common : fix noreturn compile warning (#13151)

ggml-ci

.clang-tidy

@@ -13,6 +13,7 @@ Checks: >
     -readability-magic-numbers,
     -readability-uppercase-literal-suffix,
     -readability-simplify-boolean-expr,
+    -readability-math-missing-parentheses,
     clang-analyzer-*,
     -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
     performance-*,

SECURITY.md

@@ -40,7 +40,8 @@ To protect sensitive data from potential leaks or unauthorized access, it is cru
 ### Untrusted environments or networks
 
 If you can't run your models in a secure and isolated environment or if it must be exposed to an untrusted network, make sure to take the following security precautions:
-* Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value
+* Do not use the RPC backend, [rpc-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/rpc) and [llama-server](https://github.com/ggml-org/llama.cpp/tree/master/examples/server) functionality (see https://github.com/ggml-org/llama.cpp/pull/13061).
+* Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value.
 * Encrypt your data if sending it over the network.
 
 ### Multi-Tenant environments

common/arg.cpp

@@ -38,6 +38,11 @@
 
 using json = nlohmann::ordered_json;
 
+std::initializer_list<enum llama_example> mmproj_examples = {
+    LLAMA_EXAMPLE_LLAVA,
+    // TODO: add LLAMA_EXAMPLE_SERVER when it's ready
+};
+
 common_arg & common_arg::set_examples(std::initializer_list<enum llama_example> examples) {
     this->examples = std::move(examples);
     return *this;
@@ -157,6 +162,10 @@ struct common_hf_file_res {
 
 #ifdef LLAMA_USE_CURL
 
+bool common_has_curl() {
+    return true;
+}
+
 #ifdef __linux__
 #include <linux/limits.h>
 #elif defined(_WIN32)
@@ -522,6 +531,50 @@ static bool common_download_model(
     return true;
 }
 
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params) {
+    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
+    curl_slist_ptr http_headers;
+    std::vector<char> res_buffer;
+
+    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
+    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
+    curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
+    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
+    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
+        auto data_vec = static_cast<std::vector<char> *>(data);
+        data_vec->insert(data_vec->end(), (char *)ptr, (char *)ptr + size * nmemb);
+        return size * nmemb;
+    };
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
+    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_buffer);
+#if defined(_WIN32)
+    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
+#endif
+    if (params.timeout > 0) {
+        curl_easy_setopt(curl.get(), CURLOPT_TIMEOUT, params.timeout);
+    }
+    if (params.max_size > 0) {
+        curl_easy_setopt(curl.get(), CURLOPT_MAXFILESIZE, params.max_size);
+    }
+    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
+    for (const auto & header : params.headers) {
+        http_headers.ptr = curl_slist_append(http_headers.ptr, header.c_str());
+    }
+    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+
+    CURLcode res = curl_easy_perform(curl.get());
+
+    if (res != CURLE_OK) {
+        std::string error_msg = curl_easy_strerror(res);
+        throw std::runtime_error("error: cannot make GET request: " + error_msg);
+    }
+
+    long res_code;
+    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
+
+    return { res_code, std::move(res_buffer) };
+}
+
 /**
  * Allow getting the HF file from the HF repo with tag (like ollama), for example:
  * - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
@@ -541,45 +594,26 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
         throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
     }
 
-    // fetch model info from Hugging Face Hub API
-    curl_ptr       curl(curl_easy_init(), &curl_easy_cleanup);
-    curl_slist_ptr http_headers;
-    std::string res_str;
+    std::string url = get_model_endpoint() + "v2/" + hf_repo + "/manifests/" + tag;
 
-    std::string model_endpoint = get_model_endpoint();
-
-    std::string url = model_endpoint + "v2/" + hf_repo + "/manifests/" + tag;
-    curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
-    curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
-    typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
-    auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
-        static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
-        return size * nmemb;
-    };
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
-    curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
-#if defined(_WIN32)
-    curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
-#endif
+    // headers
+    std::vector<std::string> headers;
+    headers.push_back("Accept: application/json");
     if (!bearer_token.empty()) {
-        std::string auth_header = "Authorization: Bearer " + bearer_token;
-        http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
+        headers.push_back("Authorization: Bearer " + bearer_token);
     }
     // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
-    http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
-    curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
+    // User-Agent header is already set in common_remote_get_content, no need to set it here
 
-    CURLcode res = curl_easy_perform(curl.get());
+    // make the request
+    common_remote_params params;
+    params.headers = headers;
+    auto res = common_remote_get_content(url, params);
+    long res_code = res.first;
+    std::string res_str(res.second.data(), res.second.size());
+    std::string ggufFile;
+    std::string mmprojFile;
 
-    if (res != CURLE_OK) {
-        throw std::runtime_error("error: cannot make GET request to HF API");
-    }
-
-    long res_code;
-    std::string ggufFile   = "";
-    std::string mmprojFile = "";
-    curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
     if (res_code == 200) {
         // extract ggufFile.rfilename in json, using regex
         {
@@ -613,6 +647,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
 
 #else
 
+bool common_has_curl() {
+    return false;
+}
+
 static bool common_download_file_single(const std::string &, const std::string &, const std::string &) {
     LOG_ERR("error: built without CURL, cannot download model from internet\n");
     return false;
@@ -635,17 +673,30 @@ static struct common_hf_file_res common_get_hf_file(const std::string &, const s
     return {};
 }
 
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params &) {
+    if (!url.empty()) {
+        throw std::runtime_error("error: built without CURL, cannot download model from the internet");
+    }
+
+    return {};
+}
+
 #endif // LLAMA_USE_CURL
 
 //
 // utils
 //
 
-static void common_params_handle_model(
+struct handle_model_result {
+    bool found_mmproj = false;
+    common_params_model mmproj;
+};
+
+static handle_model_result common_params_handle_model(
         struct common_params_model & model,
         const std::string & bearer_token,
-        const std::string & model_path_default,
-        bool is_mmproj = false) { // TODO: move is_mmproj to an enum when we have more files?
+        const std::string & model_path_default) {
+    handle_model_result result;
     // handle pre-fill default model path and url based on hf_repo and hf_file
     {
         if (!model.hf_repo.empty()) {
@@ -657,7 +708,12 @@ static void common_params_handle_model(
                         exit(1); // built without CURL, error message already printed
                     }
                     model.hf_repo = auto_detected.repo;
-                    model.hf_file = is_mmproj ? auto_detected.mmprojFile : auto_detected.ggufFile;
+                    model.hf_file = auto_detected.ggufFile;
+                    if (!auto_detected.mmprojFile.empty()) {
+                        result.found_mmproj   = true;
+                        result.mmproj.hf_repo = model.hf_repo;
+                        result.mmproj.hf_file = auto_detected.mmprojFile;
+                    }
                 } else {
                     model.hf_file = model.path;
                 }
@@ -694,6 +750,8 @@ static void common_params_handle_model(
             exit(1);
         }
     }
+
+    return result;
 }
 
 const std::vector<ggml_type> kv_cache_types = {
@@ -827,16 +885,25 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
-    common_params_handle_model(params.model,             params.hf_token, DEFAULT_MODEL_PATH);
-    common_params_handle_model(params.speculative.model, params.hf_token, "");
-    common_params_handle_model(params.vocoder.model,     params.hf_token, "");
-
-    // allow --mmproj to be set from -hf
-    // assuming that mmproj is always in the same repo as text model
-    if (!params.model.hf_repo.empty() && ctx_arg.ex == LLAMA_EXAMPLE_LLAVA) {
-        params.mmproj.hf_repo = params.model.hf_repo;
+    // handle model and download
+    {
+        auto res = common_params_handle_model(params.model, params.hf_token, DEFAULT_MODEL_PATH);
+        if (params.no_mmproj) {
+            params.mmproj = {};
+        } else if (res.found_mmproj && params.mmproj.path.empty() && params.mmproj.url.empty()) {
+            // optionally, handle mmproj model when -hf is specified
+            params.mmproj = res.mmproj;
+        }
+        // only download mmproj if the current example is using it
+        for (auto & ex : mmproj_examples) {
+            if (ctx_arg.ex == ex) {
+                common_params_handle_model(params.mmproj,    params.hf_token, "");
+                break;
+            }
+        }
+        common_params_handle_model(params.speculative.model, params.hf_token, "");
+        common_params_handle_model(params.vocoder.model,     params.hf_token, "");
     }
-    common_params_handle_model(params.mmproj,            params.hf_token, "", true);
 
     if (params.escape) {
         string_process_escapes(params.prompt);
@@ -968,7 +1035,6 @@ static void common_params_print_completion(common_params_context & ctx_arg) {
         "llama-embedding",
         "llama-eval-callback",
         "llama-export-lora",
-        "llama-gbnf-validator",
         "llama-gen-docs",
         "llama-gguf",
         "llama-gguf-hash",
@@ -976,20 +1042,18 @@ static void common_params_print_completion(common_params_context & ctx_arg) {
         "llama-gritlm",
         "llama-imatrix",
         "llama-infill",
-        "llama-llava-cli",
+        "llama-mtmd-cli",
         "llama-llava-clip-quantize-cli",
         "llama-lookahead",
         "llama-lookup",
         "llama-lookup-create",
         "llama-lookup-merge",
         "llama-lookup-stats",
-        "llama-minicpmv-cli",
         "llama-parallel",
         "llama-passkey",
         "llama-perplexity",
         "llama-q8dot",
         "llama-quantize",
-        "llama-quantize-stats",
         "llama-qwen2vl-cli",
         "llama-retrieval",
         "llama-run",
@@ -2096,18 +2160,32 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_NO_CONT_BATCHING"));
     add_opt(common_arg(
         {"--mmproj"}, "FILE",
-        "path to a multimodal projector file for LLaVA. see examples/llava/README.md",
+        "path to a multimodal projector file. see examples/llava/README.md",
         [](common_params & params, const std::string & value) {
             params.mmproj.path = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    ).set_examples(mmproj_examples));
     add_opt(common_arg(
         {"--mmproj-url"}, "URL",
-        "URL to a multimodal projector file for LLaVA. see examples/llava/README.md",
+        "URL to a multimodal projector file. see examples/llava/README.md",
         [](common_params & params, const std::string & value) {
             params.mmproj.url = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_LLAVA}));
+    ).set_examples(mmproj_examples));
+    add_opt(common_arg(
+        {"--no-mmproj"},
+        "explicitly disable multimodal projector, useful when using -hf",
+        [](common_params & params) {
+            params.no_mmproj = true;
+        }
+    ).set_examples(mmproj_examples));
+    add_opt(common_arg(
+        {"--no-mmproj-offload"},
+        "do not offload multimodal projector to GPU",
+        [](common_params & params) {
+            params.mmproj_use_gpu = false;
+        }
+    ).set_examples(mmproj_examples));
     add_opt(common_arg(
         {"--image"}, "FILE",
         "path to an image file. use with multimodal models. Specify multiple times for batching",
@@ -2382,6 +2460,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"-hf", "-hfr", "--hf-repo"}, "<user>/<model>[:quant]",
         "Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n"
+        "mmproj is also downloaded automatically if available. to disable, add --no-mmproj\n"
         "example: unsloth/phi-4-GGUF:q4_k_m\n"
         "(default: unused)",
         [](common_params & params, const std::string & value) {

common/arg.h

@@ -78,3 +78,12 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
 
 // function to be used by test-arg-parser
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
+bool common_has_curl();
+
+struct common_remote_params {
+    std::vector<std::string> headers;
+    long timeout = 0; // CURLOPT_TIMEOUT, in seconds ; 0 means no timeout
+    long max_size = 0; // max size of the response ; unlimited if 0 ; max is 2GB
+};
+// get remote file content, returns <http_code, raw_response_body>
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params);

common/common.h

@@ -342,6 +342,8 @@ struct common_params {
 
     // multimodal models (see examples/llava)
     struct common_params_model mmproj;
+    bool mmproj_use_gpu = true;     // use GPU for multimodal model
+    bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
 
     // embedding

common/json-schema-to-grammar.cpp

@@ -16,6 +16,9 @@ using json = nlohmann::ordered_json;
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
 
+    if (max_items == 0) {
+        return "";
+    }
     if (min_items == 0 && max_items == 1) {
         return item_rule + "?";
     }

convert_hf_to_gguf.py

@@ -419,8 +419,12 @@ class ModelBase:
     def load_hparams(dir_model: Path):
         with open(dir_model / "config.json", "r", encoding="utf-8") as f:
             hparams = json.load(f)
+            architectures = hparams.get("architectures")
             if "text_config" in hparams:
                 hparams = {**hparams, **hparams["text_config"]}
+            if architectures is not None:
+                # preserve "architectures" from root level config
+                hparams["architectures"] = architectures
             return hparams
 
     @classmethod
@@ -772,6 +776,9 @@ class TextModel(ModelBase):
         if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
             # ref: https://huggingface.co/THUDM/glm-4-9b-hf
             res = "glm4"
+        if chkhsh == "0e9433cbbb161f89e264eb32e8e64bfe69e834973ffca5d41d3948a604a3e2a3":
+            # ref: https://huggingface.co/mistral-community/pixtral-12b
+            res = "pixtral"
 
         if res is None:
             logger.warning("\n")
@@ -1061,6 +1068,8 @@ class TextModel(ModelBase):
 class VisionModel(ModelBase):
     model_arch = gguf.MODEL_ARCH.CLIP_VISION
     n_text_embd = 0
+    preprocessor_config: dict[str, Any]
+    global_config: dict[str, Any]
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
@@ -1075,24 +1084,33 @@ class VisionModel(ModelBase):
 
         if "vision_config" not in self.hparams:
             raise ValueError("vision_config not found in hparams")
-        # move vision config to the top level
+        # move vision config to the top level, while preserving the original hparams in global_config
+        self.global_config = self.hparams
         self.hparams = self.hparams["vision_config"]
 
+        # load preprocessor config
+        with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
+            self.preprocessor_config = json.load(f)
+
     def set_type(self):
         self.gguf_writer.add_type(gguf.GGUFType.CLIP_VISION)
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_file_type(self.ftype)
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.PROJECTION_DIM, self.n_embd_text)
-        self.gguf_writer.add_bool(gguf.Keys.ClipVision.HAS_VISION_ENCODER, True)
+        self.gguf_writer.add_vision_projection_dim(self.n_embd_text)
+        self.gguf_writer.add_vision_has_vision_encoder(True)
 
         # vision config
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.IMAGE_SIZE,           self.find_hparam(["image_size"]))
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.PATCH_SIZE,           self.find_hparam(["patch_size"]))
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.EMBEDDING_LENGTH,     self.find_hparam(["hidden_size"]))
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.FEED_FORWARD_LENGTH,  self.find_hparam(["intermediate_size"]))
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.BLOCK_COUNT,          self.find_hparam(["num_hidden_layers"]))
-        self.gguf_writer.add_uint32(gguf.Keys.ClipVision.Attention.HEAD_COUNT, self.find_hparam(["num_attention_heads"]))
+        self.gguf_writer.add_vision_image_size(self.find_hparam(["image_size"]))
+        self.gguf_writer.add_vision_patch_size(self.find_hparam(["patch_size"]))
+        self.gguf_writer.add_vision_embedding_length(self.find_hparam(["hidden_size"]))
+        self.gguf_writer.add_vision_feed_forward_length(self.find_hparam(["intermediate_size"]))
+        self.gguf_writer.add_vision_block_count(self.find_hparam(["num_hidden_layers"]))
+        self.gguf_writer.add_vision_head_count(self.find_hparam(["num_attention_heads"]))
+
+        # preprocessor config
+        self.gguf_writer.add_vision_image_mean(self.preprocessor_config["image_mean"])
+        self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_mean"])
 
     def write_vocab(self):
         raise ValueError("VisionModel does not support vocab writing")
@@ -1703,11 +1721,28 @@ class StableLMModel(TextModel):
                 raise ValueError(f"Unprocessed norms: {norms}")
 
 
-@ModelBase.register("LLaMAForCausalLM", "LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
+@ModelBase.register(
+    "LLaMAForCausalLM",
+    "LlamaForCausalLM",
+    "MistralForCausalLM",
+    "MixtralForCausalLM",
+    "Idefics3ForConditionalGeneration",
+    "SmolVLMForConditionalGeneration",
+    "LlavaForConditionalGeneration")
 class LlamaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.LLAMA
     undo_permute = True
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # fix for SmolVLM2, missing `num_attention_heads` in config.json
+        if self.hparams["architectures"][0] == "SmolVLMForConditionalGeneration":
+            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
+        # fix for Pixtral, missing `num_attention_heads` in config.json
+        if self.hparams["architectures"][0] == "LlavaForConditionalGeneration" \
+                and self.hparams.get("model_type") == "mistral":
+            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 32)
+
     def set_vocab(self):
         try:
             self._set_vocab_sentencepiece()
@@ -1770,6 +1805,17 @@ class LlamaModel(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         n_head = self.hparams["num_attention_heads"]
         n_kv_head = self.hparams.get("num_key_value_heads")
+        is_vision_tensor = "vision_tower" in name \
+            or "vision_model" in name \
+            or "model.connector" in name \
+            or "multi_modal_projector" in name
+
+        if is_vision_tensor:
+            return [] # skip vision tensors
+        elif name.startswith("model.text_model"):
+            name = name.replace("text_model.", "") # for SmolVLM
+        elif name.startswith("language_model."):
+            name = name.replace("language_model.", "") # for the rest
 
         if self.undo_permute:
             if name.endswith(("q_proj.weight", "q_proj.bias")):
@@ -1852,6 +1898,90 @@ class LlamaModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@ModelBase.register("LlavaForConditionalGeneration")
+class LlavaVisionModel(VisionModel):
+    img_break_tok_id = -1
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if self.hparams["model_type"] == "pixtral":
+            # fix missing config.json values
+            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 16)
+            self.hparams["num_hidden_layers"] = self.hparams.get("num_hidden_layers", 24)
+            self.hparams["intermediate_size"] = self.hparams.get("intermediate_size", 4096)
+            self.hparams["hidden_size"] = self.hparams.get("hidden_size", 1024)
+            self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5)
+            self.img_break_tok_id = 12 # see tokenizer_config.json
+        else:
+            raise ValueError(f"Unsupported model type: {self.hparams['model_type']}")
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        if hparams["model_type"] == "pixtral":
+            self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.PIXTRAL)
+            # default values below are taken from HF tranformers code
+            self.gguf_writer.add_vision_attention_layernorm_eps(hparams["layer_norm_eps"])
+            self.gguf_writer.add_vision_use_silu(True)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = n_head
+
+        if name.startswith("multi_modal_projector.") or name.startswith("vision_tower."):
+            # process vision tensors
+            if name.endswith(("q_proj.weight", "q_proj.bias")):
+                data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+            if name.endswith(("k_proj.weight", "k_proj.bias")):
+                data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+            return [(self.map_tensor_name(name), data_torch)]
+
+        if self.img_break_tok_id > 0 and "embed_tokens.weight" in name:
+            logger.info(f"Extracting [IMG_BREAK] token embedding from {name}")
+            # for pixtral model, we need to extract the [IMG_BREAK] token embedding
+            img_break_embd = data_torch[self.img_break_tok_id]
+            name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK]
+            return [(self.map_tensor_name(name), img_break_embd)]
+
+        return [] # skip other tensors
+
+
+@ModelBase.register("Idefics3ForConditionalGeneration", "SmolVLMForConditionalGeneration")
+class SmolVLMModel(VisionModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # fix for SmolVLM2, missing some keys in config.json
+        # default values are taken from transformers code
+        if self.hparams["model_type"] == "smolvlm_vision":
+            self.hparams["hidden_size"] = self.hparams.get("hidden_size", 1152)
+            self.hparams["num_attention_heads"] = self.hparams.get("num_attention_heads", 16)
+            self.hparams["intermediate_size"] = self.hparams.get("intermediate_size", 3072)
+            self.hparams["num_hidden_layers"] = self.hparams.get("num_hidden_layers", 12)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.IDEFICS3)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-5))
+        self.gguf_writer.add_vision_projector_scale_factor(self.global_config.get("scale_factor", 2))
+        self.gguf_writer.add_vision_use_gelu(True)
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        del bid, new_name, n_dims  # unused
+        if ".embeddings." in name:
+            return gguf.GGMLQuantizationType.F32
+        return False
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        is_vision_tensor = "vision_tower" in name or "vision_model" in name or "model.connector" in name
+
+        if is_vision_tensor:
+            return [(self.map_tensor_name(name), data_torch)]
+
+        return [] # skip other tensors
+
+
 @ModelBase.register("Llama4ForConditionalGeneration")
 class Llama4Model(LlamaModel):
     model_arch = gguf.MODEL_ARCH.LLAMA4
@@ -2424,11 +2554,12 @@ class Qwen2VLModel(TextModel):
         except FileNotFoundError:
             self._set_vocab_gpt2()
 
-    def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
-        for name, data in super().get_tensors():
-            if name.startswith("visual."):
-                continue
-            yield name, data
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+        if name.startswith("visual."):
+            # skip visual tensors
+            return []
+        return [(self.map_tensor_name(name), data_torch)]
 
 
 @ModelBase.register("WavTokenizerDec")
@@ -3591,12 +3722,10 @@ class Gemma3VisionModel(VisionModel):
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
-        self.gguf_writer.add_string(gguf.Keys.ClipVision.PROJECTOR_TYPE, "gemma3")
+        self.gguf_writer.add_vision_projector_type(gguf.VisionProjectorType.GEMMA3)
         # default values below are taken from HF tranformers code
-        self.gguf_writer.add_float32(gguf.Keys.ClipVision.Attention.LAYERNORM_EPS, hparams.get("layer_norm_eps", 1e-6))
-        self.gguf_writer.add_array(gguf.Keys.ClipVision.IMAGE_MEAN, [0.5, 0.5, 0.5])
-        self.gguf_writer.add_array(gguf.Keys.ClipVision.IMAGE_STD,  [0.5, 0.5, 0.5])
-        self.gguf_writer.add_bool (gguf.Keys.ClipVision.USE_GELU,   True)
+        self.gguf_writer.add_vision_attention_layernorm_eps(hparams.get("layer_norm_eps", 1e-6))
+        self.gguf_writer.add_vision_use_gelu(True)
 
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         del bid, new_name, n_dims  # unused
@@ -3614,10 +3743,6 @@ class Gemma3VisionModel(VisionModel):
                 or name.startswith("multimodal_projector.") or name.startswith("vision_model."):
             # process vision tensors
             name = name.replace("_weight", ".weight")
-            if "fc1" in name:
-                name = name.replace("fc1", "fc2")
-            else:
-                name = name.replace("fc2", "fc1")
 
             # correct norm value ; only this "soft_emb_norm" need to be corrected as it's part of Gemma projector
             # the other norm values are part of SigLIP model, and they are already correct
@@ -5017,10 +5142,25 @@ class Glm4Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GLM4
 
     def set_vocab(self):
-        self._set_vocab_gpt2()
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        tokens, toktypes, tokpre = self.get_vocab_base()
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab._set_special_token("eos", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("eot", tokenizer.get_added_vocab()["<|user|>"])
+        special_vocab._set_special_token("unk", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab._set_special_token("bos", tokenizer.get_added_vocab()["<|endoftext|>"])
+        special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
+        rope_dim = self.hparams["head_dim"]
+        self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
         if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
             if self.hparams["rope_scaling"].get("type") == "yarn":
                 self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)

convert_hf_to_gguf_update.py

@@ -115,6 +115,7 @@ models = [
     {"name": "bailingmoe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", },
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
     {"name": "glm4",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
+    {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
 ]
 
 

docs/multimodal/MobileVLM.md

@@ -9,15 +9,15 @@ The implementation is based on llava, and is compatible with llava and mobileVLM
 Notice: The overall process of model inference for both **MobileVLM** and **MobileVLM_V2** models is the same, but the process of model conversion is a little different. Therefore, using **MobileVLM-1.7B** as an example, the different conversion step will be shown.
 
 ## Usage
-Build with cmake or run `make llama-llava-cli` to build it.
 
-After building, run: `./llama-llava-cli` to see the usage. For example:
+Build the `llama-mtmd-cli` binary.
+
+After building, run: `./llama-mtmd-cli` to see the usage. For example:
 
 ```sh
-./llama-llava-cli -m MobileVLM-1.7B/ggml-model-q4_k.gguf \
+./llama-mtmd-cli -m MobileVLM-1.7B/ggml-model-q4_k.gguf \
     --mmproj MobileVLM-1.7B/mmproj-model-f16.gguf \
-    --image path/to/an/image.jpg \
-    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWho is the author of this book? Answer the question using a single word or phrase. ASSISTANT:"
+    --chat-template deepseek
 ```
 
 ## Model conversion
@@ -82,7 +82,7 @@ refer to `android/adb_run.sh`, modify resources' `name` and `path`
 ### case 1
 **input**
 ```sh
-/data/local/tmp/llama-llava-cli \
+/data/local/tmp/llama-mtmd-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -t 4 \
@@ -102,7 +102,7 @@ llama_print_timings:       total time =   34731.93 ms
 ### case 2
 **input**
 ```sh
-/data/local/tmp/llama-llava-cli \
+/data/local/tmp/llama-mtmd-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -t 4 \
@@ -123,10 +123,10 @@ llama_print_timings:       total time =   34570.79 ms
 
 ## Some result on Android with `Snapdragon 778G` chip
 ### MobileVLM-1.7B case
-#### llava-cli release-b2005
+#### mtmd-cli release-b2005
 **input**
 ```sh
-/data/local/tmp/llama-llava-cli \
+/data/local/tmp/llama-mtmd-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -t 4 \
@@ -147,7 +147,7 @@ llama_print_timings: prompt eval time =    8119.49 ms /   191 tokens (   42.51 m
 llama_print_timings:        eval time =    1005.75 ms /    14 runs   (   71.84 ms per token,    13.92 tokens per second)
 llama_print_timings:       total time =   28038.34 ms /   205 tokens
 ```
-#### llava-cli latest-version
+#### mtmd-cli latest-version
 **input**
 
 Just the same as above.
@@ -169,7 +169,7 @@ llama_print_timings:        eval time =   43894.02 ms /    13 runs   ( 3376.46 m
 llama_print_timings:       total time =  865441.76 ms /   204 tokens
 ```
 ### MobileVLM_V2-1.7B case
-#### llava-cli release-2005b
+#### mtmd-cli release-2005b
 **input**
 
 Just the same as above.
@@ -200,7 +200,7 @@ make GGML_CUDA=1 CUDA_DOCKER_ARCH=sm_87 GGML_CUDA_F16=1 -j 32
 ### case 1
 **input**
 ```sh
-./llama-llava-cli \
+./llama-mtmd-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     --image /data/local/tmp/demo.jpeg \
@@ -224,7 +224,7 @@ llama_print_timings:       total time =    1352.63 ms /   252 tokens
 ### case 2
 **input**
 ```sh
-./llama-llava-cli \
+./llama-mtmd-cli \
     -m /data/local/tmp/ggml-model-q4_k.gguf \
     --mmproj /data/local/tmp/mmproj-model-f16.gguf \
     -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:" \

docs/multimodal/gemma3.md

@@ -11,26 +11,27 @@ You can use pre-quantized model from [ggml-org](https://huggingface.co/ggml-org)
 ```bash
 # build
 cmake -B build
-cmake --build build --target llama-gemma3-cli
+cmake --build build --target llama-mtmd-cli
 
 # alternatively, install from brew (MacOS)
 brew install llama.cpp
 
 # run it
-llama-gemma3-cli -hf ggml-org/gemma-3-4b-it-GGUF
-llama-gemma3-cli -hf ggml-org/gemma-3-12b-it-GGUF
-llama-gemma3-cli -hf ggml-org/gemma-3-27b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF
 
 # note: 1B model does not support vision
 ```
 
 ## How to get mmproj.gguf?
 
+Simply to add `--mmproj` in when converting model via `convert_hf_to_gguf.py`:
+
 ```bash
 cd gemma-3-4b-it
-python ../llama.cpp/examples/llava/gemma3_convert_encoder_to_gguf.py .
-
-# output file is mmproj.gguf
+python ../llama.cpp/convert_hf_to_gguf.py --outfile model.gguf --outtype f16 --mmproj .
+# output file: mmproj-model.gguf
 ```
 
 ## How to run it?
@@ -43,8 +44,8 @@ What you need:
 ```bash
 # build
 cmake -B build
-cmake --build build --target llama-gemma3-cli
+cmake --build build --target llama-mtmd-cli
 
 # run it
-./build/bin/llama-gemma3-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
+./build/bin/llama-mtmd-cli -m {text_model}.gguf --mmproj mmproj.gguf --image your_image.jpg
 ```

docs/multimodal/glmedge.md

@@ -3,12 +3,12 @@
 Currently this implementation supports [glm-edge-v-2b](https://huggingface.co/THUDM/glm-edge-v-2b) and [glm-edge-v-5b](https://huggingface.co/THUDM/glm-edge-v-5b).
 
 ## Usage
-Build with cmake or run `make llama-llava-cli` to build it.
+Build the `llama-mtmd-cli` binary.
 
-After building, run: `./llama-llava-cli` to see the usage. For example:
+After building, run: `./llama-mtmd-cli` to see the usage. For example:
 
 ```sh
-./llama-llava-cli -m model_path/ggml-model-f16.gguf --mmproj model_path/mmproj-model-f16.gguf --image img_path/image.jpg -p "<|system|>\n system prompt <image><|user|>\n prompt <|assistant|>\n"
+./llama-mtmd-cli -m model_path/ggml-model-f16.gguf --mmproj model_path/mmproj-model-f16.gguf
 ```
 
 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.

docs/multimodal/granitevision.md

@@ -176,15 +176,11 @@ Note that currently you cannot quantize the visual encoder because granite visio
 
 
 ### 5. Running the Model in Llama cpp
-Build llama cpp normally; you should have a target binary named `llama-llava-cli`, which you can pass two binaries to. As an example, we pass the the llama.cpp banner.
+Build llama cpp normally; you should have a target binary named `llama-mtmd-cli`, which you can pass two binaries to. As an example, we pass the the llama.cpp banner.
 
