feat: llama.cpp bump (17f7f4) for SSM performance improvements (#13408)

* feat: Bump llama.cpp to the latest master (17f7f4b)

This brings in significant improvements to prefill performance for all
models using the SSM_CONV and SSM_SCAN ops (granite4, jamba, falcon-h,
nemotron-h, Qwen3 Next) on Apple Metal.

See https://github.com/ggml-org/llama.cpp/pull/17876

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Update patches 1-4

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* fix: Update patches 5-12

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Update patches 13-18

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Update patch 20

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Update patches 21-31

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

* feat: Sync vendored code

The two files I'm not sure about here are the swap from gemma3-iswa.cpp to
gemma3.cpp (I chose to include this because I think it's required), and the
inclusion of `ggml-zendnn.h` which I chose to omit.

Branch: LlamaCPPMetalSSMImprovements

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

---------

Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>

Makefile.sync

@@ -1,6 +1,6 @@
 UPSTREAM=https://github.com/ggml-org/llama.cpp.git
 WORKDIR=llama/vendor
-FETCH_HEAD=7f8ef50cce40e3e7e4526a3696cb45658190e69a
+FETCH_HEAD=17f7f4baad8b3a716ee139da7bb56ae984e8c0fa
 
 .PHONY: help
 help:

llama/build-info.cpp

@@ -1,4 +1,4 @@
 int LLAMA_BUILD_NUMBER = 0;
-char const *LLAMA_COMMIT = "7f8ef50cce40e3e7e4526a3696cb45658190e69a";
+char const *LLAMA_COMMIT = "17f7f4baad8b3a716ee139da7bb56ae984e8c0fa";
 char const *LLAMA_COMPILER = "";
 char const *LLAMA_BUILD_TARGET = "";

llama/llama.cpp/common/common.cpp

@@ -694,7 +694,7 @@ bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_over
 
 // Validate if a filename is safe to use
 // To validate a full path, split the path by the OS-specific path separator, and validate each part with this function
-bool fs_validate_filename(const std::string & filename) {
+bool fs_validate_filename(const std::string & filename, bool allow_subdirs) {
     if (!filename.length()) {
         // Empty filename invalid
         return false;
@@ -754,10 +754,14 @@ bool fs_validate_filename(const std::string & filename) {
             || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
             || c == 0xFFFD // Replacement Character (UTF-8)
             || c == 0xFEFF // Byte Order Mark (BOM)
-            || c == '/' || c == '\\' || c == ':' || c == '*' // Illegal characters
+            || c == ':' || c == '*' // Illegal characters
             || c == '?' || c == '"' || c == '<' || c == '>' || c == '|') {
             return false;
         }
+        if (!allow_subdirs && (c == '/' || c == '\\')) {
+            // Subdirectories not allowed, reject path separators
+            return false;
+        }
     }
 
     // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
@@ -782,11 +786,29 @@ bool fs_validate_filename(const std::string & filename) {
 #include <iostream>
 
 
+#ifdef _WIN32
+static std::wstring utf8_to_wstring(const std::string & str) {
+    if (str.empty()) {
+        return std::wstring();
+    }
+
+    int size = MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), NULL, 0);
+
+    if (size <= 0) {
+        return std::wstring();
+    }
+
+    std::wstring wstr(size, 0);
+    MultiByteToWideChar(CP_UTF8, 0, str.c_str(), (int)str.size(), &wstr[0], size);
+
+    return wstr;
+}
+#endif
+
 // returns true if successful, false otherwise
 bool fs_create_directory_with_parents(const std::string & path) {
 #ifdef _WIN32
-    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
-    std::wstring wpath = converter.from_bytes(path);
+    std::wstring wpath = utf8_to_wstring(path);
 
     // if the path already exists, check whether it's a directory
     const DWORD attributes = GetFileAttributesW(wpath.c_str());
@@ -859,6 +881,11 @@ bool fs_create_directory_with_parents(const std::string & path) {
 #endif // _WIN32
 }
 
+bool fs_is_directory(const std::string & path) {
+    std::filesystem::path dir(path);
+    return std::filesystem::exists(dir) && std::filesystem::is_directory(dir);
+}
+
 std::string fs_get_cache_directory() {
     std::string cache_directory = "";
     auto ensure_trailing_slash = [](std::string p) {
@@ -893,6 +920,8 @@ std::string fs_get_cache_directory() {
         cache_directory = std::getenv("HOME") + std::string("/Library/Caches/");
 #elif defined(_WIN32)
         cache_directory = std::getenv("LOCALAPPDATA");
+#elif defined(__EMSCRIPTEN__)
+        GGML_ABORT("not implemented on this platform");
 #else
 #  error Unknown architecture
 #endif
@@ -912,7 +941,7 @@ std::string fs_get_cache_file(const std::string & filename) {
     return cache_directory + filename;
 }
 
-std::vector<common_file_info> fs_list_files(const std::string & path) {
+std::vector<common_file_info> fs_list(const std::string & path, bool include_directories) {
     std::vector<common_file_info> files;
     if (path.empty()) return files;
 
@@ -929,12 +958,20 @@ std::vector<common_file_info> fs_list_files(const std::string & path) {
                 common_file_info info;
                 info.path   = p.string();
                 info.name   = p.filename().string();
+                info.is_dir = false;
                 try {
                     info.size = static_cast<size_t>(std::filesystem::file_size(p));
                 } catch (const std::filesystem::filesystem_error &) {
                     info.size = 0;
                 }
                 files.push_back(std::move(info));
+            } else if (include_directories && std::filesystem::is_directory(p)) {
+                common_file_info info;
+                info.path   = p.string();
+                info.name   = p.filename().string();
+                info.size   = 0; // Directories have no size
+                info.is_dir = true;
+                files.push_back(std::move(info));
             }
         } catch (const std::filesystem::filesystem_error &) {
             // skip entries we cannot inspect
@@ -945,6 +982,32 @@ std::vector<common_file_info> fs_list_files(const std::string & path) {
     return files;
 }
 
+//
+// TTY utils
+//
+
+bool tty_can_use_colors() {
+    // Check NO_COLOR environment variable (https://no-color.org/)
+    if (const char * no_color = std::getenv("NO_COLOR")) {
+        if (no_color[0] != '\0') {
+            return false;
+        }
+    }
+
+    // Check TERM environment variable
+    if (const char * term = std::getenv("TERM")) {
+        if (std::strcmp(term, "dumb") == 0) {
+            return false;
+        }
+    }
+
+    // Check if stdout and stderr are connected to a terminal
+    // We check both because log messages can go to either
+    bool stdout_is_tty = isatty(fileno(stdout));
+    bool stderr_is_tty = isatty(fileno(stderr));
+
+    return stdout_is_tty || stderr_is_tty;
+}
 
 //
 // Model utils

llama/llama.cpp/common/common.h

@@ -12,6 +12,10 @@
 #include <vector>
 #include <map>
 
+#if defined(_WIN32) && !defined(_WIN32_WINNT)
+#define _WIN32_WINNT 0x0A00
+#endif
+
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
 #else
@@ -26,8 +30,6 @@
     fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)
 
-#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
-
 struct common_time_meas {
     common_time_meas(int64_t & t_acc, bool disable = false);
     ~common_time_meas();
@@ -223,6 +225,7 @@ struct common_params_model {
     std::string hf_repo     = ""; // HF repo                                                // NOLINT
     std::string hf_file     = ""; // HF file                                                // NOLINT
     std::string docker_repo = ""; // Docker repo                                            // NOLINT
+    std::string name        = ""; // in format <user>/<model>[:<tag>] (tag is optional)     // NOLINT
 };
 
 struct common_params_speculative {
@@ -369,7 +372,7 @@ struct common_params {
 
     std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
 
-    int32_t verbosity                  = 0;
+    int32_t verbosity                  = 3;  // LOG_LEVEL_INFO
     int32_t control_vector_layer_start = -1; // layer range for control vector
     int32_t control_vector_layer_end   = -1; // layer range for control vector
     bool    offline                    = false;
@@ -478,9 +481,15 @@ struct common_params {
     bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
 
+    // router server configs
+    std::string models_dir = ""; // directory containing models for the router server
+    int models_max = 4;          // maximum number of models to load simultaneously
+    bool models_autoload = true; // automatically load models when requested via the router server
+
     bool log_json = false;
 
     std::string slot_save_path;
+    std::string media_path; // path to directory for loading media files
 
     float slot_prompt_similarity = 0.1f;
 
@@ -631,8 +640,9 @@ std::string string_from(const struct llama_context * ctx, const struct llama_bat
 // Filesystem utils
 //
 
-bool fs_validate_filename(const std::string & filename);
+bool fs_validate_filename(const std::string & filename, bool allow_subdirs = false);
 bool fs_create_directory_with_parents(const std::string & path);
+bool fs_is_directory(const std::string & path);
 
 std::string fs_get_cache_directory();
 std::string fs_get_cache_file(const std::string & filename);
@@ -641,8 +651,16 @@ struct common_file_info {
     std::string path;
     std::string name;
     size_t      size = 0; // in bytes
+    bool        is_dir = false;
 };
-std::vector<common_file_info> fs_list_files(const std::string & path);
+std::vector<common_file_info> fs_list(const std::string & path, bool include_directories);
+
+//
+// TTY utils
+//
+
+// Auto-detect if colors can be enabled based on terminal and environment
+bool tty_can_use_colors();
 
 //
 // Model utils

llama/llama.cpp/common/json-schema-to-grammar.cpp

@@ -974,7 +974,7 @@ public:
 
     void check_errors() {
         if (!_errors.empty()) {
-            throw std::runtime_error("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
+            throw std::invalid_argument("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
         }
         if (!_warnings.empty()) {
             fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", string_join(_warnings, "; ").c_str());

llama/llama.cpp/common/log.cpp

@@ -1,3 +1,4 @@
+#include "common.h"
 #include "log.h"
 
 #include <chrono>
@@ -26,30 +27,6 @@ void common_log_set_verbosity_thold(int verbosity) {
     common_log_verbosity_thold = verbosity;
 }
 
-// Auto-detect if colors should be enabled based on terminal and environment
-static bool common_log_should_use_colors_auto() {
-    // Check NO_COLOR environment variable (https://no-color.org/)
-    if (const char * no_color = std::getenv("NO_COLOR")) {
-        if (no_color[0] != '\0') {
-            return false;
-        }
-    }
-
-    // Check TERM environment variable
-    if (const char * term = std::getenv("TERM")) {
-        if (std::strcmp(term, "dumb") == 0) {
-            return false;
-        }
-    }
-
-    // Check if stdout and stderr are connected to a terminal
-    // We check both because log messages can go to either
-    bool stdout_is_tty = isatty(fileno(stdout));
-    bool stderr_is_tty = isatty(fileno(stderr));
-
-    return stdout_is_tty || stderr_is_tty;
-}
-
 static int64_t t_us() {
     return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
 }
@@ -391,7 +368,7 @@ struct common_log * common_log_main() {
     static std::once_flag    init_flag;
     std::call_once(init_flag, [&]() {
         // Set default to auto-detect colors
-        log.set_colors(common_log_should_use_colors_auto());
+        log.set_colors(tty_can_use_colors());
     });
 
     return &log;
@@ -422,7 +399,7 @@ void common_log_set_file(struct common_log * log, const char * file) {
 
 void common_log_set_colors(struct common_log * log, log_colors colors) {
     if (colors == LOG_COLORS_AUTO) {
-        log->set_colors(common_log_should_use_colors_auto());
+        log->set_colors(tty_can_use_colors());
         return;
     }
 
@@ -443,8 +420,22 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps) {
     log->set_timestamps(timestamps);
 }
 
+static int common_get_verbosity(enum ggml_log_level level) {
+    switch (level) {
+        case GGML_LOG_LEVEL_DEBUG: return LOG_LEVEL_DEBUG;
+        case GGML_LOG_LEVEL_INFO:  return LOG_LEVEL_INFO;
+        case GGML_LOG_LEVEL_WARN:  return LOG_LEVEL_WARN;
+        case GGML_LOG_LEVEL_ERROR: return LOG_LEVEL_ERROR;
+        case GGML_LOG_LEVEL_CONT:  return LOG_LEVEL_INFO; // same as INFO
+        case GGML_LOG_LEVEL_NONE:
+        default:
+            return LOG_LEVEL_OUTPUT;
+    }
+}
+
 void common_log_default_callback(enum ggml_log_level level, const char * text, void * /*user_data*/) {
-    if (LOG_DEFAULT_LLAMA <= common_log_verbosity_thold) {
+    auto verbosity = common_get_verbosity(level);
+    if (verbosity <= common_log_verbosity_thold) {
         common_log_add(common_log_main(), level, "%s", text);
     }
 }

llama/llama.cpp/common/log.h

@@ -21,8 +21,14 @@
 #    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
 #endif
 
-#define LOG_DEFAULT_DEBUG 1
-#define LOG_DEFAULT_LLAMA 0
+#define LOG_LEVEL_DEBUG  4
+#define LOG_LEVEL_INFO   3
+#define LOG_LEVEL_WARN   2
+#define LOG_LEVEL_ERROR  1
+#define LOG_LEVEL_OUTPUT 0 // output data from tools
+
+#define LOG_DEFAULT_DEBUG LOG_LEVEL_DEBUG
+#define LOG_DEFAULT_LLAMA LOG_LEVEL_INFO
 
 enum log_colors {
     LOG_COLORS_AUTO     = -1,
@@ -67,10 +73,11 @@ void common_log_add(struct common_log * log, enum ggml_log_level level, const ch
 //   0.00.090.578 I llm_load_tensors: offloading 32 repeating layers to GPU
 //   0.00.090.579 I llm_load_tensors: offloading non-repeating layers to GPU
 //
-// I - info    (stdout, V = 0)
-// W - warning (stderr, V = 0)
-// E - error   (stderr, V = 0)
 // D - debug   (stderr, V = LOG_DEFAULT_DEBUG)
+// I - info    (stdout, V = LOG_DEFAULT_INFO)
+// W - warning (stderr, V = LOG_DEFAULT_WARN)
+// E - error   (stderr, V = LOG_DEFAULT_ERROR)
+// O - output  (stdout, V = LOG_DEFAULT_OUTPUT)
 //
 
 void common_log_set_file      (struct common_log * log, const char * file); // not thread-safe
@@ -95,14 +102,14 @@ void common_log_set_timestamps(struct common_log * log, bool timestamps);   // w
         } \
     } while (0)
 
-#define LOG(...)             LOG_TMPL(GGML_LOG_LEVEL_NONE, 0,         __VA_ARGS__)
+#define LOG(...)             LOG_TMPL(GGML_LOG_LEVEL_NONE, LOG_LEVEL_OUTPUT, __VA_ARGS__)
 #define LOGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_NONE, verbosity,        __VA_ARGS__)
 
-#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  0,                 __VA_ARGS__)
-#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  0,                 __VA_ARGS__)
-#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0,                 __VA_ARGS__)
-#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
-#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  0,                 __VA_ARGS__)
+#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_LEVEL_DEBUG,  __VA_ARGS__)
+#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  LOG_LEVEL_INFO,   __VA_ARGS__)
+#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  LOG_LEVEL_WARN,   __VA_ARGS__)
+#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, LOG_LEVEL_ERROR,  __VA_ARGS__)
+#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  LOG_LEVEL_INFO,   __VA_ARGS__) // same as INFO
 
 #define LOG_INFV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  verbosity, __VA_ARGS__)
 #define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  verbosity, __VA_ARGS__)

llama/llama.cpp/src/llama-arch.cpp

@@ -112,6 +112,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_COGVLM,           "cogvlm"           },
     { LLM_ARCH_RND1,             "rnd1"             },
     { LLM_ARCH_PANGU_EMBED,      "pangu-embedded"   },
+    { LLM_ARCH_MISTRAL3,         "mistral3"         },
     { LLM_ARCH_UNKNOWN,          "(unknown)"        },
 };
 
@@ -205,6 +206,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
     { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
+    { LLM_KV_ATTENTION_TEMPERATURE_SCALE,            "%s.attention.temperature_scale"            },
     { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
     { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
@@ -855,7 +857,7 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_GATE_SHEXP,     "blk.%d.ffn_gate_shexp" },
             { LLM_TENSOR_FFN_DOWN_SHEXP,     "blk.%d.ffn_down_shexp" },
             { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
-            { LLM_TENSOR_SSM_A,              "blk.%d.ssm_a" },
+            { LLM_TENSOR_SSM_A_NOSCAN,       "blk.%d.ssm_a" },
             { LLM_TENSOR_SSM_CONV1D,         "blk.%d.ssm_conv1d" },
             { LLM_TENSOR_SSM_DT,             "blk.%d.ssm_dt" },
             { LLM_TENSOR_SSM_BETA_ALPHA,     "blk.%d.ssm_ba" },
@@ -2532,6 +2534,32 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
             { LLM_TENSOR_FFN_UP_EXPS,        "blk.%d.ffn_up_exps" },
         },
     },
+    {
+        LLM_ARCH_MISTRAL3,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_ATTN_ROT_EMBD,   "blk.%d.attn_rot_embd" },
+            { LLM_TENSOR_FFN_GATE_INP,    "blk.%d.ffn_gate_inp" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_GATE,        "blk.%d.ffn_gate" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_GATE_EXP,    "blk.%d.ffn_gate.%d" },
+            { LLM_TENSOR_FFN_DOWN_EXP,    "blk.%d.ffn_down.%d" },
+            { LLM_TENSOR_FFN_UP_EXP,      "blk.%d.ffn_up.%d" },
+            { LLM_TENSOR_FFN_GATE_EXPS,   "blk.%d.ffn_gate_exps" },
+            { LLM_TENSOR_FFN_DOWN_EXPS,   "blk.%d.ffn_down_exps" },
+            { LLM_TENSOR_FFN_UP_EXPS,     "blk.%d.ffn_up_exps" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -2631,6 +2659,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_FFN_ACT,                    {LLM_TENSOR_LAYER_REPEATING, GGML_OP_DIV}},
     {LLM_TENSOR_SSM_CONV1D,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}},
     {LLM_TENSOR_SSM_A,                      {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_SCAN}},
+    {LLM_TENSOR_SSM_A_NOSCAN,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, // a version of SSM_A used for MUL instead of SSM_SCAN
     {LLM_TENSOR_SSM_DT_NORM,                {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_SSM_B_NORM,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
     {LLM_TENSOR_SSM_C_NORM,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},

llama/llama.cpp/src/llama-arch.h

@@ -116,6 +116,7 @@ enum llm_arch {
     LLM_ARCH_COGVLM,
     LLM_ARCH_RND1,
     LLM_ARCH_PANGU_EMBED,
+    LLM_ARCH_MISTRAL3,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -209,6 +210,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_OUTPUT_SCALE,
     LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
+    LLM_KV_ATTENTION_TEMPERATURE_SCALE,
     LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,
     LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
@@ -379,6 +381,7 @@ enum llm_tensor {
     LLM_TENSOR_SSM_DT,
     LLM_TENSOR_SSM_DT_NORM,
     LLM_TENSOR_SSM_A,
+    LLM_TENSOR_SSM_A_NOSCAN,        // qwen3next special case with MUL instead of SSM_SCAN
     LLM_TENSOR_SSM_B_NORM,
     LLM_TENSOR_SSM_C_NORM,
     LLM_TENSOR_SSM_D,

llama/llama.cpp/src/llama-context.cpp

@@ -248,7 +248,10 @@ llama_context::llama_context(
 
         LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size());
 
-        const size_t max_nodes = this->graph_max_nodes();
+        const uint32_t n_seqs = cparams.n_seq_max;
+        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
+
+        const size_t max_nodes = this->graph_max_nodes(n_tokens);
 
         LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes);
 
@@ -300,9 +303,6 @@ llama_context::llama_context(
 
         cross.v_embd.clear();
 
-        const uint32_t n_seqs = cparams.n_seq_max;
-        const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
-
         // avoid reserving graphs with zero outputs - assume one output per sequence
         n_outputs = n_seqs;
 
@@ -1385,9 +1385,9 @@ void llama_context::output_reorder() {
 // graph
 //
 
-uint32_t llama_context::graph_max_nodes() const {
+uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
     if (model.arch == LLM_ARCH_QWEN3NEXT) {
-        return std::max<uint32_t>(8192u, 32u*model.n_tensors());
+        return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
     }
     return std::max<uint32_t>(1024u, 8u*model.n_tensors());
 }

llama/llama.cpp/src/llama-context.h

@@ -197,7 +197,7 @@ private:
     //
 
 public:
-    uint32_t graph_max_nodes() const;
+    uint32_t graph_max_nodes(uint32_t n_tokens) const;
 
     // can reuse the llm_graph_result instance of the context (for example to update a memory module)
     llm_graph_result * get_gf_res_reserve() const;

llama/llama.cpp/src/llama-grammar.cpp

@@ -181,6 +181,52 @@ static std::pair<uint32_t, const char *> parse_char(const char * src) {
     throw std::runtime_error("unexpected end of input");
 }
 
+static std::pair<uint32_t, const char *> parse_token(const llama_vocab * vocab, const char * src) {
+    const char * pos = src;
+    if (*pos != '<') {
+        throw std::runtime_error(std::string("expecting '<' at ") + pos);
+    }
+    pos++;
+
+    // Parse <[id]>
+    if (*pos == '[') {
+        pos++;
+        const char * int_end = parse_int(pos);
+        uint32_t token_id = std::stoul(std::string(pos, int_end - pos));
+        pos = int_end;
+        if (*pos != ']') {
+            throw std::runtime_error(std::string("expecting ']' at ") + pos);
+        }
+        pos++;
+        if (*pos != '>') {
+            throw std::runtime_error(std::string("expecting '>' at ") + pos);
+        }
+        pos++;
+        return std::make_pair(token_id, pos);
+    }
+
+    if (vocab == nullptr) {
+        throw std::runtime_error(std::string("no vocab to parse token at ") + src);
+    }
+
+    // Parse <token> and tokenize to obtain the token id
+    while (*pos != 0 && *pos != '>') {
+        pos++;
+    }
+    if (*pos != '>') {
+        throw std::runtime_error(std::string("expecting '>' at ") + pos);
+    }
+    pos++;
+
+    llama_token tokens[2];
+    int32_t n_tokens = vocab->tokenize(src, static_cast<int32_t>(pos - src), tokens, 2, false, true);
+    if (n_tokens != 1) {
+        // must tokenize to exactly 1 token
+        throw std::runtime_error("invalid token '" + std::string(src, pos - src) + "'");
+    }
+    return std::make_pair(tokens[0], pos);
+}
+
 static void print_grammar_char(FILE * file, uint32_t c) {
     if (0x20 <= c && c <= 0x7f) {
         fprintf(file, "%c", static_cast<char>(c));
@@ -212,6 +258,8 @@ static void print_rule_binary(FILE * file, const llama_grammar_rule & rule) {
             case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
             case LLAMA_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
             case LLAMA_GRETYPE_CHAR_ANY:       fprintf(file, "CHAR_ANY");       break;
+            case LLAMA_GRETYPE_TOKEN:          fprintf(file, "TOKEN");          break;
+            case LLAMA_GRETYPE_TOKEN_NOT:      fprintf(file, "TOKEN_NOT");      break;
         }
         switch (elem.type) {
             case LLAMA_GRETYPE_END:
@@ -228,6 +276,17 @@ static void print_rule_binary(FILE * file, const llama_grammar_rule & rule) {
                 print_grammar_char(file, elem.value);
                 fprintf(file, "\") ");
                 break;
+            case LLAMA_GRETYPE_TOKEN:
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
+            case LLAMA_GRETYPE_TOKEN_NOT:
+                fprintf(file, "!");
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
         }
     }
     fprintf(file, "\n");
@@ -284,6 +343,17 @@ static void print_rule(
             case LLAMA_GRETYPE_CHAR_ANY:
                 fprintf(file, ".");
                 break;
+            case LLAMA_GRETYPE_TOKEN:
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
+            case LLAMA_GRETYPE_TOKEN_NOT:
+                fprintf(file, "!");
+                fprintf(file, "<[");
+                fprintf(file, "%u", elem.value);
+                fprintf(file, "]> ");
+                break;
         }
         if (is_char_element(elem)) {
             switch (rule[i + 1].type) {
@@ -444,6 +514,17 @@ const char * llama_grammar_parser::parse_sequence(
                 }
             }
             pos = parse_space(pos + 1, is_nested);
+        } else if (*pos == '<' || *pos == '!') { // token
+            auto type = LLAMA_GRETYPE_TOKEN;
+            if (*pos == '!') { // token inverse
+                type = LLAMA_GRETYPE_TOKEN_NOT;
+                pos++;
+            }
+            auto token_pair = parse_token(vocab, pos);
+            const char * token_end  = token_pair.second;
+            last_sym_start = rule.size();
+            rule.push_back({type, token_pair.first});
+            pos = parse_space(token_end, is_nested);
         } else if (is_word_char(*pos)) { // rule reference
             const char * name_end    = parse_name(pos);
             uint32_t ref_rule_id = get_symbol_id(pos, name_end - pos);
@@ -691,6 +772,21 @@ static bool llama_grammar_match_partial_char(
     return !is_positive_char;
 }
 
+// returns true iff token matches the rule at pos (regular or inverse)
+// asserts that pos is pointing to a token element
+static bool llama_grammar_match_token(
+    const llama_grammar_element * pos,
+    const llama_token             token) {
+    GGML_ASSERT(pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT);
+    if (pos->type == LLAMA_GRETYPE_TOKEN) {
+        return pos->value == static_cast<uint32_t>(token);
+    }
+    if (pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+        return pos->value != static_cast<uint32_t>(token);
+    }
+    return false;
+}
+
 // transforms a grammar pushdown stack into N possible stacks, all ending
 // at a character range (terminal element)
 static void llama_grammar_advance_stack(
@@ -738,6 +834,8 @@ static void llama_grammar_advance_stack(
         case LLAMA_GRETYPE_CHAR:
         case LLAMA_GRETYPE_CHAR_NOT:
         case LLAMA_GRETYPE_CHAR_ANY:
+        case LLAMA_GRETYPE_TOKEN:
+        case LLAMA_GRETYPE_TOKEN_NOT:
             if (std::find(new_stacks.begin(), new_stacks.end(), stack) == new_stacks.end()) {
                 // only add the stack if it's not a duplicate of one we already have
                 new_stacks.emplace_back(stack);
@@ -831,26 +929,38 @@ llama_grammar_stacks & llama_grammar_get_stacks(struct llama_grammar * grammar)
     return grammar->stacks;
 }
 
+static void llama_grammar_accept_chr(
+        struct llama_grammar       & grammar,
+        const llama_grammar_stack  & stack,
+              uint32_t               chr,
+              llama_grammar_stacks & new_stacks) {
+    if (stack.empty()) {
+        return;
+    }
+
+    const llama_grammar_element * pos = stack.back();
+
+    // ignore if this turns into a token
+    if (pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+        return;
+    }
+
+    auto match = llama_grammar_match_char(pos, chr);
+    if (match.first) {
+        llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
+        if (!llama_grammar_is_end_of_sequence(match.second)) {
+            new_stack.push_back(match.second);
+        }
+        llama_grammar_advance_stack(grammar.rules, new_stack, new_stacks);
+    }
+}
+
 void llama_grammar_accept(struct llama_grammar * grammar, uint32_t chr) {
     llama_grammar_stacks stacks_new;
     stacks_new.reserve(grammar->stacks.size());
 
     for (const auto & stack : grammar->stacks) {
-        if (stack.empty()) {
-            continue;
-        }
-
-        auto match = llama_grammar_match_char(stack.back(), chr);
-        if (match.first) {
-            const llama_grammar_element * pos = match.second;
-
-            // update top of stack to next element, if any
-            llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
-            if (!llama_grammar_is_end_of_sequence(pos)) {
-                new_stack.push_back(pos);
-            }
-            llama_grammar_advance_stack(grammar->rules, new_stack, stacks_new);
-        }
+        llama_grammar_accept_chr(*grammar, stack, chr, stacks_new);
     }
 
     grammar->stacks = std::move(stacks_new);
@@ -875,6 +985,22 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
 
     const llama_grammar_element * stack_pos = stack.back();
 
+    // if the top of the stack is a token rule, then we only need to check the token id
+    if (stack_pos->type == LLAMA_GRETYPE_TOKEN || stack_pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+        for (const auto & tok : candidates) {
+            if (*tok.code_points == 0) {
+                // reached the end of a token consumed by char rules, reject iff it ended
+                // in a partial response
+                if (tok.partial_utf8.n_remain != 0) {
+                    rejects.push_back(tok);
+                }
+            } else if (!llama_grammar_match_token(stack_pos, tok.id)) {
+                rejects.push_back(tok);
+            }
+        }
+        return rejects;
+    }
+
     llama_grammar_candidates next_candidates;
     next_candidates.reserve(candidates.size());
 
@@ -887,7 +1013,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
                 rejects.push_back(tok);
             }
         } else if (llama_grammar_match_char(stack_pos, *tok.code_points).first) {
-            next_candidates.push_back({ tok.index, tok.code_points + 1, tok.partial_utf8 });
+            next_candidates.push_back({ tok.index, tok.code_points + 1, tok.partial_utf8, tok.id });
         } else {
             rejects.push_back(tok);
         }
@@ -905,7 +1031,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
 
     auto next_rejects = llama_grammar_reject_candidates(rules, next_stacks, next_candidates);
     for (const auto & tok : next_rejects) {
-        rejects.push_back({ tok.index, tok.code_points - 1, tok.partial_utf8 });
+        rejects.push_back({ tok.index, tok.code_points - 1, tok.partial_utf8, tok.id });
     }
 
     return rejects;
@@ -975,9 +1101,10 @@ struct llama_grammar * llama_grammar_init_impl(
         std::move(vec_rules),
         std::move(stacks),
         /* .partial_utf8 = */             {},
-        /* .lazy =*/              false,
+        /* .lazy = */                     false,
         /* .awaiting_trigger = */         false,
         /* .trigger_buffer = */           "",
+        /* .trigger_buffer_positions = */ {},
         /* .trigger_tokens = */           {},
         /* .trigger_patterns = */         {},
     };
@@ -993,7 +1120,7 @@ struct llama_grammar * llama_grammar_init_impl(
                             size_t num_trigger_patterns,
                const llama_token * trigger_tokens,
                             size_t num_trigger_tokens) {
-    llama_grammar_parser parser;
+    llama_grammar_parser parser(vocab);
 
     // if there is a grammar, parse it
     // rules will be empty (default) if there are parse errors
@@ -1085,6 +1212,7 @@ struct llama_grammar * llama_grammar_init_impl(
         /* .lazy = */                     lazy,
         /* .awaiting_trigger = */         lazy,
         /* .trigger_buffer = */           "",
+        /* .trigger_buffer_positions = */ {},
         std::move(vec_trigger_tokens),
         std::move(vec_trigger_patterns),
     };
@@ -1108,6 +1236,7 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
         grammar.lazy,
         grammar.awaiting_trigger,
         grammar.trigger_buffer,
+        grammar.trigger_buffer_positions,
         grammar.trigger_tokens,
         grammar.trigger_patterns,
     };
@@ -1164,7 +1293,7 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
             cur_p->data[i].logit = -INFINITY;
         } else {
             candidates_decoded.push_back(decode_utf8(piece, grammar.partial_utf8));
-            candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second });
+            candidates_grammar.push_back({ i, candidates_decoded.back().first.data(), candidates_decoded.back().second, id });
         }
     }
 
@@ -1184,10 +1313,12 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
         if (std::find(grammar.trigger_tokens.begin(), grammar.trigger_tokens.end(), token) != grammar.trigger_tokens.end()) {
             grammar.awaiting_trigger = false;
             grammar.trigger_buffer.clear();
-            llama_grammar_accept_str(grammar, piece);
+            llama_grammar_accept_token(grammar, token, piece);
             LLAMA_LOG_DEBUG("Grammar triggered on token %u (`%s`)", token, piece.c_str());
             return;
         } else {
+            auto position = std::make_pair(grammar.trigger_buffer.size(), grammar.trigger_buffer.size() + piece.size());
+            grammar.trigger_buffer_positions.push_back(std::make_pair(token, position));
             grammar.trigger_buffer += piece;
 
             std::smatch match;
@@ -1205,10 +1336,23 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
                     if (start == std::string::npos) {
                         start = match.position(0);
                     }
+
+                    // replay tokens that overlap with [start, end)
+                    for (const auto & [tok, tok_pos] : grammar.trigger_buffer_positions) {
+                        auto [tok_start, tok_end] = tok_pos;
+                        if (tok_end <= start) {
+                            continue;
+                        }
+
+                        size_t piece_start = (tok_start < start) ? start : tok_start; // allow for partial token pieces
+                        size_t piece_len = tok_end - piece_start;
+                        auto tok_piece = grammar.trigger_buffer.substr(piece_start, piece_len);
+                        llama_grammar_accept_token(grammar, tok, tok_piece);
+                    }
+
                     auto constrained_str = grammar.trigger_buffer.substr(start);
-                    // std::string constrained_str(match[1].first, grammar.trigger_buffer.end());
                     grammar.trigger_buffer.clear();
-                    llama_grammar_accept_str(grammar, constrained_str);
+                    grammar.trigger_buffer_positions.clear();
                     LLAMA_LOG_DEBUG("Grammar triggered on regex: '%s'\n", constrained_str.c_str());
                     return;
                 }
@@ -1228,7 +1372,7 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
         GGML_ABORT("grammar error: end of grammar token received but grammar stack is not empty");
     }
 
-    llama_grammar_accept_str(grammar, piece);
+    llama_grammar_accept_token(grammar, token, piece);
 }
 
 void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string & piece) {
@@ -1246,6 +1390,61 @@ void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string
     }
 }
 
+void llama_grammar_accept_token(struct llama_grammar & grammar, llama_token token, const std::string & piece) {
+    // Note terminating 0 in decoded string
+    const auto   decoded     = decode_utf8(piece, grammar.partial_utf8);
+    const auto & code_points = decoded.first;
+
+    llama_grammar_stacks stacks_new;
+    stacks_new.reserve(grammar.stacks.size());
+
+    for (const auto & stack : grammar.stacks) {
+        if (stack.empty()) {
+            continue;
+        }
+
+        const llama_grammar_element * pos = stack.back();
+
+        if (pos->type == LLAMA_GRETYPE_TOKEN || pos->type == LLAMA_GRETYPE_TOKEN_NOT) {
+            if (llama_grammar_match_token(pos, token)) {
+                llama_grammar_stack new_stack(stack.begin(), stack.end() - 1);
+                if (!llama_grammar_is_end_of_sequence(pos + 1)) {
+                    new_stack.push_back(pos + 1);
+                }
+                llama_grammar_advance_stack(grammar.rules, new_stack, stacks_new);
+            }
+        } else {
+            llama_grammar_stacks current_stacks = {stack};
+
+            for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
+                llama_grammar_stacks next_stacks;
+
+                for (const auto & cur_stack : current_stacks) {
+                    llama_grammar_accept_chr(grammar, cur_stack, *it, next_stacks);
+                }
+
+                current_stacks = std::move(next_stacks);
+                if (current_stacks.empty()) {
+                    break;
+                }
+            }
+
+            for (auto & surviving_stack : current_stacks) {
+                if (std::find(stacks_new.begin(), stacks_new.end(), surviving_stack) == stacks_new.end()) {
+                    stacks_new.emplace_back(surviving_stack);
+                }
+            }
+        }
+    }
+
+    grammar.stacks = std::move(stacks_new);
+    grammar.partial_utf8 = decoded.second;
+
+    if (grammar.stacks.empty()) {
+        throw std::runtime_error("Unexpected empty grammar stack after accepting piece: " + piece + " (" + std::to_string(token) + ")");
+    }
+}
+
 
 const std::string & ollama_vocab::token_to_piece(const uint32_t token) const {
     try {

llama/llama.cpp/src/llama-grammar.h

@@ -47,11 +47,17 @@ enum llama_gretype {
 
     // any character (.)
     LLAMA_GRETYPE_CHAR_ANY       = 7,
+
+    // terminal element: token (<[token-id]>)
+    LLAMA_GRETYPE_TOKEN          = 8,
+
+    // inverse token (!<[token-id]>)
+    LLAMA_GRETYPE_TOKEN_NOT      = 9,
 };
 
 typedef struct llama_grammar_element {
     enum llama_gretype type;
-    uint32_t           value; // Unicode code point or rule ID
+    uint32_t           value; // Unicode code point, rule ID, or token ID
 } llama_grammar_element;
 
 struct llama_partial_utf8 {
@@ -63,6 +69,7 @@ struct llama_grammar_candidate {
     size_t               index;
     const uint32_t     * code_points;
     llama_partial_utf8   partial_utf8;
+    llama_token          id;
 };
 
 using llama_grammar_rule  = std::vector<      llama_grammar_element>;
@@ -88,10 +95,13 @@ std::vector<llama_grammar_candidate> llama_grammar_reject_candidates_for_stack(
         const llama_grammar_candidates & candidates);
 
 struct llama_grammar_parser {
+    const llama_vocab * vocab;
     std::map<std::string, uint32_t> symbol_ids;
 
     llama_grammar_rules rules;
 
+    llama_grammar_parser(const struct llama_vocab * vocab = nullptr) : vocab(vocab) {}
+
     llama_grammar_stack c_rules() const;
 
     uint32_t get_symbol_id(const char * src, size_t len);
@@ -123,6 +133,9 @@ struct llama_grammar_trigger_pattern {
 };
 
 struct llama_grammar {
+    // maintain a list of llama_tokens and their positions in the trigger_buffer
+    using token_pos = std::pair<llama_token, std::pair<size_t, size_t>>;
+
     // note: allow null vocab for testing (not great)
     const llama_vocab * vocab;
     const ollama_vocab * o_vocab;
@@ -139,6 +152,7 @@ struct llama_grammar {
     bool                     lazy             = false;
     bool                     awaiting_trigger = false; // Initialized to true for lazy grammars only
     std::string              trigger_buffer;           // Output buffered by lazy grammar. Will be cleared once trigger is found.
+    std::vector<token_pos>   trigger_buffer_positions; // Tokens buffered by lazy grammar. Used to replay when a trigger is found.
     std::vector<llama_token> trigger_tokens;           // Tokens that trigger a lazy grammar, or tokens to force printing of (even if special).
     std::vector<llama_grammar_trigger_pattern>
                              trigger_patterns;         // Regular expressions that trigger a lazy grammar. Must be a full match of the entire generated
@@ -185,3 +199,8 @@ void llama_grammar_accept_impl(
 void llama_grammar_accept_str(
               struct llama_grammar & grammar,
                  const std::string & piece);
+
+void llama_grammar_accept_token(
+              struct llama_grammar & grammar,
+                       llama_token   token,
+                 const std::string & piece);

llama/llama.cpp/src/llama-graph.cpp

@@ -71,6 +71,9 @@ void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && attn_scale) {
         const int64_t n_tokens = ubatch->n_tokens;
 
+        GGML_ASSERT(f_attn_temp_scale != 0.0f);
+        GGML_ASSERT(n_attn_temp_floor_scale != 0);
+
         std::vector<float> attn_scale_data(n_tokens, 0.0f);
         for (int i = 0; i < n_tokens; ++i) {
             const float pos = ubatch->pos[i];
@@ -810,9 +813,6 @@ ggml_tensor * llm_graph_context::build_ffn(
             GGML_ABORT("fatal error");
     }
 
-    //expand here so that we can fuse ffn gate
-    ggml_build_forward_expand(gf, cur);
-
     if (gate && type_gate == LLM_FFN_PAR) {
         cur = ggml_mul(ctx0, cur, tmp);
         cb(cur, "ffn_gate_par", il);
@@ -973,7 +973,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
 
         // mask out the other groups
         selection_probs = ggml_get_rows(ctx0, selection_groups, expert_groups); // [n_exp_per_group, n_group_used, n_tokens]
-        selection_probs = ggml_set_rows(ctx0, ggml_scale_bias(ctx0, selection_groups, 0.0f, -INFINITY), selection_probs, expert_groups); // [n_exp_per_group, n_expert_groups, n_tokens]
+        selection_probs = ggml_set_rows(ctx0, ggml_fill(ctx0, selection_groups, -INFINITY), selection_probs, expert_groups); // [n_exp_per_group, n_expert_groups, n_tokens]
         selection_probs = ggml_reshape_2d(ctx0, selection_probs, n_expert, n_tokens); // [n_expert, n_tokens]
         cb(selection_probs, "ffn_moe_probs_masked", il);
     }
@@ -1093,9 +1093,6 @@ ggml_tensor * llm_graph_context::build_moe_ffn(
             GGML_ABORT("fatal error");
     }
 
