Merge branch 'master' into move-convert-py

.github/labeler.yml

@@ -62,6 +62,8 @@ server:
 ggml:
     - changed-files:
         - any-glob-to-any-file:
+            - ggml.c
+            - ggml.h
             - ggml-*.c
             - ggml-*.h
             - ggml-cuda/**
@@ -71,3 +73,6 @@ nix:
             - "**/*.nix"
             - .github/workflows/nix-*.yml
             - .devops/nix/nixpkgs-instances.nix
+embedding:
+    - changed-files:
+        - any-glob-to-any-file: examples/embedding/

.github/workflows/zig-build.yml

@@ -1,29 +0,0 @@
-name: Zig CI
-
-on:
-  pull_request:
-  push:
-    branches:
-      - master
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  build:
-    strategy:
-      fail-fast: false
-      matrix:
-        runs-on: [ubuntu-latest, macos-latest, windows-latest]
-    runs-on: ${{ matrix.runs-on }}
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          submodules: recursive
-          fetch-depth: 0
-      - uses: goto-bus-stop/setup-zig@v2
-        with:
-          version: 0.11.0
-      - name: Build Summary
-        run: zig build --summary all -freference-trace

CMakeLists.txt

@@ -124,7 +124,6 @@ set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
 set(LLAMA_METAL_STD "" CACHE STRING          "llama: metal standard version (-std flag)")
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
 option(LLAMA_RPC                             "llama: use RPC"                                   OFF)
-option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
 set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
@@ -384,10 +383,6 @@ if (LLAMA_LLAMAFILE)
     set(GGML_SOURCES_LLAMAFILE sgemm.cpp)
 endif()
 
-if (LLAMA_QKK_64)
-    add_compile_definitions(GGML_QKK_64)
-endif()
-
 if (LLAMA_CUBLAS)
     message(WARNING "LLAMA_CUBLAS is deprecated and will be removed in the future.\nUse LLAMA_CUDA instead")
     set(LLAMA_CUDA ON)
@@ -505,6 +500,12 @@ if (LLAMA_VULKAN)
 
         add_compile_definitions(GGML_USE_VULKAN)
 
+        # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
+        # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
+        if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+            add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
+        endif()
+
         if (LLAMA_VULKAN_CHECK_RESULTS)
             add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
         endif()

Makefile

@@ -389,10 +389,6 @@ else
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
 
-ifdef LLAMA_QKK_64
-	MK_CPPFLAGS += -DGGML_QKK_64
-endif
-
 ifndef LLAMA_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
 	# `-framework Accelerate` works both with Apple Silicon and Mac Intel

README.md

@@ -127,6 +127,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))
 
@@ -140,6 +141,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
 - [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
 - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
+- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 
 **HTTP server**
 

build.zig

@@ -1,172 +0,0 @@
-// Compatible with Zig Version 0.11.0
-const std = @import("std");
-const ArrayList = std.ArrayList;
-const Compile = std.Build.Step.Compile;
-const ConfigHeader = std.Build.Step.ConfigHeader;
-const Mode = std.builtin.Mode;
-const CrossTarget = std.zig.CrossTarget;
-
-const Maker = struct {
-    builder: *std.build.Builder,
-    target: CrossTarget,
-    optimize: Mode,
-    enable_lto: bool,
-
-    include_dirs: ArrayList([]const u8),
-    cflags: ArrayList([]const u8),
-    cxxflags: ArrayList([]const u8),
-    objs: ArrayList(*Compile),
-
-    fn addInclude(m: *Maker, dir: []const u8) !void {
-        try m.include_dirs.append(dir);
-    }
-    fn addProjectInclude(m: *Maker, path: []const []const u8) !void {
-        try m.addInclude(try m.builder.build_root.join(m.builder.allocator, path));
-    }
-    fn addCFlag(m: *Maker, flag: []const u8) !void {
-        try m.cflags.append(flag);
-    }
-    fn addCxxFlag(m: *Maker, flag: []const u8) !void {
-        try m.cxxflags.append(flag);
-    }
-    fn addFlag(m: *Maker, flag: []const u8) !void {
-        try m.addCFlag(flag);
-        try m.addCxxFlag(flag);
-    }
-
-    fn init(builder: *std.build.Builder) !Maker {
-        const target = builder.standardTargetOptions(.{});
-        const zig_version = @import("builtin").zig_version_string;
-        const commit_hash = try std.ChildProcess.exec(
-            .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
-        );
-        try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt(
-            \\int LLAMA_BUILD_NUMBER = {};
-            \\char const *LLAMA_COMMIT = "{s}";
-            \\char const *LLAMA_COMPILER = "Zig {s}";
-            \\char const *LLAMA_BUILD_TARGET = "{s}";
-            \\
-        , .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) }));
-        var m = Maker{
-            .builder = builder,
-            .target = target,
-            .optimize = builder.standardOptimizeOption(.{}),
-            .enable_lto = false,
-            .include_dirs = ArrayList([]const u8).init(builder.allocator),
-            .cflags = ArrayList([]const u8).init(builder.allocator),
-            .cxxflags = ArrayList([]const u8).init(builder.allocator),
-            .objs = ArrayList(*Compile).init(builder.allocator),
-        };
-
-        try m.addCFlag("-std=c11");
-        try m.addCxxFlag("-std=c++11");
-        try m.addProjectInclude(&.{});
-        try m.addProjectInclude(&.{"common"});
-        return m;
-    }
-
-    fn obj(m: *const Maker, name: []const u8, src: []const u8) *Compile {
-        const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
-        if (o.target.getAbi() != .msvc)
-            o.defineCMacro("_GNU_SOURCE", null);
-
-        if (std.mem.endsWith(u8, src, ".c")) {
-            o.addCSourceFiles(&.{src}, m.cflags.items);
-            o.linkLibC();
-        } else {
-            o.addCSourceFiles(&.{src}, m.cxxflags.items);
-            if (o.target.getAbi() == .msvc) {
-                o.linkLibC(); // need winsdk + crt
-            } else {
-                // linkLibCpp already add (libc++ + libunwind + libc)
-                o.linkLibCpp();
-            }
-        }
-        for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
-        o.want_lto = m.enable_lto;
-        return o;
-    }
-
-    fn exe(m: *const Maker, name: []const u8, src: []const u8, deps: []const *Compile) *Compile {
-        const e = m.builder.addExecutable(.{ .name = name, .target = m.target, .optimize = m.optimize });
-        e.addCSourceFiles(&.{src}, m.cxxflags.items);
-        for (deps) |d| e.addObject(d);
-        for (m.objs.items) |o| e.addObject(o);
-        for (m.include_dirs.items) |i| e.addIncludePath(.{ .path = i });
-
-        // https://github.com/ziglang/zig/issues/15448
-        if (e.target.getAbi() == .msvc) {
-            e.linkLibC(); // need winsdk + crt
-        } else {
-            // linkLibCpp already add (libc++ + libunwind + libc)
-            e.linkLibCpp();
-        }
-        m.builder.installArtifact(e);
-        e.want_lto = m.enable_lto;
-        return e;
-    }
-};
-
-pub fn build(b: *std.build.Builder) !void {
-    var make = try Maker.init(b);
-    make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
-
-    const ggml = make.obj("ggml", "ggml.c");
-    const sgemm = make.obj("sgemm", "sgemm.cpp");
-    const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
-    const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
-    const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
-    const unicode = make.obj("unicode", "unicode.cpp");
-    const unicode_data = make.obj("unicode-data", "unicode-data.cpp");
-    const llama = make.obj("llama", "llama.cpp");
-    const buildinfo = make.obj("common", "common/build-info.cpp");
-    const common = make.obj("common", "common/common.cpp");
-    const console = make.obj("console", "common/console.cpp");
-    const sampling = make.obj("sampling", "common/sampling.cpp");
-    const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
-    const json_schema_to_grammar = make.obj("json-schema-to-grammar", "common/json-schema-to-grammar.cpp");
-    const train = make.obj("train", "common/train.cpp");
-    const clip = make.obj("clip", "examples/llava/clip.cpp");
-    const llava = make.obj("llava", "examples/llava/llava.cpp");
-
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
-
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, grammar_parser, clip, llava });
-    if (server.target.isWindows()) {
-        server.linkSystemLibrary("ws2_32");
-    }
-
-    const server_assets = [_][]const u8{ "index.html", "index.js", "completion.js", "json-schema-to-grammar.mjs" };
-    for (server_assets) |asset| {
-        const input_path = b.fmt("examples/server/public/{s}", .{asset});
-        const output_path = b.fmt("examples/server/{s}.hpp", .{asset});
-
-        // Portable equivalent of `b.addSystemCommand(&.{ "xxd", "-n", asset, "-i", input_path, output_path }) })`:
-
-        const input = try std.fs.cwd().readFileAlloc(b.allocator, input_path, std.math.maxInt(usize));
-        defer b.allocator.free(input);
-
-        var buf = std.ArrayList(u8).init(b.allocator);
-        defer buf.deinit();
-
-        for (input) |byte| {
-            try std.fmt.format(buf.writer(), "0x{X:0>2}, ", .{byte});
-        }
-
-        var name = try std.mem.replaceOwned(u8, b.allocator, asset, "-", "_");
-        defer b.allocator.free(name);
-        std.mem.replaceScalar(u8, name, '.', '_');
-
-        try std.fs.cwd().writeFile(output_path, b.fmt(
-            "unsigned char {s}[] = {{{s}}};\nunsigned int {s}_len = {d};\n",
-            .{ name, buf.items, name, input.len },
-        ));
-
-        std.debug.print("Dumped hex of \"{s}\" ({s}) to {s}\n", .{ input_path, name, output_path });
-    }
-}

ci/run.sh

@@ -202,12 +202,15 @@ function gg_sum_test_scripts_release {
 }
 
 function gg_get_model {
-    local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
+    local gguf_0="$MNT/models/pythia/1.4B/ggml-model-f16.gguf"
-    local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
+    local gguf_1="$MNT/models/pythia/2.8B/ggml-model-f16.gguf"
-    if [[ -s $gguf_3b ]]; then
+    local gguf_2="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
-        echo -n "$gguf_3b"
+    if [[ -s $gguf_0 ]]; then
-    elif [[ -s $gguf_7b ]]; then
+        echo -n "$gguf_0"
-        echo -n "$gguf_7b"
+    elif [[ -s $gguf_1 ]]; then
+        echo -n "$gguf_1"
+    elif [[ -s $gguf_2 ]]; then
+        echo -n "$gguf_2"
     else
         echo >&2 "No model found. Can't run gg_run_ctest_with_model."
         exit 1
@@ -256,139 +259,6 @@ function gg_sum_ctest_with_model_release {
     gg_printf '```\n'
 }
 
-# open_llama_3b_v2
-
-function gg_run_open_llama_3b_v2 {
-    cd ${SRC}
-
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/config.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/tokenizer.model
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/tokenizer_config.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/special_tokens_map.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
-
-    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
-    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
-
-    path_models="../models-mnt/open-llama/3B-v2"
-    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
-
-    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
-
-    set -e
-
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                                             ) 2>&1 | tee -a $OUT/${ci}-make.log
-
-    python3 ../examples/convert-legacy-llama.py ${path_models}
-
-    model_f16="${path_models}/ggml-model-f16.gguf"
-    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
-    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
-    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
-    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
-    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
-    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
-    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
-    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
-    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
-    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
-
-    wiki_test_60="${path_wiki}/wiki.test-60.raw"
-
-    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
-
-    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
-    (time ./bin/save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-
-    function check_ppl {
-        qnt="$1"
-        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
-            return 20
-        fi
-
-        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
-        return 0
-    }
-
-    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-
-    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
-
-    set +e
-}
-
-function gg_sum_open_llama_3b_v2 {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'OpenLLaMA 3B-v2:\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
-    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    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)"
-    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
-    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
-    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
-    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
-    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
-    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
-    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
-    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
-    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
-    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-}
-
 # open_llama_7b_v2
 # requires: GG_BUILD_CUDA
 
