Merge branch 'ggerganov:master' into master

README.md

@@ -124,6 +124,7 @@ Typically finetunes of the base models below are supported as well.
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
+- JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
 - Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)

common/common.cpp

@@ -340,13 +340,14 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-            sparams.samplers_sequence = parse_samplers_input(argv[i]);
+            const auto sampler_names = string_split(argv[i], ';');
+            sparams.samplers_sequence = sampler_types_from_names(sampler_names);
         } else if (arg == "--sampling-seq") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            sparams.samplers_sequence = argv[i];
+            sparams.samplers_sequence = sampler_types_from_chars(argv[i]);
         } else if (arg == "--top-p") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -906,6 +907,14 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     const llama_sampling_params & sparams = params.sparams;
 
+    std::string sampler_type_chars;
+    std::string sampler_type_names;
+    for (const auto sampler_type : sparams.samplers_sequence) {
+        sampler_type_chars += static_cast<char>(sampler_type);
+        sampler_type_names += sampler_type_to_name_string(sampler_type) + ";";
+    }
+    sampler_type_names.pop_back();
+
     printf("\n");
     printf("usage: %s [options]\n", argv[0]);
     printf("\n");
@@ -947,8 +956,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
     printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\', for example: \"top_k;tfs;typical;top_p;min_p;temp\"\n");
+    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\' (default: %s)\n", sampler_type_names.c_str());
-    printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sparams.samplers_sequence.c_str());
+    printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sampler_type_chars.c_str());
     printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
     printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
     printf("  --min-p N             min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p);
@@ -1097,45 +1106,85 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
 }
 
 //
-// String parsing
+// String utils
 //
 
-std::string parse_samplers_input(std::string input) {
+std::vector<std::string> string_split(std::string input, char separator) {
-    std::string output = "";
+    std::vector<std::string> parts;
+    size_t separator_pos = input.find(separator);
+    while (separator_pos != std::string::npos) {
+        std::string part = input.substr(0, separator_pos);
+        parts.emplace_back(part);
+        input = input.substr(separator_pos + 1);
+        separator_pos = input.find(separator);
+    }
+    parts.emplace_back(input);
+    return parts;
+}
+
+std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names) {
     // since samplers names are written multiple ways
     // make it ready for both system names and input names
-    std::unordered_map<std::string, char> samplers_symbols {
+    std::unordered_map<std::string, llama_sampler_type> sampler_name_map {
-        {"top_k",      'k'},
+        {"top_k",       llama_sampler_type::TOP_K},
-        {"top-k",      'k'},
+        {"top-k",       llama_sampler_type::TOP_K},
-        {"top_p",      'p'},
+        {"top_p",       llama_sampler_type::TOP_P},
-        {"top-p",      'p'},
+        {"top-p",       llama_sampler_type::TOP_P},
-        {"nucleus",    'p'},
+        {"nucleus",     llama_sampler_type::TOP_P},
-        {"typical_p",  'y'},
+        {"typical_p",   llama_sampler_type::TYPICAL_P},
-        {"typical-p",  'y'},
+        {"typical-p",   llama_sampler_type::TYPICAL_P},
-        {"typical",    'y'},
+        {"typical",     llama_sampler_type::TYPICAL_P},
-        {"min_p",      'm'},
+        {"min_p",       llama_sampler_type::MIN_P},
-        {"min-p",      'm'},
+        {"min-p",       llama_sampler_type::MIN_P},
-        {"tfs_z",      'f'},
+        {"tfs_z",       llama_sampler_type::TFS_Z},
-        {"tfs-z",      'f'},
+        {"tfs-z",       llama_sampler_type::TFS_Z},
-        {"tfs",        'f'},
+        {"tfs",         llama_sampler_type::TFS_Z},
-        {"temp",       't'},
+        {"temp",        llama_sampler_type::TEMP},
-        {"temperature",'t'}
+        {"temperature", llama_sampler_type::TEMP}
     };
-    // expected format example: "temp;top_k;tfs_z;typical_p;top_p;min_p"
-    size_t separator = input.find(';');
-    while (separator != input.npos) {
-        std::string name = input.substr(0,separator);
-        input = input.substr(separator+1);
-        separator = input.find(';');
 
-        if (samplers_symbols.find(name) != samplers_symbols.end()) {
+    std::vector<llama_sampler_type> sampler_types;
-            output += samplers_symbols[name];
+    sampler_types.reserve(names.size());
+    for (const auto& name : names) {
+        const auto sampler_item = sampler_name_map.find(name);
+        if (sampler_item != sampler_name_map.end()) {
+            sampler_types.push_back(sampler_item->second);
         }
     }
-    if (samplers_symbols.find(input) != samplers_symbols.end()) {
+    return sampler_types;
-        output += samplers_symbols[input];
+}
+
+std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string) {
+    std::unordered_map<char, llama_sampler_type> sampler_name_map {
+        {'k', llama_sampler_type::TOP_K},
+        {'p', llama_sampler_type::TOP_P},
+        {'y', llama_sampler_type::TYPICAL_P},
+        {'m', llama_sampler_type::MIN_P},
+        {'f', llama_sampler_type::TFS_Z},
+        {'t', llama_sampler_type::TEMP}
+    };
+
+    std::vector<llama_sampler_type> sampler_types;
+    sampler_types.reserve(names_string.size());
+    for (const auto & c : names_string) {
+        const auto sampler_item = sampler_name_map.find(c);
+        if (sampler_item != sampler_name_map.end()) {
+            sampler_types.push_back(sampler_item->second);
+        }
+    }
+    return sampler_types;
+}
+
+std::string sampler_type_to_name_string(llama_sampler_type sampler_type) {
+    switch (sampler_type) {
+        case llama_sampler_type::TOP_K:     return "top_k";
+        case llama_sampler_type::TFS_Z:     return "tfs_z";
+        case llama_sampler_type::TYPICAL_P: return "typical_p";
+        case llama_sampler_type::TOP_P:     return "top_p";
+        case llama_sampler_type::MIN_P:     return "min_p";
+        case llama_sampler_type::TEMP:      return "temp";
+        default : return "";
     }
-    return output;
 }
 
 //
@@ -1550,6 +1599,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
     fprintf(stream, "cpu_has_sse3: %s\n",        ggml_cpu_has_sse3()        ? "true" : "false");
     fprintf(stream, "cpu_has_vsx: %s\n",         ggml_cpu_has_vsx()         ? "true" : "false");
+    fprintf(stream, "cpu_has_matmul_int8: %s\n", ggml_cpu_has_matmul_int8() ? "true" : "false");
 
 #ifdef NDEBUG
     fprintf(stream, "debug: false\n");

common/common.h

@@ -162,10 +162,13 @@ std::string gpt_random_prompt(std::mt19937 & rng);
 void process_escapes(std::string& input);
 
 //
-// String parsing
+// String utils
 //
 
-std::string parse_samplers_input(std::string input);
+std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names);
+std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
+std::vector<std::string> string_split(std::string input, char separator);
+std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
 
 //
 // Model utils

common/sampling.cpp

@@ -103,15 +103,10 @@ std::string llama_sampling_print(const llama_sampling_params & params) {
 std::string llama_sampling_order_print(const llama_sampling_params & params) {
     std::string result = "CFG -> Penalties ";
     if (params.mirostat == 0) {
-        for (auto s : params.samplers_sequence) {
+        for (auto sampler_type : params.samplers_sequence) {
-            switch (s) {
+            const auto sampler_type_name = sampler_type_to_name_string(sampler_type);
-                case 'k': result += "-> top_k "; break;
+            if (!sampler_type_name.empty()) {
-                case 'f': result += "-> tfs_z "; break;
+                result += "-> " + sampler_type_name + " ";
-                case 'y': result += "-> typical_p "; break;
-                case 'p': result += "-> top_p "; break;
-                case 'm': result += "-> min_p "; break;
-                case 't': result += "-> temp "; break;
-                default : break;
             }
         }
     } else {
@@ -127,8 +122,6 @@ static void sampler_queue(
             const llama_sampling_params & params,
                  llama_token_data_array & cur_p,
                                  size_t & min_keep) {
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
-
     const float         temp              = params.temp;
     const float         dynatemp_range    = params.dynatemp_range;
     const float         dynatemp_exponent = params.dynatemp_exponent;
@@ -137,16 +130,16 @@ static void sampler_queue(
     const float         min_p             = params.min_p;
     const float         tfs_z             = params.tfs_z;
     const float         typical_p         = params.typical_p;
-    const std::string & samplers_sequence = params.samplers_sequence;
+    const std::vector<llama_sampler_type> & samplers_sequence = params.samplers_sequence;
 
