mtmd : add **vision** support for Mistral Small 3.1 (#13231)

* convert ok

* load ok, missing patch merger

* ah sheet it works

* update llava/readme

* add test

* fix test

common/arg.cpp

@@ -43,6 +43,25 @@ std::initializer_list<enum llama_example> mmproj_examples = {
     // TODO: add LLAMA_EXAMPLE_SERVER when it's ready
 };
 
+static std::string read_file(const std::string & fname) {
+    std::ifstream file(fname);
+    if (!file) {
+        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
+    }
+    std::string content((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
+    file.close();
+    return content;
+}
+
+static void write_file(const std::string & fname, const std::string & content) {
+    std::ofstream file(fname);
+    if (!file) {
+        throw std::runtime_error(string_format("error: failed to open file '%s'\n", fname.c_str()));
+    }
+    file << content;
+    file.close();
+}
+
 common_arg & common_arg::set_examples(std::initializer_list<enum llama_example> examples) {
     this->examples = std::move(examples);
     return *this;
@@ -198,11 +217,11 @@ struct curl_slist_ptr {
 #define CURL_MAX_RETRY 3
 #define CURL_RETRY_DELAY_SECONDS 2
 
-static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds) {
+static bool curl_perform_with_retry(const std::string & url, CURL * curl, int max_attempts, int retry_delay_seconds, const char * method_name) {
     int remaining_attempts = max_attempts;
 
     while (remaining_attempts > 0) {
-        LOG_INF("%s: Trying to download from %s (attempt %d of %d)...\n", __func__ , url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
+        LOG_INF("%s: %s %s (attempt %d of %d)...\n", __func__ , method_name, url.c_str(), max_attempts - remaining_attempts + 1, max_attempts);
 
         CURLcode res = curl_easy_perform(curl);
         if (res == CURLE_OK) {
@@ -213,6 +232,7 @@ static bool curl_perform_with_retry(const std::string & url, CURL * curl, int ma
         LOG_WRN("%s: curl_easy_perform() failed: %s, retrying after %d milliseconds...\n", __func__, curl_easy_strerror(res), exponential_backoff_delay);
 
         remaining_attempts--;
+        if (remaining_attempts == 0) break;
         std::this_thread::sleep_for(std::chrono::milliseconds(exponential_backoff_delay));
     }
 
@@ -231,8 +251,6 @@ static bool common_download_file_single(const std::string & url, const std::stri
         return false;
     }
 
-    bool force_download = false;
-
     // Set the URL, allow to follow http redirection
     curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
     curl_easy_setopt(curl.get(), CURLOPT_FOLLOWLOCATION, 1L);
@@ -256,7 +274,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
 
     // If the file exists, check its JSON metadata companion file.
     std::string metadata_path = path + ".json";
-    nlohmann::json metadata;
+    nlohmann::json metadata; // TODO @ngxson : get rid of this json, use regex instead
     std::string etag;
     std::string last_modified;
 
@@ -266,14 +284,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
         if (metadata_in.good()) {
             try {
                 metadata_in >> metadata;
-                LOG_INF("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
-                if (metadata.contains("url") && metadata.at("url").is_string()) {
-                    auto previous_url = metadata.at("url").get<std::string>();
-                    if (previous_url != url) {
-                        LOG_ERR("%s: Model URL mismatch: %s != %s\n", __func__, url.c_str(), previous_url.c_str());
-                        return false;
-                    }
-                }
+                LOG_DBG("%s: previous metadata file found %s: %s\n", __func__, metadata_path.c_str(), metadata.dump().c_str());
                 if (metadata.contains("etag") && metadata.at("etag").is_string()) {
                     etag = metadata.at("etag");
                 }
@@ -281,10 +292,10 @@ static bool common_download_file_single(const std::string & url, const std::stri
                     last_modified = metadata.at("lastModified");
                 }
             } catch (const nlohmann::json::exception & e) {
-            LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
-                return false;
+                LOG_ERR("%s: error reading metadata file %s: %s\n", __func__, metadata_path.c_str(), e.what());
             }
         }
+        // if we cannot open the metadata file, we assume that the downloaded file is not valid (etag and last-modified are left empty, so we will download it again)
     } else {
         LOG_INF("%s: no previous model file found %s\n", __func__, path.c_str());
     }
@@ -296,7 +307,10 @@ static bool common_download_file_single(const std::string & url, const std::stri
     };
 
     common_load_model_from_url_headers headers;
+    bool head_request_ok = false;
+    bool should_download = !file_exists; // by default, we should download if the file does not exist
 
+    // get ETag to see if the remote file has changed
     {
         typedef size_t(*CURLOPT_HEADERFUNCTION_PTR)(char *, size_t, size_t, void *);
         auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
@@ -325,23 +339,28 @@ static bool common_download_file_single(const std::string & url, const std::stri
         curl_easy_setopt(curl.get(), CURLOPT_HEADERFUNCTION, static_cast<CURLOPT_HEADERFUNCTION_PTR>(header_callback));
         curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
 
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        // we only allow retrying once for HEAD requests
+        // this is for the use case of using running offline (no internet), retrying can be annoying
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), 1, 0, "HEAD");
         if (!was_perform_successful) {
-            return false;
+            head_request_ok = false;
         }
 
         long http_code = 0;
         curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-        if (http_code != 200) {
-            // HEAD not supported, we don't know if the file has changed
-            // force trigger downloading
-            force_download = true;
-            LOG_ERR("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+        if (http_code == 200) {
+            head_request_ok = true;
+        } else {
+            LOG_WRN("%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
+            head_request_ok = false;
         }
     }
 
-    bool should_download = !file_exists || force_download;
-    if (!should_download) {
+    // if head_request_ok is false, we don't have the etag or last-modified headers
+    // we leave should_download as-is, which is true if the file does not exist
+    if (head_request_ok) {
+        // check if ETag or Last-Modified headers are different
+        // if it is, we need to download the file again
         if (!etag.empty() && etag != headers.etag) {
             LOG_WRN("%s: ETag header is different (%s != %s): triggering a new download\n", __func__, etag.c_str(), headers.etag.c_str());
             should_download = true;
@@ -350,6 +369,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
             should_download = true;
         }
     }
+
     if (should_download) {
         std::string path_temporary = path + ".downloadInProgress";
         if (file_exists) {
@@ -403,7 +423,7 @@ static bool common_download_file_single(const std::string & url, const std::stri
         // start the download
         LOG_INF("%s: trying to download model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
             llama_download_hide_password_in_url(url).c_str(), path.c_str(), headers.etag.c_str(), headers.last_modified.c_str());
-        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS);
+        bool was_perform_successful = curl_perform_with_retry(url, curl.get(), CURL_MAX_RETRY, CURL_RETRY_DELAY_SECONDS, "GET");
         if (!was_perform_successful) {
             return false;
         }
@@ -424,13 +444,15 @@ static bool common_download_file_single(const std::string & url, const std::stri
             {"etag", headers.etag},
             {"lastModified", headers.last_modified}
         });
-        std::ofstream(metadata_path) << metadata.dump(4);
-        LOG_INF("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
+        write_file(metadata_path, metadata.dump(4));
+        LOG_DBG("%s: file metadata saved: %s\n", __func__, metadata_path.c_str());
 
         if (rename(path_temporary.c_str(), path.c_str()) != 0) {
             LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
             return false;
         }
+    } else {
+        LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
     }
 
     return true;
@@ -605,16 +627,37 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
     // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
     // User-Agent header is already set in common_remote_get_content, no need to set it here
 
+    // we use "=" to avoid clashing with other component, while still being allowed on windows
+    std::string cached_response_fname = "manifest=" + hf_repo + "=" + tag + ".json";
+    string_replace_all(cached_response_fname, "/", "_");
+    std::string cached_response_path = fs_get_cache_file(cached_response_fname);
+
     // 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());
+    long res_code = 0;
+    std::string res_str;
+    bool use_cache = false;
+    try {
+        auto res = common_remote_get_content(url, params);
+        res_code = res.first;
+        res_str = std::string(res.second.data(), res.second.size());
+    } catch (const std::exception & e) {
+        LOG_WRN("error: failed to get manifest: %s\n", e.what());
+        LOG_WRN("try reading from cache\n");
+        // try to read from cache
+        try {
+            res_str = read_file(cached_response_path);
+            res_code = 200;
+            use_cache = true;
+        } catch (const std::exception & e) {
+            throw std::runtime_error("error: failed to get manifest (check your internet connection)");
+        }
+    }
     std::string ggufFile;
     std::string mmprojFile;
 
-    if (res_code == 200) {
+    if (res_code == 200 || res_code == 304) {
         // extract ggufFile.rfilename in json, using regex
         {
             std::regex pattern("\"ggufFile\"[\\s\\S]*?\"rfilename\"\\s*:\\s*\"([^\"]+)\"");
@@ -631,6 +674,10 @@ static struct common_hf_file_res common_get_hf_file(const std::string & hf_repo_
                 mmprojFile = match[1].str();
             }
         }
+        if (!use_cache) {
+            // if not using cached response, update the cache file
+            write_file(cached_response_path, res_str);
+        }
     } else if (res_code == 401) {
         throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
     } else {
@@ -1142,6 +1189,9 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
         fprintf(stderr, "%s\n", ex.what());
         ctx_arg.params = params_org;
         return false;
+    } catch (std::exception & ex) {
+        fprintf(stderr, "%s\n", ex.what());
+        exit(1); // for other exceptions, we exit with status code 1
     }
 
     return true;
@@ -1442,13 +1492,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-f", "--file"}, "FNAME",
         "a file containing the prompt (default: none)",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
+            params.prompt = read_file(value);
             // store the external file name in params
             params.prompt_file = value;
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
             if (!params.prompt.empty() && params.prompt.back() == '\n') {
                 params.prompt.pop_back();
             }
@@ -1458,11 +1504,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"-sysf", "--system-prompt-file"}, "FNAME",
         "a file containing the system prompt (default: none)",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.system_prompt));
+            params.system_prompt = read_file(value);
             if (!params.system_prompt.empty() && params.system_prompt.back() == '\n') {
                 params.system_prompt.pop_back();
             }
@@ -1887,15 +1929,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--grammar-file"}, "FNAME",
         "file to read grammar from",
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(
-                std::istreambuf_iterator<char>(file),
-                std::istreambuf_iterator<char>(),
-                std::back_inserter(params.sampling.grammar)
-            );
+            params.sampling.grammar = read_file(value);
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -1905,6 +1939,23 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.sampling.grammar = json_schema_to_grammar(json::parse(value));
         }
     ).set_sparam());
+    add_opt(common_arg(
+        {"-jf", "--json-schema-file"}, "FILE",
+        "File containing a JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object\nFor schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead",
+        [](common_params & params, const std::string & value) {
+            std::ifstream file(value);
+            if (!file) {
+                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
+            }
+            std::string schema;
+            std::copy(
+                std::istreambuf_iterator<char>(file),
+                std::istreambuf_iterator<char>(),
+                std::back_inserter(schema)
+            );
+            params.sampling.grammar = json_schema_to_grammar(json::parse(schema));
+        }
+    ).set_sparam());
     add_opt(common_arg(
         {"--pooling"}, "{none,mean,cls,last,rank}",
         "pooling type for embeddings, use model default if unspecified",
@@ -2815,14 +2866,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             "list of built-in templates:\n%s", list_builtin_chat_templates().c_str()
         ),
         [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::copy(
-                std::istreambuf_iterator<char>(file),
-                std::istreambuf_iterator<char>(),
-                std::back_inserter(params.chat_template));
+            params.chat_template = read_file(value);
         }
     ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
     add_opt(common_arg(

convert_hf_to_gguf.py

@@ -16,6 +16,7 @@ from pathlib import Path
 from hashlib import sha256
 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast
 from itertools import chain
+from transformers import AutoConfig
 
 import math
 import numpy as np
@@ -66,8 +67,6 @@ class ModelBase:
     part_names: list[str]
     is_safetensors: bool
     hparams: dict[str, Any]
-    block_count: int
-    tensor_map: gguf.TensorNameMap
     tensor_names: set[str] | None
     gguf_writer: gguf.GGUFWriter
     model_name: str | None
@@ -78,6 +77,10 @@ class ModelBase:
     # subclasses should define this!
     model_arch: gguf.MODEL_ARCH
 