@@ -417,7 +287,7 @@ function gg_run_open_llama_7b_v2 {
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
     (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    python3 ../examples/convert-legacy-llama.py ${path_models}
+    python3 ../convert.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -526,6 +396,272 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
 }
 
+# pythia_1.4b
+
+function gg_run_pythia_1_4b {
+    cd ${SRC}
+
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/config.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/special_tokens_map.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/resolve/main/pytorch_model.bin
+
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
+
+    path_models="../models-mnt/pythia/1.4B"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    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} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
+
+    wiki_test_60="${path_wiki}/wiki.test-60.raw"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
+    set +e
+}
+
+function gg_sum_pythia_1_4b {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Pythia 1.4B:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    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)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
+}
+
+# pythia_2_8b
+# requires: GG_BUILD_CUDA
+
+function gg_run_pythia_2_8b {
+    cd ${SRC}
+
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/config.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/special_tokens_map.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/resolve/main/pytorch_model.bin
+
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+
+    path_models="../models-mnt/pythia/2.8B"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 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} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
+
+    wiki_test="${path_wiki}/wiki.test.raw"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/main --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
+    set +e
+}
+
+function gg_sum_pythia_2_8b {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Pythia 2.8B:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    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)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
+}
+
 # bge-small
 
 function gg_run_embd_bge_small {
@@ -552,7 +688,7 @@ function gg_run_embd_bge_small {
     (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}
+    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
     model_f16="${path_models}/ggml-model-f16.gguf"
     model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -606,9 +742,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
 
     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
+            test $ret -eq 0 && gg_run pythia_1_4b
         else
-            test $ret -eq 0 && gg_run open_llama_7b_v2
+            test $ret -eq 0 && gg_run pythia_2_8b
+            #test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
         test $ret -eq 0 && gg_run ctest_with_model_debug
         test $ret -eq 0 && gg_run ctest_with_model_release

common/common.h

@@ -41,8 +41,12 @@ extern char const *LLAMA_BUILD_TARGET;
 
 struct llama_control_vector_load_info;
 
-int get_math_cpu_count();
+//
-int32_t get_num_physical_cores();
+// CPU utils
+//
+
+int32_t cpu_get_num_physical_cores();
+int32_t cpu_get_num_math();
 
 //
 // CLI argument parsing
@@ -51,7 +55,7 @@ int32_t get_num_physical_cores();
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
-    int32_t n_threads             = get_math_cpu_count();
+    int32_t n_threads             = cpu_get_num_math();
     int32_t n_threads_draft       = -1;
     int32_t n_threads_batch       = -1;    // number of threads to use for batch processing (-1 = use n_threads)
     int32_t n_threads_batch_draft = -1;
@@ -179,33 +183,34 @@ struct gpt_params {
 
 void gpt_params_handle_model_default(gpt_params & params);
 
-bool parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
-
 bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);
-
 bool gpt_params_parse      (int argc, char ** argv, gpt_params & params);
-
-void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
-
 bool gpt_params_find_arg   (int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
+void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);
 
-std::string get_system_info(const gpt_params & params);
+std::string gpt_params_get_system_info(const gpt_params & params);
-
-std::string gpt_random_prompt(std::mt19937 & rng);
-
-void process_escapes(std::string& input);
-
-bool validate_file_name(const std::string & filename);
 
 //
 // String utils
 //
 
-std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_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 string_strip(const std::string & str);
-std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
+std::string string_get_sortable_timestamp();
+std::string string_random_prompt(std::mt19937 & rng);
+
+bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
+void string_process_escapes(std::string & input);
+
+//
+// Filesystem utils
+//
+
+bool fs_validate_filename(const std::string & filename);
+bool fs_create_directory_with_parents(const std::string & path);
+
+std::string fs_get_cache_directory();
 
 //
 // Model utils
@@ -276,29 +281,15 @@ std::string llama_detokenize_bpe(
 // defaults to true when model type is SPM, otherwise false.
 bool llama_should_add_bos_token(const llama_model * model);
 
-//
-// YAML utils
-//
-
-bool create_directory_with_parents(const std::string & path);
-void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data);
-void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data);
-void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data);
-std::string get_sortable_timestamp();
-
-void dump_non_result_info_yaml(
-    FILE * stream, const gpt_params & params, const llama_context * lctx,
-    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
-
 //
 // KV cache utils
 //
 
 // Dump the KV cache view with the number of sequences per cell.
-void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
+void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
 
 // Dump the KV cache view showing individual sequences in each cell (long output).
-void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
+void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
 
 //
 // Embedding utils
@@ -332,6 +323,20 @@ llama_control_vector_data llama_control_vector_load(const std::vector<llama_cont
 //
 // Split utils
 //
+
 static const char * const LLM_KV_SPLIT_NO            = "split.no";
 static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
+
+//
+// YAML utils
+//
+
+void yaml_dump_vector_float    (FILE * stream, const char * prop_name, const std::vector<float> & data);
+void yaml_dump_vector_int      (FILE * stream, const char * prop_name, const std::vector<int> & data);
+void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
+
+void yaml_dump_non_result_info(
+    FILE * stream, const gpt_params & params, const llama_context * lctx,
+    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
+

common/sampling.cpp

@@ -125,7 +125,7 @@ std::string llama_sampling_order_print(const llama_sampling_params & params) {
     std::string result = "CFG -> Penalties ";
     if (params.mirostat == 0) {
         for (auto sampler_type : params.samplers_sequence) {
-            const auto sampler_type_name = sampler_type_to_name_string(sampler_type);
+            const auto sampler_type_name = llama_sampling_type_to_str(sampler_type);
             if (!sampler_type_name.empty()) {
                 result += "-> " + sampler_type_name + " ";
             }
@@ -137,6 +137,87 @@ std::string llama_sampling_order_print(const llama_sampling_params & params) {
     return result;
 }
 
+std::string llama_sampling_type_to_str(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::TEMPERATURE: return "temperature";
+        default : return "";
+    }
+}
+
+std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
+    std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
+        {"top_k",       llama_sampler_type::TOP_K},
+        {"top_p",       llama_sampler_type::TOP_P},
+        {"typical_p",   llama_sampler_type::TYPICAL_P},
+        {"min_p",       llama_sampler_type::MIN_P},
+        {"tfs_z",       llama_sampler_type::TFS_Z},
+        {"temperature", llama_sampler_type::TEMPERATURE}
+    };
+
+    // since samplers names are written multiple ways
+    // make it ready for both system names and input names
+    std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
+        {"top-k",       llama_sampler_type::TOP_K},
+        {"top-p",       llama_sampler_type::TOP_P},
+        {"nucleus",     llama_sampler_type::TOP_P},
+        {"typical-p",   llama_sampler_type::TYPICAL_P},
+        {"typical",     llama_sampler_type::TYPICAL_P},
+        {"min-p",       llama_sampler_type::MIN_P},
+        {"tfs-z",       llama_sampler_type::TFS_Z},
+        {"tfs",         llama_sampler_type::TFS_Z},
+        {"temp",        llama_sampler_type::TEMPERATURE}
+    };
+
+    std::vector<llama_sampler_type> sampler_types;
+    sampler_types.reserve(names.size());
+    for (const auto & name : names)
+    {
+        auto sampler_item = sampler_canonical_name_map.find(name);
+        if (sampler_item != sampler_canonical_name_map.end())
+        {
+            sampler_types.push_back(sampler_item->second);
+        }
+        else
+        {
+            if (allow_alt_names)
+            {
+                sampler_item = sampler_alt_name_map.find(name);
+                if (sampler_item != sampler_alt_name_map.end())
+                {
+                    sampler_types.push_back(sampler_item->second);
+                }
+            }
+        }
+    }
+    return sampler_types;
+}
+
+std::vector<llama_sampler_type> llama_sampling_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::TEMPERATURE}
+    };
+
+    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;
+}
+
 // no reasons to expose this function in header
 static void sampler_queue(
                    struct llama_context * ctx_main,
@@ -179,7 +260,7 @@ static llama_token llama_sampling_sample_impl(
                   struct llama_context * ctx_main,
                   struct llama_context * ctx_cfg,
                   const int idx,
-                  bool is_resampling) {  // Add a parameter to indicate if we are resampling
+                  bool is_resampling) {
     const llama_sampling_params & params = ctx_sampling->params;
 
     const float   temp            = params.temp;
@@ -188,8 +269,8 @@ static llama_token llama_sampling_sample_impl(
     const float   mirostat_eta    = params.mirostat_eta;
 
     std::vector<float> original_logits;
-    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, !is_resampling, &original_logits);
+    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, /* apply_grammar= */ is_resampling, &original_logits);
-    if (!is_resampling) {
+    if (ctx_sampling->grammar != NULL && !is_resampling) {
         GGML_ASSERT(!original_logits.empty());
     }
     llama_token id = 0;
@@ -252,7 +333,7 @@ static llama_token llama_sampling_sample_impl(
             // Restore logits from the copy
             std::copy(original_logits.begin(), original_logits.end(), logits);
 
-            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, true);  // Pass true for is_resampling
+            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ true);
         }
     }
 
@@ -285,7 +366,8 @@ static llama_token_data_array llama_sampling_prepare_impl(
     // Get a pointer to the logits
     float * logits = llama_get_logits_ith(ctx_main, idx);
 
-    if (apply_grammar && original_logits != NULL) {
+    if (ctx_sampling->grammar != NULL && !apply_grammar) {
+        GGML_ASSERT(original_logits != NULL);
         // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
         *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
     }
@@ -342,7 +424,7 @@ llama_token llama_sampling_sample(
                   struct llama_context * ctx_cfg,
                   const int idx) {
     // Call the implementation function with is_resampling set to false by default
-    return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
+    return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ false);
 }
 
 llama_token_data_array llama_sampling_prepare(

common/sampling.h

@@ -116,6 +116,11 @@ std::string llama_sampling_print(const llama_sampling_params & params);
 // Print sampling order into a string
 std::string llama_sampling_order_print(const llama_sampling_params & params);
 
+std::string llama_sampling_type_to_str(llama_sampler_type sampler_type);
+
+std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string);
+
 // this is a common sampling function used across the examples for convenience
 // it can serve as a starting point for implementing your own sampling function
 // Note: When using multiple sequences, it is the caller's responsibility to call

common/train.cpp

@@ -1380,7 +1380,7 @@ bool consume_common_train_arg(
 
 void finish_processing_train_args(struct train_params_common * params) {
     if (params->escape) {
-        process_escapes(params->sample_start);
+        string_process_escapes(params->sample_start);
     }
 }
 

convert-hf-to-gguf-update.py

@@ -72,7 +72,7 @@ models = [
     {"name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
     {"name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
     {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
-    {"name": "stablelm",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
+    {"name": "stablelm2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
     {"name": "refact",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
     {"name": "command-r",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
     {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },

convert-hf-to-gguf.py

@@ -14,6 +14,7 @@ from pathlib import Path
 from hashlib import sha256
 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Sequence, TypeVar, cast
 