 ```bash
-$ ./build/bin/llama-llava-cli -m $LLM_GGUF_PATH \
+$ ./build/bin/llama-mtmd-cli -m $LLM_GGUF_PATH \
     --mmproj $VISUAL_GGUF_PATH \
-    --image ./media/llama0-banner.png \
     -c 16384 \
-    -p "<|system|>\nA chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions.\n<|user|>\n\<image>\nWhat does the text in this image say?\n<|assistant|>\n" \
     --temp 0
 ```
-
-Sample output: `The text in the image reads "LLAMA C++ Can it run DOOM Llama?"`

docs/multimodal/llava.md

@@ -0,0 +1,143 @@
+# LLaVA
+
+Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants,
+as well as llava-1.6 [llava-v1.6](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2) variants.
+
+The pre-converted [7b](https://huggingface.co/mys/ggml_llava-v1.5-7b)
+and [13b](https://huggingface.co/mys/ggml_llava-v1.5-13b)
+models are available.
+For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](https://huggingface.co/cmp-nct/llava-1.6-gguf)
+
+After API is confirmed, more models will be supported / uploaded.
+
+## Usage
+Build the `llama-mtmd-cli` binary.
+
+After building, run: `./llama-mtmd-cli` to see the usage. For example:
+
+```sh
+./llama-mtmd-cli -m ../llava-v1.5-7b/ggml-model-f16.gguf \
+    --mmproj ../llava-v1.5-7b/mmproj-model-f16.gguf \
+    --chat-template vicuna
+```
+
+**note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
+**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
+
+## LLaVA 1.5
+
+1. Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
+
+```sh
+git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
+
+git clone https://huggingface.co/openai/clip-vit-large-patch14-336
+```
+
+2. Install the required Python packages:
+
+```sh
+pip install -r examples/llava/requirements.txt
+```
+
+3. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
+
+```sh
+python ./examples/llava/llava_surgery.py -m ../llava-v1.5-7b
+```
+
+4. Use `convert_image_encoder_to_gguf.py` to convert the LLaVA image encoder to GGUF:
+
+```sh
+python ./examples/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
+```
+
+5. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
+
+```sh
+python ./examples/convert_legacy_llama.py ../llava-v1.5-7b --skip-unknown
+```
+
+Now both the LLaMA part and the image encoder are in the `llava-v1.5-7b` directory.
+
+## LLaVA 1.6 gguf conversion
+1) First clone a LLaVA 1.6 model:
+```console
+git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
+```
+
+2) Install the required Python packages:
+
+```sh
+pip install -r examples/llava/requirements.txt
+```
+
+3) Use `llava_surgery_v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
+```console
+python examples/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/
+```
+- you will find a llava.projector and a llava.clip file in your model directory
+
+4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
+```console
+mkdir vit
+cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
+cp ../llava-v1.6-vicuna-7b/llava.projector vit/
+curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
+```
+
+5) Create the visual gguf model:
+```console
+python ./examples/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
+```
+- This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
+
+6) Then convert the model to gguf format:
+```console
+python ./examples/convert_legacy_llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown
+```
+
+7) And finally we can run the llava cli using the 1.6 model version:
+```console
+./llama-mtmd-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf
+```
+
+**note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
+
+**note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
+
+**note** if the language model in step `6)` is incompatible with the legacy conversion script, the easiest way handle the LLM model conversion is to load the model in transformers, and export only the LLM from the llava next model.
+
+```python
+import os
+import transformers
+
+model_path = ...
+llm_export_path = ...
+
+tokenizer = transformers.AutoTokenizer.from_pretrained(model_path)
+model = transformers.AutoModelForImageTextToText.from_pretrained(model_path)
+
+tokenizer.save_pretrained(llm_export_path)
+model.language_model.save_pretrained(llm_export_path)
+```
+
+Then, you can convert the LLM using the `convert_hf_to_gguf.py` script, which handles more LLM architectures.
+
+## Chat template
+
+For llava-1.5 and llava-1.6, you need to use `vicuna` chat template. Simply add `--chat-template vicuna` to activate this template.
+
+
+## How to know if you are running in llava-1.5 or llava-1.6 mode
+
+When running llava-cli you will see a visual information right before the prompt is being processed:
+
+**Llava-1.5:**
+`encode_image_with_clip: image embedding created: 576 tokens`
+
+**Llava-1.6 (anything above 576):**
+`encode_image_with_clip: image embedding created: 2880 tokens`
+
+
+Alternatively just pay notice to how many "tokens" have been used for your prompt, it will also show 1000+ tokens for llava-1.6

docs/multimodal/minicpmo2.6.md

@@ -40,9 +40,9 @@ python ./convert_hf_to_gguf.py ../MiniCPM-o-2_6/model
 
 Inference on Linux or Mac
 ```bash
-# run f16 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-o-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
 
-# run quantized int4 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-o-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-o-2_6/mmproj-model-f16.gguf
 ```

docs/multimodal/minicpmv2.5.md

@@ -39,9 +39,9 @@ python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model
 
 Inference on Linux or Mac
 ```bash
-# run f16 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
 
-# run quantized int4 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf
 ```

docs/multimodal/minicpmv2.6.md

@@ -39,9 +39,9 @@ python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model
 
 Inference on Linux or Mac
 ```bash
-# run f16 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
 
-# run quantized int4 version
-./build/bin/llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf
 ```

examples/CMakeLists.txt

@@ -21,11 +21,6 @@ else()
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
 
-    if (NOT WIN32)
-        # disabled on Windows because it uses internal functions not exported with LLAMA_API
-        add_subdirectory(gbnf-validator)
-    endif()
-
     add_subdirectory(gguf-hash)
     add_subdirectory(gguf-split)
     add_subdirectory(gguf)
@@ -58,10 +53,6 @@ else()
         add_subdirectory(convert-llama2c-to-ggml)
         add_subdirectory(cvector-generator)
         add_subdirectory(export-lora)
-        if (NOT WIN32)
-            # disabled on Windows because it uses internal functions not exported with LLAMA_API
-            add_subdirectory(quantize-stats)
-        endif()
         add_subdirectory(llava)
         if (GGML_RPC)
             add_subdirectory(rpc)

examples/embedding/embedding.cpp

@@ -89,6 +89,13 @@ int main(int argc, char ** argv) {
     common_init();
 
     params.embedding = true;
+
+    // utilize the full context
+    if (params.n_batch < params.n_ctx) {
+        LOG_WRN("%s: setting batch size to %d\n", __func__, params.n_ctx);
+        params.n_batch = params.n_ctx;
+    }
+
     // For non-causal models, batch size must be equal to ubatch size
     params.n_ubatch = params.n_batch;
 
@@ -134,7 +141,6 @@ int main(int argc, char ** argv) {
 
     // max batch size
     const uint64_t n_batch = params.n_batch;
-    GGML_ASSERT(params.n_batch >= params.n_ctx);
 
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;

examples/gbnf-validator/CMakeLists.txt

@@ -1,5 +0,0 @@
-set(TARGET llama-gbnf-validator)
-add_executable(${TARGET} gbnf-validator.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/json_schema_to_grammar.py

@@ -10,6 +10,9 @@ from typing import Any, List, Optional, Set, Tuple, Union
 
 def _build_repetition(item_rule, min_items, max_items, separator_rule=None):
 
+    if max_items == 0:
+        return ""
+
     if min_items == 0 and max_items == 1:
         return f'{item_rule}?'
 

examples/llama-bench/llama-bench.cpp

@@ -36,6 +36,46 @@ static uint64_t get_time_ns() {
     return std::chrono::nanoseconds(clock::now().time_since_epoch()).count();
 }
 
+static bool tensor_buft_override_equal(const llama_model_tensor_buft_override& a, const llama_model_tensor_buft_override& b) {
+    if (a.pattern != b.pattern) {
+        // cString comparison that may be null
+        if (a.pattern == nullptr || b.pattern == nullptr) {
+            return false;
+        }
+        if (strcmp(a.pattern, b.pattern) != 0) {
+            return false;
+        }
+    }
+    if (a.buft != b.buft) {
+        return false;
+    }
+    return true;
+}
+
+static bool vec_tensor_buft_override_equal(const std::vector<llama_model_tensor_buft_override>& a, const std::vector<llama_model_tensor_buft_override>& b) {
+    if (a.size() != b.size()) {
+        return false;
+    }
+    for (size_t i = 0; i < a.size(); i++) {
+        if (!tensor_buft_override_equal(a[i], b[i])) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static bool vec_vec_tensor_buft_override_equal(const std::vector<std::vector<llama_model_tensor_buft_override>>& a, const std::vector<std::vector<llama_model_tensor_buft_override>>& b) {
+    if (a.size() != b.size()) {
+        return false;
+    }
+    for (size_t i = 0; i < a.size(); i++) {
+        if (!vec_tensor_buft_override_equal(a[i], b[i])) {
+            return false;
+        }
+    }
+    return true;
+}
+
 template <class T> static std::string join(const std::vector<T> & values, const std::string & delim) {
     std::ostringstream str;
     for (size_t i = 0; i < values.size(); i++) {
@@ -175,6 +215,7 @@ struct cmd_params {
     std::vector<bool>                no_kv_offload;
     std::vector<bool>                flash_attn;
     std::vector<std::vector<float>>  tensor_split;
+    std::vector<std::vector<llama_model_tensor_buft_override>> tensor_buft_overrides;
     std::vector<bool>                use_mmap;
     std::vector<bool>                embeddings;
     ggml_numa_strategy               numa;
@@ -207,6 +248,7 @@ static const cmd_params cmd_params_defaults = {
     /* no_kv_offload        */ { false },
     /* flash_attn           */ { false },
     /* tensor_split         */ { std::vector<float>(llama_max_devices(), 0.0f) },
+    /* tensor_buft_overrides*/ { std::vector<llama_model_tensor_buft_override>{{nullptr,nullptr}} },
     /* use_mmap             */ { true },
     /* embeddings           */ { false },
     /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
@@ -265,6 +307,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -embd, --embeddings <0|1>                 (default: %s)\n",
            join(cmd_params_defaults.embeddings, ",").c_str());
     printf("  -ts, --tensor-split <ts0/ts1/..>          (default: 0)\n");
+    printf("  -ot --override-tensors <tensor name pattern>=<buffer type>;... (default: disabled)\n");
     printf("  -r, --repetitions <n>                     (default: %d)\n", cmd_params_defaults.reps);
     printf("  --prio <0|1|2|3>                          (default: %d)\n", cmd_params_defaults.prio);
     printf("  --delay <0...N> (seconds)                 (default: %d)\n", cmd_params_defaults.delay);
@@ -557,6 +600,87 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 params.tensor_split.push_back(tensor_split);
             }
+        } else if (arg == "-ot" || arg == "--override-tensor") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto value = argv[i];
+            /* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list;
+            if (buft_list.empty()) {
+                // enumerate all the devices and add their buffer types to the list
+                for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+                    auto * dev = ggml_backend_dev_get(i);
+                    auto * buft = ggml_backend_dev_buffer_type(dev);
+                    if (buft) {
+                        buft_list[ggml_backend_buft_name(buft)] = buft;
+                    }
+                }
+            }
+            auto override_group_span_len = std::strcspn(value, ",");
+            bool last_group = false;
+            do {
+                if (override_group_span_len == 0) {
+                    // Adds an empty override-tensors for an empty span
+                    params.tensor_buft_overrides.push_back({{}});
+                    if (value[override_group_span_len] == '\0') {
+                        value = &value[override_group_span_len];
+                        last_group = true;
+                    } else {
+                        value = &value[override_group_span_len + 1];
+                        override_group_span_len = std::strcspn(value, ",");
+                    }
+                    continue;
+                }
+                // Stamps null terminators into the argv
+                // value for this option to avoid the
+                // memory leak present in the implementation
+                // over in arg.cpp. Acceptable because we
+                // only parse these args once in this program.
+                auto override_group = value;
+                if (value[override_group_span_len] == '\0') {
+                    value = &value[override_group_span_len];
+                    last_group = true;
+                } else {
+                    value[override_group_span_len] = '\0';
+                    value = &value[override_group_span_len + 1];
+                }
+                std::vector<llama_model_tensor_buft_override> group_tensor_buft_overrides{};
+                auto override_span_len = std::strcspn(override_group, ";");
+                while (override_span_len > 0) {
+                    auto override = override_group;
+                    if (override_group[override_span_len] != '\0') {
+                        override_group[override_span_len] = '\0';
+                        override_group = &override_group[override_span_len + 1];
+                    } else {
+                        override_group = &override_group[override_span_len];
+                    }
+                    auto tensor_name_span_len = std::strcspn(override, "=");
+                    if (tensor_name_span_len >= override_span_len) {
+                        invalid_param = true;
+                        break;
+                    }
+                    override[tensor_name_span_len] = '\0';
+                    auto tensor_name = override;
+                    auto buffer_type = &override[tensor_name_span_len + 1];
+                    if (buft_list.find(buffer_type) == buft_list.end()) {
+                        printf("Available buffer types:\n");
+                        for (const auto & it : buft_list) {
+                            printf("  %s\n", ggml_backend_buft_name(it.second));
+                        }
+                        invalid_param = true;
+                        break;
+                    }
+                    group_tensor_buft_overrides.push_back({tensor_name, buft_list.at(buffer_type)});
+                    override_span_len = std::strcspn(override_group, ";");
+                }
+                if (invalid_param) {
+                    break;
+                }
+                group_tensor_buft_overrides.push_back({nullptr,nullptr});
+                params.tensor_buft_overrides.push_back(group_tensor_buft_overrides);
+                override_group_span_len = std::strcspn(value, ",");
+            } while (!last_group);
         } else if (arg == "-r" || arg == "--repetitions") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -648,6 +772,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.tensor_split.empty()) {
         params.tensor_split = cmd_params_defaults.tensor_split;
     }
+    if (params.tensor_buft_overrides.empty()) {
+        params.tensor_buft_overrides = cmd_params_defaults.tensor_buft_overrides;
+    }
     if (params.use_mmap.empty()) {
         params.use_mmap = cmd_params_defaults.use_mmap;
     }
@@ -689,6 +816,7 @@ struct cmd_params_instance {
     bool               no_kv_offload;
     bool               flash_attn;
     std::vector<float> tensor_split;
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
     bool               use_mmap;
     bool               embeddings;
 
@@ -733,13 +861,20 @@ struct cmd_params_instance {
         mparams.tensor_split = tensor_split.data();
         mparams.use_mmap     = use_mmap;
 
+        if (tensor_buft_overrides.empty()) {
+            mparams.tensor_buft_overrides = nullptr;
+        } else {
+            GGML_ASSERT(tensor_buft_overrides.back().pattern == nullptr && "Tensor buffer overrides not terminated with empty pattern");
+            mparams.tensor_buft_overrides = tensor_buft_overrides.data();
+        }
+
         return mparams;
     }
 
     bool equal_mparams(const cmd_params_instance & other) const {
         return model == other.model && n_gpu_layers == other.n_gpu_layers && rpc_servers_str == other.rpc_servers_str &&
                split_mode == other.split_mode && main_gpu == other.main_gpu && use_mmap == other.use_mmap &&
-               tensor_split == other.tensor_split;
+               tensor_split == other.tensor_split && vec_tensor_buft_override_equal(tensor_buft_overrides, other.tensor_buft_overrides);
     }
 
     llama_context_params to_llama_cparams() const {
@@ -769,6 +904,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & sm : params.split_mode)
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
+    for (const auto & ot : params.tensor_buft_overrides)
     for (const auto & mmp : params.use_mmap)
     for (const auto & embd : params.embeddings)
     for (const auto & nb : params.n_batch)
@@ -804,6 +940,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .no_kv_offload= */ nkvo,
                 /* .flash_attn   = */ fa,
                 /* .tensor_split = */ ts,
+                /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
             };
@@ -833,6 +970,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .no_kv_offload= */ nkvo,
                 /* .flash_attn   = */ fa,
                 /* .tensor_split = */ ts,
+                /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
             };
@@ -862,6 +1000,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .no_kv_offload= */ nkvo,
                 /* .flash_attn   = */ fa,
                 /* .tensor_split = */ ts,
+                /* .tensor_buft_overrides = */ ot,
                 /* .use_mmap     = */ mmp,
                 /* .embeddings   = */ embd,
             };
@@ -896,6 +1035,7 @@ struct test {
     bool                     no_kv_offload;
     bool                     flash_attn;
     std::vector<float>       tensor_split;
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
     bool                     use_mmap;
     bool                     embeddings;
     int                      n_prompt;
@@ -927,6 +1067,7 @@ struct test {
         no_kv_offload  = inst.no_kv_offload;
         flash_attn     = inst.flash_attn;
         tensor_split   = inst.tensor_split;
+        tensor_buft_overrides = inst.tensor_buft_overrides;
         use_mmap       = inst.use_mmap;
         embeddings     = inst.embeddings;
         n_prompt       = inst.n_prompt;
@@ -972,9 +1113,9 @@ struct test {
             "build_commit", "build_number", "cpu_info",       "gpu_info",   "backends",     "model_filename",
             "model_type",   "model_size",   "model_n_params", "n_batch",    "n_ubatch",     "n_threads",
             "cpu_mask",     "cpu_strict",   "poll",           "type_k",     "type_v",       "n_gpu_layers",
-            "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "use_mmap",
-            "embeddings",   "n_prompt",     "n_gen",          "test_time",  "avg_ns",       "stddev_ns",
-            "avg_ts",       "stddev_ts",
+            "split_mode",   "main_gpu",     "no_kv_offload",  "flash_attn", "tensor_split", "tensor_buft_overrides",
+            "use_mmap",     "embeddings",   "n_prompt",       "n_gen",      "test_time",    "avg_ns",
+            "stddev_ns",    "avg_ts",       "stddev_ts",
         };
         return fields;
     }
@@ -1000,6 +1141,7 @@ struct test {
 
     std::vector<std::string> get_values() const {
         std::string tensor_split_str;
+        std::string tensor_buft_overrides_str;
         int         max_nonzero = 0;
         for (size_t i = 0; i < llama_max_devices(); i++) {
             if (tensor_split[i] > 0) {
@@ -1014,6 +1156,26 @@ struct test {
                 tensor_split_str += "/";
             }
         }
+        if (tensor_buft_overrides.size() == 1) {
+            // Last element of tensor_buft_overrides is always a null pattern
+            // so if it is only one element long, it must be a null pattern.
+            GGML_ASSERT(tensor_buft_overrides[0].pattern == nullptr);
+            tensor_buft_overrides_str += "none";
+        } else {
+            for (size_t i = 0; i < tensor_buft_overrides.size()-1; i++) {
+                // Last element of tensor_buft_overrides is always a null pattern
+                if (tensor_buft_overrides[i].pattern == nullptr) {
+                    tensor_buft_overrides_str += "none";
+                } else {
+                    tensor_buft_overrides_str += tensor_buft_overrides[i].pattern;
+                    tensor_buft_overrides_str += "=";
+                    tensor_buft_overrides_str += ggml_backend_buft_name(tensor_buft_overrides[i].buft);
+                }
+                if (i + 2 < tensor_buft_overrides.size()) {
+                    tensor_buft_overrides_str += ";";
+                }
+            }
+        }
         std::vector<std::string> values = { build_commit,
                                             std::to_string(build_number),
                                             cpu_info,
@@ -1037,6 +1199,7 @@ struct test {
                                             std::to_string(no_kv_offload),
                                             std::to_string(flash_attn),
                                             tensor_split_str,
+                                            tensor_buft_overrides_str,
                                             std::to_string(use_mmap),
                                             std::to_string(embeddings),
                                             std::to_string(n_prompt),
@@ -1254,6 +1417,9 @@ struct markdown_printer : public printer {
         if (field == "tensor_split") {
             return "ts";
         }
+        if (field == "tensor_buft_overrides") {
+            return "ot";
+        }
         return field;
     }
 
@@ -1307,6 +1473,9 @@ struct markdown_printer : public printer {
         if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
             fields.emplace_back("tensor_split");
         }
+        if (params.tensor_buft_overrides.size() > 1 || !vec_vec_tensor_buft_override_equal(params.tensor_buft_overrides, cmd_params_defaults.tensor_buft_overrides)) {
+            fields.emplace_back("tensor_buft_overrides");
+        }
         if (params.use_mmap.size() > 1 || params.use_mmap != cmd_params_defaults.use_mmap) {
             fields.emplace_back("use_mmap");
         }

examples/llava/README.md

@@ -1,158 +1,75 @@
-# LLaVA
+# Multimodal Support in llama.cpp
 
-Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants,
-as well as llava-1.6 [llava-v1.6](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2) variants.
+This directory provides multimodal capabilities for `llama.cpp`. Initially intended as a showcase for running LLaVA models, its scope has expanded significantly over time to include various other vision-capable models. As a result, LLaVA is no longer the only multimodal architecture supported.
 
-The pre-converted [7b](https://huggingface.co/mys/ggml_llava-v1.5-7b)
-and [13b](https://huggingface.co/mys/ggml_llava-v1.5-13b)
-models are available.
-For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](https://huggingface.co/cmp-nct/llava-1.6-gguf)
+> [!IMPORTANT]
+>
+> Multimodal support can be viewed as a sub-project within `llama.cpp`. It is under **very heavy development**, and **breaking changes are expected**.
 
-After API is confirmed, more models will be supported / uploaded.
+The naming and structure related to multimodal support have evolved, which might cause some confusion. Here's a brief timeline to clarify:
 
-## Usage
-Build with cmake or run `make llama-llava-cli` to build it.
+- [#3436](https://github.com/ggml-org/llama.cpp/pull/3436): Initial support for LLaVA 1.5 was added, introducing `llava.cpp` and `clip.cpp`. The `llava-cli` binary was created for model interaction.
+- [#4954](https://github.com/ggml-org/llama.cpp/pull/4954): Support for MobileVLM was added, becoming the second vision model supported. This built upon the existing `llava.cpp`, `clip.cpp`, and `llava-cli` infrastructure.
+- **Expansion & Fragmentation:** Many new models were subsequently added (e.g., [#7599](https://github.com/ggml-org/llama.cpp/pull/7599), [#10361](https://github.com/ggml-org/llama.cpp/pull/10361), [#12344](https://github.com/ggml-org/llama.cpp/pull/12344), and others). However, `llava-cli` lacked support for the increasingly complex chat templates required by these models. This led to the creation of model-specific binaries like `qwen2vl-cli`, `minicpmv-cli`, and `gemma3-cli`. While functional, this proliferation of command-line tools became confusing for users.
+- [#12849](https://github.com/ggml-org/llama.cpp/pull/12849): `libmtmd` was introduced as a replacement for `llava.cpp`. Its goals include providing a single, unified command-line interface, improving the user/developer experience (UX/DX), and supporting both audio and image inputs.
+- [#13012](https://github.com/ggml-org/llama.cpp/pull/13012): `mtmd-cli` was added, consolidating the various model-specific CLIs into a single tool powered by `libmtmd`.
 
-After building, run: `./llama-llava-cli` to see the usage. For example:
+## Pre-quantized models
+
+These are ready-to-use models, most of them come with `Q4_K_M` quantization by default:
 