-    for (auto s : samplers_sequence) {
+    for (auto sampler_type : samplers_sequence) {
-        switch (s){
+        switch (sampler_type) {
-            case 'k': llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
+            case llama_sampler_type::TOP_K    : llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
-            case 'f': llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
+            case llama_sampler_type::TFS_Z    : llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
-            case 'y': llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
+            case llama_sampler_type::TYPICAL_P: llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
-            case 'p': llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
+            case llama_sampler_type::TOP_P    : llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
-            case 'm': llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
+            case llama_sampler_type::MIN_P    : llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
-            case 't':
+            case llama_sampler_type::TEMP:
                 if (dynatemp_range > 0) {
                     float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
                     float dynatemp_max = std::max(0.0f, temp + dynatemp_range);

common/sampling.h

@@ -8,6 +8,16 @@
 #include <vector>
 #include <unordered_map>
 
+// sampler types
+enum class llama_sampler_type : char {
+    TOP_K     = 'k',
+    TOP_P     = 'p',
+    MIN_P     = 'm',
+    TFS_Z     = 'f',
+    TYPICAL_P = 'y',
+    TEMP      = 't'
+};
+
 // sampling parameters
 typedef struct llama_sampling_params {
     int32_t     n_prev                = 64;       // number of previous tokens to remember
@@ -28,7 +38,15 @@ typedef struct llama_sampling_params {
     float       mirostat_tau          = 5.00f;    // target entropy
     float       mirostat_eta          = 0.10f;    // learning rate
     bool        penalize_nl           = true;     // consider newlines as a repeatable token
-    std::string samplers_sequence     = "kfypmt"; // top_k, tail_free, typical_p, top_p, min_p, temp
+
+    std::vector<llama_sampler_type> samplers_sequence = {
+        llama_sampler_type::TOP_K,
+        llama_sampler_type::TFS_Z,
+        llama_sampler_type::TYPICAL_P,
+        llama_sampler_type::TOP_P,
+        llama_sampler_type::MIN_P,
+        llama_sampler_type::TEMP
+    };
 
     std::string grammar;  // optional BNF-like grammar to constrain sampling
 

examples/llava/README.md

@@ -29,19 +29,25 @@ git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
 git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 ```
 
-2. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
+2. Install the required Python packages:
+
+```sh
+pip install -r examples/llava/requirements.txt
+```
+
+3. Use `llava-surgery.py` to split the LLaVA model to LLaMA and multimodel projector constituents:
 
 ```sh
 python ./examples/llava/llava-surgery.py -m ../llava-v1.5-7b
 ```
 
-3. Use `convert-image-encoder-to-gguf.py` to convert the LLaVA image encoder to GGUF:
+4. Use `convert-image-encoder-to-gguf.py` to convert the LLaVA image encoder to GGUF:
 
 ```sh
 python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-patch14-336 --llava-projector ../llava-v1.5-7b/llava.projector --output-dir ../llava-v1.5-7b
 ```
 
-4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
+5. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
 python ./convert.py ../llava-v1.5-7b

examples/llava/llava-surgery.py

@@ -42,5 +42,5 @@ if len(clip_tensors) > 0:
 torch.save(checkpoint, path)
 
 print("Done!")
-print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
+print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
 print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

examples/llava/requirements.txt

@@ -0,0 +1,3 @@
+-r ../../requirements/requirements-convert.txt
+pillow~=10.2.0
+torch~=2.1.1

examples/lookup/lookup.cpp

@@ -1,7 +1,9 @@
 #include "common.h"
+#include "ggml.h"
 #include "llama.h"
 
 #include <cmath>
+#include <cstdint>
 #include <cstdio>
 #include <string>
 #include <vector>
@@ -73,6 +75,8 @@ int main(int argc, char ** argv){
     int n_drafted = 0;
     int n_accept  = 0;
 
+    int64_t t_draft_us = 0;
+
     int n_past = inp.size();
 
     bool has_eos = false;
@@ -160,7 +164,7 @@ int main(int argc, char ** argv){
 
         // generate n_pred tokens through prompt lookup
         auto prompt_lookup = [&]() -> void {
-            int inp_size = inp.size();
+            const int inp_size = inp.size();
             for (int ngram_size = ngram_max ; ngram_size > ngram_min; --ngram_size){
                 const llama_token * ngram = &inp[inp_size - ngram_size];
 
@@ -191,8 +195,12 @@ int main(int argc, char ** argv){
             return;
         };
 
+        const int64_t t_start_draft_us = ggml_time_us();
+
         prompt_lookup();
 
+        t_draft_us += ggml_time_us() - t_start_draft_us;
+
         llama_decode(ctx, batch_tgt);
         ++n_past;
 
@@ -210,6 +218,8 @@ int main(int argc, char ** argv){
     LOG_TEE("n_draft   = %d\n", n_draft);
     LOG_TEE("n_predict = %d\n", n_predict);
     LOG_TEE("n_drafted = %d\n", n_drafted);
+    LOG_TEE("t_draft   = %.2f ms, %.2f us per token, %.2f tokens per second\n",
+            t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
     LOG_TEE("n_accept  = %d\n", n_accept);
     LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
 

examples/main/main.cpp

@@ -98,7 +98,7 @@ static void write_logfile(
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
 static void sigint_handler(int signo) {
     if (signo == SIGINT) {
-        if (!is_interacting) {
+        if (!is_interacting && g_params->interactive) {
             is_interacting = true;
         } else {
             console::cleanup();
@@ -392,7 +392,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("\n");
     }
 
-    if (params.interactive) {
+    // ctrl+C handling
+    {
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
         struct sigaction sigint_action;
         sigint_action.sa_handler = sigint_handler;
@@ -405,7 +406,9 @@ int main(int argc, char ** argv) {
         };
         SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
+    }
 
+    if (params.interactive) {
         LOG_TEE("%s: interactive mode on.\n", __func__);
 
         if (!params.antiprompt.empty()) {

examples/server/README.md

@@ -185,7 +185,7 @@ node index.js
 
     `ignore_eos`: Ignore end of stream token and continue generating (default: false).
 
-    `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced (default: []).
+    `logit_bias`: Modify the likelihood of a token appearing in the generated text completion. For example, use `"logit_bias": [[15043,1.0]]` to increase the likelihood of the token 'Hello', or `"logit_bias": [[15043,-1.0]]` to decrease its likelihood. Setting the value to false, `"logit_bias": [[15043,false]]` ensures that the token `Hello` is never produced. The tokens can also be represented as strings, e.g. `[["Hello, World!",-0.5]]` will reduce the likelihood of all the individual tokens that represent the string `Hello, World!`, just like the `presence_penalty` does. (default: []).
 
     `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
 

examples/server/oai.hpp

@@ -15,9 +15,13 @@
 using json = nlohmann::json;
 
 inline static json oaicompat_completion_params_parse(
-    const json &body /* openai api json semantics */)
+    const json &body, /* openai api json semantics */
+    const std::string &chat_template)
 {
     json llama_params;
+    std::string formatted_prompt = chat_template == "chatml"
+        ? format_chatml(body["messages"])  // OpenAI 'messages' to chatml (with <|im_start|>,...)
+        : format_llama2(body["messages"]); // OpenAI 'messages' to llama2 (with [INST],...)
 
     llama_params["__oaicompat"] = true;
 
@@ -30,7 +34,7 @@ inline static json oaicompat_completion_params_parse(
     // https://platform.openai.com/docs/api-reference/chat/create
     llama_sampling_params default_sparams;
     llama_params["model"]             = json_value(body, "model", std::string("unknown"));
-    llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
+    llama_params["prompt"]            = formatted_prompt;
     llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
     llama_params["temperature"]       = json_value(body, "temperature", 0.0);
     llama_params["top_k"]             = json_value(body, "top_k", default_sparams.top_k);

examples/server/server.cpp

@@ -36,6 +36,7 @@ struct server_params
     std::string hostname = "127.0.0.1";
     std::vector<std::string> api_keys;
     std::string public_path = "examples/server/public";
+    std::string chat_template = "chatml";
     int32_t port = 8080;
     int32_t read_timeout = 600;
     int32_t write_timeout = 600;
@@ -625,18 +626,36 @@ struct llama_server_context
             const int n_vocab = llama_n_vocab(model);
             for (const auto &el : *logit_bias)
             {
-                if (el.is_array() && el.size() == 2 && el[0].is_number_integer())
+                if (el.is_array() && el.size() == 2)
+                {
+                    float bias;
+                    if (el[1].is_number())
+                    {
+                        bias = el[1].get<float>();
+                    }
+                    else if (el[1].is_boolean() && !el[1].get<bool>())
+                    {
+                        bias = -INFINITY;
+                    }
+                    else
+                    {
+                        continue;
+                    }
+
+                    if (el[0].is_number_integer())
                     {
                         llama_token tok = el[0].get<llama_token>();
                         if (tok >= 0 && tok < n_vocab)
                         {
-                        if (el[1].is_number())
+                            slot->sparams.logit_bias[tok] = bias;
-                        {
-                            slot->sparams.logit_bias[tok] = el[1].get<float>();
                         }
-                        else if (el[1].is_boolean() && !el[1].get<bool>())
+                    }
+                    else if (el[0].is_string())
                     {
-                            slot->sparams.logit_bias[tok] = -INFINITY;
+                        auto toks = llama_tokenize(model, el[0].get<std::string>(), false);
+                        for (auto tok : toks)
+                        {
+                            slot->sparams.logit_bias[tok] = bias;
                         }
                     }
                 }
@@ -1592,10 +1611,6 @@ struct llama_server_context
                         LOG_TEE("slot %d : in cache: %i tokens | to process: %i tokens\n", slot.id, slot.n_past, slot.num_prompt_tokens_processed);
                     }
 