+import math
 import numpy as np
 import torch
 
@@ -447,7 +448,7 @@ class Model:
             # ref: https://huggingface.co/openai-community/gpt2
             res = "gpt-2"
         if chkhsh == "32d85c31273f8019248f2559fed492d929ea28b17e51d81d3bb36fff23ca72b3":
-            # ref: https://huggingface.co/stabilityai/stablelm-2-1_6b
+            # ref: https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b
             res = "stablelm2"
         if chkhsh == "6221ad2852e85ce96f791f476e0b390cf9b474c9e3d1362f53a24a06dc8220ff":
             # ref: https://huggingface.co/smallcloudai/Refact-1_6-base
@@ -672,6 +673,44 @@ class GPTNeoXModel(Model):
         self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+
+        tensors: list[tuple[str, Tensor]] = []
+
+        if re.match(r"gpt_neox\.layers\.\d+\.attention\.query_key_value\.weight", name):
+            # Map bloom-style qkv_linear to gpt-style qkv_linear
+            # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
+            # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
+            qkv_weights = data_torch.reshape((n_head, 3, n_embed // n_head, n_embed))
+            data_torch = torch.cat(
+                (
+                    qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
+                    qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
+                    qkv_weights[:, 2, :, :].reshape((-1, n_embed)),
+                ),
+                dim=0,
+            )
+            logger.info("re-format attention.linear_qkv.weight")
+        elif re.match(r"gpt_neox\.layers\.\d+\.attention\.query_key_value\.bias", name):
+            qkv_bias = data_torch.reshape((n_head, 3, n_embed // n_head))
+            data_torch = torch.cat(
+                (
+                    qkv_bias[:, 0, :].reshape((n_embed,)),
+                    qkv_bias[:, 1, :].reshape((n_embed,)),
+                    qkv_bias[:, 2, :].reshape((n_embed,)),
+                ),
+                dim=0,
+            )
+            logger.info("re-format attention.linear_qkv.bias")
+
+        tensors.append((self.map_tensor_name(name), data_torch))
+
+        return tensors
+
 
 @Model.register("BloomForCausalLM")
 class BloomModel(Model):
@@ -1749,7 +1788,7 @@ class Phi3MiniModel(Model):
                     token_id = int(token_id)
                     token = foken_data["content"].encode("utf-8")
                     if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
-                        assert(tokens[token_id] == token)
+                        assert tokens[token_id] == token
                     tokens[token_id] = token
                     scores[token_id] = -1000.0
                     toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
@@ -1765,7 +1804,7 @@ class Phi3MiniModel(Model):
                     token_id = int(foken_data["id"])
                     token = foken_data["content"].encode("utf-8")
                     if toktypes[token_id] != SentencePieceTokenTypes.UNKNOWN:
-                        assert(tokens[token_id] == token)
+                        assert tokens[token_id] == token
                     tokens[token_id] = token
                     scores[token_id] = -1000.0
                     toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
@@ -1784,23 +1823,59 @@ class Phi3MiniModel(Model):
     def set_gguf_parameters(self):
         block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
 
-        rot_pct = 1.0
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        n_head_kv = self.find_hparam(["num_key_value_heads", "n_head_kv"])
         rms_eps = self.find_hparam(["rms_norm_eps"])
+        max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
+        orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
+        rope_dims = n_embd // n_head
 
         self.gguf_writer.add_name("Phi3")
-        self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
+        self.gguf_writer.add_context_length(max_pos_embds)
-
+        self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds)
         self.gguf_writer.add_embedding_length(n_embd)
-        self.gguf_writer.add_feed_forward_length(8192)
+        self.gguf_writer.add_feed_forward_length(self.find_hparam(["intermediate_size"]))
         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_head_count_kv(n_head_kv)
         self.gguf_writer.add_layer_norm_rms_eps(rms_eps)
-        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
+        self.gguf_writer.add_rope_dimension_count(rope_dims)
+        self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"]))
         self.gguf_writer.add_file_type(self.ftype)
 
+        # write rope scaling for long context (128k) model
+        rope_scaling = self.find_hparam(['rope_scaling'], True)
+        if (rope_scaling is None):
+            return
+
+        scale = max_pos_embds / orig_max_pos_embds
+
+        rope_scaling_type = rope_scaling.get('type', '').lower()
+        if len(rope_scaling_type) == 0:
+            raise KeyError('Missing the required key rope_scaling.type')
+
+        if rope_scaling_type == 'su':
+            attn_factor = math.sqrt(1 + math.log(scale) / math.log(orig_max_pos_embds)) if scale > 1.0 else 1.0
+        elif rope_scaling_type == 'yarn':
+            attn_factor = 0.1 * math.log(scale) + 1.0 if scale > 1.0 else 1.0
+        else:
+            raise NotImplementedError(f'The rope scaling type {rope_scaling_type} is not supported yet')
+
+        self.gguf_writer.add_rope_scaling_attn_factors(attn_factor)
+
+        long_factors = rope_scaling.get('long_factor', None)
+        short_factors = rope_scaling.get('short_factor', None)
+
+        if long_factors is None or short_factors is None:
+            raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
+
+        if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
+            raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
+
+        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG]  + ".weight", np.array(long_factors, dtype=np.float32))
+        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
+
 
 @Model.register("PlamoForCausalLM")
 class PlamoModel(Model):

examples/batched/batched.cpp

@@ -48,7 +48,7 @@ int main(int argc, char ** argv) {
         params.prompt = "Hello my name is";
     }
 
-    process_escapes(params.prompt);
+    string_process_escapes(params.prompt);
 
     // init LLM
 

examples/embedding/embedding.cpp

@@ -80,7 +80,7 @@ int main(int argc, char ** argv) {
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
     }
 
     llama_backend_init();
@@ -107,7 +107,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     // split the prompt into lines

examples/eval-callback/eval-callback.cpp

@@ -152,7 +152,7 @@ int main(int argc, char ** argv) {
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
     }
 
     llama_backend_init();
@@ -176,7 +176,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     bool OK = run(ctx, params);

examples/finetune/finetune.cpp

@@ -563,8 +563,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         // not capturing these, to silcence warnings
         const int rope_mode = 0;
 
-        return ggml_rope_custom(ctx,
+        return ggml_rope_ext(ctx,
-            t, KQ_pos, n_rot, rope_mode, n_ctx, 0,
+            t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0,
             rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
         );
     };
@@ -643,7 +643,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct ggml_tensor * t15 = ggml_permute      (ctx, t12, 0, 3, 1, 2);                         set_name(t15, "t15");     assert_shape_4d(t15, N, n_embd_head, n_head_kv, n_batch);
         struct ggml_tensor * t16;
         if (enable_flash_attn) {
-            t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                         set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
+            GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported");
+            //t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                         set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
         } else {
             struct ggml_tensor * t16_0 = ggml_mul_mat              (ctx, t14, t13);                  set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
             struct ggml_tensor * t16_1 = ggml_scale_inplace        (ctx, t16_0, kv_scale);           set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);

examples/imatrix/imatrix.cpp

@@ -598,7 +598,7 @@ int main(int argc, char ** argv) {
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
     }
 
     sparams.dataset = params.prompt_file;
@@ -667,7 +667,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     bool OK = compute_imatrix(ctx, params, compute_ppl, from_chunk);

examples/infill/infill.cpp

@@ -50,9 +50,9 @@ static void write_logfile(
         return;
     }
 
-    const std::string timestamp = get_sortable_timestamp();
+    const std::string timestamp = string_get_sortable_timestamp();
 
-    const bool success = create_directory_with_parents(params.logdir);
+    const bool success = fs_create_directory_with_parents(params.logdir);
     if (!success) {
         fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
                 __func__, params.logdir.c_str());
@@ -70,7 +70,7 @@ static void write_logfile(
     fprintf(logfile, "binary: infill\n");
     char model_desc[128];
     llama_model_desc(model, model_desc, sizeof(model_desc));
-    dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
 
     fprintf(logfile, "\n");
     fprintf(logfile, "######################\n");
@@ -78,8 +78,8 @@ static void write_logfile(
     fprintf(logfile, "######################\n");
     fprintf(logfile, "\n");
 
-    dump_string_yaml_multiline(logfile, "output", output.c_str());
+    yaml_dump_string_multiline(logfile, "output", output.c_str());
-    dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
+    yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
 
     llama_dump_timing_info_yaml(logfile, ctx);
     fclose(logfile);
@@ -236,7 +236,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         LOG_TEE("\n");
-        LOG_TEE("%s\n", get_system_info(params).c_str());
+        LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
     }
     const bool add_bos = llama_should_add_bos_token(model);
     GGML_ASSERT(llama_add_eos_token(model) != 1);
@@ -621,8 +621,8 @@ int main(int argc, char ** argv) {
 
                 if (params.escape) {
                     //process escape sequences, for the initial prompt this is done in common.cpp when we load the params, but for the interactive mode we need to do it here
-                    process_escapes(params.input_prefix);
+                    string_process_escapes(params.input_prefix);
-                    process_escapes(params.input_suffix);
+                    string_process_escapes(params.input_suffix);
                 }
                 suff_rm_leading_spc = params.escape;
                 if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {

examples/llama-bench/llama-bench.cpp

@@ -195,12 +195,12 @@ static const cmd_params cmd_params_defaults = {
     /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
-    /* n_pg          */ {{512, 128}},
+    /* n_pg          */ {},
     /* n_batch       */ {2048},
     /* n_ubatch      */ {512},
     /* type_k        */ {GGML_TYPE_F16},
     /* type_v        */ {GGML_TYPE_F16},
-    /* n_threads     */ {get_math_cpu_count()},
+    /* n_threads     */ {cpu_get_num_math()},
     /* n_gpu_layers  */ {99},
     /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
     /* main_gpu      */ {0},

examples/llava/llava-cli.cpp

@@ -290,7 +290,7 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
-        gpt_print_usage(argc, argv, params);
+        gpt_params_print_usage(argc, argv, params);
         show_additional_info(argc, argv);
         return 1;
     }

examples/lookahead/lookahead.cpp

@@ -174,7 +174,7 @@ int main(int argc, char ** argv) {
         // debug
         if (dump_kv_cache) {
             llama_kv_cache_view_update(ctx, &kvc_view);
-            dump_kv_cache_view_seqs(kvc_view, 40);
+            llama_kv_cache_dump_view_seqs(kvc_view, 40);
         }
 
         // build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/

examples/lookup/lookup.cpp

@@ -121,7 +121,7 @@ int main(int argc, char ** argv){
         // debug
         if (dump_kv_cache) {
             llama_kv_cache_view_update(ctx, &kvc_view);
-            dump_kv_cache_view_seqs(kvc_view, 40);
+            llama_kv_cache_dump_view_seqs(kvc_view, 40);
         }
 
         // print current draft sequence

examples/main/README.md

@@ -325,3 +325,5 @@ These options provide extra functionality and customization when running the LLa
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
+
+-   `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable  or in an OS-specific local cache.

examples/main/main.cpp

@@ -60,9 +60,9 @@ static void write_logfile(
         return;
     }
 
-    const std::string timestamp = get_sortable_timestamp();
+    const std::string timestamp = string_get_sortable_timestamp();
 
