llama : add support for Nomic Embed (#5468)

* common : don't crash if newline token is not found

* common : llama_byte_to_token: allow falling back to finding just the token byte in SPM vocabs

.flake8

@@ -1,2 +1,3 @@
 [flake8]
 max-line-length = 125
+ignore = W503

.github/workflows/python-lint.yml

@@ -16,5 +16,5 @@ jobs:
       - name: flake8 Lint
         uses: py-actions/flake8@v2
         with:
-            ignore: "E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704"
+            ignore: "E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503"
             exclude: "examples/*,examples/*/**,*/**/__init__.py"

Makefile

@@ -569,6 +569,14 @@ $(info I CC:        $(shell $(CC)   --version | head -n 1))
 $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef LLAMA_CUBLAS
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
+CUDA_VERSION := $(shell nvcc --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
+ifndef CUDA_DOCKER_ARCH
+ifndef CUDA_POWER_ARCH
+$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via CUDA_DOCKER_ARCH)
+endif # CUDA_POWER_ARCH
+endif # CUDA_DOCKER_ARCH
+endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
 endif # LLAMA_CUBLAS
 $(info )
 

Package.swift

@@ -13,17 +13,31 @@ let package = Package(
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
-    dependencies: [
-        .package(url: "https://github.com/ggerganov/ggml.git", .branch("release"))
-    ],
     targets: [
         .target(
             name: "llama",
-            dependencies: ["ggml"],
             path: ".",
-            exclude: ["ggml-metal.metal"],
+            exclude: [
+               "cmake",
+               "examples",
+               "scripts",
+               "models",
+               "tests",
+               "CMakeLists.txt",
+               "ggml-cuda.cu",
+               "ggml-cuda.h",
+               "Makefile"
+            ],
             sources: [
+                "ggml.c",
                 "llama.cpp",
+                "ggml-alloc.c",
+                "ggml-backend.c",
+                "ggml-quants.c",
+                "ggml-metal.m",
+            ],
+            resources: [
+                .process("ggml-metal.metal")
             ],
             publicHeadersPath: "spm-headers",
             cSettings: [

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)

ci/run.sh

@@ -568,6 +568,50 @@ function gg_sum_open_llama_7b_v2 {
     #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
+# bge-small
+
+function gg_run_embd_bge_small {
+    cd ${SRC}
+
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/tokenizer.model
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/tokenizer_config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/special_tokens_map.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/pytorch_model.bin
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/sentence_bert_config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/vocab.txt
+
+    path_models="../models-mnt/bge-small"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert-hf-to-gguf.py ${path_models}
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+
+    (time ./bin/embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+
+    set +e
+}
+
+function gg_sum_embd_bge_small {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'BGE Small (BERT):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+}
+
 ## main
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -591,6 +635,8 @@ test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
+    test $ret -eq 0 && gg_run embd_bge_small
+
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
         if [ -z ${GG_BUILD_CUDA} ]; then
             test $ret -eq 0 && gg_run open_llama_3b_v2

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("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sparams.samplers_sequence.c_str());
+    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", 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::string output = "";
+std::vector<std::string> string_split(std::string input, char separator) {
+    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 {
-        {"top_k",      'k'},
-        {"top-k",      'k'},
-        {"top_p",      'p'},
-        {"top-p",      'p'},
-        {"nucleus",    'p'},
-        {"typical_p",  'y'},
-        {"typical-p",  'y'},
-        {"typical",    'y'},
-        {"min_p",      'm'},
-        {"min-p",      'm'},
-        {"tfs_z",      'f'},
-        {"tfs-z",      'f'},
-        {"tfs",        'f'},
-        {"temp",       't'},
-        {"temperature",'t'}
+    std::unordered_map<std::string, llama_sampler_type> sampler_name_map {
+        {"top_k",       llama_sampler_type::TOP_K},
+        {"top-k",       llama_sampler_type::TOP_K},
+        {"top_p",       llama_sampler_type::TOP_P},
+        {"top-p",       llama_sampler_type::TOP_P},
+        {"nucleus",     llama_sampler_type::TOP_P},
+        {"typical_p",   llama_sampler_type::TYPICAL_P},
+        {"typical-p",   llama_sampler_type::TYPICAL_P},
+        {"typical",     llama_sampler_type::TYPICAL_P},
+        {"min_p",       llama_sampler_type::MIN_P},
+        {"min-p",       llama_sampler_type::MIN_P},
+        {"tfs_z",       llama_sampler_type::TFS_Z},
+        {"tfs-z",       llama_sampler_type::TFS_Z},
+        {"tfs",         llama_sampler_type::TFS_Z},
+        {"temp",        llama_sampler_type::TEMP},
+        {"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()) {
-            output += samplers_symbols[name];
+    std::vector<llama_sampler_type> sampler_types;
+    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()) {
-        output += samplers_symbols[input];
+    return sampler_types;
+}
+
+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) {
-            switch (s) {
-                case 'k': result += "-> top_k "; break;
-                case 'f': result += "-> tfs_z "; break;
-                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;
+        for (auto sampler_type : params.samplers_sequence) {
+            const auto sampler_type_name = sampler_type_to_name_string(sampler_type);
+            if (!sampler_type_name.empty()) {
+                result += "-> " + sampler_type_name + " ";
             }
         }
     } 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) {
-        switch (s){
-            case '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 'y': 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 'm': llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
-            case 't':
+    for (auto sampler_type : samplers_sequence) {
+        switch (sampler_type) {
+            case llama_sampler_type::TOP_K    : llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
+            case llama_sampler_type::TFS_Z    : llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
+            case llama_sampler_type::TYPICAL_P: llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
+            case llama_sampler_type::TOP_P    : llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
+            case llama_sampler_type::MIN_P    : llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
+            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
 

convert-hf-to-gguf.py

@@ -10,7 +10,7 @@ import re
 import sys
 from enum import IntEnum
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast
+from typing import TYPE_CHECKING, Any, ContextManager, Iterator, Sequence, cast
 
 import numpy as np
 import torch
@@ -25,15 +25,6 @@ import gguf
 from convert import HfVocab
 
 
-# check for any of the given keys in the dictionary and return the value of the first key found
-def get_key_opts(d, keys):
-    for k in keys:
-        if k in d:
-            return d[k]
-    print(f"Could not find any of {keys}")
-    sys.exit()
-
-
 ###### MODEL DEFINITIONS ######
 
 class SentencePieceTokenTypes(IntEnum):
@@ -58,6 +49,15 @@ class Model:
         self.hparams = Model.load_hparams(self.dir_model)
         self.model_arch = self._get_model_architecture()
         self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
+
+    def find_hparam(self, keys: Sequence[str], optional: bool = False) -> Any:
+        key = next((k for k in keys if k in self.hparams), None)
+        if key is not None:
+            return self.hparams[key]
+        if optional:
+            return None
+        raise KeyError(f"could not find any of: {keys}")
 
     def set_vocab(self):
         self._set_vocab_gpt2()
@@ -79,28 +79,33 @@ class Model:
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_block_count(self.hparams.get(
-            "n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")),
-        ))
-        if (n_ctx := self.hparams.get("max_position_embeddings")) is not None:
+        self.gguf_writer.add_block_count(self.block_count)
+
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
-        if (n_embd := self.hparams.get("hidden_size")) is not None:
-            self.gguf_writer.add_embedding_length(n_embd)
-        if (n_ff := self.hparams.get("intermediate_size")) is not None:
+
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        self.gguf_writer.add_embedding_length(n_embd)
+
+        if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
-        if (n_head := self.hparams.get("num_attention_heads")) is not None:
-            self.gguf_writer.add_head_count(n_head)
+
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        self.gguf_writer.add_head_count(n_head)
+
         if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
 
-        if (n_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
-            self.gguf_writer.add_layer_norm_rms_eps(n_rms_eps)
+        if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
+            self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
+        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon"], optional=True)) is not None:
+            self.gguf_writer.add_layer_norm_eps(f_norm_eps)
         if (n_experts := self.hparams.get("num_local_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
         if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
 
-        self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
+        self.gguf_writer.add_file_type(self.ftype)
 
     def write_tensors(self):
         block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
@@ -209,6 +214,10 @@ class Model:
             return InternLM2Model
         if model_architecture == "MiniCPMForCausalLM":
             return MiniCPMModel
+        if model_architecture == "BertModel":
+            return BertModel
+        if model_architecture == "NomicBertModel":
+            return NomicBertModel
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -264,6 +273,10 @@ class Model:
             return gguf.MODEL_ARCH.INTERNLM2
         if arch == "MiniCPMForCausalLM":
             return gguf.MODEL_ARCH.MINICPM
+        if arch == "BertModel":
+            return gguf.MODEL_ARCH.BERT
+        if arch == "NomicBertModel":
+            return gguf.MODEL_ARCH.NOMIC_BERT
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -1293,21 +1306,21 @@ class GPT2Model(Model):
 
 class Phi2Model(Model):
     def set_gguf_parameters(self):
-        block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"])
+        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
 
-        rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
-        n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"])
-        n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"])
+        rot_pct = self.find_hparam(["partial_rotary_factor"])
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
 
         self.gguf_writer.add_name("Phi2")
-        self.gguf_writer.add_context_length(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"]))
+        self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
 
         self.gguf_writer.add_embedding_length(n_embd)
         self.gguf_writer.add_feed_forward_length(4 * n_embd)
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head)
-        self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"]))
+        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_epsilon", "layer_norm_eps"]))
         self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_add_bos_token(False)
@@ -1629,6 +1642,127 @@ in chat mode so that the conversation can end normally.")
                 self.post_write_tensors(tensor_map, name, data_torch)
 
 
+class BertModel(Model):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.vocab_size = None
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_causal_attention(False)
+        self.gguf_writer.add_pooling_layer(True)
+
+    def set_vocab(self):
+        path = self.dir_model
+        added_tokens_path = self.dir_model if self.dir_model.exists() else None
+
+        # use huggingface vocab to get all tokens
+        vocab = HfVocab(path, added_tokens_path)
+        tokens, scores, toktypes = zip(*vocab.all_tokens())
+        assert len(tokens) == vocab.vocab_size
+        self.vocab_size = vocab.vocab_size
+
+        # we need this to validate the size of the token_type embeddings
+        # though currently we are passing all zeros to the token_type embeddings
+        n_token_types = len(set(toktypes))
+        self.gguf_writer.add_token_type_count(n_token_types)
+
+        # convert to phantom space vocab
+        def phantom(tok, typ):
+            if tok.startswith(b"[") and tok.endswith(b"]"):
+                return tok
+            if tok.startswith(b"##"):
+                return tok[2:]
+            return b"\xe2\x96\x81" + tok
+        tokens = tuple(phantom(t, y) for t, y in zip(tokens, toktypes))
+
+        # set up bos and eos tokens (cls and sep)
+        self.gguf_writer.add_bos_token_id(vocab.tokenizer.cls_token_id)
+        self.gguf_writer.add_eos_token_id(vocab.tokenizer.sep_token_id)
+
+        # add vocab to gguf
+        self.gguf_writer.add_tokenizer_model("bert")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        # handle special tokens
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def write_tensors(self):
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+        tensors = dict(self.get_tensors())
+        for name, data_torch in tensors.items():
+            # we are only using BERT for embeddings so we don't need the pooling layer
+            if name in ("embeddings.position_ids", "pooler.dense.weight", "pooler.dense.bias"):
+                continue  # we don't need these
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            data = data_torch.squeeze().numpy()
+            n_dims = len(data.shape)
+            new_dtype: type[np.floating[Any]]
+
+            if (
+                self.ftype == 1 and name.endswith(".weight") and n_dims == 2
+                and name != "embeddings.token_type_embeddings.weight"  # not used with get_rows, must be F32
+            ):
+                # if f16 desired, convert any float32 2-dim weight tensors to float16
+                new_dtype = np.float16
+            else:
+                # if f32 desired, convert any float16 to float32
+                new_dtype = np.float32
+
+            print(f"{new_name}, n_dims = {n_dims}, {data_torch.dtype} --> {new_dtype}")
+
+            if data.dtype != new_dtype:
+                data = data.astype(new_dtype)
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+
+class NomicBertModel(BertModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # the HF config claims n_ctx=8192, but it uses RoPE scaling
+        self.hparams["n_ctx"] = 2048
+
+        # SwigLU activation
+        assert self.hparams["activation_function"] == "swiglu"
+        # this doesn't do anything in the HF version
+        assert self.hparams["causal"] is False
+        # no bias tensors
+        assert self.hparams["qkv_proj_bias"] is False
+        assert self.hparams["mlp_fc1_bias"] is False
+        assert self.hparams["mlp_fc2_bias"] is False
+        # norm at end of layer
+        assert self.hparams["prenorm"] is False
+        # standard RoPE
+        assert self.hparams["rotary_emb_fraction"] == 1.0
+        assert self.hparams["rotary_emb_interleaved"] is False
+        assert self.hparams["rotary_emb_scale_base"] is None
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
+
+    def get_tensors(self):
+        assert self.vocab_size is not None
+        for name, data in super().get_tensors():
+            # Nomic Embed's token embeddings tensor is padded, but llama.cpp wants tensor sizes to match exactly.
+            if name == 'embeddings.word_embeddings.weight' and data.shape[1] != self.vocab_size:
+                rounded_vocab_size = (self.vocab_size + 63) // 64 * 64
+                assert data.shape == (rounded_vocab_size, self.hparams["n_embd"])
+                data = data[:self.vocab_size, :]
+            yield name, data
+
+
 ###### CONVERSION LOGIC ######
 
 

convert-persimmon-to-gguf.py

@@ -88,7 +88,8 @@ def main():
     gguf_writer.add_embedding_length(hidden_size)
     gguf_writer.add_block_count(block_count)
     gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size)
-    gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+    # ref: https://github.com/ggerganov/llama.cpp/pull/4889/commits/eea19039fc52ea2dbd1aab45b59ab4e3e29a3443
+    gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
     gguf_writer.add_head_count(head_count)
     gguf_writer.add_head_count_kv(head_count_kv)
     gguf_writer.add_rope_freq_base(hparams.rotary_emb_base)

examples/embedding/embedding.cpp

@@ -7,6 +7,51 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+static std::vector<std::string> split_lines(const std::string & s) {
+    std::string line;
+    std::vector<std::string> lines;
+    std::stringstream ss(s);
+    while (std::getline(ss, line)) {
+        lines.push_back(line);
+    }
+    return lines;
+}
+
+static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
+    for (size_t i = 0; i < tokens.size(); i++) {
+        llama_batch_add(batch, tokens[i], i, { seq_id }, false);
+    }
+}
+
+static void normalize(float * vec, float * out, int n) {
+    float norm = 0;
+    for (int i = 0; i < n; i++) {
+        norm += vec[i] * vec[i];
+    }
+    norm = sqrt(norm);
+    for (int i = 0; i < n; i++) {
+        out[i] = vec[i] / norm;
+    }
+}
+
+static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
+    // clear previous kv_cache values (irrelevant for embeddings)
+    llama_kv_cache_clear(ctx);
+
+    // run model
+    fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
+    if (llama_decode(ctx, batch) < 0) {
+        fprintf(stderr, "%s : failed to decode\n", __func__);
+    }
+
+    // normalize on copy
+    for (int k = 0; k < n_seq; k++) {
+        float * emb = llama_get_embeddings_ith(ctx, k);
+        float * out = output + k * n_embd;
+        normalize(emb, out, n_embd);
+    }
+}
+
 int main(int argc, char ** argv) {
     gpt_params params;
 
@@ -55,49 +100,84 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
-    int n_past = 0;
+    // split the prompt into lines
+    std::vector<std::string> prompts = split_lines(params.prompt);
 
-    // tokenize the prompt
-    auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+    // max batch size
+    const uint64_t n_batch = params.n_batch;
+    GGML_ASSERT(params.n_batch == params.n_ctx);
 
+    // tokenize the prompts and trim
+    std::vector<std::vector<int32_t>> inputs;
+    for (const auto & prompt : prompts) {
+        auto inp = ::llama_tokenize(ctx, prompt, true);
+        if (inp.size() > n_batch) {
+            inp.resize(n_batch);
+        }
+        inputs.push_back(inp);
+    }
+
+    // tokenization stats
     if (params.verbose_prompt) {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
-        fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
-        for (int i = 0; i < (int) embd_inp.size(); i++) {
-            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+        for (int i = 0; i < (int) inputs.size(); i++) {
+            fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
+            fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
+            for (int j = 0; j < (int) inputs[i].size(); j++) {
+                fprintf(stderr, "%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
+            }
+            fprintf(stderr, "\n\n");
         }
-        fprintf(stderr, "\n");
     }
 
-    if (embd_inp.size() > (size_t)n_ctx) {
-        fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
-                __func__, embd_inp.size(), n_ctx);
-        return 1;
-    }
-
-    while (!embd_inp.empty()) {
-        int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
-        if (llama_decode(ctx, llama_batch_get_one(embd_inp.data(), n_tokens, n_past, 0))) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return 1;
-        }
-        n_past += n_tokens;
-        embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
-    }
+    // initialize batch
+    const int n_prompts = prompts.size();
+    struct llama_batch batch = llama_batch_init(n_batch, 0, n_prompts);
 
+    // allocate output
     const int n_embd = llama_n_embd(model);
-    const auto * embeddings = llama_get_embeddings(ctx);
+    std::vector<float> embeddings(n_prompts * n_embd, 0);
+    float * emb = embeddings.data();
 