-    //expand here so that we can fuse ffn gate
-    ggml_build_forward_expand(gf, cur);
-
     experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens]
     cb(experts, "ffn_moe_down", il);
 

llama/llama.cpp/src/llama-hparams.h

@@ -164,8 +164,8 @@ struct llama_hparams {
     // llama4 smallthinker
     uint32_t n_moe_layer_step        = 0;
     uint32_t n_no_rope_layer_step    = 4;
-    uint32_t n_attn_temp_floor_scale = 8192;
-    float    f_attn_temp_scale       = 0.1;
+    uint32_t n_attn_temp_floor_scale = 0;
+    float    f_attn_temp_scale       = 0.0f;
 
     // gemma3n altup
     uint32_t n_altup      = 4; // altup_num_inputs

llama/llama.cpp/src/llama-impl.h

@@ -37,7 +37,7 @@ void llama_log_callback_default(ggml_log_level level, const char * text, void *
 template <typename T>
 struct no_init {
     T value;
-    no_init() { /* do nothing */ }
+    no_init() = default;
 };
 
 struct time_meas {

llama/llama.cpp/src/llama-mmap.cpp

@@ -485,7 +485,7 @@ struct llama_mlock::impl {
         if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit)) {
             suggest = false;
         }
-        if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size)) {
+        if (suggest && ((uint64_t)lock_limit.rlim_max > (uint64_t)lock_limit.rlim_cur + size)) {
             suggest = false;
         }
 #endif

llama/llama.cpp/src/llama-model.cpp

@@ -423,8 +423,8 @@ static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, llama_split_mode s
 }
 
 struct llama_model::impl {
-    impl() {}
-    ~impl() {}
+    impl() = default;
+    ~impl() = default;
 
     uint64_t n_elements = 0;
 
@@ -461,7 +461,7 @@ llama_model::llama_model(const llama_model_params & params) : params(params), pi
     pimpl->has_tensor_overrides = params.tensor_buft_overrides && params.tensor_buft_overrides[0].pattern;
 }
 
-llama_model::~llama_model() {}
+llama_model::~llama_model() = default;
 
 void llama_model::load_stats(llama_model_loader & ml) {
     pimpl->n_elements = ml.n_elements;
@@ -665,6 +665,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 } else {
                     hparams.swa_type                = LLAMA_SWA_TYPE_CHUNKED;
                     hparams.n_swa                   = 8192;
+                    hparams.n_attn_temp_floor_scale = 8192;
+                    hparams.f_attn_temp_scale       = 0.1f;
                     hparams.set_swa_pattern(4);   // pattern: 3 chunked - 1 full
                 }
 
@@ -1262,18 +1264,25 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_GEMMA3:
             {
+                const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+                if (found_swa && hparams.n_swa > 0) {
                     hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
                     hparams.set_swa_pattern(6);
 
                     hparams.rope_freq_base_train_swa  = 10000.0f;
                     hparams.rope_freq_scale_train_swa = 1.0f;
+                } else {
+                    hparams.swa_type = LLAMA_SWA_TYPE_NONE;
+                }
 
-                ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
+                hparams.f_final_logit_softcapping = 0.0f;
+                ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false);
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
                 switch (hparams.n_layer) {
                     case 18: type = LLM_TYPE_270M; break;
                     case 26: type = LLM_TYPE_1B; break;
+                    case 32: type = LLM_TYPE_8B; break; // Rnj-1
                     case 34: type = LLM_TYPE_4B; break;
                     case 48: type = LLM_TYPE_12B; break;
                     case 62: type = LLM_TYPE_27B; break;
@@ -1597,8 +1606,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
                 ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale);
 
-                switch (hparams.n_layer) {
-                    case 28: type = LLM_TYPE_20B; break;
+                switch (hparams.n_ff_exp) {
+                    case 1408: type = LLM_TYPE_16B; break;
+                    case 1792: type = LLM_TYPE_20B; break;
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
@@ -1626,6 +1636,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 }
                 ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, false);
 
+                // (optional) temperature tuning - used by mistral-large
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE,  hparams.f_attn_temp_scale,       false);
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.n_attn_temp_floor_scale, false);
+
                 switch (hparams.n_layer) {
                     case 27: type = LLM_TYPE_16B; break;
                     case 60: type = LLM_TYPE_236B; break;
@@ -2262,6 +2276,42 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                     default: type = LLM_TYPE_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_MISTRAL3:
+            {
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE, hparams.f_attn_temp_scale, false);
+
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_FAST,   hparams.yarn_beta_fast, false);
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,   hparams.yarn_beta_slow, false);
+                ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,     hparams.rope_yarn_log_mul, false);
+
+                // TODO: maybe add n_attn_temp_floor_scale as a separate KV?
+                if (hparams.f_attn_temp_scale != 0.0f) {
+                    hparams.n_attn_temp_floor_scale = hparams.n_ctx_orig_yarn;
+                    if (hparams.n_attn_temp_floor_scale == 0) {
+                        throw std::runtime_error("invalid n_ctx_orig_yarn for attention temperature scaling");
+                    }
+                }
+
+                // TODO: this seems to be correct with the case of mscale == mscale_all_dims == 1.0f
+                //       but may need further verification with other values
+                if (hparams.rope_yarn_log_mul != 0.0f) {
+                    float factor = 1.0f / hparams.rope_freq_scale_train;
+                    float mscale = 1.0f;
+                    float mscale_all_dims = hparams.rope_yarn_log_mul;
+                    static auto get_mscale = [](float scale, float mscale) {
+                        return scale <= 1.0f ? 1.0f : (0.1f * mscale * logf(scale) + 1.0f);
+                    };
+                    hparams.yarn_attn_factor = get_mscale(factor, mscale) / get_mscale(factor, mscale_all_dims);
+                }
+
+                switch (hparams.n_layer) {
+                    case 26: type = LLM_TYPE_3B; break;
+                    case 34: type = LLM_TYPE_8B; break;
+                    case 40: type = LLM_TYPE_14B; break;
+                    default: type = LLM_TYPE_UNKNOWN;
+                }
+            } break;
         default: throw std::runtime_error("unsupported model architecture");
     }
 
@@ -2575,6 +2625,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
             case LLM_ARCH_MINICPM:
             case LLM_ARCH_GRANITE:
             case LLM_ARCH_GRANITE_MOE:
+            case LLM_ARCH_MISTRAL3:
                 {
                     tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
 
@@ -6530,7 +6581,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                             layer.ssm_in         = create_tensor(tn(LLM_TENSOR_SSM_IN,         "weight", i), { n_embd, qkvz_dim }, 0);
                             layer.ssm_conv1d     = create_tensor(tn(LLM_TENSOR_SSM_CONV1D,     "weight", i), { hparams.ssm_d_conv, conv_dim }, 0);
                             layer.ssm_dt         = create_tensor(tn(LLM_TENSOR_SSM_DT,         "bias",   i), { hparams.ssm_dt_rank }, 0);
-                            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A,                    i), { hparams.ssm_dt_rank }, 0);
+                            layer.ssm_a          = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN,             i), { hparams.ssm_dt_rank }, 0);
                             layer.ssm_beta_alpha = create_tensor(tn(LLM_TENSOR_SSM_BETA_ALPHA, "weight", i), { n_embd, ba_dim }, 0);
                             layer.ssm_norm       = create_tensor(tn(LLM_TENSOR_SSM_NORM,       "weight", i), { head_v_dim }, 0);
                             layer.ssm_out        = create_tensor(tn(LLM_TENSOR_SSM_OUT,        "weight", i), { value_dim, n_embd }, 0);
@@ -7304,7 +7355,11 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             } break;
         case LLM_ARCH_GEMMA3:
             {
-                llm = std::make_unique<llm_build_gemma3_iswa>(*this, params);
+                if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
+                    llm = std::make_unique<llm_build_gemma3<true>>(*this, params);
+                } else {
+                    llm = std::make_unique<llm_build_gemma3<false>>(*this, params);
+                }
             } break;
         case LLM_ARCH_GEMMA3N:
             {
@@ -7569,6 +7624,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
             {
                 llm = std::make_unique<llm_build_qwen3next>(*this, params);
             } break;
+        case LLM_ARCH_MISTRAL3:
+            {
+                llm = std::make_unique<llm_build_mistral3>(*this, params);
+            } break;
         default:
             GGML_ABORT("fatal error");
     }
@@ -7738,6 +7797,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_ARCEE:
         case LLM_ARCH_ERNIE4_5:
         case LLM_ARCH_ERNIE4_5_MOE:
+        case LLM_ARCH_MISTRAL3:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2

llama/llama.cpp/src/llama-quant.cpp

@@ -666,7 +666,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
 
     std::map<int, std::string> mapped;
     int blk_id = 0;
-    int pruned_attention_w = 0;
 
     // make a list of weights
     std::vector<const llama_model_loader::llama_tensor_weight *> tensors;
@@ -674,11 +673,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
     for (const auto & it : ml.weights_map) {
         const std::string remapped_name(remap_layer(it.first, prune_list, mapped, blk_id));
         if (remapped_name.empty()) {
-            if (it.first.find("attn_v.weight") != std::string::npos ||
-                it.first.find("attn_qkv.weight") != std::string::npos ||
-                it.first.find("attn_kv_b.weight") != std::string::npos) {
-                    pruned_attention_w++;
-            }
             LLAMA_LOG_DEBUG("%s: pruning tensor %s\n", __func__, it.first.c_str());
             continue;
         }
@@ -703,7 +697,6 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         });
     }
 
-    bool is_clip_model = false;
     for (const auto * it : tensors) {
         const struct ggml_tensor * tensor = it->tensor;
 
@@ -717,32 +710,10 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         } else if (name == LLM_TN(model.arch)(LLM_TENSOR_OUTPUT, "weight")) {
             qs.has_output = true;
         }
-
-        is_clip_model |= name.rfind("mm.", 0) == 0; // check the "mm." prefix
     }
 
     qs.n_ffn_down = qs.n_ffn_gate = qs.n_ffn_up = (int)model.hparams.n_layer;
 
-    // sanity checks for models that have attention layers
-    if (qs.n_attention_wv != 0 && !is_clip_model)
-    {
-        const auto & n_head_kv_iter = model.hparams.n_head_kv_arr.begin();
-        // attention layers have a non-zero number of kv heads
-        int32_t n_layer_attn = model.hparams.n_layer - std::count(n_head_kv_iter, n_head_kv_iter + model.hparams.n_layer, 0);
-        if (llama_model_has_encoder(&model)) {
-            // now n_layer_attn is the number of attention layers in the encoder
-            // for each decoder block, there are 2 attention layers
-            n_layer_attn += 2 * model.hparams.dec_n_layer;
-        }
-
-        // note: for linear-attention models (such as Qwen3 Next) this is the number of linear layers
-        const int32_t n_layer_recr = std::count(model.hparams.recurrent_layer_arr.begin(), model.hparams.recurrent_layer_arr.end(), true);
-
-        LLAMA_LOG_INFO("%s: n_layer_attn = %d, n_layer_recr = %d, pruned_attention_w = %d\n", __func__, n_layer_attn, n_layer_recr, pruned_attention_w);
-
-        GGML_ASSERT((qs.n_attention_wv == n_layer_attn - pruned_attention_w - n_layer_recr) && "n_attention_wv is unexpected");
-    }
-
     size_t total_size_org = 0;
     size_t total_size_new = 0;
 

llama/llama.cpp/src/llama-vocab.cpp

@@ -3243,8 +3243,7 @@ void llama_vocab::impl::print_info() const {
 llama_vocab::llama_vocab() : pimpl(new impl(*this)) {
 }
 
-llama_vocab::~llama_vocab() {
-}
+llama_vocab::~llama_vocab() = default;
 
 void llama_vocab::load(llama_model_loader & ml, const LLM_KV & kv) {
     pimpl->load(ml, kv);

llama/llama.cpp/src/models/deepseek2.cpp

@@ -30,6 +30,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
     // {n_embd, n_tokens}
     inpL = build_inp_embd(model.tok_embd);
 
+    // (optional) temperature tuning - used by mistral-large
+    ggml_tensor * inp_attn_scale = nullptr;
+    if (hparams.f_attn_temp_scale != 0.0f) {
+        inp_attn_scale = build_inp_attn_scale();
+    }
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -128,6 +134,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 ggml_tensor * Vcur = kv_cmpr;
                 cb(Vcur, "Vcur", il);
 
+                if (inp_attn_scale) {
+                    // apply llama 4 temperature scaling
+                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                    cb(Qcur, "Qcur_attn_temp_scaled", il);
+                }
+
                 // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
                 cur = build_attn(inp_attn,
                         model.layers[il].wo, NULL,
@@ -160,6 +172,12 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr
                 ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0);
                 cb(Kcur, "Kcur", il);
 
+                if (inp_attn_scale) {
+                    // apply llama 4 temperature scaling
+                    Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                    cb(Qcur, "Qcur_attn_temp_scaled", il);
+                }
+
                 // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
                 cur = build_attn(inp_attn,
                             model.layers[il].wo, NULL,

llama/llama.cpp/src/models/gemma3.cpp

@@ -1,6 +1,7 @@
 #include "models.h"
 
-llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+template <bool iswa>
+llm_build_gemma3<iswa>::llm_build_gemma3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_k;
 
     ggml_tensor * cur;
@@ -17,13 +18,28 @@ llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const ll
     ggml_tensor * inp_pos = build_inp_pos();
 
     // TODO: is causal == true correct? might need some changes
-    auto * inp_attn = build_attn_inp_kv_iswa();
+    using inp_attn_type = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa, llm_graph_input_attn_kv>;
+    inp_attn_type * inp_attn = nullptr;
+
+    if constexpr (iswa) {
+        inp_attn = build_attn_inp_kv_iswa();
+    } else {
+        inp_attn = build_attn_inp_kv();
+    }
 
     ggml_tensor * inp_out_ids = build_inp_out_ids();
 
     for (int il = 0; il < n_layer; ++il) {
-        const float freq_base_l  = model.get_rope_freq_base (cparams, il);
-        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+        float freq_base_l  = 0.0f;
+        float freq_scale_l = 0.0f;
+
+        if constexpr (iswa) {
+            freq_base_l  = model.get_rope_freq_base (cparams, il);
+            freq_scale_l = model.get_rope_freq_scale(cparams, il);
+        } else {
+            freq_base_l  = freq_base;
+            freq_scale_l = freq_scale;
+        }
 
         // norm
         cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
@@ -102,7 +118,7 @@ llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const ll
         cur = build_norm(cur,
                 model.layers[il].ffn_post_norm, NULL,
                 LLM_NORM_RMS, -1);
-        cb(cur, "ffn_post_norm", -1);
+        cb(cur, "ffn_post_norm", il);
 
         cur = ggml_add(ctx0, cur, sa_out);
 
@@ -124,8 +140,17 @@ llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const ll
     // lm_head
     cur = build_lora_mm(model.output, cur);
 
+    if (hparams.f_final_logit_softcapping) {
+        cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
+        cur = ggml_tanh(ctx0, cur);
+        cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
+    }
+
     cb(cur, "result_output", -1);
     res->t_logits = cur;
 
     ggml_build_forward_expand(gf, cur);
 }
+
+template struct llm_build_gemma3<false>;
+template struct llm_build_gemma3<true>;

llama/llama.cpp/src/models/mistral3.cpp

@@ -0,0 +1,160 @@
+#include "models.h"
+
+llm_build_mistral3::llm_build_mistral3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v;
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+    GGML_ASSERT(n_embd_head == hparams.n_rot);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    // (optional) temperature tuning
+    ggml_tensor * inp_attn_scale = nullptr;
+    if (hparams.f_attn_temp_scale != 0.0f) {
+        inp_attn_scale = build_inp_attn_scale();
+    }
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            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);
+            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);
+            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);
+            cb(Vcur, "Vcur", il);
+            if (model.layers[il].bv) {
+                Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+                cb(Vcur, "Vcur", il);
+            }
+            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens);
+            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
+            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
+
+            Qcur = ggml_rope_ext(
+                    ctx0, Qcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            Kcur = ggml_rope_ext(
+                    ctx0, Kcur, inp_pos, rope_factors,
+                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                    );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            if (inp_attn_scale) {
+                // apply llama 4 temperature scaling
+                Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
+                cb(Qcur, "Qcur_attn_temp_scaled", il);
+            }
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].bo,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network (non-MoE)
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            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,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_norm(ffn_inp,
+                    model.layers[il].ffn_norm, NULL,
+                    LLM_NORM_RMS, il);
+            cb(cur, "ffn_norm", il);
+
+            cur = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    nullptr,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    false, 0.0,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
+                    il);
+            cb(cur, "ffn_moe_out", il);
+        }
+        cur = ggml_add(ctx0, cur, ffn_inp);
+        cb(cur, "ffn_out", il);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

llama/llama.cpp/src/models/models.h

@@ -179,8 +179,9 @@ struct llm_build_gemma2_iswa : public llm_graph_context {
     llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params);
 };
 
-struct llm_build_gemma3_iswa : public llm_graph_context {
-    llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params);
+template <bool iswa>
+struct llm_build_gemma3 : public llm_graph_context {
+    llm_build_gemma3(const llama_model & model, const llm_graph_params & params);
 };
 
 struct llm_build_gemma3n_iswa : public llm_graph_context {
@@ -322,6 +323,10 @@ struct llm_build_minimax_m2 : public llm_graph_context {
     llm_build_minimax_m2(const llama_model & model, const llm_graph_params & params);
 };
 
+struct llm_build_mistral3 : public llm_graph_context {
+    llm_build_mistral3(const llama_model & model, const llm_graph_params & params);
+};
+
 struct llm_build_mpt : public llm_graph_context {
     llm_build_mpt(const llama_model & model, const llm_graph_params & params);
 };

llama/llama.cpp/src/unicode.cpp

@@ -520,7 +520,7 @@ static std::vector<size_t> unicode_regex_split_custom_llama3(const std::string &
 
 // use std::wregex to split the text
 static std::vector<size_t> unicode_regex_split_stl(const std::wstring & wtext, const std::wstring & regex_expr, const std::vector<size_t> & offsets) {
-    std::wregex expr(regex_expr);
+    std::wregex expr(regex_expr, std::regex_constants::optimize | std::regex_constants::nosubs);
     std::vector<size_t> bpe_offsets; // store the offset of each word
     bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
     size_t start = 0;
@@ -550,7 +550,7 @@ static std::vector<size_t> unicode_regex_split_stl(const std::wstring & wtext, c
 
 // use std::regex to split the text
 static std::vector<size_t> unicode_regex_split_stl(const std::string & text, const std::string & regex_expr, const std::vector<size_t> & offsets) {
-    std::regex expr(regex_expr);
+    std::regex expr(regex_expr, std::regex_constants::optimize | std::regex_constants::nosubs);
     std::vector<size_t> bpe_offsets; // store the offset of each word
     bpe_offsets.reserve(offsets.size()); // Reserve memory for the approximate size
     size_t start = 0;

llama/llama.cpp/tools/mtmd/clip.cpp

@@ -441,6 +441,7 @@ struct clip_ctx {
     int max_nodes = 8192;
     ggml_backend_sched_ptr sched;
     clip_flash_attn_type flash_attn_type = CLIP_FLASH_ATTN_TYPE_AUTO;
+    bool is_allocated = false;
 
     // for debugging
     bool debug_graph = false;
@@ -2033,7 +2034,7 @@ private:
         ggml_tensor * pos_embd = model.position_embeddings;
         const int height       = img.ny / patch_size;
         const int width        = img.nx / patch_size;
-        const uint32_t mode    = GGML_SCALE_MODE_BILINEAR;
+        const uint32_t mode    = GGML_SCALE_MODE_BILINEAR | GGML_SCALE_FLAG_ANTIALIAS;
         const int n_per_side   = (int)std::sqrt(pos_embd->ne[1]);
 
         GGML_ASSERT(pos_embd);
@@ -2812,7 +2813,8 @@ struct clip_model_loader {
                     {
                         get_u32(KEY_PROJ_SCALE_FACTOR, hparams.n_merge, false);
                         // ref: https://huggingface.co/LiquidAI/LFM2-VL-3B/blob/main/preprocessor_config.json
-                        hparams.set_limit_image_tokens(64, 256);
+                        // config above specifies number of tokens after downsampling, while here it is before, relax lowerbound to 64
+                        hparams.set_limit_image_tokens(64, 1024);
                     } break;
                 case PROJECTOR_TYPE_PIXTRAL:
                 case PROJECTOR_TYPE_LIGHTONOCR:
@@ -3347,12 +3349,30 @@ struct clip_model_loader {
     };
 
     static void warmup(clip_ctx & ctx_clip) {
+        // create a fake batch
+        const auto & hparams = ctx_clip.model.hparams;
+        clip_image_f32_batch batch;
+        clip_image_f32_ptr img(clip_image_f32_init());
+        if (ctx_clip.model.modality == CLIP_MODALITY_VISION) {
+            img->nx = hparams.warmup_image_size;
+            img->ny = hparams.warmup_image_size;
+            LOG_INF("%s: warmup with image size = %d x %d\n", __func__, img->nx, img->ny);
+        } else {
+            img->nx = hparams.warmup_audio_size;
+            img->ny = hparams.n_mel_bins;
+            LOG_INF("%s: warmup with audio size = %d\n", __func__, img->nx);
+        }
+        batch.entries.push_back(std::move(img));
+        warmup(ctx_clip, batch);
+    }
+
+    static void warmup(clip_ctx & ctx_clip, const clip_image_f32_batch & batch) {
         support_info_graph info;
 
         if (ctx_clip.flash_attn_type == CLIP_FLASH_ATTN_TYPE_AUTO) {
             // try to enable flash attention to see if it's supported
             ctx_clip.flash_attn_type = CLIP_FLASH_ATTN_TYPE_ENABLED;
-            info = alloc_compute_meta(ctx_clip);
+            info = alloc_compute_meta(ctx_clip, batch);
             if (!info.fattn && info.fattn_op) {
                 auto op = info.fattn_op;
                 LOG_WRN("%s: *****************************************************************\n", __func__);
@@ -3371,15 +3391,17 @@ struct clip_model_loader {
                 LOG_WRN("%s: please report this on github as an issue\n", __func__);
                 LOG_WRN("%s: *****************************************************************\n", __func__);
                 ctx_clip.flash_attn_type = CLIP_FLASH_ATTN_TYPE_DISABLED;
-                alloc_compute_meta(ctx_clip);
+                alloc_compute_meta(ctx_clip, batch);
             }
         } else {
-            info = alloc_compute_meta(ctx_clip);
+            info = alloc_compute_meta(ctx_clip, batch);
             if (!info.fattn && ctx_clip.flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) {
                 LOG_WRN("%s: flash attention is not supported by the current backend; falling back to CPU (performance will be degraded)\n", __func__);
             }
         }
 
+        ctx_clip.is_allocated = true; // mark buffers as allocated
+
         LOG_INF("%s: flash attention is %s\n", __func__,
             (ctx_clip.flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) ? "enabled" : "disabled");
 
@@ -3411,24 +3433,9 @@ struct clip_model_loader {
         }
     }
 
-    static support_info_graph alloc_compute_meta(clip_ctx & ctx_clip) {
-        const auto & hparams = ctx_clip.model.hparams;
+    static support_info_graph alloc_compute_meta(clip_ctx & ctx_clip, const clip_image_f32_batch & batch) {
         ctx_clip.buf_compute_meta.resize(ctx_clip.max_nodes * ggml_tensor_overhead() + ggml_graph_overhead());
 
-        // create a fake batch
-        clip_image_f32_batch batch;
-        clip_image_f32_ptr img(clip_image_f32_init());
-        if (ctx_clip.model.modality == CLIP_MODALITY_VISION) {
-            img->nx = hparams.warmup_image_size;
-            img->ny = hparams.warmup_image_size;
-            LOG_INF("%s: warmup with image size = %d x %d\n", __func__, img->nx, img->ny);
-        } else {
-            img->nx = hparams.warmup_audio_size;
-            img->ny = hparams.n_mel_bins;
-            LOG_INF("%s: warmup with audio size = %d\n", __func__, img->nx);
-        }
-        batch.entries.push_back(std::move(img));
-
         ggml_cgraph * gf = clip_image_build_graph(&ctx_clip, batch);
         ggml_backend_sched_reserve(ctx_clip.sched.get(), gf);
 
@@ -3568,15 +3575,19 @@ struct clip_init_result clip_init(const char * fname, struct clip_context_params
             ctx_vision = new clip_ctx(ctx_params);
             loader.load_hparams(ctx_vision->model, CLIP_MODALITY_VISION);
             loader.load_tensors(*ctx_vision);
+            if (ctx_params.warmup) {
                 loader.warmup(*ctx_vision);
             }
+        }
 
         if (loader.has_audio) {
             ctx_audio = new clip_ctx(ctx_params);
             loader.load_hparams(ctx_audio->model, CLIP_MODALITY_AUDIO);
             loader.load_tensors(*ctx_audio);
+            if (ctx_params.warmup) {
                 loader.warmup(*ctx_audio);
             }
+        }
 
     } catch (const std::exception & e) {
         LOG_ERR("%s: failed to load model '%s': %s\n", __func__, fname, e.what());
@@ -3788,12 +3799,13 @@ struct img_tool {
         const int width  = inp_size.width;
         const int height = inp_size.height;
 
+        auto round_by_factor = [f = align_size](float x) { return static_cast<int>(std::round(x / static_cast<float>(f))) * f; };
         auto ceil_by_factor  = [f = align_size](float x) { return static_cast<int>(std::ceil(x / static_cast<float>(f))) * f; };
         auto floor_by_factor = [f = align_size](float x) { return static_cast<int>(std::floor(x / static_cast<float>(f))) * f; };
 
         // always align up first
-        int h_bar = std::max(align_size, ceil_by_factor(height));
-        int w_bar = std::max(align_size, ceil_by_factor(width));
+        int h_bar = std::max(align_size, round_by_factor(height));
+        int w_bar = std::max(align_size, round_by_factor(width));
 
         if (h_bar * w_bar > max_pixels) {
             const auto beta = std::sqrt(static_cast<float>(height * width) / max_pixels);
@@ -4408,7 +4420,8 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
                 const std::array<uint8_t, 3> pad_color = {122, 116, 104};
 
                 clip_image_u8 resized_img;
-                img_tool::resize(*img, resized_img, target_size, img_tool::RESIZE_ALGO_BILINEAR, true, pad_color);
+                const bool pad = (ctx->proj_type() != PROJECTOR_TYPE_LFM2);
+                img_tool::resize(*img, resized_img, target_size, img_tool::RESIZE_ALGO_BILINEAR, pad, pad_color);
                 clip_image_f32_ptr res(clip_image_f32_init());
                 normalize_image_u8_to_f32(resized_img, *res, params.image_mean, params.image_std);
                 res_imgs->entries.push_back(std::move(res));
@@ -4666,6 +4679,11 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         return false; // only support batch size of 1
     }
 
+    // if buffers are not allocated, we need to do a warmup run to allocate them
+    if (!ctx->is_allocated) {
+        clip_model_loader::warmup(*ctx, *imgs_c_ptr);
+    }
+
     // build the inference graph
     ctx->debug_print_tensors.clear();
     ggml_backend_sched_reset(ctx->sched.get());

llama/llama.cpp/tools/mtmd/clip.h

@@ -34,6 +34,7 @@ struct clip_context_params {
     enum clip_flash_attn_type flash_attn_type;
     int image_min_tokens;
     int image_max_tokens;
+    bool warmup;
 };
 
 struct clip_init_result {

llama/llama.cpp/tools/mtmd/mtmd.cpp

@@ -118,6 +118,7 @@ mtmd_context_params mtmd_context_params_default() {
         /* image_marker      */ MTMD_DEFAULT_IMAGE_MARKER,
         /* media_marker      */ mtmd_default_marker(),
         /* flash_attn_type   */ LLAMA_FLASH_ATTN_TYPE_AUTO,
+        /* warmup            */ true,
         /* image_min_tokens  */ -1,
         /* image_max_tokens  */ -1,
     };
@@ -187,6 +188,7 @@ struct mtmd_context {
             /* flash_attn_type   */ CLIP_FLASH_ATTN_TYPE_AUTO,
             /* image_min_tokens  */ ctx_params.image_min_tokens,
             /* image_max_tokens  */ ctx_params.image_max_tokens,
+            /* warmup            */ ctx_params.warmup,
         };
 
         auto res = clip_init(mmproj_fname, ctx_clip_params);
@@ -314,6 +316,10 @@ struct mtmd_context {
             img_beg = "<|im_start|>";
             img_end = "<|im_end|>";
 
+        } else if (proj == PROJECTOR_TYPE_LFM2) {
+            img_beg = "<|image_start|>";
+            img_end = "<|image_end|>";
+
         }
     }
 

llama/llama.cpp/tools/mtmd/mtmd.h

@@ -85,6 +85,7 @@ struct mtmd_context_params {
     const char * image_marker; // deprecated, use media_marker instead
     const char * media_marker;
     enum llama_flash_attn_type flash_attn_type;
+    bool warmup; // whether to run a warmup encode pass after initialization
 
     // limit number of image tokens, only for vision models with dynamic resolution
     int image_min_tokens; // minimum number of tokens for image input (default: read from metadata)

llama/patches/0001-ggml-backend-malloc-and-free-using-the-same-compiler.patch

@@ -23,7 +23,7 @@ problem.
  8 files changed, 21 insertions(+), 2 deletions(-)
 
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index 4cf377e7f..4882541c8 100644
+index 08681f35e..afde2f0b7 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
 @@ -113,7 +113,6 @@ void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
@@ -42,7 +42,7 @@ index 4cf377e7f..4882541c8 100644
  }
  
  static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-@@ -2079,6 +2079,11 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+@@ -2106,6 +2106,11 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
  static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      GGML_ASSERT(buffer);
      ggml_aligned_free(buffer->context, buffer->size);
@@ -54,7 +54,7 @@ index 4cf377e7f..4882541c8 100644
  }
  
  static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
-@@ -2131,7 +2136,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
+@@ -2158,7 +2163,7 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
  };
  
  static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
@@ -64,7 +64,7 @@ index 4cf377e7f..4882541c8 100644
      /* .init_tensor     = */ NULL, // no initialization required
      /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
 diff --git a/ggml/src/ggml-cann/ggml-cann.cpp b/ggml/src/ggml-cann/ggml-cann.cpp
-index df28d67fb..1f6a56ba2 100644
+index 81288464c..866758782 100644
 --- a/ggml/src/ggml-cann/ggml-cann.cpp
 +++ b/ggml/src/ggml-cann/ggml-cann.cpp
 @@ -831,6 +831,7 @@ static bool ggml_backend_buffer_is_cann(ggml_backend_buffer_t buffer) {
@@ -84,10 +84,10 @@ index df28d67fb..1f6a56ba2 100644
  
  /**
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index fa7e1e13a..8f3b1c173 100644
+index 279679a4e..5145c1e88 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -579,6 +579,7 @@ struct ggml_backend_cuda_buffer_context {
+@@ -583,6 +583,7 @@ struct ggml_backend_cuda_buffer_context {
  static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
      delete ctx;
@@ -95,7 +95,7 @@ index fa7e1e13a..8f3b1c173 100644
  }
  
  static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
-@@ -834,6 +835,7 @@ struct ggml_backend_cuda_split_buffer_context {
+@@ -838,6 +839,7 @@ struct ggml_backend_cuda_split_buffer_context {
  static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
      delete ctx;
@@ -103,7 +103,7 @@ index fa7e1e13a..8f3b1c173 100644
  }
  
  static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -1115,6 +1117,7 @@ static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
+@@ -1119,6 +1121,7 @@ static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
  
  static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      CUDA_CHECK(cudaFreeHost(buffer->context));
@@ -132,10 +132,10 @@ index 70bf6f3d9..f2b7fe692 100644
  
  static void * ggml_backend_metal_buffer_private_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-opencl/ggml-opencl.cpp b/ggml/src/ggml-opencl/ggml-opencl.cpp
-index e5302f455..43fa83e8f 100644
+index 0d37587f6..ff373d413 100644
 --- a/ggml/src/ggml-opencl/ggml-opencl.cpp
 +++ b/ggml/src/ggml-opencl/ggml-opencl.cpp
-@@ -3412,6 +3412,7 @@ struct ggml_backend_opencl_buffer_context {
+@@ -3417,6 +3417,7 @@ struct ggml_backend_opencl_buffer_context {
  static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
      delete ctx;
@@ -144,10 +144,10 @@ index e5302f455..43fa83e8f 100644
  
  static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cpp
-index 48fd99a76..da2aab3df 100644
+index 18a45d2d9..89041805e 100644
 --- a/ggml/src/ggml-rpc/ggml-rpc.cpp
 +++ b/ggml/src/ggml-rpc/ggml-rpc.cpp
-@@ -555,6 +555,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -556,6 +556,7 @@ static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, &request, sizeof(request), nullptr, 0);
      RPC_STATUS_ASSERT(status);
      delete ctx;
@@ -156,7 +156,7 @@ index 48fd99a76..da2aab3df 100644
  
  static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
 diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
-index 3f1bdfb9f..a95c2f305 100644
+index 7449a9160..e69a1ff5f 100644
 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp
 +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
 @@ -355,6 +355,7 @@ ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
@@ -184,10 +184,10 @@ index 3f1bdfb9f..a95c2f305 100644
  
  static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index 66dd0bfab..83cdec29e 100644
+index c6f5809cc..c801d2fd2 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -12368,6 +12368,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+@@ -12271,6 +12271,7 @@ static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
      ggml_vk_destroy_buffer(ctx->dev_buffer);
      delete ctx;
@@ -195,7 +195,7 @@ index 66dd0bfab..83cdec29e 100644
  }
  
  static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
-@@ -12511,6 +12512,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
+@@ -12414,6 +12415,7 @@ static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffe
  static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
      VK_LOG_MEMORY("ggml_backend_vk_host_buffer_free_buffer()");
      ggml_vk_host_free(vk_instance.devices[0], buffer->context);

llama/patches/0002-pretokenizer.patch

@@ -10,7 +10,7 @@ logs instead of throwing an error
  1 file changed, 3 insertions(+), 11 deletions(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index a73c4c448..b9f0631f4 100644
+index e2cca66e4..8246a0a14 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -1825,16 +1825,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {

llama/patches/0003-clip-unicode.patch

@@ -10,7 +10,7 @@ filesystems for paths that include wide characters
  1 file changed, 39 insertions(+)
 
 diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
-index 05777d2d9..f4c4d2c48 100644
+index 3ed08a0fe..6be1470ad 100644
 --- a/tools/mtmd/clip.cpp
 +++ b/tools/mtmd/clip.cpp
 @@ -24,6 +24,19 @@
@@ -33,7 +33,7 @@ index 05777d2d9..f4c4d2c48 100644
  struct clip_logger_state g_logger_state = {clip_log_callback_default, NULL};
  
  enum ffn_op_type {
-@@ -3255,7 +3268,29 @@ struct clip_model_loader {
+@@ -3257,7 +3270,29 @@ struct clip_model_loader {
          {
              std::vector<uint8_t> read_buf;
  
@@ -63,7 +63,7 @@ index 05777d2d9..f4c4d2c48 100644
              if (!fin) {
                  throw std::runtime_error(string_format("%s: failed to open %s\n", __func__, fname.c_str()));
              }
-@@ -3282,7 +3317,11 @@ struct clip_model_loader {
+@@ -3284,7 +3319,11 @@ struct clip_model_loader {
                      ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
                  }
              }

llama/patches/0004-solar-pro.patch

@@ -19,7 +19,7 @@ adds support for the Solar Pro architecture
  create mode 100644 src/models/solar.cpp
 
 diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
-index 67c7807e0..fda881640 100644
+index 4192af7c0..bd44d73e7 100644
 --- a/src/CMakeLists.txt
 +++ b/src/CMakeLists.txt
 @@ -125,6 +125,7 @@ add_library(llama
@@ -31,7 +31,7 @@ index 67c7807e0..fda881640 100644
              models/starcoder.cpp
              models/starcoder2.cpp
 diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
-index 8571a2e02..b6bde25d5 100644
+index 64ad1b776..a5fe4f66c 100644
 --- a/src/llama-arch.cpp
 +++ b/src/llama-arch.cpp
 @@ -85,6 +85,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
@@ -42,15 +42,15 @@ index 8571a2e02..b6bde25d5 100644
      { LLM_ARCH_WAVTOKENIZER_DEC, "wavtokenizer-dec" },
      { LLM_ARCH_PLM,              "plm"              },
      { LLM_ARCH_BAILINGMOE,       "bailingmoe"       },
-@@ -204,6 +205,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
-     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
+@@ -206,6 +207,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
      { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
      { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
+     { LLM_KV_ATTENTION_TEMPERATURE_SCALE,            "%s.attention.temperature_scale"            },
 +    { LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,        "%s.attention.block_skip_connection"        },
      { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },
      { LLM_KV_ATTENTION_VALUE_LENGTH_MLA,             "%s.attention.value_length_mla"             },
  
-@@ -2023,6 +2025,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
+@@ -2025,6 +2027,24 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
              { LLM_TENSOR_ATTN_K_NORM,     "blk.%d.attn_k_norm" },
          },
      },
@@ -75,7 +75,7 @@ index 8571a2e02..b6bde25d5 100644
      {
          LLM_ARCH_WAVTOKENIZER_DEC,
          {
-@@ -2681,6 +2701,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
+@@ -2710,6 +2730,7 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
      {LLM_TENSOR_LAUREL_POST_NORM,           {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      // this tensor is loaded for T5, but never used
      {LLM_TENSOR_DEC_CROSS_ATTN_REL_B,       {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
@@ -84,7 +84,7 @@ index 8571a2e02..b6bde25d5 100644
      {LLM_TENSOR_POS_NET_NORM,               {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
      {LLM_TENSOR_POS_NET_NORM1,              {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
 diff --git a/src/llama-arch.h b/src/llama-arch.h
-index 150646478..3936a4687 100644
+index e11318002..ec9e3a6df 100644
 --- a/src/llama-arch.h
 +++ b/src/llama-arch.h
 @@ -89,6 +89,7 @@ enum llm_arch {
@@ -95,15 +95,15 @@ index 150646478..3936a4687 100644
      LLM_ARCH_WAVTOKENIZER_DEC,
      LLM_ARCH_PLM,
      LLM_ARCH_BAILINGMOE,
-@@ -208,6 +209,7 @@ enum llm_kv {
-     LLM_KV_ATTENTION_SCALE,
+@@ -210,6 +211,7 @@ enum llm_kv {
      LLM_KV_ATTENTION_OUTPUT_SCALE,
      LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
+     LLM_KV_ATTENTION_TEMPERATURE_SCALE,
 +    LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
      LLM_KV_ATTENTION_KEY_LENGTH_MLA,
      LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
  
-@@ -459,6 +461,7 @@ enum llm_tensor {
+@@ -462,6 +464,7 @@ enum llm_tensor {
      LLM_TENSOR_ENC_OUTPUT_NORM,
      LLM_TENSOR_CLS,
      LLM_TENSOR_CLS_OUT,
@@ -131,7 +131,7 @@ index 8cdbaf69f..41127bf91 100644
      if (il < n_layer) {
          return swa_layers[il];
 diff --git a/src/llama-hparams.h b/src/llama-hparams.h
-index c3a53be79..2ffe7dd30 100644
+index 6eff334a5..a778fc3cf 100644
 --- a/src/llama-hparams.h
 +++ b/src/llama-hparams.h
 @@ -64,6 +64,8 @@ struct llama_hparams {
@@ -167,10 +167,10 @@ index aa3a65f87..ee303bd58 100644
  llama_model_loader::llama_model_loader(
          const std::string & fname,
 diff --git a/src/llama-model.cpp b/src/llama-model.cpp
-index c2a545531..4468de2f9 100644
+index 04fccc979..3c503b424 100644
 --- a/src/llama-model.cpp
 +++ b/src/llama-model.cpp
-@@ -1961,6 +1961,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
+@@ -1975,6 +1975,21 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                      default: type = LLM_TYPE_UNKNOWN;
                 }
              } break;
@@ -192,7 +192,7 @@ index c2a545531..4468de2f9 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
-@@ -5350,6 +5365,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
+@@ -5401,6 +5416,34 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
  