-                    LOG_TEE("slot %d : kv cache rm - [%d, end)\n", slot.id, (int) system_tokens.size() + slot.n_past);
-
-                    llama_kv_cache_seq_rm(ctx, slot.id, system_tokens.size() + slot.n_past, -1);
-
                     slot.cache_tokens = prompt_tokens;
 
                     if (slot.n_past == slot.num_prompt_tokens && slot.n_past > 0)
@@ -1609,6 +1624,10 @@ struct llama_server_context
                         }
                     }
 
+                    LOG_TEE("slot %d : kv cache rm - [%d, end)\n", slot.id, (int) system_tokens.size() + slot.n_past);
+
+                    llama_kv_cache_seq_rm(ctx, slot.id, system_tokens.size() + slot.n_past, -1);
+
                     LOG_VERBOSE("prompt ingested", {
                                                     {"n_past",  slot.n_past},
                                                     {"cached",  tokens_to_str(ctx, slot.cache_tokens.cbegin(), slot.cache_tokens.cbegin() + slot.n_past)},
@@ -1859,6 +1878,8 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("                            types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
     printf("  -gan N, --grp-attn-n N    set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`");
     printf("  -gaw N, --grp-attn-w N    set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`");
+    printf("  --chat-template FORMAT_NAME");
+    printf("                            set chat template, possible valus is: llama2, chatml (default %s)", sparams.chat_template.c_str());
     printf("\n");
 }
 
@@ -2290,6 +2311,21 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             log_set_target(stdout);
             LOG_INFO("logging to file is disabled.", {});
         }
+        else if (arg == "--chat-template")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            std::string value(argv[i]);
+            if (value != "chatml" && value != "llama2") {
+                fprintf(stderr, "error: chat template can be \"llama2\" or \"chatml\", but got: %s\n", value.c_str());
+                invalid_param = true;
+                break;
+            }
+            sparams.chat_template = value;
+        }
         else if (arg == "--override-kv")
         {
             if (++i >= argc) {
@@ -2743,13 +2779,13 @@ int main(int argc, char **argv)
 
 
     // TODO: add mount point without "/v1" prefix -- how?
-    svr.Post("/v1/chat/completions", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
+    svr.Post("/v1/chat/completions", [&llama, &validate_api_key, &sparams](const httplib::Request &req, httplib::Response &res)
             {
                 res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 if (!validate_api_key(req, res)) {
                     return;
                 }
-                json data = oaicompat_completion_params_parse(json::parse(req.body));
+                json data = oaicompat_completion_params_parse(json::parse(req.body), sparams.chat_template);
 
                 const int task_id = llama.queue_tasks.get_new_id();
                 llama.queue_results.add_waiting_task_id(task_id);

examples/server/utils.hpp

@@ -167,6 +167,34 @@ static T json_value(const json &body, const std::string &key, const T &default_v
         : default_value;
 }
 
+inline std::string format_llama2(std::vector<json> messages)
+{
+    std::ostringstream output;
+    bool is_inside_turn = false;
+
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        if (!is_inside_turn) {
+            output << "[INST] ";
+        }
+        std::string role    = json_value(*it, "role", std::string("user"));
+        std::string content = json_value(*it, "content", std::string(""));
+        if (role == "system") {
+            output << "<<SYS>>\n" << content << "\n<<SYS>>\n\n";
+            is_inside_turn = true;
+        } else if (role == "user") {
+            output << content << " [/INST]";
+            is_inside_turn = true;
+        } else {
+            output << " " << content << " </s>";
+            is_inside_turn = false;
+        }
+    }
+
+    LOG_VERBOSE("format_llama2", {{"text", output.str()}});
+
+    return output.str();
+}
+
 inline std::string format_chatml(std::vector<json> messages)
 {
     std::ostringstream chatml_msgs;
@@ -180,6 +208,8 @@ inline std::string format_chatml(std::vector<json> messages)
 
     chatml_msgs << "<|im_start|>assistant" << '\n';
 
+    LOG_VERBOSE("format_chatml", {{"text", chatml_msgs.str()}});
+
     return chatml_msgs.str();
 }
 

ggml-backend.c

@@ -653,6 +653,9 @@ struct ggml_backend_cpu_context {
     int n_threads;
     void * work_data;
     size_t work_size;
+
+    ggml_abort_callback abort_callback;
+    void *              abort_callback_data;
 };
 
 GGML_CALL static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
@@ -691,6 +694,9 @@ GGML_CALL static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(gg
         cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
     }
 
+    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
+    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
     return cpu_plan;
 }
 
@@ -721,9 +727,11 @@ GGML_CALL static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, str
         cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size);
         cpu_ctx->work_size = cplan.work_size;
     }
-
     cplan.work_data = cpu_ctx->work_data;
 
+    cplan.abort_callback      = cpu_ctx->abort_callback;
+    cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
     ggml_graph_compute(cgraph, &cplan);
     return true;
 }
@@ -762,6 +770,8 @@ ggml_backend_t ggml_backend_cpu_init(void) {
     ctx->n_threads           = GGML_DEFAULT_N_THREADS;
     ctx->work_data           = NULL;
     ctx->work_size           = 0;
+    ctx->abort_callback      = NULL;
+    ctx->abort_callback_data = NULL;
 
     ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
 
@@ -783,6 +793,14 @@ void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
     ctx->n_threads = n_threads;
 }
 
+void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->abort_callback = abort_callback;
+    ctx->abort_callback_data = abort_callback_data;
+}
+
 GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
     return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size);
 }

ggml-backend.h

@@ -85,6 +85,7 @@ extern "C" {
 
     GGML_API GGML_CALL bool ggml_backend_is_cpu                (ggml_backend_t backend);
     GGML_API           void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_API           void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
 
     // Create a backend buffer from an existing pointer
     GGML_API GGML_CALL ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);

ggml-metal.m

@@ -687,6 +687,7 @@ static bool ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
 
+    @autoreleasepool {
     MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
     edesc.dispatchType = MTLDispatchTypeSerial;
 
@@ -2272,6 +2273,7 @@ static bool ggml_metal_graph_compute(
         [[MTLCaptureManager sharedCaptureManager] stopCapture];
     }
 
+    }
     return true;
 }
 

ggml-quants.c

@@ -49,6 +49,8 @@
 #define MIN(a, b) ((a) < (b) ? (a) : (b))
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
+#define UNUSED GGML_UNUSED
+
 #define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
 
 #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
@@ -268,6 +270,17 @@ static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128
 #endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
 
 #if defined(__ARM_NEON)
+
+#ifdef _MSC_VER
+
+#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+
+#else
+
+#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+
+#endif
+
 #if !defined(__aarch64__)
 
 // 64-bit compatibility
@@ -3666,15 +3679,92 @@ static inline __m128i get_scale_shuffle(int i) {
 }
 #endif
 
-void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
 
     assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q4_0 * restrict x = vx;
     const block_q8_0 * restrict y = vy;
 
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q4_0 * restrict vx0 = vx;
+        const block_q4_0 * restrict vx1 = vx + bx;
+
+        const block_q8_0 * restrict vy0 = vy;
+        const block_q8_0 * restrict vy1 = vy + by;
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q4_0 * restrict b_x0 = &vx0[i];
+            const block_q4_0 * restrict b_x1 = &vx1[i];
+            const block_q8_0 * restrict b_y0 = &vy0[i];
+            const block_q8_0 * restrict b_y1 = &vy1[i];
+
+            const uint8x16_t m4b = vdupq_n_u8(0x0F);
+            const int8x16_t  s8b = vdupq_n_s8(0x8);
+
+            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
+            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
+
+            // 4-bit -> 8-bit
+            const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+            const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+            const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+            const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+            // sub 8
+            const int8x16_t x0_l = vsubq_s8(v0_0l, s8b);
+            const int8x16_t x0_h = vsubq_s8(v0_0h, s8b);
+            const int8x16_t x1_l = vsubq_s8(v0_1l, s8b);
+            const int8x16_t x1_h = vsubq_s8(v0_1h, s8b);
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                 GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                                 GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                 GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+
+        vst1_f32(s, vget_low_f32(sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+        return;
+    }
+#endif
 #if defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
@@ -3956,15 +4046,93 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx,
 #endif
 }
 