+    # subclasses should initialize this!
+    block_count: int
+    tensor_map: gguf.TensorNameMap
+
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, *, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
@@ -113,8 +116,6 @@ class ModelBase:
             if not self.is_safetensors:
                 self.part_names = ModelBase.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
         self.hparams = ModelBase.load_hparams(self.dir_model) if hparams is None else hparams
-        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
-        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
         self.metadata_override = metadata_override
         self.model_name = model_name
@@ -417,15 +418,13 @@ class ModelBase:
 
     @staticmethod
     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
+        try:
+            return AutoConfig.from_pretrained(dir_model).to_dict()
+        except Exception as e:
+            logger.warning(f"Failed to load model config from {dir_model}: {e}")
+            logger.warning("Trying to load config.json instead")
+            with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+                return json.load(f)
 
     @classmethod
     def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
@@ -454,6 +453,23 @@ class ModelBase:
 
 
 class TextModel(ModelBase):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        if "text_config" in self.hparams:
+            # move the text_config to the root level
+            self.hparams = {**self.hparams, **self.hparams["text_config"]}
+
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    @classmethod
+    def __init_subclass__(cls):
+        # can't use an abstract property, because overriding it without type errors
+        # would require using decorated functions instead of simply defining the property
+        if "model_arch" not in cls.__dict__:
+            raise TypeError(f"Missing property 'model_arch' for {cls.__name__!r}")
+
     def set_vocab(self):
         self._set_vocab_gpt2()
 
@@ -1070,9 +1086,9 @@ class VisionModel(ModelBase):
         if self.model_arch != gguf.MODEL_ARCH.CLIP_VISION:
             raise TypeError("VisionModel must be subclassed with model_arch = gguf.MODEL_ARCH.CLIP_VISION")
 
-        # small hack to correct the number of layers
-        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, 128)
-        self.n_embd_text = self.find_hparam(["hidden_size", "n_embd"])
+        # get n_embd of the text model
+        text_config = {**self.hparams, **self.hparams["text_config"]}
+        self.n_embd_text = text_config.get("hidden_size", text_config.get("n_embd", 0))
         assert self.n_embd_text > 0, "n_embd not found in hparams"
 
         if "vision_config" not in self.hparams:
@@ -1081,6 +1097,9 @@ class VisionModel(ModelBase):
         self.global_config = self.hparams
         self.hparams = self.hparams["vision_config"]
 
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers", "depth"])
+        self.tensor_map = gguf.get_tensor_name_map(gguf.MODEL_ARCH.CLIP_VISION, self.block_count)
+
         # load preprocessor config
         with open(self.dir_model / "preprocessor_config.json", "r", encoding="utf-8") as f:
             self.preprocessor_config = json.load(f)
@@ -1098,12 +1117,12 @@ class VisionModel(ModelBase):
         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_block_count(self.block_count)
         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"])
+        self.gguf_writer.add_vision_image_std(self.preprocessor_config["image_std"])
 
     def write_vocab(self):
         raise ValueError("VisionModel does not support vocab writing")
@@ -1719,23 +1738,12 @@ class StableLMModel(TextModel):
     "LlamaForCausalLM",
     "MistralForCausalLM",
     "MixtralForCausalLM",
-    "Idefics3ForConditionalGeneration",
-    "SmolVLMForConditionalGeneration",
+    "VLlama3ForCausalLM",
     "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()
@@ -1891,31 +1899,50 @@ class LlamaModel(TextModel):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
-@ModelBase.register("LlavaForConditionalGeneration")
+@ModelBase.register(
+    "LlavaForConditionalGeneration", # pixtral
+    "Mistral3ForConditionalGeneration", # mistral small 3.1
+)
 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)
+            # layer_norm_eps is not in config.json, it is hard-coded in modeling_pixtral.py
             self.hparams["layer_norm_eps"] = self.hparams.get("layer_norm_eps", 1e-5)
-            self.img_break_tok_id = 12 # see tokenizer_config.json
+            self.img_break_tok_id = self.get_token_id("[IMG_BREAK]")
+            logger.info(f"Image break token id: {self.img_break_tok_id}")
         else:
             raise ValueError(f"Unsupported model type: {self.hparams['model_type']}")
 
+    def get_token_id(self, token: str) -> int:
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+            added_tokens_decoder = json.load(f)['added_tokens_decoder']
+            for id_, token_data in added_tokens_decoder.items():
+                if token_data["content"] == token:
+                    return int(id_)
+        raise ValueError(f"Token '{token}' not found in tokenizer config.")
+
     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)
+
+            # hidden_act
+            if hparams["hidden_act"] == "silu":
+                self.gguf_writer.add_vision_use_silu(True)
+            elif hparams["hidden_act"] == "gelu":
+                self.gguf_writer.add_vision_use_gelu(True)
+            else:
+                raise ValueError(f"Unsupported hidden_act: {hparams['hidden_act']}")
+
+            # spatial_merge_size
+            if "spatial_merge_size" in self.global_config:
+                self.gguf_writer.add_vision_spatial_merge_size(self.global_config["spatial_merge_size"])
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
@@ -1944,13 +1971,12 @@ class LlavaVisionModel(VisionModel):
 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":
+            # fix for SmolVLM2, missing some keys in config.json
+            # default values are taken from transformers code
             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()
@@ -3505,6 +3531,8 @@ class RobertaModel(BertModel):
 
 @ModelBase.register("NomicBertModel")
 class NomicBertModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, **kwargs: Any):
         hparams = kwargs.pop("hparams", None)
         if hparams is None:
@@ -5849,6 +5877,19 @@ def split_str_to_n_bytes(split_str: str) -> int:
     return n
 
 
+def get_model_architecture(dir_model: Path, model_type: ModelType, hparams: Any = None) -> str:
+    hparams = ModelBase.load_hparams(dir_model) if hparams is None else hparams
+    text_config = hparams.get("text_config", {})
+    vision_config = hparams.get("vision_config", {})
+    arch = hparams["architectures"][0]
+    # if "architectures" is found in the sub-config, use that instead
+    if model_type == ModelType.TEXT and text_config.get("architectures") is not None:
+        arch = text_config["architectures"][0]
+    elif model_type == ModelType.VISION and vision_config.get("architectures") is not None:
+        arch = vision_config["architectures"][0]
+    return arch
+
+
 def main() -> None:
     args = parse_args()
 
@@ -5901,16 +5942,15 @@ def main() -> None:
 
     logger.info(f"Loading model: {dir_model.name}")
 
-    hparams = ModelBase.load_hparams(dir_model)
-
     if args.mmproj:
         if "mmproj" not in fname_out.name:
             fname_out = ModelBase.add_prefix_to_filename(fname_out, "mmproj-")
 
     with torch.inference_mode():
         output_type = ftype_map[args.outtype]
-        model_architecture = hparams["architectures"][0]
         model_type = ModelType.VISION if args.mmproj else ModelType.TEXT
+        model_architecture = get_model_architecture(dir_model, model_type)
+        logger.info(f"Model architecture: {model_architecture}")
         try:
             model_class = ModelBase.from_model_architecture(model_architecture, model_type=model_type)
         except NotImplementedError:

examples/llava/README.md

@@ -34,6 +34,9 @@ llama-mtmd-cli -hf ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
 