 ```sh
-./llama-llava-cli -m ../llava-v1.5-7b/ggml-model-f16.gguf --mmproj ../llava-v1.5-7b/mmproj-model-f16.gguf --image path/to/an/image.jpg
+# Gemma 3
+llama-mtmd-cli -hf ggml-org/gemma-3-4b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-12b-it-GGUF
+llama-mtmd-cli -hf ggml-org/gemma-3-27b-it-GGUF
+
+# SmolVLM
+llama-mtmd-cli -hf ggml-org/SmolVLM-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM-256M-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM-500M-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-2.2B-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-256M-Video-Instruct-GGUF
+llama-mtmd-cli -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
+
+# Pixtral 12B
+llama-mtmd-cli -hf ggml-org/pixtral-12b-GGUF
 ```
 
-**note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
-**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
+## How it works and what is `mmproj`?
 
-## LLaVA 1.5
+Multimodal support in `llama.cpp` works by encoding images into embeddings using a separate model component, and then feeding these embeddings into the language model.
 
-1. Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
+This approach keeps the multimodal components distinct from the core `libllama` library. Separating these allows for faster, independent development cycles. While many modern vision models are based on Vision Transformers (ViTs), their specific pre-processing and projection steps can vary significantly. Integrating this diverse complexity directly into `libllama` is currently challenging.
 
-```sh
-git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
+Consequently, running a multimodal model typically requires two GGUF files:
+1.  The standard language model file.
+2.  A corresponding **multimodal projector (`mmproj`)** file, which handles the image encoding and projection.
 
-git clone https://huggingface.co/openai/clip-vit-large-patch14-336
-```
+## What is `libmtmd`?
 
-2. Install the required Python packages:
+As outlined in the history, `libmtmd` is the modern library designed to replace the original `llava.cpp` implementation for handling multimodal inputs.
 
-```sh
-pip install -r examples/llava/requirements.txt
-```
+Built upon `clip.cpp` (similar to `llava.cpp`), `libmtmd` offers several advantages:
+- **Unified Interface:** Aims to consolidate interaction for various multimodal models.
+- **Improved UX/DX:** Features a more intuitive API, inspired by the `Processor` class in the Hugging Face `transformers` library.
+- **Flexibility:** Designed to support multiple input types (text, audio, images) while respecting the wide variety of chat templates used by different models.
 
-3. Use `llava_surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
+## How to obtain `mmproj`
 
-```sh
-python ./examples/llava/llava_surgery.py -m ../llava-v1.5-7b
-```
+Multimodal projector (`mmproj`) files are specific to each model architecture. Please refer to the relevant guide for instructions on how to obtain or create them:
 
-4. Use `convert_image_encoder_to_gguf.py` to convert the LLaVA image encoder to GGUF:
+- [LLaVA](../../docs/multimodal/llava.md)
+- [MobileVLM](../../docs/multimodal/MobileVLM.md)
+- [GLM-Edge](../../docs/multimodal/glmedge.md)
+- [MiniCPM-V 2.5](../../docs/multimodal/minicpmv2.5.md)
+- [MiniCPM-V 2.6](../../docs/multimodal/minicpmv2.6.md)
+- [MiniCPM-o 2.6](../../docs/multimodal/minicpmo2.6.md)
+- [IBM Granite Vision](../../docs/multimodal/granitevision.md)
+- [Google Gemma 3](../../docs/multimodal/gemma3.md)
 
-```sh
-python ./examples/llava/convert_image_encoder_to_gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
-```
-
-5. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
-
-```sh
-python ./examples/convert_legacy_llama.py ../llava-v1.5-7b --skip-unknown
-```
-
-Now both the LLaMA part and the image encoder are in the `llava-v1.5-7b` directory.
-
-## LLaVA 1.6 gguf conversion
-1) First clone a LLaVA 1.6 model:
-```console
-git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
-```
-
-2) Install the required Python packages:
-
-```sh
-pip install -r examples/llava/requirements.txt
-```
-
-3) Use `llava_surgery_v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
-```console
-python examples/llava/llava_surgery_v2.py -C -m ../llava-v1.6-vicuna-7b/
-```
-- you will find a llava.projector and a llava.clip file in your model directory
-
-4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
-```console
-mkdir vit
-cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
-cp ../llava-v1.6-vicuna-7b/llava.projector vit/
-curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
-```
-
-5) Create the visual gguf model:
-```console
-python ./examples/llava/convert_image_encoder_to_gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
-```
-- This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
-
-6) Then convert the model to gguf format:
-```console
-python ./examples/convert_legacy_llama.py ../llava-v1.6-vicuna-7b/ --skip-unknown
-```
-
-7) And finally we can run the llava cli using the 1.6 model version:
-```console
-./llama-llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
-```
-
-**note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
-
-**note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
-
-**note** if the language model in step `6)` is incompatible with the legacy conversion script, the easiest way handle the LLM model conversion is to load the model in transformers, and export only the LLM from the llava next model.
-
-```python
-import os
-import transformers
-
-model_path = ...
-llm_export_path = ...
-
-tokenizer = transformers.AutoTokenizer.from_pretrained(model_path)
-model = transformers.AutoModelForImageTextToText.from_pretrained(model_path)
-
-tokenizer.save_pretrained(llm_export_path)
-model.language_model.save_pretrained(llm_export_path)
-```
-
-Then, you can convert the LLM using the `convert_hf_to_gguf.py` script, which handles more LLM architectures.
-
-## llava-cli templating and llava-1.6 prompting
-
-llava-1.5 models all use the same vicuna prompt, here you can just add your image question like `-p "Provide a full description."`
-For llava-1.5 models which are not vicuna (mistral and Yi) you need to adapt system prompt as well as user prompt, for this purpose llava-cli has a basic templating system:
-
-**For Mistral and using llava-cli binary:**
-Add this: `-p "<image>\nUSER:\nProvide a full description.\nASSISTANT:\n"`
-The mistral template for llava-1.6 seems to be no system print and a USER/ASSISTANT role
-
-**For the 34B this should work:**
-Add this: `-e -p <|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nProvide a full description.<|im_end|><|im_start|>assistant\n`
-
-
-## How to know if you are running in llava-1.5 or llava-1.6 mode
-
-When running llava-cli you will see a visual information right before the prompt is being processed:
-
-**Llava-1.5:**
-`encode_image_with_clip: image embedding created: 576 tokens`
-
-**Llava-1.6 (anything above 576):**
-`encode_image_with_clip: image embedding created: 2880 tokens`
-
-
-Alternatively just pay notice to how many "tokens" have been used for your prompt, it will also show 1000+ tokens for llava-1.6
-
-
-
-
-## TODO
-
-- [x] Support non-CPU backend for the image encoding part.
-- [ ] Support different sampling methods.
-- [ ] Support more model variants.
+For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` flag to get the `mmproj` file:
+- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) - Note: 1B variant does not have vision support
+- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
+- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint

examples/llava/android/adb_run.sh

@@ -10,7 +10,7 @@ prompt="A chat between a curious user and an artificial intelligence assistant.
 # prompt="A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:"
 
 program_dir="build_64/bin"
-binName="llama-llava-cli"
+binName="llama-mtmd-cli"
 n_threads=4
 
 

examples/llava/clip-impl.h

@@ -17,33 +17,31 @@
 #define KEY_FTYPE               "general.file_type"
 #define KEY_NAME                "general.name"
 #define KEY_DESCRIPTION         "general.description"
-#define KEY_HAS_TEXT_ENC        "clip.has_text_encoder"
-#define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
-#define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
-#define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
-#define KEY_HAS_GLM_PROJ        "clip.has_glm_projector"
 #define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
-#define KEY_HAS_QWEN2VL_MERGER  "clip.has_qwen2vl_merger"
 #define KEY_USE_GELU            "clip.use_gelu"
 #define KEY_USE_SILU            "clip.use_silu"
-#define KEY_N_EMBD              "clip.%s.embedding_length"
-#define KEY_N_FF                "clip.%s.feed_forward_length"
-#define KEY_N_BLOCK             "clip.%s.block_count"
-#define KEY_N_HEAD              "clip.%s.attention.head_count"
-#define KEY_LAYER_NORM_EPS      "clip.%s.attention.layer_norm_epsilon"
-#define KEY_PROJ_DIM            "clip.%s.projection_dim"
-#define KEY_TOKENS              "tokenizer.ggml.tokens"
-#define KEY_N_POSITIONS         "clip.text.context_length"
+#define KEY_N_EMBD              "clip.vision.embedding_length"
+#define KEY_N_FF                "clip.vision.feed_forward_length"
+#define KEY_N_BLOCK             "clip.vision.block_count"
+#define KEY_N_HEAD              "clip.vision.attention.head_count"
+#define KEY_LAYER_NORM_EPS      "clip.vision.attention.layer_norm_epsilon"
+#define KEY_PROJ_DIM            "clip.vision.projection_dim"
 #define KEY_IMAGE_SIZE          "clip.vision.image_size"
 #define KEY_PATCH_SIZE          "clip.vision.patch_size"
 #define KEY_IMAGE_MEAN          "clip.vision.image_mean"
 #define KEY_IMAGE_STD           "clip.vision.image_std"
-#define KEY_PROJ_TYPE           "clip.projector_type"
 #define KEY_FEATURE_LAYER       "clip.vision.feature_layer"
+#define KEY_PROJ_SCALE_FACTOR   "clip.vision.projector.scale_factor"
+#define KEY_PROJ_TYPE           "clip.projector_type"
+
+#define KEY_USE_GLU_MLP         "clip.use_glu_mlp"  // for qwen2.5vl
+#define KEY_USE_RMS_NORM        "clip.use_rms_norm" // for qwen2.5vl
 
 #define KEY_MM_PATCH_MERGE_TYPE   "clip.vision.mm_patch_merge_type"
 #define KEY_IMAGE_GRID_PINPOINTS  "clip.vision.image_grid_pinpoints"
 #define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
+#define KEY_WIN_ATTN_PATTERN      "clip.vision.n_wa_pattern"
+#define KEY_ATTN_WINDOW_SIZE      "clip.vision.window_size"
 
 
 //
@@ -60,7 +58,9 @@
 #define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
 #define TN_ATTN_OUTPUT     "%s.blk.%d.attn_out.%s"
 #define TN_FFN_DOWN        "%s.blk.%d.ffn_down.%s"
+#define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
 #define TN_FFN_UP          "%s.blk.%d.ffn_up.%s"
+#define TN_FFN_GATE        "%s.blk.%d.ffn_gate.%s"
 #define TN_LN_1            "%s.blk.%d.ln1.%s"
 #define TN_LN_2            "%s.blk.%d.ln2.%s"
 #define TN_LN_PRE          "%s.pre_ln.%s"
@@ -72,6 +72,8 @@
 #define TN_IMAGE_NEWLINE   "model.image_newline"
 #define TN_MM_INP_PROJ     "mm.input_projection.weight" // gemma3
 #define TN_MM_SOFT_EMB_N   "mm.soft_emb_norm.weight"    // gemma3
+#define TN_MM_PROJECTOR    "mm.model.fc.weight"         // idefics3
+#define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
 
 // mimicpmv
 #define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
@@ -87,18 +89,19 @@
 #define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
 #define TN_GLM_ADAPTER_GATE     "adapter.linear.gate.%s"
 #define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
-#define TN_GLM_BOI_W            "adapter.boi"
-#define TN_GLM_EOI_W            "adapter.eoi"
 
 enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
     PROJECTOR_TYPE_LDP,
     PROJECTOR_TYPE_LDPV2,
-    PROJECTOR_TYPE_RESAMPLER,
+    PROJECTOR_TYPE_MINICPMV,
     PROJECTOR_TYPE_GLM_EDGE,
-    PROJECTOR_TYPE_MERGER,
+    PROJECTOR_TYPE_QWEN2VL,
     PROJECTOR_TYPE_GEMMA3,
+    PROJECTOR_TYPE_IDEFICS3,
+    PROJECTOR_TYPE_PIXTRAL,
+    PROJECTOR_TYPE_QWEN25VL,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -106,10 +109,13 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MLP,       "mlp" },
     { PROJECTOR_TYPE_LDP,       "ldp" },
     { PROJECTOR_TYPE_LDPV2,     "ldpv2"},
-    { PROJECTOR_TYPE_RESAMPLER, "resampler"},
+    { PROJECTOR_TYPE_MINICPMV,  "resampler"},
     { PROJECTOR_TYPE_GLM_EDGE,  "adapter"},
-    { PROJECTOR_TYPE_MERGER,    "qwen2vl_merger"},
+    { PROJECTOR_TYPE_QWEN2VL,   "qwen2vl_merger"},
+    { PROJECTOR_TYPE_QWEN25VL,  "qwen2.5vl_merger"},
     { PROJECTOR_TYPE_GEMMA3,    "gemma3"},
+    { PROJECTOR_TYPE_IDEFICS3,  "idefics3"},
+    { PROJECTOR_TYPE_PIXTRAL,   "pixtral"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {

examples/llava/clip.h

@@ -114,8 +114,6 @@ CLIP_API bool clip_is_qwen2vl(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_llava(const struct clip_ctx * ctx);
 CLIP_API bool clip_is_gemma3(const struct clip_ctx * ctx);
 
-CLIP_API int get_deepest_feature_layer(const struct clip_ctx * ctx);
-
 CLIP_API bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
 
 

examples/llava/gemma3_convert_encoder_to_gguf.py

@@ -1,307 +0,0 @@
-import gguf
-import argparse
-import logging
-import sys
-import torch
-import json
-import os
-import numpy as np
-from typing import cast, ContextManager, Any, Iterator
-from pathlib import Path
-from torch import Tensor
-
-logger = logging.getLogger("gemma3-mmproj")
-
-
-# (copied from convert_hf_to_gguf.py)
-# tree of lazy tensors
-class LazyTorchTensor(gguf.LazyBase):
-    _tensor_type = torch.Tensor
-    # to keep the type-checker happy
-    dtype: torch.dtype
-    shape: torch.Size
-
-    # only used when converting a torch.Tensor to a np.ndarray
-    _dtype_map: dict[torch.dtype, type] = {
-        torch.float16: np.float16,
-        torch.float32: np.float32,
-    }
-
-    # used for safetensors slices
-    # ref: https://github.com/huggingface/safetensors/blob/079781fd0dc455ba0fe851e2b4507c33d0c0d407/bindings/python/src/lib.rs#L1046
-    # TODO: uncomment U64, U32, and U16, ref: https://github.com/pytorch/pytorch/issues/58734
-    _dtype_str_map: dict[str, torch.dtype] = {
-        "F64": torch.float64,
-        "F32": torch.float32,
-        "BF16": torch.bfloat16,
-        "F16": torch.float16,
-        # "U64": torch.uint64,
-        "I64": torch.int64,
-        # "U32": torch.uint32,
-        "I32": torch.int32,
-        # "U16": torch.uint16,
-        "I16": torch.int16,
-        "U8": torch.uint8,
-        "I8": torch.int8,
-        "BOOL": torch.bool,
-        "F8_E4M3": torch.float8_e4m3fn,
-        "F8_E5M2": torch.float8_e5m2,
-    }
-
-    def numpy(self) -> gguf.LazyNumpyTensor:
-        dtype = self._dtype_map[self.dtype]
-        return gguf.LazyNumpyTensor(
-            meta=gguf.LazyNumpyTensor.meta_with_dtype_and_shape(dtype, self.shape),
-            args=(self,),
-            func=(lambda s: s.numpy())
-        )
-
-    @classmethod
-    def meta_with_dtype_and_shape(cls, dtype: torch.dtype, shape: tuple[int, ...]) -> Tensor:
-        return torch.empty(size=shape, dtype=dtype, device="meta")
-
-    @classmethod
-    def from_safetensors_slice(cls, st_slice: Any) -> Tensor:
-        dtype = cls._dtype_str_map[st_slice.get_dtype()]
-        shape: tuple[int, ...] = tuple(st_slice.get_shape())
-        lazy = cls(meta=cls.meta_with_dtype_and_shape(dtype, shape), args=(st_slice,), func=lambda s: s[:])
-        return cast(torch.Tensor, lazy)
-
-    @classmethod
-    def __torch_function__(cls, func, types, args=(), kwargs=None):
-        del types  # unused
-
-        if kwargs is None:
-            kwargs = {}
-
-        if func is torch.Tensor.numpy:
-            return args[0].numpy()
-
-        return cls._wrap_fn(func)(*args, **kwargs)
-
-
-class Gemma3VisionTower:
-    hparams: dict
-    gguf_writer: gguf.GGUFWriter
-    fname_out: Path
-    ftype: gguf.LlamaFileType
-
-    @staticmethod
-    def load_hparams(dir_model: Path):
-        with open(dir_model / "config.json", "r", encoding="utf-8") as f:
-            return json.load(f)
-
-    @staticmethod
-    def get_model_part_names(dir_model: Path, prefix: str, suffix: str) -> list[str]:
-        part_names: list[str] = []
-        for filename in os.listdir(dir_model):
-            if filename.startswith(prefix) and filename.endswith(suffix):
-                part_names.append(filename)
-        part_names.sort()
-        return part_names
-
-    def __init__(self,
-                 dir_model: Path,
-                 fname_out: Path,
-                 ftype: gguf.LlamaFileType,
-                 is_big_endian: bool,):
-        hparams = Gemma3VisionTower.load_hparams(dir_model)
-        self.hparams = hparams
-        self.fname_out = fname_out
-        self.ftype = ftype
-        endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
-        self.gguf_writer = gguf.GGUFWriter(path=None, arch="clip", endianess=endianess)
-
-        text_config = hparams["text_config"]
-        vision_config = hparams["vision_config"]
-
-        assert hparams["architectures"][0] == "Gemma3ForConditionalGeneration"
-        assert text_config is not None
-        assert vision_config is not None
-
-        self.gguf_writer.add_string ("clip.projector_type",              "gemma3")
-        self.gguf_writer.add_bool   ("clip.has_text_encoder",            False)
-        self.gguf_writer.add_bool   ("clip.has_vision_encoder",          True)
-        self.gguf_writer.add_bool   ("clip.has_llava_projector",         False) # legacy
-        self.gguf_writer.add_uint32 ("clip.vision.image_size",           vision_config["image_size"])
-        self.gguf_writer.add_uint32 ("clip.vision.patch_size",           vision_config["patch_size"])
-        self.gguf_writer.add_uint32 ("clip.vision.embedding_length",     vision_config["hidden_size"])
-        self.gguf_writer.add_uint32 ("clip.vision.feed_forward_length",  vision_config["intermediate_size"])
-        self.gguf_writer.add_uint32 ("clip.vision.projection_dim",       text_config["hidden_size"])
-        self.gguf_writer.add_uint32 ("clip.vision.block_count",          vision_config["num_hidden_layers"])
-        self.gguf_writer.add_uint32 ("clip.vision.attention.head_count", vision_config["num_attention_heads"])
-        self.gguf_writer.add_float32("clip.vision.attention.layer_norm_epsilon", vision_config.get("layer_norm_eps", 1e-6))
-        # default values taken from HF tranformers code
-        self.gguf_writer.add_array  ("clip.vision.image_mean", [0.5, 0.5, 0.5])
-        self.gguf_writer.add_array  ("clip.vision.image_std",  [0.5, 0.5, 0.5])
-        self.gguf_writer.add_bool   ("clip.use_gelu", True)
-
-        # load tensors
-        for name, data_torch in self.get_tensors(dir_model):
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-            self.add_tensor(name, data_torch)
-
-    def get_tensors(self, dir_model: Path) -> Iterator[tuple[str, Tensor]]:
-        part_names = Gemma3VisionTower.get_model_part_names(dir_model, "model", ".safetensors")
-        tensor_names_from_parts: set[str] = set()
-        for part_name in part_names:
-            logger.info(f"gguf: loading model part '{part_name}'")
-            from safetensors import safe_open
-            ctx = cast(ContextManager[Any], safe_open(dir_model / part_name, framework="pt", device="cpu"))
-            with ctx as model_part:
-                tensor_names_from_parts.update(model_part.keys())
-
-                for name in model_part.keys():
-                    data = model_part.get_slice(name)
-                    data = LazyTorchTensor.from_safetensors_slice(data)
-                    yield name, data
-
-    def add_tensor(self, name: str, data_torch: Tensor):
-        is_1d = len(data_torch.shape) == 1
-        is_embd = ".embeddings." in name
-        old_dtype = data_torch.dtype
-        can_quantize = not is_1d and not is_embd
-        data_qtype = gguf.GGMLQuantizationType.F32
-
-        # this is to support old checkpoint
-        # TODO: remove this when we have the final model
-        name = name.replace("vision_model.vision_model.", "vision_tower.vision_model.")
-        name = name.replace("multimodal_projector.", "multi_modal_projector.")
-
-        # filter only vision tensors
-        if not name.startswith("vision_tower.vision_model.") and not name.startswith("multi_modal_projector."):
-            return
-        # prefix
-        name = name.replace("vision_tower.vision_model.encoder.layers.", "v.blk.")
-        name = name.replace("vision_tower.vision_model.", "v.")
-        # projector and input embd
-        name = name.replace(".embeddings.patch_embedding.", ".patch_embd.")
-        name = name.replace(".embeddings.position_embedding.", ".position_embd.")
-        name = name.replace(
-            "multi_modal_projector.mm_input_projection_weight",
-            "mm.input_projection.weight"
-        )
-        name = name.replace(
-            "multi_modal_projector.mm_soft_emb_norm.weight",
-            "mm.soft_emb_norm.weight"
-        )
-        name = name.replace("post_layernorm.", "post_ln.")
-        # each block
-        name = name.replace(".self_attn.k_proj.", ".attn_k.")
-        name = name.replace(".self_attn.v_proj.", ".attn_v.")
-        name = name.replace(".self_attn.q_proj.", ".attn_q.")
-        name = name.replace(".self_attn.out_proj.", ".attn_out.")
-        name = name.replace(".layer_norm1.", ".ln1.")
-        name = name.replace(".layer_norm2.", ".ln2.")
-        name = name.replace(".mlp.fc1.", ".ffn_down.")
-        name = name.replace(".mlp.fc2.", ".ffn_up.")
-
-        if can_quantize:
-            if self.ftype == gguf.LlamaFileType.ALL_F32:
-                data_qtype = gguf.GGMLQuantizationType.F32
-            elif self.ftype == gguf.LlamaFileType.MOSTLY_F16:
-                data_qtype = gguf.GGMLQuantizationType.F16
-            elif self.ftype == gguf.LlamaFileType.MOSTLY_BF16:
-                data_qtype = gguf.GGMLQuantizationType.BF16
-            elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0:
-                data_qtype = gguf.GGMLQuantizationType.Q8_0
-            else:
-                raise ValueError(f"Unsupported file type: {self.ftype}")
-
-        # corrent norm value ; only this "soft_emb_norm" need to be corrected as it's part of Gemma projector
-        # the other norm values are part of SigLIP model, and they are already correct
-        # ref code: Gemma3RMSNorm
-        if "soft_emb_norm.weight" in name:
-            logger.info(f"Correcting norm value for '{name}'")
-            data_torch = data_torch + 1
-
-        data = data_torch.numpy()
-
-        try:
-            data = gguf.quants.quantize(data, data_qtype)
-        except Exception as e:
-            logger.error(f"Error quantizing tensor '{name}': {e}, fallback to F16")
-            data_qtype = gguf.GGMLQuantizationType.F16
-            data = gguf.quants.quantize(data, data_qtype)
-
-        # reverse shape to make it similar to the internal ggml dimension order
-        shape_str = f"{{{', '.join(str(n) for n in reversed(data_torch.shape))}}}"
-        logger.info(f"{f'%-32s' % f'{name},'} {old_dtype} --> {data_qtype.name}, shape = {shape_str}")
-
-        self.gguf_writer.add_tensor(name, data, raw_dtype=data_qtype)
-
-    def write(self):
-        self.gguf_writer.write_header_to_file(path=self.fname_out)
-        self.gguf_writer.write_kv_data_to_file()
-        self.gguf_writer.write_tensors_to_file(progress=True)
-        self.gguf_writer.close()
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(
-        description="Convert Gemma 3 vision tower safetensors to GGUF format",)
-    parser.add_argument(
-        "--outfile", type=Path, default="mmproj.gguf",
-        help="path to write to",
-    )
-    parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0"], default="f16",
-        help="output format",
-    )
-    parser.add_argument(
-        "--bigendian", action="store_true",
-        help="model is executed on big endian machine",
-    )
-    parser.add_argument(
-        "model", type=Path,
-        help="directory containing model file",
-        nargs="?",
-    )
-    parser.add_argument(
-        "--verbose", action="store_true",
-        help="increase output verbosity",
-    )
-
-    args = parser.parse_args()
-    if args.model is None:
-        parser.error("the following arguments are required: model")
-    return args
-
-
-def main() -> None:
-    args = parse_args()
-
-    if args.verbose:
-        logging.basicConfig(level=logging.DEBUG)
-    else:
-        logging.basicConfig(level=logging.INFO)
-
-    dir_model = args.model
-
-    if not dir_model.is_dir():
-        logger.error(f'Error: {args.model} is not a directory')
-        sys.exit(1)
-
-    ftype_map: dict[str, gguf.LlamaFileType] = {
-        "f32": gguf.LlamaFileType.ALL_F32,
-        "f16": gguf.LlamaFileType.MOSTLY_F16,
-        "bf16": gguf.LlamaFileType.MOSTLY_BF16,
-        "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
-    }
-
-    logger.info(f"Loading model: {dir_model.name}")
-
-    with torch.inference_mode():
-        gemma3_vision_tower = Gemma3VisionTower(
-            dir_model=dir_model,
-            fname_out=args.outfile,
-            ftype=ftype_map[args.outtype],
-            is_big_endian=args.bigendian,
-        )
-        gemma3_vision_tower.write()
-
-
-if __name__ == '__main__':
-    main()
-

examples/llava/mtmd-cli.cpp

@@ -24,7 +24,9 @@
 #include <signal.h>
 #endif
 
-static bool g_is_generating = false;
+// volatile, because of signal being an interrupt
+static volatile bool g_is_generating = false;
+static volatile bool g_is_interrupted = false;
 
 /**
  * Please note that this is NOT a production-ready stuff.
@@ -38,7 +40,8 @@ static void show_additional_info(int /*argc*/, char ** argv) {
         "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> -p <prompt>\n\n"
         "  -m and --mmproj are required\n"
         "  -hf user/repo can replace both -m and --mmproj in most cases\n"
-        "  --image and -p are optional, if NOT provided, the CLI will run in chat mode\n",
+        "  --image and -p are optional, if NOT provided, the CLI will run in chat mode\n"
+        "  to disable using GPU for mmproj model, add --no-mmproj-offload\n",
         argv[0]
     );
 }
@@ -50,8 +53,10 @@ static void sigint_handler(int signo) {
             g_is_generating = false;
         } else {
             console::cleanup();
-            LOG("\nInterrupted by user\n");
-            _exit(130);
+            if (g_is_interrupted) {
+                _exit(1);
+            }
+            g_is_interrupted = true;
         }
     }
 }
@@ -108,10 +113,10 @@ struct mtmd_cli_context {
     void init_vision_context(common_params & params) {
         const char * clip_path = params.mmproj.path.c_str();
         ctx_vision.reset(mtmd_init_from_file(clip_path, model, mtmd_context_params{
-            /* use_gpu */   true,
+            /* use_gpu */   params.mmproj_use_gpu,
             /* timings */   true,
             /* n_threads */ params.cpuparams.n_threads,
-            /* verbosity */ GGML_LOG_LEVEL_INFO,
+            /* verbosity */ params.verbosity > 0 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_INFO,
         }));
         if (!ctx_vision.get()) {
             LOG_ERR("Failed to load vision model from %s\n", clip_path);
@@ -167,7 +172,7 @@ struct decode_embd_batch {
 static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int n_predict) {
     llama_tokens generated_tokens;
     for (int i = 0; i < n_predict; i++) {
-        if (i > n_predict || !g_is_generating) {
+        if (i > n_predict || !g_is_generating || g_is_interrupted) {
             printf("\n");
             break;
         }
@@ -184,6 +189,11 @@ static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int
         printf("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
         fflush(stdout);
 
+        if (g_is_interrupted) {
+            printf("\n");
+            break;
+        }
+
         // eval the token
         common_batch_clear(ctx.batch);
         common_batch_add(ctx.batch, token_id, ctx.n_past++, {0}, true);
@@ -219,6 +229,9 @@ static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vect
     text.add_special   = add_bos;
     text.parse_special = true;
     mtmd_input_chunks chunks;
+
+    if (g_is_interrupted) return 0;
+
     int32_t res = mtmd_tokenize(ctx.ctx_vision.get(), chunks, text, bitmaps);
     if (res != 0) {
         LOG_ERR("Unable to tokenize prompt, res = %d\n", res);
@@ -249,6 +262,7 @@ int main(int argc, char ** argv) {
 
     if (params.mmproj.path.empty()) {
         show_additional_info(argc, argv);
+        LOG_ERR("ERR: Missing --mmproj argument\n");
         return 1;
     }
 
@@ -276,6 +290,8 @@ int main(int argc, char ** argv) {
 #endif
     }
 
+    if (g_is_interrupted) return 130;
+
     if (is_single_turn) {
         g_is_generating = true;
         if (params.prompt.find("<__image__>") == std::string::npos) {
@@ -287,7 +303,7 @@ int main(int argc, char ** argv) {
         if (eval_message(ctx, msg, params.image, true)) {
             return 1;
         }
-        if (generate_response(ctx, smpl, n_predict)) {
+        if (!g_is_interrupted && generate_response(ctx, smpl, n_predict)) {
             return 1;
         }
 
@@ -302,12 +318,13 @@ int main(int argc, char ** argv) {
         std::vector<std::string> images_fname;
         std::string content;
 
-        while (true) {
+        while (!g_is_interrupted) {
             g_is_generating = false;
             LOG("\n> ");
             console::set_display(console::user_input);
             std::string line;
             console::readline(line, false);
+            if (g_is_interrupted) break;
             console::set_display(console::reset);
             line = string_strip(line);
             if (line.empty()) {
@@ -335,6 +352,7 @@ int main(int argc, char ** argv) {
             msg.role = "user";
             msg.content = content;
             int ret = eval_message(ctx, msg, images_fname, is_first_msg);
+            if (g_is_interrupted) break;
             if (ret == 2) {
                 // non-fatal error
                 images_fname.clear();
@@ -352,6 +370,7 @@ int main(int argc, char ** argv) {
             is_first_msg = false;
         }
     }
+    if (g_is_interrupted) LOG("\nInterrupted by user\n");
     llama_perf_context_print(ctx.lctx);
-    return 0;
+    return g_is_interrupted ? 130 : 0;
 }

examples/llava/mtmd.cpp

@@ -176,6 +176,8 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
 
     std::string prompt_modified(text.text);
     std::string marker_modified(ctx->image_marker);
+    projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
+
     // a bit hacky here, but works for now
     // for some models, we need to add prefix and suffix to the image embeddings
     if (clip_is_gemma3(ctx->ctx_clip)) {
@@ -183,6 +185,21 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
         // <start_of_image> ... (image embeddings) ... <end_of_image>
         marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
         string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+
+    } else if (proj_type == PROJECTOR_TYPE_GLM_EDGE) {
+        // <|begin_of_image|> ... (image embeddings) ... <|end_of_image|>
+        marker_modified = "<|begin_of_image|>" + ctx->image_marker + "<|end_of_image|>";
+        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+
+    } else if (proj_type == PROJECTOR_TYPE_IDEFICS3) {
+        // https://github.com/huggingface/transformers/blob/a42ba80fa520c784c8f11a973ca9034e5f859b79/src/transformers/models/idefics3/processing_idefics3.py#L192-L215
+        marker_modified = "<fake_token_around_image><global-img>" + ctx->image_marker + "<fake_token_around_image>";
+        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
+
+    } else if (proj_type == PROJECTOR_TYPE_PIXTRAL) {
+        // https://github.com/huggingface/transformers/blob/1cd110c6cb6a6237614130c470e9a902dbc1a4bd/docs/source/en/model_doc/pixtral.md
+        marker_modified = ctx->image_marker + "[IMG_END]";
+        string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
     }
 
     // llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
@@ -212,7 +229,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
 
         for (auto & entry : batch_f32.entries) {
             mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
-            image_tokens->nx = clip_n_patches(ctx->ctx_clip);
+            image_tokens->nx = clip_n_patches_by_img(ctx->ctx_clip, entry.get());
             image_tokens->ny = 1;
             image_tokens->batch_f32.entries.push_back(std::move(entry));
             image_tokens->id = id;
@@ -306,8 +323,13 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
                 }
 
             } else {
+                size_t n_tokens = 0;
+                for (const auto & entry : batch_f32.entries) {
+                    n_tokens += clip_n_patches_by_img(ctx->ctx_clip, entry.get());
+                }
+
                 mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
-                image_tokens->nx = clip_n_patches(ctx->ctx_clip) * batch_f32.entries.size(); // TODO @ngxson : use clip_n_patches_by_image
+                image_tokens->nx = n_tokens;
                 image_tokens->ny = 1; // TODO