-    const bool success = create_directory_with_parents(params.logdir);
+    const bool success = fs_create_directory_with_parents(params.logdir);
     if (!success) {
         fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
                 __func__, params.logdir.c_str());
@@ -80,7 +80,7 @@ static void write_logfile(
     fprintf(logfile, "binary: main\n");
     char model_desc[128];
     llama_model_desc(model, model_desc, sizeof(model_desc));
-    dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);
 
     fprintf(logfile, "\n");
     fprintf(logfile, "######################\n");
@@ -88,8 +88,8 @@ static void write_logfile(
     fprintf(logfile, "######################\n");
     fprintf(logfile, "\n");
 
-    dump_string_yaml_multiline(logfile, "output", output.c_str());
+    yaml_dump_string_multiline(logfile, "output", output.c_str());
-    dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
+    yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
 
     llama_dump_timing_info_yaml(logfile, ctx);
     fclose(logfile);
@@ -181,7 +181,7 @@ int main(int argc, char ** argv) {
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
     }
 
     LOG("%s: llama backend init\n", __func__);
@@ -219,7 +219,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         LOG_TEE("\n");
-        LOG_TEE("%s\n", get_system_info(params).c_str());
+        LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     std::string path_session = params.path_prompt_cache;
@@ -476,10 +476,10 @@ int main(int argc, char ** argv) {
     if (params.interactive) {
         const char * control_message;
         if (params.multiline_input) {
-            control_message = " - To return control to LLaMa, end your input with '\\'.\n"
+            control_message = " - To return control to the AI, end your input with '\\'.\n"
                               " - To return control without starting a new line, end your input with '/'.\n";
         } else {
-            control_message = " - Press Return to return control to LLaMa.\n"
+            control_message = " - Press Return to return control to the AI.\n"
                               " - To return control without starting a new line, end your input with '/'.\n"
                               " - If you want to submit another line, end your input with '\\'.\n";
         }
@@ -707,7 +707,7 @@ int main(int argc, char ** argv) {
 
             const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            llama_sampling_accept(ctx_sampling, ctx, id, /* apply_grammar= */ true);
 
             LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
 
@@ -728,7 +728,7 @@ int main(int argc, char ** argv) {
 
                 // push the prompt in the sampling context in order to apply repetition penalties later
                 // for the prompt, we don't apply grammar rules
-                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
+                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], /* apply_grammar= */ false);
 
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
@@ -879,7 +879,7 @@ int main(int argc, char ** argv) {
                         embd_inp.insert(embd_inp.end(), cml_pfx.begin(), cml_pfx.end());
                     }
                     if (params.escape) {
-                        process_escapes(buffer);
+                        string_process_escapes(buffer);
                     }
 
                     const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);

examples/parallel/parallel.cpp

@@ -210,7 +210,7 @@ int main(int argc, char ** argv) {
     while (true) {
         if (dump_kv_cache) {
             llama_kv_cache_view_update(ctx, &kvc_view);
-            dump_kv_cache_view_seqs(kvc_view, 40);
+            llama_kv_cache_dump_view_seqs(kvc_view, 40);
         }
 
         llama_batch_clear(batch);

examples/perplexity/perplexity.cpp

@@ -44,9 +44,9 @@ static void write_logfile(
         return;
     }
 
-    const std::string timestamp = get_sortable_timestamp();
+    const std::string timestamp = string_get_sortable_timestamp();
 
-    const bool success = create_directory_with_parents(params.logdir);
+    const bool success = fs_create_directory_with_parents(params.logdir);
     if (!success) {
         fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
                 __func__, params.logdir.c_str());
@@ -64,7 +64,7 @@ static void write_logfile(
     fprintf(logfile, "binary: main\n");
     char model_desc[128];
     llama_model_desc(model, model_desc, sizeof(model_desc));
-    dump_non_result_info_yaml(logfile, params, ctx, timestamp, results.tokens, model_desc);
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, results.tokens, model_desc);
 
     fprintf(logfile, "\n");
     fprintf(logfile, "######################\n");
@@ -72,9 +72,9 @@ static void write_logfile(
     fprintf(logfile, "######################\n");
     fprintf(logfile, "\n");
 
-    dump_vector_float_yaml(logfile, "logits", results.logits);
+    yaml_dump_vector_float(logfile, "logits", results.logits);
     fprintf(logfile, "ppl_value: %f\n", results.ppl_value);
-    dump_vector_float_yaml(logfile, "probs", results.probs);
+    yaml_dump_vector_float(logfile, "probs", results.probs);
 
     llama_dump_timing_info_yaml(logfile, ctx);
     fclose(logfile);
@@ -2007,7 +2007,7 @@ int main(int argc, char ** argv) {
 
     std::mt19937 rng(params.seed);
     if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
     }
 
     llama_backend_init();
@@ -2035,7 +2035,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     struct results_perplexity results;

examples/quantize/quantize.cpp

@@ -259,7 +259,7 @@ int main(int argc, char ** argv) {
                 usage(argv[0]);
             }
         } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
-            if (arg_idx == argc-1 || !parse_kv_override(argv[++arg_idx], kv_overrides)) {
+            if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
                 usage(argv[0]);
             }
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {

examples/retrieval/retrieval.cpp

@@ -11,7 +11,7 @@ struct retrieval_params {
 };
 
 static void retrieval_params_print_usage(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & params) {
-    gpt_print_usage(argc, argv, gpt_params);
+    gpt_params_print_usage(argc, argv, gpt_params);
     printf("retrieval options:\n");
     printf("  --context-file FNAME  file containing context to embed.\n");
     printf("                        specify multiple files by providing --context-file option multiple times.\n");
@@ -226,7 +226,7 @@ int main(int argc, char ** argv) {
     // print system information
     {
         fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     // max batch size

examples/server/public_simplechat/index.html

@@ -0,0 +1,52 @@
+<!DOCTYPE html>
+<html lang="en">
+    <head>
+        <title>SimpleChat (LlamaCPP, ...) </title>
+        <meta charset="UTF-8" />
+        <meta name="viewport" content="width=device-width, initial-scale=1" />
+        <meta name="message" content="Save Nature Save Earth" />
+        <meta name="description" content="SimpleChat: trigger LLM web service endpoints /chat/completions and /completions, single/multi chat sessions" />
+        <meta name="author" content="by Humans for All" />
+        <meta http-equiv="Cache-Control" content="no-cache, no-store, must-revalidate" />
+        <script src="simplechat.js" defer></script>
+        <link rel="stylesheet" href="simplechat.css" />
+    </head>
+    <body>
+        <div class="samecolumn" id="fullbody">
+
+            <div class="sameline">
+                <p class="heading flex-grow" > <b> SimpleChat </b> </p>
+                <div class="sameline">
+                    <label for="api-ep">Mode:</label>
+                    <select name="api-ep" id="api-ep">
+                    <option value="chat" selected>Chat</option>
+                    <option value="completion">Completion</option>
+                    </select>
+                </div>
+            </div>
+
+            <div id="sessions-div" class="sameline"></div>
+
+            <hr>
+            <div class="sameline">
+                <label for="system-in">System</label>
+                <input type="text" name="system" id="system-in" class="flex-grow"/>
+            </div>
+
+            <hr>
+            <div id="chat-div">
+                <p> Enter the system prompt above, before entering/submitting any user query.</p>
+                <p> Enter your text to the ai assistant below.</p>
+                <p> Use shift+enter for inserting enter.</p>
+                <p> Refresh the page to start over fresh.</p>
+            </div>
+
+            <hr>
+            <div class="sameline">
+                <textarea id="user-in" class="flex-grow" rows="3"></textarea>
+                <button id="user-btn">submit</button>
+            </div>
+
+        </div>
+    </body>
+</html>

examples/server/public_simplechat/readme.md

@@ -0,0 +1,81 @@
+
+# SimpleChat
+
+by Humans for All.
+
+
+## overview
+
+This simple web frontend, allows triggering/testing the server's /completions or /chat/completions endpoints
+in a simple way with minimal code from a common code base. Inturn additionally it tries to allow single or
+multiple independent back and forth chatting to an extent, with the ai llm model at a basic level, with their
+own system prompts.
+
+The UI follows a responsive web design so that the layout can adapt to available display space in a usable
+enough manner, in general.
+
+NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and
+culling of old messages from the chat.
+
+NOTE: It doesnt set any parameters other than temperature for now. However if someone wants they can update
+the js file as needed.
+
+
+## usage
+
+One could run this web frontend directly using server itself or if anyone is thinking of adding a built in web
+frontend to configure the server over http(s) or so, then run this web frontend using something like python's
+http module.
+
+### running using examples/server
+
+bin/server -m path/model.gguf --path ../examples/server/public_simplechat [--port PORT]
+
+### running using python3's server module
+
+first run examples/server
+* bin/server -m path/model.gguf
+
+next run this web front end in examples/server/public_simplechat
+* cd ../examples/server/public_simplechat
+* python3 -m http.server PORT
+
+### using the front end
+
+Open this simple web front end from your local browser
+* http://127.0.0.1:PORT/index.html
+
+Once inside
+* Select between chat and completion mode. By default it is set to chat mode.
+* If you want to provide a system prompt, then ideally enter it first, before entering any user query.
+  * if chat.add_system_begin is used
+    * you cant change the system prompt, after it is has been submitted once along with user query.
+    * you cant set a system prompt, after you have submitted any user query
+  * if chat.add_system_anytime is used
+    * one can change the system prompt any time during chat, by changing the contents of system prompt.
+    * inturn the updated/changed system prompt will be inserted into the chat session.
+    * this allows for the subsequent user chatting to be driven by the new system prompt set above.
+* Enter your query and either press enter or click on the submit button.
+  If you want to insert enter (\n) as part of your chat/query to ai model, use shift+enter.
+* Wait for the logic to communicate with the server and get the response.
+  * the user is not allowed to enter any fresh query during this time.
+  * the user input box will be disabled and a working message will be shown in it.
+* just refresh the page, to reset wrt the chat history and or system prompt and start afresh.
+* Using NewChat one can start independent chat sessions.
+  * two independent chat sessions are setup by default.
+
+
+## Devel note
+
+Sometimes the browser may be stuborn with caching of the file, so your updates to html/css/js
+may not be visible. Also remember that just refreshing/reloading page in browser or for that
+matter clearing site data, dont directly override site caching in all cases. Worst case you may
+have to change port. Or in dev tools of browser, you may be able to disable caching fully.
+
+Concept of multiple chat sessions with different servers, as well as saving and restoring of
+those across browser usage sessions, can be woven around the SimpleChat/MultiChatUI class and
+its instances relatively easily, however given the current goal of keeping this simple, it has
+not been added, for now.
+
+By switching between chat.add_system_begin/anytime, one can control whether one can change
+the system prompt, anytime during the conversation or only at the beginning.