 # Pixtral 12B
 llama-mtmd-cli -hf ggml-org/pixtral-12b-GGUF
+
+# Mistral Small 3.1 24B (IQ2_M quantization)
+llama-mtmd-cli -hf ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF --chat-template mistral-v7
 ```
 
 ## How it works and what is `mmproj`?
@@ -73,3 +76,4 @@ For the following models, you can use `convert_hf_to_gguf.py`with `--mmproj` fla
 - 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
+- [Mistral Small 3.1 24B](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503)

examples/llava/clip-impl.h

@@ -2,8 +2,6 @@
 #include "gguf.h"
 #include "clip.h"
 
-#include "clip.h"
-
 #include <climits>
 #include <cstdarg>
 #include <string>
@@ -33,6 +31,7 @@
 #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_SPATIAL_MERGE_SIZE  "clip.vision.spatial_merge_size"
 
 #define KEY_USE_GLU_MLP         "clip.use_glu_mlp"  // for qwen2.5vl
 #define KEY_USE_RMS_NORM        "clip.use_rms_norm" // for qwen2.5vl
@@ -70,9 +69,11 @@
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
 #define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
+#define TN_MM_INP_NORM     "mm.input_norm.weight"
 #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_MM_PATCH_MERGER "mm.patch_merger.weight"     // mistral small 3.1
 #define TN_TOK_IMG_BREAK   "v.token_embd.img_break"     // pixtral
 
 // mimicpmv

examples/llava/clip.cpp

@@ -172,6 +172,7 @@ struct clip_hparams {
     std::unordered_set<int32_t> vision_feature_layer;
     int32_t attn_window_size = 0;
     int32_t n_wa_pattern = 0;
+    int32_t spatial_merge_size = 0;
 };
 
 struct clip_layer {
@@ -232,6 +233,7 @@ struct clip_vision_model {
     struct ggml_tensor * projection;
 
     // LLaVA projection
+    struct ggml_tensor * mm_input_norm_w = nullptr;
     struct ggml_tensor * mm_0_w = nullptr;
     struct ggml_tensor * mm_0_b = nullptr;
     struct ggml_tensor * mm_2_w = nullptr;
@@ -311,6 +313,7 @@ struct clip_vision_model {
 
     // pixtral
     struct ggml_tensor * token_embd_img_break = nullptr;
+    struct ggml_tensor * mm_patch_merger_w = nullptr;
 };
 
 struct clip_ctx {
@@ -637,6 +640,7 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
     const int d_head      = hidden_size / n_head;
     const int n_layer     = hparams.n_layer;
     const float eps       = hparams.eps;
+    const int n_merge     = hparams.spatial_merge_size;
 
     struct ggml_init_params params = {
         /*.mem_size   =*/ ctx->buf_compute_meta.size(),
@@ -721,7 +725,13 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
         {
             ggml_tensor * gate_proj = ggml_mul_mat(ctx0, model.layers[il].ff_gate_w, cur);
             ggml_tensor * up_proj   = ggml_mul_mat(ctx0, model.layers[il].ff_up_w,   cur);
-            gate_proj = ggml_silu(ctx0, gate_proj); // pixtral uses silu
+            if (ctx->use_silu) {
+                gate_proj = ggml_silu(ctx0, gate_proj);
+            } else if (ctx->use_gelu) {
+                gate_proj = ggml_gelu(ctx0, gate_proj);
+            } else {
+                GGML_ABORT("Pixtral: Unsupported activation");
+            }
             cur = ggml_mul(ctx0, up_proj, gate_proj);
             cur = ggml_mul_mat(ctx0, model.layers[il].ff_down_w, cur);
         }
@@ -732,14 +742,42 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
         embeddings = cur;
     }
 
-    // LlavaMultiModalProjector (with GELU activation)
+    // mistral small 3.1 patch merger
+    // ref: https://github.com/huggingface/transformers/blob/7a3e208892c06a5e278144eaf38c8599a42f53e7/src/transformers/models/mistral3/modeling_mistral3.py#L67
+    if (model.mm_patch_merger_w) {
+        GGML_ASSERT(hparams.spatial_merge_size > 0);
+
+        ggml_tensor * cur = embeddings;
+        cur = ggml_mul(ctx0, ggml_rms_norm(ctx0, cur, eps), model.mm_input_norm_w);
+
+        // reshape image tokens to 2D grid
+        cur = ggml_reshape_3d(ctx0, cur, hidden_size, n_patches_x, n_patches_y);
+        cur = ggml_permute(ctx0, cur, 2, 0, 1, 3); // [x, y, hidden_size]
+        cur = ggml_cont(ctx0, cur);
+
+        // torch.nn.functional.unfold is just an im2col under the hood
+        // we just need a dummy kernel to make it work
+        ggml_tensor * kernel = ggml_view_3d(ctx0, cur, n_merge, n_merge, cur->ne[2], 0, 0, 0);
+        cur = ggml_im2col(ctx0, kernel, cur, n_merge, n_merge, 0, 0, 1, 1, true, inp->type);
+
+        // project to hidden_size
+        cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], cur->ne[1] * cur->ne[2]);
+        cur = ggml_mul_mat(ctx0, model.mm_patch_merger_w, cur);
+        embeddings = cur;
+    }
+
+    // LlavaMultiModalProjector (always using GELU activation)
     {
         embeddings = ggml_mul_mat(ctx0, model.mm_1_w, embeddings);
-        embeddings = ggml_add(ctx0, embeddings, model.mm_1_b);
+        if (model.mm_1_b) {
+            embeddings = ggml_add(ctx0, embeddings, model.mm_1_b);
+        }
 
         embeddings = ggml_gelu(ctx0, embeddings);
         embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
-        embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+        if (model.mm_2_b) {
+            embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
+        }
     }
 
     // arrangement of the [IMG_BREAK] token
@@ -749,11 +787,14 @@ static ggml_cgraph * clip_image_build_graph_pixtral(clip_ctx * ctx, const clip_i
         // and then concatenate the [IMG_BREAK] token to the end of each row, aka n_patches_per_row dimension
         // after the concatenation, we have a tensor with shape [hidden_size, n_patches_per_row + 1, n_rows]
 
+        const int p_y             = n_merge > 0 ? n_patches_y / n_merge : n_patches_y;
+        const int p_x             = n_merge > 0 ? n_patches_x / n_merge : n_patches_x;
+        const int p_total         = p_x * p_y;
         const int n_embd_text     = embeddings->ne[0];
-        const int n_tokens_output = num_patches + n_patches_y - 1; // one [IMG_BREAK] per row, except the last row
+        const int n_tokens_output = p_total + p_y - 1; // one [IMG_BREAK] per row, except the last row
 
-        ggml_tensor * cur = ggml_reshape_3d(ctx0, embeddings, n_embd_text, n_patches_x, n_patches_y);
-        ggml_tensor * tok = ggml_new_tensor_3d(ctx0, embeddings->type, n_embd_text, 1, n_patches_y);
+        ggml_tensor * cur = ggml_reshape_3d(ctx0, embeddings, n_embd_text, p_x, p_y);
+        ggml_tensor * tok = ggml_new_tensor_3d(ctx0, embeddings->type, n_embd_text, 1, p_y);
         tok = ggml_scale(ctx0, tok, 0.0); // clear the tensor
         tok = ggml_add(ctx0, tok, model.token_embd_img_break);
         cur = ggml_concat(ctx0, cur, tok, 1);
@@ -1734,6 +1775,7 @@ struct clip_model_loader {
                 case PROJECTOR_TYPE_PIXTRAL:
                     {
                         hparams.rope_theta = 10000.0f;
+                        get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.spatial_merge_size, false);
                     } break;
                 case PROJECTOR_TYPE_QWEN25VL:
                     {
@@ -1957,11 +1999,14 @@ struct clip_model_loader {
             case PROJECTOR_TYPE_PIXTRAL:
                 {
                     vision_model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
-                    vision_model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
+                    vision_model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"), false);
                     vision_model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
-                    vision_model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
+                    vision_model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"), false);
                     // [IMG_BREAK] token embedding
                     vision_model.token_embd_img_break = get_tensor(TN_TOK_IMG_BREAK);
+                    // for mistral small 3.1
+                    vision_model.mm_input_norm_w   = get_tensor(TN_MM_INP_NORM,     false);
+                    vision_model.mm_patch_merger_w = get_tensor(TN_MM_PATCH_MERGER, false);
                 } break;
             default:
                 GGML_ASSERT(false && "unknown projector type");
@@ -2926,8 +2971,9 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
     } else if (ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) {
         n_patches /= ctx->vision_model.hparams.proj_scale_factor;
     } else if (ctx->proj_type == PROJECTOR_TYPE_PIXTRAL) {
-        int n_patches_x = img->nx / params.patch_size;
-        int n_patches_y = img->ny / params.patch_size;
+        int n_merge = ctx->vision_model.hparams.spatial_merge_size;
+        int n_patches_x = img->nx / params.patch_size / (n_merge > 0 ? n_merge : 1);
+        int n_patches_y = img->ny / params.patch_size / (n_merge > 0 ? n_merge : 1);
         n_patches = n_patches_y*n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
     }
 
@@ -3484,7 +3530,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
             return ctx->vision_model.mm_model_peg_0_b->ne[0];
         case PROJECTOR_TYPE_MLP:
         case PROJECTOR_TYPE_PIXTRAL:
-            return ctx->vision_model.mm_2_b->ne[0];
+            return ctx->vision_model.mm_2_w->ne[1];
         case PROJECTOR_TYPE_MLP_NORM:
             return ctx->vision_model.mm_3_b->ne[0];
         case PROJECTOR_TYPE_MINICPMV:

examples/llava/mtmd-cli.cpp

@@ -94,6 +94,7 @@ struct mtmd_cli_context {
             LOG_ERR("Model does not have chat template.\n");
             LOG_ERR("  For old llava models, you may need to use '--chat-template vicuna'\n");
             LOG_ERR("  For MobileVLM models, use '--chat-template deepseek'\n");
+            LOG_ERR("  For Mistral Small 3.1, use '--chat-template mistral-v7'\n");
             exit(1);
         }
 

examples/llava/tests.sh

@@ -59,6 +59,7 @@ add_test "llama-mtmd-cli"  "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
 
 # to test the big models, run: ./tests.sh big
 add_test_big "llama-mtmd-cli" "ggml-org/pixtral-12b-GGUF:Q4_K_M"
+add_test_big "llama-mtmd-cli" "ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF" "mistral-v7"
 
 # these models always give the wrong answer, not sure why
 # add_test "llama-mtmd-cli"  "ggml-org/SmolVLM-Instruct-GGUF:Q4_K_M"

ggml/include/ggml-cpp.h

@@ -24,7 +24,7 @@ typedef std::unique_ptr<gguf_context, gguf_context_deleter> gguf_context_ptr;
 
 struct ggml_gallocr_deleter { void operator()(ggml_gallocr_t galloc) { ggml_gallocr_free(galloc); } };
 
-typedef std::unique_ptr<ggml_gallocr_t, ggml_gallocr_deleter> ggml_gallocr_ptr;
+typedef std::unique_ptr<ggml_gallocr, ggml_gallocr_deleter> ggml_gallocr_ptr;
 
 // ggml-backend
 

ggml/src/ggml-cpu/CMakeLists.txt

@@ -352,10 +352,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         # TODO: Separation to determine activation of VX/VXE/VXE2
         if (${S390X_M} MATCHES "8561|8562")
             message(STATUS "z15 target")
-            list(APPEND ARCH_FLAGS -march=z15 -mtune=z15)
+            list(APPEND ARCH_FLAGS -march=z15)
         elseif (${S390X_M} MATCHES "3931")
             message(STATUS "z16 target")
-            list(APPEND ARCH_FLAGS -march=z16 -mtune=z16)
+            list(APPEND ARCH_FLAGS -march=z16)
+        elseif (${S390X_M} MATCHES "9175|9176")
+            # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
+            message(STATUS "z17 target")
+            list(APPEND ARCH_FLAGS -march=z17)
         else()
             message(STATUS "Unknown target")
             message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")

ggml/src/ggml-cpu/simd-mappings.h

@@ -341,7 +341,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
 #define GGML_F32_EPR  4
 
 #define GGML_F32x4              vector float
-#define GGML_F32x4_ZERO         0.0f
+#define GGML_F32x4_ZERO         {0.0f}
 #define GGML_F32x4_SET1         vec_splats
 #define GGML_F32x4_LOAD(p)      vec_xl(0, p)
 #define GGML_F32x4_STORE(p, r)  vec_xst(r, 0, p)

ggml/src/ggml-cuda/cpy.cu

@@ -592,6 +592,8 @@ void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, gg
         dest_ptrs_d = ctx.cuda_graph->dest_ptrs_d;
         graph_cpynode_index = ctx.cuda_graph->graph_cpynode_index;
     }
+#else
+    GGML_UNUSED(disable_indirection_for_this_node);
 #endif
     if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
         GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));

ggml/src/ggml-cuda/getrows.cu

@@ -33,8 +33,8 @@ static __global__ void k_get_rows(
     dfloat2 v;
     dequantize_kernel(src0_row, ib, iqs, v);
 
-    dst_row[iybs + iqs + 0]        = v.x;
-    dst_row[iybs + iqs + y_offset] = v.y;
+    dst_row[iybs + iqs + 0]        = float(v.x);
+    dst_row[iybs + iqs + y_offset] = float(v.y);
 }
 
 template<typename src0_t, typename dst_t>
@@ -60,7 +60,7 @@ static __global__ void k_get_rows_float(
     dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
     const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
 
-    dst_row[i00] = src0_row[i00];
+    dst_row[i00] = float(src0_row[i00]);
 }
 
 template<typename grad_t, typename dst_t>
@@ -86,120 +86,159 @@ static __global__ void k_get_rows_back_float(
     dst[dst_row*ncols + col] = sum;
 }
 
-template<int qk, int qr, dequantize_kernel_t dq>
-static void get_rows_cuda(
-        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-        const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
+template<int qk, int qr, dequantize_kernel_t dq, typename dst_t>
+static void get_rows_cuda_q(
+        const void * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
     const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
     const dim3 block_nums(block_num_x, ne10, ne11*ne12);
 
     // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
+    // const size_t s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);
 
-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);
 
     GGML_ASSERT(ne00 % 2 == 0);
 
     k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
-        src0_dd, src1_dd, dst_dd,
+        src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
         /*ne10, ne11,*/ ne12, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);
-
-    GGML_UNUSED(dst);
 }
 
-template<typename src0_t>
+template<typename src0_t, typename dst_t>
 static void get_rows_cuda_float(