-    for (int i = 0; i < n_embd; i++) {
-        printf("%f ", embeddings[i]);
+    // break into batches
+    int p = 0; // number of prompts processed already
+    int s = 0; // number of prompts in current batch
+    for (int k = 0; k < n_prompts; k++) {
+        // clamp to n_batch tokens
+        auto & inp = inputs[k];
+        const uint64_t n_toks = inp.size();
+
+        // encode if at capacity
+        if (batch.n_tokens + n_toks > n_batch) {
+            float * out = emb + p * n_embd;
+            batch_decode(ctx, batch, out, s, n_embd);
+            llama_batch_clear(batch);
+            p += s;
+            s = 0;
+        }
+
+        // add to batch
+        batch_add_seq(batch, inp, s);
+        s += 1;
     }
-    printf("\n");
 
+    // final batch
+    float * out = emb + p * n_embd;
+    batch_decode(ctx, batch, out, s, n_embd);
+
+    // print first 3 embeddings
+    for (int j = 0; j < std::min(3, n_prompts); j++) {
+        fprintf(stderr, "embedding %d: ", j);
+        for (int i = 0; i < n_embd; i++) {
+            fprintf(stderr, "%f ", emb[j * n_embd + i]);
+        }
+        fprintf(stderr, "\n\n");
+    }
+    fprintf(stderr, "\n");
+
+    // clean up
     llama_print_timings(ctx);
     llama_free(ctx);
     llama_free_model(model);
-
     llama_backend_free();
 
     return 0;

examples/export-lora/export-lora.cpp

@@ -337,24 +337,14 @@ static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int
     params.mem_buffer = NULL;
     params.no_alloc   = true;
     struct ggml_context * ctx = NULL;
-    struct ggml_allocr * alloc = NULL;
-    struct ggml_cgraph * gf = NULL;
+    struct ggml_gallocr * alloc = NULL;
+    struct ggml_cgraph  * gf = NULL;
 
     ctx   = ggml_init(params);
-    alloc = ggml_allocr_new_measure(tensor_alignment);
+    alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
     gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
-    size_t alloc_size = ggml_allocr_alloc_graph(alloc, gf);
-    ggml_allocr_free(alloc);
-    ggml_free(ctx);
 
-    static std::vector<uint8_t> data_compute;
-    data_compute.resize(alloc_size + tensor_alignment);
-
-    ctx   = ggml_init(params);
-    alloc = ggml_allocr_new(data_compute.data(), data_compute.size(), tensor_alignment);
-    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
-    ggml_allocr_alloc_graph(alloc, gf);
-    ggml_allocr_free(alloc);
+    ggml_gallocr_alloc_graph(alloc, gf);
 
     struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads);
     static std::vector<uint8_t> data_work;
@@ -363,6 +353,7 @@ static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int
 
     ggml_graph_compute(gf, &cplan);
 
