start right away
+    // Flow: Since no start token in sequence, state stays IDLE (no start/end configured means passthrough)
+    // This test needs start token to be in the sequence or use activate_immediately with start token present
+    {
+        const std::vector<llama_token> start = {100};
+        const std::vector<llama_token> end = {101};
+        const std::vector<llama_token> forced = {102, 101};
+        const std::vector<llama_token> sequence = {100, 50, 51, 52}; // start token first, then 3 tokens
+
+        test_reasoning_budget("activate immediately budget=0", sequence, start, end, forced,
+            0,      // budget of 0 tokens
+            REASONING_BUDGET_COUNTING, // starts counting, promoted to FORCING since budget=0
+            0,      // forcing starts at i=0 (after accept(100), budget=0 goes straight to FORCING)
+            1);     // forcing continues through i=1 (at i=2 state becomes DONE)
+    }
+
+    // Test 4: No start/end tokens configured - passthrough (no forcing)
+    {
+        const std::vector<llama_token> start = {};
+        const std::vector<llama_token> end = {};
+        const std::vector<llama_token> forced = {102};
+        const std::vector<llama_token> sequence = {50, 51, 52, 53};
+
+        test_reasoning_budget("no start/end configured", sequence, start, end, forced,
+            2,      // budget
+            REASONING_BUDGET_IDLE,
+            SIZE_MAX, SIZE_MAX); // no forcing (no start/end configured)
+    }
+
+    // Test 5: Activate immediately with budget > 0, count down then force
+    // Flow: i=0 accept(50)->remaining=1, i=1 accept(51)->remaining=0->FORCING
+    // So forcing starts at i=1 (apply after accept sees FORCING with force_pos=0)
+    {
+        const std::vector<llama_token> start = {100};
+        const std::vector<llama_token> end = {101};
+        const std::vector<llama_token> forced = {102, 101};
+        const std::vector<llama_token> sequence = {50, 51, 52, 53};
+
+        test_reasoning_budget("activate immediately with budget", sequence, start, end, forced,
+            2,      // budget of 2 tokens
+            REASONING_BUDGET_COUNTING,
+            1,      // forcing starts at i=1 (after 2 accepts deplete budget)
+            2);     // forcing continues through i=2
+    }
+
+    printf("OK (5 tests passed)\n");
+
+    printf("Testing UTF-8 boundary detection... ");
+    test_utf8_boundary_detection();
+    printf("OK\n");
+
+    return 0;
+}

tools/cli/cli.cpp

@@ -57,6 +57,8 @@ struct cli_context {
     std::vector<raw_buffer> input_files;
     task_params defaults;
     bool verbose_prompt;
+    int reasoning_budget = -1;
+    std::string reasoning_budget_message;
 
     // thread for showing "loading" animation
     std::atomic<bool> loading_show;
@@ -73,6 +75,8 @@ struct cli_context {
         // defaults.return_progress = true; // TODO: show progress
 
         verbose_prompt = params.verbose_prompt;
+        reasoning_budget = params.reasoning_budget;
+        reasoning_budget_message = params.reasoning_budget_message;
     }
 
     std::string generate_completion(result_timings & out_timings) {
@@ -95,6 +99,24 @@ struct cli_context {
                 task.params.chat_parser_params.parser.load(chat_params.parser);
             }
 
+            // reasoning budget sampler
+            if (reasoning_budget >= 0 && !chat_params.thinking_end_tag.empty()) {
+                const llama_vocab * vocab = llama_model_get_vocab(
+                    llama_get_model(ctx_server.get_llama_context()));
+
+                task.params.sampling.reasoning_budget_tokens = reasoning_budget;
+                task.params.sampling.reasoning_budget_activate_immediately = chat_params.thinking_forced_open;
+
+                if (!chat_params.thinking_start_tag.empty()) {
+                    task.params.sampling.reasoning_budget_start =
+                        common_tokenize(vocab, chat_params.thinking_start_tag, false, true);
+                }
+                task.params.sampling.reasoning_budget_end =
+                    common_tokenize(vocab, chat_params.thinking_end_tag, false, true);
+                task.params.sampling.reasoning_budget_forced =
+                    common_tokenize(vocab, reasoning_budget_message + chat_params.thinking_end_tag, false, true);
+            }
+
             rd.post_task({std::move(task)});
         }
 