-        const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-        const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    GGML_ASSERT(ne13 == 1);
-
+        const src0_t * src0_d, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
     const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
     const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
     const dim3 block_nums(block_num_x, ne10, ne11*ne12);
 
     // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
+    // const size_t s0 = nb0 / sizeof(dst_t);
+    const size_t s1 = nb1 / sizeof(dst_t);
+    const size_t s2 = nb2 / sizeof(dst_t);
+    const size_t s3 = nb3 / sizeof(dst_t);
 
-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
+    const size_t s10 = nb10 / sizeof(int32_t);
+    const size_t s11 = nb11 / sizeof(int32_t);
+    const size_t s12 = nb12 / sizeof(int32_t);
+    // const size_t s13 = nb13 / sizeof(int32_t);
 
     k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
-        src0_dd, src1_dd, dst_dd,
+        src0_d, src1_d, dst_d,
         ne00, /*ne01, ne02, ne03,*/
         /*ne10, ne11,*/ ne12, /*ne13,*/
         /* s0,*/ s1, s2, s3,
         /* nb00,*/ nb01, nb02, nb03,
         s10, s11, s12/*, s13*/);
+}
 
-    GGML_UNUSED(dst);
+template <typename dst_t>
+static void ggml_cuda_get_rows_switch_src0_type(
+        const void * src0_d, const ggml_type src0_type, const int32_t * src1_d, dst_t * dst_d,
+        const int64_t ne00, const size_t nb01, const size_t nb02, const size_t nb03,
+        const int64_t ne10, const int64_t ne11, const int64_t ne12, const size_t nb10, const size_t nb11, const size_t nb12,
+        const size_t nb1, const size_t nb2, const size_t nb3,
+        cudaStream_t stream) {
+    switch (src0_type) {
+        case GGML_TYPE_F16:
+            get_rows_cuda_float((const half *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F32:
+            get_rows_cuda_float((const float *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            get_rows_cuda_q<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            get_rows_cuda_q<QK4_1, QR4_1, dequantize_q4_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            get_rows_cuda_q<QK5_0, QR5_0, dequantize_q5_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            get_rows_cuda_q<QK5_1, QR5_1, dequantize_q5_1>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            get_rows_cuda_q<QK8_0, QR8_0, dequantize_q8_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            // TODO: k-quants
+            GGML_ABORT("%s: unsupported src0 type: %s\n", __func__, ggml_type_name(src0_type));
+            break;
+    }
+}
+
+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream) {
+    switch (dst_type) {
+        case GGML_TYPE_F32:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (float *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_F16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (half *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        case GGML_TYPE_BF16:
+            ggml_cuda_get_rows_switch_src0_type(src0_d, src0_type, src1_d, (nv_bfloat16 *) dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
+        default:
+            GGML_ABORT("%s: unsupported dst type: %s\n", __func__, ggml_type_name(dst_type));
+            break;
+    }
 }
 
 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
 
-    const void    * src0_d = (const void    *) src0->data;
-    const int32_t * src1_d = (const int32_t *) src1->data;
-    float         * dst_d  = (float         *) dst->data;
-
     cudaStream_t stream = ctx.stream();
 
+    GGML_TENSOR_BINARY_OP_LOCALS
+
     GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(ne13 == 1);
 
     GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
     GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
     GGML_ASSERT(dst->nb[0]  == ggml_type_size(dst->type));
 
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            get_rows_cuda_float(src0, src1, dst, (const half *) src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_F32:
-            get_rows_cuda_float(src0, src1, dst, (const float *) src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_0:
-            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_d, dst_d, stream);
-            break;
-        default:
-            // TODO: k-quants
-            GGML_ABORT("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
-            break;
-    }
+    get_rows_cuda(src0->data, src0->type, (const int32_t *) src1->data, dst->data, dst->type,
+        ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
 }
 
 void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/getrows.cuh

@@ -3,6 +3,13 @@
 #define CUDA_GET_ROWS_BLOCK_SIZE 256
 #define CUDA_GET_ROWS_BACK_BLOCK_SIZE 256
 
+void get_rows_cuda(
+        const void * src0_d, ggml_type src0_type, const int32_t * src1_d, void * dst_d, ggml_type dst_type,
+        int64_t ne00, size_t nb01, size_t nb02, size_t nb03,
+        int64_t ne10, int64_t ne11, int64_t ne12, size_t nb10, size_t nb11, size_t nb12,
+        size_t nb1, size_t nb2, size_t nb3,
+        cudaStream_t stream);
+
 void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
 
 void ggml_cuda_op_get_rows_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

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

@@ -1551,7 +1551,7 @@ static void ggml_cuda_op_mul_mat(
 
             if (src1_on_device && src1_is_contiguous) {
                 quantize_src1(
-                    dev[id].src1_ddf, dev[id].src1_ddq, src0->type, ne10,
+                    dev[id].src1_ddf, nullptr, 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());
@@ -1649,7 +1649,7 @@ static void ggml_cuda_op_mul_mat(
 
                 if (quantize_src1 && !src1_is_contiguous) {
                     quantize_src1(
-                        src1_ddf_i, src1_ddq_i, src0->type, ne10, ne10, ne11*ne10, ne12*ne11*ne10,
+                        src1_ddf_i, nullptr, src1_ddq_i, src0->type, ne10, ne10, ne11*ne10, ne12*ne11*ne10,
                         src1_padded_col_size, src1_ncols, 1, 1, stream);
                     CUDA_CHECK(cudaGetLastError());
                 }
@@ -1949,6 +1949,8 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         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 && use_mul_mat_q) {
+        ggml_cuda_mul_mat_q(ctx, src0, src1, nullptr, 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) {
         // general KQ + KQV multi-batch without FlashAttention
@@ -1964,183 +1966,145 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     }
 }
 
-struct mmid_row_mapping {
-    int32_t i1;
-    int32_t i2;
-};
-
-static __global__ void k_copy_src1_to_contiguous(const char * __restrict__ src1_original, char * __restrict__ src1_contiguous,
-                                                 int * __restrict__ cur_src1_row, mmid_row_mapping * __restrict__ row_mapping,
-                                                 const char * __restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
-                                                 int64_t ne11, int64_t ne10,
-                                                 size_t nb11, size_t nb12) {
-    int32_t iid1 = blockIdx.x;
-    int32_t id = blockIdx.y;
-
-    const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);
-
-    if (row_id_i != i02) {
-        return;
-    }
-
-    const int64_t i11 = id % ne11;
-    const int64_t i12 = iid1;
-
-    __shared__ int src1_row;
-    if (threadIdx.x == 0) {
-        src1_row = atomicAdd(cur_src1_row, 1);
-        row_mapping[src1_row] = {id, iid1};
-    }
-    __syncthreads();
-
-    const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
-    float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);
-
-    for (int i = threadIdx.x; i < ne10; i += blockDim.x) {
-        src1_row_contiguous[i] = src1_row_original[i];
-    }
-}
-
-static __global__ void k_copy_dst_from_contiguous(char * __restrict__ dst_original, const char * __restrict__ dst_contiguous,
-                                                  const mmid_row_mapping * __restrict__ row_mapping,
-                                                  int64_t ne0,
-                                                  size_t nb1, size_t nb2) {
-    int32_t i = blockIdx.x;
-
-    const int32_t i1 = row_mapping[i].i1;
-    const int32_t i2 = row_mapping[i].i2;
-
-    const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
-    float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);
-
-    for (int j = threadIdx.x; j < ne0; j += blockDim.x) {
-        dst_row_original[j] = dst_row_contiguous[j];
-    }
-}
-
 static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
     const ggml_tensor * ids  = dst->src[2];
 
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
+
     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;
-    }
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
 
-    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft) && "mul_mat_id does not support split buffers");
+    if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        if (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;
+        }
+
+        if (ggml_cuda_should_use_mmq(src0->type, cc, ne12)) {
+            ggml_cuda_mul_mat_q(ctx, src0, src1, ids, dst);
+            return;
+        }
+    }
 
     cudaStream_t stream = ctx.stream();
 
-    const int64_t n_as = ne02;
-    const int64_t n_ids = ids->ne[0];
+    GGML_ASSERT(nb12 % nb11 == 0);
+    GGML_ASSERT(nb2  % nb1  == 0);
+
+    const ggml_type type_src1_sorted = (src0->type == GGML_TYPE_F16 && !fast_fp16_hardware_available(cc))
+        || ggml_is_quantized(src0->type) ? GGML_TYPE_F32 : src0->type;
+    const ggml_type type_dst_sorted  = GGML_TYPE_F32;
+    const size_t ts_src1_sorted = ggml_type_size(type_src1_sorted);
+    const size_t ts_dst_sorted  = ggml_type_size(type_dst_sorted);
+
+    const int64_t n_expert_used = ids->ne[0];
+    const int64_t ne_get_rows = ne12 * n_expert_used;
+
+    std::vector<int32_t> ids_to_sorted_host;
+    ids_to_sorted_host.reserve(2*ne_get_rows);
+    std::vector<int32_t> ids_from_sorted_host(ne_get_rows);
+
+    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool(), 2*ne_get_rows);
+
+    std::vector<int32_t> tokens_per_expert(ne02);
+
+    ggml_cuda_pool_alloc<char> src1_sorted(ctx.pool(), ne12*n_expert_used*ne10*ts_src1_sorted);
+    ggml_cuda_pool_alloc<char>  dst_sorted(ctx.pool(), ne2 *n_expert_used* ne0*ts_dst_sorted);
 
     std::vector<char> ids_host(ggml_nbytes(ids));
-    const char * ids_dev = (const char *) ids->data;
-    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
     CUDA_CHECK(cudaStreamSynchronize(stream));
 
-    ggml_tensor src0_row = *src0;
-    ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row  = *dst;
-
-    char * src0_original = (char *) src0->data;
-    char * src1_original = (char *) src1->data;
-    char * dst_original  = (char *)  dst->data;
-
-    src0_row.ne[2] = 1;
-    src0_row.ne[3] = 1;
-    src0_row.nb[3] = nb02;
-
-    src1_row.ne[1] = 1;
-    src1_row.ne[2] = 1;
-    src1_row.ne[3] = 1;
-    src1_row.nb[2] = nb11;
-    src1_row.nb[3] = nb11;
-
-    dst_row.ne[1] = 1;
-    dst_row.ne[2] = 1;
-    dst_row.ne[3] = 1;
-    dst_row.nb[2] = nb1;
-    dst_row.nb[3] = nb1;
-
-    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 row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
-
-                GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
-
-                if (row_id_i != i02) {
-                    continue;
+    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
+        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
+            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
+                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
+                assert(expert_to_use >= 0 && expert_to_use < ne02);
+                if (expert_to_use == i02) {
+                    ids_from_sorted_host[i12*n_expert_used + iex] = ids_to_sorted_host.size();
+                    ids_to_sorted_host.push_back(i12*ne11 + iex % ne11);
+                    tokens_per_expert[i02]++;
+                    break;
                 }
-
-                num_src1_rows++;
             }
         }
+    }
+    GGML_ASSERT(ids_to_sorted_host.size() == size_t(ne_get_rows));
 
-        if (num_src1_rows == 0) {
+    ids_to_sorted_host.insert(ids_to_sorted_host.end(), ids_from_sorted_host.begin(), ids_from_sorted_host.end());
+
+    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_to_sorted_host.data(), 2*ne_get_rows*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    const int32_t * ids_to_sorted   = ids_buf_dev.ptr + 0*ne_get_rows;
+    const int32_t * ids_from_sorted = ids_buf_dev.ptr + 1*ne_get_rows;
+
+    get_rows_cuda(src1->data, src1->type, ids_to_sorted, src1_sorted.ptr, type_src1_sorted,
+        ne10, nb11, nb12, nb13,
+        ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t),
+        ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, ne_get_rows*ne10*ts_src1_sorted, stream);
+    CUDA_CHECK(cudaGetLastError());
+
+    char * src1_data_cur = (char *) src1_sorted.ptr;
+    char *  dst_data_cur = (char *)  dst_sorted.ptr;
+    for (int64_t i02 = 0; i02 < ne02; ++i02) {
+        if (tokens_per_expert[i02] == 0) {
             continue;
         }
 
-        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_tensor src0_slice = *src0;
+        src0_slice.ne[2] = 1;
+        src0_slice.nb[3] = src0_slice.nb[2];
+        src0_slice.data  = (char *) src0->data + i02*nb02;
 
-        {
-            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_tensor src1_slice;
+        memset(&src1_slice, 0, sizeof(src1_slice));
+        src1_slice.buffer = src1->buffer;
+        src1_slice.type   = type_src1_sorted;
+        src1_slice.ne[0]  = ne10;
+        src1_slice.ne[1]  = tokens_per_expert[i02];
+        src1_slice.ne[2]  = 1;
+        src1_slice.ne[3]  = 1;
+        src1_slice.nb[0]  = ts_src1_sorted;
+        src1_slice.nb[1]  = src1_slice.ne[0] * src1_slice.nb[0];
+        src1_slice.nb[2]  = src1_slice.ne[1] * src1_slice.nb[1];
+        src1_slice.nb[3]  = src1_slice.ne[2] * src1_slice.nb[2];
+        src1_slice.data   = src1_data_cur;
 