+    ggml_gallocr_free(alloc);
     ggml_free(ctx);
     return true;
 }

examples/finetune/README.md

@@ -80,9 +80,9 @@ The LORA rank can be configured for each model tensor type separately with these
   --rank-wk N                LORA rank for wk tensor (default 4)
   --rank-wv N                LORA rank for wv tensor (default 4)
   --rank-wo N                LORA rank for wo tensor (default 4)
-  --rank-w1 N                LORA rank for w1 tensor (default 4)
-  --rank-w2 N                LORA rank for w2 tensor (default 4)
-  --rank-w3 N                LORA rank for w3 tensor (default 4)
+  --rank-ffn_gate N          LORA rank for ffn_gate tensor (default 4)
+  --rank-ffn_down N          LORA rank for ffn_down tensor (default 4)
+  --rank-ffn_up N            LORA rank for ffn_up tensor (default 4)
 ```
 
 The LORA rank of 'norm' tensors should always be 1.

examples/finetune/finetune.cpp

@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "ggml-alloc.h"
+#include "ggml-backend.h"
 #include "llama.h"
 #include "common.h"
 #include "train.h"
@@ -13,8 +14,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static const size_t tensor_alignment = 32;
-
 struct my_llama_hparams {
     uint32_t n_vocab    = 32000;
     uint32_t n_ctx      = 512;
@@ -61,9 +60,9 @@ struct my_llama_layer {
     struct ggml_tensor * ffn_norm;
 
     // ff
-    struct ggml_tensor * w1;
-    struct ggml_tensor * w2;
-    struct ggml_tensor * w3;
+    struct ggml_tensor * ffn_gate; // w1
+    struct ggml_tensor * ffn_down; // w2
+    struct ggml_tensor * ffn_up;   // w3
 };
 
 struct my_llama_model {
@@ -86,9 +85,9 @@ struct my_llama_lora_hparams {
     uint32_t n_rank_wv = 4;
     uint32_t n_rank_wo = 4;
     uint32_t n_rank_ffn_norm = 1;
-    uint32_t n_rank_w1 = 4;
-    uint32_t n_rank_w2 = 4;
-    uint32_t n_rank_w3 = 4;
+    uint32_t n_rank_ffn_gate = 4;
+    uint32_t n_rank_ffn_down = 4;
+    uint32_t n_rank_ffn_up = 4;
     uint32_t n_rank_tok_embeddings = 4;
     uint32_t n_rank_norm = 1;
     uint32_t n_rank_output = 4;
@@ -118,17 +117,17 @@ struct my_llama_lora_layer {
     struct ggml_tensor * ffn_norm_b;
 
     // ff
-    struct ggml_tensor * w1_a;
-    struct ggml_tensor * w1_b;
-    struct ggml_tensor * w2_a;
-    struct ggml_tensor * w2_b;
-    struct ggml_tensor * w3_a;
-    struct ggml_tensor * w3_b;
+    struct ggml_tensor * ffn_gate_a;
+    struct ggml_tensor * ffn_gate_b;
+    struct ggml_tensor * ffn_down_a;
+    struct ggml_tensor * ffn_down_b;
+    struct ggml_tensor * ffn_up_a;
+    struct ggml_tensor * ffn_up_b;
 };
 
 struct my_llama_lora {
     struct ggml_context * ctx = NULL;
-    std::vector<uint8_t> data;
+    ggml_backend_buffer_t data;
 
     my_llama_lora_hparams hparams;
 
@@ -209,9 +208,9 @@ static void print_lora_params(struct my_llama_lora_hparams * params) {
     printf("%s: n_rank_wv             : %u\n", __func__, params->n_rank_wv);
     printf("%s: n_rank_wo             : %u\n", __func__, params->n_rank_wo);
     printf("%s: n_rank_ffn_norm       : %u\n", __func__, params->n_rank_ffn_norm);
-    printf("%s: n_rank_w1             : %u\n", __func__, params->n_rank_w1);
-    printf("%s: n_rank_w2             : %u\n", __func__, params->n_rank_w2);
-    printf("%s: n_rank_w3             : %u\n", __func__, params->n_rank_w3);
+    printf("%s: n_rank_ffn_gate       : %u\n", __func__, params->n_rank_ffn_gate);
+    printf("%s: n_rank_ffn_down       : %u\n", __func__, params->n_rank_ffn_down);
+    printf("%s: n_rank_ffn_up         : %u\n", __func__, params->n_rank_ffn_up);
     printf("%s: n_rank_tok_embeddings : %u\n", __func__, params->n_rank_tok_embeddings);
     printf("%s: n_rank_norm           : %u\n", __func__, params->n_rank_norm);
     printf("%s: n_rank_output         : %u\n", __func__, params->n_rank_output);
@@ -320,9 +319,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         layer.wv             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_V, i));
         layer.wo             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_OUT, i));
         layer.ffn_norm       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_NORM, i));
-        layer.w1             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
-        layer.w2             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
-        layer.w3             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
+        layer.ffn_gate       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
+        layer.ffn_down       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
+        layer.ffn_up         = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
 
         assert_shape_1d(layer.attention_norm, hparams.n_embd);
         assert_shape_2d(layer.wq,             hparams.n_embd, hparams.n_embd);
@@ -330,9 +329,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         assert_shape_2d(layer.wv,             hparams.n_embd, hparams.n_embd_gqa());
         assert_shape_2d(layer.wo,             hparams.n_embd, hparams.n_embd);
         assert_shape_1d(layer.ffn_norm,       hparams.n_embd);
-        assert_shape_2d(layer.w1,             hparams.n_embd, hparams.n_ff);
-        assert_shape_2d(layer.w2,             hparams.n_ff,   hparams.n_embd);
-        assert_shape_2d(layer.w3,             hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.ffn_gate,       hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.ffn_down,       hparams.n_ff,   hparams.n_embd);
+        assert_shape_2d(layer.ffn_up,         hparams.n_embd, hparams.n_ff);
     }
 }
 
@@ -363,69 +362,12 @@ static void set_param_lora(struct my_llama_lora * lora) {
         ggml_set_param(ctx, layer.wo_b);
         ggml_set_param(ctx, layer.ffn_norm_a);
         ggml_set_param(ctx, layer.ffn_norm_b);
-        ggml_set_param(ctx, layer.w1_a);
-        ggml_set_param(ctx, layer.w1_b);
-        ggml_set_param(ctx, layer.w2_a);
-        ggml_set_param(ctx, layer.w2_b);
-        ggml_set_param(ctx, layer.w3_a);
-        ggml_set_param(ctx, layer.w3_b);
-    }
-}
-
-static void alloc_lora(struct ggml_allocr * alloc, struct my_llama_lora * lora) {
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_a);
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_b);
-    ggml_allocr_alloc(alloc, lora->norm_a);
-    ggml_allocr_alloc(alloc, lora->norm_b);
-    ggml_allocr_alloc(alloc, lora->output_a);
-    ggml_allocr_alloc(alloc, lora->output_b);
-    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
-        auto & layer = lora->layers[i];
-        ggml_allocr_alloc(alloc, layer.attention_norm_a);
-        ggml_allocr_alloc(alloc, layer.attention_norm_b);
-        ggml_allocr_alloc(alloc, layer.wq_a);
-        ggml_allocr_alloc(alloc, layer.wq_b);
-        ggml_allocr_alloc(alloc, layer.wk_a);
-        ggml_allocr_alloc(alloc, layer.wk_b);
-        ggml_allocr_alloc(alloc, layer.wv_a);
-        ggml_allocr_alloc(alloc, layer.wv_b);
-        ggml_allocr_alloc(alloc, layer.wo_a);
-        ggml_allocr_alloc(alloc, layer.wo_b);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_a);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_b);
-        ggml_allocr_alloc(alloc, layer.w1_a);
-        ggml_allocr_alloc(alloc, layer.w1_b);
-        ggml_allocr_alloc(alloc, layer.w2_a);
-        ggml_allocr_alloc(alloc, layer.w2_b);
-        ggml_allocr_alloc(alloc, layer.w3_a);
-        ggml_allocr_alloc(alloc, layer.w3_b);
-    }
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_a->grad);
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_b->grad);
-    ggml_allocr_alloc(alloc, lora->norm_a->grad);
-    ggml_allocr_alloc(alloc, lora->norm_b->grad);
-    ggml_allocr_alloc(alloc, lora->output_a->grad);
-    ggml_allocr_alloc(alloc, lora->output_b->grad);
-    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
-        auto & layer = lora->layers[i];
-        ggml_allocr_alloc(alloc, layer.attention_norm_a->grad);
-        ggml_allocr_alloc(alloc, layer.attention_norm_b->grad);
-        ggml_allocr_alloc(alloc, layer.wq_a->grad);
-        ggml_allocr_alloc(alloc, layer.wq_b->grad);
-        ggml_allocr_alloc(alloc, layer.wk_a->grad);
-        ggml_allocr_alloc(alloc, layer.wk_b->grad);
-        ggml_allocr_alloc(alloc, layer.wv_a->grad);
-        ggml_allocr_alloc(alloc, layer.wv_b->grad);
-        ggml_allocr_alloc(alloc, layer.wo_a->grad);
-        ggml_allocr_alloc(alloc, layer.wo_b->grad);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_a->grad);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_b->grad);
-        ggml_allocr_alloc(alloc, layer.w1_a->grad);
-        ggml_allocr_alloc(alloc, layer.w1_b->grad);
-        ggml_allocr_alloc(alloc, layer.w2_a->grad);
-        ggml_allocr_alloc(alloc, layer.w2_b->grad);
-        ggml_allocr_alloc(alloc, layer.w3_a->grad);
-        ggml_allocr_alloc(alloc, layer.w3_b->grad);
+        ggml_set_param(ctx, layer.ffn_gate_a);
+        ggml_set_param(ctx, layer.ffn_gate_b);
+        ggml_set_param(ctx, layer.ffn_down_a);
+        ggml_set_param(ctx, layer.ffn_down_b);
+        ggml_set_param(ctx, layer.ffn_up_a);
+        ggml_set_param(ctx, layer.ffn_up_b);
     }
 }
 
@@ -493,12 +435,12 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         layer.ffn_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, n_embd);
         layer.ffn_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, 1);
 
-        layer.w1_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_embd);
-        layer.w1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_ff);
-        layer.w2_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_ff);
-        layer.w2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_embd);
-        layer.w3_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_embd);
-        layer.w3_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_ff);
+        layer.ffn_gate_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_embd);
+        layer.ffn_gate_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_ff);
+        layer.ffn_down_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_ff);
+        layer.ffn_down_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_embd);
+        layer.ffn_up_a   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_embd);
+        layer.ffn_up_b   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_ff);
 
         ggml_set_name(layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_a", i));
         ggml_set_name(layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_b", i));
@@ -512,28 +454,18 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         ggml_set_name(layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  ".weight.lora_b", i));
         ggml_set_name(layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_a", i));
         ggml_set_name(layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_b", i));
-        ggml_set_name(layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
-        ggml_set_name(layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
-        ggml_set_name(layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
-        ggml_set_name(layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
-        ggml_set_name(layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
-        ggml_set_name(layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
     }
 
     set_param_lora(lora);
 
-    // measure data size
-    size_t size = 0;
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        size += GGML_PAD(ggml_nbytes(t), tensor_alignment);
-    }
-
-    // allocate data
-    struct ggml_allocr * alloc = NULL;
-    lora->data.resize(size + tensor_alignment);
-    alloc = ggml_allocr_new(lora->data.data(), lora->data.size(), tensor_alignment);
-    alloc_lora(alloc, lora);
-    ggml_allocr_free(alloc);
+    // allocate data for lora tensors
+    lora->data = ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_cpu_buffer_type());
 }
 
 static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, float std, float min, float max) {
@@ -565,12 +497,12 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
         randomize_tensor_normal(layer.ffn_norm_a, rnd);
         ggml_set_zero(layer.ffn_norm_b);
 
-        randomize_tensor_normal(layer.w1_a, rnd);
-        ggml_set_zero(layer.w1_b);
-        randomize_tensor_normal(layer.w2_a, rnd);
-        ggml_set_zero(layer.w2_b);
-        randomize_tensor_normal(layer.w3_a, rnd);
-        ggml_set_zero(layer.w3_b);
+        randomize_tensor_normal(layer.ffn_gate_a, rnd);
+        ggml_set_zero(layer.ffn_gate_b);
+        randomize_tensor_normal(layer.ffn_down_a, rnd);
+        ggml_set_zero(layer.ffn_down_b);
+        randomize_tensor_normal(layer.ffn_up_a, rnd);
+        ggml_set_zero(layer.ffn_up_b);
     }
 
     free_random_normal_distribution(rnd);
@@ -579,7 +511,7 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
 static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct my_llama_model * model,
         struct my_llama_lora  * lora,
-        struct ggml_allocr    * alloc,
+        ggml_gallocr_t          alloc,
         struct ggml_context   * ctx,
         struct ggml_cgraph    * gf,
         struct ggml_cgraph    * gb,
@@ -590,7 +522,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         const  int              n_tokens,
         const  int              n_batch,
         const  bool             enable_flash_attn,
-        const  bool             enable_checkpointing) {
+        const  bool             enable_checkpointing,
+        const  bool             measure_only) {
 
     ggml_set_scratch(ctx, { 0, 0, nullptr, });
     const int n_past = 0;
@@ -622,13 +555,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
 
     // KQ_pos - contains the positions
     struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
-    ggml_allocr_alloc(alloc, KQ_pos);
-    if (!ggml_allocr_is_measure(alloc)) {
-        int * data = (int *) KQ_pos->data;
-        for (int i = 0; i < N; ++i) {
-            data[i] = n_past + i;
-        }
-    }
+    ggml_set_input(KQ_pos);
 
     // rope has so much parameters that we make a custom function for it
     auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
@@ -683,13 +610,13 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
 
         struct ggml_tensor * attention_norm = add_to_f32(ctx, layer.attention_norm, ggml_mul_mat(ctx, llayer.attention_norm_a, llayer.attention_norm_b));
         struct ggml_tensor * ffn_norm = add_to_f32(ctx, layer.ffn_norm, ggml_mul_mat(ctx, llayer.ffn_norm_a, llayer.ffn_norm_b));
-        struct ggml_tensor * wq = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
-        struct ggml_tensor * wk = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
-        struct ggml_tensor * wv = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
-        struct ggml_tensor * wo = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
-        struct ggml_tensor * w1 = add_to_f32(ctx, layer.w1, ggml_mul_mat(ctx, llayer.w1_a, llayer.w1_b));
-        struct ggml_tensor * w2 = add_to_f32(ctx, layer.w2, ggml_mul_mat(ctx, llayer.w2_a, llayer.w2_b));
-        struct ggml_tensor * w3 = add_to_f32(ctx, layer.w3, ggml_mul_mat(ctx, llayer.w3_a, llayer.w3_b));
+        struct ggml_tensor * wq       = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
+        struct ggml_tensor * wk       = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
+        struct ggml_tensor * wv       = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
+        struct ggml_tensor * wo       = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
+        struct ggml_tensor * ffn_gate = add_to_f32(ctx, layer.ffn_gate, ggml_mul_mat(ctx, llayer.ffn_gate_a, llayer.ffn_gate_b));
+        struct ggml_tensor * ffn_down = add_to_f32(ctx, layer.ffn_down, ggml_mul_mat(ctx, llayer.ffn_down_a, llayer.ffn_down_b));
+        struct ggml_tensor * ffn_up   = add_to_f32(ctx, layer.ffn_up, ggml_mul_mat(ctx, llayer.ffn_up_a, llayer.ffn_up_b));
 
         struct ggml_tensor * t02 = ggml_rms_norm     (ctx, cur, rms_norm_eps);                       set_name(t02, "t02");     assert_shape_2d(t02, n_embd, N*n_batch);
         struct ggml_tensor * t03 = ggml_repeat       (ctx, attention_norm, t02);                     set_name(t03, "t03");     assert_shape_2d(t03, n_embd, N*n_batch);
@@ -732,11 +659,11 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, rms_norm_eps);                       set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
         struct ggml_tensor * t23 = ggml_repeat       (ctx, ffn_norm, t22);                           set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
         struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                                set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
-        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, w3, t24);                                 set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
-        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, w1, t24);                                 set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, ffn_up, t24);                             set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, ffn_gate, t24);                           set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
         struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                     set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
         struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                                set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
-        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, w2, t28);                                 set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, ffn_down, t28);                           set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
         struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                                set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
         cur = t30;
         if (enable_checkpointing) {
@@ -780,7 +707,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     // input gradient
     ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
     GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
-    ggml_allocr_alloc(alloc, t36->grad);
+    ggml_set_input(t36->grad);
     // KQ_pos
     ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
 
@@ -796,20 +723,32 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f));
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f));
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_gate, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_down, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_up, 1.0f));
     }
 
     // allocating checkpoints in one block to reduce memory fragmentation
     // note: they will be freed in reverse order
     for (unsigned int i = 0; i < checkpoints.size(); ++i) {
         if (checkpoints[i]->data == NULL && checkpoints[i]->view_src == NULL) {
-            ggml_allocr_alloc(alloc, checkpoints[i]);
+            ggml_set_input(checkpoints[i]);
         }
     }
 
-    ggml_allocr_alloc_graph(alloc, gb);
+    if (measure_only) {
+        ggml_gallocr_reserve(alloc, gb);
+    } else {
+        ggml_gallocr_alloc_graph(alloc, gb);
+
+        // set KQ_pos
+        {
+            int * data = (int *) KQ_pos->data;
+            for (int i = 0; i < N; ++i) {
+                data[i] = n_past + i;
+            }
+        }
+    }
 
     // remove the additional nodes and leafs
     for (int i = n_leafs_before; i < gb->n_leafs; ++i) {
@@ -859,9 +798,9 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wv,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_V);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wo,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_norm,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_NORM);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w1,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w2,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w3,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_gate,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_down,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_up,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
 
     init_lora(model, lora);
 
@@ -886,12 +825,12 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
         copy_tensor_by_name(layer.wo_b,             f_ggml_ctx, ggml_get_name(layer.wo_b));
         copy_tensor_by_name(layer.ffn_norm_a,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_a));
         copy_tensor_by_name(layer.ffn_norm_b,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_b));
-        copy_tensor_by_name(layer.w1_a,             f_ggml_ctx, ggml_get_name(layer.w1_a));
-        copy_tensor_by_name(layer.w1_b,             f_ggml_ctx, ggml_get_name(layer.w1_b));
-        copy_tensor_by_name(layer.w2_a,             f_ggml_ctx, ggml_get_name(layer.w2_a));
-        copy_tensor_by_name(layer.w2_b,             f_ggml_ctx, ggml_get_name(layer.w2_b));
-        copy_tensor_by_name(layer.w3_a,             f_ggml_ctx, ggml_get_name(layer.w3_a));
-        copy_tensor_by_name(layer.w3_b,             f_ggml_ctx, ggml_get_name(layer.w3_b));
+        copy_tensor_by_name(layer.ffn_gate_a,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_a));
+        copy_tensor_by_name(layer.ffn_gate_b,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_b));
+        copy_tensor_by_name(layer.ffn_down_a,       f_ggml_ctx, ggml_get_name(layer.ffn_down_a));
+        copy_tensor_by_name(layer.ffn_down_b,       f_ggml_ctx, ggml_get_name(layer.ffn_down_b));
+        copy_tensor_by_name(layer.ffn_up_a,         f_ggml_ctx, ggml_get_name(layer.ffn_up_a));
+        copy_tensor_by_name(layer.ffn_up_b,         f_ggml_ctx, ggml_get_name(layer.ffn_up_b));
     }
 }
 
@@ -929,9 +868,9 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_V,       lora->hparams.n_rank_wv);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,     lora->hparams.n_rank_wo);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_NORM,     lora->hparams.n_rank_ffn_norm);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_w1);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_w2);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_w3);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_ffn_gate);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_ffn_down);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_ffn_up);
 