-void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     const int qk = QK8_1;
     const int nb = n / qk;
 
     assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q4_1 * restrict x = vx;
     const block_q8_1 * restrict y = vy;
 
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q4_1 * restrict vx0 = vx;
+        const block_q4_1 * restrict vx1 = vx + bx;
+        const block_q8_1 * restrict vy0 = vy;
+        const block_q8_1 * restrict vy1 = vy + by;
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+        float32x4_t summs0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q4_1 * restrict b_x0 = &vx0[i];
+            const block_q4_1 * restrict b_x1 = &vx1[i];
+            const block_q8_1 * restrict b_y0 = &vy0[i];
+            const block_q8_1 * restrict b_y1 = &vy1[i];
+
+            float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * b_y0->s,
+                                   GGML_FP16_TO_FP32(b_x1->m) * b_y0->s,
+                                   GGML_FP16_TO_FP32(b_x0->m) * b_y1->s,
+                                   GGML_FP16_TO_FP32(b_x1->m) * b_y1->s};
+            summs0 += summs_t;
+
+            const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
+            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
+
+            // 4-bit -> 8-bit
+            const int8x16_t x0_l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+            const int8x16_t x0_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+            const int8x16_t x1_l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+            const int8x16_t x1_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            // mmla into int32x4_t
+            float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                 GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                                 GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                 GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+
+        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+        sumv2 = sumv2 + summs0;
+
+        vst1_f32(s, vget_low_f32(sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+        return;
+    }
+#endif
     // TODO: add WASM SIMD
 #if defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
@@ -4096,12 +4264,17 @@ void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restri
 #endif
 }
 
-void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
 
     assert(n % qk == 0);
     assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q5_0 * restrict x = vx;
     const block_q8_0 * restrict y = vy;
@@ -4382,12 +4555,17 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri
 #endif
 }
 
-void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     const int qk = QK8_1;
     const int nb = n / qk;
 
     assert(n % qk == 0);
     assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q5_1 * restrict x = vx;
     const block_q8_1 * restrict y = vy;
@@ -4681,15 +4859,79 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri
 #endif
 }
 
-void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;
 
     assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q8_0 * restrict x = vx;
     const block_q8_0 * restrict y = vy;
 
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q8_0 * restrict vx0 = vx;
+        const block_q8_0 * restrict vx1 = vx + bx;
+        const block_q8_0 * restrict vy0 = vy;
+        const block_q8_0 * restrict vy1 = vy + by;
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q8_0 * restrict b_x0 = &vx0[i];
+            const block_q8_0 * restrict b_y0 = &vy0[i];
+
+            const block_q8_0 * restrict b_x1 = &vx1[i];
+            const block_q8_0 * restrict b_y1 = &vy1[i];
+
+            const int8x16_t x0_l = vld1q_s8(b_x0->qs);
+            const int8x16_t x0_h = vld1q_s8(b_x0->qs + 16);
+            const int8x16_t x1_l = vld1q_s8(b_x1->qs);
+            const int8x16_t x1_h = vld1q_s8(b_x1->qs + 16);
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            float32x4_t scale = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                             GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                             GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                             GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                                                       l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+
+        vst1_f32(s, vget_low_f32(sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+        return;
+    }
+#endif
 #if defined(__ARM_NEON)
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
@@ -4784,7 +5026,12 @@ void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restri
 }
 
 #if QK_K == 256
-void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q2_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -5160,7 +5407,12 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 
 #else
 
-void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q2_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -5418,8 +5670,13 @@ void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restri
 #endif
 
 #if QK_K == 256
-void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const uint32_t kmask1 = 0x03030303;
     const uint32_t kmask2 = 0x0f0f0f0f;
@@ -5938,8 +6195,13 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 
 #else
 
-void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q3_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -6281,8 +6543,13 @@ void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restri
 #endif
 
 #if QK_K == 256
-void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q4_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -6637,8 +6904,13 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 #endif
 }
 #else
-void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q4_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -6880,8 +7152,13 @@ void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restri
 #endif
 
 #if QK_K == 256
-void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy,  size_t by, int nrc) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q5_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -7300,8 +7577,13 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
 #else
 
-void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q5_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -7566,8 +7848,13 @@ void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restri
 
 
 #if QK_K == 256
-void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q6_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -7998,8 +8285,13 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 
 #else
 
-void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_q6_K * restrict x = vx;
     const block_q8_K * restrict y = vy;
@@ -8328,8 +8620,13 @@ static const int8_t keven_signs_q2xs[1024] = {
      1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
 };
 
-void ggml_vec_dot_iq2_xxs_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_iq2_xxs * restrict x = vx;
     const block_q8_K    * restrict y = vy;
@@ -8451,8 +8748,13 @@ void ggml_vec_dot_iq2_xxs_q8_K(const int n, float * restrict s, const void * res
 #endif
 }
 
-void ggml_vec_dot_iq2_xs_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_iq2_xs * restrict x = vx;
     const block_q8_K   * restrict y = vy;
@@ -8671,8 +8973,13 @@ void ggml_vec_dot_iq2_xs_q8_K(const int n, float * restrict s, const void * rest
 }
 
 // TODO
-void ggml_vec_dot_iq3_xxs_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
     assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
 
     const block_iq3_xxs * restrict x = vx;
     const block_q8_K    * restrict y = vy;
@@ -8698,10 +9005,10 @@ void ggml_vec_dot_iq3_xxs_q8_K(const int n, float * restrict s, const void * res
         for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
             q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
             memcpy(aux32, gas, 2*sizeof(uint32_t)); gas += 2*sizeof(uint32_t);
-            const uint32x4_t aux32x4_0 = {iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]};
+            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]);
-            const uint32x4_t aux32x4_1 = {iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]};
+            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]);
-            const uint32x4_t aux32x4_2 = {iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]};
+            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]);
-            const uint32x4_t aux32x4_3 = {iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]};
+            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]);
             q3 += 16;
             q3s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >>  7) & 127))));
             q3s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >> 21) & 127))));

ggml-quants.h

@@ -245,20 +245,20 @@ void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_
 void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 
 // Dot product
-void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
-void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
 //
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")

ggml-vulkan.cpp

@@ -27,6 +27,7 @@
 #define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
 
 #define VK_VENDOR_ID_AMD 0x1002
+#define VK_VENDOR_ID_APPLE 0x106b
 #define VK_VENDOR_ID_INTEL 0x8086
 #define VK_VENDOR_ID_NVIDIA 0x10de
 
@@ -744,6 +745,8 @@ static vk_buffer ggml_vk_create_buffer(ggml_backend_vk_context * ctx, size_t siz
     }
 
     if (memory_type_index >= mem_props.memoryTypeCount) {
+        ctx->device.lock()->device.destroyBuffer(buf->buffer);
+        buf->size = 0;
         throw vk::OutOfDeviceMemoryError("No suitable memory type found");
     }
 
@@ -2032,18 +2035,100 @@ static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ct
     return ctx->pipeline_matmul_f32_aligned_l.align;
 }
 