-        src0_row.data = src0_original + i02*nb02;
+        ggml_tensor dst_slice;
+        memset(&dst_slice, 0, sizeof(dst_slice));
+        dst_slice.buffer = dst->buffer;
+        dst_slice.type   = type_dst_sorted;
+        dst_slice.ne[0]  = ne0;
+        dst_slice.ne[1]  = tokens_per_expert[i02];
+        dst_slice.ne[2]  = 1;
+        dst_slice.ne[3]  = 1;
+        dst_slice.nb[0]  = ts_dst_sorted;
+        dst_slice.nb[1]  = dst_slice.ne[0] * dst_slice.nb[0];
+        dst_slice.nb[2]  = dst_slice.ne[1] * dst_slice.nb[1];
+        dst_slice.nb[3]  = dst_slice.ne[2] * dst_slice.nb[2];
+        dst_slice.data   = dst_data_cur;
 
-        GGML_ASSERT(nb11 == sizeof(float)*ne10);
-        GGML_ASSERT(nb1 == sizeof(float)*ne0);
+        ggml_cuda_mul_mat(ctx, &src0_slice, &src1_slice, &dst_slice);
+        CUDA_CHECK(cudaGetLastError());
 
-        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());
-        }
+        src1_data_cur += src1_slice.nb[2];
+        dst_data_cur  +=  dst_slice.nb[2];
     }
+
+    get_rows_cuda(dst_sorted.ptr, type_dst_sorted, ids_from_sorted, dst->data, dst->type,
+        ne0, ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted, ne_get_rows*ne0*ts_dst_sorted,
+        ne_get_rows, 1, 1, sizeof(int32_t), ne_get_rows*sizeof(int32_t), ne_get_rows*sizeof(int32_t),
+        nb1, nb2, nb3, stream);
 }
 
 static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) {

ggml/src/ggml-cuda/mmq.cu

@@ -1,37 +1,10 @@
 #include "mmq.cuh"
+#include "quantize.cuh"
 
-void ggml_cuda_op_mul_mat_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,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
+#include <vector>
 
-    const int64_t ne00 = src0->ne[0];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    GGML_ASSERT(ne10 % QK8_1 == 0);
-
-    const int64_t ne0 = dst->ne[0];
-
-    const int64_t row_diff = row_high - row_low;
-    const int64_t stride00 = ne00 / ggml_blck_size(src0->type);
-
-    int id = ggml_cuda_get_device();
-    const int cc = ggml_cuda_info().devices[id].cc;
-
-    // the main device has a larger memory buffer to hold the results from all GPUs
-    // nrows_dst == nrows of the matrix that the kernel writes into
-    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
-
-    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
-    // Also its fixup needs to allocate a temporary buffer in the memory pool.
-    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
-    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
-        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
-    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
-
-    switch (src0->type) {
+static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
+    switch (args.type_x) {
         case GGML_TYPE_Q4_0:
             mul_mat_q_case<GGML_TYPE_Q4_0>(ctx, args, stream);
             break;
@@ -90,10 +63,195 @@ void ggml_cuda_op_mul_mat_q(
             GGML_ABORT("fatal error");
             break;
     }
+}
+
+void ggml_cuda_mul_mat_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.
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    cudaStream_t stream = ctx.stream();
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+
+    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);
+
+    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));
+
+    const char  * src0_d = (const char  *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       *  dst_d = (float       *)  dst->data;
+
+    const int64_t ne10_padded = GGML_PAD(ne10, MATRIX_ROW_PADDING);
+
+    const int64_t s01 = src0->nb[1] / ts_src0;
+    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;
+
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
+
+    if (!ids) {
+        const size_t nbytes_src1_q8_1 = ne13*ne12 * ne11*ne10_padded * sizeof(block_q8_1)/QK8_1 +
+            get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
+        ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_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_mmq_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type,
+                ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
+        }
+
+        const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+        const int64_t s13 = ne12*s12;
+
+        const mmq_args args = {
+            src0_d, src0->type, (const int *) src1_q8_1.ptr, nullptr, nullptr, dst_d,
+            ne00, ne01, ne1, s01, s1,
+            ne02, ne12, s02, s12, s2,
+            ne03, ne13, s03, s13, s3,
+            use_stream_k};
+        ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
+        return;
+    }
+
+    GGML_ASSERT(ne13 == 1);
+    GGML_ASSERT(nb12 % nb11 == 0);
+    GGML_ASSERT(nb2  % nb1  == 0);
+
+    const int64_t n_expert_used = ids->ne[0];
+    const int64_t ne_get_rows = ne12 * n_expert_used;
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    std::vector<int32_t> ids_src1_host;
+    ids_src1_host.reserve(ne_get_rows);
+    std::vector<int32_t> ids_dst_host;
+    ids_dst_host.reserve(ne_get_rows);
+    std::vector<int32_t> tokens_per_expert_host(ne02);
+    std::vector<int32_t> expert_bounds_host(ne02 + 1);
+    ggml_cuda_pool_alloc<int32_t> ids_buf_dev(ctx.pool());
+
+    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids->data, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    for (int64_t i02 = 0; i02 < ne02; ++i02) { // expert matrices
+        for (int64_t i12 = 0; i12 < ne12; ++i12) { // tokens
+            for (int64_t iex = 0; iex < n_expert_used; ++iex) {
+                const int32_t expert_to_use = *(const int32_t *)(ids_host.data() + i12*ids->nb[1] + iex*ids->nb[0]);
+                assert(expert_to_use >= 0 && expert_to_use < ne02);
+                if (expert_to_use == i02) {
+                    ids_src1_host.push_back(i12*(nb12/nb11) + iex % ne11);
+                    ids_dst_host.push_back(i12*ne1 + iex);
+                    tokens_per_expert_host[i02]++;
+                    break;
+                }
+            }
+        }
+    }
+
+    int32_t cumsum = 0;
+    for (int64_t i = 0; i < ne02; ++i) {
+        expert_bounds_host[i] = cumsum;
+        cumsum += tokens_per_expert_host[i];
+    }
+    expert_bounds_host[ne02] = cumsum;
+
+    std::vector<int32_t> ids_buf_host;
+    ids_buf_host.reserve(ids_src1_host.size() + ids_dst_host.size() + expert_bounds_host.size());
+    ids_buf_host.insert(ids_buf_host.end(), ids_src1_host.begin(), ids_src1_host.end());
+    ids_buf_host.insert(ids_buf_host.end(), ids_dst_host.begin(), ids_dst_host.end());
+    ids_buf_host.insert(ids_buf_host.end(), expert_bounds_host.begin(), expert_bounds_host.end());
+    ids_buf_dev.alloc(ids_buf_host.size() + get_mmq_x_max_host(cc)); // Expert bounds are padded on device.
+    CUDA_CHECK(cudaMemcpyAsync(ids_buf_dev.ptr, ids_buf_host.data(), ids_buf_host.size()*sizeof(int32_t), cudaMemcpyHostToDevice, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    const int32_t * ids_src1_dev      = ids_buf_dev.ptr;
+    const int32_t * ids_dst_dev       = ids_src1_dev + ids_src1_host.size();
+    const int32_t * expert_bounds_dev = ids_dst_dev + ids_dst_host.size();
+
+    const size_t nbytes_src1_q8_1 = ne12*n_expert_used*ne10_padded * sizeof(block_q8_1)/QK8_1 +
+        get_mmq_x_max_host(cc)*sizeof(block_q8_1_mmq);
+    ggml_cuda_pool_alloc<char> src1_q8_1(ctx.pool(), nbytes_src1_q8_1);
+
+    const int64_t ne11_flat = ne12*n_expert_used;
+    const int64_t ne12_flat = 1;
+    const int64_t ne13_flat = 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[2] / ts_src1;
+        quantize_mmq_q8_1_cuda(src1_d, ids_src1_dev, src1_q8_1.get(), src0->type,
+            ne10, s11, s12, s13, ne10_padded, ne11_flat, ne12_flat, ne13_flat, stream);
+    }
+
+    const int64_t s12 = ne11*ne10_padded * sizeof(block_q8_1)/(QK8_1*sizeof(int));
+    const int64_t s13 = ne12*s12;
+
+    // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.
+    const mmq_args args = {
+        src0_d, src0->type, (const int *) src1_q8_1.ptr, ids_dst_dev, expert_bounds_dev, dst_d,
+        ne00, ne01, ne_get_rows, s01, s1,
+        ne02, ne02, s02, s12, s2,
+        ne03, ne13, s03, s13, s3,
+        use_stream_k};
+
+    ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
+}
+
+void ggml_cuda_op_mul_mat_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,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    const int64_t ne00 = src0->ne[0];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    const int64_t row_diff = row_high - row_low;
+    const int64_t stride01 = ne00 / ggml_blck_size(src0->type);
+
+    const int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
+
+    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
+    // Also its fixup needs to allocate a temporary buffer in the memory pool.
+    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
+    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) &&
+        ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA && src1_ncols == ne11;
+    const mmq_args args = {
+        src0_dd_i, src0->type, (const int *) src1_ddq_i, nullptr, nullptr, dst_dd_i,
+        ne00, row_diff, src1_ncols, stride01, nrows_dst,
+        1, 1, 0, 0, 0,
+        1, 1, 0, 0, 0,
+        use_stream_k};
+
+    ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
 
     GGML_UNUSED(src1);
     GGML_UNUSED(dst);
     GGML_UNUSED(src1_ddf_i);
+    GGML_UNUSED(src1_padded_row_size);
 }
 
 bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {

ggml/src/ggml-cuda/mmq.cuh

@@ -13,9 +13,10 @@ using namespace ggml_cuda_mma;
 #define MMQ_ITER_K 256
 #define MMQ_NWARPS 8
 
-typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int & kbx0, const int & i_max, const int & stride);
-typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00);
-typedef void (*mmq_write_back_t)(const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max);
+typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int kbx0, const int i_max, const int stride);
+typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00);
+typedef void (*mmq_write_back_t)(const float * __restrict__ sum, const int32_t * __restrict__ get_rows_to_sorted,
+    float * __restrict__ dst, const int stride, const int i_max, const int j_max);
 
 enum mmq_q8_1_ds_layout {
     MMQ_Q8_1_DS_LAYOUT_D4,
@@ -233,7 +234,7 @@ static constexpr __device__ int mmq_get_granularity_device(const int /* mmq_x */
 // ------------------------------------------------------------
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -289,7 +290,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -328,7 +329,7 @@ static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_1(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -384,7 +385,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -423,7 +424,7 @@ static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_0(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -495,7 +496,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_1(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -565,7 +566,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q8_0(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -621,7 +622,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -651,7 +652,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps, mmq_q8_1_ds_layout ds_layout>
 static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     typedef tile<16, 8, int> tile_A;
     typedef tile< 8, 8, int> tile_B;
@@ -732,7 +733,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -762,7 +763,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     typedef tile<16, 8, int> tile_A;
     typedef tile< 8, 8, int> tile_B;
@@ -839,7 +840,7 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16;
     const int   * x_qs = (const int   *) x;
@@ -871,7 +872,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 #ifdef NEW_MMA_AVAILABLE
 
     typedef tile<16, 4, int> tile_A;
@@ -955,7 +956,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q2_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1011,7 +1012,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1074,7 +1075,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 #ifdef NEW_MMA_AVAILABLE
 
     typedef tile<16, 4, int> tile_A;
@@ -1201,7 +1202,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q3_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1298,7 +1299,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q3_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1340,7 +1341,7 @@ static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, co
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1437,7 +1438,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1469,7 +1470,7 @@ static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1578,7 +1579,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1610,7 +1611,7 @@ static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q6_K(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1693,7 +1694,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 
     constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y);
     const int   * x_qs = (const int   *) x;
@@ -1726,7 +1727,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
 
 template <int mmq_x, int mmq_y, int nwarps>
 static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
-    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
 #ifdef NEW_MMA_AVAILABLE
 
     typedef tile<16, 4, int> tile_A;
@@ -1835,7 +1836,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_nl(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1893,7 +1894,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_xxs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -1951,7 +1952,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_xs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2007,7 +2008,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_s(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2070,7 +2071,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq3_xxs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2126,7 +2127,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq3_s(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2189,7 +2190,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq1_s(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2245,7 +2246,7 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 }
 