examples/server/public_simplechat/simplechat.css

@@ -0,0 +1,61 @@
+/**
+ * the styling of the simplechat web frontend
+ * by Humans for All
+ */
+
+#fullbody {
+    height: 98vh;
+}
+
+.heading {
+    background-color: lightgray;
+}
+
+.session-selected {
+    background-color: lightblue;
+}
+
+.role-system {
+    background-color: lightblue;
+}
+.role-user {
+    background-color: lightgray;
+}
+
+.flex-grow {
+    flex-grow: 1;
+}
+.float-right {
+    float: right;
+}
+
+#chat-div {
+    overflow: scroll;
+    flex-grow: 1;
+    flex-shrink: 1;
+    min-height: 40vh;
+}
+button {
+    min-width: 8vw;
+}
+
+.sameline {
+    display: flex;
+    flex-direction: row;
+}
+.samecolumn {
+    display: flex;
+    flex-direction: column;
+}
+
+* {
+    margin: 0.6vmin;
+}
+
+@media print {
+
+    #fullbody {
+        height: auto;
+    }
+
+}

examples/server/public_simplechat/simplechat.js

@@ -0,0 +1,478 @@
+// @ts-check
+// A simple completions and chat/completions test related web front end logic
+// by Humans for All
+
+class Roles {
+    static System = "system";
+    static User = "user";
+    static Assistant = "assistant";
+}
+
+class ApiEP {
+    static Chat = "chat";
+    static Completion = "completion";
+}
+
+let gUsageMsg = `
+    <p> Enter the system prompt above, before entering/submitting any user query.</p>
+    <p> Enter your text to the ai assistant below.</p>
+    <p> Use shift+enter for inserting enter.</p>
+    <p> Refresh the page to start over fresh.</p>
+`;
+
+class SimpleChat {
+
+    constructor() {
+        /**
+         * Maintain in a form suitable for common LLM web service chat/completions' messages entry
+         * @type {{role: string, content: string}[]}
+         */
+        this.xchat = [];
+        this.iLastSys = -1;
+    }
+
+    /**
+     * Add an entry into xchat
+     * @param {string} role
+     * @param {string|undefined|null} content
+     */
+    add(role, content) {
+        if ((content == undefined) || (content == null) || (content == "")) {
+            return false;
+        }
+        this.xchat.push( {role: role, content: content} );
+        if (role == Roles.System) {
+            this.iLastSys = this.xchat.length - 1;
+        }
+        return true;
+    }
+
+    /**
+     * Show the contents in the specified div
+     * @param {HTMLDivElement} div
+     * @param {boolean} bClear
+     */
+    show(div, bClear=true) {
+        if (bClear) {
+            div.replaceChildren();
+        }
+        let last = undefined;
+        for(const x of this.xchat) {
+            let entry = document.createElement("p");
+            entry.className = `role-${x.role}`;
+            entry.innerText = `${x.role}: ${x.content}`;
+            div.appendChild(entry);
+            last = entry;
+        }
+        if (last !== undefined) {
+            last.scrollIntoView(false);
+        } else {
+            if (bClear) {
+                div.innerHTML = gUsageMsg;
+            }
+        }
+    }
+
+    /**
+     * Add needed fields wrt json object to be sent wrt LLM web services completions endpoint
+     * Convert the json into string.
+     * @param {Object} obj
+     */
+    request_jsonstr(obj) {
+        obj["temperature"] = 0.7;
+        return JSON.stringify(obj);
+    }
+
+    /**
+     * Return a string form of json object suitable for chat/completions
+     */
+    request_messages_jsonstr() {
+        let req = {
+            messages: this.xchat,
+        }
+        return this.request_jsonstr(req);
+    }
+
+    /**
+     * Return a string form of json object suitable for /completions
+     */
+    request_prompt_jsonstr() {
+        let prompt = "";
+        for(const chat of this.xchat) {
+            prompt += `${chat.role}: ${chat.content}\n`;
+        }
+        let req = {
+            prompt: prompt,
+        }
+        return this.request_jsonstr(req);
+    }
+
+    /**
+     * Allow setting of system prompt, but only at begining.
+     * @param {string} sysPrompt
+     * @param {string} msgTag
+     */
+    add_system_begin(sysPrompt, msgTag) {
+        if (this.xchat.length == 0) {
+            if (sysPrompt.length > 0) {
+                return this.add(Roles.System, sysPrompt);
+            }
+        } else {
+            if (sysPrompt.length > 0) {
+                if (this.xchat[0].role !== Roles.System) {
+                    console.error(`ERRR:SimpleChat:SC:${msgTag}:You need to specify system prompt before any user query, ignoring...`);
+                } else {
+                    if (this.xchat[0].content !== sysPrompt) {
+                        console.error(`ERRR:SimpleChat:SC:${msgTag}:You cant change system prompt, mid way through, ignoring...`);
+                    }
+                }
+            }
+        }
+        return false;
+    }
+
+    /**
+     * Allow setting of system prompt, at any time.
+     * @param {string} sysPrompt
+     * @param {string} msgTag
+     */
+    add_system_anytime(sysPrompt, msgTag) {
+        if (sysPrompt.length <= 0) {
+            return false;
+        }
+
+        if (this.iLastSys < 0) {
+            return this.add(Roles.System, sysPrompt);
+        }
+
+        let lastSys = this.xchat[this.iLastSys].content;
+        if (lastSys !== sysPrompt) {
+            return this.add(Roles.System, sysPrompt);
+        }
+        return false;
+    }
+
+    /**
+     * Retrieve the latest system prompt.
+     */
+    get_system_latest() {
+        if (this.iLastSys == -1) {
+            return "";
+        }
+        let sysPrompt = this.xchat[this.iLastSys].content;
+        return sysPrompt;
+    }
+
+}
+
+
+let gBaseURL = "http://127.0.0.1:8080";
+let gChatURL = {
+    'chat': `${gBaseURL}/chat/completions`,
+    'completion': `${gBaseURL}/completions`,
+}
+const gbCompletionFreshChatAlways = true;
+
+
+/**
+ * Set the class of the children, based on whether it is the idSelected or not.
+ * @param {HTMLDivElement} elBase
+ * @param {string} idSelected
+ * @param {string} classSelected
+ * @param {string} classUnSelected
+ */
+function el_children_config_class(elBase, idSelected, classSelected, classUnSelected="") {
+    for(let child of elBase.children) {
+        if (child.id == idSelected) {
+            child.className = classSelected;
+        } else {
+            child.className = classUnSelected;
+        }
+    }
+}
+
+/**
+ * Create button and set it up.
+ * @param {string} id
+ * @param {(this: HTMLButtonElement, ev: MouseEvent) => any} callback
+ * @param {string | undefined} name
+ * @param {string | undefined} innerText
+ */
+function el_create_button(id, callback, name=undefined, innerText=undefined) {
+    if (!name) {
+        name = id;
+    }
+    if (!innerText) {
+        innerText = id;
+    }
+    let btn = document.createElement("button");
+    btn.id = id;
+    btn.name = name;
+    btn.innerText = innerText;
+    btn.addEventListener("click", callback);
+    return btn;
+}
+
+
+class MultiChatUI {
+
+    constructor() {
+        /** @type {Object<string, SimpleChat>} */
+        this.simpleChats = {};
+        /** @type {string} */
+        this.curChatId = "";
+
+        // the ui elements
+        this.elInSystem = /** @type{HTMLInputElement} */(document.getElementById("system-in"));
+        this.elDivChat = /** @type{HTMLDivElement} */(document.getElementById("chat-div"));
+        this.elBtnUser = /** @type{HTMLButtonElement} */(document.getElementById("user-btn"));
+        this.elInUser = /** @type{HTMLInputElement} */(document.getElementById("user-in"));
+        this.elSelectApiEP = /** @type{HTMLSelectElement} */(document.getElementById("api-ep"));
+        this.elDivSessions = /** @type{HTMLDivElement} */(document.getElementById("sessions-div"));
+
+        this.validate_element(this.elInSystem, "system-in");
+        this.validate_element(this.elDivChat, "chat-div");
+        this.validate_element(this.elInUser, "user-in");
+        this.validate_element(this.elSelectApiEP, "api-ep");
+        this.validate_element(this.elDivChat, "sessions-div");
+    }
+
+    /**
+     * Check if the element got
+     * @param {HTMLElement | null} el
+     * @param {string} msgTag
+     */
+    validate_element(el, msgTag) {
+        if (el == null) {
+            throw Error(`ERRR:SimpleChat:MCUI:${msgTag} element missing in html...`);
+        } else {
+            console.debug(`INFO:SimpleChat:MCUI:${msgTag} Id[${el.id}] Name[${el["name"]}]`);
+        }
+    }
+
+    /**
+     * Reset user input ui.
+     * * clear user input
+     * * enable user input
+     * * set focus to user input
+     */
+    ui_reset_userinput() {
+        this.elInUser.value = "";
+        this.elInUser.disabled = false;
+        this.elInUser.focus();
+    }
+
+    /**
+     * Setup the needed callbacks wrt UI, curChatId to defaultChatId and
+     * optionally switch to specified defaultChatId.
+     * @param {string} defaultChatId
+     * @param {boolean} bSwitchSession
+     */
+    setup_ui(defaultChatId, bSwitchSession=false) {
+
+        this.curChatId = defaultChatId;
+        if (bSwitchSession) {
+            this.handle_session_switch(this.curChatId);
+        }
+
+        this.elBtnUser.addEventListener("click", (ev)=>{
+            if (this.elInUser.disabled) {
+                return;
+            }
+            this.handle_user_submit(this.curChatId, this.elSelectApiEP.value).catch((/** @type{Error} */reason)=>{
+                let msg = `ERRR:SimpleChat\nMCUI:HandleUserSubmit:${this.curChatId}\n${reason.name}:${reason.message}`;
+                console.debug(msg.replace("\n", ":"));
+                alert(msg);
+                this.ui_reset_userinput();
+            });
+        });
+
+        this.elInUser.addEventListener("keyup", (ev)=> {
+            // allow user to insert enter into their message using shift+enter.
+            // while just pressing enter key will lead to submitting.
+            if ((ev.key === "Enter") && (!ev.shiftKey)) {
+                this.elBtnUser.click();
+                ev.preventDefault();
+            }
+        });
+
+        this.elInSystem.addEventListener("keyup", (ev)=> {
+            // allow user to insert enter into the system prompt using shift+enter.
+            // while just pressing enter key will lead to setting the system prompt.
+            if ((ev.key === "Enter") && (!ev.shiftKey)) {
+                let chat = this.simpleChats[this.curChatId];
+                chat.add_system_anytime(this.elInSystem.value, this.curChatId);
+                chat.show(this.elDivChat);
+                ev.preventDefault();
+            }
+        });
+
+    }
+
+    /**
+     * Setup a new chat session and optionally switch to it.
+     * @param {string} chatId
+     * @param {boolean} bSwitchSession
+     */
+    new_chat_session(chatId, bSwitchSession=false) {
+        this.simpleChats[chatId] = new SimpleChat();
+        if (bSwitchSession) {
+            this.handle_session_switch(chatId);
+        }
+    }
+
+    /**
+     * Handle user query submit request, wrt specified chat session.
+     * @param {string} chatId
+     * @param {string} apiEP
+     */
+    async handle_user_submit(chatId, apiEP) {
+
+        let chat = this.simpleChats[chatId];
+
+        chat.add_system_anytime(this.elInSystem.value, chatId);
+
+        let content = this.elInUser.value;
+        if (!chat.add(Roles.User, content)) {
+            console.debug(`WARN:SimpleChat:MCUI:${chatId}:HandleUserSubmit:Ignoring empty user input...`);
+            return;
+        }
+        chat.show(this.elDivChat);
+
+        let theBody;
+        let theUrl = gChatURL[apiEP]
+        if (apiEP == ApiEP.Chat) {
+            theBody = chat.request_messages_jsonstr();
+        } else {
+            theBody = chat.request_prompt_jsonstr();
+        }
+
+        this.elInUser.value = "working...";
+        this.elInUser.disabled = true;
+        console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:${theUrl}:ReqBody:${theBody}`);
+        let resp = await fetch(theUrl, {
+            method: "POST",
+            headers: {
+                "Content-Type": "application/json",
+            },
+            body: theBody,
+        });
+
+        let respBody = await resp.json();
+        console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:RespBody:${JSON.stringify(respBody)}`);
+        let assistantMsg;
+        if (apiEP == ApiEP.Chat) {
+            assistantMsg = respBody["choices"][0]["message"]["content"];
+        } else {
+            try {
+                assistantMsg = respBody["choices"][0]["text"];
+            } catch {
+                assistantMsg = respBody["content"];
+            }
+        }
+        chat.add(Roles.Assistant, assistantMsg);
+        if (chatId == this.curChatId) {
+            chat.show(this.elDivChat);
+        } else {
+            console.debug(`DBUG:SimpleChat:MCUI:HandleUserSubmit:ChatId has changed:[${chatId}] [${this.curChatId}]`);
+        }
+        // Purposefully clear at end rather than begin of this function
+        // so that one can switch from chat to completion mode and sequece
+        // in a completion mode with multiple user-assistant chat data
+        // from before to be sent/occur once.
+        if ((apiEP == ApiEP.Completion) && (gbCompletionFreshChatAlways)) {
+            chat.xchat.length = 0;
+        }
+        this.ui_reset_userinput();
+    }
+
+    /**
+     * Show buttons for NewChat and available chat sessions, in the passed elDiv.
+     * If elDiv is undefined/null, then use this.elDivSessions.
+     * Take care of highlighting the selected chat-session's btn.
+     * @param {HTMLDivElement | undefined} elDiv
+     */
+    show_sessions(elDiv=undefined) {
+        if (!elDiv) {
+            elDiv = this.elDivSessions;
+        }
+        elDiv.replaceChildren();
+        // Btn for creating new chat session
+        let btnNew = el_create_button("New CHAT", (ev)=> {
+            if (this.elInUser.disabled) {
+                console.error(`ERRR:SimpleChat:MCUI:NewChat:Current session [${this.curChatId}] awaiting response, ignoring request...`);
+                alert("ERRR:SimpleChat\nMCUI:NewChat\nWait for response to pending query, before starting new chat session");
+                return;
+            }
+            let chatId = `Chat${Object.keys(this.simpleChats).length}`;
+            let chatIdGot = prompt("INFO:SimpleChat\nMCUI:NewChat\nEnter id for new chat session", chatId);
+            if (!chatIdGot) {
+                console.error("ERRR:SimpleChat:MCUI:NewChat:Skipping based on user request...");
+                return;
+            }
+            this.new_chat_session(chatIdGot, true);
+            this.create_session_btn(elDiv, chatIdGot);
+            el_children_config_class(elDiv, chatIdGot, "session-selected", "");
+        });
+        elDiv.appendChild(btnNew);
+        // Btns for existing chat sessions
+        let chatIds = Object.keys(this.simpleChats);
+        for(let cid of chatIds) {
+            let btn = this.create_session_btn(elDiv, cid);
+            if (cid == this.curChatId) {
+                btn.className = "session-selected";
+            }
+        }
+    }
+
+    create_session_btn(elDiv, cid) {
+        let btn = el_create_button(cid, (ev)=>{
+            let target = /** @type{HTMLButtonElement} */(ev.target);
+            console.debug(`DBUG:SimpleChat:MCUI:SessionClick:${target.id}`);
+            if (this.elInUser.disabled) {
+                console.error(`ERRR:SimpleChat:MCUI:SessionClick:${target.id}:Current session [${this.curChatId}] awaiting response, ignoring switch...`);
+                alert("ERRR:SimpleChat\nMCUI:SessionClick\nWait for response to pending query, before switching");
+                return;
+            }
+            this.handle_session_switch(target.id);
+            el_children_config_class(elDiv, target.id, "session-selected", "");
+        });
+        elDiv.appendChild(btn);
+        return btn;
+    }
+
+    /**
+     * Switch ui to the specified chatId and set curChatId to same.
+     * @param {string} chatId
+     */
+    async handle_session_switch(chatId) {
+        let chat = this.simpleChats[chatId];
+        if (chat == undefined) {
+            console.error(`ERRR:SimpleChat:MCUI:HandleSessionSwitch:${chatId} missing...`);
+            return;
+        }
+        this.elInSystem.value = chat.get_system_latest();
+        this.elInUser.value = "";
+        chat.show(this.elDivChat);
+        this.elInUser.focus();
+        this.curChatId = chatId;
+        console.log(`INFO:SimpleChat:MCUI:HandleSessionSwitch:${chatId} entered...`);
+    }
+
+}
+
+
+let gMuitChat;
+const gChatIds = [ "Default", "Other" ];
+
+function startme() {
+    console.log("INFO:SimpleChat:StartMe:Starting...");
+    gMuitChat = new MultiChatUI();
+    for (let cid of gChatIds) {
+        gMuitChat.new_chat_session(cid);
+    }
+    gMuitChat.setup_ui(gChatIds[0]);
+    gMuitChat.show_sessions();
+}
+
+document.addEventListener("DOMContentLoaded", startme);