tools/mtmd/clip-impl.h

@@ -216,6 +216,7 @@ enum projector_type {
     PROJECTOR_TYPE_GEMMA3,
     PROJECTOR_TYPE_GEMMA3NV,
     PROJECTOR_TYPE_GEMMA3NA,
+    PROJECTOR_TYPE_PHI4,
     PROJECTOR_TYPE_IDEFICS3,
     PROJECTOR_TYPE_PIXTRAL,
     PROJECTOR_TYPE_QWEN25VL,
@@ -253,6 +254,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
     { PROJECTOR_TYPE_GEMMA3NV,  "gemma3nv"},
     { PROJECTOR_TYPE_GEMMA3NA,  "gemma3na"},
+    { PROJECTOR_TYPE_PHI4,      "phi4"},
     { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
     { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
     { PROJECTOR_TYPE_ULTRAVOX,  "ultravox"},

tools/mtmd/clip.cpp

@@ -792,6 +792,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         case PROJECTOR_TYPE_IDEFICS3:
         case PROJECTOR_TYPE_LFM2:
         case PROJECTOR_TYPE_JANUS_PRO:
+        case PROJECTOR_TYPE_PHI4:
             {
                 builder = std::make_unique<clip_graph_siglip>(ctx, img);
             } break;
@@ -1144,6 +1145,13 @@ struct clip_model_loader {
                         // ref: https://huggingface.co/LiquidAI/LFM2.5-VL-1.6B/blob/main/processor_config.json
                         hparams.set_limit_image_tokens(64, 256);
                     } break;
+                case PROJECTOR_TYPE_PHI4:
+                    {
+                        hparams.n_merge = 1;
+                        get_u32(KEY_IMAGE_MIN_PIXELS, hparams.image_min_pixels);
+                        get_u32(KEY_IMAGE_MAX_PIXELS, hparams.image_max_pixels);
+                        hparams.set_warmup_n_tokens(16*16);
+                    } break;
                 case PROJECTOR_TYPE_PIXTRAL:
                 case PROJECTOR_TYPE_LIGHTONOCR:
                     {
@@ -1841,6 +1849,13 @@ struct clip_model_loader {
                     model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
                     model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
                 } break;
+            case PROJECTOR_TYPE_PHI4:
+                {
+                    model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight"));
+                    model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias"));
+                    model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
+                    model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
+                } break;
             case PROJECTOR_TYPE_LFM2A:
                 {
                     for (int i : {0, 2, 3, 5, 6}) {
@@ -3157,6 +3172,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
                 res_imgs->entries.push_back(std::move(img_f32));
             } break;
 
+        case PROJECTOR_TYPE_PHI4:
         case PROJECTOR_TYPE_PIXTRAL:
         case PROJECTOR_TYPE_LIGHTONOCR:
             {
@@ -3383,6 +3399,7 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
         case PROJECTOR_TYPE_MLP:
         case PROJECTOR_TYPE_MLP_NORM:
         case PROJECTOR_TYPE_JANUS_PRO:
+        case PROJECTOR_TYPE_PHI4:
             {
                 // do nothing
             } break;
@@ -3884,6 +3901,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         case PROJECTOR_TYPE_VOXTRAL:
         case PROJECTOR_TYPE_MUSIC_FLAMINGO:
         case PROJECTOR_TYPE_JANUS_PRO:
+        case PROJECTOR_TYPE_PHI4:
         case PROJECTOR_TYPE_COGVLM:
             {
                 // do nothing
@@ -4013,6 +4031,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_LDPV2:
             return ctx->model.mm_model_peg_0_b->ne[0];
         case PROJECTOR_TYPE_MLP:
+        case PROJECTOR_TYPE_PHI4:
         case PROJECTOR_TYPE_PIXTRAL:
         case PROJECTOR_TYPE_LIGHTONOCR:
             return ctx->model.mm_2_w->ne[1];

tools/mtmd/models/siglip.cpp

@@ -4,7 +4,7 @@ ggml_cgraph * clip_graph_siglip::build() {
     ggml_tensor * inp = build_inp();
 
     ggml_tensor * learned_pos_embd = model.position_embeddings;
-    if (proj_type == PROJECTOR_TYPE_LFM2) {
+    if (proj_type == PROJECTOR_TYPE_LFM2 || proj_type == PROJECTOR_TYPE_PHI4) {
         learned_pos_embd = resize_position_embeddings();
     }
 
@@ -75,6 +75,14 @@ ggml_cgraph * clip_graph_siglip::build() {
             hparams.ffn_op,
             -1);
 
+    } else if (proj_type == PROJECTOR_TYPE_PHI4) {
+        cur = build_ffn(cur,
+            model.mm_0_w, model.mm_0_b,
+            nullptr, nullptr,
+            model.mm_2_w, model.mm_2_b,
+            FFN_GELU,
+            -1);
+
     } else {
         GGML_ABORT("SigLIP: Unsupported projector type");
     }

tools/mtmd/mtmd.cpp

@@ -290,6 +290,9 @@ struct mtmd_context {
             img_beg = "<|vision_start|>";
             img_end = "<|vision_end|>";
 