                 image_tokens->batch_f32 = std::move(batch_f32);
                 image_tokens->id = bitmaps[i_img].id; // optional
@@ -375,7 +397,7 @@ int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens)
         // TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
         const auto & entries = image_tokens->batch_f32.entries;
         for (size_t i = 0; i < entries.size(); i++) {
-            int n_tokens_per_image = clip_n_patches(ctx->ctx_clip);
+            int n_tokens_per_image = clip_n_patches_by_img(ctx->ctx_clip, entries[i].get());
             ok = clip_image_encode(
                 ctx->ctx_clip,
                 ctx->n_threads,

examples/llava/qwen2_vl_surgery.py

@@ -1,14 +1,16 @@
 import argparse
-from typing import Dict
+from typing import Dict, List, Optional
 
 import torch
 import numpy as np
 from gguf import *
 from transformers import (
-    Qwen2VLForConditionalGeneration,
-    Qwen2VLProcessor,
     AutoProcessor,
-    Qwen2VLConfig
+    Qwen2VLConfig,
+    Qwen2VLProcessor,
+    Qwen2VLForConditionalGeneration,
+    Qwen2_5_VLConfig, # type: ignore[reportAttributeAccessIssue]
+    Qwen2_5_VLForConditionalGeneration, # type: ignore[reportAttributeAccessIssue]
 )
 
 
@@ -19,61 +21,93 @@ def k(raw_key: str, arch: str) -> str:
     return raw_key.format(arch=arch)
 
 
-def to_gguf_name(name: str) -> str:
-    og = name
-    name = name.replace("text_model", "t").replace("vision_model", "v")
-    name = name.replace("blocks", "blk").replace("embeddings.", "")
-    name = name.replace("attn.", "attn_")
-    name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
-    # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
-    name = name.replace("norm1", "ln1").replace("norm2", "ln2")
-    name = name.replace("merger.mlp", 'mm')
-    print(f"[to_gguf_name] {og} --> {name}")
-    return name
+def get_n_wa_pattern(fullatt_block_indexes: Optional[List[int]]):
+    if fullatt_block_indexes is None:
+        return 0
+    n_wa = fullatt_block_indexes[0]
+    for a, b in zip(fullatt_block_indexes, fullatt_block_indexes[1:]):
+        if b - a - 1 != n_wa:
+            raise ValueError(
+                f"window/full attention layer should have fix pattern of "
+                f"for each full-attention layer followed by {n_wa} window-attention layers"
+            )
+    return n_wa + 1
 
 
-def find_vision_tensors(qwen2vl, dtype) -> Dict[str, np.ndarray]:
-    vision_model = qwen2vl.visual
-    tensor_map = {}
-    for name, ten in vision_model.state_dict().items():
-        ten = ten.numpy()
-        if 'qkv' in name:
-            if ten.ndim == 2: # weight
-                c3, _ = ten.shape
-            else:             # bias
-                c3 = ten.shape[0]
-            assert c3 % 3 == 0
-            c = c3 // 3
-            wq = ten[:c]
-            wk = ten[c: c * 2]
-            wv = ten[c * 2:]
-            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
-            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
-            tensor_map[to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
-        elif 'merger' in name:
-            if name.endswith("ln_q.weight"):
-                tensor_map['v.post_ln.weight'] = ten
-            elif name.endswith("ln_q.bias"):
-                tensor_map['v.post_ln.bias'] = ten
+class VL2:
+
+    @staticmethod
+    def to_gguf_name(name: str) -> str:
+        og = name
+        name = name.replace("text_model", "t").replace("vision_model", "v")
+        name = name.replace("blocks", "blk").replace("embeddings.", "")
+        name = name.replace("attn.", "attn_")
+        name = name.replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("proj.", "out.")
+        # name = name.replace("layrnorm", "ln").replace("layer_norm", "ln").replace("layernorm", "ln")
+        name = name.replace("norm1", "ln1").replace("norm2", "ln2")
+        name = name.replace("merger.mlp", 'mm')
+        print(f"[to_gguf_name] {og} --> {name}")
+        return name
+
+    @classmethod
+    def find_vision_tensors(cls, qwen2vl, dtype) -> Dict[str, np.ndarray]:
+        vision_model = qwen2vl.visual
+        tensor_map = {}
+        for name, ten in vision_model.state_dict().items():
+            ten = ten.numpy()
+            if 'qkv' in name:
+                if ten.ndim == 2: # weight
+                    c3, _ = ten.shape
+                else:             # bias
+                    c3 = ten.shape[0]
+                assert c3 % 3 == 0
+                c = c3 // 3
+                wq = ten[:c]
+                wk = ten[c: c * 2]
+                wv = ten[c * 2:]
+                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "q")] = wq
+                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "k")] = wk
+                tensor_map[cls.to_gguf_name(f"vision_model.{name}").replace("qkv", "v")] = wv
+            elif 'merger' in name:
+                if name.endswith("ln_q.weight"):
+                    tensor_map['v.post_ln.weight'] = ten
+                elif name.endswith("ln_q.bias"):
+                    tensor_map['v.post_ln.bias'] = ten
+                else:
+                    # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
+                    tensor_map[cls.to_gguf_name(name)] = ten
+            elif 'patch_embed.proj.weight' in name:
+                # NOTE: split Conv3D into Conv2Ds
+                c1, c2, kt, kh, kw = ten.shape
+                assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
+                tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
+                tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
             else:
-                # "merger.mlp.%d.weight/bias" --> "mm.%d.weight/bias"
-                tensor_map[to_gguf_name(name)] = ten
-        elif 'patch_embed.proj.weight' in name:
-            # NOTE: split Conv3D into Conv2Ds
-            c1, c2, kt, kh, kw = ten.shape
-            assert kt == 2, "Current implmentation only support temporal_patch_size of 2"
-            tensor_map["v.patch_embd.weight"] = ten[:, :, 0, ...]
-            tensor_map["v.patch_embd.weight.1"] = ten[:, :, 1, ...]
-        else:
-            tensor_map[to_gguf_name(f"vision_model.{name}")] = ten
+                tensor_map[cls.to_gguf_name(f"vision_model.{name}")] = ten
 
-    for new_name, ten in tensor_map.items():
-        if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
-            tensor_map[new_name] = ten.astype(np.float32)
-        else:
-            tensor_map[new_name] = ten.astype(dtype)
-    tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32)  # dummy tensor, just here as a placeholder
-    return tensor_map
+        for new_name, ten in tensor_map.items():
+            if ten.ndim <= 1 or new_name.endswith("_norm.weight"):
+                tensor_map[new_name] = ten.astype(np.float32)
+            else:
+                tensor_map[new_name] = ten.astype(dtype)
+        tensor_map["v.position_embd.weight"] = np.zeros([10, 10], dtype=np.float32)  # dummy tensor, just here as a placeholder
+        return tensor_map
+
+
+class VL25(VL2):
+
+    @staticmethod
+    def to_gguf_name(name: str) -> str:
+        og = name
+        name = name.replace("text_model", "t").replace("vision_model", "v")
+        name = name.replace("blocks", "blk").replace("embeddings.", "")
+        name = name.replace("attn.", "attn_")
+        name = name.replace("mlp.down_proj", "ffn_down").replace("mlp.up_proj", "ffn_up")
+        name = name.replace("mlp.gate_proj", "ffn_gate").replace("proj.", "out.")
+        name = name.replace("norm1", "ln1").replace("norm2", "ln2")
+        name = name.replace("merger.mlp", 'mm')
+        print(f"[vl25][to_gguf_name] {og} --> {name}")
+        return name
 
 
 def main(args):
@@ -82,7 +116,7 @@ def main(args):
         np_dtype = np.float32
         ftype = 0
     elif args.data_type == 'fp16':
-        dtype = torch.float32
+        dtype = torch.float16
         np_dtype = np.float16
         ftype = 1
     else:
@@ -92,11 +126,18 @@ def main(args):
     model_path = ""
     model_name = args.model_name
     print("model_name: ", model_name)
-    qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
-        model_name, torch_dtype=dtype, device_map="cpu"
-    )
-    cfg: Qwen2VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
-    vcfg = cfg.vision_config
+    if args.model_type == "qwen2vl":
+        qwen2vl = Qwen2VLForConditionalGeneration.from_pretrained(
+            model_name, torch_dtype=dtype, device_map="cpu"
+        )
+        cfg: Qwen2VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
+        vcfg = cfg.vision_config
+    else:
+        qwen2vl = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+            model_name, torch_dtype=dtype, device_map="cpu"
+        )
+        cfg: Qwen2_5_VLConfig = qwen2vl.config  # type: ignore[reportAssignmentType]
+        vcfg = cfg.vision_config
 
     if os.path.isdir(model_name):
         local_model = True
@@ -113,7 +154,6 @@ def main(args):
     fout.add_bool("clip.has_text_encoder", False)
     fout.add_bool("clip.has_vision_encoder", True)
     fout.add_bool("clip.has_qwen2vl_merger", True)
-    fout.add_string("clip.projector_type", "qwen2vl_merger")
 
     print(cfg.vision_config)
     if 'silu' in cfg.vision_config.hidden_act.lower():
@@ -125,14 +165,25 @@ def main(args):
     else:
         raise ValueError()
 
-    tensor_map = find_vision_tensors(qwen2vl, np_dtype)
+    if args.model_type == "qwen2.5vl":
+        fout.add_uint32("clip.vision.n_wa_pattern", get_n_wa_pattern(vcfg.fullatt_block_indexes))
+        fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.hidden_size)
+        fout.add_uint32("clip.vision.projection_dim", vcfg.out_hidden_size)
+        fout.add_string("clip.projector_type", "qwen2.5vl_merger")
+    else:
+        fout.add_string("clip.projector_type", "qwen2vl_merger")
+        fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
+        fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
+
+    if args.model_type == "qwen2.5vl":
+        tensor_map = VL25.find_vision_tensors(qwen2vl, np_dtype)
+    else:
+        tensor_map = VL2.find_vision_tensors(qwen2vl, np_dtype)
     for name, data in tensor_map.items():
         fout.add_tensor(name, data)
 
     fout.add_uint32("clip.vision.patch_size", vcfg.patch_size)
     fout.add_uint32("clip.vision.image_size", 14 * 40)  # some reasonable size that is divable by (14*2)
-    fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.embed_dim)
-    fout.add_uint32("clip.vision.projection_dim", vcfg.hidden_size)
     fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), vcfg.num_heads)
     fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
     fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), vcfg.depth)
@@ -160,6 +211,7 @@ def main(args):
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()
     parser.add_argument("model_name", nargs='?', default="Qwen/Qwen2-VL-2B-Instruct")
+    parser.add_argument("--model_type", nargs='?', choices=['qwen2vl', 'qwen2.5vl'], default="qwen2vl")
     parser.add_argument("--data_type", nargs='?', choices=['fp32', 'fp16'], default="fp32")
     args = parser.parse_args()
     main(args)

examples/llava/qwen2vl-cli.cpp

@@ -23,6 +23,9 @@
 #include <algorithm>
 #include <iostream>
 #include <fstream>
+#include <limits>
+#include <cassert>
+#include <cmath>
 
 
 static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed,
@@ -367,14 +370,14 @@ static void debug_test_mrope_2d() {
     // 1. Initialize backend
     ggml_backend_t backend = NULL;
     std::string backend_name = "";
-#ifdef GGML_USE_CUDA
-    fprintf(stderr, "%s: using CUDA backend\n", __func__);
-    backend = ggml_backend_cuda_init(0); // init device 0
-    backend_name = "cuda";
-    if (!backend) {
-        fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
-    }
-#endif
+// #ifdef GGML_USE_CUDA
+//     fprintf(stderr, "%s: using CUDA backend\n", __func__);
+//     backend = ggml_backend_cuda_init(0); // init device 0
+//     backend_name = "cuda";
+//     if (!backend) {
+//         fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
+//     }
+// #endif
     // if there aren't GPU Backends fallback to CPU backend
     if (!backend) {
         backend = ggml_backend_cpu_init();
@@ -483,28 +486,82 @@ static void debug_test_mrope_2d() {
     ggml_backend_free(backend);
 }
 
-static void debug_dump_img_embed(struct llava_context * ctx_llava) {
-    int n_embd  = llama_model_n_embd(llama_get_model(ctx_llava->ctx_llama));
-    int ne = n_embd * 4;
-    float vals[56 * 56 * 3];
+enum model_output_type {
+    conv3d,
+    patch_embed,
+    patch_win_attn_scatter,
+    first_attn_layer,
+    last_attn_layer,
+    attn_softmax,
+    final_layer,
+};
+
+static void debug_dump_img_embed(struct llava_context * ctx_llava, model_output_type output_type) {
+    constexpr int ih = 140;
+    constexpr int iw = 196;
+    // constexpr int ih = 56;
+    // constexpr int iw = 56;
+    // int n_embd  = llama_model_n_embd(llama_get_model(ctx_llava->ctx_llama));
+    int n_embd  = 1280;
+    int merge = 1;
+    if (output_type == model_output_type::final_layer) {
+        n_embd  = 2048;
+        merge = 2;
+    }
+    else if (output_type == model_output_type::attn_softmax) {
+        merge = 1;
+        n_embd = (ih/14/merge) * (iw/14/merge) * 16;
+    }
+
+    int ne = (ih/14/merge) * (iw/14/merge) * n_embd;
+    float vals[iw * ih * 3];
     // float embd[ne];
     std::vector<float> embd;
     embd.resize(ne);
 
-    for (int i = 0; i < 56*56; i++)
+    for (int i = 0; i < iw*ih; i++)
     {
         for (int c = 0; c < 3; c++)
-            vals[i * 3 + c] = (float)(i % (56 * 56)) / (56*56);
+            vals[i * 3 + c] = (float)i / (iw*ih);
     }
 
-    clip_encode_float_image(ctx_llava->ctx_clip, 16, vals, 56, 56, embd.data());
+    clip_encode_float_image(ctx_llava->ctx_clip, 8, vals, ih, iw, embd.data());
 
-    std::ofstream outFile("img_embed.bin", std::ios::binary);
+    std::string file_postfix = "";
+    switch (output_type)
+    {
+    case model_output_type::conv3d:
+        file_postfix = "conv3d";
+        break;
+    case model_output_type::patch_embed:
+        file_postfix = "patch_embed";
+        break;
+    case model_output_type::patch_win_attn_scatter:
+        file_postfix = "scatter";
+        break;
+    case model_output_type::first_attn_layer:
+        file_postfix = "first_attn";
+        break;
+    case model_output_type::last_attn_layer:
+        file_postfix = "last_attn";
+        break;
+    case model_output_type::attn_softmax:
+        file_postfix = "attn_softmax";
+        break;
+    case model_output_type::final_layer:
+        file_postfix = "final";
+        break;
+    default:
+        break;
+    }
+    auto output_path = "img_embed_" + file_postfix + ".bin";
+
+    std::ofstream outFile(output_path, std::ios::binary);
     if (outFile.is_open()) {
         outFile.write(reinterpret_cast<const char*>(embd.data()), ne * sizeof(float));
 
         outFile.close();
-        std::cout << "Data successfully written to mrope.bin" << std::endl;
+        std::cout << "Data successfully written to ::[ " << output_path << std::endl;
     } else {
         std::cerr << "Error opening file!" << std::endl;
     }
@@ -551,8 +608,9 @@ int main(int argc, char ** argv) {
     } else if (params.image[0].empty()) {
         auto ctx_llava = llava_init_context(&params, model);
 
-        debug_test_mrope_2d();
-        debug_dump_img_embed(ctx_llava);
+        // debug_test_mrope_2d();
+        debug_dump_img_embed(ctx_llava, model_output_type::final_layer);
+        // debug_dump_img_embed(ctx_llava, model_output_type::last_attn_layer);
 
         llama_perf_context_print(ctx_llava->ctx_llama);
         ctx_llava->model = NULL;

examples/llava/tests.sh

@@ -13,6 +13,14 @@ mkdir -p $SCRIPT_DIR/output
 PROJ_ROOT="$SCRIPT_DIR/../.."
 cd $PROJ_ROOT
 
+# Check if the first argument is "big", then run test with big models
+# This is useful if we're running the script on a larger machine, so we can test the big models
+RUN_BIG_TESTS=false
+if [ "${1:-}" = "big" ]; then
+    RUN_BIG_TESTS=true
+    echo "Include BIG models..."
+fi
+
 ###############
 
 arr_bin=()
@@ -28,6 +36,15 @@ add_test() {
     arr_tmpl+=("$tmpl")
 }
 
+add_test_big() {
+    if [ "$RUN_BIG_TESTS" = true ]; then
+        add_test "$@"
+    fi
+}
+
+add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-500M-Instruct-GGUF:Q8_0"
+add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-2.2B-Instruct-GGUF:Q4_K_M"
+add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-500M-Video-Instruct-GGUF:Q8_0"
 add_test "llama-mtmd-cli"  "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
 add_test "llama-mtmd-cli"  "guinmoon/MobileVLM-3B-GGUF:Q4_K_M"               "deepseek"
 add_test "llama-mtmd-cli"  "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
@@ -38,8 +55,18 @@ add_test "llama-mtmd-cli"  "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K"  # mode
 add_test "llama-mtmd-cli"  "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
 add_test "llama-mtmd-cli"  "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
 add_test "llama-qwen2vl-cli"  "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
+add_test "llama-qwen2vl-cli"  "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
 
-# add_test "llama-mtmd-cli"  "cmp-nct/Yi-VL-6B-GGUF:Q5_K"  # this model has broken chat template, not usable
+# to test the big models, run: ./tests.sh big
+add_test_big "llama-mtmd-cli" "ggml-org/pixtral-12b-GGUF:Q4_K_M"
+
+# these models always give the wrong answer, not sure why
+# add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-Instruct-GGUF:Q4_K_M"
+# add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-256M-Instruct-GGUF:Q8_0"
+# add_test "llama-mtmd-cli"  "ggml-org/SmolVLM2-256M-Video-Instruct-GGUF:Q8_0"
+
+# this model has broken chat template, not usable
+# add_test "llama-mtmd-cli"  "cmp-nct/Yi-VL-6B-GGUF:Q5_K"
 
 ###############
 

examples/quantize-stats/CMakeLists.txt

@@ -1,6 +0,0 @@
-set(TARGET llama-quantize-stats)
-add_executable(${TARGET} quantize-stats.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
-target_include_directories(${TARGET} PRIVATE ../../common)
-target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/rpc/rpc-server.cpp

@@ -22,6 +22,7 @@
 
 #include "ggml-rpc.h"
 #ifdef _WIN32
+#  define NOMINMAX
 #  define DIRECTORY_SEPARATOR '\\'
 #  include <locale>
 #  include <windows.h>
@@ -37,6 +38,8 @@
 #include <stdio.h>
 #include <vector>
 #include <filesystem>
+#include <algorithm>
+#include <thread>
 
 namespace fs = std::filesystem;
 
@@ -150,12 +153,14 @@ struct rpc_server_params {
     int         port        = 50052;
     size_t      backend_mem = 0;
     bool        use_cache   = false;
+    int         n_threads   = std::max(1U, std::thread::hardware_concurrency()/2);
 };
 
 static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
     fprintf(stderr, "Usage: %s [options]\n\n", argv[0]);
     fprintf(stderr, "options:\n");
     fprintf(stderr, "  -h, --help                show this help message and exit\n");
+    fprintf(stderr, "  -t,      --threads        number of threads for the CPU backend (default: %d)\n", params.n_threads);
     fprintf(stderr, "  -H HOST, --host HOST      host to bind to (default: %s)\n", params.host.c_str());
     fprintf(stderr, "  -p PORT, --port PORT      port to bind to (default: %d)\n", params.port);
     fprintf(stderr, "  -m MEM,  --mem MEM        backend memory size (in MB)\n");
@@ -172,6 +177,15 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
                 return false;
             }
             params.host = argv[i];
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.n_threads = std::stoi(argv[i]);
+            if (params.n_threads <= 0) {
+                fprintf(stderr, "error: invalid number of threads: %d\n", params.n_threads);
+                return false;
+            }
         } else if (arg == "-p" || arg == "--port") {
             if (++i >= argc) {
                 return false;
@@ -199,7 +213,7 @@ static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params &
     return true;
 }
 
-static ggml_backend_t create_backend() {
+static ggml_backend_t create_backend(const rpc_server_params & params) {
     ggml_backend_t backend = NULL;
 #ifdef GGML_USE_CUDA
     fprintf(stderr, "%s: using CUDA backend\n", __func__);
@@ -231,6 +245,7 @@ static ggml_backend_t create_backend() {
     if (!backend) {
         fprintf(stderr, "%s: using CPU backend\n", __func__);
         backend = ggml_backend_cpu_init();
+        ggml_backend_cpu_set_n_threads(backend, params.n_threads);
     }
     return backend;
 }
@@ -275,7 +290,7 @@ int main(int argc, char * argv[]) {
         fprintf(stderr, "\n");
     }
 
-    ggml_backend_t backend = create_backend();
+    ggml_backend_t backend = create_backend(params);
     if (!backend) {
         fprintf(stderr, "Failed to create backend\n");
         return 1;

examples/server/public_legacy/json-schema-to-grammar.mjs

@@ -2,6 +2,9 @@
 const SPACE_RULE = '| " " | "\\n"{1,2} [ \\t]{0,20}';
 
 function _buildRepetition(itemRule, minItems, maxItems, opts={}) {
+  if (maxItems == 0) {
+    return '';
+  }
   if (minItems === 0 && maxItems === 1) {
     return `${itemRule}?`;
   }

ggml/include/ggml-cpu.h

@@ -133,6 +133,11 @@ extern "C" {
 
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
+    GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);
+    GGML_BACKEND_API void ggml_cpu_bf16_to_fp32(const ggml_bf16_t *, float *, int64_t);
+
 #ifdef __cplusplus
 }
 #endif

ggml/include/ggml-rpc.h

@@ -7,7 +7,7 @@
 extern "C" {
 #endif
 
-#define RPC_PROTO_MAJOR_VERSION    1
+#define RPC_PROTO_MAJOR_VERSION    2
 #define RPC_PROTO_MINOR_VERSION    0
 #define RPC_PROTO_PATCH_VERSION    0
 #define GGML_RPC_MAX_SERVERS       16

ggml/include/ggml.h

@@ -393,8 +393,8 @@ extern "C" {
 
     // precision
     enum ggml_prec {
-        GGML_PREC_DEFAULT,
-        GGML_PREC_F32,
+        GGML_PREC_DEFAULT =  0, // stored as ggml_tensor.op_params, 0 by default
+        GGML_PREC_F32     = 10,
     };
 
     // model file types
@@ -481,6 +481,7 @@ extern "C" {
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
         GGML_OP_IM2COL_BACK,
+        GGML_OP_CONV_2D_DW,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
@@ -677,6 +678,9 @@ extern "C" {
     GGML_API bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
     GGML_API bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
 
+    // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
+    GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
+
     GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
     GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
@@ -1660,7 +1664,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // depthwise
+    // depthwise (via im2col and mul_mat)
     GGML_API struct ggml_tensor * ggml_conv_2d_dw(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,  // convolution kernel
@@ -1672,6 +1676,22 @@ extern "C" {
             int                  d0,  // dilation dimension 0
             int                  d1); // dilation dimension 1
 
+    // Depthwise 2D convolution
+    // may be faster than ggml_conv_2d_dw, but not available in all backends
+    // a:   KW    KH    1    C    convolution kernel
+    // b:   W     H     C    N    input data
+    // res: W_out H_out C    N
+    GGML_API struct ggml_tensor * ggml_conv_2d_dw_direct(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            int                   stride0,
+            int                   stride1,
+            int                   pad0,
+            int                   pad1,
+            int                   dilation0,
+            int                   dilation1);
+
     GGML_API struct ggml_tensor * ggml_conv_transpose_2d_p0(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,

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

@@ -215,7 +215,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .nrows                    = 1,
     },
     [GGML_TYPE_F16] = {
-        .from_float               = (ggml_from_float_t) ggml_fp32_to_fp16_row,
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_fp16,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f16,
         .vec_dot_type             = GGML_TYPE_F16,
         .nrows                    = 1,
@@ -356,7 +356,7 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .from_float               = quantize_row_q8_K,
     },
     [GGML_TYPE_BF16] = {
-        .from_float               = (ggml_from_float_t) ggml_fp32_to_bf16_row,
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_bf16,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_bf16,
         .vec_dot_type             = GGML_TYPE_BF16,
         .nrows                    = 1,
@@ -1932,6 +1932,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_im2col_back_f32(params, tensor);
             } break;
+        case GGML_OP_CONV_2D_DW:
+            {
+                ggml_compute_forward_conv_2d_dw(params, tensor);
+            } break;
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
                 ggml_compute_forward_conv_transpose_2d(params, tensor);
@@ -2268,6 +2272,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_IM2COL:
         case GGML_OP_IM2COL_BACK:
+        case GGML_OP_CONV_2D_DW:
         case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
@@ -3161,6 +3166,93 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
     return ggml_graph_compute(cgraph, &cplan);
 }
 
+void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
+    int64_t i = 0;
+#if defined(__F16C__)
+#if defined(__AVX512F__)
+    for (; i + 15 < n; i += 16) {
+        __m512 x_vec = _mm512_loadu_ps(x + i);
+        __m256i y_vec = _mm512_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
+        _mm256_storeu_si256((__m256i *)(y + i), y_vec);
+    }
+#endif
+    for (; i + 7 < n; i += 8) {
+        __m256 x_vec = _mm256_loadu_ps(x + i);
+        __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
+        _mm_storeu_si128((__m128i *)(y + i), y_vec);
+    }
+    for (; i + 3 < n; i += 4) {
+        __m128 x_vec = _mm_loadu_ps(x + i);
+        __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
+        _mm_storel_epi64((__m128i *)(y + i), y_vec);
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(x[i]);
+    }
+}
+
+void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
+    int64_t i = 0;
+#if defined(__F16C__)
+#if defined(__AVX512F__)
+    for (; i + 15 < n; i += 16) {
+        __m256i x_vec = _mm256_loadu_si256((const __m256i *)(x + i));
+        __m512 y_vec = _mm512_cvtph_ps(x_vec);
+        _mm512_storeu_ps(y + i, y_vec);
+    }
+#endif
+    for (; i + 7 < n; i += 8) {
+        __m128i x_vec = _mm_loadu_si128((const __m128i *)(x + i));
+        __m256 y_vec = _mm256_cvtph_ps(x_vec);
+        _mm256_storeu_ps(y + i, y_vec);
+    }
+    for (; i + 3 < n; i += 4) {
+        __m128i x_vec = _mm_loadl_epi64((const __m128i *)(x + i));
+        __m128 y_vec = _mm_cvtph_ps(x_vec);
+        _mm_storeu_ps(y + i, y_vec);
+    }
+#endif
+    for (; i < n; ++i) {
+        y[i] = GGML_FP16_TO_FP32(x[i]);
+    }
+}
+
+void ggml_cpu_fp32_to_bf16(const float * x, ggml_bf16_t * y, int64_t n) {
+    int64_t i = 0;
+    for (; i < n; ++i) {
+        y[i] = GGML_FP32_TO_BF16(x[i]);
+    }
+}
+
+void ggml_cpu_bf16_to_fp32(const ggml_bf16_t * x, float * y, int64_t n) {
+    int64_t i = 0;
+#if defined(__AVX2__)
+#if defined(__AVX512F__)
+    for (; i + 15 < n; i += 16) {
+        _mm512_storeu_ps(y + i,
+                        _mm512_castsi512_ps(
+                            _mm512_slli_epi32(
+                                _mm512_cvtepu16_epi32(
+                                    _mm256_loadu_si256(
+                                        (const __m256i *)(x + i))),
+                                16)));
+    }
+#endif
+    for (; i + 7 < n; i += 8) {
+        _mm256_storeu_ps(y + i,
+                        _mm256_castsi256_ps(
+                            _mm256_slli_epi32(
+                                _mm256_cvtepu16_epi32(
+                                    _mm_loadu_si128(
+                                        (const __m128i *)(x + i))),
+                                16)));
+    }
+#endif
+    for (; i < n; i++) {
+        y[i] = GGML_BF16_TO_FP32(x[i]);
+    }
+}
 
 int ggml_cpu_has_avx(void) {
 #if defined(__AVX__)

ggml/src/ggml-cpu/ops.cpp

@@ -4222,7 +4222,7 @@ static void ggml_compute_forward_get_rows_f16(
 
         GGML_ASSERT(i01 >= 0 && i01 < ne01);
 
-        ggml_fp16_to_fp32_row(
+        ggml_cpu_fp16_to_fp32(
             (const ggml_fp16_t*) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
                        (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
     }
@@ -4263,7 +4263,7 @@ static void ggml_compute_forward_get_rows_bf16(
 
         GGML_ASSERT(i01 >= 0 && i01 < ne01);
 
-        ggml_bf16_to_fp32_row(
+        ggml_cpu_bf16_to_fp32(
             (const ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
                         (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
     }
@@ -6064,6 +6064,178 @@ void ggml_compute_forward_conv_transpose_2d(
     }
 }
 
+// ggml_compute_forward_conv_2d_dw
+
+struct ggml_conv_2d_dw_params {
+    int64_t channels;
+    int64_t batch;
+    int64_t src_w;
+    int64_t src_h;
+    int64_t dst_w;
+    int64_t dst_h;
+    int64_t knl_w;
+    int64_t knl_h;
+    int stride_x;
+    int stride_y;
+    int pad_x;
+    int pad_y;
+    int dilation_x;
+    int dilation_y;
+};
+
+static void ggml_compute_forward_conv_2d_dw_cwhn(
+        const ggml_compute_params * params,
+        const ggml_tensor * src,
+        const ggml_tensor * kernel,
+        ggml_tensor * dst,
+        const ggml_conv_2d_dw_params & p) {
+
+    const int64_t c = p.channels;
+    const float * knl_data = (const float *)kernel->data;
+
+    const int64_t rows_total = p.dst_h * p.batch;
+    const int64_t rows_per_thread = (rows_total + params->nth - 1) / params->nth;
+    const int64_t row_start = params->ith * rows_per_thread;
+    const int64_t row_end = MIN(row_start + rows_per_thread, rows_total);
+
+#ifdef GGML_SIMD
+    const int64_t pkg_size = GGML_F32_EPR;
+    const int64_t pkg_count = c / pkg_size;
+    const int64_t c_pkg_end = pkg_count * pkg_size;
+#else
+    const int64_t c_pkg_end = 0;
+#endif
+
+    for (int64_t row = row_start; row < row_end; ++row) {
+        const int64_t dst_y = row % p.dst_h;
+        const float * src_data = (const float *)src->data + (row / p.dst_h) * p.src_w * p.src_h * c;
+        for (int64_t dst_x = 0; dst_x < p.dst_w; ++dst_x) {
+            float * dst_data = (float *)dst->data + (row * p.dst_w + dst_x) * c;
+            const int64_t src_y_base = dst_y * p.stride_y - p.pad_y;
+            const int64_t src_x_base = dst_x * p.stride_x - p.pad_x;
+
+#ifdef GGML_SIMD
+            // Vectorized loop
+            for (int64_t c_i = 0; c_i < c_pkg_end; c_i += pkg_size) {
+                GGML_F32_VEC sum = GGML_F32_VEC_ZERO;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = src_y_base + knl_y * p.dilation_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = src_x_base + knl_x * p.dilation_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        GGML_F32_VEC k = GGML_F32_VEC_LOAD(knl_data + (knl_y * p.knl_w + knl_x) * c + c_i);
+                        GGML_F32_VEC s = GGML_F32_VEC_LOAD(src_data + (src_y * p.src_w + src_x) * c + c_i);
+                        sum = GGML_F32_VEC_FMA(sum, k, s);
+                    }
+                }
+                GGML_F32_VEC_STORE(dst_data + c_i, sum);
+            }
+#endif
+            // Scalar loop
+            for (int64_t c_i = c_pkg_end; c_i < c; ++c_i) {
+                float sum = 0.0f;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = src_y_base + knl_y * p.dilation_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = src_x_base + knl_x * p.dilation_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        sum += knl_data[(knl_y * p.knl_w + knl_x) * c + c_i]
+                             * src_data[(src_y * p.src_w + src_x) * c + c_i];
+                    }
+                }
+                dst_data[c_i] = sum;
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_conv_2d_dw_whcn(
+        const ggml_compute_params * params,
+        const ggml_tensor * src,
+        const ggml_tensor * kernel,
+        ggml_tensor * dst,
+        const ggml_conv_2d_dw_params & p) {
+
+    const int64_t n = p.channels * p.batch;
+    const int64_t per_thread = (n + params->nth - 1) / params->nth;
+    const int64_t start = params->ith * per_thread;
+    const int64_t end = MIN(start + per_thread, n);
+
+    for (int64_t i = start; i < end; ++i) {
+        const float * knl_data = (const float *)kernel->data + (i % p.channels) * p.knl_w * p.knl_h;
+        const float * src_data = (const float *)src->data + i * p.src_w * p.src_h;
+        float * dst_data = (float *)dst->data + i * p.dst_w * p.dst_h;
+
+        for (int64_t dst_y = 0; dst_y < p.dst_h; ++dst_y) {
+            for (int64_t dst_x = 0; dst_x < p.dst_w; ++dst_x) {
+
+                float sum = 0.0f;
+                for (int64_t knl_y = 0; knl_y < p.knl_h; ++knl_y) {
+                    const int64_t src_y = dst_y * p.stride_y + knl_y * p.dilation_y - p.pad_y;
+                    if (src_y < 0 || src_y >= p.src_h) {