                          layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
  
@@ -227,7 +227,7 @@ index c2a545531..4468de2f9 100644
                          layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
                          layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
                          layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
-@@ -7425,6 +7468,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
+@@ -7480,6 +7523,10 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const {
              {
                  llm = std::make_unique<llm_build_chameleon>(*this, params);
              } break;
@@ -238,7 +238,7 @@ index c2a545531..4468de2f9 100644
          case LLM_ARCH_WAVTOKENIZER_DEC:
              {
                  llm = std::make_unique<llm_build_wavtokenizer_dec>(*this, params);
-@@ -7684,6 +7731,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
+@@ -7743,6 +7790,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
          case LLM_ARCH_GRANITE_MOE:
          case LLM_ARCH_GRANITE_HYBRID:
          case LLM_ARCH_CHAMELEON:
@@ -268,10 +268,10 @@ index f8342cf2c..cbf4e1bfa 100644
  
      struct llama_layer_convnext convnext;
 diff --git a/src/models/models.h b/src/models/models.h
-index 7ba225b47..71fea796d 100644
+index 6494f5450..e0aec822c 100644
 --- a/src/models/models.h
 +++ b/src/models/models.h
-@@ -510,6 +510,11 @@ struct llm_build_smollm3 : public llm_graph_context {
+@@ -515,6 +515,11 @@ struct llm_build_smollm3 : public llm_graph_context {
      llm_build_smollm3(const llama_model & model, const llm_graph_params & params);
  };
  

llama/patches/0005-fix-deepseek-deseret-regex.patch

@@ -12,7 +12,7 @@ regex
  2 files changed, 22 insertions(+), 1 deletion(-)
 
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index b9f0631f4..1525283d7 100644
+index 8246a0a14..dfba7778b 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -299,7 +299,7 @@ struct llm_tokenizer_bpe : llm_tokenizer {
@@ -25,7 +25,7 @@ index b9f0631f4..1525283d7 100644
                      "\\s+$",
                      "[一-龥ࠀ-一가-퟿]+",
 diff --git a/src/unicode.cpp b/src/unicode.cpp
-index 77ba4fc46..040518e1e 100644
+index bb44edfad..13ced055f 100644
 --- a/src/unicode.cpp
 +++ b/src/unicode.cpp
 @@ -2,6 +2,11 @@

llama/patches/0006-maintain-ordering-for-rules-for-grammar.patch

@@ -8,7 +8,7 @@ Subject: [PATCH] maintain ordering for rules for grammar
  1 file changed, 1 insertion(+), 1 deletion(-)
 
 diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp
-index c8421e1e8..cb659915d 100644
+index c3b4e5d9d..6be552826 100644
 --- a/common/json-schema-to-grammar.cpp
 +++ b/common/json-schema-to-grammar.cpp
 @@ -310,7 +310,7 @@ private:

llama/patches/0007-sort-devices-by-score.patch

@@ -11,10 +11,10 @@ with the fastest acceleration is loaded
  1 file changed, 13 insertions(+), 8 deletions(-)
 
 diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
-index e96b5c403..a55d9b280 100644
+index 4181a714a..079dba211 100644
 --- a/ggml/src/ggml-backend-reg.cpp
 +++ b/ggml/src/ggml-backend-reg.cpp
-@@ -179,7 +179,7 @@ struct ggml_backend_reg_entry {
+@@ -183,7 +183,7 @@ struct ggml_backend_reg_entry {
  
  struct ggml_backend_registry {
      std::vector<ggml_backend_reg_entry> backends;
@@ -23,7 +23,7 @@ index e96b5c403..a55d9b280 100644
  
      ggml_backend_registry() {
  #ifdef GGML_USE_CUDA
-@@ -230,7 +230,7 @@ struct ggml_backend_registry {
+@@ -237,7 +237,7 @@ struct ggml_backend_registry {
          }
      }
  
@@ -32,7 +32,7 @@ index e96b5c403..a55d9b280 100644
          if (!reg) {
              return;
          }
-@@ -241,15 +241,20 @@ struct ggml_backend_registry {
+@@ -248,15 +248,20 @@ struct ggml_backend_registry {
  #endif
          backends.push_back({ reg, std::move(handle) });
          for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
@@ -56,7 +56,7 @@ index e96b5c403..a55d9b280 100644
      }
  
      ggml_backend_reg_t load_backend(const fs::path & path, bool silent) {
-@@ -293,7 +298,7 @@ struct ggml_backend_registry {
+@@ -300,7 +305,7 @@ struct ggml_backend_registry {
  
          GGML_LOG_INFO("%s: loaded %s backend from %s\n", __func__, ggml_backend_reg_name(reg), path_str(path).c_str());
  
@@ -65,7 +65,7 @@ index e96b5c403..a55d9b280 100644
  
          return reg;
      }
-@@ -316,7 +321,7 @@ struct ggml_backend_registry {
+@@ -323,7 +328,7 @@ struct ggml_backend_registry {
          // remove devices
          devices.erase(
              std::remove_if(devices.begin(), devices.end(),
@@ -74,7 +74,7 @@ index e96b5c403..a55d9b280 100644
              devices.end());
  
          // remove backend
-@@ -374,7 +379,7 @@ size_t ggml_backend_dev_count() {
+@@ -381,7 +386,7 @@ size_t ggml_backend_dev_count() {
  
  ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
      GGML_ASSERT(index < ggml_backend_dev_count());

llama/patches/0008-add-phony-target-ggml-cpu-for-all-cpu-variants.patch

@@ -8,10 +8,10 @@ Subject: [PATCH] add phony target ggml-cpu for all cpu variants
  1 file changed, 2 insertions(+)
 
 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index d93664b8b..800f98b65 100644
+index 4c04c3300..f4747f262 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
-@@ -349,6 +349,7 @@ function(ggml_add_cpu_backend_variant tag_name)
+@@ -345,6 +345,7 @@ function(ggml_add_cpu_backend_variant tag_name)
      endif()
  
      ggml_add_cpu_backend_variant_impl(${tag_name})
@@ -19,7 +19,7 @@ index d93664b8b..800f98b65 100644
  endfunction()
  
  ggml_add_backend(CPU)
-@@ -359,6 +360,7 @@ if (GGML_CPU_ALL_VARIANTS)
+@@ -355,6 +356,7 @@ if (GGML_CPU_ALL_VARIANTS)
      elseif (GGML_CPU_ARM_ARCH)
          message(FATAL_ERROR "Cannot use both GGML_CPU_ARM_ARCH and GGML_CPU_ALL_VARIANTS")
      endif()

llama/patches/0009-remove-amx.patch

@@ -9,10 +9,10 @@ disable amx as it reduces performance on some systems
  1 file changed, 4 deletions(-)
 
 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index 800f98b65..6d493a4ff 100644
+index f4747f262..d55aed348 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
-@@ -369,10 +369,6 @@ if (GGML_CPU_ALL_VARIANTS)
+@@ -365,10 +365,6 @@ if (GGML_CPU_ALL_VARIANTS)
          ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
          ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
          ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)

llama/patches/0010-fix-string-arr-kv-loading.patch

@@ -25,7 +25,7 @@ index 79ee20206..3efb22f01 100644
      // get ith C string from array with given key_id
      GGML_API const char * gguf_get_arr_str (const struct gguf_context * ctx, int64_t key_id, size_t i);
 diff --git a/ggml/src/gguf.cpp b/ggml/src/gguf.cpp
-index 8cc4ef1cf..d950dbdf5 100644
+index b165d8bdc..f91d4faba 100644
 --- a/ggml/src/gguf.cpp
 +++ b/ggml/src/gguf.cpp
 @@ -805,10 +805,14 @@ enum gguf_type gguf_get_arr_type(const struct gguf_context * ctx, int64_t key_id
@@ -53,7 +53,7 @@ index 8cc4ef1cf..d950dbdf5 100644
  }
  
 diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
-index 1525283d7..ea450c361 100644
+index dfba7778b..f72f321b9 100644
 --- a/src/llama-vocab.cpp
 +++ b/src/llama-vocab.cpp
 @@ -1781,9 +1781,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {

llama/patches/0011-ollama-debug-tensor.patch

@@ -8,7 +8,7 @@ Subject: [PATCH] ollama debug tensor
  1 file changed, 6 insertions(+)
 
 diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
-index 3247af8bb..5be08d6f4 100644
+index b468b115a..bb65985b4 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.c
 +++ b/ggml/src/ggml-cpu/ggml-cpu.c
 @@ -15,6 +15,8 @@
@@ -20,7 +20,7 @@ index 3247af8bb..5be08d6f4 100644
  #if defined(_MSC_VER) || defined(__MINGW32__)
  #include <malloc.h> // using malloc.h with MSC/MINGW
  #elif !defined(__FreeBSD__) && !defined(__NetBSD__) && !defined(__OpenBSD__)
-@@ -2922,6 +2924,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
+@@ -2928,6 +2930,10 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
  
          ggml_compute_forward(&params, node);
  

llama/patches/0012-add-ollama-vocab-for-grammar-support.patch

@@ -4,16 +4,16 @@ Date: Mon, 21 Apr 2025 13:30:31 -0700
 Subject: [PATCH] add ollama vocab for grammar support
 
 ---
- src/llama-grammar.cpp  | 49 ++++++++++++++++++++++++++++++++++++------
+ src/llama-grammar.cpp  | 48 ++++++++++++++++++++++++++++++++++++------
  src/llama-grammar.h    | 14 ++++++++++++
  src/llama-sampling.cpp |  6 +++---
- 3 files changed, 59 insertions(+), 10 deletions(-)
+ 3 files changed, 58 insertions(+), 10 deletions(-)
 
 diff --git a/src/llama-grammar.cpp b/src/llama-grammar.cpp
-index b3c5eb571..a7307c47f 100644
+index 75d5d750c..a0299d181 100644
 --- a/src/llama-grammar.cpp
 +++ b/src/llama-grammar.cpp
-@@ -915,6 +915,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
+@@ -1041,6 +1041,7 @@ llama_grammar_candidates llama_grammar_reject_candidates_for_stack(
  
  struct llama_grammar * llama_grammar_init_impl(
          const struct llama_vocab * vocab,
@@ -21,7 +21,7 @@ index b3c5eb571..a7307c47f 100644
          const llama_grammar_element ** rules,
          size_t n_rules,
          size_t start_rule_index) {
-@@ -970,6 +971,7 @@ struct llama_grammar * llama_grammar_init_impl(
+@@ -1096,6 +1097,7 @@ struct llama_grammar * llama_grammar_init_impl(
      // then the pointers would be invalidated when the local vec_rules goes out of scope.
      return new llama_grammar {
          vocab,
@@ -29,7 +29,7 @@ index b3c5eb571..a7307c47f 100644
          std::move(vec_rules),
          std::move(stacks),
          /* .partial_utf8 = */             {},
-@@ -983,6 +985,7 @@ struct llama_grammar * llama_grammar_init_impl(
+@@ -1110,6 +1112,7 @@ struct llama_grammar * llama_grammar_init_impl(
  
  struct llama_grammar * llama_grammar_init_impl(
          const struct llama_vocab * vocab,
@@ -37,7 +37,7 @@ index b3c5eb571..a7307c47f 100644
                        const char * grammar_str,
                        const char * grammar_root,
                                bool lazy,
-@@ -1075,6 +1078,7 @@ struct llama_grammar * llama_grammar_init_impl(
+@@ -1202,6 +1205,7 @@ struct llama_grammar * llama_grammar_init_impl(
      // then the pointers would be invalidated when the local vec_rules goes out of scope.
      return new llama_grammar {
          vocab,
@@ -45,7 +45,7 @@ index b3c5eb571..a7307c47f 100644
          std::move(vec_rules),
          std::move(stacks),
          /* .partial_utf8 = */             {},
-@@ -1097,6 +1101,7 @@ void llama_grammar_free_impl(struct llama_grammar * grammar) {
+@@ -1225,6 +1229,7 @@ void llama_grammar_free_impl(struct llama_grammar * grammar) {
  struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & grammar) {
      auto * result = new llama_grammar {
          grammar.vocab,
@@ -53,7 +53,7 @@ index b3c5eb571..a7307c47f 100644
          grammar.rules,
          grammar.stacks,
          grammar.partial_utf8,
-@@ -1124,7 +1129,6 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
+@@ -1253,7 +1258,6 @@ struct llama_grammar * llama_grammar_clone_impl(const struct llama_grammar & gra
  }
  
  void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_data_array * cur_p) {
@@ -61,7 +61,7 @@ index b3c5eb571..a7307c47f 100644
  
      if (grammar.awaiting_trigger) {
          return;
-@@ -1146,9 +1150,13 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
+@@ -1275,9 +1279,13 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
  
      for (size_t i = 0; i < cur_p->size; ++i) {
          const llama_token id      = cur_p->data[i].id;
@@ -77,7 +77,7 @@ index b3c5eb571..a7307c47f 100644
              if (!allow_eog) {
                  cur_p->data[i].logit = -INFINITY;
              }
-@@ -1167,9 +1175,10 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
+@@ -1296,9 +1304,10 @@ void llama_grammar_apply_impl(const struct llama_grammar & grammar, llama_token_
  }
  
  void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token) {
@@ -90,7 +90,7 @@ index b3c5eb571..a7307c47f 100644
  
      if (grammar.awaiting_trigger) {
          if (std::find(grammar.trigger_tokens.begin(), grammar.trigger_tokens.end(), token) != grammar.trigger_tokens.end()) {
-@@ -1209,13 +1218,14 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
+@@ -1353,13 +1362,14 @@ void llama_grammar_accept_impl(struct llama_grammar & grammar, llama_token token
          }
      }
  
@@ -106,12 +106,11 @@ index b3c5eb571..a7307c47f 100644
 +        GGML_ABORT("grammar error: end of grammar token received but grammar stack is not empty");
      }
  
-     llama_grammar_accept_str(grammar, piece);
-@@ -1235,3 +1245,28 @@ void llama_grammar_accept_str(struct llama_grammar & grammar, const std::string
-         throw std::runtime_error("Unexpected empty grammar stack after accepting piece: " + piece);
+     llama_grammar_accept_token(grammar, token, piece);
+@@ -1435,3 +1445,27 @@ void llama_grammar_accept_token(struct llama_grammar & grammar, llama_token toke
      }
  }
-+
+ 
 +
 +const std::string & ollama_vocab::token_to_piece(const uint32_t token) const {
 +    try {
@@ -137,7 +136,7 @@ index b3c5eb571..a7307c47f 100644
 +    }
 +}
 diff --git a/src/llama-grammar.h b/src/llama-grammar.h
-index f8c291de9..2a3a62db3 100644
+index a4c978ac1..5c0da4049 100644
 --- a/src/llama-grammar.h
 +++ b/src/llama-grammar.h
 @@ -6,8 +6,19 @@
@@ -160,15 +159,15 @@ index f8c291de9..2a3a62db3 100644
  
  // grammar element type
  enum llama_gretype {
-@@ -114,6 +125,7 @@ struct llama_grammar_trigger_pattern {
- struct llama_grammar {
+@@ -127,6 +138,7 @@ struct llama_grammar {
+ 
      // note: allow null vocab for testing (not great)
      const llama_vocab * vocab;
 +    const ollama_vocab * o_vocab;
  
      const llama_grammar_rules  rules;  // TODO: shared ptr
            llama_grammar_stacks stacks;
-@@ -141,12 +153,14 @@ struct llama_grammar {
+@@ -155,12 +167,14 @@ struct llama_grammar {
  // note: needed for tests (not great)
  struct llama_grammar * llama_grammar_init_impl(
          const struct llama_vocab * vocab,

llama/patches/0013-add-argsort-and-cuda-copy-for-i32.patch

@@ -12,10 +12,10 @@ Subject: [PATCH] add argsort and cuda copy for i32
  5 files changed, 414 insertions(+), 12 deletions(-)
 
 diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
-index 2745fc54e..40666bab6 100644
+index 303278397..7d1733adb 100644
 --- a/ggml/src/ggml-cpu/ops.cpp
 +++ b/ggml/src/ggml-cpu/ops.cpp
-@@ -7846,6 +7846,45 @@ static void ggml_compute_forward_argsort_f32(
+@@ -7932,6 +7932,45 @@ static void ggml_compute_forward_argsort_f32(
      }
  }
  
@@ -61,7 +61,7 @@ index 2745fc54e..40666bab6 100644
  void ggml_compute_forward_argsort(
      const ggml_compute_params * params,
      ggml_tensor * dst) {
-@@ -7857,6 +7896,10 @@ void ggml_compute_forward_argsort(
+@@ -7943,6 +7982,10 @@ void ggml_compute_forward_argsort(
              {
                  ggml_compute_forward_argsort_f32(params, dst);
              } break;
@@ -292,10 +292,10 @@ index c4ceb4fc5..0e53ecc39 100644
          if (can_be_transposed) {
              ggml_cpy_scalar_cuda<nv_bfloat16, nv_bfloat16, true>
 diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
-index 73b45c762..8a6c834d1 100644
+index 51bcbae30..236838e9e 100644
 --- a/ggml/src/ggml-metal/ggml-metal.metal
 +++ b/ggml/src/ggml-metal/ggml-metal.metal
-@@ -4721,8 +4721,77 @@ kernel void kernel_argsort_f32_i32(
+@@ -4954,8 +4954,77 @@ kernel void kernel_argsort_f32_i32(
      }
  }
  
@@ -373,7 +373,7 @@ index 73b45c762..8a6c834d1 100644
  
  typedef void (argsort_merge_t)(
          constant   ggml_metal_kargs_argsort_merge & args,
-@@ -4877,8 +4946,154 @@ kernel void kernel_argsort_merge_f32_i32(
+@@ -5110,8 +5179,154 @@ kernel void kernel_argsort_merge_f32_i32(
      }
  }
  

llama/patches/0014-graph-memory-reporting-on-failure.patch

@@ -35,10 +35,10 @@ index f1b740785..c54ff98bf 100644
      GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
      GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 diff --git a/ggml/src/ggml-alloc.c b/ggml/src/ggml-alloc.c
-index 218222ece..06ee502ab 100644
+index a5995fdc2..dbfd8b5b2 100644
 --- a/ggml/src/ggml-alloc.c
 +++ b/ggml/src/ggml-alloc.c
-@@ -493,6 +493,7 @@ struct node_alloc {
+@@ -494,6 +494,7 @@ struct node_alloc {
  struct ggml_gallocr {
      ggml_backend_buffer_type_t * bufts; // [n_buffers]
      struct vbuffer ** buffers; // [n_buffers]
@@ -46,7 +46,7 @@ index 218222ece..06ee502ab 100644
      struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
      int n_buffers;
  
-@@ -516,6 +517,9 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
+@@ -517,6 +518,9 @@ ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs
      galloc->buffers = calloc(n_bufs, sizeof(struct vbuffer *));
      GGML_ASSERT(galloc->buffers != NULL);
  
@@ -56,7 +56,7 @@ index 218222ece..06ee502ab 100644
      galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
      GGML_ASSERT(galloc->buf_tallocs != NULL);
  
-@@ -583,6 +587,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
+@@ -584,6 +588,7 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
      ggml_hash_set_free(&galloc->hash_set);
      free(galloc->hash_values);
      free(galloc->bufts);
@@ -64,7 +64,7 @@ index 218222ece..06ee502ab 100644
      free(galloc->buffers);
      free(galloc->buf_tallocs);
      free(galloc->node_allocs);
-@@ -898,6 +903,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
+@@ -899,6 +904,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
          }
      }
  
@@ -73,7 +73,7 @@ index 218222ece..06ee502ab 100644
      // reallocate buffers if needed
      for (int i = 0; i < galloc->n_buffers; i++) {
          // if the buffer type is used multiple times, we reuse the same buffer
-@@ -932,14 +939,19 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
+@@ -933,14 +940,19 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
  #endif
              ggml_vbuffer_free(galloc->buffers[i]);
              galloc->buffers[i] = ggml_vbuffer_alloc(galloc->bufts[i], galloc->buf_tallocs[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
@@ -96,7 +96,7 @@ index 218222ece..06ee502ab 100644
  }
  
  bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
-@@ -1094,6 +1106,22 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
+@@ -1095,6 +1107,22 @@ size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
      return ggml_vbuffer_size(galloc->buffers[buffer_id]);
  }
  
@@ -120,10 +120,10 @@ index 218222ece..06ee502ab 100644
  
  static void free_buffers(ggml_backend_buffer_t ** buffers, const size_t * n_buffers) {
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index 4882541c8..ff41c7712 100644
+index afde2f0b7..dbf8486a0 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
-@@ -1813,6 +1813,13 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
+@@ -1840,6 +1840,13 @@ size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backe
      return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
  }
  

llama/patches/0015-ggml-Export-GPU-UUIDs.patch

@@ -22,10 +22,10 @@ index c54ff98bf..229bf387b 100644
          size_t memory_total;
          // device type
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 8f3b1c173..e803f4af6 100644
+index 5145c1e88..f641c1016 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -185,6 +185,51 @@ static int ggml_cuda_parse_id(char devName[]) {
+@@ -189,6 +189,51 @@ static int ggml_cuda_parse_id(char devName[]) {
  }
  #endif // defined(GGML_USE_HIP)
  
@@ -77,7 +77,7 @@ index 8f3b1c173..e803f4af6 100644
  static ggml_cuda_device_info ggml_cuda_init() {
      ggml_cuda_device_info info = {};
  
-@@ -251,22 +296,24 @@ static ggml_cuda_device_info ggml_cuda_init() {
+@@ -255,22 +300,24 @@ static ggml_cuda_device_info ggml_cuda_init() {
                  info.devices[id].cc += prop.minor * 0x10;
              }
          }
@@ -108,7 +108,7 @@ index 8f3b1c173..e803f4af6 100644
          std::string device_name(prop.name);
          if (device_name == "NVIDIA GeForce MX450") {
              turing_devices_without_mma.push_back({ id, device_name });
-@@ -4048,6 +4095,7 @@ struct ggml_backend_cuda_device_context {
+@@ -4110,6 +4157,7 @@ struct ggml_backend_cuda_device_context {
      std::string name;
      std::string description;
      std::string pci_bus_id;
@@ -116,7 +116,7 @@ index 8f3b1c173..e803f4af6 100644
  };
  
  static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
-@@ -4136,6 +4184,11 @@ static bool ggml_backend_cuda_get_available_uma_memory(long * available_memory_k
+@@ -4198,6 +4246,11 @@ static bool ggml_backend_cuda_get_available_uma_memory(long * available_memory_k
  }
  #endif // defined(__linux__)
  
@@ -128,7 +128,7 @@ index 8f3b1c173..e803f4af6 100644
  static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
      ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
      ggml_cuda_set_device(ctx->device);
-@@ -4176,6 +4229,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
+@@ -4238,6 +4291,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
  
      props->name        = ggml_backend_cuda_device_get_name(dev);
      props->description = ggml_backend_cuda_device_get_description(dev);
@@ -136,7 +136,7 @@ index 8f3b1c173..e803f4af6 100644
      props->type        = ggml_backend_cuda_device_get_type(dev);
      props->device_id   = ctx->pci_bus_id.empty() ? nullptr : ctx->pci_bus_id.c_str();
      ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
-@@ -4767,6 +4821,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+@@ -4833,6 +4887,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                  cudaDeviceProp prop;
                  CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
                  dev_ctx->description = prop.name;

llama/patches/0016-add-C-API-for-mtmd_input_text.patch

@@ -10,7 +10,7 @@ Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
  2 files changed, 13 insertions(+)
 
 diff --git a/tools/mtmd/mtmd.cpp b/tools/mtmd/mtmd.cpp
-index dfad9cd79..9858de630 100644
+index d06fa42e6..0f5712e21 100644
 --- a/tools/mtmd/mtmd.cpp
 +++ b/tools/mtmd/mtmd.cpp
 @@ -87,6 +87,16 @@ enum mtmd_slice_tmpl {
@@ -31,7 +31,7 @@ index dfad9cd79..9858de630 100644
      return "<__media__>";
  }
 diff --git a/tools/mtmd/mtmd.h b/tools/mtmd/mtmd.h
-index 015119be8..8d3fa5d34 100644
+index b3df24c29..a6a1af3b8 100644
 --- a/tools/mtmd/mtmd.h
 +++ b/tools/mtmd/mtmd.h
 @@ -75,6 +75,9 @@ typedef struct mtmd_input_chunk  mtmd_input_chunk;

llama/patches/0017-no-power-throttling-win32-with-gnuc.patch

@@ -8,10 +8,10 @@ Subject: [PATCH] no power throttling win32 with gnuc
  1 file changed, 1 insertion(+), 1 deletion(-)
 
 diff --git a/ggml/src/ggml-cpu/ggml-cpu.c b/ggml/src/ggml-cpu/ggml-cpu.c
-index 5be08d6f4..7a0df30c3 100644
+index bb65985b4..47089a62e 100644
 --- a/ggml/src/ggml-cpu/ggml-cpu.c
 +++ b/ggml/src/ggml-cpu/ggml-cpu.c
-@@ -2463,7 +2463,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
+@@ -2464,7 +2464,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
          // Newer Windows 11 versions aggresively park (offline) CPU cores and often place
          // all our threads onto the first 4 cores which results in terrible performance with
          // n_threads > 4

llama/patches/0018-ggml-Add-batch-size-hint.patch

@@ -58,7 +58,7 @@ index 6792ba986..0f5b03cef 100644
          // (optional) event synchronization
          // record an event on this stream
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index ff41c7712..f511e8d76 100644
+index dbf8486a0..312ca873c 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
 @@ -348,14 +348,14 @@ enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_ba
@@ -86,9 +86,9 @@ index ff41c7712..f511e8d76 100644
 +    int batch_size; // a hint on the batch size to optimize processing, -1 to use heuristics
 +
      int debug;
- };
  
-@@ -814,7 +816,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
+     // used for debugging graph reallocations [GGML_SCHED_DEBUG_REALLOC]
+@@ -820,7 +822,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
          if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
              int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
              // check if a backend with higher prio wants to offload the op
@@ -97,7 +97,7 @@ index ff41c7712..f511e8d76 100644
                  for (int b = 0; b < src_backend_id; b++) {
                      if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
                          SET_CAUSE(tensor, "1.off");
-@@ -1556,7 +1558,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
+@@ -1572,7 +1574,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
          }
  
          if (!sched->callback_eval) {
@@ -106,7 +106,7 @@ index ff41c7712..f511e8d76 100644
              if (ec != GGML_STATUS_SUCCESS) {
                  return ec;
              }
-@@ -1578,7 +1580,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
+@@ -1594,7 +1596,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
  
                  struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);
  
@@ -115,7 +115,7 @@ index ff41c7712..f511e8d76 100644
                  if (ec != GGML_STATUS_SUCCESS) {
                      return ec;
                  }
-@@ -1657,6 +1659,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
+@@ -1684,6 +1686,7 @@ ggml_backend_sched_t ggml_backend_sched_new(
  
      sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
      sched->op_offload = op_offload;
@@ -123,7 +123,7 @@ index ff41c7712..f511e8d76 100644
  
      ggml_backend_sched_reset(sched);
  
-@@ -1688,6 +1691,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+@@ -1715,6 +1718,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
      free(sched);
  }
  
@@ -178,10 +178,10 @@ index 3191faaa4..32f14c811 100644
  
  static const struct ggml_backend_i ggml_backend_cpu_i = {
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index e803f4af6..78fb2d8b3 100644
+index f641c1016..17062697b 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -2885,7 +2885,7 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
+@@ -2901,7 +2901,7 @@ static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
  
  #ifdef USE_CUDA_GRAPH
  static bool check_node_graph_compatibility(ggml_cgraph * cgraph,
@@ -190,7 +190,7 @@ index e803f4af6..78fb2d8b3 100644
  
      // Loop over nodes in GGML graph to obtain info needed for CUDA graph
  
-@@ -2918,24 +2918,34 @@ static bool check_node_graph_compatibility(ggml_cgraph * cgraph,
+@@ -2934,24 +2934,34 @@ static bool check_node_graph_compatibility(ggml_cgraph * cgraph,
  #endif
          }
  
@@ -241,7 +241,7 @@ index e803f4af6..78fb2d8b3 100644
          }
  
          if (!use_cuda_graph) {
-@@ -3679,7 +3689,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
+@@ -3742,7 +3752,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
      }
  }
  
@@ -250,7 +250,7 @@ index e803f4af6..78fb2d8b3 100644
      ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
  
      ggml_cuda_set_device(cuda_ctx->device);
-@@ -3717,7 +3727,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
+@@ -3780,7 +3790,7 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
      if (use_cuda_graph) {
          cuda_graph_update_required = is_cuda_graph_update_required(cuda_ctx, cgraph);
  
@@ -278,10 +278,10 @@ index 8fc1c2fb5..ba95b4acc 100644
  
  static void ggml_backend_metal_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph) {
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index 83cdec29e..a36c6560c 100644
+index c801d2fd2..b2c0d0cee 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-@@ -13103,7 +13103,7 @@ static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const stru
+@@ -13006,7 +13006,7 @@ static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const stru
      return num_adds;
  }
  
@@ -290,7 +290,7 @@ index 83cdec29e..a36c6560c 100644
      VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
      ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
  
-@@ -13320,6 +13320,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
+@@ -13241,6 +13241,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
      return GGML_STATUS_SUCCESS;
  
      UNUSED(backend);

llama/patches/0020-ggml-No-alloc-mode.patch

@@ -75,7 +75,7 @@ index 0f5b03cef..7bdf9d81f 100644
  
      struct ggml_backend {
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index f511e8d76..74b7f070c 100644
+index 312ca873c..4092dfe8a 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
 @@ -41,6 +41,19 @@ ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t
@@ -121,10 +121,10 @@ index f511e8d76..74b7f070c 100644
      void * base = buffer->iface.get_base(buffer);
  
      GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
-@@ -725,6 +745,12 @@ struct ggml_backend_sched {
-     int batch_size; // a hint on the batch size to optimize processing, -1 to use heuristics
- 
-     int debug;
+@@ -731,6 +751,12 @@ struct ggml_backend_sched {
+     int debug_realloc;
+     int debug_graph_size;
+     int debug_prev_graph_size;
 +
 +    // allocate buffers on attached ggml_backend_buffer_type_t's and during reservation
 +    // if false, dummy buffers are used for faster memory sizing calculations
@@ -134,7 +134,7 @@ index f511e8d76..74b7f070c 100644
  };
  
  #define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
-@@ -1614,6 +1640,17 @@ ggml_backend_sched_t ggml_backend_sched_new(
+@@ -1630,6 +1656,17 @@ ggml_backend_sched_t ggml_backend_sched_new(
          size_t graph_size,
          bool parallel,
          bool op_offload) {
@@ -152,7 +152,7 @@ index f511e8d76..74b7f070c 100644
      GGML_ASSERT(n_backends > 0);
      GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
      GGML_ASSERT(ggml_backend_dev_type(ggml_backend_get_device(backends[n_backends - 1])) == GGML_BACKEND_DEVICE_TYPE_CPU);
-@@ -1655,11 +1692,14 @@ ggml_backend_sched_t ggml_backend_sched_new(
+@@ -1682,11 +1719,14 @@ ggml_backend_sched_t ggml_backend_sched_new(
                  sched->events[b][c] = ggml_backend_event_new(backends[b]->device);
              }
          }
@@ -167,7 +167,7 @@ index f511e8d76..74b7f070c 100644
  
      ggml_backend_sched_reset(sched);
  
-@@ -1674,6 +1714,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+@@ -1701,6 +1741,10 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
          for (int c = 0; c < sched->n_copies; c++) {
              ggml_backend_event_free(sched->events[b][c]);
          }
@@ -178,7 +178,7 @@ index f511e8d76..74b7f070c 100644
      }
      ggml_gallocr_free(sched->galloc);
      ggml_free(sched->ctx);
-@@ -1719,6 +1763,24 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
+@@ -1746,6 +1790,24 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
          return false;
      }
  
@@ -203,7 +203,7 @@ index f511e8d76..74b7f070c 100644
      ggml_backend_sched_reset(sched);
  
      return true;
-@@ -1824,7 +1886,13 @@ size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched,
+@@ -1851,7 +1913,13 @@ size_t ggml_backend_sched_get_attempted_buffer_size(ggml_backend_sched_t sched,
      int backend_index = ggml_backend_sched_backend_id(sched, backend);
      GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
  
@@ -219,7 +219,7 @@ index f511e8d76..74b7f070c 100644
  
  void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
 diff --git a/ggml/src/ggml-cuda/common.cuh b/ggml/src/ggml-cuda/common.cuh
-index 611341deb..ee463af9c 100644
+index c4529f5d9..8b0fb5d42 100644
 --- a/ggml/src/ggml-cuda/common.cuh
 +++ b/ggml/src/ggml-cuda/common.cuh
 @@ -37,6 +37,41 @@
@@ -264,7 +264,7 @@ index 611341deb..ee463af9c 100644
  #define STRINGIZE_IMPL(...) #__VA_ARGS__
  #define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
  
-@@ -891,6 +926,9 @@ struct ggml_cuda_pool {
+@@ -938,6 +973,9 @@ struct ggml_cuda_pool {
  
      virtual void * alloc(size_t size, size_t * actual_size) = 0;
      virtual void free(void * ptr, size_t size) = 0;
@@ -274,7 +274,7 @@ index 611341deb..ee463af9c 100644
  };
  
  template<typename T>
-@@ -1179,11 +1217,15 @@ struct ggml_backend_cuda_context {
+@@ -1229,11 +1267,15 @@ struct ggml_backend_cuda_context {
      // pool
      std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS];
  
@@ -292,7 +292,7 @@ index 611341deb..ee463af9c 100644
          }
          return *pools[device][curr_stream_no];
      }
-@@ -1191,6 +1233,22 @@ struct ggml_backend_cuda_context {
+@@ -1241,6 +1283,22 @@ struct ggml_backend_cuda_context {
      ggml_cuda_pool & pool() {
          return pool(device);
      }
@@ -316,10 +316,10 @@ index 611341deb..ee463af9c 100644
  
  struct ggml_cuda_mm_fusion_args_host {
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 78fb2d8b3..f1c178f31 100644
+index 17062697b..ede1d089a 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -361,6 +361,8 @@ const ggml_cuda_device_info & ggml_cuda_info() {
+@@ -365,6 +365,8 @@ const ggml_cuda_device_info & ggml_cuda_info() {
  
  // #define DEBUG_CUDA_MALLOC
  
@@ -328,7 +328,7 @@ index 78fb2d8b3..f1c178f31 100644
  // buffer pool for cuda (legacy)
  struct ggml_cuda_pool_leg : public ggml_cuda_pool {
      static const int MAX_BUFFERS = 256;
-@@ -373,9 +375,12 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
+@@ -377,9 +379,12 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
  
      ggml_cuda_buffer buffer_pool[MAX_BUFFERS] = {};
      size_t pool_size = 0;
@@ -343,7 +343,7 @@ index 78fb2d8b3..f1c178f31 100644
      }
  
      ~ggml_cuda_pool_leg() {
-@@ -383,7 +388,9 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
+@@ -387,7 +392,9 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
          for (int i = 0; i < MAX_BUFFERS; ++i) {
              ggml_cuda_buffer & b = buffer_pool[i];
              if (b.ptr != nullptr) {
@@ -354,7 +354,7 @@ index 78fb2d8b3..f1c178f31 100644
                  pool_size -= b.size;
              }
          }
-@@ -431,8 +438,15 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
+@@ -435,8 +442,15 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
          void * ptr;
          size_t look_ahead_size = (size_t) (1.05 * size);
          look_ahead_size = 256 * ((look_ahead_size + 255)/256);
@@ -372,7 +372,7 @@ index 78fb2d8b3..f1c178f31 100644
          *actual_size = look_ahead_size;
          pool_size += look_ahead_size;
  #ifdef DEBUG_CUDA_MALLOC
-@@ -452,10 +466,20 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
+@@ -456,10 +470,20 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
              }
          }
          GGML_LOG_DEBUG(GGML_CUDA_NAME " buffer pool full, increase MAX_CUDA_BUFFERS\n");
@@ -395,7 +395,7 @@ index 78fb2d8b3..f1c178f31 100644
  };
  
  // pool with virtual memory
-@@ -467,18 +491,24 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
+@@ -471,18 +495,24 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
      CUdeviceptr pool_addr = 0;
      size_t pool_used = 0;
      size_t pool_size = 0;
@@ -423,7 +423,7 @@ index 78fb2d8b3..f1c178f31 100644
  #if defined(GGML_USE_HIP)
              // Workaround for https://github.com/ROCm/ROCR-Runtime/issues/285
              for (std::pair<CUdeviceptr, size_t> & mapping : mappings) {
-@@ -505,35 +535,49 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
+@@ -509,35 +539,49 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
  
              GGML_ASSERT(pool_size + reserve_size <= CUDA_POOL_VMM_MAX_SIZE);
  
@@ -499,7 +499,7 @@ index 78fb2d8b3..f1c178f31 100644
  
              // add to the pool
              pool_size += reserve_size;
-@@ -566,17 +610,27 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
+@@ -570,17 +614,27 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
          // all deallocations must be in reverse order of the allocations
          GGML_ASSERT(ptr == (void *) ((char *)(pool_addr) + pool_used));
      }
@@ -530,7 +530,7 @@ index 78fb2d8b3..f1c178f31 100644
  }
  
  // destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error
-@@ -760,11 +814,20 @@ static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_bac
+@@ -764,11 +818,20 @@ static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_bac
  }
  
  static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
@@ -552,7 +552,7 @@ index 78fb2d8b3..f1c178f31 100644
  static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
      size_t size = ggml_nbytes(tensor);
      int64_t ne0 = tensor->ne[0];
-@@ -788,6 +851,7 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface
+@@ -792,6 +855,7 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface
      /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
      /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
      /* .is_host          = */ NULL,
@@ -560,7 +560,7 @@ index 78fb2d8b3..f1c178f31 100644
  };
  
  ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
-@@ -3258,6 +3322,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
+@@ -3274,6 +3338,7 @@ static bool ggml_cuda_can_fuse(const struct ggml_cgraph * cgraph, int node_idx,
  
  static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
      bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
@@ -568,7 +568,7 @@ index 78fb2d8b3..f1c178f31 100644
      // flag used to determine whether it is an integrated_gpu
      const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
  
-@@ -3347,6 +3412,10 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
+@@ -3410,6 +3475,10 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                      continue;
                  }
  
@@ -579,7 +579,7 @@ index 78fb2d8b3..f1c178f31 100644
  
                  // start of fusion operations
                  static bool disable_fusion = (getenv("GGML_CUDA_DISABLE_FUSION") != nullptr);
-@@ -3691,6 +3760,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
+@@ -3754,6 +3823,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
  
  static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph, int batch_size) {
      ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
@@ -587,7 +587,7 @@ index 78fb2d8b3..f1c178f31 100644
  
      ggml_cuda_set_device(cuda_ctx->device);
  
-@@ -3766,6 +3836,77 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
+@@ -3829,6 +3899,77 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
      return GGML_STATUS_SUCCESS;
  }
  
@@ -665,7 +665,7 @@ index 78fb2d8b3..f1c178f31 100644
  static void ggml_backend_cuda_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
      ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
  
-@@ -4035,6 +4176,9 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
+@@ -4097,6 +4238,9 @@ static const ggml_backend_i ggml_backend_cuda_interface = {
      /* .event_record            = */ ggml_backend_cuda_event_record,
      /* .event_wait              = */ ggml_backend_cuda_event_wait,
      /* .graph_optimize          = */ ggml_backend_cuda_graph_optimize,

llama/patches/0021-decode-disable-output_all.patch

@@ -8,7 +8,7 @@ Subject: [PATCH] decode: disable output_all
  1 file changed, 1 insertion(+), 2 deletions(-)
 
 diff --git a/src/llama-context.cpp b/src/llama-context.cpp
-index e04f0fc4f..1359c614b 100644
+index 417140071..87f407f99 100644
 --- a/src/llama-context.cpp
 +++ b/src/llama-context.cpp
 @@ -999,8 +999,7 @@ int llama_context::decode(const llama_batch & batch_inp) {

llama/patches/0022-ggml-Enable-resetting-backend-devices.patch

@@ -43,7 +43,7 @@ index 7bdf9d81f..21b35ac5c 100644
  
      struct ggml_backend_device {
 diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
-index 74b7f070c..8d2cc167f 100644
+index 4092dfe8a..a1a19fe51 100644
 --- a/ggml/src/ggml-backend.cpp
 +++ b/ggml/src/ggml-backend.cpp
 @@ -526,6 +526,14 @@ ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * par
@@ -62,10 +62,10 @@ index 74b7f070c..8d2cc167f 100644
      GGML_ASSERT(device);
      return device->iface.get_buffer_type(device);
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index f1c178f31..1110ca372 100644
+index ede1d089a..ec63cadab 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -109,6 +109,11 @@ int ggml_cuda_get_device() {
+@@ -113,6 +113,11 @@ int ggml_cuda_get_device() {
      return id;
  }
  