+        } else if (proj == PROJECTOR_TYPE_PHI4) {
+            // Phi-4 uses media marker insertion only. Keep image boundary text empty.
+
         } else if (proj == PROJECTOR_TYPE_LLAMA4) {
             // (more details in mtmd_context constructor)
             img_beg = "<|image_start|>";

tools/server/server-common.cpp

@@ -1101,6 +1101,22 @@ json oaicompat_chat_params_parse(
         llama_params["chat_parser"] = chat_params.parser;
     }
 
+    // Reasoning budget: pass parameters through to sampling layer
+    {
+        int reasoning_budget = opt.reasoning_budget;
+        if (reasoning_budget == -1 && body.contains("thinking_budget_tokens")) {
+            reasoning_budget = json_value(body, "thinking_budget_tokens", -1);
+        }
+
+        if (reasoning_budget >= 0 && !chat_params.thinking_end_tag.empty()) {
+            llama_params["reasoning_budget_tokens"] = reasoning_budget;
+            llama_params["reasoning_budget_start_tag"] = chat_params.thinking_start_tag;
+            llama_params["reasoning_budget_end_tag"] = chat_params.thinking_end_tag;
+            llama_params["reasoning_budget_message"] = opt.reasoning_budget_message;
+            llama_params["reasoning_budget_activate_immediately"] = chat_params.thinking_forced_open;
+        }
+    }
+
     // Handle "logprobs" field
     // TODO: The response format of this option is not yet OAI-compatible, but seems like no one really using it; We may need to fix it in the future
     if (json_value(body, "logprobs", false)) {

tools/server/server-common.h

@@ -287,6 +287,8 @@ struct server_chat_params {
     bool allow_image;
     bool allow_audio;
     bool enable_thinking = true;
+    int  reasoning_budget = -1;
+    std::string reasoning_budget_message;
     std::string media_path;
 };
 

tools/server/server-context.cpp

@@ -893,9 +893,10 @@ private:
             }
 
             // thinking is enabled if:
-            // 1. It's not explicitly disabled (reasoning_budget == 0)
+            // 1. It's not explicitly disabled via --reasoning off
             // 2. The chat template supports it
-            const bool enable_thinking = params_base.use_jinja && params_base.reasoning_budget != 0 && common_chat_templates_support_enable_thinking(chat_templates.get());
+            const bool template_supports_thinking = params_base.use_jinja && common_chat_templates_support_enable_thinking(chat_templates.get());
+            const bool enable_thinking = params_base.enable_reasoning != 0 && template_supports_thinking;
             SRV_INF("%s: chat template, thinking = %d\n", __func__, enable_thinking);
 
             chat_params = {
@@ -907,6 +908,8 @@ private:
                 /* allow_image           */ mctx ? mtmd_support_vision(mctx) : false,
                 /* allow_audio           */ mctx ? mtmd_support_audio (mctx) : false,
                 /* enable_thinking       */ enable_thinking,
+                /* reasoning_budget      */ params_base.reasoning_budget,
+                /* reasoning_budget_msg  */ params_base.reasoning_budget_message,
                 /* media_path            */ params_base.media_path,
             };
         }

tools/server/server-task.cpp

@@ -462,6 +462,34 @@ task_params server_task::params_from_json_cmpl(
         }
     }
 
+    // Parse reasoning budget sampler parameters
+    {
+        const int32_t budget = json_value(data, "reasoning_budget_tokens", (int32_t) -1);
+        if (budget >= 0) {
+            const auto start_tag = json_value(data, "reasoning_budget_start_tag", std::string());
+            const auto end_tag   = json_value(data, "reasoning_budget_end_tag", std::string());
+            const auto message   = json_value(data, "reasoning_budget_message", std::string());
+            const bool activate_imm   = json_value(data, "reasoning_budget_activate_immediately", false);
+
+            params.sampling.reasoning_budget_tokens = budget;
+            params.sampling.reasoning_budget_activate_immediately = activate_imm;
+
+            if (!start_tag.empty()) {
+                params.sampling.reasoning_budget_start = common_tokenize(vocab, start_tag, false, true);
+            }
+            if (!end_tag.empty()) {
+                params.sampling.reasoning_budget_end = common_tokenize(vocab, end_tag, false, true);
+                params.sampling.reasoning_budget_forced = common_tokenize(vocab, message + end_tag, false, true);
+            }
+
+            SRV_DBG("reasoning budget: tokens=%d, activate_immediately=%s, start=%zu toks, end=%zu toks, forced=%zu toks\n",
+                budget, activate_imm ? "true" : "false",
+                params.sampling.reasoning_budget_start.size(),
+                params.sampling.reasoning_budget_end.size(),
+                params.sampling.reasoning_budget_forced.size());
+        }
+    }
+
     {
         params.sampling.logit_bias.clear();