+                        continue;
+                    }
+                    for (int64_t knl_x = 0; knl_x < p.knl_w; ++knl_x) {
+                        const int64_t src_x = dst_x * p.stride_x + knl_x * p.dilation_x - p.pad_x;
+                        if (src_x < 0 || src_x >= p.src_w) {
+                            continue;
+                        }
+                        sum += knl_data[knl_y * p.knl_w + knl_x]
+                             * src_data[src_y * p.src_w + src_x];
+                    }
+                }
+                dst_data[dst_y * p.dst_w + dst_x] = sum;
+            }
+        }
+    }
+}
+
+void ggml_compute_forward_conv_2d_dw(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * src = dst->src[1];
+    ggml_conv_2d_dw_params p;
+    p.channels = src->ne[2];
+    p.batch = src->ne[3];
+    p.src_w = src->ne[0];
+    p.src_h = src->ne[1];
+    p.dst_w = dst->ne[0];
+    p.dst_h = dst->ne[1];
+    p.knl_w = kernel->ne[0];
+    p.knl_h = kernel->ne[1];
+    p.stride_x = dst->op_params[0];
+    p.stride_y = dst->op_params[1];
+    p.pad_x = dst->op_params[2];
+    p.pad_y = dst->op_params[3];
+    p.dilation_x = dst->op_params[4];
+    p.dilation_y = dst->op_params[5];
+
+    GGML_ASSERT(kernel->ne[3] == p.channels);
+    GGML_ASSERT(dst->ne[3] == p.batch);
+
+    if (ggml_is_contiguous(src)) {
+        ggml_compute_forward_conv_2d_dw_whcn(params, src, kernel, dst, p);
+    } else if (ggml_is_contiguous_channels(src)) {
+        // kernel should also have channels most contiguous in memory
+        GGML_ASSERT(kernel->nb[0] >= kernel->nb[2] && kernel->nb[1] >= kernel->nb[0]);
+        ggml_compute_forward_conv_2d_dw_cwhn(params, src, kernel, dst, p);
+    } else {
+        GGML_ABORT("non-contiguous memory layout not supported");
+    }
+}
+
 // ggml_compute_forward_pool_1d_sk_p0
 
 static void ggml_compute_forward_pool_1d_sk_p0(

ggml/src/ggml-cpu/ops.h

@@ -65,6 +65,7 @@ void ggml_compute_forward_conv_transpose_1d(const struct ggml_compute_params * p
 void ggml_compute_forward_im2col(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_im2col_back_f32(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_conv_transpose_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_conv_2d_dw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_2d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cuda/common.cuh

@@ -78,13 +78,13 @@
 // Moore Threads
 #define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210)
 
-#define GGML_CUDA_CC_QY1  (GGML_MUSA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
-#define GGML_CUDA_CC_QY2  (GGML_MUSA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
-#define GGML_CUDA_CC_NG   (GGML_MUSA_CC_OFFSET_MTHREADS + 0x310) // TBD
+#define GGML_CUDA_CC_QY1  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
+#define GGML_CUDA_CC_QY2  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
+#define GGML_CUDA_CC_NG   (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
 
 #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
 #define GGML_CUDA_CC_IS_QY1(cc)      (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
-#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NEXT)
+#define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
 #define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)
 
 #ifdef __CUDA_ARCH_LIST__

ggml/src/ggml-cuda/cpy.cu

@@ -639,6 +639,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
     if(ctx.cuda_graph->use_cpy_indirection && !disable_indirection_for_this_node) {
         ctx.cuda_graph->graph_cpynode_index = graph_cpynode_index;
     }
+#else
+    GGML_UNUSED(disable_indirection_for_this_node);
 #endif
 
 }

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -1410,6 +1410,11 @@ static void ggml_cuda_op_mul_mat(
     const int64_t ne0 = dst->ne[0];
     const int64_t ne1 = dst->ne[1];
 
+    // const int64_t nb10 = src1->nb[0];
+    const int64_t nb11 = src1->nb[1];
+    const int64_t nb12 = src1->nb[2];
+    const int64_t nb13 = src1->nb[3];
+
     const int64_t nb2 = dst->nb[2];
     const int64_t nb3 = dst->nb[3];
 
@@ -1545,7 +1550,10 @@ static void ggml_cuda_op_mul_mat(
             dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(ctx.pool(id), src_1_ddq_size);
 
             if (src1_on_device && src1_is_contiguous) {
-                quantize_src1(dev[id].src1_ddf, dev[id].src1_ddq, ne10, ne11, ne12*ne13, src1_padded_col_size, src0->type, stream);
+                quantize_src1(
+                    dev[id].src1_ddf, dev[id].src1_ddq, src0->type, ne10,
+                    nb11/sizeof(float), nb12/sizeof(float), nb13/sizeof(float),
+                    src1_padded_col_size, ne11, ne12, ne13, stream);
                 CUDA_CHECK(cudaGetLastError());
             }
         }
@@ -1640,7 +1648,9 @@ static void ggml_cuda_op_mul_mat(
                 }
 
                 if (quantize_src1 && !src1_is_contiguous) {
-                    quantize_src1(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, 1, src1_padded_col_size, src0->type, stream);
+                    quantize_src1(
+                        src1_ddf_i, src1_ddq_i, src0->type, ne10, ne10, ne11*ne10, ne12*ne11*ne10,
+                        src1_padded_col_size, src1_ncols, 1, 1, stream);
                     CUDA_CHECK(cudaGetLastError());
                 }
 
@@ -1878,7 +1888,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
 static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const bool split = ggml_backend_buft_is_cuda_split(src0->buffer->buft);
 
-    bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
+    bool use_mul_mat_vec   = (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_BF16)
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
         && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
     bool use_mul_mat_vec_q = ggml_is_quantized(src0->type)
@@ -1919,12 +1929,14 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && use_mul_mat_vec && (src0->ne[1] < MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+    if (!split && use_mul_mat_vec && (src0->ne[1] <= MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
         // the custom F16 vector kernel can be used over batched cuBLAS GEMM
         // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
-        ggml_cuda_mul_mat_vec(ctx, src0, src1, dst);
-    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
-               && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+        ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst);
+    } else if (!split && use_mul_mat_vec_q) {
+        ggml_cuda_mul_mat_vec_q(ctx, src0, src1, nullptr, dst);
+    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16) &&
+            dst->op_params[0] == GGML_PREC_DEFAULT && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // general KQ + KQV multi-batch without FlashAttention
         ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
     } else if (use_mul_mat_vec) {
@@ -1999,6 +2011,15 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
+    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32 && ne2 == 1) {
+        if (ggml_is_quantized(src0->type)) {
+            ggml_cuda_mul_mat_vec_q(ctx, src0, src1, ids, dst);
+        } else {
+            ggml_cuda_mul_mat_vec(ctx, src0, src1, ids, dst);
+        }
+        return;
+    }
+
     GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
 
     cudaStream_t stream = ctx.stream();
@@ -2035,97 +2056,75 @@ static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor *
     dst_row.nb[2] = nb1;
     dst_row.nb[3] = nb1;
 
-    if (ne12 == 1) {
+    ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
+    ggml_cuda_pool_alloc<char>  dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));
+
+    src1_row.data = src1_contiguous.get();
+    dst_row.data  =  dst_contiguous.get();
+
+    for (int64_t i02 = 0; i02 < n_as; i02++) {
+        int64_t num_src1_rows = 0;
+
         for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
             for (int64_t id = 0; id < n_ids; id++) {
-                const int32_t i02 = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+                const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
 
-                GGML_ASSERT(i02 >= 0 && i02 < n_as);
+                GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
 
-                const int64_t i11 = id % ne11;
-                const int64_t i12 = iid1;
+                if (row_id_i != i02) {
+                    continue;
+                }
 
-                const int64_t i1 = id;
-                const int64_t i2 = i12;
-
-                src0_row.data = src0_original + i02*nb02;
-                src1_row.data = src1_original + i11*nb11 + i12*nb12;
-                dst_row.data  =  dst_original + i1*nb1   + i2*nb2;
-
-                ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
+                num_src1_rows++;
             }
         }
-    } else {
-        ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
-        ggml_cuda_pool_alloc<char>  dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));
 
-        src1_row.data = src1_contiguous.get();
-        dst_row.data  =  dst_contiguous.get();
+        if (num_src1_rows == 0) {
+            continue;
+        }
 
-        for (int64_t i02 = 0; i02 < n_as; i02++) {
-            int64_t num_src1_rows = 0;
+        ggml_cuda_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
+        ggml_cuda_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
+        CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream));
 
-            for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
-                for (int64_t id = 0; id < n_ids; id++) {
-                    const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+        {
+            dim3 block_dims(std::min((unsigned int)ne10, 768u));
+            dim3 grid_dims(ids->ne[1], n_ids);
+            k_copy_src1_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
+                    src1_original, src1_contiguous.get(),
+                    dev_cur_src1_row.get(), dev_row_mapping.get(),
+                    ids_dev, i02, ids->nb[1], ids->nb[0],
+                    ne11, ne10,
+                    nb11, nb12);
+            CUDA_CHECK(cudaGetLastError());
+        }
 
-                    GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
+        src0_row.data = src0_original + i02*nb02;
 
-                    if (row_id_i != i02) {
-                        continue;
-                    }
+        GGML_ASSERT(nb11 == sizeof(float)*ne10);
+        GGML_ASSERT(nb1 == sizeof(float)*ne0);
 
-                    num_src1_rows++;
-                }
-            }
+        src1_row.ne[1] = num_src1_rows;
+        src1_row.nb[1] = nb11;
+        src1_row.nb[2] = num_src1_rows*nb11;
+        src1_row.nb[3] = num_src1_rows*nb11;
 
-            if (num_src1_rows == 0) {
-                continue;
-            }
+        dst_row.ne[1] = num_src1_rows;
+        dst_row.nb[1] = nb1;
+        dst_row.nb[2] = num_src1_rows*nb1;
+        dst_row.nb[3] = num_src1_rows*nb1;
 
-            ggml_cuda_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
-            ggml_cuda_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
-            CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream));
+        ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
 
-            {
-                dim3 block_dims(std::min((unsigned int)ne10, 768u));
-                dim3 grid_dims(ids->ne[1], n_ids);
-                k_copy_src1_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                        src1_original, src1_contiguous.get(),
-                        dev_cur_src1_row.get(), dev_row_mapping.get(),
-                        ids_dev, i02, ids->nb[1], ids->nb[0],
-                        ne11, ne10,
-                        nb11, nb12);
-                CUDA_CHECK(cudaGetLastError());
-            }
-
-            src0_row.data = src0_original + i02*nb02;
-
-            GGML_ASSERT(nb11 == sizeof(float)*ne10);
-            GGML_ASSERT(nb1 == sizeof(float)*ne0);
-
-            src1_row.ne[1] = num_src1_rows;
-            src1_row.nb[1] = nb11;
-            src1_row.nb[2] = num_src1_rows*nb11;
-            src1_row.nb[3] = num_src1_rows*nb11;
-
-            dst_row.ne[1] = num_src1_rows;
-            dst_row.nb[1] = nb1;
-            dst_row.nb[2] = num_src1_rows*nb1;
-            dst_row.nb[3] = num_src1_rows*nb1;
-
-            ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
-
-            {
-                dim3 block_dims(std::min((unsigned int)ne0, 768u));
-                dim3 grid_dims(num_src1_rows);
-                k_copy_dst_from_contiguous<<<grid_dims, block_dims, 0, stream>>>(
-                        dst_original, dst_contiguous.get(),
-                        dev_row_mapping.get(),
-                        ne0,
-                        nb1, nb2);
-                CUDA_CHECK(cudaGetLastError());
-            }
+        {
+            dim3 block_dims(std::min((unsigned int)ne0, 768u));
+            dim3 grid_dims(num_src1_rows);
+            k_copy_dst_from_contiguous<<<grid_dims, block_dims, 0, stream>>>(
+                    dst_original, dst_contiguous.get(),
+                    dev_row_mapping.get(),
+                    ne0,
+                    nb1, nb2);
+            CUDA_CHECK(cudaGetLastError());
         }
     }
 }
@@ -2489,7 +2488,7 @@ static bool check_node_graph_compatibility_and_refresh_copy_ops(ggml_backend_cud
 #endif
         }
 
-        if (node->op == GGML_OP_MUL_MAT_ID) {
+        if (node->op == GGML_OP_MUL_MAT_ID && node->ne[2] != 1) {
             use_cuda_graph = false; // This node type is not supported by CUDA graph capture
 #ifndef NDEBUG
             GGML_LOG_DEBUG("%s: disabling CUDA graphs due to unsupported node type\n", __func__);
@@ -3203,9 +3202,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         }
         case GGML_OP_ROPE:
         case GGML_OP_ROPE_BACK: {
-            const size_t ts = ggml_type_size(op->src[0]->type);
-            const int64_t ne0_012 = op->src[0]->ne[0] * op->src[0]->ne[1] * op->src[0]->ne[2];
-            return op->src[0]->nb[0] == ts && op->src[0]->nb[3] == ne0_012*ts;
+            return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]);
         }
         case GGML_OP_IM2COL:
         case GGML_OP_POOL_2D:

ggml/src/ggml-cuda/mmq.cuh

@@ -155,25 +155,27 @@ static constexpr __device__ int get_mmq_y_device() {
 #define MMQ_DP4A_TXS_Q6_K    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE/QI6_K   + mmq_y/QI6_K,     mmq_y*WARP_SIZE/8 + mmq_y/8}
 
 static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int mmq_y) {
-    return type == GGML_TYPE_Q4_0 ? MMQ_DP4A_TXS_Q4_0 :
-        type == GGML_TYPE_Q4_1    ? MMQ_DP4A_TXS_Q4_1 :
-        type == GGML_TYPE_Q5_0    ? MMQ_DP4A_TXS_Q8_0 :
-        type == GGML_TYPE_Q5_1    ? MMQ_DP4A_TXS_Q8_1 :
-        type == GGML_TYPE_Q8_0    ? MMQ_DP4A_TXS_Q8_0 :
-        type == GGML_TYPE_Q2_K    ? MMQ_DP4A_TXS_Q2_K :
-        type == GGML_TYPE_Q3_K    ? MMQ_DP4A_TXS_Q3_K :
-        type == GGML_TYPE_Q4_K    ? MMQ_DP4A_TXS_Q4_K :
-        type == GGML_TYPE_Q5_K    ? MMQ_DP4A_TXS_Q5_K :
-        type == GGML_TYPE_Q6_K    ? MMQ_DP4A_TXS_Q6_K :
-        type == GGML_TYPE_IQ2_XXS ? MMQ_DP4A_TXS_Q8_0 :
-        type == GGML_TYPE_IQ2_XS  ? MMQ_DP4A_TXS_Q8_0_16 :
-        type == GGML_TYPE_IQ2_S   ? MMQ_DP4A_TXS_Q8_0_16 :
-        type == GGML_TYPE_IQ3_XXS ? MMQ_DP4A_TXS_Q8_0 :
-        type == GGML_TYPE_IQ3_S   ? MMQ_DP4A_TXS_Q8_0 :
-        type == GGML_TYPE_IQ1_S   ? MMQ_DP4A_TXS_Q8_0 :
-        type == GGML_TYPE_IQ4_XS  ? MMQ_DP4A_TXS_Q8_0 :
-        type == GGML_TYPE_IQ4_NL  ? MMQ_DP4A_TXS_Q8_0 :
-        tile_x_sizes{0, 0, 0};
+    switch (type) {
+        case GGML_TYPE_Q4_0:    return MMQ_DP4A_TXS_Q4_0;
+        case GGML_TYPE_Q4_1:    return MMQ_DP4A_TXS_Q4_1;
+        case GGML_TYPE_Q5_0:    return MMQ_DP4A_TXS_Q8_0;
+        case GGML_TYPE_Q5_1:    return MMQ_DP4A_TXS_Q8_1;
+        case GGML_TYPE_Q8_0:    return MMQ_DP4A_TXS_Q8_0;
+        case GGML_TYPE_Q2_K:    return MMQ_DP4A_TXS_Q2_K;
+        case GGML_TYPE_Q3_K:    return MMQ_DP4A_TXS_Q3_K;
+        case GGML_TYPE_Q4_K:    return MMQ_DP4A_TXS_Q4_K;
+        case GGML_TYPE_Q5_K:    return MMQ_DP4A_TXS_Q5_K;
+        case GGML_TYPE_Q6_K:    return MMQ_DP4A_TXS_Q6_K;
+        case GGML_TYPE_IQ2_XXS: return MMQ_DP4A_TXS_Q8_0;
+        case GGML_TYPE_IQ2_XS:  return MMQ_DP4A_TXS_Q8_0_16;
+        case GGML_TYPE_IQ2_S:   return MMQ_DP4A_TXS_Q8_0_16;
+        case GGML_TYPE_IQ3_XXS: return MMQ_DP4A_TXS_Q8_0;
+        case GGML_TYPE_IQ3_S:   return MMQ_DP4A_TXS_Q8_0;
+        case GGML_TYPE_IQ1_S:   return MMQ_DP4A_TXS_Q8_0;
+        case GGML_TYPE_IQ4_XS:  return MMQ_DP4A_TXS_Q8_0;
+        case GGML_TYPE_IQ4_NL:  return MMQ_DP4A_TXS_Q8_0;
+        default:                return tile_x_sizes{0, 0, 0};
+    }
 }
 
 #define MMQ_MMA_TILE_X_K_Q8_0 (2*WARP_SIZE + 2*WARP_SIZE/QI8_0                 + 4)
@@ -189,25 +191,27 @@ static_assert(MMQ_MMA_TILE_X_K_Q3_K % 8 == 4, "Wrong padding.");
 static_assert(MMQ_MMA_TILE_X_K_Q6_K % 8 == 4, "Wrong padding.");
 
 static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
-    return type == GGML_TYPE_Q4_0 ? MMQ_MMA_TILE_X_K_Q8_0 :
-        type == GGML_TYPE_Q4_1    ? MMQ_MMA_TILE_X_K_Q8_1 :
-        type == GGML_TYPE_Q5_0    ? MMQ_MMA_TILE_X_K_Q8_0 :
-        type == GGML_TYPE_Q5_1    ? MMQ_MMA_TILE_X_K_Q8_1 :
-        type == GGML_TYPE_Q8_0    ? MMQ_MMA_TILE_X_K_Q8_0 :
-        type == GGML_TYPE_Q2_K    ? MMQ_MMA_TILE_X_K_Q2_K :
-        type == GGML_TYPE_Q3_K    ? MMQ_MMA_TILE_X_K_Q3_K :
-        type == GGML_TYPE_Q4_K    ? MMQ_MMA_TILE_X_K_Q8_1 :
-        type == GGML_TYPE_Q5_K    ? MMQ_MMA_TILE_X_K_Q8_1 :
-        type == GGML_TYPE_Q6_K    ? MMQ_MMA_TILE_X_K_Q6_K :
-        type == GGML_TYPE_IQ2_XXS ? MMQ_MMA_TILE_X_K_Q8_0 :
-        type == GGML_TYPE_IQ2_XS  ? MMQ_MMA_TILE_X_K_Q3_K :
-        type == GGML_TYPE_IQ2_S   ? MMQ_MMA_TILE_X_K_Q3_K :
-        type == GGML_TYPE_IQ3_XXS ? MMQ_MMA_TILE_X_K_Q8_0 :
-        type == GGML_TYPE_IQ3_S   ? MMQ_MMA_TILE_X_K_Q8_0 :
-        type == GGML_TYPE_IQ1_S   ? MMQ_MMA_TILE_X_K_Q8_0 :
-        type == GGML_TYPE_IQ4_XS  ? MMQ_MMA_TILE_X_K_Q8_0 :
-        type == GGML_TYPE_IQ4_NL  ? MMQ_MMA_TILE_X_K_Q8_0 :
-        0;
+    switch (type) {
+        case GGML_TYPE_Q4_0:    return MMQ_MMA_TILE_X_K_Q8_0;
+        case GGML_TYPE_Q4_1:    return MMQ_MMA_TILE_X_K_Q8_1;
+        case GGML_TYPE_Q5_0:    return MMQ_MMA_TILE_X_K_Q8_0;
+        case GGML_TYPE_Q5_1:    return MMQ_MMA_TILE_X_K_Q8_1;
+        case GGML_TYPE_Q8_0:    return MMQ_MMA_TILE_X_K_Q8_0;
+        case GGML_TYPE_Q2_K:    return MMQ_MMA_TILE_X_K_Q2_K;
+        case GGML_TYPE_Q3_K:    return MMQ_MMA_TILE_X_K_Q3_K;
+        case GGML_TYPE_Q4_K:    return MMQ_MMA_TILE_X_K_Q8_1;
+        case GGML_TYPE_Q5_K:    return MMQ_MMA_TILE_X_K_Q8_1;
+        case GGML_TYPE_Q6_K:    return MMQ_MMA_TILE_X_K_Q6_K;
+        case GGML_TYPE_IQ2_XXS: return MMQ_MMA_TILE_X_K_Q8_0;
+        case GGML_TYPE_IQ2_XS:  return MMQ_MMA_TILE_X_K_Q3_K;
+        case GGML_TYPE_IQ2_S:   return MMQ_MMA_TILE_X_K_Q3_K;
+        case GGML_TYPE_IQ3_XXS: return MMQ_MMA_TILE_X_K_Q8_0;
+        case GGML_TYPE_IQ3_S:   return MMQ_MMA_TILE_X_K_Q8_0;
+        case GGML_TYPE_IQ1_S:   return MMQ_MMA_TILE_X_K_Q8_0;
+        case GGML_TYPE_IQ4_XS:  return MMQ_MMA_TILE_X_K_Q8_0;
+        case GGML_TYPE_IQ4_NL:  return MMQ_MMA_TILE_X_K_Q8_0;
+        default:                return 0;
+    }
 }
 
 #define MMQ_TILE_Y_K (WARP_SIZE + WARP_SIZE/QI8_1)

ggml/src/ggml-cuda/mmv.cu

@@ -4,18 +4,23 @@
 
 template <typename T, typename type_acc, int block_size>
 static __global__ void mul_mat_vec(
-        const T * __restrict__ x, const float * __restrict__ y, float * __restrict__ dst, const int64_t ncols2, const int64_t stride_row,
+        const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const int64_t ncols2, const int64_t nchannels_y, const int64_t stride_row,
         const int64_t channel_ratio, const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst,
         const int64_t sample_ratio, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst) {
-    const int64_t row       = blockIdx.x;
-    const int64_t channel   = blockIdx.y;
-    const int64_t sample    = blockIdx.z;
-    const int     tid       = threadIdx.x;
-    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+    const int64_t row         = blockIdx.x;
+    const int64_t channel_dst = blockIdx.y;
+    const int64_t channel_x   = ids ? ids[channel_dst]          : channel_dst / channel_ratio;
+    const int64_t channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
+    const int64_t sample_dst  = blockIdx.z;
+    const int64_t sample_x    = sample_dst / sample_ratio;
+    const int64_t sample_y    = sample_dst;
+    const int     tid         = threadIdx.x;
+    constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
 
-    x   +=  (sample/sample_ratio)*stride_sample_x   + (channel/channel_ratio)*stride_channel_x + row*stride_row;
-    y   +=   sample              *stride_sample_y   +  channel               *stride_channel_y;
-    dst +=   sample              *stride_sample_dst +  channel               *stride_channel_dst;
+    x   += sample_x  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
+    y   += sample_y  *stride_sample_y   + channel_y  *stride_channel_y;
+    dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst;
 
     const float2 * y2 = (const float2 *) y;
 
@@ -31,12 +36,19 @@ static __global__ void mul_mat_vec(
 
     float sumf = 0.0f;
 
-    if constexpr (std::is_same<T, half>::value) {
+    if constexpr (std::is_same<T, float>::value) {
+        const float2 * x2 = (const float2 *) x;
+
+        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            const float2 tmpx = x2[col2];
+            const float2 tmpy = y2[col2];
+            sumf += tmpx.x*tmpy.x;
+            sumf += tmpx.y*tmpy.y;
+        }
+    } else if constexpr (std::is_same<T, half>::value) {
         const half2 * x2 = (const half2 *) x;
 
         if (std::is_same<type_acc, float>::value) {
-            sumf = 0.0f;
-
             for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
                 const float2 tmpx = __half22float2(x2[col2]);
                 const float2 tmpy = y2[col2];
@@ -59,8 +71,6 @@ static __global__ void mul_mat_vec(
         }
     } else if constexpr (std::is_same<T, nv_bfloat16>::value) {
         const int * x2 = (const int *) x;
-        sumf = 0.0f;
-
         for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
             const int    tmpx = x2[col2];
             const float2 tmpy = y2[col2];
@@ -92,17 +102,17 @@ static __global__ void mul_mat_vec(
 
 template <typename T, typename type_acc>
 static void launch_mul_mat_vec_cuda(
-        const T * x, const float * y, float * dst,
-        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
-        const int64_t nsamples_y, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         cudaStream_t stream) {
     GGML_ASSERT(ncols      % 2 == 0);
     GGML_ASSERT(stride_row % 2 == 0);
-    GGML_ASSERT(nchannels_y % nchannels_x == 0);
-    GGML_ASSERT(nsamples_y  % nsamples_x  == 0);
-    const int64_t channel_ratio = nchannels_y / nchannels_x;
-    const int64_t sample_ratio  = nsamples_y  / nsamples_x;
+    GGML_ASSERT(ids || nchannels_dst % nchannels_x == 0);
+    GGML_ASSERT(       nsamples_dst  % nsamples_x  == 0);
+    const int64_t channel_ratio = nchannels_dst / nchannels_x;
+    const int64_t sample_ratio  = nsamples_dst  / nsamples_x;
     int device;
     int warp_size;
 
@@ -124,48 +134,48 @@ static void launch_mul_mat_vec_cuda(
     }
 
     const int smem = warp_size*sizeof(float);
-    const dim3 block_nums(nrows, nchannels_y, nsamples_y);
+    const dim3 block_nums(nrows, nchannels_dst, nsamples_dst);
     const dim3 block_dims(block_size_best, 1, 1);
     switch (block_size_best) {
         case   32: {
             mul_mat_vec<T, type_acc,  32><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case   64: {
             mul_mat_vec<T, type_acc,  64><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case   96: {
             mul_mat_vec<T, type_acc,  96><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  128: {
             mul_mat_vec<T, type_acc, 128><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  160: {
             mul_mat_vec<T, type_acc, 160><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  192: {
             mul_mat_vec<T, type_acc, 192><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  224: {
             mul_mat_vec<T, type_acc, 224><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         case  256: {
             mul_mat_vec<T, type_acc, 256><<<block_nums, block_dims, smem, stream>>>
-                (x, y, dst, ncols/2, stride_row, channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, channel_ratio, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -175,28 +185,28 @@ static void launch_mul_mat_vec_cuda(
 
 template<typename T>
 static void mul_mat_vec_cuda(
-        const T * x, const float * y, float * dst,
-        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y,
+        const T * x, const float * y, const int32_t * ids, float * dst,
+        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_dst,
         const int64_t stride_channel_x, const int64_t stride_channel_y, const int64_t stride_channel_dst, const int64_t nsamples_x,
-        const int64_t nsamples_y, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
+        const int64_t nsamples_dst, const int64_t stride_sample_x, const int64_t stride_sample_y, const int64_t stride_sample_dst,
         enum ggml_prec prec, cudaStream_t stream) {
-    switch (prec) {
-        case GGML_PREC_DEFAULT: {
+    if constexpr(std::is_same<T, half>::value) {
+        if (prec == GGML_PREC_DEFAULT) {
             launch_mul_mat_vec_cuda<T, half>
-                (x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
-        } break;
-        case GGML_PREC_F32: {
-            launch_mul_mat_vec_cuda<T, float>
-                (x, y, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
-        } break;
+                (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+                 stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+            return;
+        }
     }
+    launch_mul_mat_vec_cuda<T, float>
+        (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+         stride_channel_dst, nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
 }
 
-void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(!ids ||  ids->type == GGML_TYPE_I32);
+    GGML_ASSERT(         dst->type == GGML_TYPE_F32);
 
     GGML_TENSOR_BINARY_OP_LOCALS;
 
@@ -204,21 +214,24 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
     const size_t ts_src1 = ggml_type_size(src1->type);
     const size_t ts_dst  = ggml_type_size(dst->type);
 
-    GGML_ASSERT(ne11 == 1);
-    GGML_ASSERT(ne12 == ne2);
+    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for  batch size 1.
     GGML_ASSERT(ne13 == ne3);
 
-    GGML_ASSERT(nb00 == ts_src0);
-    GGML_ASSERT(nb10 == ts_src1);
-    GGML_ASSERT(nb0  == ts_dst);
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
+    GGML_ASSERT(        nb0        == ts_dst);
 
     const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
     const enum ggml_prec prec = fast_fp16_available(cc) ? ggml_prec(dst->op_params[0]) : GGML_PREC_F32;
 
-    const float * src1_d = (const float *) src1->data;
-    float       *  dst_d = (float       *)  dst->data;
+    const float   * src1_d =       (const float   *) src1->data;
+    const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
+    float         *  dst_d =       (float         *)  dst->data;
 
     const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s11 = src1->nb[1] / ts_src1;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
     const int64_t s02 = src0->nb[2] / ts_src0;
     const int64_t s12 = src1->nb[2] / ts_src1;
     const int64_t s2  =  dst->nb[2] / ts_dst;
@@ -226,14 +239,33 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
     const int64_t s13 = src1->nb[3] / ts_src1;
     const int64_t s3  =  dst->nb[3] / ts_dst;
 
+    // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
+    const int64_t ncols_dst          = ids ? ne2  : ne1;
+    const int64_t nchannels_y        = ids ? ne11 : ne12;
+    const int64_t nchannels_dst      = ids ? ne1  : ne2;
+    const int64_t stride_channel_dst = ids ? s1   : s2;
+    const int64_t stride_channel_y   = ids ? s11  : s12;
+
+    GGML_ASSERT(ncols_dst == 1);
+
     switch (src0->type) {
+        case GGML_TYPE_F32: {
+            const float * src0_d = (const float *) src0->data;
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
+        } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0->data;
-            mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, s01, ne02, ne12, s02, s12, s2, ne03, ne13, s03, s13, s3, prec, ctx.stream());
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0->data;
-            mul_mat_vec_cuda(src0_d, src1_d, dst_d, ne00, ne01, s01, ne02, ne12, s02, s12, s2, ne03, ne13, s03, s13, s3, prec, ctx.stream());
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+                ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+                ne03,              ne3,           s03, s13,              s3,                 prec, ctx.stream());
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));
@@ -262,27 +294,34 @@ void ggml_cuda_op_mul_mat_vec(
     const int64_t stride_row         = ne00;
     const int64_t nchannels_x        = 1;
     const int64_t nchannels_y        = 1;
+    const int64_t nchannels_dst      = 1;
     const int64_t stride_channel_x   = 0;
     const int64_t stride_channel_y   = 0;
     const int64_t stride_channel_dst = 0;
     const int64_t nsamples_x         = 1;
-    const int64_t nsamples_y         = 1;
+    const int64_t nsamples_dst       = 1;
     const int64_t stride_sample_x    = 0;
     const int64_t stride_sample_y    = 0;
     const int64_t stride_sample_dst  = 0;
 
     switch (src0->type) {
+        case GGML_TYPE_F32: {
+            const float * src0_d = (const float *) src0_dd_i;
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+        } break;
         case GGML_TYPE_F16: {
             const half * src0_d = (const half *) src0_dd_i;
-            mul_mat_vec_cuda(src0_d, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
-                nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
         case GGML_TYPE_BF16: {
             const nv_bfloat16 * src0_d = (const nv_bfloat16 *) src0_dd_i;
-            mul_mat_vec_cuda(src0_d, src1_ddf_i, dst_dd_i, ne00, row_diff, stride_row,
-                nchannels_x, nchannels_y, stride_channel_x, stride_channel_y, stride_channel_dst,
-                nsamples_x, nsamples_y, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+                nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, prec, stream);
         } break;
         default:
             GGML_ABORT("unsupported type: %s", ggml_type_name(src0->type));

ggml/src/ggml-cuda/mmv.cuh

@@ -3,7 +3,7 @@
 // maximum number of src0 rows with which to use mul_mat_vec over cuBLAS if FP16 tensor cores are available
 #define MMV_MAX_ROWS 512
 
-void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
 
 void ggml_cuda_op_mul_mat_vec(
     ggml_backend_cuda_context & ctx,

ggml/src/ggml-cuda/mmvq.cu

@@ -1,50 +1,57 @@
 #include "mmvq.cuh"
+#include "quantize.cuh"
 #include "vecdotq.cuh"
 
+#include <cstdint>
+
 typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);
 
 static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
-    return type == GGML_TYPE_Q4_0 ? vec_dot_q4_0_q8_1 :
-        type == GGML_TYPE_Q4_1 ? vec_dot_q4_1_q8_1 :
-        type == GGML_TYPE_Q5_0 ? vec_dot_q5_0_q8_1 :
-        type == GGML_TYPE_Q5_1 ? vec_dot_q5_1_q8_1 :
-        type == GGML_TYPE_Q8_0 ? vec_dot_q8_0_q8_1 :
-        type == GGML_TYPE_Q2_K ? vec_dot_q2_K_q8_1 :
-        type == GGML_TYPE_Q3_K ? vec_dot_q3_K_q8_1 :
-        type == GGML_TYPE_Q4_K ? vec_dot_q4_K_q8_1 :
-        type == GGML_TYPE_Q5_K ? vec_dot_q5_K_q8_1 :
-        type == GGML_TYPE_Q6_K ? vec_dot_q6_K_q8_1 :
-        type == GGML_TYPE_IQ2_XXS ? vec_dot_iq2_xxs_q8_1 :
-        type == GGML_TYPE_IQ2_XS ? vec_dot_iq2_xs_q8_1 :
-        type == GGML_TYPE_IQ2_S ? vec_dot_iq2_s_q8_1 :
-        type == GGML_TYPE_IQ3_XXS ? vec_dot_iq3_xxs_q8_1 :
-        type == GGML_TYPE_IQ1_S ? vec_dot_iq1_s_q8_1 :
-        type == GGML_TYPE_IQ1_M ? vec_dot_iq1_m_q8_1 :
-        type == GGML_TYPE_IQ4_NL ? vec_dot_iq4_nl_q8_1 :
-        type == GGML_TYPE_IQ4_XS ? vec_dot_iq4_xs_q8_1 :
-        type == GGML_TYPE_IQ3_S ? vec_dot_iq3_s_q8_1 :
-        nullptr;
+    switch (type) {
+        case GGML_TYPE_Q4_0:    return vec_dot_q4_0_q8_1;
+        case GGML_TYPE_Q4_1:    return vec_dot_q4_1_q8_1;
+        case GGML_TYPE_Q5_0:    return vec_dot_q5_0_q8_1;
+        case GGML_TYPE_Q5_1:    return vec_dot_q5_1_q8_1;
+        case GGML_TYPE_Q8_0:    return vec_dot_q8_0_q8_1;
+        case GGML_TYPE_Q2_K:    return vec_dot_q2_K_q8_1;
+        case GGML_TYPE_Q3_K:    return vec_dot_q3_K_q8_1;
+        case GGML_TYPE_Q4_K:    return vec_dot_q4_K_q8_1;
+        case GGML_TYPE_Q5_K:    return vec_dot_q5_K_q8_1;