-static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, int m, int n, bool aligned) {
+static vk_pipeline* ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, int m, int n, bool aligned) {
-#ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_guess_matmul_pipeline(" << bit16_x << ", " << bit16_y << ", " << m << ", " << n << ", " << aligned << ")";
-#endif
     if (bit16_x && bit16_y) {
-        if (ctx->device.lock()->vendor_id == VK_VENDOR_ID_INTEL || m <= 32 || n <= 32) {
+        if (m <= 32 || n <= 32) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " S" << std::endl;
 #endif
             return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s;
         }
-        if (ctx->device.lock()->subgroup_size == 64 || m <= 64 || n <= 64) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " M" << std::endl;
+#endif
+        return aligned ? &ctx->pipeline_matmul_f16_aligned_m : &ctx->pipeline_matmul_f16_m;
+    }
+    if (bit16_x && !bit16_y) {
+        if (m <= 32 || n <= 32) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " S" << std::endl;
+#endif
+            return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s;
+        }
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " M" << std::endl;
+#endif
+        return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_m : &ctx->pipeline_matmul_f16_f32_m;
+    }
+    if (!bit16_x && bit16_y) {
+        GGML_ASSERT(false);
+    }
+
+    if (m <= 32 || n <= 32) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " S" << std::endl;
+#endif
+        return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s;
+    }
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " M" << std::endl;
+#endif
+    return aligned ? &ctx->pipeline_matmul_f32_aligned_m : &ctx->pipeline_matmul_f32_m;
+}
+
+static vk_pipeline* ggml_vk_guess_matmul_pipeline_apple(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, bool aligned) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " M" << std::endl;
+#endif
+    if (bit16_x && bit16_y) {
+        return aligned ? &ctx->pipeline_matmul_f16_aligned_m : &ctx->pipeline_matmul_f16_m;
+    }
+    if (bit16_x && !bit16_y) {
+        return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_m : &ctx->pipeline_matmul_f16_f32_m;
+    }
+    if (!bit16_x && bit16_y) {
+        GGML_ASSERT(false);
+    }
+    return aligned ? &ctx->pipeline_matmul_f32_aligned_m : &ctx->pipeline_matmul_f32_m;
+}
+
+static vk_pipeline* ggml_vk_guess_matmul_pipeline_intel(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, bool aligned) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " S" << std::endl;
+#endif
+    if (bit16_x && bit16_y) {
+        return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s;
+    }
+    if (bit16_x && !bit16_y) {
+        return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s;
+    }
+    if (!bit16_x && bit16_y) {
+        GGML_ASSERT(false);
+    }
+    return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s;
+}
+
+static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, bool bit16_x, bool bit16_y, int m, int n, bool aligned) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << "ggml_vk_guess_matmul_pipeline(" << bit16_x << ", " << bit16_y << ", " << m << ", " << n << ", " << aligned << ")";
+#endif
+    switch (ctx->device.lock()->vendor_id) {
+    case VK_VENDOR_ID_AMD:
+        return ggml_vk_guess_matmul_pipeline_amd(ctx, bit16_x, bit16_y, m, n, aligned);
+    case VK_VENDOR_ID_APPLE:
+        return ggml_vk_guess_matmul_pipeline_apple(ctx, bit16_x, bit16_y, aligned);
+    case VK_VENDOR_ID_INTEL:
+        return ggml_vk_guess_matmul_pipeline_intel(ctx, bit16_x, bit16_y, aligned);
+    }
+
+    if (bit16_x && bit16_y) {
+        if (m <= 32 || n <= 32) {
+#ifdef GGML_VULKAN_DEBUG
+    std::cerr << " S" << std::endl;
+#endif
+            return aligned ? &ctx->pipeline_matmul_f16_aligned_s : &ctx->pipeline_matmul_f16_s;
+        }
+        if (m <= 64 || n <= 64) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " M" << std::endl;
 #endif
@@ -2055,13 +2140,13 @@ static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
         return aligned ? &ctx->pipeline_matmul_f16_aligned_l : &ctx->pipeline_matmul_f16_l;
     }
     if (bit16_x && !bit16_y) {
-        if (ctx->device.lock()->vendor_id == VK_VENDOR_ID_INTEL || m <= 32 || n <= 32) {
+        if (m <= 32 || n <= 32) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " S" << std::endl;
 #endif
             return aligned ? &ctx->pipeline_matmul_f16_f32_aligned_s : &ctx->pipeline_matmul_f16_f32_s;
         }
-        if (ctx->device.lock()->subgroup_size == 64 || m <= 64 || n <= 64) {
+        if (m <= 64 || n <= 64) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " M" << std::endl;
 #endif
@@ -2076,13 +2161,13 @@ static vk_pipeline* ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx,
         GGML_ASSERT(false);
     }
 
-    if (ctx->device.lock()->vendor_id == VK_VENDOR_ID_INTEL || m <= 32 || n <= 32) {
+    if (m <= 32 || n <= 32) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " S" << std::endl;
 #endif
         return aligned ? &ctx->pipeline_matmul_f32_aligned_s : &ctx->pipeline_matmul_f32_s;
     }
-    if (ctx->device.lock()->subgroup_size == 64 || m <= 64 || n <= 64) {
+    if (m <= 64 || n <= 64) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << " M" << std::endl;
 #endif
@@ -3875,7 +3960,7 @@ static ggml_tensor * ggml_vk_find_last_use(const ggml_tensor * node, ggml_cgraph
 
 static void ggml_vk_preallocate_buffers_graph(ggml_backend_vk_context * ctx, ggml_tensor * node){
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_ctx->preallocate_buffers_graph(" << node << ")" << std::endl;
+    std::cerr << "ggml_vk_preallocate_buffers_graph(" << node << ")" << std::endl;
 #endif
     const bool any_on_device = node->backend == GGML_BACKEND_GPU
         || (node->src[0] != nullptr && (node->src[0]->backend == GGML_BACKEND_GPU || node->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
@@ -3994,8 +4079,7 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
         return;
     }
 #ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_ctx->preallocate_buffers()" << std::endl;
+    std::cerr << "ggml_vk_preallocate_buffers(qx_size: " << ctx->prealloc_size_qx << " qy_size: " << ctx->prealloc_size_qy << " x_size: " << ctx->prealloc_size_x << " y_size: " << ctx->prealloc_size_y << " split_k_size: " << ctx->prealloc_size_split_k << ")" << std::endl;
-    std::cerr << "qx_size: " << ctx->prealloc_size_qx << " qy_size: " << ctx->prealloc_size_qy << " x_size: " << ctx->prealloc_size_x << " y_size: " << ctx->prealloc_size_y << " split_k_size: " << ctx->prealloc_size_split_k << std::endl;
 #endif
 #if defined(GGML_VULKAN_RUN_TESTS)
     ctx->staging = ggml_vk_create_buffer_check(ctx, 100ul * 1024ul * 1024ul, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached);

ggml.c

@@ -428,8 +428,8 @@ int64_t ggml_cycles_per_ms(void) {
 
 static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 
-static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y);
+static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc);
-static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y);
+static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc);
 
 static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
     [GGML_TYPE_I8] = {
@@ -457,6 +457,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = false,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
         .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
     },
     [GGML_TYPE_F16] = {
         .type_name                = "f16",
@@ -468,6 +469,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t) ggml_fp32_to_fp16_row,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f16,
         .vec_dot_type             = GGML_TYPE_F16,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q4_0] = {
         .type_name                = "q4_0",
@@ -479,6 +481,11 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t) quantize_row_q4_0_reference,
         .vec_dot                  = ggml_vec_dot_q4_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
+#if defined (__ARM_FEATURE_MATMUL_INT8)
+        .nrows                    = 2,
+#else
+        .nrows                    = 1,
+#endif
     },
     [GGML_TYPE_Q4_1] = {
         .type_name                = "q4_1",
@@ -490,6 +497,11 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t) quantize_row_q4_1_reference,
         .vec_dot                  = ggml_vec_dot_q4_1_q8_1,
         .vec_dot_type             = GGML_TYPE_Q8_1,
+#if defined (__ARM_FEATURE_MATMUL_INT8)
+        .nrows                    = 2,
+#else
+        .nrows                    = 1,
+#endif
     },
     [4] = { // GGML_TYPE_Q4_2
         .type_name                = "DEPRECATED",
@@ -501,6 +513,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = NULL,
         .vec_dot                  = NULL,
         .vec_dot_type             = GGML_TYPE_COUNT,
+        .nrows                    = 1,
     },
     [5] = { // GGML_TYPE_Q4_3
         .type_name                = "DEPRECATED",
@@ -512,6 +525,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = NULL,
         .vec_dot                  = NULL,
         .vec_dot_type             = GGML_TYPE_COUNT,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q5_0] = {
         .type_name                = "q5_0",
@@ -523,6 +537,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t) quantize_row_q5_0_reference,
         .vec_dot                  = ggml_vec_dot_q5_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q5_1] = {
         .type_name                = "q5_1",
@@ -534,6 +549,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t) quantize_row_q5_1_reference,
         .vec_dot                  = ggml_vec_dot_q5_1_q8_1,
         .vec_dot_type             = GGML_TYPE_Q8_1,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q8_0] = {
         .type_name                = "q8_0",
@@ -545,6 +561,11 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t) quantize_row_q8_0_reference,
         .vec_dot                  = ggml_vec_dot_q8_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
+#if defined (__ARM_FEATURE_MATMUL_INT8)
+        .nrows                    = 2,
+#else
+        .nrows                    = 1,
+#endif
     },
     [GGML_TYPE_Q8_1] = {
         .type_name                = "q8_1",
@@ -554,6 +575,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float               = quantize_row_q8_1,
         .from_float_reference     = (ggml_from_float_t) quantize_row_q8_1_reference,
         .vec_dot_type             = GGML_TYPE_Q8_1,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q2_K] = {
         .type_name                = "q2_K",
@@ -565,6 +587,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t) quantize_row_q2_K_reference,
         .vec_dot                  = ggml_vec_dot_q2_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q3_K] = {
         .type_name                = "q3_K",
@@ -576,6 +599,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t) quantize_row_q3_K_reference,
         .vec_dot                  = ggml_vec_dot_q3_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q4_K] = {
         .type_name                = "q4_K",
@@ -587,6 +611,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t) quantize_row_q4_K_reference,
         .vec_dot                  = ggml_vec_dot_q4_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q5_K] = {
         .type_name                = "q5_K",
@@ -598,6 +623,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t) quantize_row_q5_K_reference,
         .vec_dot                  = ggml_vec_dot_q5_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q6_K] = {
         .type_name                = "q6_K",
@@ -609,6 +635,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t) quantize_row_q6_K_reference,
         .vec_dot                  = ggml_vec_dot_q6_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_IQ2_XXS] = {
         .type_name                = "iq2_xxs",
@@ -620,6 +647,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = NULL,
         .vec_dot                  = ggml_vec_dot_iq2_xxs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_IQ2_XS] = {
         .type_name                = "iq2_xs",
@@ -631,6 +659,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = NULL,
         .vec_dot                  = ggml_vec_dot_iq2_xs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_IQ3_XXS] = {
         .type_name                = "iq3_xxs",
@@ -642,6 +671,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .from_float_reference     = (ggml_from_float_t)quantize_row_iq3_xxs_reference,
         .vec_dot                  = ggml_vec_dot_iq3_xxs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q8_K] = {
         .type_name                = "q8_K",
@@ -1212,7 +1242,13 @@ inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x)
 inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]*y[i];   }
 inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]/y[i];   }
 