@@ -77,7 +77,7 @@ index f1c178f31..1110ca372 100644
  static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
      ggml_cuda_set_device(device);
      cudaError_t err;
-@@ -4386,7 +4391,10 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
+@@ -4448,7 +4453,10 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
      props->id          = ggml_backend_cuda_device_get_id(dev);
      props->type        = ggml_backend_cuda_device_get_type(dev);
      props->device_id   = ctx->pci_bus_id.empty() ? nullptr : ctx->pci_bus_id.c_str();
@@ -89,7 +89,7 @@ index f1c178f31..1110ca372 100644
  
      bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
  #ifdef GGML_CUDA_NO_PEER_COPY
-@@ -4841,6 +4849,11 @@ static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, g
+@@ -4907,6 +4915,11 @@ static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, g
      CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
  }
  
@@ -101,7 +101,7 @@ index f1c178f31..1110ca372 100644
  static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
      /* .get_name                = */ ggml_backend_cuda_device_get_name,
      /* .get_description         = */ ggml_backend_cuda_device_get_description,
-@@ -4857,6 +4870,7 @@ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
+@@ -4923,6 +4936,7 @@ static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
      /* .event_new               = */ ggml_backend_cuda_device_event_new,
      /* .event_free              = */ ggml_backend_cuda_device_event_free,
      /* .event_synchronize       = */ ggml_backend_cuda_device_event_synchronize,

llama/patches/0024-GPU-discovery-enhancements.patch

@@ -45,10 +45,10 @@ index 69223c488..6510e0cba 100644
  
      GGML_API const char *                  ggml_backend_dev_name(ggml_backend_dev_t device);
 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index 6d493a4ff..ac8f38464 100644
+index d55aed348..99ae293cc 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
-@@ -209,6 +209,8 @@ add_library(ggml-base
+@@ -205,6 +205,8 @@ add_library(ggml-base
              ggml-threading.h
              ggml-quants.c
              ggml-quants.h
@@ -58,10 +58,10 @@ index 6d493a4ff..ac8f38464 100644
  
  set_target_properties(ggml-base PROPERTIES
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index 1110ca372..c1bfadb3e 100644
+index ec63cadab..cd71902df 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -263,6 +263,16 @@ static ggml_cuda_device_info ggml_cuda_init() {
+@@ -267,6 +267,16 @@ static ggml_cuda_device_info ggml_cuda_init() {
      for (int id = 0; id < info.device_count; ++id) {
          int device_vmm = 0;
  
@@ -78,7 +78,7 @@ index 1110ca372..c1bfadb3e 100644
  #if defined(GGML_USE_VMM)
          CUdevice device;
          CU_CHECK(cuDeviceGet(&device, id));
-@@ -316,6 +326,11 @@ static ggml_cuda_device_info ggml_cuda_init() {
+@@ -320,6 +330,11 @@ static ggml_cuda_device_info ggml_cuda_init() {
  #else
          info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
          info.devices[id].cc = 100*prop.major + 10*prop.minor;
@@ -90,7 +90,7 @@ index 1110ca372..c1bfadb3e 100644
          GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s, ID: %s\n",
                          id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no",
                          ggml_cuda_parse_uuid(prop, id).c_str());
-@@ -4255,6 +4270,11 @@ struct ggml_backend_cuda_device_context {
+@@ -4317,6 +4332,11 @@ struct ggml_backend_cuda_device_context {
      std::string description;
      std::string pci_bus_id;
      std::string id;
@@ -102,7 +102,7 @@ index 1110ca372..c1bfadb3e 100644
  };
  
  static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
-@@ -4351,6 +4371,28 @@ static const char * ggml_backend_cuda_device_get_id(ggml_backend_dev_t dev) {
+@@ -4413,6 +4433,28 @@ static const char * ggml_backend_cuda_device_get_id(ggml_backend_dev_t dev) {
  static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
      ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
      ggml_cuda_set_device(ctx->device);
@@ -131,7 +131,7 @@ index 1110ca372..c1bfadb3e 100644
      CUDA_CHECK(cudaMemGetInfo(free, total));
  
  // ref: https://github.com/ggml-org/llama.cpp/pull/17368
-@@ -4383,6 +4425,7 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
+@@ -4445,6 +4487,7 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
      return GGML_BACKEND_DEVICE_TYPE_GPU;
  }
  
@@ -139,7 +139,7 @@ index 1110ca372..c1bfadb3e 100644
  static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
      ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
  
-@@ -4396,6 +4439,19 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
+@@ -4458,6 +4501,19 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
      // If you need the memory data, call ggml_backend_dev_memory() explicitly.
      props->memory_total = props->memory_free = 0;
  
@@ -159,7 +159,7 @@ index 1110ca372..c1bfadb3e 100644
      bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
  #ifdef GGML_CUDA_NO_PEER_COPY
      bool events = false;
-@@ -4980,6 +5036,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+@@ -5046,6 +5102,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
          std::lock_guard<std::mutex> lock(mutex);
          if (!initialized) {
              ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
@@ -167,7 +167,7 @@ index 1110ca372..c1bfadb3e 100644
  
              for (int i = 0; i < ggml_cuda_info().device_count; i++) {
                  ggml_backend_cuda_device_context * dev_ctx = new ggml_backend_cuda_device_context;
-@@ -4995,6 +5052,14 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
+@@ -5061,6 +5118,14 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                  snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.0", prop.pciDomainID, prop.pciBusID, prop.pciDeviceID);
                  dev_ctx->pci_bus_id = pci_bus_id;
  
@@ -243,7 +243,7 @@ index ba95b4acc..f6f8f7a10 100644
          /* .async                 = */ true,
          /* .host_buffer           = */ false,
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index a36c6560c..a234eda2e 100644
+index b2c0d0cee..d9f4d34f5 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 @@ -236,6 +236,7 @@ class vk_memory_logger;
@@ -254,7 +254,7 @@ index a36c6560c..a234eda2e 100644
  
  static constexpr uint32_t mul_mat_vec_max_cols = 8;
  static constexpr uint32_t p021_max_gqa_ratio = 8;
-@@ -12353,6 +12354,29 @@ static void ggml_vk_get_device_description(int device, char * description, size_
+@@ -12256,6 +12257,29 @@ static void ggml_vk_get_device_description(int device, char * description, size_
      snprintf(description, description_size, "%s", props.deviceName.data());
  }
  
@@ -284,7 +284,7 @@ index a36c6560c..a234eda2e 100644
  // backend interface
  
  #define UNUSED GGML_UNUSED
-@@ -13614,15 +13638,72 @@ void ggml_backend_vk_get_device_description(int device, char * description, size
+@@ -13535,15 +13559,72 @@ void ggml_backend_vk_get_device_description(int device, char * description, size
      ggml_vk_get_device_description(dev_idx, description, description_size);
  }
  
@@ -361,7 +361,7 @@ index a36c6560c..a234eda2e 100644
  
      if (membudget_supported) {
          memprops.pNext = &budgetprops;
-@@ -13674,8 +13755,13 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
+@@ -13595,8 +13676,13 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
          }
      }
  
@@ -376,7 +376,7 @@ index a36c6560c..a234eda2e 100644
      }
  
      vk::PhysicalDeviceProperties2 props = {};
-@@ -13692,19 +13778,24 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
+@@ -13613,19 +13699,24 @@ static std::string ggml_backend_vk_get_device_pci_id(int device_idx) {
  
      char pci_bus_id[16] = {};
      snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.%x", pci_domain, pci_bus, pci_device, pci_function);
@@ -410,7 +410,7 @@ index a36c6560c..a234eda2e 100644
  
  static const char * ggml_backend_vk_device_get_name(ggml_backend_dev_t dev) {
      ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
-@@ -13716,9 +13807,14 @@ static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t de
+@@ -13637,9 +13728,14 @@ static const char * ggml_backend_vk_device_get_description(ggml_backend_dev_t de
      return ctx->description.c_str();
  }
  
@@ -426,7 +426,7 @@ index a36c6560c..a234eda2e 100644
  }
  
  static ggml_backend_buffer_type_t ggml_backend_vk_device_get_buffer_type(ggml_backend_dev_t dev) {
-@@ -13742,8 +13838,9 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
+@@ -13663,8 +13759,9 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
  
      props->name        = ggml_backend_vk_device_get_name(dev);
      props->description = ggml_backend_vk_device_get_description(dev);
@@ -437,7 +437,7 @@ index a36c6560c..a234eda2e 100644
      ggml_backend_vk_device_get_memory(dev, &props->memory_free, &props->memory_total);
      props->caps = {
          /* .async                 = */ false,
-@@ -13751,6 +13848,13 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
+@@ -13672,6 +13769,13 @@ static void ggml_backend_vk_device_get_props(ggml_backend_dev_t dev, struct ggml
          /* .buffer_from_host_ptr  = */ false,
          /* .events                = */ false,
      };
@@ -451,7 +451,7 @@ index a36c6560c..a234eda2e 100644
  }
  
  static ggml_backend_t ggml_backend_vk_device_init(ggml_backend_dev_t dev, const char * params) {
-@@ -14319,6 +14423,8 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+@@ -14236,6 +14340,8 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
          static std::mutex mutex;
          std::lock_guard<std::mutex> lock(mutex);
          if (!initialized) {
@@ -460,7 +460,7 @@ index a36c6560c..a234eda2e 100644
              for (int i = 0; i < ggml_backend_vk_get_device_count(); i++) {
                  ggml_backend_vk_device_context * ctx = new ggml_backend_vk_device_context;
                  char desc[256];
-@@ -14327,12 +14433,41 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+@@ -14244,12 +14350,41 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
                  ctx->name = GGML_VK_NAME + std::to_string(i);
                  ctx->description = desc;
                  ctx->is_integrated_gpu = ggml_backend_vk_get_device_type(i) == vk::PhysicalDeviceType::eIntegratedGpu;

llama/patches/0026-report-LoadLibrary-failures.patch

@@ -8,10 +8,10 @@ Subject: [PATCH] report LoadLibrary failures
  1 file changed, 12 insertions(+)
 
 diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
-index a55d9b280..ec6f7f1e9 100644
+index 079dba211..2474e0ed6 100644
 --- a/ggml/src/ggml-backend-reg.cpp
 +++ b/ggml/src/ggml-backend-reg.cpp
-@@ -122,6 +122,18 @@ static dl_handle * dl_load_library(const fs::path & path) {
+@@ -126,6 +126,18 @@ static dl_handle * dl_load_library(const fs::path & path) {
      SetErrorMode(old_mode | SEM_FAILCRITICALERRORS);
  
      HMODULE handle = LoadLibraryW(path.wstring().c_str());

llama/patches/0027-interleave-multi-rope.patch

@@ -13,7 +13,7 @@ interleaved version used for qwen3vl
  4 files changed, 16 insertions(+), 16 deletions(-)
 
 diff --git a/ggml/src/ggml-cpu/ops.cpp b/ggml/src/ggml-cpu/ops.cpp
-index 40666bab6..3155cb4bb 100644
+index 7d1733adb..f4aae5332 100644
 --- a/ggml/src/ggml-cpu/ops.cpp
 +++ b/ggml/src/ggml-cpu/ops.cpp
 @@ -5599,14 +5599,14 @@ static void ggml_mrope_cache_init(
@@ -59,10 +59,10 @@ index 88ed79111..71ca60214 100644
      } else {
          if (sector < sections.v[0]) {
 diff --git a/ggml/src/ggml-metal/ggml-metal.metal b/ggml/src/ggml-metal/ggml-metal.metal
-index 8a6c834d1..761b57a26 100644
+index 236838e9e..c98d269d1 100644
 --- a/ggml/src/ggml-metal/ggml-metal.metal
 +++ b/ggml/src/ggml-metal/ggml-metal.metal
-@@ -4009,14 +4009,14 @@ kernel void kernel_rope_multi(
+@@ -4242,14 +4242,14 @@ kernel void kernel_rope_multi(
  
              float theta_base;
              if (FC_rope_is_imrope) {

llama/patches/0028-Add-memory-detection-using-DXGI-PDH.patch

@@ -12,10 +12,10 @@ Subject: [PATCH] Add memory detection using DXGI + PDH
  create mode 100644 ggml/src/mem_dxgi_pdh.cpp
 
 diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt
-index ac8f38464..faa1beed2 100644
+index 99ae293cc..9a134b7af 100644
 --- a/ggml/src/CMakeLists.txt
 +++ b/ggml/src/CMakeLists.txt
-@@ -211,6 +211,7 @@ add_library(ggml-base
+@@ -207,6 +207,7 @@ add_library(ggml-base
              ggml-quants.h
              mem_hip.cpp
              mem_nvml.cpp
@@ -38,7 +38,7 @@ index 1c07e767a..0da3e065b 100644
  #ifdef __cplusplus
  }
 diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
-index a234eda2e..c98f98c73 100644
+index d9f4d34f5..8a83427fb 100644
 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
 @@ -74,6 +74,7 @@ DispatchLoaderDynamic & ggml_vk_default_dispatcher();
@@ -49,7 +49,7 @@ index a234eda2e..c98f98c73 100644
  
  typedef struct VkPhysicalDeviceShaderBfloat16FeaturesKHR {
      VkStructureType                       sType;
-@@ -13655,6 +13656,7 @@ struct ggml_backend_vk_device_context {
+@@ -13576,6 +13577,7 @@ struct ggml_backend_vk_device_context {
      std::string pci_id;
      std::string id;
      std::string uuid;
@@ -57,7 +57,7 @@ index a234eda2e..c98f98c73 100644
      int major;
      int minor;
      int driver_major;
-@@ -13673,6 +13675,20 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
+@@ -13594,6 +13596,20 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
      
      vk::PhysicalDeviceProperties2 props2;
      vkdev.getProperties2(&props2);
@@ -78,7 +78,7 @@ index a234eda2e..c98f98c73 100644
  
      if (!is_integrated_gpu)
      {
-@@ -13704,7 +13720,6 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
+@@ -13625,7 +13641,6 @@ void ggml_backend_vk_get_device_memory(ggml_backend_vk_device_context *ctx, size
      }
      // else fallback to memory budget if supported
  
@@ -86,7 +86,7 @@ index a234eda2e..c98f98c73 100644
      if (membudget_supported) {
          memprops.pNext = &budgetprops;
      }
-@@ -14440,7 +14455,6 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+@@ -14357,7 +14372,6 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
                      /* .reg     = */ reg,
                      /* .context = */ ctx,
                  });
@@ -94,7 +94,7 @@ index a234eda2e..c98f98c73 100644
                  // Gather additional information about the device
                  int dev_idx = vk_instance.device_indices[i];
                  vk::PhysicalDeviceProperties props1;
-@@ -14463,6 +14477,14 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
+@@ -14380,6 +14394,14 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
                      }
                  }
                  ctx->uuid = oss.str();

llama/patches/0029-ggml-cuda-skip-large-batches.patch

@@ -10,10 +10,10 @@ fallback to cpu
  1 file changed, 3 insertions(+)
 
 diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
-index c1bfadb3e..16c166a08 100644
+index cd71902df..d69d62193 100644
 --- a/ggml/src/ggml-cuda/ggml-cuda.cu
 +++ b/ggml/src/ggml-cuda/ggml-cuda.cu
-@@ -4570,6 +4570,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
+@@ -4632,6 +4632,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                  if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
                      return false;
                  }

llama/patches/0030-win-exit-instead-of-abort.patch

@@ -8,10 +8,10 @@ Subject: [PATCH] win: exit instead of abort
  1 file changed, 6 insertions(+), 1 deletion(-)
 
 diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
-index b99345a2e..1c9e0bc05 100644
+index 530ff7b95..fc0196eb7 100644
 --- a/ggml/src/ggml.c
 +++ b/ggml/src/ggml.c
-@@ -229,8 +229,13 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) {
+@@ -250,8 +250,13 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) {
          fprintf(stderr, "%s\n", message);
          ggml_print_backtrace();
      }

llama/patches/0031-fix-bakllava-regression.patch

@@ -9,10 +9,10 @@ Rever to prior logic of assuming an empty projector type is mlp
  1 file changed, 4 insertions(+)
 
 diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
-index f4c4d2c48..3334ff25b 100644
+index 6be1470ad..2a325c726 100644
 --- a/tools/mtmd/clip.cpp
 +++ b/tools/mtmd/clip.cpp
-@@ -2648,6 +2648,10 @@ struct clip_model_loader {
+@@ -2649,6 +2649,10 @@ struct clip_model_loader {
              if (proj_type.empty()) {
                  if (modality == CLIP_MODALITY_VISION) {
                      get_string(KEY_VISION_PROJ_TYPE, proj_type, false);

ml/backend/ggml/ggml/include/ggml-rpc.h

@@ -1,6 +1,5 @@
 #pragma once
 
-#include "ggml.h"
 #include "ggml-backend.h"
 
 #ifdef  __cplusplus
@@ -8,7 +7,7 @@ extern "C" {
 #endif
 
 #define RPC_PROTO_MAJOR_VERSION    3
-#define RPC_PROTO_MINOR_VERSION    5
+#define RPC_PROTO_MINOR_VERSION    6
 #define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16
 

ml/backend/ggml/ggml/include/ggml.h

@@ -204,6 +204,10 @@
 #    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
 #endif
 
+#if defined(_WIN32) && !defined(_WIN32_WINNT)
+#    define _WIN32_WINNT 0x0A00
+#endif
+
 #include <stdbool.h>
 #include <stddef.h>
 #include <stdint.h>
@@ -2148,7 +2152,8 @@ extern "C" {
     };
 
     enum ggml_scale_flag {
-        GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8)
+        GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8),
+        GGML_SCALE_FLAG_ANTIALIAS     = (1 << 9),
     };
 
     // interpolate
@@ -2191,6 +2196,15 @@ extern "C" {
             int                  p2,
             int                  p3);
 
+    // pad each dimension with values on the other side of the torus (looping around)
+    GGML_API struct ggml_tensor * ggml_pad_circular(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   p0,
+            int                   p1,
+            int                   p2,
+            int                   p3);
+
     GGML_API struct ggml_tensor * ggml_pad_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -2204,6 +2218,19 @@ extern "C" {
             int                  rp3
             );
 
+    // pad each dimension with values on the other side of the torus (looping around)
+    GGML_API struct ggml_tensor * ggml_pad_ext_circular(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   lp0,
+            int                   rp0,
+            int                   lp1,
+            int                   rp1,
+            int                   lp2,
+            int                   rp2,
+            int                   lp3,
+            int                   rp3);
+
     // pad each dimension with reflection: [a, b, c, d] -> [b, a, b, c, d, c]
     GGML_API struct ggml_tensor * ggml_pad_reflect_1d(
             struct ggml_context * ctx,
@@ -2278,7 +2305,7 @@ extern "C" {
             float                 stop,
             float                 step);
 
-#define GGML_KQ_MASK_PAD 64
+#define GGML_KQ_MASK_PAD 1
 
     // q:    [n_embd_k, n_batch,     n_head,    ne3 ]
     // k:    [n_embd_k, n_kv,        n_head_kv, ne3 ]

ml/backend/ggml/ggml/src/CMakeLists.txt

@@ -127,10 +127,6 @@ if (NOT MSVC)
     endif()
 endif()
 
-if (MINGW)
-    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
-endif()
-
 #
 # POSIX conformance
 #
@@ -445,6 +441,7 @@ ggml_add_backend(WebGPU)
 ggml_add_backend(zDNN)
 ggml_add_backend(OpenCL)
 ggml_add_backend(Hexagon)
+ggml_add_backend(ZenDNN)
 
 foreach (target ggml-base ggml)
     target_include_directories(${target} PUBLIC    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../include> $<INSTALL_INTERFACE:include>)

ml/backend/ggml/ggml/src/ggml-alloc.c

@@ -25,6 +25,7 @@ static bool ggml_is_view(const struct ggml_tensor * t) {
 // ops that return true for this function must not use restrict pointers for their backend implementations
 bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {
+        case GGML_OP_FILL:
         case GGML_OP_SCALE:
         case GGML_OP_DIAG_MASK_ZERO:
         case GGML_OP_DIAG_MASK_INF:

ml/backend/ggml/ggml/src/ggml-backend-reg.cpp

@@ -73,6 +73,10 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef GGML_USE_ZENDNN
+#include "ggml-zendnn.h"
+#endif
+
 // disable C++17 deprecation warning for std::codecvt_utf8
 #if defined(__clang__)
 #    pragma clang diagnostic push
@@ -215,6 +219,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_OPENCL
         register_backend(ggml_backend_opencl_reg());
 #endif
+#ifdef GGML_USE_ZENDNN
+        register_backend(ggml_backend_zendnn_reg());
+#endif
 #ifdef GGML_USE_HEXAGON
         register_backend(ggml_backend_hexagon_reg());
 #endif
@@ -551,8 +558,12 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
     fs::path best_path;
 
     for (const auto & search_path : search_paths) {
-        if (!fs::exists(search_path)) {
+        if (std::error_code ec; !fs::exists(search_path, ec)) {
+            if (ec) {
+                GGML_LOG_DEBUG("%s: posix_stat(%s) failure, error-message: %s\n", __func__, path_str(search_path).c_str(), ec.message().c_str());
+            } else {
                 GGML_LOG_DEBUG("%s: search path %s does not exist\n", __func__, path_str(search_path).c_str());
+            }
             continue;
         }
         fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
@@ -592,8 +603,12 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
         for (const auto & search_path : search_paths) {
             fs::path filename = backend_filename_prefix().native() + name_path.native() + backend_filename_extension().native();
             fs::path path = search_path / filename;
-            if (fs::exists(path)) {
+            if (std::error_code ec; fs::exists(path, ec)) {
                 return get_reg().load_backend(path, silent);
+            } else {
+                if (ec) {
+                    GGML_LOG_DEBUG("%s: posix_stat(%s) failure, error-message: %s\n", __func__, path_str(path).c_str(), ec.message().c_str());
+                }
             }
         }
         return nullptr;
@@ -614,6 +629,7 @@ void ggml_backend_load_all_from_path(const char * dir_path) {
 #endif
 
     ggml_backend_load_best("blas", silent, dir_path);
+    ggml_backend_load_best("zendnn", silent, dir_path);
     ggml_backend_load_best("cann", silent, dir_path);
     ggml_backend_load_best("cuda", silent, dir_path);
     ggml_backend_load_best("hip", silent, dir_path);

ml/backend/ggml/ggml/src/ggml-backend.cpp

@@ -754,6 +754,12 @@ struct ggml_backend_sched {
 
     int debug;
 
+    // used for debugging graph reallocations [GGML_SCHED_DEBUG_REALLOC]
+    // ref: https://github.com/ggml-org/llama.cpp/pull/17617
+    int debug_realloc;
+    int debug_graph_size;
+    int debug_prev_graph_size;
+
     // allocate buffers on attached ggml_backend_buffer_type_t's and during reservation
     // if false, dummy buffers are used for faster memory sizing calculations
     // the scheduler needs to be recreated with allocated buffers before it can be used
@@ -1270,10 +1276,8 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
                                 tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
                                 ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
                             }
-                            if (sched->n_copies > 1) {
                             ggml_set_input(tensor_copy);
                             ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
-                            }
                             tensor_id_copy(src_id, src_backend_id, c) = tensor_copy;
                             SET_CAUSE(tensor_copy, "4.cpy");
                         }
@@ -1325,6 +1329,11 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
     }
 
     int graph_size = std::max(graph->n_nodes, graph->n_leafs) + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sched->n_copies;
+
+    // remember the actual graph_size for performing reallocation checks later [GGML_SCHED_DEBUG_REALLOC]
+    sched->debug_prev_graph_size = sched->debug_graph_size;
+    sched->debug_graph_size = graph_size;
+
     if (sched->graph.size < graph_size) {
         sched->graph.size = graph_size;
         sched->graph.nodes = (ggml_tensor **) realloc(sched->graph.nodes, graph_size * sizeof(struct ggml_tensor *));
@@ -1431,14 +1440,21 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
 
     // allocate graph
     if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
-#ifdef GGML_SCHED_NO_REALLOC
-        GGML_ABORT("%s: failed to allocate graph, but graph re-allocation is disabled by GGML_SCHED_NO_REALLOC\n", __func__);
-#endif
-
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
 #endif
 
+        if (sched->debug_realloc > 0) {
+            // we are interested only in situations where the graph was reallocated even though its size remained the same [GGML_SCHED_DEBUG_REALLOC]
+            // example: https://github.com/ggml-org/llama.cpp/pull/17143
+            const bool unexpected = !backend_ids_changed && sched->debug_prev_graph_size == sched->debug_graph_size;
+
+            if (unexpected || sched->debug_realloc > 1) {
+                GGML_ABORT("%s: unexpected graph reallocation (graph size = %d, nodes = %d, leafs = %d), debug_realloc = %d\n", __func__,
+                        sched->debug_graph_size, sched->graph.n_nodes, sched->graph.n_leafs, sched->debug_realloc);
+            }
+        }
+
         // the re-allocation may cause the split inputs to be moved to a different address
         // synchronize without ggml_backend_sched_synchronize to avoid changing cur_copy
         for (int i = 0; i < sched->n_backends; i++) {
@@ -1667,6 +1683,14 @@ ggml_backend_sched_t ggml_backend_sched_new_ext(
 
     const char * GGML_SCHED_DEBUG = getenv("GGML_SCHED_DEBUG");
     sched->debug = GGML_SCHED_DEBUG ? atoi(GGML_SCHED_DEBUG) : 0;
+
+    sched->debug_realloc = 0;
+#ifdef GGML_SCHED_NO_REALLOC
+    sched->debug_realloc = 1;
+#endif
+    const char * GGML_SCHED_DEBUG_REALLOC = getenv("GGML_SCHED_DEBUG_REALLOC");
+    sched->debug_realloc = GGML_SCHED_DEBUG_REALLOC ? atoi(GGML_SCHED_DEBUG_REALLOC) : sched->debug_realloc;
+
     sched->n_backends = n_backends;
     sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
 
@@ -1683,6 +1707,9 @@ ggml_backend_sched_t ggml_backend_sched_new_ext(
     sched->prev_node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
     sched->prev_leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
 
+    sched->debug_graph_size = 0;
+    sched->debug_prev_graph_size = 0;
+
     sched->context_buffer_size = ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
     sched->context_buffer = (char *) malloc(sched->context_buffer_size);
 

ml/backend/ggml/ggml/src/ggml-cpu/CMakeLists.txt

@@ -469,6 +469,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             if (GGML_RV_ZICBOP)
                 string(APPEND MARCH_STR "_zicbop")
             endif()
+            if (GGML_RV_ZIHINTPAUSE)
+                string(APPEND MARCH_STR "_zihintpause")
+            endif()
             list(APPEND ARCH_FLAGS "-march=${MARCH_STR}" -mabi=lp64d)
         else()
             # Begin with the lowest baseline

ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/cpu-feats.cpp

@@ -8,6 +8,10 @@
 #include <sys/sysctl.h>
 #endif
 
+#if !defined(HWCAP2_SVE2)
+#define HWCAP2_SVE2 (1 << 1)
+#endif
+
 #if !defined(HWCAP2_I8MM)
 #define HWCAP2_I8MM (1 << 13)
 #endif

ml/backend/ggml/ggml/src/ggml-cpu/arch/arm/repack.cpp

@@ -505,7 +505,6 @@ void ggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     constexpr int blocklen          = 8;
 
     assert(n % qk == 0);
-    assert(nr % 4 == 0);
     assert(nc % ncols_interleaved == 0);
 
     UNUSED(nb);
@@ -645,7 +644,6 @@ void ggml_gemv_q4_K_8x8_q8_K(int                        n,
     constexpr int blocklen          = 8;
 
     assert(n % qk == 0);
-    assert(nr % 4 == 0);
     assert(nc % ncols_interleaved == 0);
 
     UNUSED(nb);

ml/backend/ggml/ggml/src/ggml-cpu/ggml-cpu.c

@@ -492,6 +492,15 @@ static inline void ggml_thread_cpu_relax(void) {
 static inline void ggml_thread_cpu_relax(void) {
     _mm_pause();
 }
+#elif defined(__riscv)
+static inline void ggml_thread_cpu_relax(void) {
+    #ifdef __riscv_zihintpause
+        __asm__ __volatile__ ("pause");
+    #else
+        /* Encoding of the pause instruction */
+        __asm__ __volatile__ (".4byte 0x100000F");
+    #endif
+}
 #else
 static inline void ggml_thread_cpu_relax(void) {;}
 #endif
@@ -685,22 +694,14 @@ bool ggml_is_numa(void) {
 }
 
 #if defined(__ARM_ARCH)
-
-#if defined(__linux__) && defined(__aarch64__)
-#include <sys/auxv.h>
-#endif
-
-static void ggml_init_arm_arch_features(void) {
 #if defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
-#if defined(__linux__)
-    ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
-#else
-    // TODO: add support of SVE for non-linux systems
-#error "TODO: SVE is not supported on this platform. To use SVE, sve_cnt needs to be initialized here."
-#endif
-#endif
+#include <arm_sve.h>
+static void ggml_init_arm_arch_features(void) {
+    ggml_arm_arch_features.sve_cnt = svcntb();
 }
-
+#else
+static void ggml_init_arm_arch_features(void) {}
+#endif
 #endif // __ARM_ARCH
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
@@ -2708,6 +2709,11 @@ struct ggml_cplan ggml_graph_plan(
         n_threads = threadpool ? threadpool->n_threads_max : GGML_DEFAULT_N_THREADS;
     }
 
+#if defined(__EMSCRIPTEN__) && !defined(__EMSCRIPTEN_PTHREADS__)
+    // Emscripten without pthreads support can only use a single thread
+    n_threads = 1;
+#endif
+
     size_t work_size = 0;
 
     struct ggml_cplan cplan;

ml/backend/ggml/ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -117,8 +117,7 @@ inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); }
 #endif
 
 #if defined(__MMA__)
-typedef vector unsigned char vec_t;
-typedef __vector_quad acc_t;
+#include "sgemm-ppc.h"
 #endif
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED FUSED MULTIPLY ADD
@@ -1573,95 +1572,35 @@ class tinyBLAS_BF16_PPC {
     const int nth;
 };
 
-template <typename TA>
-class tinyBLAS_Q0_PPC {
-  public:
-    tinyBLAS_Q0_PPC(int64_t k,
+    template <typename TA>
+    tinyBLAS_Q0_PPC<TA>::tinyBLAS_Q0_PPC(int64_t k,
         const TA *A, int64_t lda,
         const block_q8_0 *B, int64_t ldb,
         float *C, int64_t ldc,
         int ith, int nth)
         : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+                kc = 64;
     }
 
-    void matmul(int64_t m, int64_t n) {
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::matmul(int64_t m, int64_t n) {
+        int mc = 64; int nc = 64;
+        if (n % 8 == 0 && n < nc) {
+                nc = n;
+                mc = 32 ;
+                kc = 32;
+        }
+        const bool is_aligned = ((m & (mc - 1)) == 0) & ((n & (nc - 1)) == 0) & ((k & (kc - 1)) == 0);
+        if (is_aligned) {
+            this->matmul_tiled_q0(m, n, mc, nc, kc);
+        } else {
             mnpack(0, m, 0, n);
         }
-
-  private:
-
-    inline void save_res(int ii, int jj, int idx, vector float* fin_res, int RM=4, int RN=4) {
-       for (int I = 0; I < RM; I++) {
-          for (int J = 0; J < RN; J++) {
-             *((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&fin_res[idx+I]+J);
-          }
-       }
     }
 
+   template<typename TA>
    template<int size>
-    inline void compute(acc_t* ACC, int c_idx, int s_idx, std::array<int, size>& comparray, vector float* vs, vector float* fin_res) {
-       vector signed int vec_C[4];
-       vector float CA[4] = {0};
-       vector float res[4] = {0};
-       __builtin_mma_disassemble_acc(vec_C, ACC);
-       for (int i = 0; i < 4; i++) {
-          CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0));
-          res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
-          fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]);
-       }
-    }
-    /* This function processes quantized data from block_q4_0 elements.
-     * First the we try to extract the two int4 values stored in single int8_t into two signed int8.
-     * And then we subtract each of the resultant element with 8, to convert signed int8 to unsigned int8.
-     * Also compute the rowsum which is required to compensate the above conversion. */
-    inline void process_q4_elements(vector signed char (&c)[2], int* ca) {
-        const vector signed char lowMask = vec_splats((signed char)0xF);
-        const vector unsigned char v4 = vec_splats((unsigned char)0x4);
-        const vector signed char v8 = vec_splats((signed char)0x8);
-        vector signed int vsum = {0};
-        vector signed int vsum2 = {0};
-        c[0] = vec_and(c[1], lowMask);
-        c[1] = vec_sr(c[1], v4);
-        c[0] = vec_sub(c[0], v8);
-        c[1] = vec_sub(c[1], v8);
-        vsum = vec_sum4s(c[0], vsum);
-        vsum2 = vec_sum4s(c[1], vsum2);
-        vsum = vec_add(vsum, vsum2);
-        *(ca) = vsum[0] + vsum[1] + vsum[2] + vsum[3];
-    }
-
-    template <typename V1, typename V2>
-    inline void vector_permute_store(V2 &s1, V2 &s2, V2 &s3, V2 &s4, V1 *vecOffset, bool flip) {
-        vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
-        vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
-        vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
-        vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
-        V2 t1, t2, t3, t4, t5, t6, t7, t8;
-        vector unsigned char xor_vector;
-        uint8_t flip_vec = 0x80;
-        xor_vector = vec_splats(flip_vec);
-        t1 = vec_perm(s1, s2, swiz1);
-        t2 = vec_perm(s1, s2, swiz2);
-        t3 = vec_perm(s3, s4, swiz1);
-        t4 = vec_perm(s3, s4, swiz2);
-        t5 = vec_perm(t1, t3, swiz3);
-        t6 = vec_perm(t1, t3, swiz4);
-        t7 = vec_perm(t2, t4, swiz3);
-        t8 = vec_perm(t2, t4, swiz4);
-        if (flip == true) {
-            t5 = vec_xor(t5, xor_vector);
-            t6 = vec_xor(t6, xor_vector);
-            t7 = vec_xor(t7, xor_vector);
-            t8 = vec_xor(t8, xor_vector);
-        }
-        vec_xst(t5, 0, vecOffset);
-        vec_xst(t6, 0, vecOffset+16);
-        vec_xst(t7, 0, vecOffset+32);
-        vec_xst(t8, 0, vecOffset+48);
-    }
-
-    template<int size>
-    void packNormalInt4(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, std::array<int, size>& comparray) {
+   void tinyBLAS_Q0_PPC<TA>::packNormalInt4(const TA* a, int64_t lda, int rows, int cols, int8_t* vec, std::array<int, size>& comparray) {
         int64_t i, j;
         TA *aoffset = NULL;
         int8_t *vecOffset = NULL;
@@ -1781,8 +1720,10 @@ class tinyBLAS_Q0_PPC {
             }
         }
     }
+
+    template<typename TA>
     template<typename VA, typename VB>
-    void packNormal(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
+    void tinyBLAS_Q0_PPC<TA>::packNormal(const block_q8_0* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
         int64_t i, j;
         block_q8_0 *aoffset = NULL;
         VA *vecOffset = NULL;
@@ -1822,7 +1763,6 @@ class tinyBLAS_Q0_PPC {
             j--;
         } while(j > 0);
     }
-
     if (rows & 4) {
             aoffsets[0]  = aoffset;
             for (int it = 1; it < 4; it++ )
@@ -1878,7 +1818,8 @@ class tinyBLAS_Q0_PPC {
         }
     }
 
-    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
         int m_rem = MIN(m - m0, 16);
         int n_rem = MIN(n - n0, 16);
 
@@ -1915,7 +1856,8 @@ class tinyBLAS_Q0_PPC {
     }
 
 
-    void KERNEL_4x8(int64_t ii, int64_t jj) {
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::KERNEL_4x8(int64_t ii, int64_t jj) {
         vec_t vec_A[8], vec_B[16] = {0};
         acc_t acc_0, acc_1;
         std::array<int, 4> comparray {};
@@ -1953,14 +1895,15 @@ class tinyBLAS_Q0_PPC {
                     aoffset += lda;
                 }
             }
-            compute<4>(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute<4>(&acc_1, 0, 4, comparray, vs, fin_res);
+            compute(&acc_0, 0, 0, comparray, vs, fin_res);
+            compute(&acc_1, 0, 4, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
         save_res(ii, jj+4, 4, fin_res);
     }
 
-    void KERNEL_8x4(int64_t ii, int64_t jj) {
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::KERNEL_8x4(int64_t ii, int64_t jj) {
         vec_t vec_A[16], vec_B[8] = {0};
         acc_t acc_0, acc_1;
         std::array<int, 8> comparray {};
@@ -1997,16 +1940,18 @@ class tinyBLAS_Q0_PPC {
                     aoffset += lda;
                 }
             }
-            compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
+            compute(&acc_0, 0, 0, comparray, vs, fin_res);
+            compute(&acc_1, 4, 4, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
         save_res(ii+4, jj, 4, fin_res);
     }
 
-    void KERNEL_8x8(int64_t ii, int64_t jj) {
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::KERNEL_8x8(int64_t ii, int64_t jj) {
         vec_t vec_A[16], vec_B[16] = {0};
         acc_t acc_0, acc_1, acc_2, acc_3;
+        acc_t acc_4, acc_5, acc_6, acc_7;
         std::array<int, 8> comparray {};
         vector float fin_res[16] = {0};
         vector float vs[16] = {0};
@@ -2046,10 +1991,10 @@ class tinyBLAS_Q0_PPC {
                     aoffset += lda;
                 }
             }
-            compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
-            compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
-            compute<8>(&acc_2, 0, 8, comparray, vs, fin_res);
-            compute<8>(&acc_3, 4, 12, comparray, vs, fin_res);
+            compute(&acc_0, 0, 0, comparray, vs, fin_res);
+            compute(&acc_1, 4, 4, comparray, vs, fin_res);
+            compute(&acc_2, 0, 8, comparray, vs, fin_res);
+            compute(&acc_3, 4, 12, comparray, vs, fin_res);
         }
         save_res(ii, jj, 0, fin_res);
         save_res(ii+4, jj, 4, fin_res);
@@ -2057,7 +2002,8 @@ class tinyBLAS_Q0_PPC {
         save_res(ii+4, jj+4, 12, fin_res);
     }
 
-    void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
+    template<typename TA>
+    void tinyBLAS_Q0_PPC<TA>::gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
         int64_t ytiles = (m - m0) / RM;
         int64_t xtiles = (n - n0) / RN;
         int64_t tiles = xtiles * ytiles;
@@ -2125,21 +2071,9 @@ class tinyBLAS_Q0_PPC {
         }
     }
 
-    template<int RM, int RN>
-    inline void kernel(int64_t ii, int64_t jj) {
-       if constexpr(RM == 4 && RN == 8) {
-          KERNEL_4x8(ii,jj);
-       } else if constexpr(RM == 8 && RN == 4) {
-          KERNEL_8x4(ii,jj);
-       } else if constexpr(RM == 8 && RN == 8) {
-          KERNEL_8x8(ii,jj);
-       } else {
-          assert(false && "RN/RM values not supported");
-       }
-    }
-
+    template<typename TA>
     template <int RM, int RN>
-    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+    NOINLINE void tinyBLAS_Q0_PPC<TA>::gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
         int64_t ytiles = (m - m0) / RM;
         int64_t xtiles = (n - n0) / RN;
         int64_t tiles = xtiles * ytiles;
@@ -2151,20 +2085,12 @@ class tinyBLAS_Q0_PPC {
         for (int64_t job = start; job < end; ++job) {
             int64_t ii = m0 + job / xtiles * RM;
             int64_t jj = n0 + job % xtiles * RN;
-            kernel<RM, RN>(ii, jj);
+            this->kernel<RM, RN>(ii, jj);
         }
     }
 