examples/server/server.cpp

@@ -1019,7 +1019,7 @@ struct server_context {
                         sampler_names.emplace_back(sampler_name);
                     }
                 }
-                slot.sparams.samplers_sequence = sampler_types_from_names(sampler_names, false);
+                slot.sparams.samplers_sequence = llama_sampling_types_from_names(sampler_names, false);
             } else {
                 slot.sparams.samplers_sequence = default_sparams.samplers_sequence;
             }
@@ -1256,7 +1256,7 @@ struct server_context {
         std::vector<std::string> samplers_sequence;
         samplers_sequence.reserve(slot.sparams.samplers_sequence.size());
         for (const auto & sampler_type : slot.sparams.samplers_sequence) {
-            samplers_sequence.emplace_back(sampler_type_to_name_string(sampler_type));
+            samplers_sequence.emplace_back(llama_sampling_type_to_str(sampler_type));
         }
 
         return json {
@@ -2852,7 +2852,7 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                 invalid_param = true;
                 break;
             }
-            if (!parse_kv_override(argv[i], params.kv_overrides)) {
+            if (!string_parse_kv_override(argv[i], params.kv_overrides)) {
                 fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
                 invalid_param = true;
                 break;
@@ -3310,7 +3310,7 @@ int main(int argc, char ** argv) {
     const auto handle_slots_save = [&ctx_server, &res_error, &sparams](const httplib::Request & req, httplib::Response & res, int id_slot) {
         json request_data = json::parse(req.body);
         std::string filename = request_data.at("filename");
-        if (!validate_file_name(filename)) {
+        if (!fs_validate_filename(filename)) {
             res_error(res, format_error_response("Invalid filename", ERROR_TYPE_INVALID_REQUEST));
             return;
         }
@@ -3340,7 +3340,7 @@ int main(int argc, char ** argv) {
     const auto handle_slots_restore = [&ctx_server, &res_error, &sparams](const httplib::Request & req, httplib::Response & res, int id_slot) {
         json request_data = json::parse(req.body);
         std::string filename = request_data.at("filename");
-        if (!validate_file_name(filename)) {
+        if (!fs_validate_filename(filename)) {
             res_error(res, format_error_response("Invalid filename", ERROR_TYPE_INVALID_REQUEST));
             return;
         }

examples/server/tests/features/server.feature

@@ -37,8 +37,8 @@ Feature: llama.cpp server
 
     Examples: Prompts
       | prompt                                                                    | n_predict | re_content                                  | n_prompt | n_predicted | truncated |
-      | I believe the meaning of life is                                          | 8         | (read\|going\|pretty)+                      | 18       | 8           | not       |
+      | I believe the meaning of life is                                          | 8         | (read\|going)+                              | 18       | 8           | not       |
-      | Write a joke about AI from a very long prompt which will not be truncated | 256       | (princesses\|everyone\|kids\|Anna\|forest)+ | 45       | 64          | not       |
+      | Write a joke about AI from a very long prompt which will not be truncated | 256       | (princesses\|everyone\|kids\|Anna\|forest)+ | 46       | 64          | not       |
 
   Scenario: Completion prompt truncated
     Given a prompt:
@@ -67,8 +67,8 @@ Feature: llama.cpp server
 
     Examples: Prompts
       | model        | system_prompt               | user_prompt                          | max_tokens | re_content                        | n_prompt | n_predicted | enable_streaming | truncated |
-      | llama-2      | Book                        | What is the best book                | 8          | (Here\|what)+                     | 76       | 8           | disabled         | not       |
+      | llama-2      | Book                        | What is the best book                | 8          | (Here\|what)+                     | 77       | 8           | disabled         | not       |
-      | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 128        | (thanks\|happy\|bird\|fireplace)+ | -1       | 64          | enabled          |           |
+      | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 128        | (thanks\|happy\|bird\|Annabyear)+ | -1       | 64          | enabled          |           |
 
 
   Scenario Outline: OAI Compatibility w/ response format
@@ -84,7 +84,7 @@ Feature: llama.cpp server
       | response_format                                                     | n_predicted | re_content             |
       | {"type": "json_object", "schema": {"const": "42"}}                  | 5           | "42"                   |
       | {"type": "json_object", "schema": {"items": [{"type": "integer"}]}} | 10          | \[ -300 \]             |
-      | {"type": "json_object"}                                             | 10          | \{ " Saragine.         |
+      | {"type": "json_object"}                                             | 10          | \{ " Jacky.            |
 
 
   Scenario: Tokenize / Detokenize

examples/server/tests/features/slotsave.feature

@@ -26,7 +26,7 @@ Feature: llama.cpp server slot management
     # Since we have cache, this should only process the last tokens
     Given a user prompt "What is the capital of Germany?"
     And   a completion request with no api error
-    Then  24 tokens are predicted matching (Thank|special|Lily)
+    Then  24 tokens are predicted matching (Thank|special)
     And   7 prompt tokens are processed
     # Loading the original cache into slot 0,
     # we should only be processing 1 prompt token and get the same output
@@ -41,7 +41,7 @@ Feature: llama.cpp server slot management
     Given a user prompt "What is the capital of Germany?"
     And   using slot id 1
     And   a completion request with no api error
-    Then  24 tokens are predicted matching (Thank|special|Lily)
+    Then  24 tokens are predicted matching (Thank|special)
     And   1 prompt tokens are processed
 
   Scenario: Erase Slot

examples/sycl/win-build-sycl.bat

@@ -13,10 +13,10 @@ if %errorlevel% neq 0 goto ERROR
 
 ::  for FP16
 ::  faster for long-prompt inference
-::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
+::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
 ::  for FP32
-cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
 if %errorlevel% neq 0 goto ERROR
 ::  build example/main only
 ::  make main

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

@@ -301,8 +301,8 @@ static struct ggml_tensor * llama_build_train_graphs(
         // not capturing these, to silcence warnings
         const int rope_mode = 0;
 
-        return ggml_rope_custom(
+        return ggml_rope_ext(
-            ctx, t, KQ_pos, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
+            ctx, t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
         );
     };
 