     gguf_add_tensor(fctx, lora->tok_embeddings_a);
     gguf_add_tensor(fctx, lora->tok_embeddings_b);
@@ -955,12 +894,12 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
         gguf_add_tensor(fctx, layer.wo_b);
         gguf_add_tensor(fctx, layer.ffn_norm_a);
         gguf_add_tensor(fctx, layer.ffn_norm_b);
-        gguf_add_tensor(fctx, layer.w1_a);
-        gguf_add_tensor(fctx, layer.w1_b);
-        gguf_add_tensor(fctx, layer.w2_a);
-        gguf_add_tensor(fctx, layer.w2_b);
-        gguf_add_tensor(fctx, layer.w3_a);
-        gguf_add_tensor(fctx, layer.w3_b);
+        gguf_add_tensor(fctx, layer.ffn_gate_a);
+        gguf_add_tensor(fctx, layer.ffn_gate_b);
+        gguf_add_tensor(fctx, layer.ffn_down_a);
+        gguf_add_tensor(fctx, layer.ffn_down_b);
+        gguf_add_tensor(fctx, layer.ffn_up_a);
+        gguf_add_tensor(fctx, layer.ffn_up_b);
     }
 }
 
@@ -1165,12 +1104,12 @@ static void save_as_llama_lora(const char * filename, struct my_llama_lora * lor
         write_tensor(&file, layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  i, ".weight.loraB"));
         write_tensor(&file, layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraA"));
         write_tensor(&file, layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
-        write_tensor(&file, layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
-        write_tensor(&file, layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
-        write_tensor(&file, layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
     }
 }
 
@@ -1200,9 +1139,9 @@ struct train_params {
     uint32_t n_rank_wv;
     uint32_t n_rank_wo;
     uint32_t n_rank_ffn_norm;
-    uint32_t n_rank_w1;
-    uint32_t n_rank_w2;
-    uint32_t n_rank_w3;
+    uint32_t n_rank_ffn_gate;
+    uint32_t n_rank_ffn_down;
+    uint32_t n_rank_ffn_up;
     uint32_t n_rank_tok_embeddings;
     uint32_t n_rank_norm;
     uint32_t n_rank_output;
@@ -1213,9 +1152,9 @@ struct train_params {
     bool custom_n_rank_wv;
     bool custom_n_rank_wo;
     bool custom_n_rank_ffn_norm;
-    bool custom_n_rank_w1;
-    bool custom_n_rank_w2;
-    bool custom_n_rank_w3;
+    bool custom_n_rank_ffn_gate;
+    bool custom_n_rank_ffn_down;
+    bool custom_n_rank_ffn_up;
     bool custom_n_rank_tok_embeddings;
     bool custom_n_rank_norm;
     bool custom_n_rank_output;
@@ -1247,9 +1186,9 @@ static struct train_params get_default_train_params() {
     params.n_rank_wv             = 4;
     params.n_rank_wo             = 4;
     params.n_rank_ffn_norm       = 1;
-    params.n_rank_w1             = 4;
-    params.n_rank_w2             = 4;
-    params.n_rank_w3             = 4;
+    params.n_rank_ffn_gate       = 4;
+    params.n_rank_ffn_down       = 4;
+    params.n_rank_ffn_up         = 4;
     params.n_rank_tok_embeddings = 4;
     params.n_rank_norm           = 1;
     params.n_rank_output         = 4;
@@ -1260,9 +1199,9 @@ static struct train_params get_default_train_params() {
     params.custom_n_rank_wv             = false;
     params.custom_n_rank_wo             = false;
     params.custom_n_rank_ffn_norm       = false;
-    params.custom_n_rank_w1             = false;
-    params.custom_n_rank_w2             = false;
-    params.custom_n_rank_w3             = false;
+    params.custom_n_rank_ffn_gate       = false;
+    params.custom_n_rank_ffn_down       = false;
+    params.custom_n_rank_ffn_up         = false;
     params.custom_n_rank_tok_embeddings = false;
     params.custom_n_rank_norm           = false;
     params.custom_n_rank_output         = false;
@@ -1293,9 +1232,9 @@ static void train_print_usage(int argc, char ** argv, const struct train_params
     fprintf(stderr, "  --rank-wk N                LORA rank for wk tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wv N                LORA rank for wv tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wo N                LORA rank for wo tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w1 N                LORA rank for w1 tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w2 N                LORA rank for w2 tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w3 N                LORA rank for w3 tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_gate N          LORA rank for ffn_gate tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_down N          LORA rank for ffn_down tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_up N            LORA rank for ffn_up tensor, overrides default rank.\n");
 
     print_common_train_usage(argc, argv, &params->common);
 }
@@ -1430,27 +1369,27 @@ static bool train_params_parse(int argc, char ** argv, struct train_params * par
             }
             params->n_rank_wo = std::stoi(argv[i]);
             params->custom_n_rank_wo = true;
-        } else if (arg == "--rank-w1") {
+        } else if (arg == "--rank-ffn_gate") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w1 = std::stoi(argv[i]);
-            params->custom_n_rank_w1 = true;
-        } else if (arg == "--rank-w2") {
+            params->n_rank_ffn_gate = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_gate = true;
+        } else if (arg == "--rank-ffn_down") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w2 = std::stoi(argv[i]);
-            params->custom_n_rank_w2 = true;
-        } else if (arg == "--rank-w3") {
+            params->n_rank_ffn_down = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_down = true;
+        } else if (arg == "--rank-ffn_up") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w3 = std::stoi(argv[i]);
-            params->custom_n_rank_w3 = true;
+            params->n_rank_ffn_up = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_up = true;
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             train_print_usage(argc, argv, &default_params);
@@ -1513,12 +1452,12 @@ static int64_t get_parameter_count(struct my_llama_lora* lora) {
         nx += ggml_nelements(layer.wo_b);
         nx += ggml_nelements(layer.ffn_norm_a);
         nx += ggml_nelements(layer.ffn_norm_b);
-        nx += ggml_nelements(layer.w1_a);
-        nx += ggml_nelements(layer.w1_b);
-        nx += ggml_nelements(layer.w2_a);
-        nx += ggml_nelements(layer.w2_b);
-        nx += ggml_nelements(layer.w3_a);
-        nx += ggml_nelements(layer.w3_b);
+        nx += ggml_nelements(layer.ffn_gate_a);
+        nx += ggml_nelements(layer.ffn_gate_b);
+        nx += ggml_nelements(layer.ffn_down_a);
+        nx += ggml_nelements(layer.ffn_down_b);
+        nx += ggml_nelements(layer.ffn_up_a);
+        nx += ggml_nelements(layer.ffn_up_b);
     }
     return nx;
 }
@@ -1572,9 +1511,9 @@ int main(int argc, char ** argv) {
     uint32_t n_rank_wv                 = params.custom_n_rank_wv             ? params.n_rank_wv             : params.lora_r;
     uint32_t n_rank_wo                 = params.custom_n_rank_wo             ? params.n_rank_wo             : params.lora_r;
     uint32_t n_rank_ffn_norm           = params.custom_n_rank_ffn_norm       ? params.n_rank_ffn_norm       : 1;
-    uint32_t n_rank_w1                 = params.custom_n_rank_w1             ? params.n_rank_w1             : params.lora_r;
-    uint32_t n_rank_w2                 = params.custom_n_rank_w2             ? params.n_rank_w2             : params.lora_r;
-    uint32_t n_rank_w3                 = params.custom_n_rank_w3             ? params.n_rank_w3             : params.lora_r;
+    uint32_t n_rank_ffn_gate           = params.custom_n_rank_ffn_gate       ? params.n_rank_ffn_gate       : params.lora_r;
+    uint32_t n_rank_ffn_down           = params.custom_n_rank_ffn_down       ? params.n_rank_ffn_down       : params.lora_r;
+    uint32_t n_rank_ffn_up             = params.custom_n_rank_ffn_up         ? params.n_rank_ffn_up         : params.lora_r;
     uint32_t n_rank_tok_embeddings     = params.custom_n_rank_tok_embeddings ? params.n_rank_tok_embeddings : params.lora_r;
     uint32_t n_rank_norm               = params.custom_n_rank_norm           ? params.n_rank_norm           : 1;
     uint32_t n_rank_output             = params.custom_n_rank_output         ? params.n_rank_output         : params.lora_r;
@@ -1584,9 +1523,9 @@ int main(int argc, char ** argv) {
     lora.hparams.n_rank_wv             = n_rank_wv;
     lora.hparams.n_rank_wo             = n_rank_wo;
     lora.hparams.n_rank_ffn_norm       = n_rank_ffn_norm;
-    lora.hparams.n_rank_w1             = n_rank_w1;
-    lora.hparams.n_rank_w2             = n_rank_w2;
-    lora.hparams.n_rank_w3             = n_rank_w3;
+    lora.hparams.n_rank_ffn_gate       = n_rank_ffn_gate;
+    lora.hparams.n_rank_ffn_down       = n_rank_ffn_down;
+    lora.hparams.n_rank_ffn_up         = n_rank_ffn_up;
     lora.hparams.n_rank_tok_embeddings = n_rank_tok_embeddings;
     lora.hparams.n_rank_norm           = n_rank_norm;
     lora.hparams.n_rank_output         = n_rank_output;
@@ -1627,9 +1566,9 @@ int main(int argc, char ** argv) {
         || (lora.hparams.n_rank_wv             != n_rank_wv)
         || (lora.hparams.n_rank_wo             != n_rank_wo)
         || (lora.hparams.n_rank_ffn_norm       != n_rank_ffn_norm)
-        || (lora.hparams.n_rank_w1             != n_rank_w1)
-        || (lora.hparams.n_rank_w2             != n_rank_w2)
-        || (lora.hparams.n_rank_w3             != n_rank_w3)
+        || (lora.hparams.n_rank_ffn_gate       != n_rank_ffn_gate)
+        || (lora.hparams.n_rank_ffn_down       != n_rank_ffn_down)
+        || (lora.hparams.n_rank_ffn_up         != n_rank_ffn_up)
         || (lora.hparams.n_rank_tok_embeddings != n_rank_tok_embeddings)
         || (lora.hparams.n_rank_norm           != n_rank_norm)
         || (lora.hparams.n_rank_output         != n_rank_output)
@@ -1663,7 +1602,7 @@ int main(int argc, char ** argv) {
     printf("%s: seen train_samples     %llu\n", __func__, (long long unsigned) train->train_samples);
     printf("%s: seen train_tokens      %llu\n", __func__, (long long unsigned) train->train_tokens);
     printf("%s: completed train_epochs %llu\n", __func__, (long long unsigned) train->train_epochs);
-    printf("%s: lora_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(lora.ctx) + lora.data.size()), (float) (ggml_used_mem(lora.ctx) + lora.data.size()) / (1024.0f*1024.0f));
+    printf("%s: lora_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(lora.ctx) + ggml_backend_buffer_get_size(lora.data)), (float) (ggml_used_mem(lora.ctx) + ggml_backend_buffer_get_size(lora.data)) / (1024.0f*1024.0f));
 
     if (params.only_write_lora) {
         save_train_files_data save_data;
@@ -1690,10 +1629,6 @@ int main(int argc, char ** argv) {
     int n_vocab  = model.hparams.n_vocab;
     int n_batch  = params.common.n_batch;
 
-
-    std::vector<uint8_t> mem_input_data;
-    std::vector<uint8_t> mem_compute_data;
-
     // context for input tensors without their data
     struct ggml_init_params ctx_input_params = {
         ggml_tensor_overhead() * 2, // mem_size
@@ -1706,17 +1641,11 @@ int main(int argc, char ** argv) {
     struct ggml_tensor * tokens_input  = ggml_new_tensor_2d(ctx_input, GGML_TYPE_I32, n_tokens, n_batch);
     struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
 
-    // measure required memory for input tensors
-    size_t max_input_size = GGML_PAD(ggml_nbytes(tokens_input), tensor_alignment) +
-                            GGML_PAD(ggml_nbytes(target_probs), tensor_alignment) +
-                            tensor_alignment;
-    printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
-
     // allocate input tensors
-    mem_input_data.resize(max_input_size);
-    ggml_allocr_t alloc_inps = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
-    ggml_allocr_alloc(alloc_inps, tokens_input);
-    ggml_allocr_alloc(alloc_inps, target_probs);
+    // measure required memory for input tensors
+    ggml_backend_buffer_t input_data = ggml_backend_alloc_ctx_tensors_from_buft(ctx_input, ggml_backend_cpu_buffer_type());
+    size_t max_input_size = ggml_backend_buffer_get_size(input_data);
+    printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
 
     // context for compute tensors without their data
     const size_t estimated_compute_size_wo_data = (
@@ -1743,7 +1672,7 @@ int main(int argc, char ** argv) {
     // find best evaluation order
     for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
         ctx_compute = ggml_init(ctx_compute_params);
-        ggml_allocr_t alloc = ggml_allocr_new_measure(tensor_alignment);
+        ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
         gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
         gf->order = (enum ggml_cgraph_eval_order) order;
         gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@@ -1756,14 +1685,15 @@ int main(int argc, char ** argv) {
             &logits, tokens_input, target_probs,
             n_tokens, n_batch,
             params.common.use_flash,
-            params.common.use_checkpointing
+            params.common.use_checkpointing,
+            true
         );
-        size_t max_compute_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+        size_t max_compute_size = ggml_gallocr_get_buffer_size(alloc, 0); // FIXME: this will still allocate the buffer
         if (max_compute_size < best_compute_size) {
             best_compute_size = max_compute_size;
             best_order = gf->order;
         }
-        ggml_allocr_free(alloc);
+        ggml_gallocr_free(alloc);
         ggml_free(ctx_compute);
     }
     size_t max_compute_size = best_compute_size;
@@ -1774,9 +1704,8 @@ int main(int argc, char ** argv) {
         "invalid");
 
     // allocate compute tensors
-    mem_compute_data.resize(max_compute_size);
     ctx_compute = ggml_init(ctx_compute_params);
-    ggml_allocr_t alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
+    ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
     gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
     gf->order = best_order;
     gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@@ -1789,11 +1718,9 @@ int main(int argc, char ** argv) {
         &logits, tokens_input, target_probs,
         n_tokens, n_batch,
         params.common.use_flash,
-        params.common.use_checkpointing
+        params.common.use_checkpointing,
+        false
     );
-    ggml_allocr_free(alloc);
-    ggml_allocr_free(alloc_inps);
-
 
     // tokenize data
     std::vector<llama_token> train_tokens;
@@ -1908,6 +1835,8 @@ int main(int argc, char ** argv) {
     ggml_free(ctx_work);
     ggml_free(ctx_compute);
     ggml_free(ctx_input);
+    ggml_gallocr_free(alloc);
+
 
     int64_t t1 = ggml_time_ms();
     printf("%s: total training time: ", __func__);

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/clip.cpp

@@ -367,7 +367,7 @@ struct clip_ctx {
     ggml_backend_buffer_t params_buffer = NULL;
     ggml_backend_buffer_t compute_buffer = NULL;
     ggml_backend_t backend = NULL;
-    ggml_allocr * compute_alloc = NULL;
+    ggml_gallocr_t compute_alloc = NULL;
 };
 
 static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
@@ -405,31 +405,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     struct ggml_cgraph * gf = ggml_new_graph(ctx0);
 
     struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size, image_size, 3, batch_size);
-    ggml_allocr_alloc(ctx->compute_alloc, inp_raw);
-
-    if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
-        float * data = (float *)malloc(ggml_nbytes(inp_raw));
-
-        for (size_t i = 0; i < imgs->size; i++) {
-            const int nx = imgs->data[i].nx;
-            const int ny = imgs->data[i].ny;
-            GGML_ASSERT(nx == image_size && ny == image_size);
-
-            const int n = nx * ny;
-
-            for (int b = 0; b < batch_size; b++) {
-                for (int k = 0; k < 3; k++) {
-                    for (int y = 0; y < ny; y++) {
-                        for (int x = 0; x < nx; x++) {
-                            data[(b * 3 * n) + k * n + y * nx + x] = imgs->data[b].buf[3 * (y * nx + x) + k];
-                        }
-                    }
-                }
-            }
-        }
-        ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw));
-        free(data);
-    }
+    ggml_set_name(inp_raw, "inp_raw");
+    ggml_set_input(inp_raw);
 
     struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
 
@@ -438,13 +415,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
 
     // concat class_embeddings and patch_embeddings
     struct ggml_tensor * embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
-    ggml_allocr_alloc(ctx->compute_alloc, embeddings);
-    if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
-        void* zero_mem = malloc(ggml_nbytes(embeddings));
-        memset(zero_mem, 0, ggml_nbytes(embeddings));
-        ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
-        free(zero_mem);
-    }
+    ggml_set_name(embeddings, "embeddings");
+    ggml_set_input(embeddings);
 
     embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
             embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
@@ -453,15 +425,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
 
     struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
-    ggml_allocr_alloc(ctx->compute_alloc, positions);
-    if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
-        int* positions_data = (int*)malloc(ggml_nbytes(positions));
-        for (int i = 0; i < num_positions; i++) {
-            positions_data[i] = i;
-        }
-        ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
-        free(positions_data);
-    }
+    ggml_set_name(positions, "positions");
+    ggml_set_input(positions);
 
     embeddings =
         ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
@@ -560,15 +525,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
 
         struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches);
-        ggml_allocr_alloc(ctx->compute_alloc, patches);
-        if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
-            int* patches_data = (int*)malloc(ggml_nbytes(patches));
-            for (int i = 0; i < num_patches; i++) {
-                patches_data[i] = i + 1;
-            }
-            ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
-            free(patches_data);
-        }
+        ggml_set_name(patches, "patches");
+        ggml_set_input(patches);
 
         // shape [1, 576, 1024]
         // ne is whcn, ne = [1024, 576, 1, 1]
@@ -809,7 +767,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     }
 
     // data
-    size_t buffer_size = 0;
+    size_t model_size = 0;
     {
         for (int i = 0; i < n_tensors; ++i) {
             const char * name = gguf_get_tensor_name(ctx, i);
@@ -817,7 +775,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             enum ggml_type type = gguf_get_tensor_type(ctx, i);
             struct ggml_tensor * cur = ggml_get_tensor(meta, name);
             size_t tensor_size = ggml_nbytes(cur);
-            buffer_size += tensor_size;
+            model_size += tensor_size;
             if (verbosity >= 3) {
                 printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
                        __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type));
@@ -825,8 +783,6 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-    buffer_size += n_tensors * 128 /* CLIP PADDING */;
-
     clip_ctx * new_clip = new clip_ctx;
 
     // update projector type
@@ -886,12 +842,12 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             printf("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
             printf("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
             printf("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
-            printf("%s: model size:     %.2f MB\n", __func__, buffer_size / 1024.0 / 1024.0);
+            printf("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
             printf("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
         }
     }
 
-    printf("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, buffer_size / (1024.0 * 1024.0), n_tensors);
+    printf("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
 
     // load tensors
     {
@@ -925,12 +881,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
 
         // alloc memory and offload data
-        new_clip->params_buffer = ggml_backend_alloc_buffer(new_clip->backend, buffer_size);
-        ggml_allocr* alloc = ggml_allocr_new_from_buffer(new_clip->params_buffer);
+        new_clip->params_buffer = ggml_backend_alloc_ctx_tensors(new_clip->ctx_data, new_clip->backend);
         for (int i = 0; i < n_tensors; ++i) {
             const char * name = gguf_get_tensor_name(ctx, i);
             struct ggml_tensor * cur = ggml_get_tensor(new_clip->ctx_data, name);
-            ggml_allocr_alloc(alloc, cur);
             const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
             fin.seekg(offset, std::ios::beg);
             if (!fin) {
@@ -949,7 +903,6 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
                 ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
             }
         }
-        ggml_allocr_free(alloc);
         fin.close();
     }
 
@@ -1077,15 +1030,12 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     // measure mem requirement and allocate
     {
         new_clip->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead());
-        new_clip->compute_alloc = ggml_allocr_new_measure_from_backend(new_clip->backend);
+        new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
         clip_image_f32_batch batch;
         batch.size = 1;
         ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
-        size_t compute_memory_buffer_size = ggml_allocr_alloc_graph(new_clip->compute_alloc, gf);
-        ggml_allocr_free(new_clip->compute_alloc);
-        new_clip->compute_buffer = ggml_backend_alloc_buffer(new_clip->backend, compute_memory_buffer_size);
-        new_clip->compute_alloc = ggml_allocr_new_from_buffer(new_clip->compute_buffer);
-
+        ggml_gallocr_reserve(new_clip->compute_alloc, gf);
+        size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