-    const TA *const A;
-    const block_q8_0 *const B;
-    float *C;
-    const int64_t k;
-    const int64_t lda;
-    const int64_t ldb;
-    const int64_t ldc;
-    const int ith;
-    const int nth;
-};
+template class tinyBLAS_Q0_PPC<block_q4_0>;
+template class tinyBLAS_Q0_PPC<block_q8_0>;
 
 class tinyBLAS_PPC {
   public:

ml/backend/ggml/ggml/src/ggml-cpu/llamafile/sgemm.h

@@ -6,6 +6,12 @@
 #include <vecintrin.h>
 #endif
 
+#ifdef _MSC_VER
+#define NOINLINE __declspec(noinline)
+#else
+#define NOINLINE __attribute__((__noinline__))
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif

ml/backend/ggml/ggml/src/ggml-cpu/ops.cpp

@@ -6383,7 +6383,7 @@ static void ggml_compute_forward_im2col_3d_f16(
                                         const int64_t iih = ioh*s1 + ikh*d1 - p1;
                                         const int64_t iid = iod*s2 + ikd*d2 - p2;
 
-                                        if (iid < 0 || iid >= ID || iih < 0 || iih >= IH || iiw < 0 || iiw >= IW || iid < 0 || iid >= ID) {
+                                        if (iid < 0 || iid >= ID || iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
                                             dst_data[iic*KD_KH_KW + ikd * KH_KW + ikh*KW + ikw] = 0;
                                         } else {
                                             const float * const s = (const float *) ((const char *)src_data + iid*nb12 + iih*nb11 + iiw*nb10); // [ID, IH, IW]
@@ -6554,8 +6554,13 @@ static void ggml_call_mul_mat(ggml_type type, const ggml_compute_params * params
     ggml_compute_forward_mul_mat(params, &dst);
 }
 
+static inline int64_t ggml_wrap_around(int64_t coord, int64_t size) {
+    return (coord  + size) % size; // adding size avoids negative number weirdness
+}
+
 // ggml_compute_forward_conv_2d
 
+
 static void ggml_compute_forward_conv_2d_impl(const ggml_compute_params * params,
                                               const ggml_tensor *         kernel,  // [KW, KH, IC, OC]
                                               const ggml_tensor *         src,     // [W, H, C, N]
@@ -7420,6 +7425,65 @@ static void ggml_compute_forward_upscale_f32(
                 }
             }
         }
+    } else if (mode == GGML_SCALE_MODE_BILINEAR && (mode_flags & GGML_SCALE_FLAG_ANTIALIAS)) {
+        // Similar to F.interpolate(..., mode="bilinear", align_corners=False, antialias=True)
+        // https://github.com/pytorch/pytorch/blob/8871ff29b743948d1225389d5b7068f37b22750b/aten/src/ATen/native/cpu/UpSampleKernel.cpp
+        auto triangle_filter = [](float x) -> float {
+            return std::max(1.0f - fabsf(x), 0.0f);
+        };
+
+        // support and invscale, minimum 1 pixel for bilinear
+        const float support1  = std::max(1.0f, 1.0f / sf1);
+        const float invscale1 = 1.0f / support1;
+        const float support0  = std::max(1.0f, 1.0f / sf0);
+        const float invscale0 = 1.0f / support0;
+
+        for (int64_t i3 = 0; i3 < ne3; i3++) {
+            const int64_t i03 = i3 / sf3;
+            for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+                const int64_t i02 = i2 / sf2;
+                for (int64_t i1 = 0; i1 < ne1; i1++) {
+                    const float y = ((float) i1 + pixel_offset) / sf1;
+                    for (int64_t i0 = 0; i0 < ne0; i0++) {
+                        const float x = ((float) i0 + pixel_offset) / sf0;
+
+                        // the range of source pixels that contribute
+                        const int64_t x_min = std::max<int64_t>(x - support0 + pixel_offset, 0);
+                        const int64_t x_max = std::min<int64_t>(x + support0 + pixel_offset, ne00);
+                        const int64_t y_min = std::max<int64_t>(y - support1 + pixel_offset, 0);
+                        const int64_t y_max = std::min<int64_t>(y + support1 + pixel_offset, ne01);
+
+                        // bilinear filter with antialiasing
+                        float val = 0.0f;
+                        float total_weight = 0.0f;
+
+                        for (int64_t sy = y_min; sy < y_max; sy++) {
+                            const float weight_y = triangle_filter((sy - y + pixel_offset) * invscale1);
+
+                            for (int64_t sx = x_min; sx < x_max; sx++) {
+                                const float weight_x = triangle_filter((sx - x + pixel_offset) * invscale0);
+                                const float weight = weight_x * weight_y;
+
+                                if (weight <= 0.0f) {
+                                    continue;
+                                }
+
+                                const float pixel = *(const float *)((const char *)src0->data + sx*nb00 + sy*nb01 + i02*nb02 + i03*nb03);
+                                val += pixel * weight;
+                                total_weight += weight;
+                            }
+                        }
+
+                        if (total_weight > 0.0f) {
+                            val /= total_weight;
+                        }
+
+                        float * dst_ptr = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+                        *dst_ptr = val;
+                    }
+                }
+            }
+        }
     } else if (mode == GGML_SCALE_MODE_BILINEAR) {
         for (int64_t i3 = 0; i3 < ne3; i3++) {
             const int64_t i03 = i3 / sf3;
@@ -7532,6 +7596,7 @@ void ggml_compute_forward_upscale(
 
 // ggml_compute_forward_pad
 
+template<bool circular_t>
 static void ggml_compute_forward_pad_f32(
     const ggml_compute_params * params,
           ggml_tensor * dst) {
@@ -7556,13 +7621,29 @@ static void ggml_compute_forward_pad_f32(
     const int32_t lp3 = ggml_get_op_params_i32(dst, 6);
     const int32_t rp3 = ggml_get_op_params_i32(dst, 7);
 
-
     // TODO: optimize
 
     for (int64_t i2 = 0; i2 < ne2; ++i2) {
         for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
             for (int64_t i0 = 0; i0 < ne0; ++i0) {
                 for (int64_t i3 = 0; i3 < ne3; ++i3) {
+                    // circular means wrap around on a torus, so x and y loop around
+                    if constexpr (circular_t) {
+                        const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
+                        const int64_t src_i0 = ggml_wrap_around(i0 - lp0, ne00);
+                        const int64_t src_i1 = ggml_wrap_around(i1 - lp1, ne01);
+                        const int64_t src_i2 = ggml_wrap_around(i2 - lp2, ne02);
+                        const int64_t src_i3 = ggml_wrap_around(i3 - lp3, ne03);
+
+                        const int64_t src_idx =
+                            src_i3*nb03 +
+                            src_i2*nb02 +
+                            src_i1*nb01 +
+                            src_i0*nb00;
+
+                        const float * src_ptr = (const float *)((char *) src0->data + src_idx);
+                        dst_ptr[dst_idx] = *src_ptr;
+                    } else {
                         const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
                         if ((i0 >= lp0 && i0 < ne0 - rp0) \
                             && (i1 >= lp1 && i1 < ne1 - rp1) \
@@ -7578,18 +7659,23 @@ static void ggml_compute_forward_pad_f32(
                 }
             }
         }
+    }
 }
 
+
 void ggml_compute_forward_pad(
     const ggml_compute_params * params,
     ggml_tensor * dst) {
-
     const ggml_tensor * src0 = dst->src[0];
-
+    const bool circular = (bool) ggml_get_op_params_i32(dst, 8);
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_pad_f32(params, dst);
+                if (circular) {
+                    ggml_compute_forward_pad_f32<true>(params, dst);
+                } else {
+                    ggml_compute_forward_pad_f32<false>(params, dst);
+                }
             } break;
         default:
             {

ml/backend/ggml/ggml/src/ggml-cuda/common.cuh

@@ -261,7 +261,7 @@ static const char * cu_get_error_str(CUresult err) {
 #define AMD_MFMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 
-#if defined(GGML_USE_HIP) && defined(RDNA4)
+#if defined(GGML_USE_HIP) && (defined(RDNA4) || defined(RDNA3))
 #define AMD_WMMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(RDNA4)
 
@@ -329,7 +329,7 @@ static bool amd_mfma_available(const int cc) {
 }
 
 static bool amd_wmma_available(const int cc) {
-    return GGML_CUDA_CC_IS_RDNA4(cc);
+    return (GGML_CUDA_CC_IS_RDNA4(cc) || GGML_CUDA_CC_IS_RDNA3(cc));
 }
 
 static bool volta_mma_available(const int cc) {
@@ -498,6 +498,53 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
     return x;
 }
 
+template<typename T, int width = WARP_SIZE>
+static __device__ __forceinline__ T warp_prefix_inclusive_sum(T x) {
+    const int lane_id = threadIdx.x % width;
+#pragma unroll
+    for (int offset = 1; offset < width; offset <<= 1) {
+        const T t = __shfl_up_sync(0xffffffff, x, offset, width);
+        if (lane_id >= offset) {
+            x += t;
+        }
+    }
+    return x;
+}
+
+template<int width = WARP_SIZE>
+static __device__ __forceinline__ float2 warp_prefix_inclusive_sum(float2 a) {
+    const int lane_id = threadIdx.x % width;
+#pragma unroll
+    for (int offset = 1; offset < width; offset <<= 1) {
+        const float t_x = __shfl_up_sync(0xffffffff, a.x, offset, width);
+        const float t_y = __shfl_up_sync(0xffffffff, a.y, offset, width);
+        if (lane_id >= offset) {
+            a.x += t_x;
+            a.y += t_y;
+        }
+    }
+    return a;
+}
+
+template<int width = WARP_SIZE>
+static __device__ __forceinline__ half2 warp_prefix_inclusive_sum(half2 a) {
+#ifdef FP16_AVAILABLE
+    const int lane_id = threadIdx.x % width;
+#pragma unroll
+    for (int offset = 1; offset < width; offset <<= 1) {
+        const half2 t = __shfl_up_sync(0xffffffff, a, offset, width);
+        if (lane_id >= offset) {
+            a = __hadd2(a, t);
+        }
+    }
+    return a;
+
+#else
+    NO_DEVICE_CODE;
+    return a;
+#endif // FP16_AVAILABLE
+}
+
 static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
 #ifdef FP16_AVAILABLE
 
@@ -1027,6 +1074,10 @@ struct ggml_cuda_concurrent_event {
     int                                          n_streams = 0;
     std::unordered_map<const ggml_tensor *, int> stream_mapping;
 
+    // Original order of nodes in this concurrent region (before interleaving)
+    // Used to restore grouping for fusion within streams
+    std::vector<const ggml_tensor *> original_order;
+
     const ggml_tensor * join_node;
 
     ggml_cuda_concurrent_event() = default;
@@ -1049,6 +1100,7 @@ struct ggml_cuda_concurrent_event {
     , fork_event(other.fork_event)
     , n_streams(other.n_streams)
     , stream_mapping(std::move(other.stream_mapping))
+    , original_order(std::move(other.original_order))
     , join_node(other.join_node) {
         other.fork_event = nullptr;
     }
@@ -1159,11 +1211,9 @@ struct ggml_cuda_concurrent_event {
 };
 
 struct ggml_cuda_stream_context {
-    std::vector<const ggml_tensor *>                                    original_nodes;
     std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> concurrent_events;
 
     void reset() {
-        original_nodes.clear();
         concurrent_events.clear();
     }
 };

ml/backend/ggml/ggml/src/ggml-cuda/cumsum.cu

@@ -0,0 +1,237 @@
+#include <algorithm>
+#include "cumsum.cuh"
+#include "convert.cuh"
+#include "ggml-cuda/common.cuh"
+#include "ggml.h"
+
+#ifdef GGML_CUDA_USE_CUB
+#   include <cub/device/device_scan.cuh>
+#endif // GGML_CUDA_USE_CUB
+
+template<typename T, int BLOCK_SIZE>
+static __global__ void cumsum_cub_kernel(
+        const T * __restrict__ src,
+        T * __restrict__ dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t  s01, const int64_t  s02, const int64_t  s03,
+        const int64_t   s1,  const int64_t   s2,  const int64_t   s3) {
+#ifdef GGML_CUDA_USE_CUB
+    using BlockScan = cub::BlockScan<T, BLOCK_SIZE>;
+
+    __shared__ typename BlockScan::TempStorage temp_storage;
+    __shared__ T block_carry;      // carry from previous tile
+
+    const int tid = threadIdx.x;
+
+    const int64_t i1 = blockIdx.x;
+    const int64_t i2 = blockIdx.y;
+    const int64_t i3 = blockIdx.z;
+
+    if (i1 >= ne01 || i2 >= ne02 || i3 >= ne03) {
+        return;
+    }
+
+    const T * src_row = src + i1 * s01 + i2 * s02 + i3 * s03;
+    T *       dst_row = dst + i1 * s1  + i2 * s2  + i3 * s3;
+
+    if (tid == 0) {
+        block_carry = 0;
+    }
+    __syncthreads();
+
+    for (int64_t start = 0; start < ne00; start += BLOCK_SIZE) {
+        int64_t idx = start + tid;
+        T x = (idx < ne00) ? src_row[idx] : T(0);
+
+        T inclusive;
+        T block_total;
+        BlockScan(temp_storage).InclusiveSum(x, inclusive, block_total);
+
+        __syncthreads();
+
+        T final_val = inclusive + block_carry;
+
+        // store result
+        if (idx < ne00) {
+            dst_row[idx] = final_val;
+        }
+
+        __syncthreads();
+
+        if (tid == 0) {
+            block_carry += block_total;
+        }
+
+        __syncthreads();
+    }
+#else
+    NO_DEVICE_CODE;
+#endif // GGML_CUDA_USE_CUB
+}
+
+// Fallback kernel implementation (original)
+template<typename T>
+static __global__ void cumsum_kernel(
+        const T * src, T * dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t  s00, const int64_t  s01, const int64_t  s02, const int64_t  s03,
+        const int64_t   s0, const int64_t   s1, const int64_t   s2, const int64_t   s3) {
+
+    GGML_UNUSED_VARS(s00, s0);
+
+    const int tid = threadIdx.x;
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    const int lane = tid % warp_size;
+    const int warp = tid / warp_size;
+    const int warps_per_block = blockDim.x / warp_size;
+
+    extern __shared__ float smem[];
+    float * s_vals = smem;
+    float * s_warp_sums = smem + blockDim.x;
+    float * s_carry = smem + blockDim.x + warps_per_block;
+    float * s_chunk_total = s_carry + 1;
+
+    // Initialize carry
+    if (tid == 0) {
+        *s_carry = 0.0f;
+    }
+    __syncthreads();
+
+    const int64_t i3 = blockIdx.z;
+    const int64_t i2 = blockIdx.y;
+    const int64_t i1 = blockIdx.x;
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    const T * src_row = src + i1 * s01 + i2 * s02 + i3 * s03;
+    T       * dst_row = dst + i1 * s1  + i2 * s2  + i3 * s3;
+
+    for (int64_t start = 0; start < ne00; start += blockDim.x) {
+        int64_t idx = start + tid;
+        float val = (idx < ne00) ? ggml_cuda_cast<float, T>(src_row[idx]) : 0.0f;
+
+        // 1. Warp inclusive scan
+        val = warp_prefix_inclusive_sum<T, warp_size>(val);
+        s_vals[tid] = val;
+
+        // Store warp total
+        if (lane == warp_size - 1) {
+            s_warp_sums[warp] = val;
+        }
+        __syncthreads();
+
+        // 2. Exclusive scan of warp sums (warp 0 only)
+        if (warp == 0) {
+            float w = (tid < warps_per_block) ? s_warp_sums[tid] : 0.0f;
+            float inc = warp_prefix_inclusive_sum<T, warp_size>(w);
+            if (tid < warps_per_block) {
+                s_warp_sums[tid] = inc - w;   // exclusive sum
+            }
+            if (tid == warps_per_block - 1) {
+                *s_chunk_total = inc;          // total sum of this chunk
+            }
+        }
+        __syncthreads();
+
+        float carry = *s_carry;
+        float final_val = s_vals[tid] + s_warp_sums[warp] + carry;
+        if (idx < ne00) {
+            dst_row[idx] = ggml_cuda_cast<T, float>(final_val);
+        }
+        __syncthreads();
+
+        // Update carry for next chunk
+        if (tid == 0) {
+            *s_carry += *s_chunk_total;
+        }
+        __syncthreads();
+    }
+}
+
+template<typename T>
+static void cumsum_cuda(
+        const T * src, T * dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
+        const int64_t  nb0,  const int64_t nb1, const int64_t  nb2, const int64_t  nb3,
+        cudaStream_t stream) {
+
+    const size_t type_size = sizeof(T);
+    bool use_cub = false;
+#ifdef GGML_CUDA_USE_CUB
+    // Check if we can use CUB (data must be contiguous along innermost dimension)
+    const bool is_contiguous = (nb00 == type_size) && (nb0 == type_size);
+
+    if (is_contiguous) {
+        use_cub = true;
+    }
+#endif // GGML_CUDA_USE_CUB
+    dim3 grid_dims(ne01, ne02, ne03);
+    const auto &info = ggml_cuda_info().devices[ggml_cuda_get_device()];
+    const int warp_size = info.warp_size;
+    const int num_warps = (ne00 + warp_size - 1) / warp_size;
+    int block_size = num_warps * warp_size;
+    block_size = std::min(block_size, CUDA_CUMSUM_BLOCK_SIZE);
+    dim3 block_dims(block_size, 1, 1);
+    const int warps_per_block = block_size / warp_size;
+    const size_t shmem_size = (block_size + warps_per_block + 2) * sizeof(float);
+
+    if (use_cub) {
+        cumsum_cub_kernel<T, CUDA_CUMSUM_BLOCK_SIZE><<<grid_dims, CUDA_CUMSUM_BLOCK_SIZE, 0, stream>>>(
+            src, dst,
+            ne00, ne01, ne02, ne03,
+            nb01 / type_size, nb02 / type_size, nb03 / type_size,
+            nb1 / type_size,  nb2 / type_size,  nb3 / type_size
+        );
+    } else {
+        cumsum_kernel<<<grid_dims, block_dims, shmem_size, stream>>>(
+            src, dst,
+            ne00, ne01, ne02, ne03,
+            nb00 / type_size, nb01 / type_size, nb02 / type_size, nb03 / type_size,
+            nb0 / type_size, nb1 / type_size, nb2 / type_size, nb3 / type_size
+        );
+    }
+}
+
+void ggml_cuda_op_cumsum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == dst->type);
+    switch(src0->type) {
+        case GGML_TYPE_F32:
+            {
+                cumsum_cuda(
+                    (const float *)src0->data, (float *)dst->data,
+                    src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+                    src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                    dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
+                    stream
+                );
+            } break;
+        // We do not support those on CPU for now anyway, so comment them out because they cause errors on some CI platforms
+        /*case GGML_TYPE_F16:
+            {
+                cumsum_cuda(
+                    (const half *)src0->data, (half *)dst->data,
+                    src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+                    src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                    dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
+                    stream
+                );
+            } break;
+        case GGML_TYPE_BF16:
+            {
+                cumsum_cuda(
+                    (const nv_bfloat16 *)src0->data, (nv_bfloat16 *)dst->data,
+                    src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+                    src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                    dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
+                    stream
+                );
+            } break;*/
+        default:
+            GGML_ABORT("fatal error");
+    }
+}

ml/backend/ggml/ggml/src/ggml-cuda/cumsum.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_CUMSUM_BLOCK_SIZE 256
+
+void ggml_cuda_op_cumsum(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ml/backend/ggml/ggml/src/ggml-cuda/diag.cu

@@ -0,0 +1,77 @@
+#include "convert.cuh"
+#include "diag.cuh"
+#include "ggml.h"
+
+template <typename T>
+static __global__ void diag_kernel(T * __restrict__ dst,
+                                   const T * __restrict__ src,
+                                   const int64_t ne0,
+                                   const int64_t ne1,
+                                   const int64_t ne2,
+                                   const int64_t ne3,
+                                   const int64_t total_elements) {
+    const int64_t global_idx = blockIdx.x * blockDim.x + threadIdx.x;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    const int64_t i0 = global_idx % ne0;
+    const int64_t i1 = (global_idx / ne0) % ne1;
+    const int64_t i2 = (global_idx / (ne0 * ne1)) % ne2;
+    const int64_t i3 = global_idx / (ne0 * ne1 * ne2);
+
+    const int64_t dst_idx = ((i3 * ne2 + i2) * ne1 + i1) * ne0 + i0;
+
+    if (i0 == i1) {
+        const int64_t batch_idx = i3 * ne2 + i2;
+        const int64_t src_idx   = batch_idx * ne0 + i0;
+        dst[dst_idx]            = src[src_idx];
+    } else {
+        dst[dst_idx] = ggml_cuda_cast<T>(0);
+    }
+    GGML_UNUSED_VARS(ne3);
+}
+
+void ggml_cuda_op_diag(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    void *       dst_d  = dst->data;
+    const void * src0_d = src0->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+    const int64_t ne2 = dst->ne[2];
+    const int64_t ne3 = dst->ne[3];
+
+    GGML_ASSERT(ne00 == ne0);
+    GGML_ASSERT(ne01 == 1);
+    GGML_ASSERT(ne02 == ne2);
+    GGML_ASSERT(ne03 == ne3);
+
+    const int64_t n_elems    = ggml_nelements(dst);
+    const int64_t num_blocks = (n_elems + CUDA_DIAG_BLOCK_SIZE - 1) / CUDA_DIAG_BLOCK_SIZE;
+
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            diag_kernel<<<num_blocks, CUDA_DIAG_BLOCK_SIZE, 0, stream>>>((float *) dst_d, (const float *) src0_d, ne0,
+                                                                         ne1, ne2, ne3, n_elems);
+            break;
+        case GGML_TYPE_F16:
+            diag_kernel<<<num_blocks, CUDA_DIAG_BLOCK_SIZE, 0, stream>>>((half *) dst_d, (const half *) src0_d, ne0,
+                                                                         ne1, ne2, ne3, n_elems);
+            break;
+        default:
+            GGML_ABORT("unsupported type");
+    }
+}

ml/backend/ggml/ggml/src/ggml-cuda/diag.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_DIAG_BLOCK_SIZE 256
+
+void ggml_cuda_op_diag(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ml/backend/ggml/ggml/src/ggml-cuda/fattn-common.cuh

@@ -10,6 +10,12 @@
 #define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
 #define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
 
+// log(2) = 0.6931, by adding this to the KQ maximum used for the softmax the numerical range representable
+//     by the VKQ accumulators is effectively being shifted up by a factor of 8.
+// This reduces issues with numerical overflow but also causes larger values to be flushed to zero.
+// However, as the output from FlashAttention will usually be used as an input for a matrix multiplication this should be negligible.
+#define FATTN_KQ_MAX_OFFSET 0.6931f
+
 typedef void (* fattn_kernel_t)(
         const char * __restrict__ Q,
         const char * __restrict__ K,
@@ -25,7 +31,7 @@ typedef void (* fattn_kernel_t)(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+        const int32_t ne00, const uint3   ne01, const int32_t ne02, const int32_t ne03,
                             const int32_t nb01, const int32_t nb02, const int32_t nb03,
         const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
                             const int32_t nb11, const int32_t nb12, const int64_t nb13,
@@ -621,7 +627,8 @@ static __global__ void flash_attn_mask_to_KV_max(
 template<int D, int ncols1, int ncols2> // D == head size
 __launch_bounds__(D, 1)
 static __global__ void flash_attn_stream_k_fixup(
-        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03, const int ne11) {
+        float * __restrict__ dst, const float2 * __restrict__ dst_fixup, const int ne01, const int ne02, const int ne03, const int ne11,
+        const int nbatch_fa) {
     constexpr int ncols = ncols1*ncols2;
 
     const int bidx0 = blockIdx.x;
@@ -632,7 +639,7 @@ static __global__ void flash_attn_stream_k_fixup(
 
     const float * dst_fixup_data = ((const float *) dst_fixup) + gridDim.x*(2*2*ncols);
 
-    const int iter_k = ne11 / FATTN_KQ_STRIDE;
+    const int iter_k = (ne11 + (nbatch_fa - 1)) / nbatch_fa;
     const int iter_j = (ne01 + (ncols1    - 1)) / ncols1;
 
     const int kbc0      = (bidx0 + 0)*(iter_k*iter_j*(ne02/ncols2)*ne03) / gridDim.x;
@@ -765,7 +772,7 @@ static __global__ void flash_attn_combine_results(
 template <int DV, int ncols1, int ncols2>
 void launch_fattn(
     ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared,
-    const int KQ_row_granularity, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE
+    const int nbatch_fa, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE
 ) {
     constexpr int ncols = ncols1 * ncols2;
 
@@ -790,8 +797,6 @@ void launch_fattn(
     GGML_ASSERT(!V || V->nb[0] == ggml_element_size(V));
 
     GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
-    GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
-        "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
 
     ggml_cuda_pool & pool = ctx.pool();
     cudaStream_t main_stream = ctx.stream();
@@ -915,7 +920,7 @@ void launch_fattn(
 
         dst_tmp_meta.alloc(blocks_num.x*ncols * (2*2 + DV) * sizeof(float));
     } else {
-        const int ntiles_KQ = (K->ne[1] + KQ_row_granularity - 1) / KQ_row_granularity; // Max. number of parallel blocks limited by tensor size.
+        const int ntiles_KQ = (K->ne[1] + nbatch_fa - 1) / nbatch_fa; // Max. number of parallel blocks limited by tensor size.
 
         // parallel_blocks must not be larger than what the tensor size allows:
         parallel_blocks = std::min(parallel_blocks, ntiles_KQ);
@@ -970,6 +975,9 @@ void launch_fattn(
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
+    // TODO other tensor dimensions after removal of WMMA kernel:
+    const uint3 ne01 = init_fastdiv_values(Q->ne[1]);
+
     GGML_ASSERT(block_dim.x % warp_size == 0);
     fattn_kernel<<<blocks_num, block_dim, nbytes_shared, main_stream>>>(
         (const char *) Q->data,
@@ -980,7 +988,7 @@ void launch_fattn(
         KV_max.ptr,
         !stream_k && parallel_blocks > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
         scale, max_bias, m0, m1, n_head_log2, logit_softcap,
-        Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3], Q->nb[1], Q->nb[2], Q->nb[3],
+        Q->ne[0], ne01,     Q->ne[2], Q->ne[3], Q->nb[1], Q->nb[2], Q->nb[3],
         K->ne[0], K->ne[1], K->ne[2], K->ne[3], nb11, nb12, nb13,
         nb21, nb22, nb23,
         mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0, mask ? mask->ne[3] : 0,
@@ -995,7 +1003,7 @@ void launch_fattn(
 
             flash_attn_stream_k_fixup<DV, ncols1, ncols2>
                 <<<blocks_num_combine, block_dim_combine, 0, main_stream>>>
-                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1]);
+                ((float *) KQV->data, dst_tmp_meta.ptr, Q->ne[1], Q->ne[2], Q->ne[3], K->ne[1], nbatch_fa);
         }
     } else if (parallel_blocks > 1) {
         const dim3 block_dim_combine(DV, 1, 1);

ml/backend/ggml/ggml/src/ggml-cuda/fattn-tile.cuh

@@ -501,6 +501,7 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
         const half2 * const __restrict__ K_h2,
         const half2 * const __restrict__ V_h2,
         const half  * const __restrict__ mask,
+        const uint3 ne01,
         const float logit_softcap,
         const float slope,
         T_KQ      * const KQ,
@@ -512,7 +513,8 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
         float * const KQ_sum,
         T_acc * const VKQ,
         const int k_VKQ_0,
-        const int k_VKQ_max) {
+        const int k_VKQ_max,
+        const int col_Q_0) {
     constexpr int cpy_nb = ggml_cuda_get_max_cpy_bytes();
     constexpr int cpy_ne = cpy_nb / 4;
 
@@ -556,12 +558,18 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
     // Apply logit softcap + mask, update KQ_max:
 #pragma unroll
     for (int jc0 = 0; jc0 < cpw; ++jc0) {
-        const int j = (jc0 + (threadIdx.y / np)*cpw)/ncols2;
+        const int j = fastmodulo(col_Q_0 + (jc0 + (threadIdx.y / np)*cpw)/ncols2, ne01);
 
 #pragma unroll
         for (int i_KQ_0 = 0; i_KQ_0 < nbatch_fa; i_KQ_0 += np*warp_size) {
             const int i_KQ = i_KQ_0 + (threadIdx.y % np)*warp_size + threadIdx.x;
 
+#if defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
+            // Without the v_dot2_f32_f16 instruction there is a higher risk of numerical overflow in the KQ calculation.
+            // Therefore, scale down Q values and apply the inverse scale the FP32 KQ values afterwards again.
+            KQ_acc[i_KQ_0/(np*warp_size)*cpw + jc0] *= 4.0f;
+#endif // defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
+
             if (use_logit_softcap) {
                 KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] = logit_softcap * tanhf(KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0]);
             }
@@ -570,7 +578,7 @@ static __device__ __forceinline__ void flash_attn_tile_iter(
                 KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] += (ncols2 > 1 || mask) ?
                     slope*__half2float(mask[j*stride_mask + k_VKQ_0 + i_KQ]) : 0.0f;
 
-                KQ_max_new[jc0] = fmaxf(KQ_max_new[jc0], KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0]);
+                KQ_max_new[jc0] = fmaxf(KQ_max_new[jc0], KQ_acc[(i_KQ_0/(np*warp_size))*cpw + jc0] + FATTN_KQ_MAX_OFFSET);
             }
         }
 
@@ -736,7 +744,7 @@ static __global__ void flash_attn_tile(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+        const int32_t ne00, const uint3   ne01, const int32_t ne02, const int32_t ne03,
                             const int32_t nb01, const int32_t nb02, const int32_t nb03,
         const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
                             const int32_t nb11, const int32_t nb12, const int64_t nb13,
@@ -781,11 +789,11 @@ static __global__ void flash_attn_tile(
     const int sequence = blockIdx.z / (ne02/ncols2);
     const int head0 = blockIdx.z*ncols2 - sequence*ne02; // == blockIdx.z % (ne02/ncols2)
     const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
-    const float * Q_f  = (const float *) (Q + nb03*sequence + nb02* head0              + nb01*col_Q_0);
+    const float * Q_f  = (const float *) (Q + nb03*sequence + nb02* head0);
     const half2 * K_h2 = (const half2 *) (K + nb13*sequence + nb12*(head0 / gqa_ratio));
     const half2 * V_h2 = (const half2 *) (V + nb23*sequence + nb22*(head0 / gqa_ratio)); // K and V have same shape
 
-    const half * maskh = mask ? (const half *) (mask + nb33*(sequence % ne33) + nb31*col_Q_0) : nullptr;
+    const half * maskh = mask ? (const half *) (mask + nb33*(sequence % ne33)) : nullptr;
 
     const int stride_K2   = nb11 / sizeof(half2);
     const int stride_V2   = nb21 / sizeof(half2);
@@ -842,11 +850,9 @@ static __global__ void flash_attn_tile(
         for (int i0 = 0; i0 < DKQp; i0 += np*warp_size*cpy_ne_D) {
             if (i0 + np*warp_size*cpy_ne_D <= DKQ || i0 + (threadIdx.y % np)*(warp_size*cpy_ne_D) + threadIdx.x*cpy_ne_D < DKQ) {
                 float tmp_f[cpy_ne_D] = {0.0f};
-                if (ncols1 == 1 || col_Q_0 + j < ne01) {
                 ggml_cuda_memcpy_1<sizeof(tmp_f)>
-                        (tmp_f, &Q_f[c*(nb02/sizeof(float)) + j*(nb01/sizeof(float))
+                    (tmp_f, &Q_f[c*(nb02/sizeof(float)) + fastmodulo(col_Q_0 + j, ne01)*(nb01/sizeof(float))
                                  + i0 + (threadIdx.y % np)*(warp_size*cpy_ne_D) + threadIdx.x*cpy_ne_D]);
-                }
 
 #pragma unroll
                 for (int i1 = 0; i1 < cpy_ne_D; ++i1) {
@@ -858,6 +864,11 @@ static __global__ void flash_attn_tile(
 #pragma unroll
                 for (int i1 = 0; i1 < cpy_ne_D; i1 += 2) {
                     tmp_h2[i1/2] = make_half2(tmp_f[i1 + 0], tmp_f[i1 + 1]);
+#if defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
+                    // Without the v_dot2_f32_f16 instruction there is a higher risk of numerical overflow in the KQ calculation.
+                    // Therefore, scale down Q values and apply the inverse scale the FP32 KQ values afterwards again.
+                    tmp_h2[i1/2] *= make_half2(0.25f, 0.25f);
+#endif // defined(FAST_FP16_AVAILABLE) && !defined(V_DOT2_F32_F16_AVAILABLE)
                 }
                 ggml_cuda_memcpy_1<sizeof(tmp_h2)>(
                     &Q_tmp[jc*(DKQ/2) + i0/2 + (threadIdx.y % np)*(warp_size*cpy_ne_D/2) + threadIdx.x*(cpy_ne_D/2)],
@@ -881,23 +892,23 @@ static __global__ void flash_attn_tile(
         while (k_VKQ_0 < k_VKQ_max - nbatch_fa) {
             constexpr bool oob_check = false;
             flash_attn_tile_iter<warp_size, nwarps, ncols1, ncols2, DKQ, DV, nbatch_fa, nbatch_K, use_logit_softcap, oob_check>
-                (Q_tmp, K_h2, V_h2, maskh, logit_softcap, slope, KQ, KV_tmp,
-                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max);
+                (Q_tmp, K_h2, V_h2, maskh, ne01, logit_softcap, slope, KQ, KV_tmp,
+                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max, col_Q_0);
             k_VKQ_0 += gridDim.y*nbatch_fa;
         }
         if (k_VKQ_0 < k_VKQ_max) {
             constexpr bool oob_check = true;
             flash_attn_tile_iter<warp_size, nwarps, ncols1, ncols2, DKQ, DV, nbatch_fa, nbatch_K, use_logit_softcap, oob_check>
-                (Q_tmp, K_h2, V_h2, maskh, logit_softcap, slope, KQ, KV_tmp,
-                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max);
+                (Q_tmp, K_h2, V_h2, maskh, ne01, logit_softcap, slope, KQ, KV_tmp,
+                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max, col_Q_0);
         }
     } else {
         // Branch without out-of-bounds checks.
         for (int k_VKQ_0 = blockIdx.y*nbatch_fa; k_VKQ_0 < k_VKQ_max; k_VKQ_0 += gridDim.y*nbatch_fa) {
             constexpr bool oob_check = false;
             flash_attn_tile_iter<warp_size, nwarps, ncols1, ncols2, DKQ, DV, nbatch_fa, nbatch_K, use_logit_softcap, oob_check>
-                (Q_tmp, K_h2, V_h2, maskh, logit_softcap, slope, KQ, KV_tmp,
-                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max);
+                (Q_tmp, K_h2, V_h2, maskh, ne01, logit_softcap, slope, KQ, KV_tmp,
+                stride_K2, stride_V2, stride_mask, KQ_max, KQ_sum, VKQ, k_VKQ_0, k_VKQ_max, col_Q_0);
         }
     }
 
@@ -1010,13 +1021,13 @@ static __global__ void flash_attn_tile(
         const int j = jc / ncols2;
         const int c = jc % ncols2;
 
-        if (ncols1 > 1 && col_Q_0 + j >= ne01) {
+        if (ncols1 > 1 && col_Q_0 + j >= int(ne01.z)) {
             return;
         }
 
         const float scale = gridDim.y == 1 ? 1.0f/KQ_sum[jc0] : 1.0f;
 
-        const int j_dst_unrolled = ((sequence*ne01 + col_Q_0 + j)*ne02 + head0 + c)*gridDim.y + blockIdx.y;
+        const int j_dst_unrolled = ((sequence*int(ne01.z) + col_Q_0 + j)*ne02 + head0 + c)*gridDim.y + blockIdx.y;
 
 #ifdef FAST_FP16_AVAILABLE
         constexpr int cpy_ne_D = cpy_ne/2 < (DVp/2)/warp_size ? cpy_ne/2 : (DVp/2)/warp_size;

ml/backend/ggml/ggml/src/ggml-cuda/fattn-vec.cuh

@@ -33,7 +33,7 @@ static __global__ void flash_attn_ext_vec(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+        const int32_t ne00, const uint3   ne01, const int32_t ne02, const int32_t ne03,
                             const int32_t nb01, const int32_t nb02, const int32_t nb03,
         const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
                             const int32_t nb11, const int32_t nb12, const int64_t nb13,
@@ -150,7 +150,7 @@ static __global__ void flash_attn_ext_vec(
             float2 * tmp_q_ds  = (float2 *) (tmp_q_i32 + D/sizeof(int));
 
             // Set memory to zero if out of bounds:
-            if (ncols > 1 && ic0 + j >= ne01) {
+            if (ncols > 1 && ic0 + j >= int(ne01.z)) {
 #pragma unroll
                 for (int i0 = 0; i0 < int(D/sizeof(int)); i0 += WARP_SIZE) {
                     const int i = i0 + threadIdx.x;
@@ -201,7 +201,7 @@ static __global__ void flash_attn_ext_vec(
                 const int i = i0 + (nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ)*cpy_ne;
 
                 float2 tmp[cpy_ne] = {{0.0f, 0.0f}};
-                if (ncols == 1 || ic0 + j < ne01) {
+                if (ncols == 1 || ic0 + j < int(ne01.z)) {
                     ggml_cuda_memcpy_1<cpy_nb>(tmp,            &Q_j[i]);
                     ggml_cuda_memcpy_1<cpy_nb>(tmp + cpy_ne/2, &Q_j[i + cpy_ne/2]);
                 }
@@ -222,7 +222,7 @@ static __global__ void flash_attn_ext_vec(
 #pragma unroll
             for (int i0 = 0; i0 < D/2; i0 += nthreads_KQ*cpy_ne) {
                 const int i = i0 + (nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ)*cpy_ne;
-                if (ncols == 1 || ic0 + j < ne01) {
+                if (ncols == 1 || ic0 + j < int(ne01.z)) {
                     ggml_cuda_memcpy_1<cpy_nb>(&Q_reg[j][i0/nthreads_KQ],            &Q_j[i]);
                     ggml_cuda_memcpy_1<cpy_nb>(&Q_reg[j][i0/nthreads_KQ + cpy_ne/2], &Q_j[i + cpy_ne/2]);
                 }
@@ -266,11 +266,11 @@ static __global__ void flash_attn_ext_vec(
                     sum = logit_softcap*tanhf(sum);
                 }
 
-                if (mask) {
+                if (mask && (ncols == 1 || ic0 + j < int(ne01.z))) {
                     sum += slope*__half2float(maskh[j*ne11 + i_KQ]);
                 }
 
-                KQ_max_new[j] = fmaxf(KQ_max_new[j], sum);
+                KQ_max_new[j] = fmaxf(KQ_max_new[j], sum + FATTN_KQ_MAX_OFFSET);
 
                 if ((nthreads_KQ == WARP_SIZE ? threadIdx.x : threadIdx.x % nthreads_KQ) == uint32_t(i_KQ_0)) {
                     KQ_reg[j] = sum;
@@ -412,7 +412,7 @@ static __global__ void flash_attn_ext_vec(
 
 #pragma unroll
     for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
-        if (ncols > 1 && ic0 + j_VKQ >= ne01) {
+        if (ncols > 1 && ic0 + j_VKQ >= int(ne01.z)) {
             break;
         }
 
@@ -479,7 +479,7 @@ static __global__ void flash_attn_ext_vec(
                 if (gridDim.y == 1) {
                     dst_val /= KQ_sum[j_VKQ];
                 }
-                dst[(((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + i0 + tid] = dst_val;
+                dst[(((sequence*int(ne01.z) + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y)*D + i0 + tid] = dst_val;
             }
         }
 