-static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y) {
+static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc) {
+   assert(nrc == 1);
+   UNUSED(nrc);
+   UNUSED(bx);
+   UNUSED(by);
+   UNUSED(bs);
+
 #ifdef GGML_SIMD
     float sumf = 0.0f;
     const int np = (n & ~(GGML_F32_STEP - 1));
@@ -1249,7 +1285,13 @@ static void ggml_vec_dot_f32(const int n, float * restrict s, const float * rest
     *s = sumf;
 }
 
-static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y) {
+static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
     ggml_float sumf = 0.0;
 
 #if defined(GGML_SIMD)
@@ -1455,7 +1497,7 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 #endif
 }
 
-inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, x, x); *s = sqrtf(*s);   }
+inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, 0, x, 0, x, 0, 1); *s = sqrtf(*s);   }
 inline static void ggml_vec_sqr_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i];   }
 inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); }
 inline static void ggml_vec_log_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]);   }
@@ -9992,6 +10034,7 @@ static void ggml_compute_forward_mul_mat(
     ggml_vec_dot_t    const vec_dot               = type_traits[type].vec_dot;
     enum ggml_type    const vec_dot_type          = type_traits[type].vec_dot_type;
     ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float;
+    int64_t           const vec_dot_num_rows      = type_traits[type].nrows;
 
     GGML_ASSERT(ne0 == ne01);
     GGML_ASSERT(ne1 == ne11);
@@ -10159,12 +10202,23 @@ static void ggml_compute_forward_mul_mat(
     const int64_t blck_0 = 16;
     const int64_t blck_1 = 16;
 
+    // dot kernels can handle 1 row and col at a time, but mmla kernels can process 2 rows and cols
+    int64_t nrc = vec_dot_num_rows;
+    // TODO: currently the mmla kernels support only even numbered rows/cols.
+    // this check can be removed once they are extended to support odd numbered rows/cols too
+    if ((nr0 % 2 != 0) || (ne11 % 2 != 0)) {
+        nrc = 1;
+    }
+
+    const size_t src1_col_stride = src1_cont || src1->type != vec_dot_type ? row_size : nb11;
+
     // attempt to reduce false-sharing (does not seem to make a difference)
-    float tmp[16];
+    // 16 * 2, accounting for mmla kernels
+    float tmp[32];
 
     for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
         for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
-            for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
+            for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ir1 += nrc) {
                 const int64_t i13 = (ir1/(ne12*ne1));
                 const int64_t i12 = (ir1 - i13*ne12*ne1)/ne1;
                 const int64_t i11 = (ir1 - i13*ne12*ne1 - i12*ne1);
@@ -10187,17 +10241,19 @@ static void ggml_compute_forward_mul_mat(
                     (src1_cont || src1->type != vec_dot_type
                      ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
                      : (i11*nb11 + i12*nb12 + i13*nb13));
-
                 float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
 
                 //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
                 //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
                 //}
 
-                for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
+                for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ir0 += nrc) {
-                    vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
+                    vec_dot(ne00, &tmp[ir0 - iir0], (nrc>1 ? 16 : 0), src0_row + ir0*nb01, (nrc>1 ? nb01 : 0), src1_col, (nrc>1 ? src1_col_stride : 0), nrc);
+                }
+
+                for (int cn = 0; cn < nrc; ++cn) {
+                    memcpy(&dst_col[iir0 + cn*nb1/nb0], tmp + (cn*16), (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
                 }
-                memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
             }
         }
     }
@@ -10386,7 +10442,7 @@ static void ggml_compute_forward_mul_mat_id(
                     //}
 
                     for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
-                        vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
+                        vec_dot(ne00, &tmp[ir0 - iir0], 0, src0_row + ir0*nb01, 0, src1_col, 0, 1);
                     }
                     memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
                 }
@@ -11568,7 +11624,7 @@ static void ggml_compute_forward_soft_max_back_f32(
 
         // linear runtime, no additional memory
         float dot_y_dy = 0;
-        ggml_vec_dot_f32 (nc, &dot_y_dy, y, dy);
+        ggml_vec_dot_f32 (nc, &dot_y_dy, 0, y, 0, dy, 0, 1);
         ggml_vec_cpy_f32 (nc, dx, dy);
         ggml_vec_acc1_f32(nc, dx, -dot_y_dy);
         ggml_vec_mul_f32 (nc, dx, dx, y);
@@ -12369,9 +12425,9 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32(
             const int i1n = i10*ne11;
             for (int i00 = 0; i00 < ne00; i00++) {
                 float v = 0;
-                ggml_vec_dot_f16(ne02, &v,
+                ggml_vec_dot_f16(ne02, &v, 0,
-                        (ggml_fp16_t *)    wdata_src + i1n,
+                        (ggml_fp16_t *)    wdata_src + i1n, 0,
-                        (ggml_fp16_t *) wdata_kernel + i00*ne02);
+                        (ggml_fp16_t *) wdata_kernel + i00*ne02, 0, 1);
                 dst_data[i10*s0 + i00] += v;
             }
         }
@@ -12466,9 +12522,9 @@ static void ggml_compute_forward_conv_transpose_1d_f32(
             const int i1n = i10*ne11;
             for (int i00 = 0; i00 < ne00; i00++) {
                 float v = 0;
-                ggml_vec_dot_f32(ne02, &v,
+                ggml_vec_dot_f32(ne02, &v, 0,
-                        wdata_src + i1n,
+                        wdata_src + i1n, 0,
-                        wdata_kernel + i00*ne02);
+                        wdata_kernel + i00*ne02, 0, 1);
                 dst_data[i10*s0 + i00] += v;
             }
         }
@@ -12783,9 +12839,9 @@ static void ggml_compute_forward_conv_transpose_2d(
                 for (int i01 = 0; i01 < ne01; i01++) {
                     for (int i00 = 0; i00 < ne00; i00++) {
                         float v = 0;
-                        ggml_vec_dot_f16(ne03, &v,
+                        ggml_vec_dot_f16(ne03, &v, 0,
-                                wdata_src + i1n,
+                                wdata_src + i1n, 0,
-                                wdata_kernel + i01*ne00*ne03 + i00*ne03);
+                                wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
                         dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v;
                     }
                 }
@@ -13214,9 +13270,9 @@ static void ggml_compute_forward_flash_attn_f32(
             const int i1 = ik1;
 
             ggml_vec_dot_f32(neq0,
-                    S + i1,
+                    S + i1, 0,
-                    (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
+                    (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)), 0,
-                    (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
+                    (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), 0, 1);
         }
 
         // scale
@@ -13299,9 +13355,9 @@ static void ggml_compute_forward_flash_attn_f32(
             const int iv3 = iq3;
 
             ggml_vec_dot_f32(masked_begin,
-                    (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)),
+                    (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)), 0,
-                    (float *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)),
+                    (float *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)), 0,
-                    S);
+                    S, 0, 1);
         }
     }
 }
@@ -13404,9 +13460,9 @@ static void ggml_compute_forward_flash_attn_f16(
                 const int i1 = ik1;
 
                 ggml_vec_dot_f16(neq0,
-                        S + i1,
+                        S + i1, 0,
-                        (ggml_fp16_t *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
+                        (ggml_fp16_t *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)), 0,
-                        (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
+                        (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), 0, 1);
             }
         } else {
             for (int64_t ic = 0; ic < nek1; ic += GGML_VEC_DOT_UNROLL) {
@@ -13508,9 +13564,9 @@ static void ggml_compute_forward_flash_attn_f16(
                 const int iv3 = iq3;
 
                 ggml_vec_dot_f16(nev0,
-                        (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)),
+                        (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)), 0,
-                        (ggml_fp16_t *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)),
+                        (ggml_fp16_t *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)), 0,
-                        S16);
+                        S16, 0, 1);
             }
         } else {
             for (int64_t ic = 0; ic < nev1; ic += GGML_VEC_DOT_UNROLL) {
@@ -13652,9 +13708,9 @@ static void ggml_compute_forward_flash_ff_f16(
             const int i1 = ib01;
 