         printf("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
     }
 
@@ -1267,12 +1217,72 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         GGML_ASSERT(batch_size == 1); // TODO: support multiple images
     }
 
-    // reset alloc buffer to clean the memory from previous invocations
-    ggml_allocr_reset(ctx->compute_alloc);
-
     // build the inference graph
     ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
-    ggml_allocr_alloc_graph(ctx->compute_alloc, gf);
+    ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
+
+    // set inputs
+    const auto & model = ctx->vision_model;
+    const auto & hparams = model.hparams;
+    const int image_size = hparams.image_size;
+    const int patch_size = hparams.patch_size;
+    const int num_patches = ((image_size / patch_size) * (image_size / patch_size));
+    const int num_positions = num_patches + 1;
+
+    {
+        struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
+        float * data = (float *)malloc(ggml_nbytes(inp_raw));
+
+        for (size_t i = 0; i < imgs->size; i++) {
+            const int nx = imgs->data[i].nx;
+            const int ny = imgs->data[i].ny;
+            GGML_ASSERT(nx == image_size && ny == image_size);
+
+            const int n = nx * ny;
+
+            for (int b = 0; b < batch_size; b++) {
+                for (int k = 0; k < 3; k++) {
+                    for (int y = 0; y < ny; y++) {
+                        for (int x = 0; x < nx; x++) {
+                            data[(b * 3 * n) + k * n + y * nx + x] = imgs->data[b].buf[3 * (y * nx + x) + k];
+                        }
+                    }
+                }
+            }
+        }
+        ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw));
+        free(data);
+    }
+
+    {
+        struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
+
+        void* zero_mem = malloc(ggml_nbytes(embeddings));
+        memset(zero_mem, 0, ggml_nbytes(embeddings));
+        ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
+        free(zero_mem);
+    }
+
+    {
+        struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
+
+        int* positions_data = (int*)malloc(ggml_nbytes(positions));
+        for (int i = 0; i < num_positions; i++) {
+            positions_data[i] = i;
+        }
+        ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+        free(positions_data);
+    }
+
+    {
+        struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
+        int* patches_data = (int*)malloc(ggml_nbytes(patches));
+        for (int i = 0; i < num_patches; i++) {
+            patches_data[i] = i + 1;
+        }
+        ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+        free(patches_data);
+    }
 
     if (ggml_backend_is_cpu(ctx->backend)) {
         ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);

examples/llava/convert-image-encoder-to-gguf.py

@@ -71,7 +71,7 @@ def bytes_to_unicode():
     return dict(zip(bs, cs))
 
 
-ap = argparse.ArgumentParser(prog="convert_hf_to_gguf.py")
+ap = argparse.ArgumentParser()
 ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
 ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
 ap.add_argument("--text-only", action="store_true", required=False,

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)
                 {
-                    llama_token tok = el[0].get<llama_token>();
-                    if (tok >= 0 && tok < n_vocab)
+                    float bias;
+                    if (el[1].is_number())
                     {
-                        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)
                         {
-                            slot->sparams.logit_bias[tok] = el[1].get<float>();
+                            slot->sparams.logit_bias[tok] = bias;
                         }
-                        else if (el[1].is_boolean() && !el[1].get<bool>())
+                    }
+                    else if (el[0].is_string())
+                    {
+                        auto toks = llama_tokenize(model, el[0].get<std::string>(), false);
+                        for (auto tok : toks)
                         {
-                            slot->sparams.logit_bias[tok] = -INFINITY;
+                            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();
 }
 

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "ggml-alloc.h"
+#include "ggml-backend.h"
 #include "common.h"
 #include "train.h"
 #include "llama.h"
@@ -19,8 +20,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static const size_t tensor_alignment = 32;
-
 struct my_llama_hparams {
     uint32_t n_vocab = 32000;
     uint32_t n_ctx   = 512;
@@ -51,14 +50,14 @@ struct my_llama_layer {
     struct ggml_tensor * ffn_norm;
 
     // ff
-    struct ggml_tensor * w1;
-    struct ggml_tensor * w2;
-    struct ggml_tensor * w3;
+    struct ggml_tensor * ffn_gate; // w1
+    struct ggml_tensor * ffn_down; // w2
+    struct ggml_tensor * ffn_up;   // w3
 };
 
 struct my_llama_model {
     struct ggml_context * ctx = NULL;
-    std::vector<uint8_t> data;
+    ggml_backend_buffer_t data = NULL;
 
     my_llama_hparams hparams;
 
@@ -141,42 +140,9 @@ static void set_param_model(struct my_llama_model * model) {
         ggml_set_param(ctx, layer.wv);
         ggml_set_param(ctx, layer.wo);
         ggml_set_param(ctx, layer.ffn_norm);
-        ggml_set_param(ctx, layer.w1);
-        ggml_set_param(ctx, layer.w2);
-        ggml_set_param(ctx, layer.w3);
-    }
-}
-
-static void alloc_model(struct ggml_allocr * alloc, struct my_llama_model * model) {
-    ggml_allocr_alloc(alloc, model->tok_embeddings);
-    ggml_allocr_alloc(alloc, model->norm);
-    ggml_allocr_alloc(alloc, model->output);
-    for (uint32_t i = 0; i < model->layers.size(); ++i) {
-        auto & layer = model->layers[i];
-        ggml_allocr_alloc(alloc, layer.attention_norm);
-        ggml_allocr_alloc(alloc, layer.wq);
-        ggml_allocr_alloc(alloc, layer.wk);
-        ggml_allocr_alloc(alloc, layer.wv);
-        ggml_allocr_alloc(alloc, layer.wo);
-        ggml_allocr_alloc(alloc, layer.ffn_norm);
-        ggml_allocr_alloc(alloc, layer.w1);
-        ggml_allocr_alloc(alloc, layer.w2);
-        ggml_allocr_alloc(alloc, layer.w3);
-    }
-    ggml_allocr_alloc(alloc, model->tok_embeddings->grad);
-    ggml_allocr_alloc(alloc, model->norm->grad);
-    ggml_allocr_alloc(alloc, model->output->grad);
-    for (uint32_t i = 0; i < model->layers.size(); ++i) {
-        auto & layer = model->layers[i];
-        ggml_allocr_alloc(alloc, layer.attention_norm->grad);
-        ggml_allocr_alloc(alloc, layer.wq->grad);
-        ggml_allocr_alloc(alloc, layer.wk->grad);
-        ggml_allocr_alloc(alloc, layer.wv->grad);
-        ggml_allocr_alloc(alloc, layer.wo->grad);
-        ggml_allocr_alloc(alloc, layer.ffn_norm->grad);
-        ggml_allocr_alloc(alloc, layer.w1->grad);
-        ggml_allocr_alloc(alloc, layer.w2->grad);
-        ggml_allocr_alloc(alloc, layer.w3->grad);
+        ggml_set_param(ctx, layer.ffn_gate);
+        ggml_set_param(ctx, layer.ffn_down);
+        ggml_set_param(ctx, layer.ffn_up);
     }
 }
 
@@ -232,9 +198,9 @@ static void init_model(struct my_llama_model * model) {
 
         layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
 
-        layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
-        layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,   n_ff, n_embd);
-        layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
+        layer.ffn_gate = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
+        layer.ffn_down = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,   n_ff, n_embd);
+        layer.ffn_up   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
 
         ggml_set_name(layer.attention_norm, tni(LLM_TENSOR_ATTN_NORM, i));
 
@@ -245,24 +211,15 @@ static void init_model(struct my_llama_model * model) {
 
         ggml_set_name(layer.ffn_norm,       tni(LLM_TENSOR_FFN_NORM, i));
 
-        ggml_set_name(layer.w1,             tni(LLM_TENSOR_FFN_GATE, i));
-        ggml_set_name(layer.w2,             tni(LLM_TENSOR_FFN_DOWN, i));
-        ggml_set_name(layer.w3,             tni(LLM_TENSOR_FFN_UP, i));
+        ggml_set_name(layer.ffn_gate,       tni(LLM_TENSOR_FFN_GATE, i));
+        ggml_set_name(layer.ffn_down,       tni(LLM_TENSOR_FFN_DOWN, i));
+        ggml_set_name(layer.ffn_up,         tni(LLM_TENSOR_FFN_UP, i));
     }
 
     set_param_model(model);
 
-    // measure data size
-    size_t size = 0;
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        size += GGML_PAD(ggml_nbytes(t), tensor_alignment);
-    }
-
     // allocate data
-    struct ggml_allocr * alloc = NULL;
-    model->data.resize(size + tensor_alignment);
-    alloc = ggml_allocr_new(model->data.data(), model->data.size(), tensor_alignment);
-    alloc_model(alloc, model);
+    model->data = ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_cpu_buffer_type());
 }
 
 static void randomize_model(struct my_llama_model * model, int seed, float mean, float std, float min, float max) {
@@ -287,9 +244,9 @@ static void randomize_model(struct my_llama_model * model, int seed, float mean,
 
         randomize_tensor_normal(layer.ffn_norm, rnd);
 
-        randomize_tensor_normal(layer.w1, rnd);
-        randomize_tensor_normal(layer.w2, rnd);
-        randomize_tensor_normal(layer.w3, rnd);
+        randomize_tensor_normal(layer.ffn_gate, rnd);
+        randomize_tensor_normal(layer.ffn_down, rnd);
+        randomize_tensor_normal(layer.ffn_up,   rnd);
     }
 
     free_random_normal_distribution(rnd);
@@ -297,7 +254,7 @@ static void randomize_model(struct my_llama_model * model, int seed, float mean,
 
 static struct ggml_tensor * llama_build_train_graphs(
         struct my_llama_model * model,
-        struct ggml_allocr    * alloc,
+        ggml_gallocr_t          alloc,
         struct ggml_context   * ctx,
         struct ggml_cgraph    * gf,
         struct ggml_cgraph    * gb,
@@ -308,7 +265,8 @@ static struct ggml_tensor * llama_build_train_graphs(
         const  int              n_tokens,
         const  int              n_batch,
         const  bool             enable_flash_attn,
-        const  bool             enable_checkpointing) {
+        const  bool             enable_checkpointing,
+        const  bool             measure_only) {
 
     ggml_set_scratch(ctx, { 0, 0, nullptr, });
     const int n_past = 0;
@@ -334,13 +292,7 @@ static struct ggml_tensor * llama_build_train_graphs(
 
     // KQ_pos - contains the positions
     struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
-    ggml_allocr_alloc(alloc, KQ_pos);
-    if (!ggml_allocr_is_measure(alloc)) {
-        int * data = (int *) KQ_pos->data;
-        for (int i = 0; i < N; ++i) {
-            data[i] = n_past + i;
-        }
-    }
+    ggml_set_input(KQ_pos);
 
     // rope has so much parameters that we make a custom function for it
     auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
@@ -404,11 +356,11 @@ static struct ggml_tensor * llama_build_train_graphs(
         struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, f_norm_rms_eps);                    set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
         struct ggml_tensor * t23 = ggml_repeat       (ctx, layer.ffn_norm, t22);                    set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
         struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                               set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
-        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, layer.w3, t24);                          set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
-        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, layer.w1, t24);                          set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, layer.ffn_up, t24);                      set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, layer.ffn_gate, t24);                    set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
         struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                    set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
         struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                               set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
-        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, layer.w2, t28);                          set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, layer.ffn_down, t28);                    set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
         struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                               set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
         cur = t30;
         checkpoints.push_back(cur);
@@ -448,21 +400,31 @@ static struct ggml_tensor * llama_build_train_graphs(
         // KQ_pos
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
         GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
-
-        ggml_allocr_alloc(alloc, t36->grad);
+        ggml_set_input(t36->grad);
 
         // allocating checkpoints in one block to reduce memory fragmentation
         // note: they will be freed in reverse order
         for (int i = 0; i < (int) checkpoints.size(); ++i) {
             if (checkpoints[i]->data == NULL && checkpoints[i]->view_src == NULL) {
-                ggml_allocr_alloc(alloc, checkpoints[i]);
+                ggml_set_input(checkpoints[i]);
             }
         }
 
         //int n_leafs_after = gb->n_leafs;
         //int n_nodes_after = gb->n_nodes;
+        if (measure_only) {
+            // FIXME: will still allocate
+            ggml_gallocr_reserve(alloc, gb);
+        } else {
+            ggml_gallocr_alloc_graph(alloc, gb);
 
-        ggml_allocr_alloc_graph(alloc, gb);
+            if (!measure_only) {
+                int * data = (int *) KQ_pos->data;
+                for (int i = 0; i < N; ++i) {
+                    data[i] = n_past + i;
+                }
+            }
+        }
 
         // remove the additional nodes and leafs
         for (int i = n_leafs_before; i < gb->n_leafs; ++i) {
@@ -559,9 +521,9 @@ static void load_llama_model_gguf(struct gguf_context * fctx, struct ggml_contex
         copy_tensor_by_name(layer.wv,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_V, i));
         copy_tensor_by_name(layer.wo,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_OUT, i));
         copy_tensor_by_name(layer.ffn_norm,       f_ggml_ctx, tni(LLM_TENSOR_FFN_NORM, i));
-        copy_tensor_by_name(layer.w1,             f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
-        copy_tensor_by_name(layer.w2,             f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
-        copy_tensor_by_name(layer.w3,             f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
+        copy_tensor_by_name(layer.ffn_gate,       f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
+        copy_tensor_by_name(layer.ffn_down,       f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
+        copy_tensor_by_name(layer.ffn_up,         f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
     }
 }
 
@@ -702,9 +664,9 @@ static void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vo
         gguf_add_tensor(fctx, layer.wv);
         gguf_add_tensor(fctx, layer.wo);
         gguf_add_tensor(fctx, layer.ffn_norm);
-        gguf_add_tensor(fctx, layer.w1);
-        gguf_add_tensor(fctx, layer.w2);
-        gguf_add_tensor(fctx, layer.w3);
+        gguf_add_tensor(fctx, layer.ffn_gate);
+        gguf_add_tensor(fctx, layer.ffn_down);
+        gguf_add_tensor(fctx, layer.ffn_up);
     }
 }
 
@@ -953,9 +915,9 @@ static int64_t get_parameter_count(struct my_llama_model* model) {
         nx += ggml_nelements(layer.wv);
         nx += ggml_nelements(layer.wo);
         nx += ggml_nelements(layer.ffn_norm);
-        nx += ggml_nelements(layer.w1);
-        nx += ggml_nelements(layer.w2);
-        nx += ggml_nelements(layer.w3);
+        nx += ggml_nelements(layer.ffn_gate);
+        nx += ggml_nelements(layer.ffn_down);
+        nx += ggml_nelements(layer.ffn_up);
     }
     return nx;
 }
@@ -1046,7 +1008,7 @@ int main(int argc, char ** argv) {
     printf("%s: seen train_samples     %llu\n", __func__, (long long unsigned) train->train_samples);
     printf("%s: seen train_tokens      %llu\n", __func__, (long long unsigned) train->train_tokens);
     printf("%s: completed train_epochs %llu\n", __func__, (long long unsigned) train->train_epochs);
-    printf("%s: model_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(model.ctx) + model.data.size()), (float) (ggml_used_mem(model.ctx) + model.data.size()) / (1024.0f*1024.0f));
+    printf("%s: model_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(model.ctx) + ggml_backend_buffer_get_size(model.data)), (float) (ggml_used_mem(model.ctx) + ggml_backend_buffer_get_size(model.data)) / (1024.0f*1024.0f));
 
     if (params.only_write_model) {
         save_train_files_data save_data;
@@ -1073,11 +1035,6 @@ int main(int argc, char ** argv) {
     int n_vocab  = model.hparams.n_vocab;
     int n_batch  = params.common.n_batch;
 
-    std::vector<uint8_t> mem_input_data;
-    std::vector<uint8_t> mem_compute_data;
-
-    ggml_allocr * alloc = NULL;
-
     // context for input tensors without their data
     struct ggml_init_params ctx_input_params = {
         ggml_tensor_overhead() * 2, // mem_size
@@ -1091,16 +1048,10 @@ int main(int argc, char ** argv) {
     struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
 
     // measure required memory for input tensors
-    size_t max_input_size = GGML_PAD(ggml_nbytes(tokens_input), tensor_alignment) +
-                            GGML_PAD(ggml_nbytes(target_probs), tensor_alignment) +
-                            tensor_alignment;
-    printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
-
     // allocate input tensors
-    mem_input_data.resize(max_input_size);
-    alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
-    ggml_allocr_alloc(alloc, tokens_input);
-    ggml_allocr_alloc(alloc, target_probs);
+    ggml_backend_buffer_t input_data = ggml_backend_alloc_ctx_tensors_from_buft(ctx_input, ggml_backend_cpu_buffer_type());
+    size_t max_input_size = ggml_backend_buffer_get_size(input_data);
+    printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
 
     // context for compute tensors without their data
     const size_t estimated_compute_size_wo_data = (
@@ -1127,7 +1078,7 @@ int main(int argc, char ** argv) {
     // find best evaluation order
     for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
         ctx_compute = ggml_init(ctx_compute_params);
-        alloc = ggml_allocr_new_measure(tensor_alignment);
+        ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
         gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
         gf->order = (enum ggml_cgraph_eval_order) order;
         gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@@ -1140,9 +1091,10 @@ int main(int argc, char ** argv) {
             &logits, tokens_input, target_probs,
             n_tokens, n_batch,
             params.common.use_flash,
-            params.common.use_checkpointing
+            params.common.use_checkpointing,
+            true
         );
-        size_t max_compute_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+        size_t max_compute_size = ggml_gallocr_get_buffer_size(alloc, 0); // FIXME: this will still allocate the buffer
         if (max_compute_size < best_compute_size) {
             best_compute_size = max_compute_size;
             best_order = gf->order;
@@ -1157,9 +1109,8 @@ int main(int argc, char ** argv) {
         "invalid");
 
     // allocate compute tensors
-    mem_compute_data.resize(max_compute_size);
     ctx_compute = ggml_init(ctx_compute_params);
-    alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
+    ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
     gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
     gf->order = best_order;
     gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
@@ -1172,7 +1123,8 @@ int main(int argc, char ** argv) {
         &logits, tokens_input, target_probs,
         n_tokens, n_batch,
         params.common.use_flash,
-        params.common.use_checkpointing
+        params.common.use_checkpointing,
+        false
     );
 
     std::vector<llama_token> train_tokens;

flake.lock

@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1706732774,
-        "narHash": "sha256-hqJlyJk4MRpcItGYMF+3uHe8HvxNETWvlGtLuVpqLU0=",
+        "lastModified": 1707268954,
+        "narHash": "sha256-2en1kvde3cJVc3ZnTy8QeD2oKcseLFjYPLKhIGDanQ0=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "b8b232ae7b8b144397fdb12d20f592e5e7c1a64d",
+        "rev": "f8e2ebd66d097614d51a56a755450d4ae1632df1",
         "type": "github"
       },
       "original": {

ggml-alloc.h

@@ -6,88 +6,62 @@
 extern "C" {
 #endif
 
-struct ggml_backend;
-struct ggml_backend_buffer;
-struct ggml_backend_buffer_type;
-
-//
-// Legacy API
-//
-
-typedef struct ggml_allocr * ggml_allocr_t;
-
-// initialize allocator for use with CPU backend only
-GGML_API ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment);
-GGML_API ggml_allocr_t ggml_allocr_new_measure(size_t alignment);
-
-// initialize allocator for use with ggml-backend
-GGML_API ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer);
-GGML_API ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
-GGML_API ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend);
-
-GGML_API struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc);
-
-// tell the allocator to parse nodes following the order described in the list
-// you should call this if your graph are optimized to execute out-of-order
-GGML_API void   ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n);
-
-GGML_API void   ggml_allocr_free       (ggml_allocr_t alloc);
-GGML_API bool   ggml_allocr_is_measure (ggml_allocr_t alloc);
-GGML_API void   ggml_allocr_reset      (ggml_allocr_t alloc);
-GGML_API void   ggml_allocr_alloc      (ggml_allocr_t alloc, struct ggml_tensor * tensor);
-GGML_API size_t ggml_allocr_max_size   (ggml_allocr_t alloc);
-
-GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph);
-
-//