@@ -489,8 +489,8 @@ static __global__ void flash_attn_ext_vec(
 
     }
 
-    if (gridDim.y != 1 && tid < ncols && (ncols == 1 || ic0 + tid < ne01)) {
-        dst_meta[((sequence*ne01 + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(KQ_max[tid], KQ_sum[tid]);
+    if (gridDim.y != 1 && tid < ncols && (ncols == 1 || ic0 + tid < int(ne01.z))) {
+        dst_meta[((sequence*int(ne01.z) + ic0 + tid)*ne02 + head)*gridDim.y + blockIdx.y] = make_float2(KQ_max[tid], KQ_sum[tid]);
     }
 #else
     GGML_UNUSED_VARS(Q, K, V, mask, sinks, KV_max, dst, dst_meta, scale,

ml/backend/ggml/ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -38,14 +38,14 @@ static __global__ void flash_attn_ext_f16(
         const float m1,
         const uint32_t n_head_log2,
         const float logit_softcap,
-        const int32_t ne00, const int32_t ne01, const int32_t ne02, const int32_t ne03,
+        const int32_t ne00, const uint3   ne01, const int32_t ne02, const int32_t ne03,
                             const int32_t nb01, const int32_t nb02, const int32_t nb03,
         const int32_t ne10, const int32_t ne11, const int32_t ne12, const int32_t ne13,
                             const int32_t nb11, const int32_t nb12, const int64_t nb13,
                             const int32_t nb21, const int32_t nb22, const int64_t nb23,
                             const int32_t ne31, const int32_t ne32, const int32_t ne33,
                             const int32_t nb31, const int32_t nb32, const int64_t nb33) {
-#if defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN)))
+#if defined(FLASH_ATTN_AVAILABLE) && (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN))
     // Skip unused kernel variants for faster compilation:
     if (use_logit_softcap && !(D == 128 || D == 256)) {
         NO_DEVICE_CODE;
@@ -149,7 +149,7 @@ static __global__ void flash_attn_ext_f16(
             if (i0 + warp_size > D && i >= D) {
                 break;
             }
-            KQ[j*D_padded + i] = ic0 + j < ne01 ? Q_f[j*stride_Q + i] * scale : 0.0f;
+            KQ[j*D_padded + i] = ic0 + j < int(ne01.z) ? Q_f[j*stride_Q + i] * scale : 0.0f;
         }
     }
 
@@ -218,8 +218,9 @@ static __global__ void flash_attn_ext_f16(
                 for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += warp_size) {
                     const int k = k0 + threadIdx.x;
 
-                    KQ_f_tmp[k0/warp_size] += mask ? __half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f;
-                    KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/warp_size]);
+                    KQ_f_tmp[k0/warp_size] += mask && ic0 + j < int(ne01.z) ?
+                        __half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f;
+                    KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/warp_size] + FATTN_KQ_MAX_OFFSET);
                 }
                 KQ_max_new = warp_reduce_max<warp_size>(KQ_max_new);
 
@@ -270,7 +271,7 @@ static __global__ void flash_attn_ext_f16(
                 for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += warp_size) {
                     const int k = k0 + threadIdx.x;
 
-                    KQ2_tmp[k0/warp_size] += mask ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
+                    KQ2_tmp[k0/warp_size] += mask && ic0 + j < int(ne01.z) ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
                     KQ_max_new = ggml_cuda_hmax2(KQ_max_new, KQ2_tmp[k0/warp_size]);
                 }
                 KQ_max_new = __half2half2(warp_reduce_max<warp_size>(ggml_cuda_hmax(__low2half(KQ_max_new), __high2half(KQ_max_new))));
@@ -431,7 +432,7 @@ static __global__ void flash_attn_ext_f16(
 #pragma unroll
     for (int j0 = 0; j0 < ncols; j0 += nwarps) {
         const int j_VKQ = j0 + threadIdx.y;
-        if (ic0 + j_VKQ >= ne01) {
+        if (ic0 + j_VKQ >= int(ne01.z)) {
             return;
         }
 
@@ -442,7 +443,7 @@ static __global__ void flash_attn_ext_f16(
             KQ_rowsum_j = __low2float(KQ_rowsum_h2[j0/nwarps]) + __high2float(KQ_rowsum_h2[j0/nwarps]);
         }
 
-        const int j_dst_unrolled = ((sequence*ne01 + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
+        const int j_dst_unrolled = ((sequence*int(ne01.z) + ic0 + j_VKQ)*ne02 + head)*gridDim.y + blockIdx.y;
 
 #pragma unroll
         for (int i0 = 0; i0 < D; i0 += warp_size) {
@@ -481,7 +482,7 @@ static __global__ void flash_attn_ext_f16(
               ne31, ne32, ne33,
               nb31, nb32, nb33);
     NO_DEVICE_CODE;
-#endif // defined(FLASH_ATTN_AVAILABLE) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN)))
+#endif // defined(FLASH_ATTN_AVAILABLE) && (defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_USE_WMMA_FATTN))
 }
 
 constexpr int get_max_power_of_2(int x) {

ml/backend/ggml/ggml/src/ggml-cuda/fattn-wmma-f16.cuh

@@ -2,9 +2,9 @@
 
 #include "common.cuh"
 
-#if (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
+#if defined(GGML_USE_MUSA)
 #define GGML_USE_WMMA_FATTN
-#endif // (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
+#endif // defined(GGML_USE_MUSA)
 
 #if defined(GGML_HIP_ROCWMMA_FATTN)
 #if defined(CDNA) && (ROCWMMA_VERSION_MAJOR < 2 || ROCWMMA_VERSION_MINOR > 0 || ROCWMMA_VERSION_PATCH > 0)

ml/backend/ggml/ggml/src/ggml-cuda/fattn.cu

@@ -12,13 +12,13 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
     const ggml_tensor * Q = dst->src[0];
 
     if constexpr (ncols2 <= 8) {
-        if (Q->ne[1] <= 8/ncols2) {
+        if (turing_mma_available(cc) && Q->ne[1] <= 8/ncols2) {
             ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 8/ncols2, ncols2>(ctx, dst);
             return;
         }
     }
 
-    if (Q->ne[1] <= 16/ncols2) {
+    if (turing_mma_available(cc) && Q->ne[1] <= 16/ncols2) {
         ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
         return;
     }
@@ -36,12 +36,26 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols2(ggml_backend_cuda_con
     const ggml_tensor * KQV  = dst;
     const ggml_tensor * Q    = dst->src[0];
     const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * V    = dst->src[2];
     const ggml_tensor * mask = dst->src[3];
 
     float max_bias = 0.0f;
     memcpy(&max_bias, (const float *) KQV->op_params + 1, sizeof(float));
 
-    const bool use_gqa_opt = mask && max_bias == 0.0f;
+    // Edge cases like no mask, ALiBi, unpadded K/V, or misaligned addresses for large data transfers
+    //     are put into the template specialization without GQA optimizations.
+    bool use_gqa_opt = mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    for (const ggml_tensor * t : {Q, K, V, mask}) {
+        if (t == nullptr) {
+            continue;
+        }
+        for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
+            if (t->nb[i] % 16 != 0) {
+                use_gqa_opt = false;
+                break;
+            }
+        }
+    }
 
     GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
     const int gqa_ratio = Q->ne[2] / K->ne[2];
@@ -275,8 +289,8 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
     // For small batch sizes the vector kernel may be preferable over the kernels optimized for large batch sizes:
     const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % 64 == 0 && K->ne[1] % FATTN_KQ_STRIDE == 0;
 
-    // If Turing tensor cores available, use them:
-    if (turing_mma_available(cc) && K->ne[1] % FATTN_KQ_STRIDE == 0 && Q->ne[0] != 40 && Q->ne[0] != 72) {
+    // If Turing tensor cores are available, use them:
+    if (turing_mma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72) {
         if (can_use_vector_kernel) {
             if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
                 if (cc >= GGML_CUDA_CC_ADA_LOVELACE && Q->ne[1] == 1 && Q->ne[3] == 1 && !(gqa_ratio > 4 && K->ne[1] >= 8192)) {
@@ -297,7 +311,21 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
                 return BEST_FATTN_KERNEL_VEC;
             }
         }
+        return BEST_FATTN_KERNEL_MMA_F16;
+    }
 
+    if (volta_mma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72) {
+        int gqa_ratio_eff = 1;
+        const int ncols2_max = Q->ne[0] == 576 ? 16 : 8;
+        while (gqa_ratio % (2*gqa_ratio_eff) == 0 && gqa_ratio_eff < ncols2_max) {
+            gqa_ratio_eff *= 2;
+        }
+        if (can_use_vector_kernel && Q->ne[1] * gqa_ratio_eff <= 2) {
+            return BEST_FATTN_KERNEL_VEC;
+        }
+        if (Q->ne[1] * gqa_ratio_eff <= 16) {
+            return BEST_FATTN_KERNEL_TILE; // On Volta tensor cores are only faster for sufficiently large matrices.
+        }
         return BEST_FATTN_KERNEL_MMA_F16;
     }
 

ml/backend/ggml/ggml/src/ggml-cuda/fill.cu

@@ -0,0 +1,37 @@
+#include "fill.cuh"
+#include "convert.cuh"
+
+#define CUDA_FILL_BLOCK_SIZE 256
+
+template <typename T>
+static __global__ void fill_kernel(T * dst, const int64_t k, const T value) {
+    const int64_t i = (int64_t)blockDim.x * blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = value;
+}
+
+void ggml_cuda_op_fill(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    void * dst_d = dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    float value;
+    memcpy(&value, dst->op_params, sizeof(float));
+
+    const int64_t k = ggml_nelements(dst);
+    const int64_t num_blocks = (k + CUDA_FILL_BLOCK_SIZE - 1) / CUDA_FILL_BLOCK_SIZE;
+
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            fill_kernel<<<num_blocks, CUDA_FILL_BLOCK_SIZE, 0, stream>>>((float *)dst_d, k, value);
+            break;
+        case GGML_TYPE_F16:
+            fill_kernel<<<num_blocks, CUDA_FILL_BLOCK_SIZE, 0, stream>>>((half *)dst_d, k, ggml_cuda_cast<half>(value));
+            break;
+        default:
+            GGML_ABORT("unsupported type");
+    }
+}

ml/backend/ggml/ggml/src/ggml-cuda/fill.cuh

@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_fill(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu

@@ -20,6 +20,7 @@
 #include "ggml-cuda/cpy.cuh"
 #include "ggml-cuda/cross-entropy-loss.cuh"
 #include "ggml-cuda/diagmask.cuh"
+#include "ggml-cuda/diag.cuh"
 #include "ggml-cuda/fattn.cuh"
 #include "ggml-cuda/getrows.cuh"
 #include "ggml-cuda/im2col.cuh"
@@ -54,6 +55,9 @@
 #include "ggml-cuda/set-rows.cuh"
 #include "ggml-cuda/pad_reflect_1d.cuh"
 #include "ggml-cuda/solve_tri.cuh"
+#include "ggml-cuda/tri.cuh"
+#include "ggml-cuda/cumsum.cuh"
+#include "ggml-cuda/fill.cuh"
 #include "ggml.h"
 
 #include <algorithm>
@@ -2772,6 +2776,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
                 break;
+        case GGML_OP_DIAG:
+            ggml_cuda_op_diag(ctx, dst);
+            break;
         case GGML_OP_DIAG_MASK_INF:
             ggml_cuda_op_diag_mask_inf(ctx, dst);
             break;
@@ -2835,6 +2842,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_CROSS_ENTROPY_LOSS:
             ggml_cuda_cross_entropy_loss(ctx, dst);
             break;
+        case GGML_OP_CUMSUM:
+            ggml_cuda_op_cumsum(ctx, dst);
+            break;
+        case GGML_OP_TRI:
+            ggml_cuda_op_tri(ctx, dst);
+            break;
         case GGML_OP_RWKV_WKV6:
             ggml_cuda_op_rwkv_wkv6(ctx, dst);
             break;
@@ -2856,6 +2869,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SOLVE_TRI:
             ggml_cuda_op_solve_tri(ctx, dst);
             break;
+        case GGML_OP_FILL:
+            ggml_cuda_op_fill(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -3383,9 +3399,56 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                 }
             }
             if (should_launch_concurrent_events) {
-                //Restore the original graph to enable fusion within the streams
-                cgraph->nodes   = const_cast<ggml_tensor **>(stream_ctx.original_nodes.data());
-                cgraph->n_nodes = (int) stream_ctx.original_nodes.size();
+                // Restore original node order within each concurrent region to enable fusion within streams
+
+                std::unordered_map<const ggml_tensor *, int> node_to_idx;
+                node_to_idx.reserve(cgraph->n_nodes);
+                for (int i = 0; i < cgraph->n_nodes; ++i) {
+                    node_to_idx[cgraph->nodes[i]] = i;
+                }
+
+                for (auto & [fork_node, event] : stream_ctx.concurrent_events) {
+                    // Find positions of all nodes from this event in the current graph
+                    std::vector<int> positions;
+                    positions.reserve(event.original_order.size());
+
+                    bool all_found = true;
+                    for (const ggml_tensor * orig_node : event.original_order) {
+                        auto it = node_to_idx.find(orig_node);
+                        if (it != node_to_idx.end()) {
+                            positions.push_back(it->second);
+                        } else {
+                            all_found = false;
+                            break;
+                        }
+                    }
+
+                    if (!all_found || positions.size() != event.original_order.size()) {
+                        continue;
+                    }
+
+                    // Sort positions to get contiguous range
+                    std::vector<int> sorted_positions = positions;
+                    std::sort(sorted_positions.begin(), sorted_positions.end());
+
+                    bool is_contiguous = true;
+                    for (size_t i = 1; i < sorted_positions.size(); ++i) {
+                        if (sorted_positions[i] != sorted_positions[i-1] + 1) {
+                            is_contiguous = false;
+                            break;
+                        }
+                    }
+
+                    if (!is_contiguous) {
+                        continue;
+                    }
+
+                    // Restore original order at the sorted positions
+                    int start_pos = sorted_positions[0];
+                    for (size_t i = 0; i < event.original_order.size(); ++i) {
+                        cgraph->nodes[start_pos + i] = const_cast<ggml_tensor *>(event.original_order[i]);
+                    }
+                }
             }
 
             for (int i = 0; i < cgraph->n_nodes; i++) {
@@ -3419,7 +3482,6 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                         GGML_LOG_DEBUG("Setting stream no to %d for node %s\n", cuda_ctx->curr_stream_no, node->name);
                     }
                 }
-                prev_i = i;
 
 #ifdef GGML_CUDA_DEBUG
                 const int nodes_fused = i - prev_i - 1;
@@ -3427,6 +3489,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                     GGML_LOG_INFO("nodes_fused: %d\n", nodes_fused);
                 }
 #endif
+                prev_i = i;
 
                 if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
                     continue;
@@ -4026,14 +4089,6 @@ static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph
     // store {fork_idx, join_idx}
     std::vector<std::pair<int, int>> concurrent_node_ranges;
 
-    // save the original nodes
-    std::vector<const ggml_tensor *> original_nodes;
-    original_nodes.reserve(cgraph->n_nodes);
-    for (int i = 0; i < cgraph->n_nodes; ++i) {
-        original_nodes.push_back(cgraph->nodes[i]);
-    }
-    cuda_ctx->stream_context().original_nodes = std::move(original_nodes);
-
     for (const auto & [root_node, count] : fan_out) {
         if (count >= min_fan_out && count <= max_fan_out) {
             const int root_node_idx = node_indices[root_node];
@@ -4138,6 +4193,13 @@ static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph
                     continue;
                 }
 
+                // Save the original order of nodes in this region before interleaving
+                // This is used later to restore grouping for fusion within streams
+                concurrent_event.original_order.reserve(total_branch_nodes);
+                for (int i = fork_node_idx + 1; i < join_node_idx; ++i) {
+                    concurrent_event.original_order.push_back(cgraph->nodes[i]);
+                }
+
                 std::unordered_map<const ggml_tensor *, ggml_cuda_concurrent_event> & concurrent_events = cuda_ctx->stream_context().concurrent_events;
                 GGML_ASSERT(concurrent_events.find(root_node) == concurrent_events.end());
                 concurrent_events.emplace(root_node, std::move(concurrent_event));
@@ -4841,6 +4903,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:
         case GGML_OP_OPT_STEP_SGD:
+        case GGML_OP_FILL:
+        case GGML_OP_CUMSUM:
+        case GGML_OP_TRI:
+        case GGML_OP_DIAG:
             return true;
         case GGML_OP_SOLVE_TRI:
             return op->src[0]->ne[0] <= 64 && op->src[1]->ne[0] <= 32;

ml/backend/ggml/ggml/src/ggml-cuda/mma.cuh

@@ -68,10 +68,31 @@ static __device__ __forceinline__ half2 ggml_cuda_movmatrix(const half2 x) {
 
 namespace ggml_cuda_mma {
 
+    // Some architectures like Volta or CDNA3 perform multiple matrix multiplications per warp in parallel,
+    //     effectively the warp is being split into subgroups of threads that each perform a single mma instruction.
+    // In those cases the data can be split in different ways across the warp.
+    enum data_layout {
+        // By default the data uses the I direction as its major dimension and the J direction as its minor dimension.
+        // For the A/C matrices this means I major == row major, J major == column major.
+        // For the B matrix this means I major == column major, J major == row major.
+        // MIRRORED == Each data value is held exactly once per thread subgroup.
+        DATA_LAYOUT_I_MAJOR           =  0, // Always used for Turing, Ampere, Ada Lovelace, consumer Blackwell.
+        DATA_LAYOUT_I_MAJOR_MIRRORED  = 10,
+        DATA_LAYOUT_J_MAJOR_MIRRORED  = 20,
+    };
+    // Implemented mma combinations are:
+    //   - (I_MAJOR, I_MAJOR)          -> I_MAJOR
+    //   - (I_MAJOR, I_MAJOR_MIRRORED) -> I_MAJOR
+    //   - (I_MAJOR, J_MAJOR_MIRRORED) -> I_MAJOR
+
+    template <int I_, int J_, typename T, data_layout ds_=DATA_LAYOUT_I_MAJOR>
+    struct tile {};
+
     template <int I_, int J_, typename T>
-    struct tile {
+    struct tile<I_, J_, T, DATA_LAYOUT_I_MAJOR> {
         static constexpr int         I  = I_;
         static constexpr int         J  = J_;
+        static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
 
 #if defined(AMD_MFMA_AVAILABLE)
         static constexpr int ne = I * J / 64;
@@ -131,9 +152,9 @@ namespace ggml_cuda_mma {
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 32 && J == 8) {
 #ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
-                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (l & 2) | (threadIdx.x % 2);
+                return (((threadIdx.x % 16) / 4) * 8) + ((threadIdx.x / 16) * 4) + (l & 2) + (threadIdx.x % 2);
 #else
-                return (l & 2) | (threadIdx.x & ~2);
+                return (l & 2) + (threadIdx.x & ~2);
 #endif // GGML_CUDA_MMA_NO_VOLTA_PERM
             } else {
                 NO_DEVICE_CODE;
@@ -143,7 +164,7 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 32 && J == 8) {
-                return (threadIdx.x & 2) | (l & (4 + 1));
+                return (threadIdx.x & 2) + (l & (4 + 1));
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -152,6 +173,9 @@ namespace ggml_cuda_mma {
 #elif defined(AMD_WMMA_AVAILABLE)
 #if defined(RDNA4)
         static constexpr int ne = I * J / 32;
+#elif defined(RDNA3)
+        static constexpr int ne = (I == 16 && J == 16) ? I * J / 32 : I * J / 16;
+#endif // defined(RDNA4)
         T x[ne] = {0};
 
         static constexpr __device__ bool supported() {
@@ -161,7 +185,11 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_i(const int l) {
             if constexpr (I == 16 && J == 16) {
+#if defined(RDNA4)
                 return 8 * (threadIdx.x / 16) + l;
+#elif defined(RDNA3)
+                return 2 * l + (threadIdx.x / 16);
+#endif // defined(RDNA4)
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -176,7 +204,6 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#endif
 #else
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
@@ -196,9 +223,9 @@ namespace ggml_cuda_mma {
             } else if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 8) {
-                return ((l / 2) * 8) | (threadIdx.x / 4);
+                return ((l / 2) * 8) + (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 16) {
-                return (((l / 2) % 2) * 8) | (threadIdx.x / 4);
+                return (((l / 2) % 2) * 8) + (threadIdx.x / 4);
             } else if constexpr (I == 32 && J == 8) {
                 return tile<16, 8, T>::get_i(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
@@ -211,11 +238,11 @@ namespace ggml_cuda_mma {
             if constexpr (I == 8 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 8 && J == 8) {
-                return (l * 4) | (threadIdx.x % 4);
+                return (l * 4) + (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 8) {
-                return ((threadIdx.x % 4) * 2) | (l % 2);
+                return ((threadIdx.x % 4) * 2) + (l % 2);
             } else if constexpr (I == 16 && J == 16) {
-                return ((l / 4) * 8) | ((threadIdx.x % 4) * 2) | (l % 2);
+                return ((l / 4) * 8) + ((threadIdx.x % 4) * 2) + (l % 2);
             } else if constexpr (I == 32 && J == 8) {
                 return tile<16, 8, T>::get_j(l); // Memory layout simply repeated with same pattern in i direction.
             } else {
@@ -227,26 +254,24 @@ namespace ggml_cuda_mma {
     };
 
     template <int I_, int J_>
-    struct tile<I_, J_, half2> {
+    struct tile<I_, J_, half2, DATA_LAYOUT_I_MAJOR> {
         static constexpr int         I  = I_;
         static constexpr int         J  = J_;
+        static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
 
 #if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
-        static constexpr int ne = I == 8 && J == 8 ? I * J / (WARP_SIZE/4) : I * J / WARP_SIZE;
+        static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
 
         static constexpr __device__ bool supported() {
-            if (I ==  8 && J ==  8) return true;
-            if (I == 32 && J ==  8) return true;
+            if (I == 32 && J ==  4) return true;
             return false;
         }
 
         static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 8 && J == 8) {
-                return ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
-            } else if constexpr (I == 32 && J == 8) {
+            if constexpr (I == 32 && J == 4) {
 #ifdef GGML_CUDA_MMA_NO_VOLTA_PERM
-                return (((threadIdx.x % 16) / 4) * 8) | ((threadIdx.x / 16) * 4) | (threadIdx.x % 4);
+                return (((threadIdx.x % 16) / 4) * 8) + ((threadIdx.x / 16) * 4) + (threadIdx.x % 4);
 #else
                 return threadIdx.x;
 #endif // GGML_CUDA_MMA_NO_VOLTA_PERM
@@ -257,7 +282,7 @@ namespace ggml_cuda_mma {
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
-            if constexpr ((I == 8 || I == 32) && J == 8) {
+            if constexpr (I == 32 && J == 4) {
                 return l;
             } else {
                 NO_DEVICE_CODE;
@@ -265,6 +290,7 @@ namespace ggml_cuda_mma {
             }
         }
 #elif defined(AMD_WMMA_AVAILABLE)
+
         static constexpr int ne = I * J / 32;
         half2 x[ne] = {{0.0f, 0.0f}};
 
@@ -307,11 +333,11 @@ namespace ggml_cuda_mma {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return (l * 8) | (threadIdx.x / 4);
+                return (l * 8) + (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return ((l % 2) * 8) | (threadIdx.x / 4);
+                return ((l % 2) * 8) + (threadIdx.x / 4);
             } else if constexpr (I == 32 && J == 8) {
-                return ((l / 4) * 16) | ((l % 2) * 8) | (threadIdx.x / 4);
+                return ((l / 4) * 16) + ((l % 2) * 8) + (threadIdx.x / 4);
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -320,13 +346,13 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return (l * 4) | (threadIdx.x % 4);
+                return (l * 4) + (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return ((l / 2) * 4) | (threadIdx.x % 4);
+                return ((l / 2) * 4) + (threadIdx.x % 4);
             } else if constexpr (I == 32 && J == 8) {
-                return ((l & 2) * 2) | (threadIdx.x % 4);
+                return ((l & 2) * 2) + (threadIdx.x % 4);
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -336,14 +362,15 @@ namespace ggml_cuda_mma {
     };
 
     template <int I_, int J_>
-    struct tile<I_, J_, nv_bfloat162> {
+    struct tile<I_, J_, nv_bfloat162, DATA_LAYOUT_I_MAJOR> {
         static constexpr int         I  = I_;
         static constexpr int         J  = J_;
+        static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
+        static constexpr int         ne = I * J / WARP_SIZE;
 
-#if defined(AMD_WMMA_AVAILABLE)
-        static constexpr int ne = I * J / 32;
         nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
 
+#if defined(AMD_WMMA_AVAILABLE)
         static constexpr __device__ bool supported() {
             if (I == 16 && J == 8) return true;
             return false;
@@ -367,9 +394,6 @@ namespace ggml_cuda_mma {
             }
         }
 #else
-        static constexpr int ne = I * J / WARP_SIZE;
-        nv_bfloat162 x[ne] = {{0.0f, 0.0f}};
-
         static constexpr __device__ bool supported() {
             if (I ==  8 && J ==  8) return true;
             if (I == 16 && J ==  4) return true;
@@ -381,9 +405,9 @@ namespace ggml_cuda_mma {
             if constexpr (I == 8 && J == 8) {
                 return threadIdx.x / 4;
             } else if constexpr (I == 16 && J == 4) {
-                return (l * 8) | (threadIdx.x / 4);
+                return (l * 8) + (threadIdx.x / 4);
             } else if constexpr (I == 16 && J == 8) {
-                return ((l % 2) * 8) | (threadIdx.x / 4);
+                return ((l % 2) * 8) + (threadIdx.x / 4);
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -392,11 +416,11 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 8 && J == 8) {
-                return (l * 4) | (threadIdx.x % 4);
+                return (l * 4) + (threadIdx.x % 4);
             } else if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 4;
             } else if constexpr (I == 16 && J == 8) {
-                return ((l / 2) * 4) | (threadIdx.x % 4);
+                return ((l / 2) * 4) + (threadIdx.x % 4);
             } else {
                 NO_DEVICE_CODE;
                 return -1;
@@ -405,6 +429,73 @@ namespace ggml_cuda_mma {
 #endif  // defined(AMD_WMMA_AVAILABLE)
     };
 
+    template <int I_, int J_>
+    struct tile<I_, J_, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> {
+        static constexpr int         I  = I_;
+        static constexpr int         J  = J_;
+        static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR_MIRRORED;
+        static constexpr int         ne = I * J / (WARP_SIZE/4);
+
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int /*l*/) {
+            if constexpr (I == 8 && J == 4) {
+                return ((threadIdx.x / 16) * 4) + (threadIdx.x % 4);
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 8 && J == 4) {
+                return l;
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+    };
+
+    template <int I_, int J_>
+    struct tile<I_, J_, half2, DATA_LAYOUT_J_MAJOR_MIRRORED> {
+        static constexpr int         I  = I_;
+        static constexpr int         J  = J_;
+        static constexpr data_layout dl = DATA_LAYOUT_J_MAJOR_MIRRORED;
+        static constexpr int         ne = I * J / (WARP_SIZE/4);
+
+        half2 x[ne] = {{0.0f, 0.0f}};
+
+        static constexpr __device__ bool supported() {
+            if (I ==  8 && J ==  4) return true;
+            return false;
+        }
+
+        static __device__ __forceinline__ int get_i(const int l) {
+            if constexpr (I == 8 && J == 4) {
+                return ((l / 2) * 4) + (threadIdx.x % 4);
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+
+        static __device__ __forceinline__ int get_j(const int l) {
+            if constexpr (I == 8 && J == 4) {
+                return ((threadIdx.x / 16) * 2) + (l % 2);
+            } else {
+                NO_DEVICE_CODE;
+                return -1;
+            }
+        }
+    };
+
+#if defined(TURING_MMA_AVAILABLE)
     template <int I, int J>
     static __device__ __forceinline__ tile<I, J/2, half2> get_half2(const tile<I, J, float> & tile_float) {
         tile<I, J/2, half2> ret;
@@ -422,9 +513,26 @@ namespace ggml_cuda_mma {
 
         return ret;
     }
+#else // Volta
+    template <int I, int J>
+    static __device__ __forceinline__ tile<I, J/2, half2> get_half2(const tile<I, J, float> & tile_float) {
+        tile<I, J/2, half2> ret;
+#pragma unroll
+        for (int l0 = 0; l0 < tile_float.ne; l0 += 4) {
+            ret.x[l0/2 + 0] = make_half2(tile_float.x[l0 + 0], tile_float.x[l0 + 1]);
+            ret.x[l0/2 + 1] = make_half2(tile_float.x[l0 + 2], tile_float.x[l0 + 3]);
 
-    template <int I, int J, typename T>
-    static __device__ __forceinline__ void load_generic(tile<I, J, T> & t, const T * __restrict__ xs0, const int stride) {
+            // On Volta FP16 and FP32 tiles have a different memory layout,
+            //     for the conversion threads with an offset of 2 need to exchange half their values:
+            ret.x[l0/2 + (((threadIdx.x % 4) / 2) ^ 1)] = __shfl_xor_sync(
+                0xFFFFFFFF, ret.x[l0/2 + (((threadIdx.x % 4) / 2) ^ 1)], 2, WARP_SIZE);
+        }
+        return ret;
+    }
+#endif // defined(TURING_MMA_AVAILABLE)
+
+    template <int I, int J, typename T, data_layout dl>
+    static __device__ __forceinline__ void load_generic(tile<I, J, T, dl> & t, const T * __restrict__ xs0, const int stride) {
 #if defined(AMD_MFMA_AVAILABLE)
         if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
 #pragma unroll
@@ -443,18 +551,34 @@ namespace ggml_cuda_mma {
         } else if constexpr (std::is_same_v<T, int>) {
             if constexpr (I == 16 && J == 4) {
                 int64_t * xi = (int64_t *) t.x;
+#if defined(RDNA4)
                 const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 2 * (threadIdx.x / t.I));
                 xi[0] = xs[0];
-
-            }else if constexpr (I == 16 && J == 8) {
+#elif defined(RDNA3)
+                static_assert(tile<I,J,T>::ne >= 4, "fragment too small");
+                const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride);
+                xi[0] = xs[0];
+                xi[1] = xs[1];
+#endif // defined(RDNA4)
+            } else if constexpr (I == 16 && J == 8) {
                 int64_t * xi = (int64_t *) t.x;
+#if defined(RDNA4)
                 const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 4 * (threadIdx.x / t.I));
                 xi[0] = xs[0];
 
                 const int64_t * xs1 = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride + 4 * (threadIdx.x / t.I) + 2);
                 xi[1] = xs1[0];
-
-            }else{
+#elif defined(RDNA3)
+                static_assert(tile<I,J,T>::ne >= 8, "fragment too small");
+                const int64_t * xs = (int64_t *) ((const int *) xs0 + (threadIdx.x % t.I) * stride);
+                // contiguous four 64-bit chunks per lane for the wider RDNA3 fragment
+                xi[0] = xs[0];
+                xi[1] = xs[1];
+                const int64_t * xs1 = xs + 2;
+                xi[2] = xs1[0];
+                xi[3] = xs1[1];
+#endif // defined(RDNA4)
+            } else {
                 NO_DEVICE_CODE;
             }
         } else {
@@ -511,18 +635,6 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
             : "l"(xs));
 #else
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
-        GGML_UNUSED_VARS(t, xs0, stride);
-        NO_DEVICE_CODE;
-#else
-        load_generic(t, xs0, stride);
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
-#endif // TURING_MMA_AVAILABLE
-    }
-
-    template <typename T>
-    static __device__ __forceinline__ void load_ldmatrix(
-            tile<32, 8, T> & t, const T * __restrict__ xs0, const int stride) {
 #if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #if 1
         // TODO: more generic handling
@@ -533,9 +645,31 @@ namespace ggml_cuda_mma {
         load_generic(t, xs0, stride);
 #endif // 1
 #else
-        tile<16, 8, T> * t16 = (tile<16, 8, T> *) &t;
-        load_ldmatrix(t16[0], xs0 +  0*stride, stride);
-        load_ldmatrix(t16[1], xs0 + 16*stride, stride);
+        load_generic(t, xs0, stride);
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // TURING_MMA_AVAILABLE
+    }
+
+    static __device__ __forceinline__ void load_ldmatrix(
+            tile<8, 4, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & t, const half2 * __restrict__ xs0, const int stride) {
+        ggml_cuda_memcpy_1<4*sizeof(half2)>(t.x, xs0 + t.get_i(0)*stride);
+    }
+
+    static __device__ __forceinline__ void load_ldmatrix(
+            tile<8, 4, half2, DATA_LAYOUT_J_MAJOR_MIRRORED> & t, const half2 * __restrict__ xs0, const int stride) {
+#pragma unroll
+        for (int l0 = 0; l0 < t.ne; l0 += 2) {
+            ggml_cuda_memcpy_1<2*sizeof(half2)>(t.x + l0, xs0 + t.get_i(l0)*stride + t.get_j(l0));
+        }
+    }
+
+    static __device__ __forceinline__ void load_ldmatrix(
+            tile<32, 4, half2> & t, const half2 * __restrict__ xs0, const int stride) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        ggml_cuda_memcpy_1<4*sizeof(half2)>(t.x, xs0 + t.get_i(0)*stride);
+#else
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
     }
 
@@ -747,12 +881,14 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #elif defined(AMD_WMMA_AVAILABLE)
+#if defined(RDNA4)
         using halfx8_t = __attribute__((ext_vector_type(8))) _Float16;
         using floatx8_t = __attribute__((ext_vector_type(8))) float;
         floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
         const halfx8_t& a_frag = reinterpret_cast<const halfx8_t&>(A.x[0]);
         const halfx8_t& b_frag = reinterpret_cast<const halfx8_t&>(B.x[0]);
         acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32_gfx12(a_frag, b_frag, acc_frag);
+#endif // RDNA4
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -762,12 +898,14 @@ namespace ggml_cuda_mma {
     static __device__ __forceinline__ void mma(
             tile<16, 16, float> & D, const tile<16, 8, nv_bfloat162> & A, const tile<16, 8, nv_bfloat162> & B) {
 #if defined(AMD_WMMA_AVAILABLE)
+#if defined(RDNA4)
         using bf16x8_t = __attribute__((ext_vector_type(8))) __bf16;
         using floatx8_t = __attribute__((ext_vector_type(8))) float;
         floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
         const bf16x8_t& a_frag = reinterpret_cast<const bf16x8_t&>(A.x[0]);
         const bf16x8_t& b_frag = reinterpret_cast<const bf16x8_t&>(B.x[0]);
         acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12(a_frag, b_frag, acc_frag);
+#endif // RDNA4
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -796,14 +934,14 @@ namespace ggml_cuda_mma {
 #endif // defined(CDNA3)
 
 #elif defined(AMD_WMMA_AVAILABLE)
-        using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
-        int32x2_t * a_vec = (int32x2_t *) A.x;
-        int32x2_t * b_vec = (int32x2_t *) B.x;
 
         using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
         int32x8_t * acc = (int32x8_t *) D.x;
 
 #if defined(RDNA4)
+        using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
+        int32x2_t * a_vec = (int32x2_t *) A.x;
+        int32x2_t * b_vec = (int32x2_t *) B.x;
 
         acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
             true,
@@ -822,7 +960,30 @@ namespace ggml_cuda_mma {
             acc[0],
             true
         );
-#endif // defined(RDNA4)
+
+#elif defined(RDNA3)
+        using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
+        int32x4_t * a_vec = (int32x4_t *) A.x;
+        int32x4_t * b_vec = (int32x4_t *) B.x;
+
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
+            true,
+            a_vec[0],
+            true,
+            b_vec[0],
+            acc[0],
+            true
+        );
+
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
+            true,
+            a_vec[1],
+            true,
+            b_vec[1],
+            acc[0],
+            true
+        );
+#endif // RDNA4
 
 #else
         GGML_UNUSED_VARS(D, A, B);
@@ -860,14 +1021,14 @@ namespace ggml_cuda_mma {
     template <typename T1, typename T2, int J, int K>
     static __device__ __forceinline__ void mma(
             tile<32, J, T1> & D, const tile<32, K, T2> & A, const tile<J, K, T2> & B) {
-        tile<16, J, T1> * D16 = (tile<16, J, T1> *) &D;
-        tile<16, K, T2> * A16 = (tile<16, K, T2> *) &A;
+        tile      <16, J, T1> * D16 = reinterpret_cast<      tile<16, J, T1> *>(&D);
+        const tile<16, K, T2> * A16 = reinterpret_cast<const tile<16, K, T2> *>(&A);
         mma(D16[0], A16[0], B);
         mma(D16[1], A16[1], B);
     }
 
     static __device__ __forceinline__ void mma(
-            tile<32, 8, float> & D, const tile<32, 8, half2> & A, const tile<8, 8, half2> & B) {
+            tile<32, 8, float> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & B) {
 #if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
@@ -880,32 +1041,42 @@ namespace ggml_cuda_mma {
             "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
             : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]), "r"(Bxi[3]));
-        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
-            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
-            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
-            : "r"(Axi[4]), "r"(Axi[5]), "r"(Bxi[4]), "r"(Bxi[5]));
-        asm("mma.sync.aligned.m8n8k4.row.col.f32.f16.f16.f32 "
-            "{%0, %1, %2, %3, %4, %5, %6, %7}, {%8, %9}, {%10, %11}, {%0, %1, %2, %3, %4, %5, %6, %7};"
-            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3]), "+r"(Dxi[4]), "+r"(Dxi[5]), "+r"(Dxi[6]), "+r"(Dxi[7])
-            : "r"(Axi[6]), "r"(Axi[7]), "r"(Bxi[6]), "r"(Bxi[7]));
 #else
-        tile      <16, 8, float> * D16 = reinterpret_cast<tile      <16, 8, float> *>(&D);
-        const tile<16, 8, half2> * A16 = reinterpret_cast<const tile<16, 8, half2> *>(&A);
-        mma(D16[0], A16[0], B);
-        mma(D16[1], A16[1], B);
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+    }
+
+    static __device__ __forceinline__ void mma(
+            tile<32, 4, half2> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_J_MAJOR_MIRRORED> & B) {
+#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        const int * Axi = (const int *) A.x;
+        const int * Bxi = (const int *) B.x;
+        int       * Dxi = (int       *) D.x;
+        asm("mma.sync.aligned.m8n8k4.row.row.f16.f16.f16.f16 "
+            "{%0, %1, %2, %3}, {%4, %5}, {%6, %7}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[0]), "r"(Axi[1]), "r"(Bxi[0]), "r"(Bxi[1]));
+        asm("mma.sync.aligned.m8n8k4.row.row.f16.f16.f16.f16 "
+            "{%0, %1, %2, %3}, {%4, %5}, {%6, %7}, {%0, %1, %2, %3};"
+            : "+r"(Dxi[0]), "+r"(Dxi[1]), "+r"(Dxi[2]), "+r"(Dxi[3])
+            : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[2]), "r"(Bxi[3]));
+#else
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
     }
 
 static __device__ __forceinline__ void mma(
             tile<16, 16, int> & D, const tile<16, 4, int> & A, const tile<16, 4, int> & B) {
 #if defined(AMD_WMMA_AVAILABLE)
+        using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
+        int32x8_t * acc = (int32x8_t *) D.x;
+#if defined(RDNA4)
         using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
         int32x2_t * a_vec = (int32x2_t *) A.x;
         int32x2_t * b_vec = (int32x2_t *) B.x;
 