 template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_xs(
-    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
 
 #ifdef NEW_MMA_AVAILABLE
     int   * x_qs = (int   *)  x_tile;
@@ -2306,8 +2307,8 @@ template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinlin
 
 template<int mmq_x, int mmq_y, int nwarps, bool need_check>
 static __device__ __forceinline__ void mmq_write_back_dp4a(
-    const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
-
+        const float * __restrict__ sum, const int32_t * __restrict__ ids_dst, float * __restrict__ dst,
+        const int stride, const int i_max, const int j_max) {
 #pragma unroll
     for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
         const int j = j0 + threadIdx.y;
@@ -2324,15 +2325,15 @@ static __device__ __forceinline__ void mmq_write_back_dp4a(
                 continue;
             }
 
-            dst[j*stride + i] = sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+            dst[ids_dst[j]*stride + i] = sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
         }
     }
 }
 
 template<int mmq_x, int mmq_y, int nwarps, bool need_check>
 static __device__ __forceinline__ void mmq_write_back_mma(
-    const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
-
+        const float * __restrict__ sum, const int * __restrict__ ids_dst, float * __restrict__ dst,
+        const int stride, const int i_max, const int j_max) {
     typedef tile<16, 8, int> tile_C;
 
     constexpr int granularity = mmq_get_granularity_device(mmq_x);
@@ -2362,7 +2363,7 @@ static __device__ __forceinline__ void mmq_write_back_mma(
                     continue;
                 }
 
-                dst[j*stride + i] = sum[(j0/tile_C::J + n)*tile_C::ne + l];
+                dst[ids_dst[j]*stride + i] = sum[(j0/tile_C::J + n)*tile_C::ne + l];
             }
         }
     }
@@ -2518,17 +2519,18 @@ struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_XS> {
 };
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check, bool fixup>
-static __device__ void mul_mat_q_process_tile(
-    const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
-    const int & ne00, const int & ne01, const int & stride01, const int & ne10, const int & ne11, const int & stride11, const int & ne0,
-    const int & it, const int & jt, const int & kb0_start, const int & kb0_stop) {
+static __device__ __forceinline__ void mul_mat_q_process_tile(
+        const char * __restrict__ x, const int offset_x, const int * __restrict__ y,
+        const int * __restrict__ ids_dst, float * __restrict__ dst, float * __restrict__ tmp_fixup,
+        const int nrows_x, const int ncols_y, const int stride_row_x, const int stride_col_dst,
+        const int tile_x_max_i, const int tile_y_max_j, const int kb0_start, const int kb0_stop) {
 
     constexpr int              qk         = ggml_cuda_type_traits<type>::qk;
     constexpr int              mmq_y      = get_mmq_y_device();
     constexpr load_tiles_mmq_t load_tiles = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::load_tiles;
 
-    extern __shared__ char data_mul_mat_q[];
-    int * tile_y = (int *) data_mul_mat_q;
+    extern __shared__ int data_mul_mat_q[];
+    int * tile_y = data_mul_mat_q + mmq_x;
     int * tile_x = tile_y + GGML_PAD(mmq_x*(WARP_SIZE + WARP_SIZE/QI8_1), nwarps*WARP_SIZE);
 
 #ifdef NEW_MMA_AVAILABLE
@@ -2543,16 +2545,11 @@ static __device__ void mul_mat_q_process_tile(
 
     float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
 
-    const int tile_x_max_i = ne01 - it*mmq_y - 1;
-    const int tile_y_max_j = ne11 - jt*mmq_x - 1;
-
-    const int * y = (const int *) yc + jt*(mmq_x*sizeof(block_q8_1_mmq)/sizeof(int));
-
     for (int kb0 = kb0_start; kb0 < kb0_stop; kb0 += blocks_per_iter) {
-        load_tiles(x, tile_x, stride01*it*mmq_y + kb0, tile_x_max_i, stride01);
+        load_tiles(x, tile_x, offset_x + kb0, tile_x_max_i, stride_row_x);
 
         {
-            const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int));
+            const int * by0 = y + ncols_y*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int));
 #pragma unroll
             for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) {
                 int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x;
@@ -2568,7 +2565,7 @@ static __device__ void mul_mat_q_process_tile(
         __syncthreads();
 
         {
-            const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 1*sizeof(block_q8_1_mmq)/sizeof(int));
+            const int * by0 = y + ncols_y*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 1*sizeof(block_q8_1_mmq)/sizeof(int));
 #pragma unroll
             for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) {
                 int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x;
@@ -2585,12 +2582,10 @@ static __device__ void mul_mat_q_process_tile(
     }
 
     if (fixup) {
-        write_back(sum, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x);
+        write_back(sum, ids_dst, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x);
     } else {
-        write_back(sum, dst + jt*mmq_x*ne0 + it*mmq_y, ne0, tile_x_max_i, tile_y_max_j);
+        write_back(sum, ids_dst, dst, stride_col_dst, tile_x_max_i, tile_y_max_j);
     }
-
-    GGML_UNUSED(ne00); GGML_UNUSED(ne10);
 }
 
 
@@ -2609,8 +2604,11 @@ template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 static __global__ void mul_mat_q(
-    const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
-    const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) {
+        const char * __restrict__ x, const int * __restrict__ y, const int32_t * __restrict__ ids_dst,
+        const int32_t * __restrict__ expert_bounds, float * __restrict__ dst, float * __restrict__ tmp_fixup,
+        const int ncols_x, const int nrows_x, const int ncols_y, const int stride_row_x, const int stride_col_dst,
+        const int channel_ratio, const int nchannels_y, const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
+        const int sample_ratio, const int nsamples_y, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst) {
 
     // Skip unused template specializations for faster compilation:
     if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) {
@@ -2621,26 +2619,85 @@ static __global__ void mul_mat_q(
     constexpr int qk    = ggml_cuda_type_traits<type>::qk;
     constexpr int mmq_y = get_mmq_y_device();
 
+    const int ntx = (ncols_y + mmq_x - 1) / mmq_x; // Number of tiles x
+    const int nty = (nrows_x + mmq_y - 1) / mmq_y; // Number of tiles y
+
+    // Initialize the ids for writing back data with just the index.
+    // For regular matrix multiplications this is never changed.
+    // For MoE the correct indices are loaded from ids_dst.
+    extern __shared__ int ids_dst_shared[]; // Stored at beginning of shared memory.
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+        const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+        if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+            break;
+        }
+
+        ids_dst_shared[j] = j;
+    }
+
     // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
 #if (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
     {
+        const int wt = blockIdx.z / nchannels_y;
+        const int zt = blockIdx.z - wt*nchannels_y;
+        const int jt = blockIdx.y;
+        const int it = blockIdx.x;
+
+        // Defaults for regular matrix multiplication:
+        int col_low    = 0;
+        int col_high   = ncols_y;
+        int col_diff   = ncols_y;
+        int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
+        int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
+
+        if (ids_dst) {
+            col_low  = expert_bounds[zt + 0];
+            col_high = expert_bounds[zt + 1];
+            col_diff = col_high - col_low;
+
+            offset_y   = 0;
+            offset_dst = 0;
+
+            if (jt*mmq_x >= col_diff) {
+                return;
+            }
+
+#pragma unroll
+            for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+                const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+                if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+                    break;
+                }
+
+                ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j];
+            }
+        }
+
+        offset_y   += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int));
+        offset_dst += it*mmq_y;
+
+        const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
+        const int tile_y_max_j = col_diff - jt*mmq_x - 1;
+
+        const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
+
         constexpr bool fixup = false;
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-            (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
-                blockIdx.x, blockIdx.y, 0, ne00/qk);
+            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+             tile_x_max_i, tile_y_max_j, 0, ncols_x/qk);
         return;
     }
 #endif // (defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < GGML_CUDA_CC_VOLTA
 
-    const     int64_t blocks_per_ne00 = ne00 / qk;
+    const     int64_t blocks_per_ne00 = ncols_x / qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
 
-    const int ntx = (ne11 + mmq_x - 1) / mmq_x; // Number of tiles x
-    const int nty = (ne01 + mmq_y - 1) / mmq_y; // Number of tiles y
-
     // kbc == k block continuous, current index in continuous ijk space.
-    int64_t kbc      = (int64_t) blockIdx.x     *blocks_per_ne00*ntx*nty / gridDim.x;
-    int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*blocks_per_ne00*ntx*nty / gridDim.x;
+    int64_t kbc      = (int64_t) blockIdx.x     *nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+    int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
 
     kbc      -= (kbc      % blocks_per_ne00) % blocks_per_iter;
     kbc_stop -= (kbc_stop % blocks_per_ne00) % blocks_per_iter;
@@ -2649,13 +2706,64 @@ static __global__ void mul_mat_q(
     int kb0_start = kbc % blocks_per_ne00;
     int kb0_stop  = min(blocks_per_ne00, kb0_start + kbc_stop - kbc);
     while (kbc < kbc_stop && kb0_stop == blocks_per_ne00) {
-        const int jt =  kbc /    (blocks_per_ne00*nty);                    // j index of current tile.
-        const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00; // i index of current tile.
+        int tmp = kbc;
+        const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+        tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+        const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
+        tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
+        const int zt = tmp / (ntx*blocks_per_ne00);
+        tmp -= zt * (ntx*blocks_per_ne00);
+        const int jt = tmp / blocks_per_ne00;
+
+        // Defaults for regular matrix multiplication:
+        int col_low    = 0;
+        int col_high   = ncols_y;
+        int col_diff   = ncols_y;
+        int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
+        int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
+
+        if (ids_dst) {
+            col_low  = expert_bounds[zt + 0];
+            col_high = expert_bounds[zt + 1];
+            col_diff = col_high - col_low;
+
+            offset_y   = 0;
+            offset_dst = 0;
+
+            if (jt*mmq_x >= col_diff) {
+                kbc += blocks_per_ne00;
+                kbc -= kbc % blocks_per_ne00;
+
+                kb0_start = 0;
+                kb0_stop  = min(blocks_per_ne00, kbc_stop - kbc);
+
+                continue;
+            }
+
+#pragma unroll
+            for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+                const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+                if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+                    break;
+                }
+
+                ids_dst_shared[j] = ids_dst[col_low + jt*mmq_x + j];
+            }
+        }
+
+        offset_y   += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int));
+        offset_dst += it*mmq_y;
+
+        const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
+        const int tile_y_max_j = col_diff - jt*mmq_x - 1;
+
+        const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
 
         constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
         mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-            (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
-             it, jt, kb0_start, kb0_stop);
+            (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+             tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 
         kbc += blocks_per_ne00;
         kbc -= kbc % blocks_per_ne00;
@@ -2668,55 +2776,106 @@ static __global__ void mul_mat_q(
         return;
     }
 
-    const int jt =  kbc /    (blocks_per_ne00*nty);
-    const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
+    int tmp = kbc;
+    const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
+    tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
+    const int zt = tmp / (ntx*blocks_per_ne00);
+    tmp -= zt * (ntx*blocks_per_ne00);
+    const int jt = tmp / blocks_per_ne00;
+
+    // Defaults for regular matrix multiplication:
+    int col_low    = 0;
+    int col_high   = ncols_y;
+    int col_diff   = ncols_y;
+    int offset_y   = wt*stride_sample_y   + zt*stride_channel_y;
+    int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst;
+
+    if (ids_dst) {
+        col_low  = expert_bounds[zt + 0];
+        col_high = expert_bounds[zt + 1];
+        col_diff = col_high - col_low;
+
+        offset_y   = 0;
+        offset_dst = 0;
+
+        if (jt*mmq_x >= col_diff) {
+            return;
+        }
+
+        // The memory layout for the fixup buffer is always contiguous, therefore reset ids:
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps*WARP_SIZE) {
+            const int j = j0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+            if (j0 + nwarps*WARP_SIZE > mmq_x && j >= mmq_x) {