-// ggml-backend v2 API
-//
-
-// Separate tensor and graph allocator objects
-// This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
-// The original API is kept as a wrapper around the new API
+typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+typedef struct ggml_backend * ggml_backend_t;
 
 // Tensor allocator
 typedef struct ggml_tallocr * ggml_tallocr_t;
 
-GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment);
-GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment);
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_buft(struct ggml_backend_buffer_type * buft, size_t size);
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
-GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_buft(struct ggml_backend_buffer_type * buft);
-GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend);
-
-GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc);
-
-GGML_API void   ggml_tallocr_free       (ggml_tallocr_t talloc);
-GGML_API bool   ggml_tallocr_is_measure (ggml_tallocr_t talloc);
-GGML_API void   ggml_tallocr_reset      (ggml_tallocr_t talloc);
-GGML_API void   ggml_tallocr_alloc      (ggml_tallocr_t talloc, struct ggml_tensor * tensor);
-GGML_API size_t ggml_tallocr_max_size   (ggml_tallocr_t talloc);
-
+GGML_API ggml_tallocr_t ggml_tallocr_new(ggml_backend_buffer_t buffer);
+GGML_API void           ggml_tallocr_free(ggml_tallocr_t talloc);
+GGML_API void           ggml_tallocr_alloc(ggml_tallocr_t talloc, struct ggml_tensor * tensor);
 
 // Graph allocator
+/*
+  Example usage:
+    ggml_gallocr_t galloc = ggml_gallocr_new(ggml_bacckend_cpu_buffer_type());
+
+    // optional: create a worst-case graph and reserve the buffers to avoid reallocations
+    ggml_gallocr_reserve(galloc, build_graph(max_batch));
+
+    // allocate the graph
+    struct ggml_cgraph * graph = build_graph(batch);
+    ggml_gallocr_alloc_graph(galloc, graph);
+
+    printf("compute buffer size: %zu bytes\n", ggml_gallocr_get_buffer_size(galloc, 0));
+
+    // evaluate the graph
+    ggml_backend_graph_compute(backend, graph);
+*/
+
+// special tensor flags for use with the graph allocator:
+//   ggml_set_input(): all input tensors are allocated at the beginning of the graph in non-overlapping addresses
+//   ggml_set_output(): output tensors are never freed and never overwritten
+
 typedef struct ggml_gallocr * ggml_gallocr_t;
 
-GGML_API ggml_gallocr_t ggml_gallocr_new(void);
-GGML_API void   ggml_gallocr_free(ggml_gallocr_t galloc);
+GGML_API ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft);
+GGML_API ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs);
+GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
 
-GGML_API void   ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n);
-GGML_API size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph);
+// pre-allocate buffers from a measure graph - does not allocate or modify the graph
+// call with a worst-case graph to avoid buffer reallocations
+// not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
+// returns false if the buffer allocation failed
+GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+GGML_API bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids);
 
-// Allocate tensors from the allocators given by the hash table
-GGML_API void   ggml_gallocr_alloc_graph_n(
-                    ggml_gallocr_t galloc,
-                    struct ggml_cgraph * graph,
-                    struct ggml_hash_set hash_set,
-                    ggml_tallocr_t * hash_node_talloc);
+// automatic reallocation if the topology changes when using a single buffer
+// returns false if using multiple buffers and a re-allocation is needed (call ggml_gallocr_reserve_n first to set the node buffers)
+GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
 
+GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
 
 // Utils
 // Create a buffer and allocate all the tensors in a ggml_context
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, struct ggml_backend_buffer_type * buft);
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, struct ggml_backend * backend);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);
 
 #ifdef  __cplusplus
 }

ggml-backend.h

@@ -83,8 +83,9 @@ extern "C" {
 
     GGML_API ggml_backend_t ggml_backend_cpu_init(void);
 
-    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 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);
@@ -129,11 +130,7 @@ extern "C" {
 
         // in build_graph:
         build_graph(...) {
-            // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer)
-            alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu);
-            ggml_allocr_alloc(alloc_cpu, tensor);
-
-            // manually assigning nodes to a backend (optional, shouldn't be needed in most cases)
+            // manually assign nodes to a backend (optional, should not be needed in most cases)
             struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
             ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
         }
@@ -163,20 +160,19 @@ extern "C" {
     GGML_API ggml_backend_sched_t  ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size);
     GGML_API void                  ggml_backend_sched_free(ggml_backend_sched_t sched);
     // Initialize backend buffers from a measure graph
-    GGML_API void                  ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+    GGML_API bool                  ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
     // Get the number of splits of the last graph
     GGML_API int                   ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
 
-    GGML_API ggml_tallocr_t        ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
-    GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);
+    GGML_API size_t                ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
 
     GGML_API void                  ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
     GGML_API ggml_backend_t        ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 
     // Allocate and compute graph on the backend scheduler
-    GGML_API void                  ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API bool                  ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
 
-    // Reset all assignments and allocators - must be called before using the sched allocators to allocate inputs
+    // Reset all assignments and allocators - must be called before changing the node backends
     GGML_API void                  ggml_backend_sched_reset(ggml_backend_sched_t sched);
 
     // Set a callback to be called for each resulting node during graph compute

ggml-cuda.cu

@@ -150,8 +150,8 @@
 #define CUDA_USE_TENSOR_CORES
 #endif
 
-// max batch size to use MMQ kernels when tensor cores are available
-#define MMQ_MAX_BATCH_SIZE 32
+#define MMVQ_MAX_BATCH_SIZE  8 // max batch size to use MMVQ kernels
+#define  MMQ_MAX_BATCH_SIZE 32 // max batch size to use MMQ kernels when tensor cores are available
 
 #if defined(GGML_USE_HIPBLAS)
 #define __CUDA_ARCH__ 1300
@@ -5310,49 +5310,80 @@ template <bool need_check> static __global__ void
 #endif // __CUDA_ARCH__ >= CC_VOLTA
 }
 
-template <int ncols_y_template, int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot_q_cuda>
+template <int ncols_y, int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot_q_cuda>
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+// tell the compiler to use as many registers as it wants, see nwarps definition below
+__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
 static __global__ void mul_mat_vec_q(
     const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y_par, const int nrows_dst) {
+    const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
 
-    const int ncols_y = ncols_y_template != 0 ? ncols_y_template : ncols_y_par;
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+    constexpr int nwarps              = 1;
+    constexpr int rows_per_cuda_block = 1;
+#else
+    constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
+    constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
 
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-
-    if (row >= nrows_x) {
-        return;
-    }
-
-    const int blocks_per_row_x = ncols_x / qk;
-    const int blocks_per_col_y = nrows_y / QK8_1;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+    const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    const     int row0 = rows_per_cuda_block*blockIdx.x;
+    const     int blocks_per_row_x = ncols_x / qk;
+    const     int blocks_per_col_y = nrows_y / QK8_1;
+    constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi;
 
 // partial sum for each thread
-    float tmp[ncols_y_template != 0 ? ncols_y_template : 8] = {0.0f};
+    float tmp[ncols_y][rows_per_cuda_block] = {0.0f};
 
     const block_q_t  * x = (const block_q_t  *) vx;
     const block_q8_1 * y = (const block_q8_1 *) vy;
 
-    for (int i = threadIdx.x / (qi/vdr); i < blocks_per_row_x; i += blocks_per_warp) {
-        const int ibx = row*blocks_per_row_x + i; // x block index
+    for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
+        const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx
 
-        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
-
-        const int iqs  = vdr * (threadIdx.x % (qi/vdr)); // x block quant index when casting the quants to int
+        // x block quant index when casting the quants to int
+        const int kqs = vdr * (tid % (qi/vdr));
 
 #pragma unroll
         for (int j = 0; j < ncols_y; ++j) {
-            tmp[j] += vec_dot_q_cuda(&x[ibx], &y[j*blocks_per_col_y + iby], iqs);
+#pragma unroll
+            for (int i = 0; i < rows_per_cuda_block; ++i) {
+                tmp[j][i] += vec_dot_q_cuda(
+                    &x[kbx + (row0 + i)*blocks_per_row_x], &y[j*blocks_per_col_y + kby], kqs);
+            }
         }
     }
 
+    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][WARP_SIZE];
+    if (threadIdx.y > 0) {
+#pragma unroll
+        for (int j = 0; j < ncols_y; ++j) {
+#pragma unroll
+            for (int i = 0; i < rows_per_cuda_block; ++i) {
+                tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
+            }
+        }
+    }
+    __syncthreads();
+    if (threadIdx.y > 0) {
+        return;
+    }
+
     // sum up partial sums and write back result
 #pragma unroll
     for (int j = 0; j < ncols_y; ++j) {
-        tmp[j] = warp_reduce_sum(tmp[j]);
+#pragma unroll
+        for (int i = 0; i < rows_per_cuda_block; ++i) {
+#pragma unroll
+            for (int l = 0; l < nwarps-1; ++l) {
+                tmp[j][i] += tmp_shared[l][j][i][threadIdx.x];
+            }
+            tmp[j][i] = warp_reduce_sum(tmp[j][i]);
+        }
 
-        if (threadIdx.x == 0) {
-            dst[j*nrows_dst + row] = tmp[j];
+        if (threadIdx.x < rows_per_cuda_block) {
+            dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
         }
     }
 }
@@ -6831,48 +6862,77 @@ static void mul_mat_vec_q_cuda(
     const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
 
     GGML_ASSERT(ncols_x % qk == 0);
-    GGML_ASSERT(ncols_y <= 4);
+    GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE);
+
+    int id;
+    CUDA_CHECK(cudaGetDevice(&id));
+
+    int64_t nwarps = 1;
+    int64_t rows_per_cuda_block = 1;
+
+    if (g_device_caps[id].cc < CC_RDNA2) { // NVIDIA and AMD older than RDNA2
+        switch(ncols_y) {
+            case 1:
+                nwarps = 4;
+                rows_per_cuda_block = 1;
+                break;
+            case 2:
+            case 3:
+            case 4:
+                nwarps = 4;
+                rows_per_cuda_block = 2;
+                break;
+            case 5:
+            case 6:
+            case 7:
+            case 8:
+                nwarps = 2;
+                rows_per_cuda_block = 2;
+                break;
+            default:
+                GGML_ASSERT(false);
+                break;
+        }
+    }
+    const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block;
+    const dim3 block_nums(nblocks, 1, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
 
-    const int block_num_y = (nrows_x + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
     switch (ncols_y) {
         case 1:
             mul_mat_vec_q<1, qk, qi, block_q_t, vdr, vec_dot>
-                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
         case 2:
             mul_mat_vec_q<2, qk, qi, block_q_t, vdr, vec_dot>
-                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
         case 3:
             mul_mat_vec_q<3, qk, qi, block_q_t, vdr, vec_dot>
-                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
         case 4:
             mul_mat_vec_q<4, qk, qi, block_q_t, vdr, vec_dot>
-                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 5:
+            mul_mat_vec_q<5, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 6:
+            mul_mat_vec_q<6, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 7:
+            mul_mat_vec_q<7, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 8:
+            mul_mat_vec_q<8, qk, qi, block_q_t, vdr, vec_dot>
+                <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
             break;
-        // case 5:
-        //     mul_mat_vec_q<5, qk, qi, block_q_t, vdr, vec_dot>
-        //         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-        //     break;
-        // case 6:
-        //     mul_mat_vec_q<6, qk, qi, block_q_t, vdr, vec_dot>
-        //         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-        //     break;
-        // case 7:
-        //     mul_mat_vec_q<7, qk, qi, block_q_t, vdr, vec_dot>
-        //         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-        //     break;
-        // case 8:
-        //     mul_mat_vec_q<8, qk, qi, block_q_t, vdr, vec_dot>
-        //         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
-        //     break;
         default:
             GGML_ASSERT(false);
-            // mul_mat_vec_q<0, qk, qi, block_q_t, vdr, vec_dot>
-            //     <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst);
             break;
     }
 }
@@ -9696,7 +9756,7 @@ static __global__ void k_compute_batched_ptrs(
     ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2 + i13*nbd3;
 }
 
-static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_cuda_mul_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
 
@@ -9854,39 +9914,69 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
 
     int64_t min_compute_capability = INT_MAX;
 
+    bool any_pascal_with_slow_fp16 = false;
     if (split) {
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
         auto & tensor_split = buft_ctx->tensor_split;
         for (int id = 0; id < g_device_count; ++id) {
-            if (min_compute_capability > g_device_caps[id].cc && tensor_split[id] < (id + 1 < g_device_count ? tensor_split[id + 1] : 1.0f)) {
+            // skip devices that are not going to do any work:
+            if (tensor_split[id] >= (id + 1 < g_device_count ? tensor_split[id + 1] : 1.0f)) {
+                continue;
+            }
+
+            if (min_compute_capability > g_device_caps[id].cc) {
                 min_compute_capability = g_device_caps[id].cc;
             }
+            if (g_device_caps[id].cc == 610) {
+                any_pascal_with_slow_fp16 = true;
+            }
         }
     } else {
-        min_compute_capability = g_device_caps[g_main_device].cc;
+        min_compute_capability    = g_device_caps[g_main_device].cc;
+        any_pascal_with_slow_fp16 = g_device_caps[g_main_device].cc == 610;
     }
 
+    // check data types and tensor shapes for custom matrix multiplication kernels:
+    bool use_dequantize_mul_mat_vec = (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src1->ne[1] == 1;
+
+    bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+
+    bool              use_mul_mat_q =  ggml_cuda_supports_mmq(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
 #if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
 
     const bool fp16_performance_good = min_compute_capability >= CC_RDNA1;
-    bool               use_mul_mat_q = ggml_is_quantized(src0->type);
+
 #ifdef CUDA_USE_TENSOR_CORES
     use_mul_mat_q = use_mul_mat_q && min_compute_capability < CC_RDNA3;
 #endif // CUDA_USE_TENSOR_CORES
 
 #else
 
-    const bool fp16_performance_good = min_compute_capability >= CC_VOLTA;
-    bool               use_mul_mat_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
+    // fp16 performance is good on Volta or newer and on P100 (compute capability 6.0)
+    const bool fp16_performance_good = min_compute_capability >= CC_PASCAL && !any_pascal_with_slow_fp16;
+
+    // mmvq and mmq need the __dp4a instruction which on NVIDIA is only available for CC >= 6.1
+    use_mul_mat_vec_q = use_mul_mat_vec_q && min_compute_capability >= MIN_CC_DP4A;
+    use_mul_mat_q     = use_mul_mat_q     && min_compute_capability >= MIN_CC_DP4A;
+
 #ifdef CUDA_USE_TENSOR_CORES
     // when tensor cores are available, use them for large batch size
     // ref: https://github.com/ggerganov/llama.cpp/pull/3776
-    use_mul_mat_q = use_mul_mat_q && !(fp16_performance_good && src1->ne[1] > MMQ_MAX_BATCH_SIZE);
+    use_mul_mat_q     = use_mul_mat_q     && (!fp16_performance_good || src1->ne[1] <= MMQ_MAX_BATCH_SIZE);
 #endif // CUDA_USE_TENSOR_CORES
 
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
 
-    use_mul_mat_q = use_mul_mat_q && ggml_cuda_supports_mmq(src0->type);
+    // if mmvq is available it's a better choice than dmmv:
+#ifndef GGML_CUDA_FORCE_DMMV
+    use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
+#endif // GGML_CUDA_FORCE_DMMV
 
     // debug helpers
     //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
@@ -9904,33 +9994,15 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
         ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
     } else if (!split && all_on_device && fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
         // KQ + KQV multi-batch
-        ggml_cuda_mul_mat_mat_batched_cublas(src0, src1, dst);
-    } else if (src0->type == GGML_TYPE_F32) {
-        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
-    } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) {
-        if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src1->type == GGML_TYPE_F32) {
-#ifdef GGML_CUDA_FORCE_DMMV
-            const bool use_mul_mat_vec_q = false;
-#else
-            const bool use_mul_mat_vec_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
-#endif // GGML_CUDA_FORCE_DMMV
-
-            if (use_mul_mat_vec_q) {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
-            } else {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
-            }
-        } else {
-            if (src1->ne[1] <= 4 && min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type) && src1->type == GGML_TYPE_F32) {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
-            } else if (use_mul_mat_q) {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
-            } else {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
-            }
-        }
+        ggml_cuda_mul_mat_batched_cublas(src0, src1, dst);
+    } else if (use_dequantize_mul_mat_vec) {
+        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
+    } else if (use_mul_mat_vec_q) {
+        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
+    } else if (use_mul_mat_q) {
+        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
     } else {
-        GGML_ASSERT(false);
+        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
     }
 }
 

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);
@@ -3729,15 +3819,15 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx,
         /* Compute combined scale for the block */
         const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
 
-        __m256i bx = bytes_from_nibbles_32(x[i].qs);
+        __m256i qx = bytes_from_nibbles_32(x[i].qs);
 
         // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
         const __m256i off = _mm256_set1_epi8( 8 );
-        bx = _mm256_sub_epi8( bx, off );
+        qx = _mm256_sub_epi8( qx, off );
 
-        __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
-        const __m256 q = mul_sum_i8_pairs_float(bx, by);
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
         /* Multiply q with scale and accumulate */
         acc = _mm256_fmadd_ps( d, q, acc );
@@ -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);
@@ -4028,10 +4196,10 @@ void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restri
         const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
 
         // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
-        const __m256i bx = bytes_from_nibbles_32(x[i].qs);
-        const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs );
+        const __m256i qx = bytes_from_nibbles_32(x[i].qs);
+        const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[i].qs );
 
-        const __m256 xy = mul_sum_us8_pairs_float(bx, by);
+        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
 
         // Accumulate d0*d1*x*y
 #if defined(__AVX2__)
@@ -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;
@@ -4245,14 +4418,14 @@ void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restri
         /* Compute combined scale for the block */
         const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
 
-        __m256i bx = bytes_from_nibbles_32(x[i].qs);
+        __m256i qx = bytes_from_nibbles_32(x[i].qs);
         __m256i bxhi = bytes_from_bits_32(x[i].qh);
         bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
-        bx = _mm256_or_si256(bx, bxhi);
+        qx = _mm256_or_si256(qx, bxhi);
 
-        __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
-        const __m256 q = mul_sum_i8_pairs_float(bx, by);
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
         /* Multiply q with scale and accumulate */
         acc = _mm256_fmadd_ps(d, q, acc);
@@ -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;
@@ -4544,15 +4722,15 @@ void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restri
 
         summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
 