+        case GGML_TYPE_Q6_K:    return vec_dot_q6_K_q8_1;
+        case GGML_TYPE_IQ2_XXS: return vec_dot_iq2_xxs_q8_1;
+        case GGML_TYPE_IQ2_XS:  return vec_dot_iq2_xs_q8_1;
+        case GGML_TYPE_IQ2_S:   return vec_dot_iq2_s_q8_1;
+        case GGML_TYPE_IQ3_XXS: return vec_dot_iq3_xxs_q8_1;
+        case GGML_TYPE_IQ1_S:   return vec_dot_iq1_s_q8_1;
+        case GGML_TYPE_IQ1_M:   return vec_dot_iq1_m_q8_1;
+        case GGML_TYPE_IQ4_NL:  return vec_dot_iq4_nl_q8_1;
+        case GGML_TYPE_IQ4_XS:  return vec_dot_iq4_xs_q8_1;
+        case GGML_TYPE_IQ3_S:   return vec_dot_iq3_s_q8_1;
+        default:                return nullptr;
+    }
 }
 
 static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
-    return type == GGML_TYPE_Q4_0 ? VDR_Q4_0_Q8_1_MMVQ :
-        type == GGML_TYPE_Q4_1    ? VDR_Q4_1_Q8_1_MMVQ :
-        type == GGML_TYPE_Q5_0    ? VDR_Q5_0_Q8_1_MMVQ :
-        type == GGML_TYPE_Q5_1    ? VDR_Q5_1_Q8_1_MMVQ :
-        type == GGML_TYPE_Q8_0    ? VDR_Q8_0_Q8_1_MMVQ :
-        type == GGML_TYPE_Q2_K    ? VDR_Q2_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q3_K    ? VDR_Q3_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q4_K    ? VDR_Q4_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q5_K    ? VDR_Q5_K_Q8_1_MMVQ :
-        type == GGML_TYPE_Q6_K    ? VDR_Q6_K_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ2_XXS ? VDR_IQ2_XXS_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ2_XS  ? VDR_IQ2_XS_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ2_S   ? VDR_IQ2_S_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ3_XXS ? VDR_IQ3_XXS_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ3_S   ? VDR_IQ3_S_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ4_NL  ? VDR_IQ4_NL_Q8_1_MMVQ :
-        type == GGML_TYPE_IQ4_XS  ? VDR_IQ4_XS_Q8_1_MMVQ :
-        1;
+    switch (type) {
+        case GGML_TYPE_Q4_0:    return VDR_Q4_0_Q8_1_MMVQ;
+        case GGML_TYPE_Q4_1:    return VDR_Q4_1_Q8_1_MMVQ;
+        case GGML_TYPE_Q5_0:    return VDR_Q5_0_Q8_1_MMVQ;
+        case GGML_TYPE_Q5_1:    return VDR_Q5_1_Q8_1_MMVQ;
+        case GGML_TYPE_Q8_0:    return VDR_Q8_0_Q8_1_MMVQ;
+        case GGML_TYPE_Q2_K:    return VDR_Q2_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q3_K:    return VDR_Q3_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q4_K:    return VDR_Q4_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q5_K:    return VDR_Q5_K_Q8_1_MMVQ;
+        case GGML_TYPE_Q6_K:    return VDR_Q6_K_Q8_1_MMVQ;
+        case GGML_TYPE_IQ2_XXS: return VDR_IQ2_XXS_Q8_1_MMVQ;
+        case GGML_TYPE_IQ2_XS:  return VDR_IQ2_XS_Q8_1_MMVQ;
+        case GGML_TYPE_IQ2_S:   return VDR_IQ2_S_Q8_1_MMVQ;
+        case GGML_TYPE_IQ3_XXS: return VDR_IQ3_XXS_Q8_1_MMVQ;
+        case GGML_TYPE_IQ3_S:   return VDR_IQ3_S_Q8_1_MMVQ;
+        case GGML_TYPE_IQ4_NL:  return VDR_IQ4_NL_Q8_1_MMVQ;
+        case GGML_TYPE_IQ4_XS:  return VDR_IQ4_XS_Q8_1_MMVQ;
+        default:                return 1;
+    }
 }
 
 enum mmvq_parameter_table_id {
@@ -73,9 +80,9 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
     return MMVQ_PARAMETERS_GENERIC;
 }
 
-static constexpr __host__ __device__ int calc_nwarps(int ncols_y,  mmvq_parameter_table_id table_id) {
+static constexpr __host__ __device__ int calc_nwarps(int ncols_dst,  mmvq_parameter_table_id table_id) {
     if (table_id == MMVQ_PARAMETERS_GENERIC) {
-        switch (ncols_y) {
+        switch (ncols_dst) {
             case 1:
             case 2:
             case 3:
@@ -90,7 +97,7 @@ static constexpr __host__ __device__ int calc_nwarps(int ncols_y,  mmvq_paramete
                 return 1;
         }
     } else if (table_id == MMVQ_PARAMETERS_GCN) {
-        switch (ncols_y) {
+        switch (ncols_dst) {
             case 1:
             case 2:
             case 3:
@@ -107,9 +114,9 @@ static constexpr __host__ __device__ int calc_nwarps(int ncols_y,  mmvq_paramete
     return 1;
 }
 
-static constexpr __host__ __device__ int calc_rows_per_block(int ncols_y, int table_id) {
+static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id) {
     if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN) {
-        switch (ncols_y) {
+        switch (ncols_dst) {
             case 1:
                 return 1;
             case 2:
@@ -127,19 +134,21 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_y, int ta
     return 1;
 }
 
-template <ggml_type type, int ncols_y>
+template <ggml_type type, int ncols_dst>
 // tell the compiler to use as many registers as it wants, see nwarps definition below
-__launch_bounds__(calc_nwarps(ncols_y, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
+__launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
+        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, float * __restrict__ dst,
+        const int ncols_x, const int nchannels_y, const int stride_row_x, const int stride_col_y, const int stride_col_dst,
+        const int channel_ratio, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
 
     constexpr int qk  = ggml_cuda_type_traits<type>::qk;
     constexpr int qi  = ggml_cuda_type_traits<type>::qi;
     constexpr int vdr = get_vdr_mmvq(type);
     constexpr mmvq_parameter_table_id table_id = get_device_table_id();
-    constexpr int nwarps = calc_nwarps(ncols_y, table_id);
-    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_y, table_id);
+    constexpr int nwarps = calc_nwarps(ncols_dst, table_id);
+    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id);
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
@@ -147,13 +156,21 @@ static __global__ void mul_mat_vec_q(
     const     int tid = warp_size*threadIdx.y + threadIdx.x;
     const     int row0 = rows_per_cuda_block*blockIdx.x;
     const     int blocks_per_row_x = ncols_x / qk;
-    const     int blocks_per_col_y = nrows_y / QK8_1;
     constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
 
-    // partial sum for each thread
-    float tmp[ncols_y][rows_per_cuda_block] = {{0.0f}};
+    // The MUL_MAT_ID code path with ids != nullptr is only implemetned for ncols_dst == 1.
+    const int channel_dst = blockIdx.y;
+    const int channel_x   = ncols_dst == 1 && ids ? ids[channel_dst]          : channel_dst / channel_ratio;
+    const int channel_y   = ncols_dst == 1 && ids ? channel_dst % nchannels_y : channel_dst;
+    const int sample_dst  = blockIdx.z;
+    const int sample_x    = sample_dst / sample_ratio;
+    const int sample_y    = sample_dst;
 
-    const block_q8_1 * y = (const block_q8_1 *) vy;
+    // partial sum for each thread
+    float tmp[ncols_dst][rows_per_cuda_block] = {{0.0f}};
+
+    const block_q8_1 * y = ((const block_q8_1 *) vy) + sample_y*stride_sample_y + channel_y*stride_channel_y;
+    const int kbx_offset = sample_x*stride_sample_x + channel_x*stride_channel_x + row0*stride_row_x;
 
     for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
         const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx
@@ -162,18 +179,19 @@ static __global__ void mul_mat_vec_q(
         const int kqs = vdr * (tid % (qi/vdr));
 
 #pragma unroll
-        for (int j = 0; j < ncols_y; ++j) {
+        for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
             for (int i = 0; i < rows_per_cuda_block; ++i) {
-                tmp[j][i] += vec_dot_q_cuda(vx, &y[j*blocks_per_col_y + kby], (row0 + i)*blocks_per_row_x + kbx, kqs);
+                tmp[j][i] += vec_dot_q_cuda(
+                    vx, &y[j*stride_col_y + kby], kbx_offset + i*stride_row_x + kbx, kqs);
             }
         }
     }
 
-    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][warp_size];
+    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
     if (threadIdx.y > 0) {
 #pragma unroll
-        for (int j = 0; j < ncols_y; ++j) {
+        for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
             for (int i = 0; i < rows_per_cuda_block; ++i) {
                 tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
@@ -185,9 +203,11 @@ static __global__ void mul_mat_vec_q(
         return;
     }
 
+    dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst + row0;
+
     // sum up partial sums and write back result
 #pragma unroll
-    for (int j = 0; j < ncols_y; ++j) {
+    for (int j = 0; j < ncols_dst; ++j) {
 #pragma unroll
         for (int i = 0; i < rows_per_cuda_block; ++i) {
 #pragma unroll
@@ -197,88 +217,121 @@ static __global__ void mul_mat_vec_q(
             tmp[j][i] = warp_reduce_sum<warp_size>(tmp[j][i]);
         }
 
-        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < (unsigned)nrows_dst)) {
-            dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
+        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + int(threadIdx.x) < stride_col_dst)) {
+            dst[j*stride_col_dst + threadIdx.x] = tmp[j][threadIdx.x];
         }
     }
-
-    GGML_UNUSED(nrows_x);
 }
 
-static std::pair<dim3, dim3> calc_launch_params(const int ncols_y, const int nrows_x, const int warp_size, const mmvq_parameter_table_id table_id) {
-    const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_y, table_id) - 1) / calc_rows_per_block(ncols_y, table_id);
-    const dim3 block_nums(nblocks, 1, 1);
-    const dim3 block_dims(warp_size, calc_nwarps(ncols_y, table_id), 1);
+static std::pair<dim3, dim3> calc_launch_params(
+        const int ncols_dst, const int nrows_x, const int nchannels_y, const int nsamples_y,
+        const int warp_size, const mmvq_parameter_table_id table_id) {
+    const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_dst, table_id) - 1) / calc_rows_per_block(ncols_dst, table_id);
+    const dim3 block_nums(nblocks, nchannels_y, nsamples_y);
+    const dim3 block_dims(warp_size, calc_nwarps(ncols_dst, table_id), 1);
     return {block_nums, block_dims};
 }
 
 template <ggml_type type>
-static void mul_mat_vec_q_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+static void mul_mat_vec_q_switch_ncols_dst(
+        const void * vx, const void * vy, const int32_t * ids, float * dst,
+        const int ncols_x, const int nrows_x, const int ncols_dst,
+        const int stride_row_x, const int stride_col_y, const int stride_col_dst,
+        const int nchannels_x, const int nchannels_y, const int nchannels_dst,
+        const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        cudaStream_t stream) {
 
     GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
-    GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE);
+    GGML_ASSERT(ncols_dst <= MMVQ_MAX_BATCH_SIZE);
+
+    const int channel_ratio = nchannels_dst / nchannels_x;
+    const int sample_ratio  = nsamples_dst  / nsamples_x;
 
     const int device = ggml_cuda_get_device();
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
     const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
 
-    switch (ncols_y) {
+    GGML_ASSERT(!ids || ncols_dst == 1);
+    switch (ncols_dst) {
         case 1:
         {
-            constexpr int c_ncols_y = 1;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 1;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 2:
         {
-            constexpr int c_ncols_y = 2;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 2;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 3:
         {
-            constexpr int c_ncols_y = 3;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 3;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 4:
         {
-            constexpr int c_ncols_y = 4;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 4;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 5:
         {
-            constexpr int c_ncols_y = 5;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 5;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 6:
         {
-            constexpr int c_ncols_y = 6;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 6;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 7:
         {
-            constexpr int c_ncols_y = 7;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 7;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         case 8:
         {
-            constexpr int c_ncols_y = 8;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_y, nrows_x, warp_size, table_id);
-            mul_mat_vec_q<type, c_ncols_y><<<dims.first, dims.second, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            constexpr int c_ncols_dst = 8;
+            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            mul_mat_vec_q<type, c_ncols_dst><<<dims.first, dims.second, 0, stream>>>
+                (vx, vy, ids, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+                 channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
             break;
         }
         default:
@@ -287,137 +340,213 @@ static void mul_mat_vec_q_cuda(
     }
 }
 
-static void mul_mat_vec_q4_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q4_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+static void mul_mat_vec_q_switch_type(
+        const void * vx, const ggml_type type_x, const void * vy, const int32_t * ids, float * dst,
+        const int ncols_x, const int nrows_x, const int ncols_dst,
+        const int stride_row_x, const int stride_col_y, const int stride_col_dst,
+        const int nchannels_x, const int nchannels_y, const int nchannels_dst,
+        const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
+        cudaStream_t stream) {
+    switch (type_x) {
+        case GGML_TYPE_Q4_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ2_XXS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ2_S:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ3_XXS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ1_S:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ1_M:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ4_XS:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
+                (vx, vy, ids, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst,
+                 stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
 }
 
-static void mul_mat_vec_q4_1_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+void ggml_cuda_mul_mat_vec_q(
+        ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst) {
+    GGML_ASSERT(        src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(        dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ids || ids->type  == GGML_TYPE_I32); // Optional, used for batched GGML_MUL_MAT_ID.
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q4_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    GGML_TENSOR_BINARY_OP_LOCALS;
 
-static void mul_mat_vec_q5_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    cudaStream_t stream = ctx.stream();
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q5_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    const size_t ts_src0 = ggml_type_size(src0->type);
+    const size_t ts_src1 = ggml_type_size(src1->type);
+    const size_t ts_dst  = ggml_type_size(dst->type);
 
-static void mul_mat_vec_q5_1_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    GGML_ASSERT(        nb00       == ts_src0);
+    GGML_ASSERT(        nb10       == ts_src1);
+    GGML_ASSERT(        nb0        == ts_dst);
+    GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q5_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for  batch size 1.
 
-static void mul_mat_vec_q8_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    const float   * src1_d =       (const float   *) src1->data;
+    const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
+    float         *  dst_d =       (float         *)  dst->data;
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q8_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING);
+    ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1);
+    {
+        const int64_t s11 = src1->nb[1] / ts_src1;
+        const int64_t s12 = src1->nb[2] / ts_src1;
+        const int64_t s13 = src1->nb[3] / ts_src1;
+        quantize_row_q8_1_cuda(src1_d, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+    }
 
-static void mul_mat_vec_q2_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    const int64_t s01 = src0->nb[1] / ts_src0;
+    const int64_t s11 = ne10_padded / QK8_1;
+    const int64_t s1  =  dst->nb[1] / ts_dst;
+    const int64_t s02 = src0->nb[2] / ts_src0;
+    const int64_t s2  =  dst->nb[2] / ts_dst;
+    const int64_t s03 = src0->nb[3] / ts_src0;
+    const int64_t s3  =  dst->nb[3] / ts_dst;
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q2_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
+    const int64_t s12 = ne11*s11;
+    const int64_t s13 = ne12*s12;
 
-static void mul_mat_vec_q3_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+    // For MUL_MAT_ID the memory layout is different than for MUL_MAT:
+    const int64_t ncols_dst          = ids ? ne2  : ne1;
+    const int64_t nchannels_y        = ids ? ne11 : ne12;
+    const int64_t nchannels_dst      = ids ? ne1  : ne2;
+    const int64_t stride_col_dst     = ids ? s2   : s1;
+    const int64_t stride_col_y       = ids ? s12  : s11;
+    const int64_t stride_channel_dst = ids ? s1   : s2;
+    const int64_t stride_channel_y   = ids ? s11  : s12;
 
-    mul_mat_vec_q_cuda<GGML_TYPE_Q3_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_q4_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q4_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_q5_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q5_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_q6_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_Q6_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq2_xxs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq2_xs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq2_s_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq3_xxs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ3_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq1_s_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ1_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq1_m_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ1_M>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq4_nl_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ4_NL>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq4_xs_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ4_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
-}
-
-static void mul_mat_vec_iq3_s_q8_1_cuda(
-    const void * vx, const void * vy, float * dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
-
-    mul_mat_vec_q_cuda<GGML_TYPE_IQ3_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+    mul_mat_vec_q_switch_type(
+        src0->data, src0->type, src1_q8_1.get(), ids_d, dst_d, ne00,
+        ne01,              ncols_dst,     s01, stride_col_y,     stride_col_dst,
+        ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
+        ne03,              ne3,           s03, s13,              s3,                 stream);
 }
 
 void ggml_cuda_op_mul_mat_vec_q(
@@ -440,68 +569,12 @@ void ggml_cuda_op_mul_mat_vec_q(
     // nrows_dst == nrows of the matrix that the kernel writes into
     const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
 
-    switch (src0->type) {
-        case GGML_TYPE_Q4_0:
-            mul_mat_vec_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            mul_mat_vec_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            mul_mat_vec_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            mul_mat_vec_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            mul_mat_vec_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q2_K:
-            mul_mat_vec_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q3_K:
-            mul_mat_vec_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q4_K:
-            mul_mat_vec_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_K:
-            mul_mat_vec_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q6_K:
-            mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ2_XXS:
-            mul_mat_vec_iq2_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ2_XS:
-            mul_mat_vec_iq2_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ2_S:
-            mul_mat_vec_iq2_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ3_XXS:
-            mul_mat_vec_iq3_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ1_S:
-            mul_mat_vec_iq1_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ1_M:
-            mul_mat_vec_iq1_m_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ4_NL:
-            mul_mat_vec_iq4_nl_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ4_XS:
-            mul_mat_vec_iq4_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        case GGML_TYPE_IQ3_S:
-            mul_mat_vec_iq3_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
-            break;
-        default:
-            GGML_ABORT("fatal error");
-            break;
-    }
+    const int stride_row_x = ne00 / ggml_blck_size(src0->type);
+    const int stride_col_y = src1_padded_row_size / QK8_1;
+
+    mul_mat_vec_q_switch_type(
+        src0_dd_i, src0->type, src1_ddq_i, nullptr, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
+        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
 
     GGML_UNUSED(src1);
     GGML_UNUSED(dst);

ggml/src/ggml-cuda/mmvq.cuh

@@ -2,6 +2,9 @@
 
 #define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
 
+void ggml_cuda_mul_mat_vec_q(ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
+
 void ggml_cuda_op_mul_mat_vec_q(
     ggml_backend_cuda_context & ctx,
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,

ggml/src/ggml-cuda/quantize.cu

@@ -1,30 +1,40 @@
 #include "quantize.cuh"
 #include <cstdint>
 
-static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int64_t kx, const int64_t kx0_padded) {
-    const int64_t ix0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
+static __global__ void quantize_q8_1(
+        const float * __restrict__ x, void * __restrict__ vy,
+        const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+        const int64_t ne0, const int ne1, const int ne2) {
+    const int64_t i0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
 
-    if (ix0 >= kx0_padded) {
+    if (i0 >= ne0) {
         return;
     }
 
-    const int64_t ix1 = blockIdx.y;
+    const int64_t i1 = blockIdx.y;
+    const int64_t i2 = blockIdx.z % ne2;
+    const int64_t i3 = blockIdx.z / ne2;
 
-    const int64_t i_padded = ix1*kx0_padded + ix0;
+    const int64_t & i00 = i0;
+    const int64_t & i01 = i1;
+    const int64_t & i02 = i2;
+    const int64_t & i03 = i3;
+
+    const int64_t i_cont = ((i3*ne2 + i2) * ne1 + i1) * ne0 + i0;
 
     block_q8_1 * y = (block_q8_1 *) vy;
 
-    const int64_t ib = i_padded / QK8_1; // block index
-    const int64_t iqs = i_padded % QK8_1; // quant index
+    const int64_t ib  = i_cont / QK8_1; // block index
+    const int64_t iqs = i_cont % QK8_1; // quant index
 
-    const float xi = ix0 < kx ? x[ix1*kx + ix0] : 0.0f;
+    const float xi = i0 < ne00 ? x[i03*s03 + i02*s02 + i01*s01 + i00] : 0.0f;
     float amax = fabsf(xi);
     float sum = xi;
 
     amax = warp_reduce_max(amax);
-    sum = warp_reduce_sum(sum);
+    sum  = warp_reduce_sum(sum);
 
-    const float d = amax / 127;
+    const float  d = amax / 127;
     const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
 
     y[ib].qs[iqs] = q;
@@ -127,43 +137,45 @@ static __global__ void quantize_mmq_q8_1(
 }
 
 void quantize_row_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels,
-    const int64_t kx0_padded, const ggml_type type_x, cudaStream_t stream) {
+    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
 
-    GGML_ASSERT(kx0_padded % QK8_1 == 0);
+    GGML_ASSERT(ne0 % QK8_1 == 0);
 
-    const int64_t block_num_x = (kx0_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
-    const dim3 num_blocks(block_num_x, kx1*channels, 1);
+    const int64_t block_num_x = (ne0 + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
+    const dim3 num_blocks(block_num_x, ne1, ne2*ne3);
     const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
-    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx0_padded);
-
-    GGML_UNUSED(type_x);
+    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
+    GGML_UNUSED(type_src0);
 }
 
 void quantize_mmq_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels,
-    const int64_t kx0_padded, const ggml_type type_x, cudaStream_t stream) {
+    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream) {
 
-    GGML_ASSERT(kx0_padded % (4*QK8_1) == 0);
+    GGML_ASSERT(ne0 % (4*QK8_1) == 0);
 
-    const int64_t block_num_x = (kx0_padded + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
-    const dim3 num_blocks(block_num_x, kx1, channels);
+    const int64_t block_num_x = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
+    const dim3 num_blocks(block_num_x, ne1, ne2*ne3);
     const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE_MMQ, 1, 1);
-    switch (mmq_get_q8_1_ds_layout(type_x)) {
+    switch (mmq_get_q8_1_ds_layout(type_src0)) {
         case MMQ_Q8_1_DS_LAYOUT_D4:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D4>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+                <<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
             break;
         case MMQ_Q8_1_DS_LAYOUT_DS4:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_DS4>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+                <<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
             break;
         case MMQ_Q8_1_DS_LAYOUT_D2S6:
             quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D2S6>
-                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+                <<<num_blocks, block_size, 0, stream>>>(x, vy, ne00, ne1, ne0);
             break;
         default:
             GGML_ABORT("fatal error");
             break;
     }
+    GGML_UNUSED(s01);
+    GGML_UNUSED(s02);
+    GGML_UNUSED(s03);
 }

ggml/src/ggml-cuda/quantize.cuh

@@ -12,13 +12,13 @@ static_assert(MATRIX_ROW_PADDING %    CUDA_QUANTIZE_BLOCK_SIZE      == 0, "Risk
 static_assert(MATRIX_ROW_PADDING % (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ) == 0, "Risk of out-of-bounds access.");
 
 typedef void (*quantize_cuda_t)(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
-    const ggml_type type_x, cudaStream_t stream);
+    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream);
 
 void quantize_row_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
-    const ggml_type type_x, cudaStream_t stream);
+    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream);
 
 void quantize_mmq_q8_1_cuda(
-    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
-    const ggml_type type_x, cudaStream_t stream);
+    const float * x, void * vy, const ggml_type type_src0, const int64_t ne00, const int64_t s01, const int64_t s02, const int64_t s03,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2, const int64_t ne3, cudaStream_t stream);

ggml/src/ggml-cuda/vecdotq.cuh

@@ -1,3 +1,5 @@
+#pragma once
+
 #include "common.cuh"
 #include <cstdint>
 

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

@@ -44,8 +44,8 @@ static struct ggml_backend_device g_ggml_backend_metal_device;
 // note: assumes single GPU device - the default one
 // TODO: support multiple GPU devices
 static struct ggml_backend_metal_device_context {
-    id<MTLDevice> mtl_device;
-    int           mtl_device_ref_count;
+    id<MTLDevice>  mtl_device;
+    int            mtl_device_ref_count;
     id<MTLLibrary> mtl_library;
 
     bool has_simdgroup_reduction;
@@ -490,7 +490,259 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_COUNT
 };
 
+//
+// ggml_metal_heap
+//
+
+struct ggml_metal_heap {
+    // number of times the heap was unused
+    int n_unused;
+
+    // total number of buffer allocations in this heap across all computes
+    int64_t n_alloc;
+
+    // current offset in the heap - we reset this after each node in order to reuse the memory
+    size_t offs;
+
+    // the currently allocated MTLBuffer objects in this heap
+    id<MTLHeap> obj;
+
+    NSMutableArray * bufs;
+};
+
+static struct ggml_metal_heap * ggml_metal_heap_init(id<MTLDevice> device, size_t size) {
+    struct ggml_metal_heap * heap = calloc(1, sizeof(struct ggml_metal_heap));
+
+    MTLHeapDescriptor * desc = [[MTLHeapDescriptor alloc] init];
+    desc.storageMode  = MTLStorageModePrivate;
+    desc.cpuCacheMode = MTLCPUCacheModeDefaultCache;
+    desc.type         = MTLHeapTypePlacement;
+    desc.size         = size;
+
+    heap->n_unused = 0;
+    heap->n_alloc = 0;
+
+    heap->obj = [device newHeapWithDescriptor:desc];
+    if (!heap->obj) {
+        GGML_LOG_ERROR("%s: error: failed to create MTLHeap with size %zu\n", __func__, size);
+
+        free(heap);
+
+        return false;
+    }
+
+    [desc release];
+
+    heap->bufs = [[NSMutableArray alloc] init];
+
+    return heap;
+}
+
+static void ggml_metal_heap_reset(struct ggml_metal_heap * heap) {
+    heap->offs = 0;
+
+    // count how many graph computes the heap ended up being unused
+    if ([heap->bufs count] > 0) {
+        heap->n_unused = 0;
+    } else {
+        heap->n_unused++;
+    }
+
+    for (id<MTLBuffer> buf in heap->bufs) {
+        [buf release];
+    }
+    [heap->bufs removeAllObjects];
+
+    // tell the OS that it can reuse this memory if needed
+    // ref: https://developer.apple.com/documentation/metal/mtlpurgeablestate?language=objc
+    [heap->obj setPurgeableState:MTLPurgeableStateVolatile];
+}
+
+static void ggml_metal_heap_free(struct ggml_metal_heap * heap) {
+    if (heap == nil) {
+        return;
+    }
+
+    ggml_metal_heap_reset(heap);
+
+    [heap->obj  release];
+    [heap->bufs release];
+
+    free(heap);
+}
+
+@interface ggml_metal_heap_ptr : NSObject
+
+@property (nonatomic, assign) struct ggml_metal_heap * data;
+
+@end
+
+@implementation ggml_metal_heap_ptr
+@end
+
+//
+// ggml_metal_mem_pool
+//
+
+struct ggml_metal_mem_pool {
+    id<MTLDevice> device;
+
+    int n_heaps; // total number of heaps ever created (including those that were removed)
+
+    NSMutableArray * heaps;
+    NSMutableArray * heaps_to_remove;
+};
+
+static struct ggml_metal_mem_pool * ggml_metal_mem_pool_init(void) {
+    struct ggml_metal_mem_pool * mem_pool = calloc(1, sizeof(struct ggml_metal_mem_pool));
+
+    mem_pool->n_heaps = 0;
+
+    mem_pool->heaps           = [[NSMutableArray alloc] init];
+    mem_pool->heaps_to_remove = [[NSMutableArray alloc] init];
+
+    return mem_pool;
+}
+
+static void ggml_metal_mem_pool_free(struct ggml_metal_mem_pool * mem_pool) {
+    GGML_LOG_DEBUG("%s: freeing memory pool, num heaps = %zu (total = %d)\n", __func__, [mem_pool->heaps count], mem_pool->n_heaps);
+
+    size_t size_all = 0;
+    size_t size_cur = 0;
+
+    for (ggml_metal_heap_ptr * ptr in mem_pool->heaps) {
+        GGML_LOG_DEBUG("%s:   heap: %p\n",                __func__, (void *) ptr.data);
+        GGML_LOG_DEBUG("%s:     n_alloc:  %" PRId64 "\n", __func__, ptr.data->n_alloc);
+        GGML_LOG_DEBUG("%s:     n_unused: %d\n",          __func__, ptr.data->n_unused);
+        GGML_LOG_DEBUG("%s:     size:     %.2f MiB\n",    __func__, [ptr.data->obj size] / 1024.0 / 1024.0);
+        GGML_LOG_DEBUG("%s:     bufs:     %zu\n",         __func__, [ptr.data->bufs count]);
+
+        if ([ptr.data->bufs count] > 0) {
+            size_cur += [ptr.data->obj size];
+        }
+        size_all += [ptr.data->obj size];
+
+        ggml_metal_heap_free(ptr.data);
+        [ptr release];
+    }
+    [mem_pool->heaps           release];
+    [mem_pool->heaps_to_remove release];
+
+    if (size_all > 0) {
+        GGML_LOG_DEBUG("%s:   size_all: %.2f MiB\n", __func__, size_all / 1024.0 / 1024.0);
+        GGML_LOG_DEBUG("%s:   size_cur: %.2f MiB\n", __func__, size_cur / 1024.0 / 1024.0);
+    }
+
+    free(mem_pool);
+}
+
+static void ggml_metal_mem_pool_reset(struct ggml_metal_mem_pool * mem_pool) {
+    for (NSUInteger i = 0; i < [mem_pool->heaps count]; i++) {
+        ggml_metal_heap_ptr * ptr = [mem_pool->heaps objectAtIndex:i];
+
+        struct ggml_metal_heap * heap = ptr.data;
+        ggml_metal_heap_reset(heap);
+
+        // if the heap hasn't been used for a while, remove it
+        if (heap->n_unused >= 128) {
+            [mem_pool->heaps_to_remove addObject:@(i)];
+        }
+    }
+
+    if (mem_pool->heaps_to_remove.count > 0) {
+        for (NSUInteger i = 0; i < [mem_pool->heaps_to_remove count]; i++) {
+            NSUInteger index = [[mem_pool->heaps_to_remove objectAtIndex:i] intValue];
+            ggml_metal_heap_ptr * ptr = [mem_pool->heaps objectAtIndex:index];
+
+            struct ggml_metal_heap * heap = ptr.data;
+            ggml_metal_heap_free(heap);
+
+            [mem_pool->heaps removeObjectAtIndex:index];
+            [ptr release];
+        }
+
+        [mem_pool->heaps_to_remove removeAllObjects];
+    }
+}
+
+static void ggml_metal_mem_pool_clear(struct ggml_metal_mem_pool * mem_pool) {
+    for (ggml_metal_heap_ptr * ptr in mem_pool->heaps) {
+        ptr.data->offs = 0;
+    }
+}
+
+static id<MTLBuffer> ggml_metal_mem_pool_alloc(struct ggml_metal_mem_pool * mem_pool, size_t size) {
+    const size_t alignment = 32;
+
+    const size_t size_aligned = GGML_PAD(size, alignment);
+