             ggml_vec_dot_f16(nea0,
-                    S + i1,
+                    S + i1, 0,
-                    (ggml_fp16_t *) ((char *) b0->data + (ib01*nbb01 + ib02*nbb02 + ib03*nbb03)),
+                    (ggml_fp16_t *) ((char *) b0->data + (ib01*nbb01 + ib02*nbb02 + ib03*nbb03)), 0,
-                    (ggml_fp16_t *) ((char *)  a->data + ( ia1*nba1  +  ia2*nba2  +  ia3*nba3)));
+                    (ggml_fp16_t *) ((char *)  a->data + ( ia1*nba1  +  ia2*nba2  +  ia3*nba3)), 0, 1);
         }
 
         ggml_vec_add_f32(neb01, S, S, (float *) b1->data);
@@ -13677,9 +13733,9 @@ static void ggml_compute_forward_flash_ff_f16(
             for (int64_t ic = 0; ic < nec01; ++ic) {
 
                 ggml_vec_dot_f16(neb01,
-                        (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1   + i2*nb2   + i3*nb3)),
+                        (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1   + i2*nb2   + i3*nb3)), 0,
-                        (ggml_fp16_t *) ((char *) c0->data  + (         ic*nbc01 + i2*nbc02 + i3*nbc03)),
+                        (ggml_fp16_t *) ((char *) c0->data  + (         ic*nbc01 + i2*nbc02 + i3*nbc03)), 0,
-                        S16);
+                        S16, 0, 1);
             }
 
             ggml_vec_add_f32(nec01,
@@ -13866,9 +13922,9 @@ static void ggml_compute_forward_flash_attn_back_f32(
                     const int i1 = ik1;
 
                     ggml_vec_dot_f32(neq0,
-                            S + i1,
+                            S + i1, 0,
-                            (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
+                            (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)), 0,
-                            (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
+                            (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), 0, 1);
                 }
 
                 // scale
@@ -14013,7 +14069,7 @@ static void ggml_compute_forward_flash_attn_back_f32(
 
                 // S = SM * (S - dot(SM, S))
                 float dot_SM_gradSM = 0;
-                ggml_vec_dot_f32 (masked_begin, &dot_SM_gradSM, SM, S);
+                ggml_vec_dot_f32 (masked_begin, &dot_SM_gradSM, 0, SM, 0, S, 0, 1);
                 ggml_vec_acc1_f32(M, S, -dot_SM_gradSM);
                 ggml_vec_mul_f32 (masked_begin, S, S, SM);
 
@@ -16649,7 +16705,7 @@ struct ggml_compute_state_shared {
     atomic_int node_n;    // active graph node
     atomic_int node_task; // active graph node task phase
 
-    bool (*abort_callback)(void * data); // abort ggml_graph_compute when true
+    ggml_abort_callback abort_callback; // abort ggml_graph_compute when true
     void * abort_callback_data;
 };
 
@@ -18382,7 +18438,7 @@ static enum ggml_opt_result linesearch_backtracking(
     }
 
     // compute the initial gradient in the search direction
-    ggml_vec_dot_f32(nx, &dginit, g, d);
+    ggml_vec_dot_f32(nx, &dginit, 0, g, 0, d, 0, 1);
 
     // make sure that d points to a descent direction
     if (0 < dginit) {
@@ -18432,7 +18488,7 @@ static enum ggml_opt_result linesearch_backtracking(
                 return count;
             }
 
-            ggml_vec_dot_f32(nx, &dg, g, d);
+            ggml_vec_dot_f32(nx, &dg, 0, g, 0, d, 0, 1);
 
             // check the Wolfe condition
             if (dg < params->lbfgs.wolfe * dginit) {
@@ -18693,8 +18749,8 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         //     ys = y^t \cdot s    -> 1 / \rho.
         //     yy = y^t \cdot y.
         //
-        ggml_vec_dot_f32(nx, &ys, &lm_y[end[0]*nx], &lm_s[end[0]*nx]);
+        ggml_vec_dot_f32(nx, &ys, 0, &lm_y[end[0]*nx], 0, &lm_s[end[0]*nx], 0, 1);
-        ggml_vec_dot_f32(nx, &yy, &lm_y[end[0]*nx], &lm_y[end[0]*nx]);
+        ggml_vec_dot_f32(nx, &yy, 0, &lm_y[end[0]*nx], 0, &lm_y[end[0]*nx], 0, 1);
 
         lm_ys[end[0]] = ys;
 
@@ -18713,7 +18769,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         for (int i = 0; i < bound; ++i) {
             j[0] = (j[0] + m - 1) % m;
             // \alpha_{j} = \rho_{j} s^{t}_{j} \cdot q_{k+1}
-            ggml_vec_dot_f32(nx, &lm_alpha[j[0]], &lm_s[j[0]*nx], d);
+            ggml_vec_dot_f32(nx, &lm_alpha[j[0]], 0, &lm_s[j[0]*nx], 0, d, 0, 1);
             lm_alpha[j[0]] /= lm_ys[j[0]];
             // q_{i} = q_{i+1} - \alpha_{i} y_{i}
             ggml_vec_mad_f32(nx, d, &lm_y[j[0]*nx], -lm_alpha[j[0]]);
@@ -18723,7 +18779,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
 
         for (int i = 0; i < bound; ++i) {
             // \beta_{j} = \rho_{j} y^t_{j} \cdot \gamma_{i}
-            ggml_vec_dot_f32(nx, &beta, &lm_y[j[0]*nx], d);
+            ggml_vec_dot_f32(nx, &beta, 0, &lm_y[j[0]*nx], 0, d, 0, 1);
             beta /= lm_ys[j[0]];
             // \gamma_{i+1} = \gamma_{i} + (\alpha_{j} - \beta_{j}) s_{j}
             ggml_vec_mad_f32(nx, d, &lm_s[j[0]*nx], lm_alpha[j[0]] - beta);
@@ -20611,4 +20667,12 @@ int ggml_cpu_has_vsx(void) {
 #endif
 }
 
+int ggml_cpu_has_matmul_int8(void) {
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 ////////////////////////////////////////////////////////////////////////////////

ggml.h

@@ -567,6 +567,11 @@ extern "C" {
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
 
+    // Abort callback
+    // If not NULL, called before ggml computation
+    // If it returns true, the computation is aborted
+    typedef bool (*ggml_abort_callback)(void * data);
+
     // the compute plan that needs to be prepared for ggml_graph_compute()
     // since https://github.com/ggerganov/ggml/issues/287
     struct ggml_cplan {
@@ -576,7 +581,7 @@ extern "C" {
         int n_threads;
 
         // abort ggml_graph_compute when true
-        bool (*abort_callback)(void * data);
+        ggml_abort_callback abort_callback;
         void *              abort_callback_data;
     };
 
@@ -2273,6 +2278,7 @@ extern "C" {
     GGML_API int ggml_cpu_has_ssse3      (void);
     GGML_API int ggml_cpu_has_sycl       (void);
     GGML_API int ggml_cpu_has_vsx        (void);
+    GGML_API int ggml_cpu_has_matmul_int8(void);
 
     //
     // Internal types and functions exposed for tests and benchmarks
@@ -2286,7 +2292,8 @@ extern "C" {
 #endif
     typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
     typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int k);
-    typedef void (*ggml_vec_dot_t)   (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
+    typedef void (*ggml_vec_dot_t)   (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
+                                      const void * GGML_RESTRICT y, size_t by, int nrc);
 
     typedef struct {
         const char      * type_name;
@@ -2298,6 +2305,7 @@ extern "C" {
         ggml_from_float_t from_float_reference;
         ggml_vec_dot_t    vec_dot;
         enum ggml_type    vec_dot_type;
+        int64_t           nrows; // number of rows to process simultaneously;
     } ggml_type_traits_t;
 
     GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);

ggml_vk_generate_shaders.py

@@ -2067,6 +2067,8 @@ type_names = {
 
 K_QUANTS_PER_ITERATION = 2
 
+ASYNCIO_CONCURRENCY = 64
+
 output_dir = gettempdir()
 
 lock = asyncio.Lock()
@@ -2291,7 +2293,14 @@ async def main():
     tasks.append(string_to_spv("rope_neox_f32", rope_neox_src, {"A_TYPE": "float", "D_TYPE": "float"}))
     tasks.append(string_to_spv("rope_neox_f16", rope_neox_src, {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
 
-    await asyncio.gather(*tasks)
+    # Helper to decorate tasks with semaphore acquisition.
+    async def withSemaphore(sem, task):
+        async with sem:
+            return await task
+
+    # Run tasks concurrently guarded by a concurrency limit.
+    sem = asyncio.Semaphore(ASYNCIO_CONCURRENCY)
+    await asyncio.gather(*(withSemaphore(sem, task) for task in tasks))
 
     with open("ggml-vulkan-shaders.hpp", "w") as f:
         f.write("#include <cstdint>\n\n")

llama.cpp

@@ -7285,7 +7285,9 @@ static int llama_decode_internal(
     // TODO: this is mostly important for Apple Silicon where CBLAS is still performing very well
     //       we still need some threads to process all non-mul_mat ops, but not too much to avoid interfering
     //       with the BLAS calls. need a better solution
-    if (n_tokens >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas()) {
+    // MoE Special Case: This logic applies when hparams.n_expert == 0, i.e. the model is NOT an MoE model. When an MoE is
+    //                   being processed then Accelerate/BLAS will not be involved, so capping would limit performance.
+    if (n_tokens >= 32 && hparams.n_expert == 0 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas()) {
         n_threads = std::min(4, n_threads);
     }
 