-    using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
-    int32x8_t * acc = (int32x8_t *) D.x;
-
         acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
             true,
             a_vec[0],
@@ -914,12 +1085,25 @@ static __device__ __forceinline__ void mma(
             acc[0],
             false
         );
+#elif defined(RDNA3)
+        using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
+        int32x4_t * a_vec = (int32x4_t *) A.x;
+        int32x4_t * b_vec = (int32x4_t *) B.x;
+
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
+            true,
+            a_vec[0],
+            true,
+            b_vec[0],
+            acc[0],
+            false
+        );
+#endif // RDNA4
 #else
         GGML_UNUSED(D);
         GGML_UNUSED(A);
         GGML_UNUSED(B);
         NO_DEVICE_CODE;
-#endif
+#endif // AMD_WMMA_AVAILABLE
     }
 }
-

ml/backend/ggml/ggml/src/ggml-cuda/mmf.cu

@@ -160,9 +160,9 @@ bool ggml_cuda_should_use_mmf(enum ggml_type type, int cc, int warp_size, const
         case GGML_TYPE_F32:
             return ampere_mma_available(cc);
         case GGML_TYPE_F16:
-            return volta_mma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc);
+            return volta_mma_available(cc) || turing_mma_available(cc) || (amd_wmma_available(cc) && GGML_CUDA_CC_IS_RDNA4(cc));
         case GGML_TYPE_BF16:
-            return ampere_mma_available(cc) || amd_wmma_available(cc);
+            return ampere_mma_available(cc) || (amd_wmma_available(cc) && GGML_CUDA_CC_IS_RDNA4(cc));
         default:
             return false;
     }

ml/backend/ggml/ggml/src/ggml-cuda/mmf.cuh

@@ -37,23 +37,19 @@ static __global__ void mul_mat_f(
     typedef tile<16,       8, T>     tile_A;
     typedef tile<tile_B_I, 8, T>     tile_B;
     typedef tile<16,       tile_C_J, float> tile_C;
-
-    constexpr bool a_supported = tile_A::supported();
-    constexpr bool b_supported = tile_B::supported();
-    constexpr bool c_supported = tile_C::supported();
-    constexpr bool supported = a_supported && b_supported && c_supported;
 #else
-    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
-    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
-    constexpr bool supported = I_16_supported || I_32_supported;
-
-    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work but 16 is ~1% faster.
-
-    typedef tile<I_preferred, 8, T>     tile_A;
+#ifdef VOLTA_MMA_AVAILABLE
+    if constexpr (!std::is_same_v<T, half2>) {NO_DEVICE_CODE;} else {
+    typedef tile<32, 4, T,     DATA_LAYOUT_I_MAJOR>          tile_A;
+    typedef tile< 8, 4, T,     DATA_LAYOUT_I_MAJOR_MIRRORED> tile_B;
+    typedef tile<32, 8, float, DATA_LAYOUT_I_MAJOR>          tile_C;
+#else
+    typedef tile<16, 8, T>     tile_A;
     typedef tile<8,  8, T>     tile_B;
-    typedef tile<I_preferred, 8, float> tile_C;
+    typedef tile<16, 8, float> tile_C;
+#endif // VOLTA_MMA_AVAILABLE
 #endif // defined(AMD_WMMA_AVAILABLE)
-    if constexpr (!supported) {
+    if constexpr (!tile_A::supported() || !tile_B::supported() || !tile_C::supported()) {
         NO_DEVICE_CODE;
         return;
     }
@@ -248,6 +244,9 @@ static __global__ void mul_mat_f(
             }
         }
     }
+#ifdef VOLTA_MMA_AVAILABLE
+    }
+#endif //VOLTA_MMA_AVAILABLE
 #else
     GGML_UNUSED_VARS(x, y, ids, dst,
         ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst,
@@ -278,27 +277,24 @@ static __global__ void mul_mat_f_ids(
     typedef tile<16,       8, T>     tile_A;
     typedef tile<tile_B_I, 8, T>     tile_B;
     typedef tile<16,       tile_C_J, float> tile_C;
-
-    constexpr bool a_supported = tile_A::supported();
-    constexpr bool b_supported = tile_B::supported();
-    constexpr bool c_supported = tile_C::supported();
-    constexpr bool supported = a_supported && b_supported && c_supported;
 #else
-    constexpr bool I_16_supported = tile<16, 8, T>::supported() && tile<16, 8, float>::supported();
-    constexpr bool I_32_supported = tile<32, 8, T>::supported() && tile<32, 8, float>::supported();
-    constexpr bool supported = I_16_supported || I_32_supported;
-
-    constexpr int I_preferred = I_16_supported ? 16 : 32; // For Turing MMA both work but 16 is ~1% faster.
-
-    typedef tile<I_preferred, 8, T>     tile_A;
+#ifdef VOLTA_MMA_AVAILABLE
+    if constexpr (!std::is_same_v<T, half2>) {NO_DEVICE_CODE;} else {
+    typedef tile<32, 4, T,     DATA_LAYOUT_I_MAJOR>          tile_A;
+    typedef tile< 8, 4, T,     DATA_LAYOUT_I_MAJOR_MIRRORED> tile_B;
+    typedef tile<32, 8, float, DATA_LAYOUT_I_MAJOR>          tile_C;
+#else
+    typedef tile<16, 8, T>     tile_A;
     typedef tile<8,  8, T>     tile_B;
-    typedef tile<I_preferred, 8, float> tile_C;
+    typedef tile<16, 8, float> tile_C;
+#endif // VOLTA_MMA_AVAILABLE
 #endif // defined(AMD_WMMA_AVAILABLE)
-    if constexpr (!supported) {
+    if constexpr (!tile_A::supported() || !tile_B::supported() || !tile_C::supported()) {
         NO_DEVICE_CODE;
         return;
     }
 
+
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int tile_k_padded = warp_size + 4;
     constexpr int ntA = rows_per_block / tile_A::I;
@@ -517,6 +513,9 @@ static __global__ void mul_mat_f_ids(
             }
         }
     }
+#ifdef VOLTA_MMA_AVAILABLE
+    }
+#endif // VOLTA_MMA_AVAILABLE
 #else
     GGML_UNUSED_VARS(x, y, ids_src_compact, ids_dst_compact, expert_bounds, dst,
         ncols, ncols_dst_total, nchannels_dst, stride_row, stride_col_y, stride_col_dst,

ml/backend/ggml/ggml/src/ggml-cuda/mmq.cu

@@ -307,10 +307,9 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
     }
 
     if (amd_wmma_available(cc)) {
-        if (GGML_CUDA_CC_IS_RDNA4(cc)) {
         return true;
     }
-    }
 
-    return (!GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    return (!GGML_CUDA_CC_IS_CDNA(cc)) || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+
 }

ml/backend/ggml/ggml/src/ggml-cuda/mmq.cuh

@@ -1542,8 +1542,10 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
             tile_C Cm;
             if (k01 >= MMQ_TILE_NE_K * 3/4) {
                 tile_A A1;
-                A1.x[0] = 0x01010101;
-                A1.x[1] = 0x01010101;
+#pragma unroll
+                for (int l = 0; l < tile_A::ne; ++l) {
+                    A1.x[l] = 0x01010101;
+                }
                 mma(Cm, A1, B);
             }
 

ml/backend/ggml/ggml/src/ggml-cuda/pad.cu

@@ -1,9 +1,17 @@
 #include "pad.cuh"
 
+#include <stdint.h>
+
+__device__ __forceinline__ int64_t wrap_around(int64_t coord, int64_t size) {
+    // + size ensures negatives are handled properly
+    return (coord + size) % size;
+}
+
 static __global__ void pad_f32(const float * src, float * dst,
                                const int lp0, const int rp0, const int lp1, const int rp1,
                                const int lp2, const int rp2, const int lp3, const int rp3,
-                               const int ne0, const int ne1, const int ne2, const int ne3) {
+                               const int ne0, const int ne1, const int ne2, const int ne3,
+                               const bool circular) {
     // blockIdx.z: i3*ne2+i2
     // blockIdx.y: i1
     // blockIDx.x: i0 / CUDA_PAD_BLOCK_SIZE
@@ -12,15 +20,15 @@ static __global__ void pad_f32(const float * src, float * dst,
     int i1 = blockIdx.y;
     int i2 = blockIdx.z % ne2;
     int i3 = blockIdx.z / ne2;
+
     if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
         return;
     }
 
-    // operation
-    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
-    if ((i0 >= lp0 && i0 < ne0 - rp0) &&
-        (i1 >= lp1 && i1 < ne1 - rp1) &&
-        (i2 >= lp2 && i2 < ne2 - rp2) &&
+    const int64_t dst_idx = i3 * (ne0 * ne1 * ne2) + i2 * (ne0 * ne1) + i1 * ne0 + i0;
+
+    if (!circular) {
+        if ((i0 >= lp0 && i0 < ne0 - rp0) && (i1 >= lp1 && i1 < ne1 - rp1) && (i2 >= lp2 && i2 < ne2 - rp2) &&
             (i3 >= lp3 && i3 < ne3 - rp3)) {
             const int64_t i00  = i0 - lp0;
             const int64_t i01  = i1 - lp1;
@@ -30,43 +38,66 @@ static __global__ void pad_f32(const float * src, float * dst,
             const int64_t ne01 = ne1 - lp1 - rp1;
             const int64_t ne00 = ne0 - lp0 - rp0;
 
-        const int64_t src_idx = i03*(ne00*ne01*ne02) + i02*(ne00*ne01) + i01*ne00 + i00;
+            const int64_t src_idx = i03 * (ne00 * ne01 * ne02) + i02 * (ne00 * ne01) + i01 * ne00 + i00;
 
             dst[dst_idx] = src[src_idx];
         } else {
             dst[dst_idx] = 0.0f;
         }
+    }
+    // circular means on a torus, so x and y wrap around
+    else {
+        const int64_t ne00 = ne0 - lp0 - rp0;
+        const int64_t ne01 = ne1 - lp1 - rp1;
+        const int64_t ne02 = ne2 - lp2 - rp2;
+        const int64_t ne03 = ne3 - lp3 - rp3;
+
+        const int64_t i00 = wrap_around(i0 - lp0, ne00);
+        const int64_t i01 = wrap_around(i1 - lp1, ne01);
+        const int64_t i02 = wrap_around(i2 - lp2, ne02);
+        const int64_t i03 = wrap_around(i3 - lp3, ne03);
+
+        const int64_t src_idx = i03 * (ne00 * ne01 * ne02) + i02 * (ne00 * ne01) + i01 * ne00 + i00;
+
+        dst[dst_idx] = src[src_idx];
+    }
 }
 
+
 static void pad_f32_cuda(const float * src, float * dst,
     const int lp0, const int rp0, const int lp1, const int rp1,
     const int lp2, const int rp2, const int lp3, const int rp3,
-    const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
+    const int ne0, const int ne1, const int ne2, const int ne3,
+    const bool circular, cudaStream_t stream) {
     int  num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, ne1, ne2*ne3);
-    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(src, dst, lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3, ne0, ne1, ne2, ne3);
+    dim3 gridDim(num_blocks, ne1, ne2 * ne3);
+    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(src, dst,
+                                                         lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
+                                                         ne0, ne1, ne2, ne3, circular);
 }
 
 void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0   = dst->src[0];
-    const float * src0_d = (const float *)src0->data;
-    float * dst_d = (float *)dst->data;
+    const float *       src0_d = (const float *) src0->data;
+    float *             dst_d  = (float *) dst->data;
     cudaStream_t        stream = ctx.stream();
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(dst->type == GGML_TYPE_F32);
     GGML_ASSERT(ggml_is_contiguous(src0));
 
-    const int32_t lp0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t rp0 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t lp1 = ((const int32_t*)(dst->op_params))[2];
-    const int32_t rp1 = ((const int32_t*)(dst->op_params))[3];
-    const int32_t lp2 = ((const int32_t*)(dst->op_params))[4];
-    const int32_t rp2 = ((const int32_t*)(dst->op_params))[5];
-    const int32_t lp3 = ((const int32_t*)(dst->op_params))[6];
-    const int32_t rp3 = ((const int32_t*)(dst->op_params))[7];
+    const int32_t lp0      = ((const int32_t *) (dst->op_params))[0];
+    const int32_t rp0      = ((const int32_t *) (dst->op_params))[1];
+    const int32_t lp1      = ((const int32_t *) (dst->op_params))[2];
+    const int32_t rp1      = ((const int32_t *) (dst->op_params))[3];
+    const int32_t lp2      = ((const int32_t *) (dst->op_params))[4];
+    const int32_t rp2      = ((const int32_t *) (dst->op_params))[5];
+    const int32_t lp3      = ((const int32_t *) (dst->op_params))[6];
+    const int32_t rp3      = ((const int32_t *) (dst->op_params))[7];
+    const int32_t circular = ((const int32_t *) (dst->op_params))[8];
 
     pad_f32_cuda(src0_d, dst_d,
                  lp0, rp0, lp1, rp1, lp2, rp2, lp3, rp3,
-                 dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+                 dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+                 (bool) circular, stream);
 }

ml/backend/ggml/ggml/src/ggml-cuda/solve_tri.cu

@@ -3,7 +3,6 @@
 #include "solve_tri.cuh"
 
 #define MAX_N_FAST 64
-#define MAX_K_FAST 32
 
 // ======================
 // Fast Kernel (n <= 64, k <= 32) - Warp-based parallel reduction
@@ -48,65 +47,58 @@ static __global__ void solve_tri_f32_fast(const float * __restrict__ A,
     float *             X_batch = (float *) (X + i02 * nb2 + i03 * nb3);
 
     __shared__ float sA[MAX_N_FAST * MAX_N_FAST];
-    __shared__ float sXt[MAX_N_FAST * (MAX_K_FAST + 1)];
 
     const int offset = threadIdx.x + threadIdx.y * blockDim.x;
 
 #pragma unroll
     for (int i = 0; i < n * n; i += k * WARP_SIZE) {
-        int i0 = i + offset;
+        const int i0 = i + offset;
         if (i0 < n * n) {
             sA[i0] = A_batch[i0];
         }
     }
 
-    const int rows_per_warp = (n + WARP_SIZE - 1) / WARP_SIZE;
-
-#pragma unroll
-    for (int i = 0; i < rows_per_warp; i++) {
-        const int i0 = lane + i * WARP_SIZE;
-        if (i0 < n) {
-            sXt[col_idx * n + i0] = B_batch[i0 * k + col_idx];
-        }
-    }
-
     __syncthreads();
 
+    float x_low  = (lane < n) ? B_batch[lane * k + col_idx] : 0.0f;
+    float x_high = (WARP_SIZE + lane < n) ? B_batch[(WARP_SIZE + lane) * k + col_idx] : 0.0f;
+
+    const int half = WARP_SIZE;
+    const int nrows_low = (n < half) ? n : half;
+
 #pragma unroll
-    for (int row = 0; row < n; ++row) {
+    for (int row = 0; row < nrows_low; ++row) {
         float sum = 0.0f;
-
-        {
-            int j = lane;
-            if (j < row) {
-                sum += sA[row * n + j] * sXt[col_idx * n + j];
+        if (lane < row) {
+            sum += sA[row * n + lane] * x_low;
         }
-        }
-        if (row >= WARP_SIZE) {
-            int j = WARP_SIZE + lane;
-            if (j < row) {
-                sum += sA[row * n + j] * sXt[col_idx * n + j];
-            }
-        }
-
         sum = warp_reduce_sum(sum);
 
-        if (lane == 0) {
-            const float b_val      = sXt[col_idx * n + row];
-            const float a_diag     = sA[row * n + row];
-            // no safeguards for division by zero because that indicates corrupt
-            // data anyway
-            sXt[col_idx * n + row] = (b_val - sum) / a_diag;
+        if (lane == row) {
+            x_low = (x_low - sum) / sA[row * n + row];
         }
     }
 
-    __syncthreads();
-
 #pragma unroll
-    for (int i = 0; i < rows_per_warp; i++) {
-        const int i0 = lane + i * WARP_SIZE;
-        if (i0 < n) {
-            X_batch[i0 * k + col_idx] = sXt[col_idx * n + i0];
+    for (int row = half; row < n; ++row) {
+        float sum = sA[row * n + lane] * x_low;
+        const int j = half + lane;
+        if (j < row) {
+            sum += sA[row * n + j] * x_high;
+        }
+        sum = warp_reduce_sum(sum);
+
+        if (lane == row - half) {
+            x_high = (x_high - sum) / sA[row * n + row];
+        }
+    }
+
+#pragma unroll
+    for (int rr = 0; rr < 2; ++rr) {
+        const int row = rr * WARP_SIZE + lane;
+        if (row < n) {
+            const float val = (row < half) ? x_low : x_high;
+            X_batch[row * k + col_idx] = val;
         }
     }
 }

ml/backend/ggml/ggml/src/ggml-cuda/tri.cu

@@ -0,0 +1,136 @@
+#include "common.cuh"
+#include "convert.cuh"
+#include "tri.cuh"
+#include "ggml.h"
+
+template<typename T, bool prefix_keep, int add_to_split>
+static __global__ void tri_kernel(
+        const T * src, T * dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
+        const int64_t nb0,  const int64_t nb1,  const int64_t nb2,  const int64_t nb3) {
+    const int64_t i3 = blockIdx.z;
+    const int64_t i2 = blockIdx.y;
+    const int64_t i1 = blockIdx.x;
+    const int64_t split_point = i1 + add_to_split;
+
+    GGML_UNUSED_VARS(nb00, nb0);
+
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    const T * src_row = src + i1*nb01 + i2*nb02 + i3*nb03;
+    T       * dst_row = dst + i1*nb1  + i2*nb2  + i3*nb3;
+
+    if constexpr (prefix_keep) {
+        for (int64_t i0 = threadIdx.x; i0 < split_point; i0 += blockDim.x) {
+            dst_row[i0] = src_row[i0];
+        }
+        for (int64_t i0 = threadIdx.x + split_point; i0 < ne00; i0 += blockDim.x) {
+            dst_row[i0] = ggml_cuda_cast<T, float>(0.0f);
+        }
+    } else {
+        for (int64_t i0 = threadIdx.x; i0 < split_point; i0 += blockDim.x) {
+            dst_row[i0] = ggml_cuda_cast<T, float>(0.0f);
+        }
+        for (int64_t i0 = threadIdx.x + split_point; i0 < ne00; i0 += blockDim.x) {
+            dst_row[i0] = src_row[i0];
+        }
+    }
+}
+
+template<typename T>
+static void tri_cuda(
+        const T * src, T * dst,
+        const int64_t ne00, const int64_t ne01, const int64_t ne02, const int64_t ne03,
+        const int64_t nb00, const int64_t nb01, const int64_t nb02, const int64_t nb03,
+        const int64_t nb0,  const int64_t nb1,  const int64_t nb2,  const int64_t nb3,
+        const ggml_tri_type ttype,
+        cudaStream_t stream) {
+
+    dim3 block_dims(CUDA_TRI_BLOCK_SIZE, 1, 1);
+    dim3 grid_dims(ne01, ne02, ne03);
+    const size_t type_size = sizeof(T);
+
+    const int add_to_split = (ttype == GGML_TRI_TYPE_LOWER_DIAG || ttype == GGML_TRI_TYPE_UPPER) ? 1 : 0;
+    const bool prefix_keep = (ttype == GGML_TRI_TYPE_LOWER || ttype == GGML_TRI_TYPE_LOWER_DIAG);
+
+    if (prefix_keep) {
+        if (add_to_split == 0) {
+            tri_kernel<T, true, 0><<<grid_dims, block_dims, 0, stream>>>(
+                src, dst,
+                ne00, ne01, ne02, ne03,
+                nb00 / type_size, nb01 / type_size, nb02 / type_size, nb03 / type_size,
+                nb0 / type_size, nb1 / type_size, nb2 / type_size, nb3 / type_size
+            );
+        } else { // only 0 and 1 supported
+            tri_kernel<T, true, 1><<<grid_dims, block_dims, 0, stream>>>(
+                src, dst,
+                ne00, ne01, ne02, ne03,
+                nb00 / type_size, nb01 / type_size, nb02 / type_size, nb03 / type_size,
+                nb0 / type_size, nb1 / type_size, nb2 / type_size, nb3 / type_size
+            );
+        }
+    } else {
+        if (add_to_split == 0) {
+            tri_kernel<T, false, 0><<<grid_dims, block_dims, 0, stream>>>(
+                src, dst,
+                ne00, ne01, ne02, ne03,
+                nb00 / type_size, nb01 / type_size, nb02 / type_size, nb03 / type_size,
+                nb0 / type_size, nb1 / type_size, nb2 / type_size, nb3 / type_size
+            );
+        } else {
+            tri_kernel<T, false, 1><<<grid_dims, block_dims, 0, stream>>>(
+                src, dst,
+                ne00, ne01, ne02, ne03,
+                nb00 / type_size, nb01 / type_size, nb02 / type_size, nb03 / type_size,
+                nb0 / type_size, nb1 / type_size, nb2 / type_size, nb3 / type_size
+            );
+        }
+    }
+}
+
+void ggml_cuda_op_tri(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    cudaStream_t stream = ctx.stream();
+
+    const ggml_tri_type ttype = static_cast<ggml_tri_type>(ggml_get_op_params_i32(dst, 0));
+
+    GGML_ASSERT(src0->type == dst->type);
+
+    switch(src0->type) {
+        case GGML_TYPE_F32:
+            {
+                tri_cuda(
+                    (const float *)src0->data, (float *)dst->data,
+                    src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+                    src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                    dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
+                    ttype, stream
+                );
+            } break;
+        case GGML_TYPE_F16:
+            {
+                tri_cuda(
+                    (const half *)src0->data, (half *)dst->data,
+                    src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+                    src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                    dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
+                    ttype, stream
+                );
+            } break;
+        case GGML_TYPE_BF16:
+            {
+                tri_cuda(
+                    (const nv_bfloat16 *)src0->data, (nv_bfloat16 *)dst->data,
+                    src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+                    src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                    dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3],
+                    ttype, stream
+                );
+            } break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+}

ml/backend/ggml/ggml/src/ggml-cuda/tri.cuh

@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_TRI_BLOCK_SIZE 256
+
+void ggml_cuda_op_tri(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ml/backend/ggml/ggml/src/ggml-cuda/upscale.cu

@@ -81,6 +81,76 @@ static __global__ void upscale_f32_bilinear(const float * x, float * dst,
     dst[index] = result;
 }
 
+// Similar to F.interpolate(..., mode="bilinear", align_corners=False, antialias=True)
+// https://github.com/pytorch/pytorch/blob/8871ff29b743948d1225389d5b7068f37b22750b/aten/src/ATen/native/cpu/UpSampleKernel.cpp
+static __global__ void upscale_f32_bilinear_antialias(const float * src0, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne00_src, const int ne01_src,
+        const int ne10_dst, const int ne11_dst, const int ne12_dst, const int ne13_dst,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        const float pixel_offset) {
+    const int64_t index              = threadIdx.x + blockIdx.x * blockDim.x;
+    const int64_t dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    const int i10_dst = index % ne10_dst;
+    const int i11_dst = (index / ne10_dst) % ne11_dst;
+    const int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    const int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    const int i02_src = (int)(i12_dst / sf2);
+    const int i03_src = (int)(i13_dst / sf3);
+
+    const float y = ((float)i11_dst + pixel_offset) / sf1;
+    const float x = ((float)i10_dst + pixel_offset) / sf0;
+
+    // support and invscale, minimum 1 pixel for bilinear
+    const float support1  = max(1.0f / sf1, 1.0f);
+    const float invscale1 = 1.0f / support1;
+    const float support0  = max(1.0f / sf0, 1.0f);
+    const float invscale0 = 1.0f / support0;
+
+    // the range of source pixels that contribute
+    const int64_t x_min = max(int64_t(0), int64_t(x - support0 + pixel_offset));
+    const int64_t x_max = min(int64_t(ne00_src), int64_t(x + support0 + pixel_offset));
+    const int64_t y_min = max(int64_t(0), int64_t(y - support1 + pixel_offset));
+    const int64_t y_max = min(int64_t(ne01_src), int64_t(y + support1 + pixel_offset));
+
+    // bilinear filter with antialiasing
+    float val = 0.0f;
+    float total_weight = 0.0f;
+
+    auto triangle_filter = [](float x) -> float {
+        return max(1.0f - fabsf(x), 0.0f);
+    };
+
+    for (int64_t sy = y_min; sy < y_max; sy++) {
+        const float weight_y = triangle_filter((sy - y + pixel_offset) * invscale1);
+
+        for (int64_t sx = x_min; sx < x_max; sx++) {
+            const float weight_x = triangle_filter((sx - x + pixel_offset) * invscale0);
+            const float weight = weight_x * weight_y;
+
+            if (weight <= 0.0f) {
+                continue;
+            }
+
+            const float pixel = *(const float *)((const char *)src0 + sx*nb00 + sy*nb01 + i02_src*nb02 + i03_src*nb03);
+            val += pixel * weight;
+            total_weight += weight;
+        }
+    }
+
+    if (total_weight > 0.0f) {
+        val /= total_weight;
+    }
+
+    dst[index] = val;
+}
+
 namespace bicubic_interpolation {
 // https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm
 __device__ const float a = -0.75f; // use alpha = -0.75 (same as PyTorch)
@@ -161,11 +231,15 @@ static void upscale_f32_bilinear_cuda(const float * x, float * dst,
         const int ne00_src, const int ne01_src,
         const int ne10_dst, const int ne11_dst, const int ne12_dst, const int ne13_dst,
         const float sf0, const float sf1, const float sf2, const float sf3,
-        const float pixel_offset, cudaStream_t stream) {
+        const float pixel_offset, bool antialias, cudaStream_t stream) {
     const int64_t dst_size   = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
     const int64_t num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
 
+    if (antialias) {
+        upscale_f32_bilinear_antialias<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne00_src, ne01_src, ne10_dst, ne11_dst, ne12_dst, ne13_dst, sf0, sf1, sf2, sf3, pixel_offset);
+    } else {
         upscale_f32_bilinear<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne00_src, ne01_src, ne10_dst, ne11_dst, ne12_dst, ne13_dst, sf0, sf1, sf2, sf3, pixel_offset);
+    }
 }
 
 static void upscale_f32_bicubic_cuda(const float * x, float * dst,
@@ -207,9 +281,10 @@ void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     if (mode == GGML_SCALE_MODE_NEAREST) {
         upscale_f32_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
     } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        const bool antialias = (mode_flags & GGML_SCALE_FLAG_ANTIALIAS);
         upscale_f32_bilinear_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
                                  src0->ne[0], src0->ne[1], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
-                                 sf0, sf1, sf2, sf3, pixel_offset, stream);
+                                 sf0, sf1, sf2, sf3, pixel_offset, antialias, stream);
     } else if (mode == GGML_SCALE_MODE_BICUBIC) {
         upscale_f32_bicubic_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
                                  src0->ne[0], src0->ne[1], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],

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

@@ -24,9 +24,6 @@ struct ggml_metal_command_buffer {
 };
 
 struct ggml_metal {
-    id<MTLDevice>       device;
-    id<MTLCommandQueue> queue; // currently a pointer to the device queue, but might become separate queue [TAG_QUEUE_PER_BACKEND]
-
     ggml_metal_device_t  dev;
     ggml_metal_library_t lib;
 
@@ -91,15 +88,15 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
     // init context
     ggml_metal_t res = calloc(1, sizeof(struct ggml_metal));
 
-    res->device = ggml_metal_device_get_obj(dev);
+    id<MTLDevice> device = ggml_metal_device_get_obj(dev);
 
-    GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[res->device name] UTF8String]);
+    GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
     // TODO: would it be better to have one queue for the backend and one queue for the device?
     //       the graph encoders and async ops would use the backend queue while the sync ops would use the device queue?
     //res->queue = [device newCommandQueue]; [TAG_QUEUE_PER_BACKEND]
-    res->queue = ggml_metal_device_get_queue(dev);
-    if (res->queue == nil) {
+    id<MTLCommandQueue> queue = ggml_metal_device_get_queue(dev);
+    if (queue == nil) {
         GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__);
         return NULL;
     }
@@ -274,7 +271,8 @@ static struct ggml_metal_buffer_id ggml_metal_get_buffer_id(const struct ggml_te
 void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     @autoreleasepool {
         // wrap the source data into a Metal buffer
-        id<MTLBuffer> buf_src = [ctx->device newBufferWithBytes:data
+        id<MTLDevice> device = ggml_metal_device_get_obj(ctx->dev);
+        id<MTLBuffer> buf_src = [device newBufferWithBytes:data
                                                          length:size
                                                         options:MTLResourceStorageModeShared];
 
@@ -289,7 +287,8 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,
 
         // queue the copy operation into the queue of the Metal context
         // this will be queued at the end, after any currently ongoing GPU operations
-        id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBuffer];
+        id<MTLCommandQueue> queue = ggml_metal_device_get_queue(ctx->dev);
+        id<MTLCommandBuffer> cmd_buf = [queue commandBuffer];
         id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
 
         [encoder copyFromBuffer:buf_src
@@ -315,7 +314,8 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,
 
 void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     @autoreleasepool {
-        id<MTLBuffer> buf_dst = [ctx->device newBufferWithBytesNoCopy:data
+        id<MTLDevice> device = ggml_metal_device_get_obj(ctx->dev);
+        id<MTLBuffer> buf_dst = [device newBufferWithBytesNoCopy:data
                                                                length:size
                                                               options:MTLResourceStorageModeShared
                                                           deallocator:nil];
@@ -331,7 +331,8 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te
 
         // queue the copy operation into the queue of the Metal context
         // this will be queued at the end, after any currently ongoing GPU operations
-        id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBuffer];
+        id<MTLCommandQueue> queue = ggml_metal_device_get_queue(ctx->dev);
+        id<MTLCommandBuffer> cmd_buf = [queue commandBuffer];
         id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
 
         [encoder copyFromBuffer:bid_src.metal
@@ -362,6 +363,9 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
     // number of threads in addition to the main thread
     const int n_cb = ctx->n_cb;
 
+    // keep the memory wired
+    ggml_metal_device_rsets_keep_alive(ctx->dev);
+
     // submit the ggml compute graph to the GPU by creating command buffers and encoding the ops in them
     // the first n_nodes_0 are encoded and submitted for processing directly by the calling thread
     // while these nodes are processing, we start n_cb threads to enqueue the rest of the nodes
@@ -389,7 +393,8 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
 
             if (!ctx->capture_started) {
                 // create capture scope
-                ctx->capture_scope = [[MTLCaptureManager sharedCaptureManager] newCaptureScopeWithDevice:ctx->device];
+                id<MTLDevice> device = ggml_metal_device_get_obj(ctx->dev);
+                ctx->capture_scope = [[MTLCaptureManager sharedCaptureManager] newCaptureScopeWithDevice:device];
 
                 MTLCaptureDescriptor * descriptor = [MTLCaptureDescriptor new];
                 descriptor.captureObject = ctx->capture_scope;
@@ -406,10 +411,13 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
             }
         }
 
+        // short-hand
+        id<MTLCommandQueue> queue = ggml_metal_device_get_queue(ctx->dev);
+
         // the main thread commits the first few commands immediately
         // cmd_buf[n_cb]
         {
-            id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
+            id<MTLCommandBuffer> cmd_buf = [queue commandBufferWithUnretainedReferences];
             [cmd_buf retain];
 
             if (ctx->cmd_bufs[n_cb].obj) {
@@ -428,7 +436,7 @@ enum ggml_status ggml_metal_graph_compute(ggml_metal_t ctx, struct ggml_cgraph *
         // prepare the rest of the command buffers asynchronously (optional)
         // cmd_buf[0.. n_cb)
         for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
-            id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
+            id<MTLCommandBuffer> cmd_buf = [queue commandBufferWithUnretainedReferences];
             [cmd_buf retain];
 
             if (ctx->cmd_bufs[cb_idx].obj) {
@@ -589,9 +597,11 @@ void ggml_metal_set_abort_callback(ggml_metal_t ctx, ggml_abort_callback abort_c
 }
 
 bool ggml_metal_supports_family(ggml_metal_t ctx, int family) {
-    GGML_ASSERT(ctx->device != nil);
+    GGML_ASSERT(ctx->dev != nil);
 
-    return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
+    id<MTLDevice> device = ggml_metal_device_get_obj(ctx->dev);
+
+    return [device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
 }
 
 void ggml_metal_capture_next_compute(ggml_metal_t ctx) {

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

@@ -35,20 +35,6 @@ typedef struct ggml_metal_pipeline * ggml_metal_pipeline_t;
 ggml_metal_pipeline_t ggml_metal_pipeline_init(void);
 void ggml_metal_pipeline_free(ggml_metal_pipeline_t pipeline);
 
-void ggml_metal_pipeline_set_nsg(ggml_metal_pipeline_t pipeline, int nsg);
-int  ggml_metal_pipeline_get_nsg(ggml_metal_pipeline_t pipeline);
-
-void ggml_metal_pipeline_set_nr0(ggml_metal_pipeline_t pipeline, int nr0);
-int  ggml_metal_pipeline_get_nr0(ggml_metal_pipeline_t pipeline);
-
-void ggml_metal_pipeline_set_nr1(ggml_metal_pipeline_t pipeline, int nr1);
-int  ggml_metal_pipeline_get_nr1(ggml_metal_pipeline_t pipeline);
-
-void   ggml_metal_pipeline_set_smem(ggml_metal_pipeline_t pipeline, size_t smem);
-size_t ggml_metal_pipeline_get_smem(ggml_metal_pipeline_t pipeline);
-
-int ggml_metal_pipeline_max_theads_per_threadgroup(ggml_metal_pipeline_t pipeline);
-
 // a collection of pipelines
 typedef struct ggml_metal_pipelines * ggml_metal_pipelines_t;
 
@@ -58,6 +44,19 @@ void ggml_metal_pipelines_free(ggml_metal_pipelines_t ppls);
 void                  ggml_metal_pipelines_add(ggml_metal_pipelines_t ppls, const char * name, ggml_metal_pipeline_t pipeline);
 ggml_metal_pipeline_t ggml_metal_pipelines_get(ggml_metal_pipelines_t ppls, const char * name);
 
+struct ggml_metal_pipeline_with_params {
+    ggml_metal_pipeline_t pipeline;
+
+    int nsg;
+
+    int nr0;
+    int nr1;
+
+    size_t smem;
+};
+
+int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_with_params pipeline);
+
 //
 // MTLCommandBuffer wrapper
 //
@@ -76,7 +75,7 @@ void ggml_metal_encoder_free(ggml_metal_encoder_t encoder);
 void ggml_metal_encoder_debug_group_push(ggml_metal_encoder_t encoder, const char * name);
 void ggml_metal_encoder_debug_group_pop (ggml_metal_encoder_t encoder);
 
-void ggml_metal_encoder_set_pipeline(ggml_metal_encoder_t encoder, ggml_metal_pipeline_t pipeline);
+void ggml_metal_encoder_set_pipeline(ggml_metal_encoder_t encoder, struct ggml_metal_pipeline_with_params pipeline);
 
 void ggml_metal_encoder_set_bytes (ggml_metal_encoder_t encoder, void * data, size_t size, int idx);
 void ggml_metal_encoder_set_buffer(ggml_metal_encoder_t encoder, struct ggml_metal_buffer_id buffer, int idx);
@@ -100,66 +99,68 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
 
 void ggml_metal_library_free(ggml_metal_library_t lib);
 
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline    (ggml_metal_library_t lib, const char * name);
-ggml_metal_pipeline_t ggml_metal_library_compile_pipeline(ggml_metal_library_t lib, const char * base, const char * name, ggml_metal_cv_t cv);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline    (ggml_metal_library_t lib, const char * name);
+struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_metal_library_t lib, const char * base, const char * name, ggml_metal_cv_t cv);
 
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_base              (ggml_metal_library_t lib, enum ggml_op op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cpy               (ggml_metal_library_t lib, enum ggml_type tsrc, enum ggml_type tdst);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pool_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op, enum ggml_op_pool op_pool);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_get_rows          (ggml_metal_library_t lib, enum ggml_type tsrc);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_set_rows          (ggml_metal_library_t lib, enum ggml_type tidx, enum ggml_type tdst);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_repeat            (ggml_metal_library_t lib, enum ggml_type tsrc);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_unary             (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum               (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_sum_rows          (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cumsum_blk        (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_cumsum_add        (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_soft_max          (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rwkv              (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t 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);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm            (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv            (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id_map0    (ggml_metal_library_t lib, int ne02, int ne20);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mm_id         (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_mul_mv_id         (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argmax            (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort           (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_argsort_merge     (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k             (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_top_k_merge       (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, enum ggml_op op, int32_t n_fuse, bool row);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_l2_norm           (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_group_norm        (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_norm              (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_rope              (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_im2col            (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_transpose_1d (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_transpose_2d (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_conv_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_upscale           (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pad               (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_pad_reflect_1d    (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_arange            (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_timestep_embedding(ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_opt_step_adamw    (ggml_metal_library_t lib, const struct ggml_tensor * op);
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_opt_step_sgd      (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_base              (ggml_metal_library_t lib, enum ggml_op op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cpy               (ggml_metal_library_t lib, enum ggml_type tsrc, enum ggml_type tdst);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pool_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op, enum ggml_op_pool op_pool);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_get_rows          (ggml_metal_library_t lib, enum ggml_type tsrc);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_set_rows          (ggml_metal_library_t lib, enum ggml_type tidx, enum ggml_type tdst);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_repeat            (ggml_metal_library_t lib, enum ggml_type tsrc);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_unary             (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_glu               (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum               (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_sum_rows          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cumsum_blk        (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_cumsum_add        (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_tri               (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_soft_max          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_conv_batched  (ggml_metal_library_t lib, const struct ggml_tensor * op, int ssm_conv_bs);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_ssm_scan          (ggml_metal_library_t lib, const struct ggml_tensor * op);
+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_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);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv            (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm_id_map0    (ggml_metal_library_t lib, int ne02, int ne20);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mm_id         (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_id         (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argmax            (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort           (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_argsort_merge     (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k             (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_top_k_merge       (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_bin               (ggml_metal_library_t lib, enum ggml_op op, int32_t n_fuse, bool row);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_l2_norm           (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_group_norm        (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_norm              (ggml_metal_library_t lib, const struct ggml_tensor * op, int32_t n_fuse);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_rope              (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_im2col            (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_transpose_1d (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_transpose_2d (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_2d           (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_upscale           (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad               (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad_reflect_1d    (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_arange            (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_timestep_embedding(ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_opt_step_adamw    (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_opt_step_sgd      (ggml_metal_library_t lib, const struct ggml_tensor * op);
 
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_pad(
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_pad(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,
         bool    has_mask,
         int32_t ncpsg);
 
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_blk(
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_blk(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,
         int32_t nqptg,
         int32_t ncpsg);
 
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,
         bool    has_mask,
@@ -169,7 +170,7 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext(
         bool    has_kvpad,
         int32_t nsg);
 
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,
         bool    has_mask,
@@ -180,12 +181,22 @@ ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec(
         int32_t nsg,
         int32_t nwg);
 
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline_flash_attn_ext_vec_reduce(
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec_reduce(
         ggml_metal_library_t lib,
         const struct ggml_tensor * op,
         int32_t dv,
         int32_t nwg);
 
+// MTLResidencySet wrapper
+
+typedef void * ggml_metal_rset_t;
+
+// a collection of residency sets (non-owning)
+typedef struct ggml_metal_rsets * ggml_metal_rsets_t;
+
+ggml_metal_rsets_t ggml_metal_rsets_init(void);
+void ggml_metal_rsets_free(ggml_metal_rsets_t rsets);
+
 //
 // device
 //
@@ -219,6 +230,11 @@ void * ggml_metal_device_get_queue(ggml_metal_device_t dev); // id<MTLCommandQue
 
 ggml_metal_library_t ggml_metal_device_get_library(ggml_metal_device_t dev);
 
+void ggml_metal_device_rsets_add(ggml_metal_device_t dev, ggml_metal_rset_t rset);
+void ggml_metal_device_rsets_rm (ggml_metal_device_t dev, ggml_metal_rset_t rset);
+
+void ggml_metal_device_rsets_keep_alive(ggml_metal_device_t dev);
+
 void ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total);
 bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_tensor * op);
 

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

@@ -1,7 +1,6 @@
 #import "ggml-metal-device.h"
 
 #import "ggml-impl.h"
-#import "ggml-threading.h"
 
 #include <Foundation/Foundation.h>
 
@@ -75,14 +74,6 @@ void ggml_metal_cv_set_bool(ggml_metal_cv_t cv, bool value, int32_t idx) {
 
 struct ggml_metal_pipeline {
     id<MTLComputePipelineState> obj;
-
-    // suggested dispatch sizes
-    int nsg;
-
-    int nr0;
-    int nr1;
-
-    size_t smem;
 };
 
 ggml_metal_pipeline_t ggml_metal_pipeline_init(void) {
@@ -90,10 +81,6 @@ ggml_metal_pipeline_t ggml_metal_pipeline_init(void) {
 
     *res = (struct ggml_metal_pipeline) {
         /*.obj  =*/ nil,
-        /*.nsg  =*/ 0,
-        /*.nr0  =*/ 0,
-        /*.nr1  =*/ 0,
-        /*.smem =*/ 0,
     };
 
     return res;
@@ -105,40 +92,8 @@ void ggml_metal_pipeline_free(ggml_metal_pipeline_t pipeline) {
     free(pipeline);
 }
 
-void ggml_metal_pipeline_set_nsg(ggml_metal_pipeline_t pipeline, int nsg) {
-    pipeline->nsg = nsg;
-}
-
-int ggml_metal_pipeline_get_nsg(ggml_metal_pipeline_t pipeline) {
-    return pipeline->nsg;
-}
-
-void ggml_metal_pipeline_set_nr0(ggml_metal_pipeline_t pipeline, int nr0) {
-    pipeline->nr0 = nr0;
-}
-
-int ggml_metal_pipeline_get_nr0(ggml_metal_pipeline_t pipeline) {
-    return pipeline->nr0;
-}
-
-void ggml_metal_pipeline_set_nr1(ggml_metal_pipeline_t pipeline, int nr1) {
-    pipeline->nr1 = nr1;
-}
-
-int ggml_metal_pipeline_get_nr1(ggml_metal_pipeline_t pipeline) {
-    return pipeline->nr1;
-}
-
-void   ggml_metal_pipeline_set_smem(ggml_metal_pipeline_t pipeline, size_t smem) {
-    pipeline->smem = smem;
-}
-
-size_t ggml_metal_pipeline_get_smem(ggml_metal_pipeline_t pipeline) {
-    return pipeline->smem;
-}
-
-int ggml_metal_pipeline_max_theads_per_threadgroup(ggml_metal_pipeline_t pipeline) {
-    return pipeline->obj.maxTotalThreadsPerThreadgroup;
+int ggml_metal_pipeline_max_theads_per_threadgroup(struct ggml_metal_pipeline_with_params pipeline) {
+    return pipeline.pipeline->obj.maxTotalThreadsPerThreadgroup;
 }
 
 struct ggml_metal_library {
@@ -146,6 +101,8 @@ struct ggml_metal_library {
     id<MTLDevice> device;
 
     ggml_metal_pipelines_t pipelines; // cache of compiled pipelines
+
+    NSLock * lock;
 };
 
 ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
@@ -299,6 +256,7 @@ ggml_metal_library_t ggml_metal_library_init(ggml_metal_device_t dev) {
     res->obj       = library;
     res->device    = device;
     res->pipelines = ggml_metal_pipelines_init();
+    res->lock      = [NSLock new];
 
     return res;
 }
@@ -365,6 +323,7 @@ ggml_metal_library_t ggml_metal_library_init_from_source(ggml_metal_device_t dev
     res->obj       = library;
     res->device    = device;
     res->pipelines = ggml_metal_pipelines_init();
+    res->lock      = [NSLock new];
 
     return res;
 }
@@ -380,26 +339,47 @@ void ggml_metal_library_free(ggml_metal_library_t lib) {
 
     ggml_metal_pipelines_free(lib->pipelines);
 
+    [lib->lock release];
+
     free(lib);
 }
 
-ggml_metal_pipeline_t ggml_metal_library_get_pipeline(ggml_metal_library_t lib, const char * name) {
-    return ggml_metal_pipelines_get(lib->pipelines, name);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline(ggml_metal_library_t lib, const char * name) {
+    [lib->lock lock];
+
+    struct ggml_metal_pipeline_with_params res = {
+        /*.pipeline =*/ nil,
+        /*.nr0      =*/ 0,
+        /*.nr1      =*/ 0,
+        /*.nsg      =*/ 0,
+        /*.smem     =*/ 0,
+    };
+
+    res.pipeline = ggml_metal_pipelines_get(lib->pipelines, name);
+
+    [lib->lock unlock];
+
+    return res;
 }
 
-ggml_metal_pipeline_t ggml_metal_library_compile_pipeline(ggml_metal_library_t lib, const char * base, const char * name, ggml_metal_cv_t cv) {
-    // note: the pipelines are cached in the library per device, so they are shared across all metal contexts
-    ggml_critical_section_start();
+struct ggml_metal_pipeline_with_params ggml_metal_library_compile_pipeline(ggml_metal_library_t lib, const char * base, const char * name, ggml_metal_cv_t cv) {
+    struct ggml_metal_pipeline_with_params res = {
+        /*.pipeline =*/ nil,
+        /*.nr0      =*/ 0,
+        /*.nr1      =*/ 0,
+        /*.nsg      =*/ 0,
+        /*.smem     =*/ 0,
+    };
 
-    ggml_metal_pipeline_t res = ggml_metal_library_get_pipeline(lib, name);
-    if (res) {
-        ggml_critical_section_end();
+    [lib->lock lock];
+
+    res.pipeline = ggml_metal_pipelines_get(lib->pipelines, name);
+    if (res.pipeline) {
+        [lib->lock unlock];
 
         return res;
     }
 
-    res = ggml_metal_pipeline_init();
-
     @autoreleasepool {
         NSError * error = nil;
 
@@ -414,36 +394,53 @@ ggml_metal_pipeline_t ggml_metal_library_compile_pipeline(ggml_metal_library_t l
             mtl_function = [lib->obj newFunctionWithName:base_func constantValues:cv->obj error:&error];
         }
         if (!mtl_function) {
-            ggml_critical_section_end();
+            [lib->lock unlock];
 
             GGML_LOG_ERROR("%s: failed to compile pipeline: base = '%s', name = '%s'\n", __func__, base, name);
             if (error) {
                 GGML_LOG_ERROR("%s: %s\n", __func__, [[error description] UTF8String]);
             }
 
-            return nil;
+            return res;
         }
 
-        res->obj = [lib->device newComputePipelineStateWithFunction:mtl_function error:&error];
+        id<MTLComputePipelineState> obj = [lib->device newComputePipelineStateWithFunction:mtl_function error:&error];
 
         [mtl_function release];
 
-        GGML_LOG_DEBUG("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, name, (void *) res->obj,
-                (int) res->obj.maxTotalThreadsPerThreadgroup,
-                (int) res->obj.threadExecutionWidth);
+        if (!obj) {
+            [lib->lock unlock];
 
-        if (res->obj.maxTotalThreadsPerThreadgroup == 0 || res->obj.threadExecutionWidth == 0) {
-            ggml_critical_section_end();
+            GGML_LOG_ERROR("%s: failed to create pipeline state: base = '%s', name = '%s'\n", __func__, base, name);
+            if (error) {
+                GGML_LOG_ERROR("%s: %s\n", __func__, [[error description] UTF8String]);
+            }
+
+            return res;
+        }
+
+        GGML_LOG_DEBUG("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, name,
+                (void *) obj,
+                (int)    obj.maxTotalThreadsPerThreadgroup,
+                (int)    obj.threadExecutionWidth);
+
+        if (obj.maxTotalThreadsPerThreadgroup == 0 || obj.threadExecutionWidth == 0) {
+            [obj release];
+
+            [lib->lock unlock];
 
             GGML_LOG_ERROR("%s: incompatible pipeline %s\n", __func__, name);
 
-            return nil;
+            return res;
         }
 
-        ggml_metal_pipelines_add(lib->pipelines, name, res);
+        res.pipeline = ggml_metal_pipeline_init();
+        res.pipeline->obj = obj;
+
+        ggml_metal_pipelines_add(lib->pipelines, name, res.pipeline);
     }
 
-    ggml_critical_section_end();
+    [lib->lock unlock];
 
     return res;
 }
@@ -485,8 +482,8 @@ void ggml_metal_encoder_debug_group_pop (ggml_metal_encoder_t encoder) {
     [encoder->obj popDebugGroup];
 }
 
-void ggml_metal_encoder_set_pipeline(ggml_metal_encoder_t encoder, ggml_metal_pipeline_t pipeline) {
-    [encoder->obj setComputePipelineState:pipeline->obj];
+void ggml_metal_encoder_set_pipeline(ggml_metal_encoder_t encoder, struct ggml_metal_pipeline_with_params pipeline) {
+    [encoder->obj setComputePipelineState:pipeline.pipeline->obj];
 }
 
 void ggml_metal_encoder_set_bytes(ggml_metal_encoder_t encoder, void * data, size_t size, int idx) {
@@ -521,11 +518,106 @@ struct ggml_metal_device {
     // ref: https://github.com/ggml-org/llama.cpp/pull/15906
     id<MTLCommandQueue> mtl_queue;
 
+    ggml_metal_rsets_t rsets;
+
     ggml_metal_library_t library;
 
     struct ggml_metal_device_props props;
 };
 
+//
+// MTLResidenceSet wrapper
+//
+
+struct ggml_metal_rsets {
+    NSLock * lock;
+
+    NSMutableArray * data;
+
+    // number of seconds since the last graph computation
+    // keep the residency sets wired for that amount of time to avoid being collected by the OS
+    int keep_alive_s;
+
+    // background heartbeat thread to keep the residency sets alive
+    atomic_bool d_stop;
+    atomic_int  d_loop;
+
+    dispatch_group_t d_group;
+};
+
+ggml_metal_rsets_t ggml_metal_rsets_init(void) {
+    ggml_metal_rsets_t res = calloc(1, sizeof(struct ggml_metal_rsets));
+
+    res->lock = [[NSLock alloc] init];
+    res->data = [[NSMutableArray alloc] init];
+
+    // by default keep the memory wired for 3 minutes
+    res->keep_alive_s = 3*60;
+
+    const char * GGML_METAL_RESIDENCY_KEEP_ALIVE_S = getenv("GGML_METAL_RESIDENCY_KEEP_ALIVE_S");
+    if (GGML_METAL_RESIDENCY_KEEP_ALIVE_S) {
+        res->keep_alive_s = atoi(GGML_METAL_RESIDENCY_KEEP_ALIVE_S);
+    }
+
+    if (res->keep_alive_s <= 0) {
+        res->keep_alive_s = 3*60;
+    }
+
+    GGML_LOG_INFO("%s: creating a residency set collection (keep_alive = %d s)\n", __func__, res->keep_alive_s);
+
+    atomic_store_explicit(&res->d_stop, false, memory_order_relaxed);
+    atomic_store_explicit(&res->d_loop, 2*res->keep_alive_s, memory_order_relaxed);
+
+    res->d_group = dispatch_group_create();
+
+    // start a background thread that periodically requests residency for all the currently active sets in the collection
+    // the requests stop after a certain amount of time (keep_alive_s) of inactivity
+    dispatch_queue_t d_queue = dispatch_get_global_queue(QOS_CLASS_DEFAULT, 0);
+    dispatch_group_async(res->d_group, d_queue, ^{
+#if defined(GGML_METAL_HAS_RESIDENCY_SETS)
+        if (@available(macOS 15.0, iOS 18.0, tvOS 18.0, visionOS 2.0, *)) {
+              while (!atomic_load_explicit(&res->d_stop, memory_order_relaxed)) {
+                  if (atomic_load_explicit(&res->d_loop, memory_order_relaxed) > 0) {
+                      [res->lock lock];
+
+                      for (int i = 0; i < (int) res->data.count; ++i) {
+                          [res->data[i] requestResidency];
+                      }
+
+                      atomic_fetch_sub_explicit(&res->d_loop, 1, memory_order_relaxed);
+
+                      [res->lock unlock];
+                  }
+
+                  // half a second
+                  usleep(500 * 1000);
+              }
+        }
+#endif
+    });
+
+    return res;
+}
+
+void ggml_metal_rsets_free(ggml_metal_rsets_t rsets) {
+    if (rsets == NULL) {
+        return;
+    }
+
+    // note: if you hit this assert, most likely you haven't deallocated all Metal resources before exiting
+    GGML_ASSERT([rsets->data count] == 0);
+
+    atomic_store_explicit(&rsets->d_stop, true, memory_order_relaxed);
+
+    dispatch_group_wait(rsets->d_group, DISPATCH_TIME_FOREVER);
+    dispatch_release(rsets->d_group);
+
+    [rsets->data release];
+    [rsets->lock release];
+
+    free(rsets);
+}
+
 ggml_metal_device_t ggml_metal_device_init(void) {
     ggml_metal_device_t dev = calloc(1, sizeof(struct ggml_metal_device));
 
@@ -611,8 +703,8 @@ ggml_metal_device_t ggml_metal_device_init(void) {
                     GGML_LOG_WARN("%s: - the tensor API is not supported in this environment - disabling\n", __func__);
                     dev->props.has_tensor = false;
                 } else {
-                    ggml_metal_pipeline_t ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil);
-                    if (!ppl) {
+                    struct ggml_metal_pipeline_with_params ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil);
+                    if (!ppl.pipeline) {
                         GGML_LOG_WARN("%s: - the tensor API is not supported in this environment - disabling\n", __func__);
                         dev->props.has_tensor = false;
                     }
@@ -661,8 +753,8 @@ ggml_metal_device_t ggml_metal_device_init(void) {
                     GGML_LOG_WARN("%s: - the tensor API does not support bfloat - disabling bfloat support\n", __func__);
                     dev->props.has_bfloat = false;
                 } else {
-                    ggml_metal_pipeline_t ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil);
-                    if (!ppl) {
+                    struct ggml_metal_pipeline_with_params ppl = ggml_metal_library_compile_pipeline(lib, "dummy_kernel", "dummy_kernel", nil);
+                    if (!ppl.pipeline) {
                         GGML_LOG_WARN("%s: - the tensor API does not support bfloat - disabling bfloat support\n", __func__);
                         dev->props.has_bfloat = false;
                     }
@@ -694,7 +786,11 @@ ggml_metal_device_t ggml_metal_device_init(void) {
                 GGML_LOG_ERROR("%s: error: failed to create library\n", __func__);
             }
 
-            // --------------------------------------------------
+            if (dev->props.use_residency_sets) {
+                dev->rsets = ggml_metal_rsets_init();
+            } else {
+                dev->rsets = nil;
+            }
 
             // print MTL GPU family:
             GGML_LOG_INFO("%s: GPU name:   %s\n", __func__, dev->props.name);
@@ -747,6 +843,8 @@ ggml_metal_device_t ggml_metal_device_init(void) {
 void ggml_metal_device_free(ggml_metal_device_t dev) {
     assert(dev != NULL);
 
+    ggml_metal_rsets_free(dev->rsets);
+
     ggml_metal_library_free(dev->library);
     dev->library = NULL;
 
@@ -775,6 +873,42 @@ ggml_metal_library_t ggml_metal_device_get_library(ggml_metal_device_t dev) {
     return dev->library;
 }
 
+void ggml_metal_device_rsets_add(ggml_metal_device_t dev, ggml_metal_rset_t rset) {
+    if (rset == nil) {
+        return;
+    }
+
+    GGML_ASSERT(dev->rsets);
+
+    [dev->rsets->lock lock];
+
+    [dev->rsets->data addObject:rset];
+
+    [dev->rsets->lock unlock];
+}
+
+void ggml_metal_device_rsets_rm(ggml_metal_device_t dev, ggml_metal_rset_t rset) {
+    if (rset == nil) {
+        return;
+    }
+
+    GGML_ASSERT(dev->rsets);
+
+    [dev->rsets->lock lock];
+
+    [dev->rsets->data removeObject:rset];
+
+    [dev->rsets->lock unlock];
+}
+
+void ggml_metal_device_rsets_keep_alive(ggml_metal_device_t dev) {
+    if (dev->rsets == NULL) {
+        return;
+    }
+
+    atomic_store_explicit(&dev->rsets->d_loop, 2*dev->rsets->keep_alive_s, memory_order_relaxed);
+}
+
 void ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total) {
     if (@available(macOS 10.12, iOS 16.0, *)) {
         *total = dev->mtl_device.recommendedMaxWorkingSetSize;
@@ -820,6 +954,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
                 case GGML_UNARY_OP_HARDSWISH:
                 case GGML_UNARY_OP_HARDSIGMOID:
                 case GGML_UNARY_OP_EXP:
+                case GGML_UNARY_OP_SOFTPLUS:
+                case GGML_UNARY_OP_EXPM1:
                     return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                 default:
                     return false;
@@ -852,6 +988,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_ACC:
         case GGML_OP_REPEAT:
         case GGML_OP_SCALE:
+        case GGML_OP_FILL:
         case GGML_OP_CONV_TRANSPOSE_1D:
             return true;
         case GGML_OP_CONV_TRANSPOSE_2D:
@@ -869,6 +1006,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SUM:
             return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
+        case GGML_OP_TRI:
+            return ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_SUM_ROWS:
         case GGML_OP_CUMSUM:
         case GGML_OP_MEAN:
@@ -894,10 +1033,15 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_POOL_1D:
             return false;
         case GGML_OP_UPSCALE:
-            return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST;
+            return op->src[0]->type == GGML_TYPE_F32 && op->op_params[0] == GGML_SCALE_MODE_NEAREST && !(op->op_params[0] & GGML_SCALE_FLAG_ANTIALIAS);
         case GGML_OP_POOL_2D:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_PAD:
+            // TODO: add circular padding support for metal, see https://github.com/ggml-org/llama.cpp/pull/16985
+            if (ggml_get_op_params_i32(op, 8) != 0) {
+                return false;
+            }
+
             return (ggml_get_op_params_i32(op, 0) == 0) && (ggml_get_op_params_i32(op, 2) == 0) &&
                    (ggml_get_op_params_i32(op, 4) == 0) && (ggml_get_op_params_i32(op, 6) == 0);
         case GGML_OP_PAD_REFLECT_1D:
@@ -912,6 +1056,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             // for new head sizes, add checks here
             if (op->src[0]->ne[0] != 32 &&
                 op->src[0]->ne[0] != 40 &&
+                op->src[0]->ne[0] != 48 &&
                 op->src[0]->ne[0] != 64 &&
                 op->src[0]->ne[0] != 72 &&
                 op->src[0]->ne[0] != 80 &&
@@ -1062,9 +1207,8 @@ struct ggml_metal_buffer {
     // note: cannot use explicity "id<MTLResidencySet>" here because it is not available on certain OSes
     id rset;
 
-    // pointers to global device objects
-    id<MTLDevice> device;
-    id<MTLCommandQueue> queue;
+    // pointers to global device
+    ggml_metal_device_t dev;
 };
 
 static void ggml_metal_log_allocated_size(id<MTLDevice> device, size_t size_aligned) {
@@ -1107,7 +1251,7 @@ static bool ggml_metal_buffer_rset_init(ggml_metal_buffer_t buf) {
         desc.initialCapacity = buf->n_buffers;
 
         NSError * error;
-        buf->rset = [buf->device newResidencySetWithDescriptor:desc error:&error];
+        buf->rset = [buf->dev->mtl_device newResidencySetWithDescriptor:desc error:&error];
         if (error) {
             GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
             [desc release];
@@ -1168,6 +1312,8 @@ static void * ggml_metal_host_malloc(size_t n) {
 ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size, bool shared) {
     ggml_metal_buffer_t res = calloc(1, sizeof(struct ggml_metal_buffer));
 
+    res->dev = dev;
+
     const size_t size_page = sysconf(_SC_PAGESIZE);
 
     size_t size_aligned = size;
@@ -1192,9 +1338,6 @@ ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size,
 
     res->owned = true;
 
-    res->device = ggml_metal_device_get_obj(dev);
-    res->queue  = ggml_metal_device_get_queue(dev);
-
     res->n_buffers = 1;
 
     if (res->all_data != NULL) {
@@ -1203,12 +1346,12 @@ ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size,
 
         if (size_aligned > 0) {
             if (props_dev->use_shared_buffers && shared) {
-                res->buffers[0].metal = [res->device newBufferWithBytesNoCopy:res->all_data
+                res->buffers[0].metal = [res->dev->mtl_device newBufferWithBytesNoCopy:res->all_data
                                                                   length:size_aligned
                                                                  options:MTLResourceStorageModeShared
                                                              deallocator:nil];
             } else {
-                res->buffers[0].metal = [res->device newBufferWithLength:size_aligned options:MTLResourceStorageModePrivate];
+                res->buffers[0].metal = [res->dev->mtl_device newBufferWithLength:size_aligned options:MTLResourceStorageModePrivate];
             }
         }
 
@@ -1229,6 +1372,8 @@ ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size,
         return NULL;
     }
 
+    ggml_metal_device_rsets_add(dev, res->rset);
+
     //ggml_metal_log_allocated_size(device, size_aligned);
 
     return res;
@@ -1237,6 +1382,8 @@ ggml_metal_buffer_t ggml_metal_buffer_init(ggml_metal_device_t dev, size_t size,
 ggml_metal_buffer_t ggml_metal_buffer_map(ggml_metal_device_t dev, void * ptr, size_t size, size_t max_tensor_size) {
     ggml_metal_buffer_t res = calloc(1, sizeof(struct ggml_metal_buffer));
 
+    res->dev = dev;
+
     res->all_data = ptr;
     res->all_size = size;
 
@@ -1259,9 +1406,6 @@ ggml_metal_buffer_t ggml_metal_buffer_map(ggml_metal_device_t dev, void * ptr, s
         size_aligned += (size_page - (size_aligned % size_page));
     }
 
-    res->device = ggml_metal_device_get_obj(dev);
-    res->queue  = ggml_metal_device_get_queue(dev);
-
     const struct ggml_metal_device_props * props_dev = ggml_metal_device_get_props(dev);
 
     // the buffer fits into the max buffer size allowed by the device
@@ -1271,7 +1415,7 @@ ggml_metal_buffer_t ggml_metal_buffer_map(ggml_metal_device_t dev, void * ptr, s
         res->buffers[res->n_buffers].metal = nil;
 
         if (size_aligned > 0) {
-            res->buffers[res->n_buffers].metal = [res->device newBufferWithBytesNoCopy:ptr length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
+            res->buffers[res->n_buffers].metal = [res->dev->mtl_device newBufferWithBytesNoCopy:ptr length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
             if (res->buffers[res->n_buffers].metal == nil) {
                 GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
@@ -1280,7 +1424,7 @@ ggml_metal_buffer_t ggml_metal_buffer_map(ggml_metal_device_t dev, void * ptr, s
             }
         }
 
-        ggml_metal_log_allocated_size(res->device, size_aligned);
+        ggml_metal_log_allocated_size(res->dev->mtl_device, size_aligned);
 
         ++res->n_buffers;
     } else {
@@ -1298,7 +1442,7 @@ ggml_metal_buffer_t ggml_metal_buffer_map(ggml_metal_device_t dev, void * ptr, s
             res->buffers[res->n_buffers].metal = nil;
 
             if (size_step_aligned > 0) {
-                res->buffers[res->n_buffers].metal = [res->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) ptr + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
+                res->buffers[res->n_buffers].metal = [res->dev->mtl_device newBufferWithBytesNoCopy:(void *) ((uint8_t *) ptr + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
 
                 if (res->buffers[res->n_buffers].metal == nil) {
                     GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0);
@@ -1307,7 +1451,7 @@ ggml_metal_buffer_t ggml_metal_buffer_map(ggml_metal_device_t dev, void * ptr, s
                 }
             }
 
-            ggml_metal_log_allocated_size(res->device, size_step_aligned);
+            ggml_metal_log_allocated_size(res->dev->mtl_device, size_step_aligned);
 
             if (i + size_step < size) {
                 GGML_LOG_INFO("\n");
@@ -1325,10 +1469,14 @@ ggml_metal_buffer_t ggml_metal_buffer_map(ggml_metal_device_t dev, void * ptr, s
         return NULL;
     }
 
+    ggml_metal_device_rsets_add(dev, res->rset);
+
     return res;
 }
 
 void ggml_metal_buffer_free(ggml_metal_buffer_t buf) {
+    ggml_metal_device_rsets_rm(buf->dev, buf->rset);
+
     for (int i = 0; i < buf->n_buffers; i++) {
         [buf->buffers[i].metal release];
     }
@@ -1365,8 +1513,7 @@ void ggml_metal_buffer_memset_tensor(ggml_metal_buffer_t buf, struct ggml_tensor
         struct ggml_metal_buffer_id bid_dst = ggml_metal_buffer_get_id(buf, tensor);
         bid_dst.offs += offset;
 
-        id<MTLCommandQueue>  queue   = buf->queue;
-        id<MTLCommandBuffer> cmd_buf = [queue commandBufferWithUnretainedReferences];
+        id<MTLCommandBuffer> cmd_buf = [buf->dev->mtl_queue commandBufferWithUnretainedReferences];
 
         {
             id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
@@ -1392,7 +1539,7 @@ void ggml_metal_buffer_set_tensor(ggml_metal_buffer_t buf, struct ggml_tensor *
     @autoreleasepool {
         // src
         void * data_ptr = (void *)(uintptr_t) data; // "const cast" the src data
-        id<MTLBuffer> buf_src = [buf->device newBufferWithBytesNoCopy:data_ptr
+        id<MTLBuffer> buf_src = [buf->dev->mtl_device newBufferWithBytesNoCopy:data_ptr
                                                                length:size
                                                               options:MTLResourceStorageModeShared
                                                           deallocator:nil];
@@ -1407,8 +1554,7 @@ void ggml_metal_buffer_set_tensor(ggml_metal_buffer_t buf, struct ggml_tensor *
         //       this is alternative to waitUntilCompleted, which should be faster, but don't seem to make much difference
         dispatch_semaphore_t completion_semaphore = dispatch_semaphore_create(0);
 
-        id<MTLCommandQueue>  queue   = buf->queue;
-        id<MTLCommandBuffer> cmd_buf = [queue commandBufferWithUnretainedReferences];
+        id<MTLCommandBuffer> cmd_buf = [buf->dev->mtl_queue commandBufferWithUnretainedReferences];
 
         {
             id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
@@ -1450,15 +1596,14 @@ void ggml_metal_buffer_get_tensor(ggml_metal_buffer_t buf, const struct ggml_ten
         bid_src.offs += offset;
 
         // dst
-        id<MTLBuffer> buf_dst = [buf->device newBufferWithBytesNoCopy:data
+        id<MTLBuffer> buf_dst = [buf->dev->mtl_device newBufferWithBytesNoCopy:data
                                                                length:size
                                                               options:MTLResourceStorageModeShared
                                                           deallocator:nil];
 
         GGML_ASSERT(buf_dst);
 
-        id<MTLCommandQueue>  queue   = buf->queue;
-        id<MTLCommandBuffer> cmd_buf = [queue commandBufferWithUnretainedReferences];
+        id<MTLCommandBuffer> cmd_buf = [buf->dev->mtl_queue commandBufferWithUnretainedReferences];
 
         {
             id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
@@ -1484,8 +1629,7 @@ void ggml_metal_buffer_clear(ggml_metal_buffer_t buf, uint8_t value) {
     }
 
     @autoreleasepool {
-        id<MTLCommandQueue>  queue   = buf->queue;
-        id<MTLCommandBuffer> cmd_buf = [queue commandBufferWithUnretainedReferences];
+        id<MTLCommandBuffer> cmd_buf = [buf->dev->mtl_queue commandBufferWithUnretainedReferences];
 
         {
             id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];

ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal

@@ -1962,6 +1962,7 @@ GGML_TABLE_END()
 #define FC_MUL_MV                      600
 #define FC_MUL_MM                      700
 #define FC_ROPE                        800
+#define FC_SSM_CONV                    900
 
 // op-specific constants
 #define OP_FLASH_ATTN_EXT_NQPTG 8
@@ -2067,6 +2068,10 @@ typedef struct {
     float bias;
 } ggml_metal_kargs_scale;
 
+typedef struct {
+    float val;
+} ggml_metal_kargs_fill;
+
 typedef struct {
     float min;
     float max;
@@ -2716,6 +2721,25 @@ typedef struct {
     float    slope;
 } ggml_metal_kargs_leaky_relu;
 
+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  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_tri;
+
 typedef struct {
     int32_t  ne00;
     int32_t  ne01;
@@ -4026,6 +4050,22 @@ kernel void kernel_scale_f32_4(
     dst[tpig] = src0[tpig] * args.scale + args.bias;
 }
 
+kernel void kernel_fill_f32(
+        constant ggml_metal_kargs_fill & args,
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = args.val;
+}
+
+kernel void kernel_fill_f32_4(
+        constant ggml_metal_kargs_fill & args,
+        device const float4 * src0,
+        device       float4 * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = args.val;
+}
+
 kernel void kernel_clamp_f32(
         constant ggml_metal_kargs_clamp & args,
         device const float * src0,
@@ -4372,6 +4412,36 @@ kernel void kernel_exp_f32_4(
     dst[tpig] = exp(src0[tpig]);
 }
 
+kernel void kernel_softplus_f32(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+    dst[tpig] = select(log(1.0f + exp(x)), x, x > 20.0f);
+}
+
+kernel void kernel_softplus_f32_4(
+        device const float4 * src0,
+        device       float4 * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float4 & x = src0[tpig];
+    dst[tpig] = select(log(1.0f + exp(x)), x, x > 20.0f);
+}
+
+kernel void kernel_expm1_f32(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = exp(src0[tpig]) - 1.0f;
+}
+
+kernel void kernel_expm1_f32_4(
+        device const float4 * src0,
+        device       float4 * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = exp(src0[tpig]) - 1.0f;
+}
+
 kernel void kernel_reglu_f32(
         constant ggml_metal_kargs_glu & args,
         device const char * src0,
@@ -4720,6 +4790,75 @@ typedef decltype(kernel_cumsum_add<float>) kernel_cumsum_add_t;
 
 template [[host_name("kernel_cumsum_add_f32")]] kernel kernel_cumsum_add_t kernel_cumsum_add<float>;
 
+
+template<uint32_t ttype>
+bool _ggml_vec_tri_cmp(const int i, const int r);
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER */ 3>(const int i, const int r) {
+    return i < r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_LOWER_DIAG */ 2>(const int i, const int r) {
+    return i <= r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER */ 1>(const int i, const int r) {
+    return i > r;
+}
+
+template<>
+bool _ggml_vec_tri_cmp</* GGML_TRI_TYPE_UPPER_DIAG */ 0>(const int i, const int r) {
+    return i >= r;
+}
+
+template<typename T, int ttype>
+kernel void kernel_tri(
+        constant ggml_metal_kargs_tri & args,
+        device const char * src0,
+        device const char * dst,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    const int i3 = tgpig.z;
+    const int i2 = tgpig.y;
+    const int i1 = tgpig.x;
+
+    if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
+        return;
+    }
+
+    device const T * src_row = (device const T *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
+    device       T * dst_row = (device       T *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
+
+    // Each thread is a single element of the row if ne00 < max threads per
+    // threadgroup, so this will loop once for each index that this thread is
+    // responsible for
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        // Use the comparison as a mask for branchless
+        dst_row[i0] = static_cast<T>(_ggml_vec_tri_cmp<ttype>(i0, i1)) * src_row[i0];
+    }
+}
+
+typedef decltype(kernel_tri<float, 0>) kernel_tri_t;
+
+template [[host_name("kernel_tri_f32_0")]] kernel kernel_tri_t kernel_tri<float, 0>;
+template [[host_name("kernel_tri_f32_1")]] kernel kernel_tri_t kernel_tri<float, 1>;
+template [[host_name("kernel_tri_f32_2")]] kernel kernel_tri_t kernel_tri<float, 2>;
+template [[host_name("kernel_tri_f32_3")]] kernel kernel_tri_t kernel_tri<float, 3>;
+template [[host_name("kernel_tri_f16_0")]] kernel kernel_tri_t kernel_tri<half, 0>;
+template [[host_name("kernel_tri_f16_1")]] kernel kernel_tri_t kernel_tri<half, 1>;
+template [[host_name("kernel_tri_f16_2")]] kernel kernel_tri_t kernel_tri<half, 2>;
+template [[host_name("kernel_tri_f16_3")]] kernel kernel_tri_t kernel_tri<half, 3>;
+#if defined(GGML_METAL_HAS_BF16)
+template [[host_name("kernel_tri_bf16_0")]] kernel kernel_tri_t kernel_tri<bfloat, 0>;
+template [[host_name("kernel_tri_bf16_1")]] kernel kernel_tri_t kernel_tri<bfloat, 1>;
+template [[host_name("kernel_tri_bf16_2")]] kernel kernel_tri_t kernel_tri<bfloat, 2>;
+template [[host_name("kernel_tri_bf16_3")]] kernel kernel_tri_t kernel_tri<bfloat, 3>;
+#endif
+
 template<typename T>
 kernel void kernel_soft_max(
         constant ggml_metal_kargs_soft_max & args,
@@ -5005,7 +5144,102 @@ kernel void kernel_ssm_conv_f32_f32_4(
     x[0] = sumf;
 }
 
+constant short FC_ssm_conv_bs   [[function_constant(FC_SSM_CONV + 0)]];
+
+// Batched version: each threadgroup processes multiple tokens for better efficiency
+// Thread layout: each thread handles one token, threadgroup covers BATCH_SIZE tokens
+kernel void kernel_ssm_conv_f32_f32_batched(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float * c = (device const float *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float * s = (device const float *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_batched_4(
+        constant ggml_metal_kargs_ssm_conv & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    // tgpig.x = row index (ir)
+    // tgpig.y = batch of tokens (i2_base / BATCH_SIZE)
+    // tgpig.z = sequence index (i3)
+    // tpitg.x = thread within batch (0..BATCH_SIZE-1)
+    const short BATCH_SIZE = FC_ssm_conv_bs;
+
+    const int64_t ir      = tgpig.x;
+    const int64_t i2_base = tgpig.y * BATCH_SIZE;
+    const int64_t i3      = tgpig.z;
+    const int64_t i2_off  = tpitg.x;
+    const int64_t i2      = i2_base + i2_off;
+
+    const int64_t nc  = args.ne10;  // conv kernel size (typically 4)
+    const int64_t n_t = args.ne1;   // number of tokens
+
+    // Bounds check for partial batches at the end
+    if (i2 >= n_t) {
+        return;
+    }
+
+    // Load conv weights (shared across all tokens for this row)
+    device const float4 * c = (device const float4 *) ((device const char *) src1 + ir*args.nb11);
+
+    // Load source for this specific token
+    device const float4 * s = (device const float4 *) ((device const char *) src0 + ir*args.nb01 + i2*args.nb00 + i3*args.nb02);
+
+    // Output location for this token
+    device float * x = (device float *) ((device char *) dst + ir*args.nb0 + i2*args.nb1 + i3*args.nb2);
+
+    float sumf = 0.0f;
+    for (int64_t i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    x[0] = sumf;
+}
+
 // ref: ggml.c:ggml_compute_forward_ssm_scan_f32, Mamba-2 part
+// Optimized version: reduces redundant memory loads by having one thread load shared values
 kernel void kernel_ssm_scan_f32(
         constant ggml_metal_kargs_ssm_scan & args,
         device const void * src0,
@@ -5025,7 +5259,15 @@ kernel void kernel_ssm_scan_f32(
         uint3    tgpg[[threadgroups_per_grid]]) {
     constexpr short NW = N_SIMDWIDTH;
 
-    shared[tpitg.x] = 0.0f;
+    // Shared memory layout:
+    // [0..sgptg*NW-1]: partial sums for reduction (existing)
+    // [sgptg*NW..sgptg*NW+sgptg-1]: pre-computed x_dt values for each token in batch
+    // [sgptg*NW+sgptg..sgptg*NW+2*sgptg-1]: pre-computed dA values for each token in batch
+    threadgroup float * shared_sums = shared;
+    threadgroup float * shared_x_dt = shared + sgptg * NW;
+    threadgroup float * shared_dA   = shared + sgptg * NW + sgptg;
+
+    shared_sums[tpitg.x] = 0.0f;
 
     const int32_t i0 = tpitg.x;
     const int32_t i1 = tgpig.x;
@@ -5065,32 +5307,47 @@ kernel void kernel_ssm_scan_f32(
     for (int i2 = 0; i2 < n_t; i2 += sgptg) {
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        for (int t = 0; t < sgptg && i2 + t < n_t; t++) {
-            const float dt0  = dt[0];
+        // Pre-compute x_dt and dA for this batch of tokens
+        // Only first sgptg threads do the loads and expensive math
+        if (i0 < sgptg && i2 + i0 < n_t) {
+            // ns12 and ns21 are element strides (nb12/nb10, nb21/nb20)
+            device const float * x_t  = x  + i0 * args.ns12;
+            device const float * dt_t = dt + i0 * args.ns21;
+
+            const float dt0  = dt_t[0];
             const float dtsp = dt0 <= 20.0f ? log(1.0f + exp(dt0)) : dt0;
-            const float x_dt = x[0] * dtsp;
-            const float dA   = exp(dtsp * A0);
+            shared_x_dt[i0] = x_t[0] * dtsp;
+            shared_dA[i0]   = dtsp;  // Store dtsp, compute exp(dtsp * A0) per-thread since A0 varies
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (int t = 0; t < sgptg && i2 + t < n_t; t++) {
+            const float x_dt = shared_x_dt[t];
+            const float dA   = exp(shared_dA[t] * A0);
 
             s = (s0 * dA) + (B[i0] * x_dt);
 
             const float sumf = simd_sum(s * C[i0]);
 
             if (tiisg == 0) {
-                shared[t*NW + sgitg] = sumf;
+                shared_sums[t*NW + sgitg] = sumf;
             }
 
             // recurse
             s0 = s;
 
-            x  += args.ns12;
-            dt += args.ns21;
             B  += args.ns42;
             C  += args.ns52;
         }
 
+        // Advance pointers for next batch
+        x  += sgptg * args.ns12;
+        dt += sgptg * args.ns21;
+
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
-        const float sumf = simd_sum(shared[sgitg*NW + tiisg]);
+        const float sumf = simd_sum(shared_sums[sgitg*NW + tiisg]);
 
         if (tiisg == 0 && i2 + sgitg < n_t) {
             y[sgitg*nh*nr] = sumf;
@@ -8749,6 +9006,7 @@ typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, hal
 
 template [[host_name("kernel_flash_attn_ext_f32_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  32,  32>;
 template [[host_name("kernel_flash_attn_ext_f32_dk40_dv40"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  40,  40>;
+template [[host_name("kernel_flash_attn_ext_f32_dk48_dv48"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  48,  48>;
 template [[host_name("kernel_flash_attn_ext_f32_dk64_dv64"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  64,  64>;
 template [[host_name("kernel_flash_attn_ext_f32_dk72_dv72"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  72,  72>;
 template [[host_name("kernel_flash_attn_ext_f32_dk80_dv80"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_F32, float4x4,   1, dequantize_f32,  float4x4,   1, dequantize_f32,  80,  80>;
@@ -8762,6 +9020,7 @@ template [[host_name("kernel_flash_attn_ext_f32_dk576_dv512")]]  kernel flash_at
 
 template [[host_name("kernel_flash_attn_ext_f16_dk32_dv32"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  32,  32>;
 template [[host_name("kernel_flash_attn_ext_f16_dk40_dv40"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  40,  40>;
+template [[host_name("kernel_flash_attn_ext_f16_dk48_dv48"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  48,  48>;
 template [[host_name("kernel_flash_attn_ext_f16_dk64_dv64"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  64,  64>;
 template [[host_name("kernel_flash_attn_ext_f16_dk72_dv72"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  72,  72>;
 template [[host_name("kernel_flash_attn_ext_f16_dk80_dv80"  )]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  80,  80>;
@@ -8776,6 +9035,7 @@ template [[host_name("kernel_flash_attn_ext_f16_dk576_dv512")]]  kernel flash_at
 #if defined(GGML_METAL_HAS_BF16)
 template [[host_name("kernel_flash_attn_ext_bf16_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 40,  40>;
+template [[host_name("kernel_flash_attn_ext_bf16_dk48_dv48"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 48,  48>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_bf16_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
@@ -8790,6 +9050,7 @@ template [[host_name("kernel_flash_attn_ext_bf16_dk576_dv512")]] kernel flash_at
 
 template [[host_name("kernel_flash_attn_ext_q4_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 40,  40>;
+template [[host_name("kernel_flash_attn_ext_q4_0_dk48_dv48"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 48,  48>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q4_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 80,  80>;
@@ -8803,6 +9064,7 @@ template [[host_name("kernel_flash_attn_ext_q4_0_dk576_dv512")]] kernel flash_at
 
 template [[host_name("kernel_flash_attn_ext_q4_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 40,  40>;
+template [[host_name("kernel_flash_attn_ext_q4_1_dk48_dv48"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 48,  48>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q4_1_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q4_1, 2, dequantize_q4_1, block_q4_1, 2, dequantize_q4_1, 80,  80>;
@@ -8816,6 +9078,7 @@ template [[host_name("kernel_flash_attn_ext_q4_1_dk576_dv512")]] kernel flash_at
 
 template [[host_name("kernel_flash_attn_ext_q5_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 40,  40>;
+template [[host_name("kernel_flash_attn_ext_q5_0_dk48_dv48"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 48,  48>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q5_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_0, 2, dequantize_q5_0, block_q5_0, 2, dequantize_q5_0, 80,  80>;
@@ -8829,6 +9092,7 @@ template [[host_name("kernel_flash_attn_ext_q5_0_dk576_dv512")]] kernel flash_at
 
 template [[host_name("kernel_flash_attn_ext_q5_1_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 40,  40>;
+template [[host_name("kernel_flash_attn_ext_q5_1_dk48_dv48"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 48,  48>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q5_1_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q5_1, 2, dequantize_q5_1, block_q5_1, 2, dequantize_q5_1, 80,  80>;
@@ -8842,6 +9106,7 @@ template [[host_name("kernel_flash_attn_ext_q5_1_dk576_dv512")]] kernel flash_at
 
 template [[host_name("kernel_flash_attn_ext_q8_0_dk32_dv32"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 32,  32>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk40_dv40"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 40,  40>;
+template [[host_name("kernel_flash_attn_ext_q8_0_dk48_dv48"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 48,  48>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk64_dv64"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 64,  64>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk72_dv72"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 72,  72>;
 template [[host_name("kernel_flash_attn_ext_q8_0_dk80_dv80"  )]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES,    block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 80,  80>;