+                break;
+            }
+
+            ids_dst_shared[j] = j;
+        }
+    }
+
+    offset_y   += (col_low + jt*mmq_x)*(sizeof(block_q8_1_mmq)/sizeof(int));
+    offset_dst += it*mmq_y;
+
+    const int tile_x_max_i = nrows_x  - it*mmq_y - 1;
+    const int tile_y_max_j = col_diff - jt*mmq_x - 1;
+
+    const int offset_x = (wt/sample_ratio)*stride_sample_x + (zt/channel_ratio)*stride_channel_x + it*mmq_y*stride_row_x;
 
     constexpr bool fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
     mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
-        (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
-            it, jt, kb0_start, kb0_stop);
+        (x, offset_x, y + offset_y, ids_dst_shared, dst + offset_dst, tmp_fixup, nrows_x, ncols_y, stride_row_x, stride_col_dst,
+         tile_x_max_i, tile_y_max_j, kb0_start, kb0_stop);
 }
 
 
 template <ggml_type type, int mmq_x, int nwarps, bool need_check>
 static __global__ void mul_mat_q_stream_k_fixup(
-    float * __restrict__ dst, const float * __restrict__ tmp_last_tile, const int ne00, const int ne01, const int ne11, const int ne0, const int block_num_mmq) {
-
+        const int32_t * ids_dst, const int32_t * expert_bounds, float * __restrict__ dst, const float * __restrict__ tmp_last_tile,
+        const int ncols_x, const int nrows_x, const int ncols_y, const int stride_col_dst,
+        const int nchannels_y, const int stride_channel_dst, const int nsamples_y, const int stride_sample_dst) {
     constexpr int     mmq_y           = get_mmq_y_device();
     constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
     constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
-    const     int64_t blocks_per_ne00 = ne00 / qk;
+    const     int64_t blocks_per_ne00 = ncols_x / qk;
 
     float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
 
-    const int ntx = (ne11 + mmq_x - 1) / mmq_x;
-    const int nty = (ne01 + mmq_y - 1) / mmq_y;
+    const int ntx  = (ncols_y + mmq_x - 1) / mmq_x;
+    const int nty  = (nrows_x + mmq_y - 1) / mmq_y;
+
+    const int bidx0 = blockIdx.x;
+
+    // kbc == k block continuous, current index in continuous ijk space.
+    int64_t kbc0      = (int64_t) bidx0     *nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+    int64_t kbc0_stop = (int64_t)(bidx0 + 1)*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+
+    kbc0      -= (kbc0      % blocks_per_ne00) % blocks_per_iter;
+    kbc0_stop -= (kbc0_stop % blocks_per_ne00) % blocks_per_iter;
+
+    const bool did_not_have_any_data   = kbc0 == kbc0_stop;
+    const bool wrote_beginning_of_tile = kbc0 % blocks_per_ne00 == 0;
+    const bool did_not_write_last      = kbc0/blocks_per_ne00 == kbc0_stop/blocks_per_ne00 && kbc0_stop % blocks_per_ne00 != 0;
+    if (did_not_have_any_data || wrote_beginning_of_tile || did_not_write_last) {
+        return;
+    }
 
     bool any_fixup = false;
 
-    const int bidx_start = ((blockIdx.y*nty + blockIdx.x)     * block_num_mmq)                           / (gridDim.y*gridDim.x);
-    const int bidx_stop  = ((blockIdx.y*nty + blockIdx.x + 1) * block_num_mmq + gridDim.y*gridDim.x - 1) / (gridDim.y*gridDim.x);
+    // Iterate over previous blocks and sum up partial sums written to fixup buffer.
+    // All CUDA blocks that get here must have a previous block that needs a fixup.
+    int64_t bidx = bidx0 - 1;
+    int64_t kbc_stop = kbc0;
+    while(true) {
+        int64_t kbc = bidx*nsamples_y*nchannels_y*ntx*nty*blocks_per_ne00 / gridDim.x;
+        kbc -= (kbc % blocks_per_ne00) % blocks_per_iter;
 
-    int64_t kbc_0;
-    int64_t kbc_stop_0 = (int64_t) bidx_start*blocks_per_ne00*ntx*nty / block_num_mmq;
-
-    for (int bidx = bidx_start; bidx < bidx_stop; ++bidx) {
-        kbc_0 = kbc_stop_0;
-        kbc_stop_0 = (int64_t) (bidx + 1)*blocks_per_ne00*ntx*nty / block_num_mmq;
-
-        const int64_t kbc      = kbc_0      - (kbc_0      % blocks_per_ne00) % blocks_per_iter;
-        const int64_t kbc_stop = kbc_stop_0 - (kbc_stop_0 % blocks_per_ne00) % blocks_per_iter;
-
-        // Skip fixup tile if the MMQ CUDA block never wrote anything to it:
-        if (kbc == kbc_stop || kbc_stop % blocks_per_ne00 == 0) {
-            continue;
-        }
-
-        const int jt =  kbc_stop /    (blocks_per_ne00*nty);
-        const int it = (kbc_stop - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
-
-        // Skip fixup tile if it's unrelated to the output tile assigned to this CUDA block:
-        if ((unsigned)it != blockIdx.x || (unsigned)jt != blockIdx.y) {
+        if (kbc == kbc_stop) { // Did not have any data.
+            bidx--;
+            kbc_stop = kbc;
             continue;
         }
 
@@ -2733,16 +2892,71 @@ static __global__ void mul_mat_q_stream_k_fixup(
                 sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE] += tmp_last_tile[bidx*(mmq_x*mmq_y) + j*mmq_y + i];
             }
         }
+
+        // If this block started in a previous tile we are done and don't need to combine additional partial results.
+        if (kbc % blocks_per_ne00 == 0 || kbc/blocks_per_ne00 < kbc0/blocks_per_ne00) {
+            break;
+        }
+        bidx--;
+        kbc_stop = kbc;
     }
 
     if (!any_fixup) {
         return;
     }
 
-    dst += blockIdx.y*mmq_x*ne0 + blockIdx.x*mmq_y;
+    int tmp = kbc0;
+    const int it = tmp / (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    tmp -= it * (nsamples_y*nchannels_y*ntx*blocks_per_ne00);
+    const int wt = tmp / (nchannels_y*ntx*blocks_per_ne00);
+    tmp -= wt * (nchannels_y*ntx*blocks_per_ne00);
+    const int zt = tmp / (ntx*blocks_per_ne00);
+    tmp -= zt * (ntx*blocks_per_ne00);
+    const int jt = tmp / blocks_per_ne00;
 
-    const int i_max = ne01 - blockIdx.x*mmq_y - 1;
-    const int j_max = ne11 - blockIdx.y*mmq_x - 1;
+    if (!ids_dst) {
+        const int offset_dst = wt*stride_sample_dst + zt*stride_channel_dst + jt*mmq_x*stride_col_dst + it*mmq_y;
+        dst += offset_dst;
+
+        const int i_max = nrows_x - it*mmq_y - 1;
+        const int j_max = ncols_y - jt*mmq_x - 1;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (j > j_max) {
+                return;
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                if (need_check && i > i_max) {
+                    continue;
+                }
+
+                dst[j*stride_col_dst + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+            }
+        }
+        return;
+    }
+
+    __shared__ int ids_dst_shared[mmq_x];
+    const int col_low  = expert_bounds[zt + 0];
+    const int col_high = expert_bounds[zt + 1];
+    const int col_diff = col_high - col_low;
+
+    for (int j = threadIdx.y*WARP_SIZE + threadIdx.x; j < mmq_x; j += nwarps*WARP_SIZE) {
+        ids_dst_shared[j] = ids_dst[col_low + j];
+    }
+
+    const int offset_dst = it*mmq_y;
+    dst += offset_dst;
+
+    const int i_max = nrows_x  - it*mmq_y - 1;
+    const int j_max = col_diff - jt*mmq_x - 1;
 
 #pragma unroll
     for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
@@ -2760,26 +2974,27 @@ static __global__ void mul_mat_q_stream_k_fixup(
                 continue;
             }
 
-            dst[j*ne0 + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+            dst[ids_dst_shared[j]*stride_col_dst + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
         }
     }
 }
 
 struct mmq_args {
-    const char * x; const char * y; float * dst;
-    int64_t ne00; int64_t ne01; int64_t stride01;
-    int64_t ne10; int64_t ne11; int64_t stride11;
-    int64_t ne0;
+    const char * x; ggml_type type_x; const int * y; const int32_t * ids_dst; const int32_t * expert_bounds; float * dst;
+    int64_t ncols_x; int64_t nrows_x; int64_t ncols_y; int64_t stride_row_x; int64_t nrows_dst;
+    int64_t nchannels_x; int64_t nchannels_y; int64_t stride_channel_x; int64_t stride_channel_y; int64_t stride_channel_dst;
+    int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst;
     bool use_stream_k;
 };
 
 template<ggml_type type>
-static int mmq_get_shmem(const int mmq_x, const int mmq_y, const int cc) {
+static size_t mmq_get_nbytes_shared(const int mmq_x, const int mmq_y, const int cc) {
     const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y);
     const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type);
-    const int shmem_x = new_mma_available(cc) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
-    const int shmem_y = mmq_x*sizeof(block_q8_1_mmq);
-    return shmem_x + GGML_PAD(shmem_y, MMQ_NWARPS*WARP_SIZE*sizeof(int));
+    const size_t nbs_ids = mmq_x*sizeof(int);
+    const size_t nbs_x = new_mma_available(cc) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
+    const size_t nbs_y = mmq_x*sizeof(block_q8_1_mmq);
+    return nbs_ids + nbs_x + GGML_PAD(nbs_y, MMQ_NWARPS*WARP_SIZE*sizeof(int));
 }
 
 template <ggml_type type, int mmq_x>
@@ -2791,86 +3006,114 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
 
     const dim3 block_dims(WARP_SIZE, MMQ_NWARPS, 1);
 
-    const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);
+    const int nbytes_shared = mmq_get_nbytes_shared<type>(mmq_x, mmq_y, cc);
 
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
-    static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-    if (!shmem_limit_raised[id]) {
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
-        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
-        shmem_limit_raised[id] = true;
+    static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
+    if (!shared_memory_limit_raised[id]) {
+        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
+        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared));
+        shared_memory_limit_raised[id] = true;
     }
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && !defined(GGML_USE_MUSA)
 
-    const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
-    const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
-    const dim3 block_nums_xy_tiling(nty, ntx, 1);
+    const int nty  = (args.nrows_x + mmq_y - 1) / mmq_y;
+    const int ntx  = (args.ncols_y + mmq_x - 1) / mmq_x;
+    const int ntzw = args.nchannels_y * args.nsamples_y;
+    const dim3 block_nums_xy_tiling(nty, ntx, ntzw);
+
+    GGML_ASSERT(args.nchannels_y % args.nchannels_x == 0);
+    GGML_ASSERT(args.nsamples_y  % args.nsamples_x  == 0);
+    const int channel_ratio = args.nchannels_y / args.nchannels_x;
+    const int sample_ratio  = args.nsamples_y  / args.nsamples_x;
 
     if (!args.use_stream_k) {
-        if (args.ne01 % mmq_y == 0) {
+        if (args.nrows_x % mmq_y == 0) {
             constexpr bool need_check = false;
-            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
-                (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
+                (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
+                 args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+                 channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
         } else {
             constexpr bool need_check = true;
-            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
-                (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, nbytes_shared, stream>>>
+                (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, nullptr,
+                 args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+                 channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+                 sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
         }
         return;
     }
 
-    const dim3 block_nums_mmq(nsm, 1, 1);
+    const dim3 block_nums_stream_k(nsm, 1, 1);
+    const bool fixup_needed = ntx*nty*ntzw % nsm != 0;
 
     ggml_cuda_pool & pool = ctx.pool(id);
-    ggml_cuda_pool_alloc<float> tmp_fixup(pool, block_nums_mmq.x * mmq_x*mmq_y);
+    ggml_cuda_pool_alloc<float> tmp_fixup(pool);
+    if (fixup_needed) {
+        tmp_fixup.alloc(block_nums_stream_k.x * mmq_x*mmq_y);
+    }
 
-    if (args.ne01 % mmq_y == 0) {
+    if (args.nrows_x % mmq_y == 0) {
         constexpr bool need_check = false;
 
-        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
-            (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
+            (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
+             args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+             channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
 
-        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
-            (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
+        if (!fixup_needed) {
+            return;
+        }
+
+        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
+            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_y,
+             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
     } else {
         constexpr bool need_check = true;
 
-        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
-            (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, nbytes_shared, stream>>>
+            (args.x, args.y, args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr,
+             args.ncols_x, args.nrows_x, args.ncols_y, args.stride_row_x, args.nrows_dst,
+             channel_ratio, args.nchannels_y, args.stride_channel_x, args.stride_channel_y, args.stride_channel_dst,
+             sample_ratio, args.nsamples_y, args.stride_sample_x, args.stride_sample_y, args.stride_sample_dst);
 
-        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
-            (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
+        if (!fixup_needed) {
+            return;
+        }
+
+        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_stream_k, block_dims, 0, stream>>>
+            (args.ids_dst, args.expert_bounds, args.dst, tmp_fixup.ptr, args.ncols_x, args.nrows_x, args.ncols_y,
+             args.nrows_dst, args.nchannels_y, args.stride_channel_dst, args.nsamples_y, args.stride_sample_dst);
     }
 }
 
 template <ggml_type type>
 void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
-    const int id    = ggml_cuda_get_device();
-    const int cc    = ggml_cuda_info().devices[id].cc;