-        __m256i bx = bytes_from_nibbles_32(x[i].qs);
+        __m256i qx = bytes_from_nibbles_32(x[i].qs);
         __m256i bxhi = bytes_from_bits_32(x[i].qh);
         bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
-        bx = _mm256_or_si256(bx, bxhi);
+        qx = _mm256_or_si256(qx, bxhi);
 
         const __m256 dy = _mm256_set1_ps(y[i].d);
-        const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        const __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
-        const __m256 q = mul_sum_us8_pairs_float(bx, by);
+        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
 
         acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
     }
@@ -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);
@@ -4731,10 +4973,10 @@ void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restri
     for (int i = 0; i < nb; ++i) {
         // Compute combined scale for the block
         const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
-        __m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs);
-        __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
+        __m256i qx = _mm256_loadu_si256((const __m256i *)x[i].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
 
-        const __m256 q = mul_sum_i8_pairs_float(bx, by);
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
 
         // Multiply q with scale and accumulate
 #if defined(__AVX2__)
@@ -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_1 = {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_3 = {iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]};
+            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 = ggml_vld1q_u32(iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]);
+            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 = 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_1_q8_1(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, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-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_q8_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, 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, 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, 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, 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_q3_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, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-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_q6_K_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, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy);
-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_iq3_xxs_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, 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, 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, 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, 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, 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, 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, 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-sycl.cpp

@@ -11578,11 +11578,8 @@ static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst,
     }
     char * dst_ptr = (char *) dst;
 
-    const int64_t ne0 = src->ne[0];
-    const int64_t nb0 = src->nb[0];
-    const int64_t nb1 = src->nb[1];
-    const int64_t nb2 = src->nb[2];
-    const int64_t nb3 = src->nb[3];
+    GGML_TENSOR_LOCALS_1(int64_t, ne, src, ne);
+    GGML_TENSOR_LOCALS(int64_t, nb, src, nb);
     const enum ggml_type type = src->type;
     const int64_t ts = ggml_type_size(type);
     const int64_t bs = ggml_blck_size(type);
@@ -12426,9 +12423,7 @@ inline void ggml_sycl_op_alibi(const ggml_tensor *src0, const ggml_tensor *src1,
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
+    GGML_TENSOR_LOCALS_3(int64_t, ne0, src0, ne);
     const int64_t nrows = ggml_nrows(src0);
 
     //const int n_past = ((int32_t *) dst->op_params)[0];
@@ -12758,15 +12753,9 @@ static void ggml_sycl_op_mul_mat(const ggml_tensor *src0,
                                  ggml_sycl_op_mul_mat_t op,
                                  const bool convert_src1_to_q8_1) try {
 
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
 
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
     const int64_t nrows1 = ggml_nrows(src1);
 
     GGML_ASSERT(ne03 == ne13);
@@ -13337,23 +13326,13 @@ static void ggml_sycl_mul_mat_mat_batched_sycl(const ggml_tensor *src0,
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = src0->ne[0]; GGML_UNUSED(ne00);
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
 
-    const int64_t nb01 = src0->nb[1];
-    const int64_t nb02 = src0->nb[2]; GGML_UNUSED(nb02);
-    const int64_t nb03 = src0->nb[3]; GGML_UNUSED(nb03);
+    GGML_TENSOR_LOCALS(int64_t, nb0, src0, nb);
 
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
 
-    const int64_t nb11 = src1->nb[1];
-    const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
-    const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
+    GGML_TENSOR_LOCALS(int64_t, nb1, src1, nb);
 
     const int64_t ne1 = ggml_nelements(src1);
     const int64_t ne  = ggml_nelements(dst);
@@ -13655,23 +13634,15 @@ static void ggml_sycl_mul_mat_id_sycl(ggml_tensor * dst) {
     GGML_ASSERT(src00->backend != GGML_BACKEND_GPU_SPLIT);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
-    const int64_t ne00 = src00->ne[0]; GGML_UNUSED(ne00);
-    const int64_t ne01 = src00->ne[1];
-    const int64_t ne02 = src00->ne[2];
-    const int64_t ne03 = src00->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne0, src00, ne);
 
     //const int64_t nb01 = src00->nb[1];
-    const int64_t nb02 = src00->nb[2]; GGML_UNUSED(nb02);
-    const int64_t nb03 = src00->nb[3]; GGML_UNUSED(nb03);
+    GGML_TENSOR_LOCALS(int64_t, nb0, src00, nb);
 
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
 
+    GGML_TENSOR_LOCALS(int64_t, nb1, src1, nb);
     //const int64_t nb11 = src1->nb[1];
-    const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
-    const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
 
     const int64_t ne1 = ggml_nelements(src1);
     const int64_t ne  = ggml_nelements(dst);
@@ -13940,25 +13911,7 @@ static void ggml_sycl_cpy(const ggml_tensor *src0, const ggml_tensor *src1,
     GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
     GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
 
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-
-    const int64_t nb00 = src0->nb[0];
-    const int64_t nb01 = src0->nb[1];
-    const int64_t nb02 = src0->nb[2];
-    const int64_t nb03 = src0->nb[3];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-
-
-    const int64_t nb10 = src1->nb[0];
-    const int64_t nb11 = src1->nb[1];
-    const int64_t nb12 = src1->nb[2];
-    const int64_t nb13 = src1->nb[3];
+    GGML_TENSOR_BINARY_OP_LOCALS;
 
     SYCL_CHECK(ggml_sycl_set_device(g_main_device));
     dpct::queue_ptr main_stream = g_syclStreams[g_main_device_index][0];

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) {
-#ifdef GGML_VULKAN_DEBUG
-    std::cerr << "ggml_vk_guess_matmul_pipeline(" << bit16_x << ", " << bit16_y << ", " << m << ", " << n << ", " << aligned << ")";
-#endif
+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) {
     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 << "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 << "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;
 #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_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * 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(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]);   }
@@ -2607,7 +2649,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         /*.nb           =*/ { 0, 0, 0, 0 },
         /*.op           =*/ GGML_OP_NONE,
         /*.op_params    =*/ { 0 },
-        /*.is_param     =*/ false,
+        /*.flags        =*/ 0,
         /*.grad         =*/ NULL,
         /*.src          =*/ { NULL },
         /*.perf_runs    =*/ 0,
@@ -6509,7 +6551,7 @@ struct ggml_tensor * ggml_cross_entropy_loss_back(
 void ggml_set_param(
         struct ggml_context * ctx,
         struct ggml_tensor * tensor) {
-    tensor->is_param = true;
+    tensor->flags |= GGML_TENSOR_FLAG_PARAM;
 
     GGML_ASSERT(tensor->grad == NULL);
     tensor->grad = ggml_dup_tensor(ctx, tensor);
@@ -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) {
-                    vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
+                for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ir0 += nrc) {
+                    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_fp16_t *)    wdata_src + i1n,
-                        (ggml_fp16_t *) wdata_kernel + i00*ne02);
+                ggml_vec_dot_f16(ne02, &v, 0,
+                        (ggml_fp16_t *)    wdata_src + i1n, 0,
+                        (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,
-                        wdata_src + i1n,
-                        wdata_kernel + i00*ne02);
+                ggml_vec_dot_f32(ne02, &v, 0,
+                        wdata_src + i1n, 0,
+                        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,
-                                wdata_src + i1n,
-                                wdata_kernel + i01*ne00*ne03 + i00*ne03);
+                        ggml_vec_dot_f16(ne03, &v, 0,
+                                wdata_src + i1n, 0,
+                                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,
-                    (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
-                    (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
+                    S + i1, 0,
+                    (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)), 0,
+                    (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 *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)),
-                    S);
+                    (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)), 0,
+                    (float *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)), 0,
+                    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,
-                        (ggml_fp16_t *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
-                        (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
+                        S + i1, 0,
+                        (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)), 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)),
-                        (ggml_fp16_t *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)),
-                        S16);
+                        (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)), 0,
+                        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,
-                    (ggml_fp16_t *) ((char *) b0->data + (ib01*nbb01 + ib02*nbb02 + ib03*nbb03)),
-                    (ggml_fp16_t *) ((char *)  a->data + ( ia1*nba1  +  ia2*nba2  +  ia3*nba3)));
+                    S + i1, 0,
+                    (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)), 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)),
-                        (ggml_fp16_t *) ((char *) c0->data  + (         ic*nbc01 + i2*nbc02 + i3*nbc03)),
-                        S16);
+                        (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)), 0,
+                        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,
-                            (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
-                            (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
+                            S + i1, 0,
+                            (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)), 0,
+                            (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);
 
@@ -15311,7 +15367,7 @@ static struct ggml_tensor * ggml_recompute_graph_node(
         return NULL;
     }
 
-    if (node->is_param) {
+    if (node->flags & GGML_TENSOR_FLAG_PARAM) {
         return node;
     }
 
@@ -15345,7 +15401,7 @@ static struct ggml_tensor * ggml_recompute_graph_node(
 
     clone->op       = node->op;
     clone->grad     = node->grad;
-    clone->is_param = node->is_param;
+    clone->flags    = node->flags;
     clone->extra    = node->extra;
     for (int k = 0; k < GGML_MAX_DIMS; ++k) {
         clone->nb[k] = node->nb[k];
@@ -16377,7 +16433,7 @@ void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph *
     for (int i = 0; i < gf->n_nodes; i++) {
         struct ggml_tensor * node = gf->nodes[i];
 
-        if (node->is_param) {
+        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
             GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node);
             ggml_build_forward_expand(gb, node->grad);
         }
@@ -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;
 };
 
@@ -17862,7 +17918,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) {
         GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n",
                 i,
                 node->ne[0], node->ne[1], node->ne[2],
-                ggml_op_name(node->op), node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs,
+                ggml_op_name(node->op), (node->flags & GGML_TENSOR_FLAG_PARAM) ? "x" : node->grad ? "g" : " ", node->perf_runs,
                 (double) node->perf_cycles  / (double) ggml_cycles_per_ms(),
                 (double) node->perf_cycles  / (double) ggml_cycles_per_ms() / (double) node->perf_runs,
                 (double) node->perf_time_us / 1000.0,
@@ -17955,7 +18011,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
             continue;
         }
 
-        if (node->is_param) {
+        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
             snprintf(color, sizeof(color), "yellow");
         } else if (node->grad) {
             if (ggml_graph_find(gf, node)) {
@@ -18129,7 +18185,7 @@ static enum ggml_opt_result ggml_opt_adam(
     int np = 0;
     int64_t nx = 0;
     for (int i = 0; i < gf->n_nodes; ++i) {
-        if (gf->nodes[i]->is_param) {
+        if (gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) {
             GGML_PRINT_DEBUG("found param %d: grad->op = %d\n", np, gf->nodes[i]->grad->op);
 
             GGML_ASSERT(np < GGML_MAX_PARAMS);
@@ -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) {
@@ -18492,7 +18548,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
     int np = 0;
     int nx = 0;
     for (int i = 0; i < gf->n_nodes; ++i) {
-        if (gf->nodes[i]->is_param) {
+        if (gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) {
             GGML_PRINT_DEBUG("found param %d: grad->op = %d\n", np, gf->nodes[i]->grad->op);
 
             GGML_ASSERT(np < GGML_MAX_PARAMS);
@@ -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, &yy, &lm_y[end[0]*nx], &lm_y[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, 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);
@@ -18967,6 +19023,16 @@ enum ggml_opt_result ggml_opt_resume_g(
 
 ////////////////////////////////////////////////////////////////////////////////
 
+void ggml_set_input(struct ggml_tensor * tensor) {
+    tensor->flags |= GGML_TENSOR_FLAG_INPUT;
+}
+
+void ggml_set_output(struct ggml_tensor * tensor) {
+    tensor->flags |= GGML_TENSOR_FLAG_OUTPUT;
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
 void ggml_quantize_init(enum ggml_type type) {
     ggml_critical_section_start();
 
@@ -20611,4 +20677,12 @@ 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

@@ -505,11 +505,17 @@ extern "C" {
 
     enum ggml_log_level {
         GGML_LOG_LEVEL_ERROR = 2,
-        GGML_LOG_LEVEL_WARN = 3,
-        GGML_LOG_LEVEL_INFO = 4,
+        GGML_LOG_LEVEL_WARN  = 3,
+        GGML_LOG_LEVEL_INFO  = 4,
         GGML_LOG_LEVEL_DEBUG = 5
     };
 
+    enum ggml_tensor_flag {
+        GGML_TENSOR_FLAG_INPUT  = 1,
+        GGML_TENSOR_FLAG_OUTPUT = 2,
+        GGML_TENSOR_FLAG_PARAM  = 4,
+    };
+
     // ggml object
     struct ggml_object {
         size_t offs;
@@ -543,7 +549,7 @@ extern "C" {
         // op params - allocated as int32_t for alignment
         int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 
-        bool is_param;
+        int32_t flags;
 
         struct ggml_tensor * grad;
         struct ggml_tensor * src[GGML_MAX_SRC];
@@ -567,6 +573,11 @@ 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 {
@@ -576,8 +587,8 @@ extern "C" {
         int n_threads;
 
         // abort ggml_graph_compute when true
-        bool (*abort_callback)(void * data);
-        void * abort_callback_data;
+        ggml_abort_callback abort_callback;
+        void *              abort_callback_data;
     };
 
     enum ggml_cgraph_eval_order {
@@ -2087,6 +2098,12 @@ extern "C" {
             ggml_opt_callback callback,
             void * callback_data);
 
+    //
+    // tensor flags
+    //
+    GGML_API void ggml_set_input(struct ggml_tensor * tensor);
+    GGML_API void ggml_set_output(struct ggml_tensor * tensor);
+
     //
     // quantization
     //
@@ -2273,6 +2290,7 @@ 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 +2304,8 @@ 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 +2317,7 @@ 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")

gguf-py/gguf/constants.py

@@ -40,6 +40,7 @@ class Keys:
         TENSOR_DATA_LAYOUT    = "{arch}.tensor_data_layout"
         EXPERT_COUNT          = "{arch}.expert_count"
         EXPERT_USED_COUNT     = "{arch}.expert_used_count"
+        POOLING_LAYER         = "{arch}.pooling_layer"
 
     class Attention:
         HEAD_COUNT        = "{arch}.attention.head_count"
@@ -50,6 +51,7 @@ class Keys:
         VALUE_LENGTH      = "{arch}.attention.value_length"
         LAYERNORM_EPS     = "{arch}.attention.layer_norm_epsilon"
         LAYERNORM_RMS_EPS = "{arch}.attention.layer_norm_rms_epsilon"
+        CAUSAL            = "{arch}.attention.causal"
 
     class Rope:
         DIMENSION_COUNT      = "{arch}.rope.dimension_count"
@@ -60,22 +62,23 @@ class Keys:
         SCALING_FINETUNED    = "{arch}.rope.scaling.finetuned"
 
     class Tokenizer:
-        MODEL         = "tokenizer.ggml.model"
-        LIST          = "tokenizer.ggml.tokens"
-        TOKEN_TYPE    = "tokenizer.ggml.token_type"
-        SCORES        = "tokenizer.ggml.scores"
-        MERGES        = "tokenizer.ggml.merges"
-        BOS_ID        = "tokenizer.ggml.bos_token_id"
-        EOS_ID        = "tokenizer.ggml.eos_token_id"
-        UNK_ID        = "tokenizer.ggml.unknown_token_id"
-        SEP_ID        = "tokenizer.ggml.seperator_token_id"
-        PAD_ID        = "tokenizer.ggml.padding_token_id"
-        ADD_BOS       = "tokenizer.ggml.add_bos_token"
-        ADD_EOS       = "tokenizer.ggml.add_eos_token"
-        ADD_PREFIX    = "tokenizer.ggml.add_space_prefix"
-        HF_JSON       = "tokenizer.huggingface.json"
-        RWKV          = "tokenizer.rwkv.world"
-        CHAT_TEMPLATE = "tokenizer.chat_template"
+        MODEL            = "tokenizer.ggml.model"
+        LIST             = "tokenizer.ggml.tokens"
+        TOKEN_TYPE       = "tokenizer.ggml.token_type"
+        TOKEN_TYPE_COUNT = "tokenizer.ggml.token_type_count"  # for BERT-style token types
+        SCORES           = "tokenizer.ggml.scores"
+        MERGES           = "tokenizer.ggml.merges"
+        BOS_ID           = "tokenizer.ggml.bos_token_id"
+        EOS_ID           = "tokenizer.ggml.eos_token_id"
+        UNK_ID           = "tokenizer.ggml.unknown_token_id"
+        SEP_ID           = "tokenizer.ggml.seperator_token_id"
+        PAD_ID           = "tokenizer.ggml.padding_token_id"
+        ADD_BOS          = "tokenizer.ggml.add_bos_token"
+        ADD_EOS          = "tokenizer.ggml.add_eos_token"
+        ADD_PREFIX       = "tokenizer.ggml.add_space_prefix"
+        HF_JSON          = "tokenizer.huggingface.json"
+        RWKV             = "tokenizer.rwkv.world"
+        CHAT_TEMPLATE    = "tokenizer.chat_template"
 
 
 #
@@ -84,27 +87,28 @@ class Keys:
 
 
 class MODEL_ARCH(IntEnum):
-    LLAMA     = auto()
-    FALCON    = auto()
-    BAICHUAN  = auto()
-    GPT2      = auto()
-    GPTJ      = auto()
-    GPTNEOX   = auto()
-    MPT       = auto()
-    STARCODER = auto()
-    PERSIMMON = auto()
-    REFACT    = auto()
-    BERT      = auto()
-    BLOOM     = auto()
-    STABLELM  = auto()
-    QWEN      = auto()
-    QWEN2     = auto()
-    PHI2      = auto()
-    PLAMO     = auto()
-    CODESHELL = auto()
-    ORION     = auto()
+    LLAMA      = auto()
+    FALCON     = auto()
+    BAICHUAN   = auto()
+    GPT2       = auto()
+    GPTJ       = auto()
+    GPTNEOX    = auto()
+    MPT        = auto()
+    STARCODER  = auto()
+    PERSIMMON  = auto()
+    REFACT     = auto()
+    BERT       = auto()
+    NOMIC_BERT = auto()
+    BLOOM      = auto()
+    STABLELM   = auto()
+    QWEN       = auto()
+    QWEN2      = auto()
+    PHI2       = auto()
+    PLAMO      = auto()
+    CODESHELL  = auto()
+    ORION      = auto()
     INTERNLM2  = auto()
-    MINICPM   = auto()
+    MINICPM    = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -122,6 +126,7 @@ class MODEL_TENSOR(IntEnum):
     ATTN_OUT        = auto()
     ATTN_NORM       = auto()
     ATTN_NORM_2     = auto()
+    ATTN_OUT_NORM   = auto()
     ATTN_ROT_EMBD   = auto()
     FFN_GATE_INP    = auto()
     FFN_NORM        = auto()
@@ -134,6 +139,7 @@ class MODEL_TENSOR(IntEnum):
     FFN_UP_EXP      = auto()
     ATTN_Q_NORM     = auto()
     ATTN_K_NORM     = auto()
+    LAYER_OUT_NORM  = auto()
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -148,6 +154,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.PERSIMMON:      "persimmon",
     MODEL_ARCH.REFACT:         "refact",
     MODEL_ARCH.BERT:           "bert",
+    MODEL_ARCH.NOMIC_BERT:     "nomic-bert",
     MODEL_ARCH.BLOOM:          "bloom",
     MODEL_ARCH.STABLELM:       "stablelm",
     MODEL_ARCH.QWEN:           "qwen",
@@ -178,6 +185,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.ATTN_ROT_EMBD:   "blk.{bid}.attn_rot_embd",
     MODEL_TENSOR.ATTN_Q_NORM:     "blk.{bid}.attn_q_norm",
     MODEL_TENSOR.ATTN_K_NORM:     "blk.{bid}.attn_k_norm",
+    MODEL_TENSOR.ATTN_OUT_NORM:   "blk.{bid}.attn_output_norm",
     MODEL_TENSOR.FFN_GATE_INP:    "blk.{bid}.ffn_gate_inp",