+    // try one of the existing heaps
+    for (ggml_metal_heap_ptr * ptr in mem_pool->heaps) {
+        struct ggml_metal_heap * heap = ptr.data;
+        if (heap->offs + size_aligned <= [heap->obj size]) {
+            // if this is the first buffer in the heap for the current command buffer, tell the OS that
+            //   it cannot free the memory used by the heap
+            // ref: https://developer.apple.com/documentation/metal/mtlpurgeablestate?language=objc
+            if ([heap->bufs count] == 0) {
+                [heap->obj setPurgeableState:MTLPurgeableStateNonVolatile];
+            }
+
+            id<MTLBuffer> buf = [heap->obj newBufferWithLength:size_aligned options:MTLResourceStorageModePrivate offset:heap->offs];
+            if (buf == nil) {
+                GGML_LOG_ERROR("%s: error: failed to create MTLBuffer with size %zu\n", __func__, size_aligned);
+                return nil;
+            }
+
+            heap->n_alloc++;
+            heap->offs += size_aligned;
+
+            [heap->bufs addObject:buf];
+
+            return buf;
+        }
+    }
+
+    // create a new heap that can fit this buffer
+    ggml_metal_heap_ptr * heap_ptr = [ggml_metal_heap_ptr new];
+
+    struct ggml_metal_heap * heap = ggml_metal_heap_init(mem_pool->device, size_aligned);
+    if (heap == NULL) {
+        GGML_LOG_ERROR("%s: error: failed to create heap of size %zu\n", __func__, size_aligned);
+        return NULL;
+    }
+
+    //GGML_LOG_DEBUG("%s: creating new heap of size %zu, got %zu\n", __func__, size_aligned, [heap->obj size]);
+
+    heap_ptr.data = heap;
+    ggml_metal_heap_reset(heap);
+
+    [heap->obj setPurgeableState:MTLPurgeableStateNonVolatile];
+    id<MTLBuffer> buf = [heap->obj newBufferWithLength:size_aligned options:MTLResourceStorageModePrivate offset:heap->offs];
+    if (buf == nil) {
+        GGML_LOG_ERROR("%s: error: failed to create MTLBuffer with size %zu\n", __func__, size_aligned);
+        return NULL;
+    }
+
+    heap->n_alloc++;
+    heap->offs += size_aligned;
+
+    [heap->bufs addObject:buf];
+
+    [mem_pool->heaps addObject:heap_ptr];
+    mem_pool->n_heaps++;
+
+    return buf;
+}
+
+struct ggml_metal_command_buffer {
+    id<MTLCommandBuffer> obj;
+
+    // each command buffer has a memory pool from which it can allocate temporary buffers during the compute
+    struct ggml_metal_mem_pool * mem_pool;
+};
+
 struct ggml_backend_metal_context {
+    id<MTLDevice>       device;
     id<MTLCommandQueue> queue;
 
     dispatch_queue_t d_queue;
@@ -515,7 +767,7 @@ struct ggml_backend_metal_context {
     void (^encode_async)(size_t ith);
 
     // n_cb command buffers + 1 used by the main thread
-    id<MTLCommandBuffer> command_buffers[GGML_METAL_MAX_COMMAND_BUFFERS + 1];
+    struct ggml_metal_command_buffer cmd_bufs[GGML_METAL_MAX_COMMAND_BUFFERS + 1];
 
     // abort ggml_metal_graph_compute if callback returns true
     ggml_abort_callback abort_callback;
@@ -705,9 +957,11 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
     struct ggml_backend_metal_device_context * ctx_dev = dev->context;
 
     id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+
     GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
-    ctx->queue  = [device newCommandQueue];
+    ctx->device = device;
+    ctx->queue = [device newCommandQueue];
     if (ctx->queue == nil) {
         GGML_LOG_ERROR("%s: error: failed to create command queue\n", __func__);
         return NULL;
@@ -768,7 +1022,10 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
     ctx->gf = nil;
     ctx->encode_async = nil;
     for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
-        ctx->command_buffers[i] = nil;
+        ctx->cmd_bufs[i].obj = nil;
+
+        ctx->cmd_bufs[i].mem_pool = ggml_metal_mem_pool_init();
+        ctx->cmd_bufs[i].mem_pool->device = device;
     }
 
 #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
@@ -1181,6 +1438,12 @@ static void ggml_metal_free(struct ggml_backend_metal_context * ctx) {
 
     [ctx->queue release];
 
+    for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
+        // ctx->cmd_bufs[i].obj is auto released
+
+        ggml_metal_mem_pool_free(ctx->cmd_bufs[i].mem_pool);
+    }
+
     dispatch_release(ctx->d_queue);
 
     free(ctx);
@@ -1486,10 +1749,11 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
     }
 }
 
-static void ggml_metal_encode_node(
+static bool ggml_metal_encode_node(
                         ggml_backend_t   backend,
                                    int   idx,
-          id<MTLComputeCommandEncoder>   encoder) {
+          id<MTLComputeCommandEncoder>   encoder,
+            struct ggml_metal_mem_pool * mem_pool) {
     struct ggml_backend_metal_context        * ctx     = backend->context;
     struct ggml_backend_metal_device_context * ctx_dev = backend->device->context;
 
@@ -1505,7 +1769,7 @@ static void ggml_metal_encode_node(
     struct ggml_tensor * dst  = node;
 
     if (ggml_is_empty(dst)) {
-        return;
+        return true;
     }
 
     switch (dst->op) {
@@ -1516,7 +1780,7 @@ static void ggml_metal_encode_node(
         case GGML_OP_PERMUTE:
             {
                 // noop -> next node
-            } return;
+            } return true;
         default:
             {
             } break;
@@ -1527,6 +1791,8 @@ static void ggml_metal_encode_node(
         GGML_ABORT("unsupported op");
     }
 
+    ggml_metal_mem_pool_clear(mem_pool);
+
     const int64_t  ne00 = src0 ? src0->ne[0] : 0;
     const int64_t  ne01 = src0 ? src0->ne[1] : 0;
     const int64_t  ne02 = src0 ? src0->ne[2] : 0;
@@ -2173,26 +2439,76 @@ static void ggml_metal_encode_node(
                 const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
                 const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-                ggml_metal_kargs_soft_max args = {
+// use this branch to test the ggml_metal_mem_pool functionality
+#if 0
+                // cpy to tmp buffer in MTLHeap
+
+                id<MTLBuffer> h_src0 = h_src0 = ggml_metal_mem_pool_alloc(mem_pool, ggml_nbytes(src0));
+                if (!h_src0) {
+                    GGML_LOG_ERROR("%s: failed to allocate buffer from memory pool, size = %zu\n", __func__, ggml_nbytes(src0));
+                    return false;
+                }
+
+                offs_src0 = 0;
+
+                ggml_metal_kargs_cpy args_cpy = {
                     /*.ne00 =*/ ne00,
                     /*.ne01 =*/ ne01,
                     /*.ne02 =*/ ne02,
-                    /*.scale =*/ scale,
-                    /*.max_bias =*/ max_bias,
-                    /*.m0 =*/ m0,
-                    /*.m1 =*/ m1,
+                    /*.ne03 =*/ ne03,
+                    /*.nb00 =*/ nb00,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne0  =*/ ne00,
+                    /*.ne1  =*/ ne01,
+                    /*.ne2  =*/ ne02,
+                    /*.ne3  =*/ ne03,
+                    /*.nb0  =*/ nb00,
+                    /*.nb1  =*/ nb01,
+                    /*.nb2  =*/ nb02,
+                    /*.nb3  =*/ nb03,
+                };
+
+                if (src0->type == GGML_TYPE_F16) {
+                    [encoder setComputePipelineState:ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F16].pipeline];
+                } else {
+                    [encoder setComputePipelineState:ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline];
+                }
+                [encoder setBytes:&args_cpy length:sizeof(args_cpy) atIndex:0];
+                [encoder setBuffer:id_src0  offset:offs_src0        atIndex:1];
+                [encoder setBuffer:h_src0   offset:0                atIndex:2];
+
+                GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+                int nth_cpy = MIN(1024, ne00 / ggml_blck_size(src0->type));
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth_cpy, 1, 1)];
+
+#else
+                id<MTLBuffer> h_src0 = id_src0;
+#endif
+                // softmax
+
+                ggml_metal_kargs_soft_max args = {
+                    /*.ne00        =*/ ne00,
+                    /*.ne01        =*/ ne01,
+                    /*.ne02        =*/ ne02,
+                    /*.scale       =*/ scale,
+                    /*.max_bias    =*/ max_bias,
+                    /*.m0          =*/ m0,
+                    /*.m1          =*/ m1,
                     /*.n_head_log2 =*/ n_head_log2,
                 };
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
+                [encoder setBuffer:h_src0 offset:offs_src0      atIndex:0];
                 if (id_src1) {
-                    [encoder setBuffer:id_src1 offset:offs_src1   atIndex:1];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
                 } else {
-                    [encoder setBuffer:id_src0 offset:offs_src0   atIndex:1];
+                    [encoder setBuffer:h_src0 offset:offs_src0  atIndex:1];
                 }
-                [encoder setBuffer:id_dst      offset:offs_dst            atIndex:2];
-                [encoder setBytes:&args        length:sizeof(args)        atIndex:3];
+                [encoder setBuffer:id_dst offset:offs_dst       atIndex:2];
+                [encoder setBytes:&args   length:sizeof(args)   atIndex:3];
 
                 [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
@@ -4601,6 +4917,8 @@ static void ggml_metal_encode_node(
                 GGML_ABORT("fatal error");
             }
     }
+
+    return true;
 }
 
 static enum ggml_status ggml_metal_graph_compute(
@@ -4654,25 +4972,25 @@ static enum ggml_status ggml_metal_graph_compute(
         }
 
         // the main thread commits the first few commands immediately
-        // command_buffer[n_cb]
+        // cmd_buf[n_cb]
         {
-            id<MTLCommandBuffer> command_buffer = [ctx->queue commandBufferWithUnretainedReferences];
-            ctx->command_buffers[n_cb] = command_buffer;
+            id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
+            ctx->cmd_bufs[n_cb].obj = cmd_buf;
 
-            [command_buffer enqueue];
+            [cmd_buf enqueue];
             ctx->encode_async(n_cb);
         }
 
         // prepare the rest of the command buffers asynchronously
-        // command_buffer[0.. n_cb)
+        // cmd_buf[0.. n_cb)
         for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
-            id<MTLCommandBuffer> command_buffer = [ctx->queue commandBufferWithUnretainedReferences];
-            ctx->command_buffers[cb_idx] = command_buffer;
+            id<MTLCommandBuffer> cmd_buf = [ctx->queue commandBufferWithUnretainedReferences];
+            ctx->cmd_bufs[cb_idx].obj = cmd_buf;
 
             // always enqueue the first two command buffers
             // enqueue all of the command buffers if we don't need to abort
             if (cb_idx < 2 || ctx->abort_callback == NULL) {
-                [command_buffer enqueue];
+                [cmd_buf enqueue];
             }
         }
 
@@ -4681,14 +4999,14 @@ static enum ggml_status ggml_metal_graph_compute(
         // wait for completion and check status of each command buffer
         // needed to detect if the device ran out-of-memory for example (#1881)
         {
-            id<MTLCommandBuffer> command_buffer = ctx->command_buffers[n_cb];
-            [command_buffer waitUntilCompleted];
+            id<MTLCommandBuffer> cmd_buf = ctx->cmd_bufs[n_cb].obj;
+            [cmd_buf waitUntilCompleted];
 
-            MTLCommandBufferStatus status = [command_buffer status];
+            MTLCommandBufferStatus status = [cmd_buf status];
             if (status != MTLCommandBufferStatusCompleted) {
                 GGML_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, n_cb, status);
                 if (status == MTLCommandBufferStatusError) {
-                    GGML_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]);
+                    GGML_LOG_INFO("error: %s\n", [[cmd_buf error].localizedDescription UTF8String]);
                 }
 
                 return GGML_STATUS_FAILED;
@@ -4696,20 +5014,20 @@ static enum ggml_status ggml_metal_graph_compute(
         }
 
         for (int i = 0; i < n_cb; ++i) {
-            id<MTLCommandBuffer> command_buffer = ctx->command_buffers[i];
-            [command_buffer waitUntilCompleted];
+            id<MTLCommandBuffer> cmd_buf = ctx->cmd_bufs[i].obj;
+            [cmd_buf waitUntilCompleted];
 
-            MTLCommandBufferStatus status = [command_buffer status];
+            MTLCommandBufferStatus status = [cmd_buf status];
             if (status != MTLCommandBufferStatusCompleted) {
                 GGML_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
                 if (status == MTLCommandBufferStatusError) {
-                    GGML_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]);
+                    GGML_LOG_INFO("error: %s\n", [[cmd_buf error].localizedDescription UTF8String]);
                 }
 
                 return GGML_STATUS_FAILED;
             }
 
-            id<MTLCommandBuffer> next_buffer = (i + 1 < n_cb ? ctx->command_buffers[i + 1] : nil);
+            id<MTLCommandBuffer> next_buffer = (i + 1 < n_cb ? ctx->cmd_bufs[i + 1].obj : nil);
             if (!next_buffer) {
                 continue;
             }
@@ -5092,8 +5410,9 @@ static void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
 
         const int n_nodes_per_cb = ctx->n_nodes_per_cb;
 
-        id<MTLCommandBuffer> command_buffer  = ctx->command_buffers[cb_idx];
-        id<MTLComputeCommandEncoder> encoder = [command_buffer computeCommandEncoder];
+        id<MTLCommandBuffer> cmd_buf = ctx->cmd_bufs[cb_idx].obj;
+
+        id<MTLComputeCommandEncoder> encoder = [cmd_buf computeCommandEncoder];
 
         int node_start = 0;
         int node_end   = n_nodes_0;
@@ -5105,22 +5424,29 @@ static void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
 
         const bool should_capture = ctx->capture_next_compute;
 
+        struct ggml_metal_mem_pool * mem_pool = ctx->cmd_bufs[cb_idx].mem_pool;
+        ggml_metal_mem_pool_reset(mem_pool);
+
         for (int idx = node_start; idx < node_end; ++idx) {
             if (should_capture) {
                 [encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(ggml_graph_node(ctx->gf, idx)) encoding:NSUTF8StringEncoding]];
             }
 
-            ggml_metal_encode_node(backend, idx, encoder);
+            const bool res = ggml_metal_encode_node(backend, idx, encoder, mem_pool);
 
             if (should_capture) {
                 [encoder popDebugGroup];
             }
+
+            if (!res) {
+                break;
+            }
         }
 
         [encoder endEncoding];
 
         if (cb_idx < 2 || ctx->abort_callback == NULL) {
-            [command_buffer commit];
+            [cmd_buf commit];
         }
     });
 }

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

@@ -3192,7 +3192,7 @@ kernel void kernel_flash_attn_ext(
 
     {
         float S[Q] = { [0 ... Q-1] = 0.0f };
-        float M[Q] = { [0 ... Q-1] = -__FLT16_MAX__/2 };
+        float M[Q] = { [0 ... Q-1] = -__FLT_MAX__/2 };
 
         // thread indices inside the simdgroup
         // TODO: see if we can utilize quad-group functions for better performance
@@ -3452,7 +3452,7 @@ kernel void kernel_flash_attn_ext(
     // reduce the warps sequentially
     for (ushort sg = 1; sg < nsg; ++sg) {
         float S = { 0.0f };
-        float M = { -__FLT16_MAX__/2 };
+        float M = { -__FLT_MAX__/2 };
 
         threadgroup_barrier(mem_flags::mem_threadgroup);
 
@@ -3699,7 +3699,7 @@ kernel void kernel_flash_attn_ext_vec(
 
     {
         float S = 0.0f;
-        float M = -__FLT16_MAX__/2;
+        float M = -__FLT_MAX__/2;
 
         // thread indices inside the simdgroup
         const short tx = tiisg%NL;

ggml/src/ggml-rpc/ggml-rpc.cpp

@@ -378,8 +378,8 @@ static bool parse_endpoint(const std::string & endpoint, std::string & host, int
 }
 
 // RPC request : | rpc_cmd (1 byte) | request_size (8 bytes) | request_data (request_size bytes) |
-// RPC response: | response_size (8 bytes) | response_data (response_size bytes) |
-static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const void * input, size_t input_size, void * output, size_t output_size) {
+// No response
+static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const void * input, size_t input_size) {
     uint8_t cmd_byte = cmd;
     if (!send_data(sock->fd, &cmd_byte, sizeof(cmd_byte))) {
         return false;
@@ -390,6 +390,15 @@ static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cm
     if (!send_data(sock->fd, input, input_size)) {
         return false;
     }
+    return true;
+}
+
+// RPC request : | rpc_cmd (1 byte) | request_size (8 bytes) | request_data (request_size bytes) |
+// RPC response: | response_size (8 bytes) | response_data (response_size bytes) |
+static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const void * input, size_t input_size, void * output, size_t output_size) {
+    if (!send_rpc_cmd(sock, cmd, input, input_size)) {
+        return false;
+    }
     // TODO: currently the output_size is always known, do we need support for commands with variable output size?
     // even if we do, we can skip sending output_size from the server for commands with known output size
     uint64_t out_size;
@@ -555,7 +564,7 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggm
     memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
     memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
     memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
-    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input.data(), input.size(), nullptr, 0);
+    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input.data(), input.size());
     GGML_ASSERT(status);
 }
 
@@ -1428,9 +1437,6 @@ static void rpc_serve_client(ggml_backend_t backend, const char * cache_dir,
                 if (!server.set_tensor(input)) {
                     return;
                 }
-                if (!send_msg(sockfd, nullptr, 0)) {
-                    return;
-                }
                 break;
             }
             case RPC_CMD_SET_TENSOR_HASH: {

ggml/src/ggml-sycl/common.hpp

@@ -313,7 +313,6 @@ struct ggml_backend_sycl_context {
     int device;
     std::string name;
     optimize_feature opt_feature;
-    bool optimized_graph=false;
 
     queue_ptr qptrs[GGML_SYCL_MAX_DEVICES][GGML_SYCL_MAX_STREAMS] = { { nullptr } };
 

ggml/src/ggml-sycl/ggml-sycl.cpp

@@ -192,7 +192,7 @@ static void ggml_check_sycl() try {
 
     if (!initialized) {
         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
-        g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1);
+        g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 0);
         g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
         GGML_SYCL_DEBUG("[SYCL] call ggml_check_sycl\n");
         GGML_LOG_INFO("Running with Environment Variables:\n");
@@ -2852,6 +2852,64 @@ static bool ggml_sycl_supports_dmmv(enum ggml_type type) {
     }
 }
 
+static void reorder_qw(char *data_device, const int ncols, const int nrows,
+                size_t size, size_t offset, dpct::queue_ptr stream) {
+    auto tmp_buf = sycl::malloc_shared<char>(size, *stream);
+    SYCL_CHECK(
+        CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size)
+            .wait()));
+    GGML_ASSERT((size % sizeof(block_q4_0) == 0));
+    GGML_ASSERT((offset % sizeof(block_q4_0) == 0));
+    int offset_blks = offset / sizeof(block_q4_0);
+    auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2;;
+    auto d_ptr = (sycl::half*)(qs_ptr + ncols * nrows / 2) + offset_blks;
+
+    stream->parallel_for(
+        size / sizeof(block_q4_0),
+            [=](auto i) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+            const block_q4_0* x = (const block_q4_0*)tmp_buf;
+            const int ib = i;
+
+            for (int j = 0; j < QK4_0/2; j ++)
+            {
+                *(qs_ptr + ib * QK4_0 / 2 + j) = x[ib].qs[j];
+            }
+            *(d_ptr + ib) = x[ib].d;
+        });
+
+    sycl::free(tmp_buf, *stream);
+}
+
+static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
+    char*data_device = (char*)src0->data;
+    size_t ncols = src0->ne[0];
+    size_t nrows = src0->ne[1];
+    size_t size = ggml_nbytes(src0);
+
+    reorder_qw(data_device, ncols, nrows, size, 0, stream);
+}
+
+/*
+* This function could be called when the OP (mul_mat) function support reorder optimizition.
+*/
+static void opt_for_reorder(ggml_backend_sycl_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1,
+    ggml_tensor * dst) {
+    if (!g_ggml_sycl_disable_optimize && //allow optimize, controlled by $GGML_SYCL_DISABLE_OPT
+        ctx->opt_feature.reorder &&      //allow this device due to good perf, skip the devices with bad perf.
+        dst->op == GGML_OP_MUL_MAT &&    //limit to some supported cases of Q4_0, to do for more cases.
+        src0->type == GGML_TYPE_Q4_0 &&
+        src1->ne[2]==1 && src1->ne[3]==1) {
+
+        ggml_tensor_extra_gpu* extra = (ggml_tensor_extra_gpu*)src0->extra;
+        if (!extra) return; //only happen in CI/UT permute case.
+
+        if (extra->optimized_feature.reorder) return; //skip the tensor which is handled for reorder.
+
+        reorder_qw(src0, ctx->stream());
+        extra->optimized_feature.reorder = true; //used to decode/dequan in next steps.
+    }
+}
+
 static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
 
     const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
@@ -2914,6 +2972,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
         // KQ + KQV multi-batch
         ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
     } else if (use_dequantize_mul_mat_vec) {
+        opt_for_reorder(&ctx, src0, src1, dst); //the OP function in this branch support reorder.
         ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false);
         // save_tensor_txt("1/dst_1.txt", (float*) dst->data, src0->ne[1], sizeof(float), ctx.stream());
     } else if (use_mul_mat_vec_q) {
@@ -2921,6 +2980,7 @@ static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor
     } else if (use_mul_mat_q) {
         ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q, true);
     } else {
+        opt_for_reorder(&ctx, src0, src1, dst); //the OP function in this branch support reorder.
         ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false);
     }
 }
@@ -3545,71 +3605,8 @@ catch (sycl::exception const &exc) {
   std::exit(1);
 }
 
-static void reorder_qw(char *data_device, const int ncols, const int nrows,
-                size_t size, size_t offset, dpct::queue_ptr stream) {
-    auto tmp_buf = sycl::malloc_shared<char>(size, *stream);
-    SYCL_CHECK(
-        CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size)
-            .wait()));
-    GGML_ASSERT((size % sizeof(block_q4_0) == 0));
-    GGML_ASSERT((offset % sizeof(block_q4_0) == 0));
-    int offset_blks = offset / sizeof(block_q4_0);
-    auto qs_ptr = (uint8_t*)data_device + offset_blks * QK4_0 / 2;;
-    auto d_ptr = (sycl::half*)(qs_ptr + ncols * nrows / 2) + offset_blks;
-
-    stream->parallel_for(
-        size / sizeof(block_q4_0),
-            [=](auto i) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-            const block_q4_0* x = (const block_q4_0*)tmp_buf;
-            const int ib = i;
-
-            for (int j = 0; j < QK4_0/2; j ++)
-            {
-                *(qs_ptr + ib * QK4_0 / 2 + j) = x[ib].qs[j];
-            }
-            *(d_ptr + ib) = x[ib].d;
-        });
-
-    sycl::free(tmp_buf, *stream);
-}
-
-static void reorder_qw(ggml_tensor * src0, dpct::queue_ptr stream) {
-    char*data_device = (char*)src0->data;
-    size_t ncols = src0->ne[0];
-    size_t nrows = src0->ne[1];
-    size_t size = ggml_nbytes(src0);
-
-    reorder_qw(data_device, ncols, nrows, size, 0, stream);
-}
-
-static void opt_for_reorder(ggml_tensor * dst, dpct::queue_ptr stream) {
-    ggml_tensor *src0 = dst->src[0];
-    ggml_tensor *src1 = dst->src[1];
-
-    if (dst->op == GGML_OP_MUL_MAT && src0->type == GGML_TYPE_Q4_0 &&
-        src1->ne[2]==1 && src1->ne[3]==1) {
-        reorder_qw(src0, stream);
-        ggml_tensor_extra_gpu* extra = (ggml_tensor_extra_gpu*)src0->extra;
-        GGML_ASSERT(extra);
-        extra->optimized_feature.reorder = true; //used to decode/dequan in next steps.
-    }
-}
-
-static void optimize_graph_once(ggml_cgraph * cgraph, ggml_backend_sycl_context * ctx) {
-    dpct::queue_ptr stream = ctx->stream();
-    if (ctx->optimized_graph) {
-       return;
-    }
-    ctx->optimized_graph = true;
-
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        if (ctx->opt_feature.reorder) opt_for_reorder(cgraph->nodes[i], stream);
-    }
-}
-
 static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * sycl_ctx, ggml_cgraph * cgraph) {
     ggml_sycl_set_main_device(sycl_ctx->device);
-    if (!g_ggml_sycl_disable_optimize) optimize_graph_once(cgraph, sycl_ctx);
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_tensor * node = cgraph->nodes[i];

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

@@ -246,6 +246,7 @@ struct vk_device_struct {
     bool pipeline_robustness;
     vk::Device device;
     uint32_t vendor_id;
+    vk::DriverId driver_id;
     vk_device_architecture architecture;
     vk_queue compute_queue;
     vk_queue transfer_queue;
@@ -1740,6 +1741,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         m_warptile_mmq_int = { 128,  64,  64, 32, subgroup_size_8,     32, 2, 2, 2, 1, subgroup_size_8 };
         s_warptile_mmq_int = { subgroup_size_32, 32, 32, 32, 32,       32, 2, 2, 1, 1, subgroup_size_8 };
 
+        // chip specific tuning
+        if ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) {
+            m_warptile_mmq = m_warptile_mmq_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
+        }
+
         l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
         m_mmq_wg_denoms = m_wg_denoms = { 64,  64, 1 };
         s_mmq_wg_denoms = s_wg_denoms = { 32,  32, 1 };
@@ -2658,6 +2664,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->physical_device.getProperties2(&props2);
         device->properties = props2.properties;
         device->vendor_id = device->properties.vendorID;
+        device->driver_id = driver_props.driverID;
 
         const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE");
 

ggml/src/ggml.c

@@ -4,6 +4,7 @@
 #include "ggml-backend.h"
 #include "ggml-impl.h"
 #include "ggml-threading.h"
+#include "ggml-cpu.h"
 #include "ggml.h"
 
 // FIXME: required here for quantization functions
@@ -382,58 +383,16 @@ void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n) {
     }
 }
 
-// FIXME: these functions must detect the instruction set at runtime, since they are part of the core ggml library
-//        currently, the ggml_cpu_has_* functions are entirely compile-time
 void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n) {
-    int64_t i = 0;
-#if defined(__F16C__)
-    //if (ggml_cpu_has_f16c()) {
-        for (; i + 7 < n; i += 8) {
-            __m256 x_vec = _mm256_loadu_ps(x + i);
-            __m128i y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
-            _mm_storeu_si128((__m128i *)(y + i), y_vec);
-        }
-        for(; i + 3 < n; i += 4) {
-            __m128 x_vec = _mm_loadu_ps(x + i);
-            __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
-            _mm_storel_epi64((__m128i *)(y + i), y_vec);
-        }
-    //}
-#endif
-    for (; i < n; i++) {
+    int i = 0;
+    for (; i < n; ++i) {
         y[i] = GGML_FP32_TO_FP16(x[i]);
     }
 }
 
 void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
-    int64_t i = 0;
-#if defined(__AVX512F__)
-    //if (ggml_cpu_has_avx512()) {
-        for (; i + 16 <= n; i += 16) {
-            _mm512_storeu_ps(y + i,
-                            _mm512_castsi512_ps(
-                                _mm512_slli_epi32(
-                                    _mm512_cvtepu16_epi32(
-                                        _mm256_loadu_si256(
-                                            (const __m256i *)(x + i))),
-                                    16)));
-        }
-    //}
-#endif
-#if defined(__AVX2__)
-    //if (ggml_cpu_has_avx2()) {
-        for (; i + 8 <= n; i += 8) {
-            _mm256_storeu_ps(y + i,
-                            _mm256_castsi256_ps(
-                                _mm256_slli_epi32(
-                                    _mm256_cvtepu16_epi32(
-                                        _mm_loadu_si128(
-                                            (const __m128i *)(x + i))),
-                                    16)));
-        }
-    //}
-#endif
-    for (; i < n; i++) {
+    int i = 0;
+    for (; i < n; ++i) {
         y[i] = GGML_BF16_TO_FP32(x[i]);
     }
 }
@@ -956,6 +915,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CONV_TRANSPOSE_1D",
     "IM2COL",
     "IM2COL_BACK",
+    "CONV_2D_DW",
     "CONV_TRANSPOSE_2D",
     "POOL_1D",
     "POOL_2D",
@@ -993,7 +953,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "OPT_STEP_ADAMW",
 };
 
-static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1050,6 +1010,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "conv_transpose_1d(x)",
     "im2col(x)",
     "im2col_back(x)",
+    "conv_2d_dw(x)",
     "conv_transpose_2d(x)",
     "pool_1d(x)",
     "pool_2d(x)",
@@ -1087,7 +1048,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "adamw(x)",
 };
 
-static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
+static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -1344,6 +1305,13 @@ bool ggml_is_permuted(const struct ggml_tensor * tensor) {
     return tensor->nb[0] > tensor->nb[1] || tensor->nb[1] > tensor->nb[2] || tensor->nb[2] > tensor->nb[3];
 }
 
+bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor) {
+    return
+        tensor->nb[0] > tensor->nb[2] &&
+        tensor->nb[1] > tensor->nb[0] &&
+        tensor->nb[2] == ggml_type_size(tensor->type);
+}
+
 static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
@@ -4050,6 +4018,46 @@ struct ggml_tensor * ggml_conv_2d_dw(
     return result;
 }
 
+// ggml_conv_2d_dw_direct
+
+struct ggml_tensor * ggml_conv_2d_dw_direct(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int                   stride0,
+        int                   stride1,
+        int                   pad0,
+        int                   pad1,
+        int                   dilation0,
+        int                   dilation1) {
+    GGML_ASSERT(a->ne[2] == 1);
+    GGML_ASSERT(a->ne[3] == b->ne[2]);
+    int64_t ne[4];
+    ne[0] = ggml_calc_conv_output_size(b->ne[0], a->ne[0], stride0, pad0, dilation0);
+    ne[1] = ggml_calc_conv_output_size(b->ne[1], a->ne[1], stride1, pad1, dilation1);
+    ne[2] = b->ne[2];
+    ne[3] = b->ne[3];
+
+    struct ggml_tensor * result = ggml_new_tensor(ctx, b->type, 4, ne);
+
+    if (ggml_is_contiguous_channels(b)) {
+        // Result will be permuted the same way as input (CWHN order)
+        const int64_t type_size = ggml_type_size(result->type);
+        GGML_ASSERT(ggml_blck_size(result->type) == 1);
+        result->nb[0] = result->ne[2] * type_size;
+        result->nb[1] = result->ne[0] * result->nb[0];
+        result->nb[2] = type_size;
+    }
+
+    int32_t params[] = { stride0, stride1, pad0, pad1, dilation0, dilation1 };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = GGML_OP_CONV_2D_DW;
+    result->src[0] = a;
+    result->src[1] = b;
+    return result;
+}
+
 // ggml_conv_transpose_2d_p0
 
 static int64_t ggml_calc_conv_transpose_output_size(int64_t ins, int64_t ks, int s, int p) {

gguf-py/gguf/constants.py

@@ -231,11 +231,15 @@ class Keys:
         IMAGE_MEAN          = "clip.vision.image_mean"
         IMAGE_STD           = "clip.vision.image_std"
         USE_GELU            = "clip.use_gelu"
+        USE_SILU            = "clip.use_silu"
 
         class Attention:
             HEAD_COUNT      = "clip.vision.attention.head_count"
             LAYERNORM_EPS   = "clip.vision.attention.layer_norm_epsilon"
 
+        class Projector:
+            SCALE_FACTOR    = "clip.vision.projector.scale_factor"
+
 #
 # recommended mapping of model tensor names for storage in gguf
 #
@@ -481,6 +485,7 @@ class MODEL_TENSOR(IntEnum):
     V_ENC_OUTPUT         = auto()
     V_ENC_OUTPUT_NORM    = auto()
     V_ENC_FFN_UP         = auto()
+    V_ENC_FFN_GATE       = auto()
     V_ENC_FFN_DOWN       = auto()
     V_PRE_NORM           = auto()
     V_POST_NORM          = auto()
@@ -497,6 +502,7 @@ class MODEL_TENSOR(IntEnum):
     V_RESMPL_Q_NORM      = auto() # minicpmv
     V_RESMPL_PROJ        = auto() # minicpmv
     V_RESMPL_QUERY       = auto() # minicpmv
+    V_TOK_EMBD_IMG_BREAK = auto() # pixtral
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -733,6 +739,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_ENC_OUTPUT:              "v.blk.{bid}.attn_out",
     MODEL_TENSOR.V_ENC_OUTPUT_NORM:         "v.blk.{bid}.ln2",
     MODEL_TENSOR.V_ENC_FFN_UP:              "v.blk.{bid}.ffn_up",
+    MODEL_TENSOR.V_ENC_FFN_GATE:            "v.blk.{bid}.ffn_gate",
     MODEL_TENSOR.V_ENC_FFN_DOWN:            "v.blk.{bid}.ffn_down",
     MODEL_TENSOR.V_PRE_NORM:                "v.pre_ln",
     MODEL_TENSOR.V_POST_NORM:               "v.post_ln",
@@ -749,6 +756,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_RESMPL_Q_NORM:           "resampler.ln_q",