@@ -11867,6 +11869,7 @@ const char * llama_print_system_info(void) {
     s += "SSE3 = "        + std::to_string(ggml_cpu_has_sse3())        + " | ";
     s += "SSSE3 = "       + std::to_string(ggml_cpu_has_ssse3())       + " | ";
     s += "VSX = "         + std::to_string(ggml_cpu_has_vsx())         + " | ";
+    s += "MATMUL_INT8 = " + std::to_string(ggml_cpu_has_matmul_int8()) + " | ";
 
     return s.c_str();
 }

pocs/vdot/q8dot.cpp

@@ -156,8 +156,8 @@ int main(int argc, char** argv) {
 
         t1 = std::chrono::high_resolution_clock::now();
         float fs;
-        if (type == 0) funcs.vec_dot(kVecSize * QK4_1, &fs, x40.data(), y.data());
+        if (type == 0) funcs.vec_dot(kVecSize * QK4_1, &fs, 0, x40.data(), 0, y.data(), 0, 1);
-        else funcs.vec_dot(kVecSize * QK4_1, &fs, x41.data(), y.data());
+        else funcs.vec_dot(kVecSize * QK4_1, &fs, 0, x41.data(), 0, y.data(), 0, 1);
         t2 = std::chrono::high_resolution_clock::now();
         t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
         if (iloop > 3) ggml.addResult(fs, t);

pocs/vdot/vdot.cpp

@@ -284,8 +284,8 @@ int main(int argc, char** argv) {
         else {
             auto vdot = ggml_internal_get_type_traits(funcs.vec_dot_type);
             vdot.from_float(y1.data(), q8.data(), kVecSize);
-            if (useQ4_1) funcs.vec_dot(kVecSize, &result, q41.data(), q8.data());
+            if (useQ4_1) funcs.vec_dot(kVecSize, &result, 0, q41.data(), 0, q8.data(), 0, 1);
-            else funcs.vec_dot(kVecSize, &result, q40.data(), q8.data());
+            else funcs.vec_dot(kVecSize, &result, 0, q40.data(), 0, q8.data(), 0, 1);
         }
         sumq += result;
         t2 = std::chrono::high_resolution_clock::now();

scripts/sync-ggml-am.sh

@@ -97,6 +97,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-cuda.cu            -> ggml-cuda.cu
     # src/ggml-cuda.h             -> ggml-cuda.h
     # src/ggml-impl.h             -> ggml-impl.h
+    # src/ggml-kompute.cpp        -> ggml-kompute.cpp
+    # src/ggml-kompute.h          -> ggml-kompute.h
     # src/ggml-metal.h            -> ggml-metal.h
     # src/ggml-metal.m            -> ggml-metal.m
     # src/ggml-mpi.h              -> ggml-mpi.h
@@ -105,6 +107,10 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
     # src/ggml-opencl.h           -> ggml-opencl.h
     # src/ggml-quants.c           -> ggml-quants.c
     # src/ggml-quants.h           -> ggml-quants.h
+    # src/ggml-sycl.cpp           -> ggml-sycl.cpp
+    # src/ggml-sycl.h             -> ggml-sycl.h
+    # src/ggml-vulkan.cpp         -> ggml-vulkan.cpp
+    # src/ggml-vulkan.h           -> ggml-vulkan.h
     # include/ggml/ggml.h         -> ggml.h
     # include/ggml/ggml-alloc.h   -> ggml-alloc.h
     # include/ggml/ggml-backend.h -> ggml-backend.h
@@ -123,6 +129,8 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/src\/ggml-cuda\.cu/ggml-cuda.cu/g' \
         -e 's/src\/ggml-cuda\.h/ggml-cuda.h/g' \
         -e 's/src\/ggml-impl\.h/ggml-impl.h/g' \
+        -e 's/src\/ggml-kompute\.cpp/ggml-kompute.cpp/g' \
+        -e 's/src\/ggml-kompute\.h/ggml-kompute.h/g' \
         -e 's/src\/ggml-metal\.h/ggml-metal.h/g' \
         -e 's/src\/ggml-metal\.m/ggml-metal.m/g' \
         -e 's/src\/ggml-mpi\.h/ggml-mpi.h/g' \
@@ -131,6 +139,10 @@ if [ -f $SRC_LLAMA/ggml-src.patch ]; then
         -e 's/src\/ggml-opencl\.h/ggml-opencl.h/g' \
         -e 's/src\/ggml-quants\.c/ggml-quants.c/g' \
         -e 's/src\/ggml-quants\.h/ggml-quants.h/g' \
+        -e 's/src\/ggml-sycl\.cpp/ggml-sycl.cpp/g' \
+        -e 's/src\/ggml-sycl\.h/ggml-sycl.h/g' \
+        -e 's/src\/ggml-vulkan\.cpp/ggml-vulkan.cpp/g' \
+        -e 's/src\/ggml-vulkan\.h/ggml-vulkan.h/g' \
         -e 's/include\/ggml\/ggml\.h/ggml.h/g' \
         -e 's/include\/ggml\/ggml-alloc\.h/ggml-alloc.h/g' \
         -e 's/include\/ggml\/ggml-backend\.h/ggml-backend.h/g' \

scripts/sync-ggml.last

@@ -1 +1 @@
-475cbad5c1c834e31e26a2283bc1413181644360
+2c7cf49810d523b9632da393a9e8270b60bf3b24

scripts/sync-ggml.sh

@@ -7,6 +7,8 @@ cp -rpv ../ggml/src/ggml-backend.c          ./ggml-backend.c
 cp -rpv ../ggml/src/ggml-cuda.cu            ./ggml-cuda.cu
 cp -rpv ../ggml/src/ggml-cuda.h             ./ggml-cuda.h
 cp -rpv ../ggml/src/ggml-impl.h             ./ggml-impl.h
+cp -rpv ../ggml/src/ggml-kompute.cpp        ./ggml-kompute.cpp
+cp -rpv ../ggml/src/ggml-kompute.h          ./ggml-kompute.h
 cp -rpv ../ggml/src/ggml-metal.h            ./ggml-metal.h
 cp -rpv ../ggml/src/ggml-metal.m            ./ggml-metal.m
 cp -rpv ../ggml/src/ggml-metal.metal        ./ggml-metal.metal
@@ -16,6 +18,10 @@ cp -rpv ../ggml/src/ggml-opencl.cpp         ./ggml-opencl.cpp
 cp -rpv ../ggml/src/ggml-opencl.h           ./ggml-opencl.h
 cp -rpv ../ggml/src/ggml-quants.c           ./ggml-quants.c
 cp -rpv ../ggml/src/ggml-quants.h           ./ggml-quants.h
+cp -rpv ../ggml/src/ggml-sycl.cpp           ./ggml-sycl.cpp
+cp -rpv ../ggml/src/ggml-sycl.h             ./ggml-sycl.h
+cp -rpv ../ggml/src/ggml-vulkan.cpp         ./ggml-vulkan.cpp
+cp -rpv ../ggml/src/ggml-vulkan.h           ./ggml-vulkan.h
 cp -rpv ../ggml/include/ggml/ggml.h         ./ggml.h
 cp -rpv ../ggml/include/ggml/ggml-alloc.h   ./ggml-alloc.h
 cp -rpv ../ggml/include/ggml/ggml-backend.h ./ggml-backend.h

tests/test-quantize-fns.cpp

@@ -87,7 +87,7 @@ static float dot_product_error(
     vdot.from_float(test_data2, tmp_q2.data(), test_size);
 
     float result = INFINITY;
-    qfns.vec_dot(test_size, &result, tmp_q1.data(), tmp_q2.data());
+    qfns.vec_dot(test_size, &result, 0, tmp_q1.data(), 0, tmp_q2.data(), 0, 1);
 
     const float dot_ref = dot_product(test_data1, test_data2, test_size);
 

tests/test-quantize-perf.cpp

@@ -346,7 +346,7 @@ int main(int argc, char * argv[]) {
                     printf("    %zu values (%.2f MB)\n", size, 4*size/(float)(1024*1024));
                     auto quantize_fn = [&](void) -> float {
                         float result;
-                        qfns.vec_dot(size, &result, test_q1, test_q2);
+                        qfns.vec_dot(size, &result, 0, test_q1, 0, test_q2, 0, 1);
                         return result;
                     };
                     size_t quantized_size = ggml_row_size(type, size);