     MODEL_TENSOR.FFN_NORM:        "blk.{bid}.ffn_norm",
     MODEL_TENSOR.FFN_GATE:        "blk.{bid}.ffn_gate",
@@ -187,6 +195,7 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.FFN_GATE_EXP:    "blk.{bid}.ffn_gate.{xid}",
     MODEL_TENSOR.FFN_DOWN_EXP:    "blk.{bid}.ffn_down.{xid}",
     MODEL_TENSOR.FFN_UP_EXP:      "blk.{bid}.ffn_up.{xid}",
+    MODEL_TENSOR.LAYER_OUT_NORM:  "blk.{bid}.layer_output_norm",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -262,17 +271,32 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
     ],
     MODEL_ARCH.BERT: [
         MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
         MODEL_TENSOR.TOKEN_TYPES,
         MODEL_TENSOR.POS_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
-        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_OUT_NORM,
         MODEL_TENSOR.ATTN_Q,
         MODEL_TENSOR.ATTN_K,
         MODEL_TENSOR.ATTN_V,
         MODEL_TENSOR.ATTN_OUT,
-        MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.LAYER_OUT_NORM,
+    ],
+    MODEL_ARCH.NOMIC_BERT: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.TOKEN_EMBD_NORM,
+        MODEL_TENSOR.TOKEN_TYPES,
+        MODEL_TENSOR.POS_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.LAYER_OUT_NORM,
     ],
     MODEL_ARCH.MPT: [
         MODEL_TENSOR.TOKEN_EMBD,

gguf-py/gguf/gguf_writer.py

@@ -357,6 +357,12 @@ class GGUFWriter:
     def add_layer_norm_rms_eps(self, value: float) -> None:
         self.add_float32(Keys.Attention.LAYERNORM_RMS_EPS.format(arch=self.arch), value)
 
+    def add_causal_attention(self, value: bool) -> None:
+        self.add_bool(Keys.Attention.CAUSAL.format(arch=self.arch), value)
+
+    def add_pooling_layer(self, value: bool) -> None:
+        self.add_bool(Keys.LLM.POOLING_LAYER.format(arch=self.arch), value)
+
     def add_rope_dimension_count(self, count: int) -> None:
         self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
 
@@ -387,6 +393,9 @@ class GGUFWriter:
     def add_token_types(self, types: Sequence[TokenType] | Sequence[int]) -> None:
         self.add_array(Keys.Tokenizer.TOKEN_TYPE, types)
 
+    def add_token_type_count(self, value: int) -> None:
+        self.add_uint32(Keys.Tokenizer.TOKEN_TYPE_COUNT, value)
+
     def add_token_scores(self, scores: Sequence[float]) -> None:
         self.add_array(Keys.Tokenizer.SCORES, scores)
 

gguf-py/gguf/tensor_mapping.py

@@ -15,7 +15,7 @@ class TensorNameMap:
             "word_embeddings",                           # bloom
             "model.embed_tokens",                        # llama-hf
             "tok_embeddings",                            # llama-pth
-            "embeddings.word_embeddings",                # bert
+            "embeddings.word_embeddings",                # bert nomic-bert
             "language_model.embedding.word_embeddings",  # persimmon
             "wte",                                       # gpt2
             "transformer.embd.wte",                      # phi2
@@ -24,12 +24,14 @@ class TensorNameMap:
 
         # Token type embeddings
         MODEL_TENSOR.TOKEN_TYPES: (
-            "embeddings.token_type_embeddings",  # bert
+            "embeddings.token_type_embeddings",  # bert nomic-bert
         ),
 
         # Normalization of token embeddings
         MODEL_TENSOR.TOKEN_EMBD_NORM: (
             "word_embeddings_layernorm",  # bloom
+            "embeddings.LayerNorm",       # bert
+            "emb_ln",                     # nomic-bert
         ),
 
         # Position embeddings
@@ -54,7 +56,6 @@ class TensorNameMap:
             "transformer.ln_f",                        # gpt2 gpt-j falcon
             "model.norm",                              # llama-hf baichuan internlm2
             "norm",                                    # llama-pth
-            "embeddings.LayerNorm",                    # bert
             "transformer.norm_f",                      # mpt
             "ln_f",                                    # refact bloom qwen gpt2
             "language_model.encoder.final_layernorm",  # persimmon
@@ -79,7 +80,6 @@ class TensorNameMap:
             "transformer.h.{bid}.ln_mlp",                           # falcon40b
             "model.layers.{bid}.input_layernorm",                   # llama-hf
             "layers.{bid}.attention_norm",                          # llama-pth
-            "encoder.layer.{bid}.attention.output.LayerNorm",       # bert
             "language_model.encoder.layers.{bid}.input_layernorm",  # persimmon
             "model.layers.{bid}.ln1",                               # yi
             "h.{bid}.ln_1",                                         # gpt2
@@ -104,6 +104,7 @@ class TensorNameMap:
             "model.layers.{bid}.self_attn.query_key_value",                        # persimmon
             "h.{bid}.attn.c_attn",                                                 # gpt2
             "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
+            "encoder.layers.{bid}.attn.Wqkv",                                      # nomic-bert
         ),
 
         # Attention query
@@ -153,6 +154,13 @@ class TensorNameMap:
             "transformer.h.{bid}.mixer.out_proj",                        # phi2
             "model.layers.layers.{bid}.self_attn.o_proj",                # plamo
             "model.layers.{bid}.attention.wo",                           # internlm2
+            "encoder.layers.{bid}.attn.out_proj",                        # nomic-bert
+        ),
+
+        # Attention output norm
+        MODEL_TENSOR.ATTN_OUT_NORM: (
+            "encoder.layer.{bid}.attention.output.LayerNorm",  # bert
+            "encoder.layers.{bid}.norm1",                      # nomic-bert
         ),
 
         # Rotary embeddings
@@ -171,7 +179,6 @@ class TensorNameMap:
             "transformer.blocks.{bid}.norm_2",                               # mpt
             "model.layers.{bid}.post_attention_layernorm",                   # llama-hf
             "layers.{bid}.ffn_norm",                                         # llama-pth
-            "encoder.layer.{bid}.output.LayerNorm",                          # bert
             "language_model.encoder.layers.{bid}.post_attention_layernorm",  # persimmon
             "model.layers.{bid}.ln2",                                        # yi
             "h.{bid}.ln_2",                                                  # gpt2
@@ -202,6 +209,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.fc1",                             # phi2
             "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
             "model.layers.{bid}.feed_forward.w3",                     # internlm2
+            "encoder.layers.{bid}.mlp.fc11",                          # nomic-bert
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -221,6 +229,7 @@ class TensorNameMap:
             "transformer.h.{bid}.mlp.w2",                 # qwen
             "model.layers.layers.{bid}.mlp.gate_proj",    # plamo
             "model.layers.{bid}.feed_forward.w1",         # internlm2
+            "encoder.layers.{bid}.mlp.fc12",              # nomic-bert
         ),
 
         MODEL_TENSOR.FFN_GATE_EXP: (
@@ -246,6 +255,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.fc2",                             # phi2
             "model.layers.layers.{bid}.mlp.down_proj",                # plamo
             "model.layers.{bid}.feed_forward.w2",                     # internlm2
+            "encoder.layers.{bid}.mlp.fc2",                           # nomic-bert
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (
@@ -266,6 +276,11 @@ class TensorNameMap:
         MODEL_TENSOR.ROPE_FREQS: (
             "language_model.encoder.layers.{bid}.self_attention.rotary_emb.inv_freq",  # persimmon
         ),
+
+        MODEL_TENSOR.LAYER_OUT_NORM: (
+            "encoder.layer.{bid}.output.LayerNorm",  # bert
+            "encoder.layers.{bid}.norm2",            # nomic-bert
+        )
     }
 
     mapping: dict[str, tuple[MODEL_TENSOR, str]]

llama.h

@@ -61,6 +61,7 @@ extern "C" {
     enum llama_vocab_type {
         LLAMA_VOCAB_TYPE_SPM = 0, // SentencePiece
         LLAMA_VOCAB_TYPE_BPE = 1, // Byte Pair Encoding
+        LLAMA_VOCAB_TYPE_WPM = 2, // WordPiece
     };
 
     enum llama_token_type {
@@ -235,6 +236,7 @@ extern "C" {
         bool logits_all;  // the llama_eval() call computes all logits, not just the last one (DEPRECATED - set llama_batch.logits instead)
         bool embedding;   // embedding mode only
         bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
+        bool do_pooling;  // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer)
     };
 
     // model quantization parameters
@@ -627,6 +629,10 @@ extern "C" {
     // shape: [n_embd] (1-dimensional)
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 
+    // Get the embeddings for the ith sequence
+    // llama_get_embeddings(ctx) + i*n_embd
+    LLAMA_API float * llama_get_embeddings_ith(struct llama_context * ctx, int32_t i);
+
     //
     // Vocab
     //

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());
-        else funcs.vec_dot(kVecSize * QK4_1, &fs, x41.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, 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());
-            else funcs.vec_dot(kVecSize, &result, q40.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, 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
+5070f078a67c18c11736e78316ab715ca9afde16

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

spm-headers/ggml-alloc.h

@@ -0,0 +1 @@
+../ggml-alloc.h

spm-headers/ggml-backend.h

@@ -0,0 +1 @@
+../ggml-backend.h

spm-headers/ggml.h

@@ -0,0 +1 @@
+../ggml.h

tests/test-backend-ops.cpp

@@ -2129,14 +2129,13 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_pad());
     test_cases.emplace_back(new test_leaky_relu());
 
+    // these tests are disabled to save execution time, but they can be handy for debugging
+#if 0
 #if !defined(__SANITIZE_THREAD__)
     // FIXME: these tests use too much memory with thread sanitizer
     test_cases.emplace_back(new test_moe(8, 2, 1, 4096, 8*1024));
     //test_cases.emplace_back(new test_moe(8, 2, 8, 4096, 14336));
 #endif
-
-    // these tests are disabled to save execution time, but they can be handy for debugging
-#if 0
     test_cases.emplace_back(new test_llama(1));
     test_cases.emplace_back(new test_llama(2));
     test_cases.emplace_back(new test_falcon(1));

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);

tests/test-tokenizer-1-bpe.cpp

@@ -4,13 +4,13 @@
 #include "console.h"
 
 #include <cassert>
+#include <codecvt>
 #include <cstdio>
 #include <cstring>
-#include <string>
-#include <codecvt>
-#include <map>
-#include <vector>
 #include <locale>
+#include <string>
+#include <thread>
+#include <vector>
 
 int main(int argc, char **argv) {
     if (argc < 2) {
@@ -74,45 +74,46 @@ int main(int argc, char **argv) {
             }
         }
         catch (const std::invalid_argument &) {
-            fprintf(stderr, "%s : info: utf8 conversion %d '%s'\n", __func__, i, str.c_str());
+            //fprintf(stderr, "%s : info: utf8 conversion %d '%s'\n", __func__, i, str.c_str());
         }
     }
 
-    for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
-        // NOTE: these exceptions seem to be necessary, because the GPT2 tokenizer doesn't want to interfere with some ASCII control characters
-        if ((cp < 0x03 || cp > 0x05) && cp != 0x0b && cp != 0x11 && (cp < 0x13 || cp > 0x17) && cp != 0x19 && (cp < 0x1c || cp > 0x1e) && (cp < 0xd800 || cp > 0xdfff)) {
-            std::string str = " " + codepoint_to_utf8(cp);
-            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-            std::string check = llama_detokenize_bpe(ctx, tokens);
-            if (str != check) {
-                fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                    __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-                return 3;
-            }
-        }
-    }
-    // Restrict to assigned unicode planes
-    // for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
-    for (uint32_t cp = 0x10000; cp < 0x00040000; ++cp) {
-        std::string str = codepoint_to_utf8(cp);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-        std::string check = llama_detokenize_bpe(ctx, tokens);
-        if (str != check) {
-            fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-            return 4;
-        }
-    }
-    for (uint32_t cp = 0x000e0000; cp < 0x0010ffff; ++cp) {
-        std::string str = codepoint_to_utf8(cp);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-        std::string check = llama_detokenize_bpe(ctx, tokens);
-        if (str != check) {
-            fprintf(stderr, "%s : error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-            return 4;
+    // unicode
+    {
+        const int nthread = std::thread::hardware_concurrency();
+
+        std::vector<std::thread> threads(nthread);
+
+        for (int i = 0; i < nthread; ++i) {
+            threads[i] = std::thread([i, nthread, ctx]() {
+                for (uint32_t cp = i; cp < 0x0010ffff; cp += nthread) {
+                    if (!( // NOLINT
+                                (cp < 0x03       || cp >  0x05)   && cp != 0x0b && cp != 0x11 &&
+                                (cp < 0x13       || cp >  0x17)   && cp != 0x19 &&
+                                (cp < 0x1c       || cp >  0x1e)   &&
+                                (cp < 0xd800     || cp >  0xdfff) &&
+                                (cp < 0x00040000 || cp >= 0x000e0000)
+                        )) {
+                        continue;
+                    }
+
+                    std::string str = codepoint_to_utf8(cp);
+                    std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+                    std::string check = llama_detokenize_bpe(ctx, tokens);
+                    if (cp != 9601 && str != check) {
+                        fprintf(stderr, "error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                                cp, check.c_str(), check.length(), str.c_str(), str.length());
+                        std::exit(3);
+                    }
+                }
+            });
+        }
+
+        for (auto & t : threads) {
+            t.join();
         }
     }
+
     llama_free_model(model);
     llama_free(ctx);
 

tests/test-tokenizer-1-llama.cpp

@@ -4,13 +4,13 @@
 #include "console.h"
 
 #include <cassert>
+#include <codecvt>
 #include <cstdio>
 #include <cstring>
-#include <string>
-#include <codecvt>
-#include <map>
-#include <vector>
 #include <locale>
+#include <string>
+#include <thread>
+#include <vector>
 
 int main(int argc, char **argv) {
     if (argc < 2) {
@@ -72,26 +72,33 @@ int main(int argc, char **argv) {
         }
     }
 
-    for (uint32_t cp = 0x0000; cp < 0xffff; ++cp) {
-        if (cp < 0xd800 || cp > 0xdfff) {
-            std::string str = codepoint_to_utf8(cp);
-            std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-            std::string check = llama_detokenize_spm(ctx, tokens);
-            if (cp != 9601 && str != check) {
-                fprintf(stderr, "%s : error: codepoint %d detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                    __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-                return 3;
-            }
+    // unicode
+    {
+        const int nthread = std::thread::hardware_concurrency();
+
+        std::vector<std::thread> threads(nthread);
+
+        for (int i = 0; i < nthread; ++i) {
+            threads[i] = std::thread([i, nthread, ctx]() {
+                for (uint32_t cp = i; cp < 0x0010ffff; cp += nthread) {
+                    if (cp >= 0xd800 && cp <= 0xdfff) {
+                        continue;
+                    }
+
+                    std::string str = codepoint_to_utf8(cp);
+                    std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
+                    std::string check = llama_detokenize_spm(ctx, tokens);
+                    if (cp != 9601 && str != check) {
+                        fprintf(stderr, "error: codepoint %x detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
+                                cp, check.c_str(), check.length(), str.c_str(), str.length());
+                        std::exit(3);
+                    }
+                }
+            });
         }
-    }
-    for (uint32_t cp = 0x10000; cp < 0x0010ffff; ++cp) {
-        std::string str = codepoint_to_utf8(cp);
-        std::vector<llama_token> tokens = llama_tokenize(ctx, str, false);
-        std::string check = llama_detokenize_spm(ctx, tokens);
-        if (str != check) {
-            fprintf(stderr, "%s : error: codepoint %d detokenizes to '%s'(%zu) instead of '%s'(%zu)\n",
-                __func__, cp, check.c_str(), check.length(), str.c_str(), str.length());
-            return 4;
+
+        for (auto & t : threads) {
+            t.join();
         }
     }
 

unicode.h

@@ -264,26 +264,29 @@ static uint32_t codepoint_from_utf8(const std::string & utf8, size_t & offset) {
         offset += 1;
         return result;
     }
-    else if (!(utf8[offset + 0] & 0x40)) {
+    if (!(utf8[offset + 0] & 0x40)) {
         throw std::invalid_argument("invalid character");
     }
-    else if (!(utf8[offset + 0] & 0x20)) {
-        if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80))
+    if (!(utf8[offset + 0] & 0x20)) {
+        if (offset + 1 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80)) {
             throw std::invalid_argument("invalid character");
+        }
         auto result = ((utf8[offset + 0] & 0x1f) << 6) | (utf8[offset + 1] & 0x3f);
         offset += 2;
         return result;
     }
-    else if (!(utf8[offset + 0] & 0x10)) {
-        if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80))
+    if (!(utf8[offset + 0] & 0x10)) {
+        if (offset + 2 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80)) {
             throw std::invalid_argument("invalid character");
+        }
         auto result = ((utf8[offset + 0] & 0x0f) << 12) | ((utf8[offset + 1] & 0x3f) << 6) | (utf8[offset + 2] & 0x3f);
         offset += 3;
         return result;
     }
-    else if (!(utf8[offset + 0] & 0x08)) {
-        if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80))
+    if (!(utf8[offset + 0] & 0x08)) {
+        if (offset + 3 >= utf8.size() || ! ((utf8[offset + 1] & 0xc0) == 0x80) || ! ((utf8[offset + 2] & 0xc0) == 0x80) || !((utf8[offset + 3] & 0xc0) == 0x80)) {
             throw std::invalid_argument("invalid character");
+        }
         auto result = ((utf8[offset + 0] & 0x07) << 18) | ((utf8[offset + 1] & 0x3f) << 12) | ((utf8[offset + 2] & 0x3f) << 6) | (utf8[offset + 3] & 0x3f);
         offset += 4;
         return result;
@@ -331,21 +334,22 @@ static uint32_t codepoint_from_utf16(const std::vector<uint16_t> & utf16, size_t
         offset += 1;
         return result;
     }
-    else {
-        if (offset + 1 >= utf16.size() || !((utf16[1] & 0xdc00) == 0xdc00))
-            throw std::invalid_argument("invalid character");
-        auto result = 0x10000 + (((utf16[0] & 0x03ff) << 10) | (utf16[1] & 0x03ff));
-        offset += 2;
-        return result;
+
+    if (offset + 1 >= utf16.size() || !((utf16[1] & 0xdc00) == 0xdc00)) {
+        throw std::invalid_argument("invalid character");
     }
-    throw std::invalid_argument("invalid string");
+
+    auto result = 0x10000 + (((utf16[0] & 0x03ff) << 10) | (utf16[1] & 0x03ff));
+    offset += 2;
+    return result;
 }
 
 static std::vector<uint32_t> codepoints_from_utf16(const std::vector<uint16_t> & utf16) {
     std::vector<uint32_t> result;
     size_t offset = 0;
-    while (offset < utf16.size())
+    while (offset < utf16.size()) {
         result.push_back(codepoint_from_utf16(utf16, offset));
+    }
     return result;
 }
 
@@ -361,44 +365,52 @@ static std::vector<uint32_t> codepoints_from_utf16(const std::vector<uint16_t> &
 static std::unordered_map<uint32_t, int> codepoint_type_map() {
     std::unordered_map<uint32_t, int> codepoint_types;
     for (auto p : digit_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
+        for (auto i = p.first; i <= p.second; ++ i) {
             codepoint_types[i] = CODEPOINT_TYPE_DIGIT;
+        }
     }
-    for(auto p : letter_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
+    for (auto p : letter_ranges) {
+        for (auto i = p.first; i <= p.second; ++ i) {
             codepoint_types[i] = CODEPOINT_TYPE_LETTER;
+        }
     }
-    for(auto p : whitespace_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
+    for (auto p : whitespace_ranges) {
+        for (auto i = p.first; i <= p.second; ++ i) {
             codepoint_types[i] = CODEPOINT_TYPE_WHITESPACE;
+        }
     }
-    for(auto p : accent_mark_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
+    for (auto p : accent_mark_ranges) {
+        for (auto i = p.first; i <= p.second; ++ i) {
             codepoint_types[i] = CODEPOINT_TYPE_ACCENT_MARK;
+        }
     }
-    for(auto p : punctuation_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
+    for (auto p : punctuation_ranges) {
+        for (auto i = p.first; i <= p.second; ++ i) {
             codepoint_types[i] = CODEPOINT_TYPE_PUNCTUATION;
+        }
     }
-    for (auto p : symbol_ranges) {
-        for (auto i = p.first; i <= p.second; ++i)
+    for  (auto p : symbol_ranges) {
+        for (auto i = p.first; i <= p.second; ++i) {
             codepoint_types[i] = CODEPOINT_TYPE_SYMBOL;
+        }
     }
-    for(auto p : control_ranges) {
-        for(auto i = p.first; i <= p.second; ++ i)
+    for (auto p : control_ranges) {
+        for (auto i = p.first; i <= p.second; ++ i) {
             codepoint_types[i] = CODEPOINT_TYPE_CONTROL;
+        }
     }
     return codepoint_types;
 }
 
 static int codepoint_type(uint32_t cp) {
     static std::unordered_map<uint32_t, int> codepoint_types = codepoint_type_map();
-    return codepoint_types[cp];
+    return codepoint_types.find(cp) == codepoint_types.end() ? CODEPOINT_TYPE_UNIDENTIFIED : codepoint_types.at(cp);
 }
 
 static int codepoint_type(const std::string & utf8) {
-    if (utf8.length() == 0)
+    if (utf8.length() == 0) {
         return CODEPOINT_TYPE_UNIDENTIFIED;
+    }
     size_t offset = 0;
     return codepoint_type(codepoint_from_utf8(utf8, offset));
 }