@@ -341,7 +341,8 @@ static struct ggml_tensor * llama_build_train_graphs(
         struct ggml_tensor * t15 = ggml_permute      (ctx, t12, 0, 3, 1, 2);                        set_name(t15, "t15");     assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch);
         struct ggml_tensor * t16;
         if (enable_flash_attn) {
-            t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                        set_name(t16, "t16");     assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
+            GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported");
+            //t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                        set_name(t16, "t16");     assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
         } else {
             struct ggml_tensor * t16_0 = ggml_mul_mat              (ctx, t14, t13);                 set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
             struct ggml_tensor * t16_1 = ggml_scale_inplace        (ctx, t16_0, kv_scale);          set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);

ggml-common.h

@@ -65,13 +65,8 @@ typedef sycl::half2 ggml_half2;
 // QK = number of values after dequantization
 // QK_K = super-block size
 
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
 #define QK_K 256
 #define K_SCALE_SIZE 12
-#endif // GGML_QKK_64
 
 #if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
 // QR = QK / number of values before dequantization
@@ -131,13 +126,8 @@ typedef sycl::half2 ggml_half2;
 #define QI4_NL (QK4_NL / (4*QR4_NL))
 #define QR4_NL 2
 
-#if QK_K == 64
-#define QI4_XS QI4_NL
-#define QR4_XS QR4_NL
-#else
 #define QI4_XS (QK_K / (4*QR4_XS))
 #define QR4_XS 8
-#endif
 
 #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
 
@@ -228,15 +218,6 @@ static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wro
 // weight is represented as x = a * q
 // 16 blocks of 16 elements each
 // Effectively 3.4375 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8]; // quants - high bit
-    uint8_t qs[QK_K/4];    // quants - low 2 bits
-    uint8_t scales[2];
-    ggml_half d;           // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
 typedef struct {
     uint8_t hmask[QK_K/8]; // quants - high bit
     uint8_t qs[QK_K/4];    // quants - low 2 bits
@@ -244,20 +225,11 @@ typedef struct {
     ggml_half d;           // super-block scale
 } block_q3_K;
 static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif
 
 // 4-bit quantization
 // 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 4.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d[2];     // super-block scales/mins
-    uint8_t scales[2];  // 4-bit block scales/mins
-    uint8_t qs[QK_K/2]; // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
 typedef struct {
     union {
         struct {
@@ -270,21 +242,11 @@ typedef struct {
     uint8_t qs[QK_K/2];           // 4--bit quants
 } block_q4_K;
 static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif
 
 // 5-bit quantization
 // 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 5.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d;             // super-block scale
-    int8_t  scales[QK_K/16]; // 8-bit block scales
-    uint8_t qh[QK_K/8];      // quants, high bit
-    uint8_t qs[QK_K/2];      // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
 typedef struct {
     union {
         struct {
@@ -298,7 +260,6 @@ typedef struct {
     uint8_t qs[QK_K/2];           // quants, low 4 bits
 } block_q5_K;
 static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
 
 // 6-bit quantization
 // weight is represented as x = a * q
@@ -356,11 +317,7 @@ typedef struct {
 static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
 
 // 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
 #define IQ3S_N_SCALE QK_K/64
-#endif
 typedef struct {
     ggml_half d;
     uint8_t qs[QK_K/4];
@@ -381,16 +338,9 @@ static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wron
 typedef struct {
     uint8_t  qs[QK_K/8];      // grid index, low 8 bits
     uint8_t  qh[QK_K/16];     // grid index, high 3 bits + grid shift bit (for two groups of 8)
-#if QK_K == 64
-    ggml_half d;
-#endif
     uint8_t  scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
 } block_iq1_m;
-#if QK_K == 64
-static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding");
-#else
 static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
-#endif
 
 // Used by IQ1_M quants
 typedef union {
@@ -406,9 +356,6 @@ typedef struct {
 } block_iq4_nl;
 static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
 
-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-#else
 typedef struct {
     ggml_half d;
     uint16_t scales_h;
@@ -416,7 +363,6 @@ typedef struct {
     uint8_t  qs[QK_K/2];
 } block_iq4_xs;
 static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
-#endif
 
 #endif // GGML_COMMON_DECL
 #endif // GGML_COMMON_DECL

ggml-cuda/convert.cu

@@ -131,7 +131,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
     const block_q2_K * x = (const block_q2_K *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t n   = tid/32;
     const int64_t l   = tid - 32*n;
     const int64_t is  = 8*n + l/16;
@@ -145,17 +144,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
     y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
     y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
     y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
-#else
-    const int64_t is = tid/16;  // 0 or 1
-    const int64_t il = tid%16;  // 0...15
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    dst_t * y = yy + i*QK_K + 16*is + il;
-    float dall = __low2half(x[i].dm);
-    float dmin = __high2half(x[i].dm);
-    y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
-    y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
-#endif
-
 }
 
 template<typename dst_t>
@@ -164,7 +152,6 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
     const int64_t i = blockIdx.x;
     const block_q3_K * x = (const block_q3_K *) vx;
 
-#if QK_K == 256
     const int64_t r = threadIdx.x/4;
     const int64_t tid = r/2;
     const int64_t is0 = r%2;
@@ -188,31 +175,8 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
     const uint8_t * hm = x[i].hmask;
 
     for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
-#else
-    const int64_t tid = threadIdx.x;
-    const int64_t is  = tid/16;  // 0 or 1
-    const int64_t il  = tid%16;  // 0...15
-    const int64_t im  = il/8;    // 0...1
-    const int64_t in  = il%8;    // 0...7
-
-    dst_t * y = yy + i*QK_K + 16*is + il;
-
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    const uint8_t h = x[i].hmask[in] >> (2*is + im);
-    const float   d = (float)x[i].d;
-
-    if (is == 0) {
-        y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    } else {
-        y[ 0] = d * ((x[i].scales[0] >>  4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] >>  4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    }
-#endif
-
 }
 
-#if QK_K == 256
 static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
     if (j < 4) {
         d = q[j] & 63; m = q[j + 4] & 63;
@@ -221,7 +185,6 @@ static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t
         m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
     }
 }
-#endif
 
 template<typename dst_t>
 static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
@@ -229,7 +192,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
 
     const int64_t i = blockIdx.x;
 
-#if QK_K == 256
     // assume 32 threads
     const int64_t tid = threadIdx.x;
     const int64_t il  = tid/8;
@@ -253,15 +215,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
         y[l + 0] = d1 * (q[l] & 0xF) - m1;
         y[l +32] = d2 * (q[l] >>  4) - m2;
     }
-#else
-    const int64_t tid = threadIdx.x;
-    const uint8_t * q = x[i].qs;
-    dst_t * y = yy + i*QK_K;
-    const float d = (float)x[i].dm[0];
-    const float m = (float)x[i].dm[1];
-    y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
-    y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
-#endif
 }
 
 template<typename dst_t>
@@ -270,7 +223,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
 
     const int64_t i = blockIdx.x;
 
-#if QK_K == 256
     // assume 64 threads - this is very slightly better than the one below
     const int64_t tid = threadIdx.x;
     const int64_t il  = tid/16;   // il is in 0...3
@@ -297,18 +249,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
     hm <<= 1;
     y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
     y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
-#else
-    const int64_t tid = threadIdx.x;
-    const uint8_t q = x[i].qs[tid];
-    const int64_t im = tid/8;  // 0...3
-    const int64_t in = tid%8;  // 0...7
-    const int64_t is = tid/16; // 0 or 1
-    const uint8_t h = x[i].qh[in] >> im;
-    const float d = x[i].d;
-    dst_t * y = yy + i*QK_K + tid;
-    y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
-    y[32] = d * x[i].scales[is+2] * ((q >>  4) - ((h >> 4) & 1 ? 0 : 16));
-#endif
 }
 
 template<typename dst_t>
@@ -316,7 +256,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
     const block_q6_K * x = (const block_q6_K *) vx;
 
     const int64_t i = blockIdx.x;
-#if QK_K == 256
 
     // assume 64 threads - this is very slightly better than the one below
     const int64_t tid = threadIdx.x;
@@ -336,24 +275,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
     y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
     y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
     y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
-#else
-
-    // assume 32 threads
-    const int64_t tid = threadIdx.x;
-    const int64_t ip  = tid/16;         // 0 or 1
-    const int64_t il  = tid - 16*ip;    // 0...15
-
-    dst_t * y = yy + i*QK_K + 16*ip + il;
-
-    const float d = x[i].d;
-
-    const uint8_t   ql = x[i].ql[16*ip + il];
-    const uint8_t   qh = x[i].qh[il] >> (2*ip);
-    const int8_t  * sc = x[i].scales;
-
-    y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
-    y[32] = d * sc[ip+2] * ((int8_t)((ql  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
-#endif
 }
 
 template<typename dst_t>
@@ -363,7 +284,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
     const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -374,10 +294,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
     const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -387,7 +303,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
     const block_iq2_xs * x = (const block_iq2_xs *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -396,10 +311,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
     const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -409,7 +320,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
     const block_iq2_s * x = (const block_iq2_s *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -417,10 +327,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
     const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -430,7 +336,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
     const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -445,10 +350,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
         y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
         y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -458,7 +359,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
     const block_iq3_s * x = (const block_iq3_s *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -471,10 +371,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
         y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
         y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -484,7 +380,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
     const block_iq1_s * x = (const block_iq1_s  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -497,10 +392,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
     for (int j = 0; j < 8; ++j) {
         y[j] = d * (q[j] + delta);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
 template<typename dst_t>
@@ -510,7 +401,6 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
     const block_iq1_m * x = (const block_iq1_m  *) vx;
 
     const int64_t tid = threadIdx.x;
-#if QK_K == 256
     const int64_t il = tid/8; // 0...3
     const int64_t ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -527,13 +417,8 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
     for (int j = 0; j < 8; ++j) {
         y[j] = d * (q[j] + delta);
     }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }
 
-
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
 
@@ -550,10 +435,8 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst
         y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
         y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
     }
-
 }
 
-#if QK_K != 64
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
     const int64_t i   = blockIdx.x;
@@ -570,7 +453,6 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
         y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
     }
 }
-#endif
 
 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
@@ -592,21 +474,13 @@ static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half *
 template<typename dst_t>
 static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q2_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
 static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q3_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
@@ -632,21 +506,13 @@ static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k
 template<typename dst_t>
 static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q5_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
 static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q6_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template<typename dst_t>
@@ -700,11 +566,7 @@ static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t
 template<typename dst_t>
 static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
     const int nb = (k + QK_K - 1) / QK_K;
-#if QK_K == 64
-    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
-#else
     dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }
 
 template <typename src_t, typename dst_t>

ggml-cuda/dmmv.cu

@@ -22,7 +22,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
     const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
     const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
 
@@ -71,37 +70,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
         tmp += dall * sum1 - dmin * sum2;
 
     }
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;
-
-    uint32_t uaux[2];
-    const uint8_t * d = (const uint8_t *)uaux;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint32_t * s = (const uint32_t *)x[i].scales;
-
-        uaux[0] = s[0] & 0x0f0f0f0f;
-        uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;
-
-        const float2 dall = __half22float2(x[i].dm);
-
-        float sum1 = 0, sum2 = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t ql = q[l];
-            sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
-                  + y[l+16] * d[1] * ((ql >> 2) & 3)
-                  + y[l+32] * d[2] * ((ql >> 4) & 3)
-                  + y[l+48] * d[3] * ((ql >> 6) & 3);
-            sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
-        }
-        tmp += dall.x * sum1 - dall.y * sum2;
-    }
-#endif
 
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
@@ -123,8 +91,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
-
     const uint16_t kmask1 = 0x0303;
     const uint16_t kmask2 = 0x0f0f;
 