     MODEL_TENSOR.V_RESMPL_PROJ:             "resampler.proj",
     MODEL_TENSOR.V_RESMPL_QUERY:            "resampler.query",
+    MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK:      "v.token_embd.img_break", # pixtral
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -767,6 +775,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_ENC_OUTPUT,
         MODEL_TENSOR.V_ENC_OUTPUT_NORM,
         MODEL_TENSOR.V_ENC_FFN_UP,
+        MODEL_TENSOR.V_ENC_FFN_GATE,
         MODEL_TENSOR.V_ENC_FFN_DOWN,
         MODEL_TENSOR.V_PRE_NORM,
         MODEL_TENSOR.V_POST_NORM,
@@ -783,6 +792,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_RESMPL_Q_NORM,
         MODEL_TENSOR.V_RESMPL_PROJ,
         MODEL_TENSOR.V_RESMPL_QUERY,
+        MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK,
     ],
     MODEL_ARCH.LLAMA: [
         MODEL_TENSOR.TOKEN_EMBD,
@@ -2122,6 +2132,12 @@ class GGUFValueType(IntEnum):
             raise ValueError(f"Unknown type: {type(val)}")
 
 
+class VisionProjectorType:
+    GEMMA3 = "gemma3"
+    IDEFICS3 = "idefics3"
+    PIXTRAL = "pixtral"
+
+
 # Items here are (block size, type size)
 QK_K = 256
 GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {

gguf-py/gguf/gguf_writer.py

@@ -931,6 +931,53 @@ class GGUFWriter:
     def add_eom_token_id(self, id: int) -> None:
         self.add_uint32(Keys.Tokenizer.EOM_ID, id)
 
+    # for vision models
+
+    def add_vision_projection_dim(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.PROJECTION_DIM, value)
+
+    def add_vision_has_vision_encoder(self, value: bool) -> None:
+        self.add_bool(Keys.ClipVision.HAS_VISION_ENCODER, value)
+
+    def add_vision_patch_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.PATCH_SIZE, value)
+
+    def add_vision_embedding_length(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.EMBEDDING_LENGTH, value)
+
+    def add_vision_feed_forward_length(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.FEED_FORWARD_LENGTH, value)
+
+    def add_vision_block_count(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.BLOCK_COUNT, value)
+
+    def add_vision_head_count(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.Attention.HEAD_COUNT, value)
+
+    def add_vision_projector_type(self, value: str) -> None:
+        self.add_string(Keys.ClipVision.PROJECTOR_TYPE, value)
+
+    def add_vision_attention_layernorm_eps(self, value: float) -> None:
+        self.add_float32(Keys.ClipVision.Attention.LAYERNORM_EPS, value)
+
+    def add_vision_image_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.IMAGE_SIZE, value)
+
+    def add_vision_image_mean(self, values: Sequence[float]) -> None:
+        self.add_array(Keys.ClipVision.IMAGE_MEAN, values)
+
+    def add_vision_image_std(self, values: Sequence[float]) -> None:
+        self.add_array(Keys.ClipVision.IMAGE_STD, values)
+
+    def add_vision_use_gelu(self, value: bool) -> None:
+        self.add_bool(Keys.ClipVision.USE_GELU, value)
+
+    def add_vision_use_silu(self, value: bool) -> None:
+        self.add_bool(Keys.ClipVision.USE_SILU, value)
+
+    def add_vision_projector_scale_factor(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.Projector.SCALE_FACTOR, value)
+
     def _pack(self, fmt: str, value: Any, skip_pack_prefix: bool = False) -> bytes:
         pack_prefix = ''
         if not skip_pack_prefix:

gguf-py/gguf/tensor_mapping.py

@@ -914,6 +914,7 @@ class TensorNameMap:
             "vision_tower.vision_model.embeddings.patch_embedding",
             "vpm.embeddings.patch_embedding",
             "model.vision_model.embeddings.patch_embedding", # SmolVLM
+            "vision_tower.patch_conv", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_EMBD_POS: (
@@ -926,52 +927,65 @@ class TensorNameMap:
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.q_proj",
             "vpm.encoder.layers.{bid}.self_attn.q_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.q_proj", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.attention.q_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_K: (
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.k_proj",
             "vpm.encoder.layers.{bid}.self_attn.k_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.k_proj", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.attention.k_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_V: (
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.v_proj",
             "vpm.encoder.layers.{bid}.self_attn.v_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.v_proj", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.attention.v_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_INPUT_NORM: (
             "vision_tower.vision_model.encoder.layers.{bid}.layer_norm1",
             "vpm.encoder.layers.{bid}.layer_norm1",
             "model.vision_model.encoder.layers.{bid}.layer_norm1", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.attention_norm", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_OUTPUT: (
             "vision_tower.vision_model.encoder.layers.{bid}.self_attn.out_proj",
             "vpm.encoder.layers.{bid}.self_attn.out_proj",
             "model.vision_model.encoder.layers.{bid}.self_attn.out_proj", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.attention.o_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_OUTPUT_NORM: (
             "vision_tower.vision_model.encoder.layers.{bid}.layer_norm2",
             "vpm.encoder.layers.{bid}.layer_norm2",
             "model.vision_model.encoder.layers.{bid}.layer_norm2", # SmolVLM
+            "vision_tower.transformer.layers.{bid}.ffn_norm", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_FFN_UP: (
             "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc1",
             "vpm.encoder.layers.{bid}.mlp.fc1",
-            "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM
+            "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM, gemma3 (note: name is swapped)
+            "vision_tower.transformer.layers.{bid}.feed_forward.up_proj", # pixtral
+        ),
+
+        MODEL_TENSOR.V_ENC_FFN_GATE: (
+            "vision_tower.transformer.layers.{bid}.feed_forward.gate_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_ENC_FFN_DOWN: (
             "vision_tower.vision_model.encoder.layers.{bid}.mlp.fc2",
             "vpm.encoder.layers.{bid}.mlp.fc2",
-            "model.vision_model.encoder.layers.{bid}.mlp.fc2", # SmolVLM
+            "model.vision_model.encoder.layers.{bid}.mlp.fc1", # SmolVLM, gemma3 (note: name is swapped)
+            "vision_tower.transformer.layers.{bid}.feed_forward.down_proj", # pixtral
         ),
 
         MODEL_TENSOR.V_PRE_NORM: (
             "vision_tower.vision_model.pre_layrnorm",
+            "vision_tower.ln_pre", # pixtral
         ),
 
         MODEL_TENSOR.V_POST_NORM: (
@@ -1030,6 +1044,10 @@ class TensorNameMap:
         MODEL_TENSOR.V_RESMPL_QUERY: (
             "resampler.query",
         ),
+
+        MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK: (
+            "v.token_embd.img_break", # for pixtral, this is a generated vector
+        ),
     }
 
     # architecture-specific block mappings

grammars/README.md

@@ -112,7 +112,7 @@ You can use GBNF grammars:
 
 - In [llama-server](../examples/server)'s completion endpoints, passed as the `grammar` body field
 - In [llama-cli](../examples/main), passed as the `--grammar` & `--grammar-file` flags
-- With [llama-gbnf-validator](../examples/gbnf-validator) tool, to test them against strings.
+- With [test-gbnf-validator](../tests/test-gbnf-validator.cpp), to test them against strings.
 
 ## JSON Schemas → GBNF
 

include/llama.h

@@ -111,6 +111,7 @@ extern "C" {
         LLAMA_VOCAB_PRE_TYPE_TRILLION       = 31,
         LLAMA_VOCAB_PRE_TYPE_BAILINGMOE     = 32,
         LLAMA_VOCAB_PRE_TYPE_LLAMA4         = 33,
+        LLAMA_VOCAB_PRE_TYPE_PIXTRAL        = 34,
     };
 
     enum llama_rope_type {

models/ggml-vocab-pixtral.gguf.inp

@@ -0,0 +1,112 @@
+ied 4 ½ months
+__ggml_vocab_test__
+Führer
+__ggml_vocab_test__
+
+__ggml_vocab_test__
+ 
+__ggml_vocab_test__
+  
+__ggml_vocab_test__
+   
+__ggml_vocab_test__
+	
+__ggml_vocab_test__
+
+
+__ggml_vocab_test__
+
+
+
+__ggml_vocab_test__
+
+
+
+
+__ggml_vocab_test__
+	
+
+__ggml_vocab_test__
+Hello world
+__ggml_vocab_test__
+ Hello world
+__ggml_vocab_test__
+Hello World
+__ggml_vocab_test__
+ Hello World
+__ggml_vocab_test__
+ Hello World!
+__ggml_vocab_test__
+Hello, world!
+__ggml_vocab_test__
+ Hello, world!
+__ggml_vocab_test__
+ this is 🦙.cpp
+__ggml_vocab_test__
+w048 7tuijk dsdfhu
+__ggml_vocab_test__
+нещо на Български
+__ggml_vocab_test__
+កាន់តែពិសេសអាចខលចេញ
+__ggml_vocab_test__
+🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)
+__ggml_vocab_test__
+Hello
+__ggml_vocab_test__
+ Hello
+__ggml_vocab_test__
+  Hello
+__ggml_vocab_test__
+   Hello
+__ggml_vocab_test__
+    Hello
+__ggml_vocab_test__
+    Hello
+    Hello
+__ggml_vocab_test__
+ (
+__ggml_vocab_test__
+
+ =
+__ggml_vocab_test__
+' era
+__ggml_vocab_test__
+Hello, y'all! How are you 😁 ?我想在apple工作1314151天～
+__ggml_vocab_test__
+!!!!!!
+__ggml_vocab_test__
+3
+__ggml_vocab_test__
+33
+__ggml_vocab_test__
+333
+__ggml_vocab_test__
+3333
+__ggml_vocab_test__
+33333
+__ggml_vocab_test__
+333333
+__ggml_vocab_test__
+3333333
+__ggml_vocab_test__
+33333333
+__ggml_vocab_test__
+333333333
+__ggml_vocab_test__
+Cửa Việt
+__ggml_vocab_test__
+ discards
+__ggml_vocab_test__
+
+ 
+
+ 
+
+
+ 	 		 	
+  
+   
+    
+     
+🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български ''''''```````""""......!!!!!!?????? I've been 'told he's there, 'RE you sure? 'M not sure I'll make it, 'D you like some tea? We'Ve a'lL
+__ggml_vocab_test__

models/ggml-vocab-pixtral.gguf.out

@@ -0,0 +1,46 @@
+ 2014 1032 1052 1032 28504 6972
+ 1070 7088 1258
+
+ 1032
+ 1256
+ 1293
+ 1009
+ 1010
+ 1267
+ 4688
+ 1009 1010
+ 22177 4304
+ 45383 4304
+ 22177 5325
+ 45383 5325
+ 45383 5325 1033
+ 22177 1044 4304 1033
+ 45383 1044 4304 1033
+ 1593 1395 119685 1166 1153 1046 51228
+ 1119 1048 1052 1056 1032 1055 17391 23216 30203 7785 17279
+ 3337 30757 1902 4200 63073 3671
+ 1225 1158 1128 1225 1158 1182 1225 1158 1147 1225 1159 1139 1225 1158 1143 1225 1159 1130 1225 1158 1150 1225 1158 1183 1225 1158 1159 1225 21359 1225 1158 1159 1225 1158 1162 1225 1158 1182 1225 1158 1133 1225 1158 1129 1225 1158 1155 1225 1158 1133 1225 21359 1225 1158 1137
+ 1240 1159 1154 1128 1319 13052 1041 119685 1152 1182 29568 1240 1159 1140 1171 1239 1184 1143 1319 88181 1873 3659 1275 56421 1621 1041 126241 1133 1319 11234 1873 26303 1455 1934 2246 3754 10835 1041
+ 22177
+ 45383
+ 1032 45383
+ 1256 45383
+ 1293 45383
+ 1293 45383 1010 1293 45383
+ 1319
+ 1010 1376
+ 1039 4033
+ 22177 1044 1404 48054 1033 3075 1584 1636 119685 1152 1129 3082 26060 2998 63614 82278 1049 1051 1049 1052 1049 1053 1049 6434 6749
+ 7290 7290 7290
+ 1051
+ 1051 1051
+ 1051 1051 1051
+ 1051 1051 1051 1051
+ 1051 1051 1051 1051 1051
+ 1051 1051 1051 1051 1051 1051
+ 1051 1051 1051 1051 1051 1051 1051
+ 1051 1051 1051 1051 1051 1051 1051 1051
+ 1051 1051 1051 1051 1051 1051 1051 1051 1051
+ 1067 59503 28783
+ 3724 4058
+ 1010 1032 1267 1032 4688 1032 17152 1458 29356 1010 1256 1010 1293 1010 1260 1010 1652 1010 1240 1159 1154 1128 1319 13052 1041 119685 1152 1182 29568 1240 1159 1140 1171 1239 1184 1143 1319 88181 1873 3659 1275 56421 1621 1041 126241 1133 119685 1166 1153 1240 1159 1166 1153 1032 1051 1032 1051 1051 1032 1051 1051 1051 1032 1051 1051 1051 1051 1032 1051 1051 1051 1051 1051 1032 1051 1051 1051 1051 1051 1051 1032 1051 1051 1051 1051 1051 1051 1051 1032 1051 1051 1051 1051 1051 1051 1051 1051 1032 1051 1046 1051 1032 1051 1791 1051 1032 1051 2880 1051 71881 1158 1128 1225 1158 1182 1225 1158 1147 1225 1159 1139 1225 1158 1143 1225 1159 1130 1225 1158 1150 1225 1158 1183 1225 1158 1159 1225 21359 1225 1158 1159 1225 1158 1162 1225 1158 1182 1225 1158 1133 1240 1159 1152 1129 3082 26060 2998 63614 82278 1049 1051 1049 1052 1049 1053 1049 6434 6749 45577 1045 6626 43555 2843 30757 1902 4200 63073 3671 14931 20040 20040 1657 1657 1975 14135 14135 83923 7290 7290 7290 45509 45509 45509 1362 6483 2151 1576 1116 2189 1514 1681 2156 1044 1576 3609 1636 5257 1063 1576 1077 1605 5257 1362 7534 3180 1494 1044 1576 1068 1636 2479 2269 26883 1063 2837 1039 45654 1261 54297 1076

scripts/sync-ggml.last

@@ -1 +1 @@
-f71d538ece3fb32a04824dc6d1e73e360be9d22f
+13bcf9ce50651a8b4238ec6d136f46f2c1b23b6f

src/CMakeLists.txt

@@ -32,8 +32,9 @@ add_library(llama
             unicode.h
             )
 
-target_include_directories(llama PUBLIC . ../include)
-target_compile_features   (llama PUBLIC cxx_std_17) # don't bump
+target_include_directories(llama PRIVATE .)
+target_include_directories(llama PUBLIC ../include)
+target_compile_features   (llama PRIVATE cxx_std_17) # don't bump
 
 target_link_libraries(llama PUBLIC ggml)
 

src/llama-chat.cpp

@@ -50,8 +50,8 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "deepseek3",         LLM_CHAT_TEMPLATE_DEEPSEEK_3        },
     { "command-r",         LLM_CHAT_TEMPLATE_COMMAND_R         },
     { "llama3",            LLM_CHAT_TEMPLATE_LLAMA_3           },
-    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGML_3         },
-    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGML_4         },
+    { "chatglm3",          LLM_CHAT_TEMPLATE_CHATGLM_3         },
+    { "chatglm4",          LLM_CHAT_TEMPLATE_CHATGLM_4         },
     { "glmedge",           LLM_CHAT_TEMPLATE_GLMEDGE           },
     { "minicpm",           LLM_CHAT_TEMPLATE_MINICPM           },
     { "exaone3",           LLM_CHAT_TEMPLATE_EXAONE_3          },
@@ -62,6 +62,7 @@ static const std::map<std::string, llm_chat_template> LLM_CHAT_TEMPLATES = {
     { "yandex",            LLM_CHAT_TEMPLATE_YANDEX            },
     { "bailing",           LLM_CHAT_TEMPLATE_BAILING           },
     { "llama4",            LLM_CHAT_TEMPLATE_LLAMA4            },
+    { "smolvlm",           LLM_CHAT_TEMPLATE_SMOLVLM           },
 };
 
 llm_chat_template llm_chat_template_from_str(const std::string & name) {
@@ -81,7 +82,9 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
     if (tmpl_contains("<|im_start|>")) {
         return tmpl_contains("<|im_sep|>")
             ? LLM_CHAT_TEMPLATE_PHI_4
-            : LLM_CHAT_TEMPLATE_CHATML;
+            : tmpl_contains("<end_of_utterance>")
+                ? LLM_CHAT_TEMPLATE_SMOLVLM // SmolVLM uses <|im_start|> as BOS, but it is NOT chatml
+                : LLM_CHAT_TEMPLATE_CHATML;
     } else if (tmpl.find("mistral") == 0 || tmpl_contains("[INST]")) {
         if (tmpl_contains("[SYSTEM_PROMPT]")) {
             return LLM_CHAT_TEMPLATE_MISTRAL_V7;
@@ -119,6 +122,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         }
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|end|>")) {
         return LLM_CHAT_TEMPLATE_PHI_3;
+    } else if (tmpl_contains("[gMASK]<sop>")) {
+        return LLM_CHAT_TEMPLATE_CHATGLM_4;
     } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
         return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
     } else if (tmpl_contains("<|{{ item['role'] }}|>") && tmpl_contains("<|begin_of_image|>")) {
@@ -151,9 +156,7 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
         return LLM_CHAT_TEMPLATE_LLAMA_3;
     } else if (tmpl_contains("[gMASK]sop")) {
         // chatglm3-6b
-        return LLM_CHAT_TEMPLATE_CHATGML_3;
-    } else if (tmpl_contains("[gMASK]<sop>")) {
-        return LLM_CHAT_TEMPLATE_CHATGML_4;
+        return LLM_CHAT_TEMPLATE_CHATGLM_3;
     } else if (tmpl_contains(LU8("<用户>"))) {
         // MiniCPM-3B-OpenHermes-2.5-v2-GGUF
         return LLM_CHAT_TEMPLATE_MINICPM;
@@ -434,7 +437,7 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_3) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_3) {
         // chatglm3-6b
         ss << "[gMASK]" << "sop";
         for (auto message : chat) {
@@ -444,7 +447,7 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|assistant|>";
         }
-    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGML_4) {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_CHATGLM_4) {
         ss << "[gMASK]" << "<sop>";
         for (auto message : chat) {
             std::string role(message->role);
@@ -622,7 +625,23 @@ int32_t llm_chat_apply_template(
         if (add_ass) {
             ss << "<|header_start|>assistant<|header_end|>\n\n";
         }
-    }  else {
+    } else if (tmpl == LLM_CHAT_TEMPLATE_SMOLVLM) {
+        // SmolVLM
+        ss << "<|im_start|>"; // uses <|im_start|> as BOS, but the actual content is NOT chatml
+        for (auto message : chat) {
+            std::string role(message->role);
+            if (role == "system") {
+                ss << message->content << "\n\n";
+            } else if (role == "user") {
+                ss << "User: " << message->content << "<end_of_utterance>\n";
+            } else {
+                ss << "Assistant: " << message->content << "<end_of_utterance>\n";
+            }
+        }
+        if (add_ass) {
+            ss << "Assistant:";
+        }
+    } else {
         // template not supported
         return -1;
     }

src/llama-chat.h

@@ -29,8 +29,8 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_DEEPSEEK_3,
     LLM_CHAT_TEMPLATE_COMMAND_R,
     LLM_CHAT_TEMPLATE_LLAMA_3,
-    LLM_CHAT_TEMPLATE_CHATGML_3,
-    LLM_CHAT_TEMPLATE_CHATGML_4,
+    LLM_CHAT_TEMPLATE_CHATGLM_3,
+    LLM_CHAT_TEMPLATE_CHATGLM_4,
     LLM_CHAT_TEMPLATE_GLMEDGE,
     LLM_CHAT_TEMPLATE_MINICPM,
     LLM_CHAT_TEMPLATE_EXAONE_3,
@@ -41,6 +41,7 @@ enum llm_chat_template {
     LLM_CHAT_TEMPLATE_YANDEX,
     LLM_CHAT_TEMPLATE_BAILING,
     LLM_CHAT_TEMPLATE_LLAMA4,
+    LLM_CHAT_TEMPLATE_SMOLVLM,
     LLM_CHAT_TEMPLATE_UNKNOWN,
 };
 

src/llama-context.cpp

@@ -469,8 +469,7 @@ ggml_tensor * llama_context::build_rope_shift(
         ggml_tensor * shift,
         ggml_tensor * factors,
               float   freq_base,
-              float   freq_scale,
-        ggml_backend_buffer * bbuf) const {
+              float   freq_scale) const {
     const auto & n_ctx_orig = cparams.n_ctx_orig_yarn;
 
     const auto & yarn_ext_factor  = cparams.yarn_ext_factor;
@@ -492,17 +491,7 @@ ggml_tensor * llama_context::build_rope_shift(
         // dequantize to f32 -> RoPE -> quantize back
         tmp = ggml_cast(ctx0, cur, GGML_TYPE_F32);
 
-        if (bbuf) {
-            for (const auto & backend : backends) {
-                // Figure out which backend KV cache belongs to
-                if (ggml_backend_supports_buft(backend.get(), ggml_backend_buffer_get_type(bbuf))) {
-                    ggml_backend_sched_set_tensor_backend(sched.get(), tmp, backend.get());
-                    break;
-                }
-            }
-        }
-
-        tmp = ggml_rope_ext_inplace(ctx0, tmp,
+        tmp = ggml_rope_ext(ctx0, tmp,
                 shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
                 yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow);
 
@@ -582,7 +571,7 @@ llm_graph_result_ptr llama_context::build_kv_self_shift(
                 ggml_row_size(kv_self->k_l[il]->type, n_embd_k_gqa),
                 0);
 
-        ggml_tensor * cur = build_rope_shift(ctx0, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l, kv_self->k_l[il]->buffer);
+        ggml_tensor * cur = build_rope_shift(ctx0, k, inp->k_shift, rope_factors, freq_base_l, freq_scale_l);
 
         ggml_build_forward_expand(gf, cur);
     }

src/llama-context.h

@@ -170,8 +170,7 @@ private:
         ggml_tensor * shift,
         ggml_tensor * factors,
               float   freq_base,
-              float   freq_scale,
-        ggml_backend_buffer * bbuf) const;
+              float   freq_scale) const;
 
     llm_graph_result_ptr build_kv_self_shift(
             ggml_context * ctx0,

src/llama-graph.cpp

@@ -803,6 +803,10 @@ ggml_tensor * llm_graph_context::build_ffn(
 
     if (down) {
         cur = build_lora_mm(down, cur);
+        if (arch == LLM_ARCH_GLM4) {
+            // GLM4 seems to have numerical issues with half-precision accumulators
+            ggml_mul_mat_set_prec(cur, GGML_PREC_F32);
+        }
     }
 
     if (down_b) {

src/llama-model.cpp

@@ -10149,6 +10149,7 @@ struct llm_build_deepseek2 : public llm_graph_context {
 
                     // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
                     ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
+                    ggml_mul_mat_set_prec(q_nope_absorbed, GGML_PREC_F32);
                     cb(q_nope_absorbed, "q_nope_absorbed", il);
 
                     // {kv_lora_rank, n_head, n_tokens}

src/llama-vocab.cpp

@@ -1506,7 +1506,8 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "llama3"   ||
                     tokenizer_pre == "llama-v3" ||
                     tokenizer_pre == "llama-bpe"||
-                    tokenizer_pre == "falcon3") {
+                    tokenizer_pre == "falcon3"  ||
+                    tokenizer_pre == "pixtral") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
                 ignore_merges = true;
                 add_bos = true;

tests/CMakeLists.txt

@@ -1,5 +1,17 @@
 llama_add_compile_flags()
 
+function(llama_build source)
+    if (DEFINED LLAMA_TEST_NAME)
+        set(TEST_TARGET ${LLAMA_TEST_NAME})
+    else()
+        get_filename_component(TEST_TARGET ${source} NAME_WE)
+    endif()
+
+    add_executable(${TEST_TARGET} ${source})
+    target_link_libraries(${TEST_TARGET} PRIVATE common)
+    install(TARGETS ${TEST_TARGET} RUNTIME)
+endfunction()
+
 function(llama_test target)
     include(CMakeParseArguments)
     set(options)
@@ -36,7 +48,7 @@ endfunction()
 # - LABEL: label for the test (defaults to main)
 # - ARGS: arguments to pass to the test executable
 # - WORKING_DIRECTORY
-function(llama_target_and_test source)
+function(llama_build_and_test source)
     include(CMakeParseArguments)
     set(options)
     set(oneValueArgs NAME LABEL WORKING_DIRECTORY)
@@ -58,6 +70,7 @@ function(llama_target_and_test source)
     add_executable(${TEST_TARGET} ${source} get-model.cpp)
     install(TARGETS ${TEST_TARGET} RUNTIME)
     target_link_libraries(${TEST_TARGET} PRIVATE common)
+
     add_test(
         NAME ${TEST_TARGET}
         WORKING_DIRECTORY ${LLAMA_TEST_WORKING_DIRECTORY}
@@ -68,9 +81,7 @@ function(llama_target_and_test source)
 endfunction()
 
 # build test-tokenizer-0 target once and add many tests
-add_executable(test-tokenizer-0 test-tokenizer-0.cpp)
-target_link_libraries(test-tokenizer-0 PRIVATE common)
-install(TARGETS test-tokenizer-0 RUNTIME)
+llama_build(test-tokenizer-0.cpp)
 
 llama_test(test-tokenizer-0 NAME test-tokenizer-0-bert-bge          ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-bert-bge.gguf)
 llama_test(test-tokenizer-0 NAME test-tokenizer-0-command-r         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-command-r.gguf)
@@ -87,27 +98,27 @@ llama_test(test-tokenizer-0 NAME test-tokenizer-0-refact            ARGS ${CMAKE
 llama_test(test-tokenizer-0 NAME test-tokenizer-0-starcoder         ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
 
 if (LLAMA_LLGUIDANCE)
-    llama_target_and_test(test-grammar-llguidance.cpp ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf)
+    llama_build_and_test(test-grammar-llguidance.cpp ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-bpe.gguf)
 endif ()
 
 if (NOT WIN32)
     # these tests are disabled on Windows because they use internal functions not exported with LLAMA_API
-    llama_target_and_test(test-sampling.cpp)
-    llama_target_and_test(test-grammar-parser.cpp)
-    llama_target_and_test(test-grammar-integration.cpp)
-    llama_target_and_test(test-llama-grammar.cpp)
-    llama_target_and_test(test-chat.cpp)
+    llama_build_and_test(test-sampling.cpp)
+    llama_build_and_test(test-grammar-parser.cpp)
+    llama_build_and_test(test-grammar-integration.cpp)
+    llama_build_and_test(test-llama-grammar.cpp)
+    llama_build_and_test(test-chat.cpp)
     # TODO: disabled on loongarch64 because the ggml-ci node lacks Python 3.8
     if (NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
-        llama_target_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..)
+        llama_build_and_test(test-json-schema-to-grammar.cpp   WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/..)
         target_include_directories(test-json-schema-to-grammar PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/../examples/server)
     endif()
 
+    llama_build(test-quantize-stats.cpp)
+    llama_build(test-gbnf-validator.cpp)
 
     # build test-tokenizer-1-bpe target once and add many tests
-    add_executable(test-tokenizer-1-bpe test-tokenizer-1-bpe.cpp)
-    target_link_libraries(test-tokenizer-1-bpe PRIVATE common)
-    install(TARGETS test-tokenizer-1-bpe RUNTIME)
+    llama_build(test-tokenizer-1-bpe.cpp)
 
     # TODO: disabled due to slowness
     #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-aquila    ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-aquila.gguf)
@@ -120,37 +131,35 @@ if (NOT WIN32)
     #llama_test(test-tokenizer-1-bpe NAME test-tokenizer-1-starcoder ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-starcoder.gguf)
 
     # build test-tokenizer-1-spm target once and add many tests
-    add_executable(test-tokenizer-1-spm test-tokenizer-1-spm.cpp)
-    target_link_libraries(test-tokenizer-1-spm PRIVATE common)
-    install(TARGETS test-tokenizer-1-spm RUNTIME)
+    llama_build(test-tokenizer-1-spm.cpp)
 
     llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-llama-spm ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-llama-spm.gguf)
     #llama_test(test-tokenizer-1-spm  NAME test-tokenizer-1-baichuan  ARGS ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab-baichuan.gguf)
 
-    # llama_target_and_test(test-double-float.cpp) # SLOW
+    # llama_build_and_test(test-double-float.cpp) # SLOW
 endif()
 
-llama_target_and_test(test-log.cpp)
-llama_target_and_test(test-chat-template.cpp)
+llama_build_and_test(test-log.cpp)
+llama_build_and_test(test-chat-template.cpp)
 
 # this fails on windows (github hosted runner) due to curl DLL not found (exit code 0xc0000135)
 if (NOT WIN32)
-    llama_target_and_test(test-arg-parser.cpp)
+    llama_build_and_test(test-arg-parser.cpp)
 endif()
 
-# llama_target_and_test(test-opt.cpp) # SLOW
-llama_target_and_test(test-gguf.cpp)
-llama_target_and_test(test-backend-ops.cpp)
+# llama_build_and_test(test-opt.cpp) # SLOW
+llama_build_and_test(test-gguf.cpp)
+llama_build_and_test(test-backend-ops.cpp)
 
-llama_target_and_test(test-model-load-cancel.cpp  LABEL "model")
-llama_target_and_test(test-autorelease.cpp        LABEL "model")
+llama_build_and_test(test-model-load-cancel.cpp  LABEL "model")
+llama_build_and_test(test-autorelease.cpp        LABEL "model")
 
 if (NOT GGML_BACKEND_DL)
     # these tests use the backends directly and cannot be built with dynamic loading
-    llama_target_and_test(test-barrier.cpp)
-    llama_target_and_test(test-quantize-fns.cpp)
-    llama_target_and_test(test-quantize-perf.cpp)
-    llama_target_and_test(test-rope.cpp)
+    llama_build_and_test(test-barrier.cpp)
+    llama_build_and_test(test-quantize-fns.cpp)
+    llama_build_and_test(test-quantize-perf.cpp)
+    llama_build_and_test(test-rope.cpp)
 endif()
 
 

tests/test-arg-parser.cpp

@@ -126,6 +126,53 @@ int main(void) {
     assert(params.cpuparams.n_threads == 1010);
 #endif // _WIN32
 
+    if (common_has_curl()) {
+        printf("test-arg-parser: test curl-related functions\n\n");
+        const char * GOOD_URL = "https://raw.githubusercontent.com/ggml-org/llama.cpp/refs/heads/master/README.md";
+        const char * BAD_URL  = "https://www.google.com/404";
+        const char * BIG_FILE = "https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v1.bin";
+
+        {
+            printf("test-arg-parser: test good URL\n\n");
+            auto res = common_remote_get_content(GOOD_URL, {});
+            assert(res.first == 200);
+            assert(res.second.size() > 0);
+            std::string str(res.second.data(), res.second.size());
+            assert(str.find("llama.cpp") != std::string::npos);
+        }
+
+        {
+            printf("test-arg-parser: test bad URL\n\n");
+            auto res = common_remote_get_content(BAD_URL, {});
+            assert(res.first == 404);
+        }
+
+        {
+            printf("test-arg-parser: test max size error\n");
+            common_remote_params params;
+            params.max_size = 1;
+            try {
+                common_remote_get_content(GOOD_URL, params);
+                assert(false && "it should throw an error");
+            } catch (std::exception & e) {
+                printf("  expected error: %s\n\n", e.what());
+            }
+        }
+
+        {
+            printf("test-arg-parser: test timeout error\n");
+            common_remote_params params;
+            params.timeout = 1;
+            try {
+                common_remote_get_content(BIG_FILE, params);
+                assert(false && "it should throw an error");
+            } catch (std::exception & e) {
+                printf("  expected error: %s\n\n", e.what());
+            }
+        }
+    } else {
+        printf("test-arg-parser: no curl, skipping curl-related functions\n");
+    }
 
     printf("test-arg-parser: all tests OK\n\n");
 }

tests/test-backend-ops.cpp

@@ -2071,7 +2071,7 @@ struct test_mul_mat_id : public test_case {
     const ggml_type type_b;
     const int n_mats;
     const int n_used;
-    const bool b; // brodcast b matrix
+    const bool b; // broadcast b matrix
     const int64_t m;
     const int64_t n;
     const int64_t k;
@@ -2606,6 +2606,8 @@ struct test_rope : public test_case {
             } else {
                 out = ggml_rope_ext_back(ctx, a, pos, freq, n_dims, mode, 0, 10000.0f, fs, ef, af, 1.0f, 1.0f);
             }
+
+            // TODO: add test with a non-contiguous view as input ; this case is needed for build_rope_2d in clip.cpp
         }
         ggml_set_name(out, "out");
 

tests/test-chat.cpp

@@ -11,8 +11,9 @@
 #include <string>
 
 #include "chat.h"
-#include "llama-grammar.h"
-#include "unicode.h"
+
+#include "../src/unicode.h"
+#include "../src/llama-grammar.h"
 
 using json = nlohmann::ordered_json;
 

tests/test-gbnf-validator.cpp

@@ -1,5 +1,5 @@
-#include "unicode.h"
-#include "llama-grammar.h"
+#include "../src/unicode.h"
+#include "../src/llama-grammar.h"
 
 #include <cstdio>
 #include <cstdlib>

tests/test-grammar-integration.cpp

@@ -2,10 +2,11 @@
 #undef NDEBUG
 #endif
 
-#include "unicode.h"
-#include "llama-grammar.h"
 #include "json-schema-to-grammar.h"
 
+#include "../src/unicode.h"
+#include "../src/llama-grammar.h"
+
 #include <cassert>
 #include <string>
 #include <vector>

tests/test-grammar-llguidance.cpp

@@ -2,7 +2,6 @@
 #    undef NDEBUG
 #endif
 
-#include "unicode.h"
 #include "sampling.h"
 
 #include <cassert>
@@ -84,7 +83,7 @@ static void test(const std::string & test_desc, const std::string & grammar_str,
 
             fprintf(stderr,
                     "\n NOTE: Debug grammar file generated. To analyze this failure in detail, run the following "
-                    "command:     ./llama-gbnf-validator test-grammar-integration.grammar.gbnf "
+                    "command:     ./test-gbnf-validator test-grammar-integration.grammar.gbnf "
                     "test-grammar-integration.string.txt\n\n");
         } else {
             fprintf(stdout, "✅︎\n");

tests/test-grammar-parser.cpp

@@ -3,7 +3,9 @@
 #endif
 
 #include "llama.h"
-#include "llama-grammar.h"
+
+// TODO: shold not include libllama sources
+#include "../src/llama-grammar.h"
 
 #include <cassert>
 

tests/test-json-schema-to-grammar.cpp

@@ -4,7 +4,7 @@
 
 #include "json-schema-to-grammar.h"
 
-#include "llama-grammar.h"
+#include "../src/llama-grammar.h"
 
 #include <cassert>
 #include <fstream>
@@ -597,6 +597,22 @@ static void test_all(const std::string & lang, std::function<void(const TestCase
         )"""
     });
 
+    test({
+        SUCCESS,
+        "maxItems 0",
+        R"""({
+            "items": {
+                "type": "boolean"
+            },
+            "maxItems": 0
+        })""",
+        R"""(
+            boolean ::= ("true" | "false") space
+            root ::= "[" space  "]" space
+            space ::= | " " | "\n"{1,2} [ \t]{0,20}
+        )"""
+    });
+
     test({
         SUCCESS,
         "maxItems 1",

tests/test-llama-grammar.cpp

@@ -3,7 +3,8 @@
 #endif
 
 #include "llama.h"
-#include "llama-grammar.h"
+
+#include "../src/llama-grammar.h"
 
 #include <cassert>
 #include <stdexcept>

tests/test-quantize-stats.cpp

@@ -1,8 +1,9 @@
 #include "ggml.h"
 #include "llama.h"
-#include "llama-model.h"
 #include "common.h"
 
+#include "../src/llama-model.h"
+
 #include <algorithm>
 #include <cassert>
 #include <cinttypes>

tests/test-tokenizer-1-bpe.cpp

@@ -1,8 +1,9 @@
 #include "llama.h"
 #include "common.h"
-#include "unicode.h"
 #include "console.h"
 
+#include "../src/unicode.h"
+
 #include <cassert>
 #include <codecvt>
 #include <cstdio>

tests/test-tokenizer-1-spm.cpp

@@ -1,8 +1,9 @@
 #include "llama.h"
 #include "common.h"
-#include "unicode.h"
 #include "console.h"
 
+#include "../src/unicode.h"
+
 #include <cassert>
 #include <codecvt>
 #include <cstdio>