-    const int smpbo = ggml_cuda_info().devices[id].smpbo;
+    const int    id    = ggml_cuda_get_device();
+    const int    cc    = ggml_cuda_info().devices[id].cc;
+    const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
 
     const int mmq_x_max = get_mmq_x_max_host(cc);
     const int mmq_y = get_mmq_y_host(cc);
-    const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y;
-    const bool use_stream_k = GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA;
 
     int mmq_x_best  = 0;
-    int nparts_best = INT_MAX;
+    int ntiles_x_best = INT_MAX;
 
-    for (int mmq_x = 8; mmq_x <= mmq_x_max && nparts_best > 1; mmq_x += 8) {
+    for (int mmq_x = 8; mmq_x <= mmq_x_max && ntiles_x_best > 1; mmq_x += 8) {
         const int granularity = mmq_get_granularity_host(mmq_x, cc);
 
-        if (mmq_x % granularity != 0 || mmq_get_shmem<type>(mmq_x, mmq_y, cc) > smpbo) {
+        if (mmq_x % granularity != 0 || mmq_get_nbytes_shared<type>(mmq_x, mmq_y, cc) > smpbo) {
             continue;
         }
 
-        const int ntiles_x = (args.ne11 + mmq_x - 1) / mmq_x;
-        const int nwaves_xy_tiling = ntiles_x*block_num_y;
-        const int nparts = use_stream_k ? ntiles_x : nwaves_xy_tiling;
+        const int ntiles_x = (args.ncols_y + mmq_x - 1) / mmq_x;
 
-        if (nparts < nparts_best) {
-            mmq_x_best  = mmq_x;
-            nparts_best = nparts;
+        if (ntiles_x < ntiles_x_best) {
+            mmq_x_best = mmq_x;
+            ntiles_x_best = ntiles_x;
         }
     }
 
@@ -2954,6 +3197,9 @@ extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS);
 
 // -------------------------------------------------------------------------------------------------------------------------
 
+void ggml_cuda_mul_mat_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_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/mmvq.cu

@@ -158,7 +158,7 @@ static __global__ void mul_mat_vec_q(
     const     int blocks_per_row_x = ncols_x / qk;
     constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
 
-    // The MUL_MAT_ID code path with ids != nullptr is only implemetned for ncols_dst == 1.
+    // The MUL_MAT_ID code path with ids != nullptr is only implemented 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;
@@ -507,7 +507,7 @@ void ggml_cuda_mul_mat_vec_q(
     GGML_ASSERT(        nb0        == ts_dst);
     GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
 
-    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for  batch size 1.
+    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.
 
     const float   * src1_d =       (const float   *) src1->data;
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
@@ -519,7 +519,7 @@ void ggml_cuda_mul_mat_vec_q(
         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);
+        quantize_row_q8_1_cuda(src1_d, nullptr, src1_q8_1.get(), src0->type, ne10, s11, s12, s13, ne10_padded, ne11, ne12, ne13, stream);
     }
 
     const int64_t s01 = src0->nb[1] / ts_src0;

ggml/src/ggml-cuda/quantize.cu

@@ -49,29 +49,38 @@ static __global__ void quantize_q8_1(
 
 template <mmq_q8_1_ds_layout ds_layout>
 static __global__ void quantize_mmq_q8_1(
-    const float * __restrict__ x, void * __restrict__ vy, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded) {
+        const float * __restrict__ x, const int32_t * __restrict__ ids, 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) {
 
     constexpr int vals_per_scale = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 64 : 32;
     constexpr int vals_per_sum   = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 16 : 32;
 
-    const int64_t ix0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
+    const int64_t i0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
 
-    if (ix0 >= kx0_padded) {
+    if (i0 >= ne0) {
         return;
     }
 
-    const float4 * x4 = (const float4 *) x;
+    const int64_t i1 = blockIdx.y;
+    const int64_t i2 = blockIdx.z % ne2;
+    const int64_t i3 = blockIdx.z / ne2;
 
-    const int64_t ix1 = kx1*blockIdx.z + blockIdx.y;
+    const int64_t i00 = i0;
+    const int64_t i01 = ids ? ids[i1] : i1;
+    const int64_t i02 = i2;
+    const int64_t i03 = i3;
+
+    const float4 * x4 = (const float4 *) x;
 
     block_q8_1_mmq * y = (block_q8_1_mmq *) vy;
 
     const int64_t ib0 = blockIdx.z*((int64_t)gridDim.y*gridDim.x*blockDim.x/QK8_1); // first block of channel
-    const int64_t ib  = ib0 + (ix0 / (4*QK8_1))*kx1 + blockIdx.y;                   // block index in channel
-    const int64_t iqs = ix0 % (4*QK8_1);                                            // quant index in block
+    const int64_t ib  = ib0 + (i0 / (4*QK8_1))*ne1 + blockIdx.y;                    // block index in channel
+    const int64_t iqs = i0 % (4*QK8_1);                                             // quant index in block
 
     // Load 4 floats per thread and calculate max. abs. value between them:
-    const float4 xi = ix0 < kx0 ? x4[(ix1*kx0 + ix0)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+    const float4 xi = i0 < ne00 ? x4[(i03*s03 + i02*s02 + i01*s01 + i00)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
     float amax = fabsf(xi.x);
     amax = fmaxf(amax, fabsf(xi.y));
     amax = fmaxf(amax, fabsf(xi.z));
@@ -87,7 +96,7 @@ static __global__ void quantize_mmq_q8_1(
     if (ds_layout != MMQ_Q8_1_DS_LAYOUT_D4) {
         sum = xi.x + xi.y + xi.z + xi.w;
 
-        // Exchange calculate sum across vals_per_sum/4 threads.
+        // Calculate sums across vals_per_sum/4 threads.
 #pragma unroll
         for (int offset = vals_per_sum/8; offset > 0; offset >>= 1) {
             sum += __shfl_xor_sync(0xFFFFFFFF, sum, offset, WARP_SIZE);
@@ -137,9 +146,10 @@ static __global__ void quantize_mmq_q8_1(
 }
 
 void quantize_row_q8_1_cuda(
-    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) {
-
+        const float * x, const int32_t * ids, 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(!ids);
     GGML_ASSERT(ne0 % QK8_1 == 0);
 
     const int64_t block_num_x = (ne0 + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
@@ -150,9 +160,9 @@ void quantize_row_q8_1_cuda(
 }
 
 void quantize_mmq_q8_1_cuda(
-    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) {
-
+        const float * x, const int32_t * ids, 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(ne0 % (4*QK8_1) == 0);
 
     const int64_t block_num_x = (ne0 + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
@@ -161,21 +171,18 @@ void quantize_mmq_q8_1_cuda(
     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, ne00, ne1, ne0);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             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, ne00, ne1, ne0);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             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, ne00, ne1, ne0);
+                <<<num_blocks, block_size, 0, stream>>>(x, ids, vy, ne00, s01, s02, s03, ne0, ne1, ne2);
             break;
         default:
             GGML_ABORT("fatal error");
             break;
     }
-    GGML_UNUSED(s01);
-    GGML_UNUSED(s02);
-    GGML_UNUSED(s03);
 }

ggml/src/ggml-cuda/quantize.cuh

@@ -12,13 +12,16 @@ 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 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);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
 
 void quantize_row_q8_1_cuda(
-    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);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);
 
 void quantize_mmq_q8_1_cuda(
-    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);
+        const float * x, const int32_t * ids, void * vy,
+        ggml_type type_src0, int64_t ne00, int64_t s01, int64_t s02, int64_t s03,
+        int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3, cudaStream_t stream);

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp

@@ -482,7 +482,7 @@ float16_t dequantFuncIQ2_XXS(const in decodeBufIQ2_XXS bl, const in uint blockCo
     const uint ib8 = (idx & 0x18) >> 3;  // 0..3
     const uint iqs = 8 * ib32 + ib8;
 
-    const uint8_t qs = bl.block.qs[iqs];
+    const uint qs = bl.block.qs[iqs];
     const uint signscale = pack32(u16vec2(bl16.block.qs[4*ib32+2], bl16.block.qs[4*ib32+3]));
 
     const float dscale = float(bl.block.d) * 0.25 * (0.5 + float(signscale >> 28));

gguf-py/gguf/constants.py

@@ -231,6 +231,7 @@ class Keys:
         BLOCK_COUNT         = "clip.vision.block_count"
         IMAGE_MEAN          = "clip.vision.image_mean"
         IMAGE_STD           = "clip.vision.image_std"
+        SPATIAL_MERGE_SIZE  = "clip.vision.spatial_merge_size"
         USE_GELU            = "clip.use_gelu"
         USE_SILU            = "clip.use_silu"
 
@@ -491,6 +492,7 @@ class MODEL_TENSOR(IntEnum):
     V_ENC_FFN_DOWN       = auto()
     V_PRE_NORM           = auto()
     V_POST_NORM          = auto()
+    V_MM_INP_NORM        = auto()
     V_MM_INP_PROJ        = auto() # gemma3
     V_MM_SOFT_EMB_NORM   = auto() # gemma3
     V_RESMPL_POS_EMBD_K  = auto() # minicpmv
@@ -505,6 +507,7 @@ class MODEL_TENSOR(IntEnum):
     V_RESMPL_PROJ        = auto() # minicpmv
     V_RESMPL_QUERY       = auto() # minicpmv
     V_TOK_EMBD_IMG_BREAK = auto() # pixtral
+    V_MM_PATCH_MERGER    = auto() # mistral small 3.1
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -747,6 +750,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.V_PRE_NORM:                "v.pre_ln",
     MODEL_TENSOR.V_POST_NORM:               "v.post_ln",
     MODEL_TENSOR.V_MM_INP_PROJ:             "mm.input_projection",
+    MODEL_TENSOR.V_MM_INP_NORM:             "mm.input_norm",
     MODEL_TENSOR.V_MM_SOFT_EMB_NORM:        "mm.soft_emb_norm",
     MODEL_TENSOR.V_RESMPL_POS_EMBD_K:       "resampler.pos_embd_k",
     MODEL_TENSOR.V_RESMPL_ATTN_Q:           "resampler.attn.q",
@@ -760,6 +764,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     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_TENSOR.V_MM_PATCH_MERGER:         "mm.patch_merger", # mistral small 3.1
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -783,6 +788,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_PRE_NORM,
         MODEL_TENSOR.V_POST_NORM,
         MODEL_TENSOR.V_MM_INP_PROJ,
+        MODEL_TENSOR.V_MM_INP_NORM,
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM,
         MODEL_TENSOR.V_RESMPL_POS_EMBD_K,
         MODEL_TENSOR.V_RESMPL_ATTN_Q,
@@ -796,6 +802,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.V_RESMPL_PROJ,
         MODEL_TENSOR.V_RESMPL_QUERY,
         MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK,
+        MODEL_TENSOR.V_MM_PATCH_MERGER,
     ],
     MODEL_ARCH.LLAMA: [
         MODEL_TENSOR.TOKEN_EMBD,

gguf-py/gguf/gguf_writer.py

@@ -972,6 +972,9 @@ class GGUFWriter:
     def add_vision_image_std(self, values: Sequence[float]) -> None:
         self.add_array(Keys.ClipVision.IMAGE_STD, values)
 
+    def add_vision_spatial_merge_size(self, value: int) -> None:
+        self.add_uint32(Keys.ClipVision.SPATIAL_MERGE_SIZE, value)
+
     def add_vision_use_gelu(self, value: bool) -> None:
         self.add_bool(Keys.ClipVision.USE_GELU, value)
 

gguf-py/gguf/tensor_mapping.py

@@ -1001,6 +1001,10 @@ class TensorNameMap:
             "multi_modal_projector.mm_input_projection",
         ),
 
+        MODEL_TENSOR.V_MM_INP_NORM: (
+            "multi_modal_projector.norm",
+        ),
+
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM: (
             "multi_modal_projector.mm_soft_emb_norm",
         ),
@@ -1052,6 +1056,10 @@ class TensorNameMap:
         MODEL_TENSOR.V_TOK_EMBD_IMG_BREAK: (
             "v.token_embd.img_break", # for pixtral, this is a generated vector
         ),
+
+        MODEL_TENSOR.V_MM_PATCH_MERGER: (
+            "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1
+        ),
     }
 
     # architecture-specific block mappings

scripts/sync-ggml.last

@@ -1 +1 @@
-13bcf9ce50651a8b4238ec6d136f46f2c1b23b6f
+7734363ac74a14d7f2c8033c552d9d655c5dfc00

src/llama-context.cpp

@@ -114,7 +114,7 @@ llama_context::llama_context(
     }
 
     if (n_ctx_per_seq > hparams.n_ctx_train) {
-        LLAMA_LOG_WARN("%s: n_ctx_pre_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
+        LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n",
                 __func__, n_ctx_per_seq, hparams.n_ctx_train);
     }
 

src/llama-model.cpp

@@ -40,6 +40,7 @@ const char * llm_type_name(llm_type type) {
         case LLM_TYPE_335M:          return "335M";
         case LLM_TYPE_410M:          return "410M";
         case LLM_TYPE_450M:          return "450M";
+        case LLM_TYPE_475M:          return "475M";
         case LLM_TYPE_770M:          return "770M";
         case LLM_TYPE_780M:          return "780M";
         case LLM_TYPE_0_5B:          return "0.5B";
@@ -707,7 +708,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_MOE_EVERY_N_LAYERS,         hparams.moe_every_n_layers, 0);
 
                 if (hparams.n_layer == 12 && hparams.n_embd == 768) {
-                    type = LLM_TYPE_137M;
+                    if (arch == LLM_ARCH_NOMIC_BERT) {
+                        type = LLM_TYPE_137M;
+                    } else if (arch == LLM_ARCH_NOMIC_BERT_MOE && hparams.moe_every_n_layers == 2) {
+                        type = LLM_TYPE_475M;
+                    }
                 }
             } break;
         case LLM_ARCH_BLOOM:

src/llama-model.h

@@ -36,6 +36,7 @@ enum llm_type {
     LLM_TYPE_335M,
     LLM_TYPE_410M,
     LLM_TYPE_450M,
+    LLM_TYPE_475M,
     LLM_TYPE_770M,
     LLM_TYPE_780M,
     LLM_TYPE_0_5B,

tests/test-backend-ops.cpp

@@ -4184,6 +4184,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 2, 1, 3}));
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 1, 3, 2}));
             test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 256, {2, 3}, {1, 1}, {0, 3, 2, 1}));
+
+            // test cases with large ne00/ne10 to cover stream-k fixup
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  1, 1024, {3, 2}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16,  8, 1024, {3, 2}, {1, 1}));
+            test_cases.emplace_back(new test_mul_mat(type_a, type_b, 16, 16, 1024, {3, 2}, {1, 1}));
         }
     }
     for (ggml_type type_a : other_types) {