@@ -175,34 +141,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,
         tmp += d * sum;
 
     }
-#else
-
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;         // 0...15 or 0...14
-    const int in = offset/8;                                 // 0 or 1
-    const int im = offset%8;                                 // 0...7
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint8_t * s = x[i].scales;
-
-        const float dall = (float)x[i].d;
-
-        float sum = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t hl = x[i].hmask[im+l] >> in;
-            const uint8_t ql = q[l];
-            sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
-                 + y[l+16] * dall * ((s[0] >>  4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
-                 + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
-                 + y[l+48] * dall * ((s[1] >>  4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
-        }
-        tmp += sum;
-    }
-#endif
 
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
@@ -221,7 +159,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
 
     const block_q4_K * x = (const block_q4_K *)vx + ib0;
 
-#if QK_K == 256
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
     const uint16_t kmask3 = 0xc0c0;
@@ -306,36 +243,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
 #endif
 
     }
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    uint16_t aux16[2];
-    const uint8_t * s = (const uint8_t *)aux16;
-
-    float tmp = 0;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const float   * y = yy + i*QK_K + step;
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux16[0] = a[0] & 0x0f0f;
-        aux16[1] = (a[0] >> 4) & 0x0f0f;
-        const float d = (float)x[i].dm[0];
-        const float m = (float)x[i].dm[1];
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
-                 + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
-                 + y[j+32] * (d * s[1] * (q[j+ 0] >>  4) - m * s[3])
-                 + y[j+48] * (d * s[1] * (q[j+16] >>  4) - m * s[3]);
-        }
-        tmp += sum;
-    }
-
-#endif
 
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
@@ -355,7 +262,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
     const uint16_t kmask3 = 0xc0c0;
@@ -426,30 +332,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
         tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
     }
 
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
-    const int step = tid * K_QUANTS_PER_ITERATION;
-    const int im = step/8;
-    const int in = step%8;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const int8_t  * s = x[i].scales;
-        const float   * y = yy + i*QK_K + step;
-        const float     d = x[i].d;
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            const uint8_t h = x[i].qh[in+j] >> im;
-            sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
-                 + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
-                 + y[j+32] * d * s[2] * ((q[j+ 0] >>  4) - ((h >> 4) & 1 ? 0 : 16))
-                 + y[j+48] * d * s[3] * ((q[j+16] >>  4) - ((h >> 6) & 1 ? 0 : 16));
-        }
-        tmp += sum;
-    }
-#endif
-
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
 
@@ -470,8 +352,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
 
     const block_q6_K * x = (const block_q6_K *)vx + ib0;
 
-#if QK_K == 256
-
     const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
     const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
 
@@ -526,37 +406,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,
 
     }
 
-#else
-
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0...3
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + step;
-        const uint8_t * ql = x[i].ql + step;
-        const uint8_t * qh = x[i].qh + step;
-        const int8_t  * s  = x[i].scales;
-
-        const float d = x[i+0].d;
-
-        float sum = 0;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
-                 + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
-                 + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >>  4) | ((qh[j] & 0x30) >> 0)) - 32)
-                 + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >>  4) | ((qh[j] & 0xc0) >> 2)) - 32);
-        }
-        tmp += sum;
-
-    }
-
-#endif
-
     // sum up partial sums and write back result
     tmp = warp_reduce_sum(tmp);
 

ggml-cuda/fattn-tile-f16.cu

@@ -83,7 +83,7 @@ static __global__ void flash_attn_tile_ext_f16(
         for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
             const int i = i0 + threadIdx.x;
 
-            const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
+            const float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
             Q_h2[j][i] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
         }
     }
@@ -238,6 +238,10 @@ static __global__ void flash_attn_tile_ext_f16(
     for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
         const int j_VKQ = j_VKQ_0 + threadIdx.y;
 
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+
         half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
         kqsum_j = warp_reduce_sum(kqsum_j);
 

ggml-cuda/fattn-tile-f32.cu

@@ -79,7 +79,7 @@ static __global__ void flash_attn_tile_ext_f32(
 
 #pragma unroll
         for (int i0 = 0; i0 < D; i0 += 2*WARP_SIZE) {
-            float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i0/2 + threadIdx.x];
+            float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i0/2 + threadIdx.x] : make_float2(0.0f, 0.0f);
             Q_f[j][i0 + 0*WARP_SIZE + threadIdx.x] = tmp.x * scale;
             Q_f[j][i0 + 1*WARP_SIZE + threadIdx.x] = tmp.y * scale;
         }
@@ -237,6 +237,10 @@ static __global__ void flash_attn_tile_ext_f32(
     for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
         const int j_VKQ = j_VKQ_0 + threadIdx.y;
 
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+
         float kqsum_j = kqsum[j_VKQ_0/nwarps];
         kqsum_j = warp_reduce_sum(kqsum_j);
 
@@ -283,12 +287,8 @@ void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor *
 }
 
 void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * KQV = dst;
     const ggml_tensor * Q = dst->src[0];
 
-    const int32_t precision = KQV->op_params[2];
-    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
-
     if (Q->ne[1] <= 16) {
         constexpr int cols_per_block = 16;
         constexpr int parallel_blocks = 4;

ggml-cuda/fattn-vec-f16.cu

@@ -94,7 +94,7 @@ static __global__ void flash_attn_vec_ext_f16(
         for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
             const int i = i0 + threadIdx.x;
 
-            const float2 tmp = Q_f2[j*(nb01/sizeof(float2)) + i];
+            const float2 tmp = ncols <= 2 || ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
             Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
         }
     }
@@ -212,6 +212,10 @@ static __global__ void flash_attn_vec_ext_f16(
 
 #pragma unroll
     for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        if (ncols > 2 && ic0 + j_VKQ >= ne01) {
+            break;
+        }
+
         kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
         kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
 
@@ -223,7 +227,7 @@ static __global__ void flash_attn_vec_ext_f16(
         dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
     }
 
-    if (parallel_blocks != 1 && tid < ncols) {
+    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
         dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
     }
 #else

ggml-cuda/fattn-vec-f32.cu

@@ -91,7 +91,7 @@ static __global__ void flash_attn_vec_ext_f32(
         for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
             const int i = i0 + threadIdx.x;
 
-            Q_h2[j][i0/WARP_SIZE]    = Q_f2[j*(nb01/sizeof(float2)) + i];
+            Q_h2[j][i0/WARP_SIZE]    = ncols <= 2 || ic0 + j ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
             Q_h2[j][i0/WARP_SIZE].x *= scale;
             Q_h2[j][i0/WARP_SIZE].y *= scale;
         }
@@ -200,6 +200,10 @@ static __global__ void flash_attn_vec_ext_f32(
 
 #pragma unroll
     for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        if (ncols > 2 && ic0 + j_VKQ >= ne01) {
+            break;
+        }
+
         kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
         kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
 
@@ -211,7 +215,7 @@ static __global__ void flash_attn_vec_ext_f32(
         dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
     }
 
-    if (parallel_blocks != 1 && tid < ncols) {
+    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
         dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
     }
 }

ggml-cuda/rope.cu

@@ -58,10 +58,10 @@ static __global__ void rope(
     dst[i + 1] = x0*sin_theta + x1*cos_theta;
 }
 
-template<typename T, bool has_pos>
+template<typename T, bool has_pos, bool has_freq_facs>
 static __global__ void rope_neox(
     const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims, const float * freq_factors
 ) {
     const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
 
@@ -88,7 +88,9 @@ static __global__ void rope_neox(
     float cur_rot = inv_ndims * ic - ib;
 
     const int p = has_pos ? pos[i2] : 0;
-    const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f);
+    const float freq_factor = has_freq_facs ? freq_factors[ic/2] : 1.0f;
+
+    const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f)/freq_factor;
 
     float cos_theta, sin_theta;
     rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
@@ -164,7 +166,7 @@ static void rope_cuda(
 template<typename T>
 static void rope_neox_cuda(
     const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream
 ) {
     GGML_ASSERT(ncols % 2 == 0);
     const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
@@ -175,16 +177,30 @@ static void rope_neox_cuda(
     const float inv_ndims = -1.0f / n_dims;
 
     if (pos == nullptr) {
-        rope_neox<T, false><<<block_nums, block_dims, 0, stream>>>(
+        if (freq_factors == nullptr) {
+            rope_neox<T, false, false><<<block_nums, block_dims, 0, stream>>>(
                 x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
-            theta_scale, inv_ndims
+                theta_scale, inv_ndims, freq_factors
                 );
         } else {
-        rope_neox<T, true><<<block_nums, block_dims, 0, stream>>>(
+            rope_neox<T, false, true><<<block_nums, block_dims, 0, stream>>>(
                 x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
-            theta_scale, inv_ndims
+                theta_scale, inv_ndims, freq_factors
                 );
         }
+    } else {
+        if (freq_factors == nullptr) {
+            rope_neox<T, true, false><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, inv_ndims, freq_factors
+                );
+        } else {
+            rope_neox<T, true, true><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, inv_ndims, freq_factors
+                );
+        }
+    }
 }
 
 static void rope_glm_f32_cuda(
@@ -214,24 +230,27 @@ static void rope_cuda_f32(
 
 static void rope_neox_cuda_f16(
     const half * x, half * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream) {
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
 
-    rope_neox_cuda<half>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
+    rope_neox_cuda<half>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
 }
 
 static void rope_neox_cuda_f32(
     const float * x, float * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream
 ) {
 
-    rope_neox_cuda<float>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, stream);
+    rope_neox_cuda<float>(x, dst, ncols, n_dims, nrows, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
 }
 
 void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0 = dst->src[0];
     const ggml_tensor * src1 = dst->src[1];
+    const ggml_tensor * src2 = dst->src[2];
+
     const float * src0_d = (const float *)src0->data;
     const float * src1_d = (const float *)src1->data;
+
     float * dst_d = (float *)dst->data;
     cudaStream_t stream = ctx.stream();
 
@@ -241,7 +260,6 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
 
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
-    const int64_t ne2 = dst->ne[2];
     const int64_t nrows = ggml_nrows(src0);
 
     //const int n_past      = ((int32_t *) dst->op_params)[0];
@@ -259,16 +277,22 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
+    const float * freq_factors = nullptr;
     const int32_t * pos = nullptr;
-    if ((mode & 1) == 0) {
-        GGML_ASSERT(src1->type == GGML_TYPE_I32);
-        GGML_ASSERT(src1->ne[0] == ne2);
-        pos = (const int32_t *) src1_d;
-    }
 
     const bool is_neox = mode & 2;
     const bool is_glm  = mode & 4;
 
+    pos = (const int32_t *) src1_d;
+
+    if (is_neox) {
+        if (src2 != nullptr) {
+            freq_factors = (const float *) src2->data;
+        }
+    } else {
+        GGML_ASSERT(src2 == nullptr && "TODO: freq_factors not implemented for !is_neox");
+    }
+
     rope_corr_dims corr_dims;
     ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v);
 
@@ -280,12 +304,12 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
         if (src0->type == GGML_TYPE_F32) {
             rope_neox_cuda_f32(
                 (const float *)src0_d, (float *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, stream
+                attn_factor, corr_dims, freq_factors, stream
             );
         } else if (src0->type == GGML_TYPE_F16) {
             rope_neox_cuda_f16(
                 (const half *)src0_d, (half *)dst_d, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, stream
+                attn_factor, corr_dims, freq_factors, stream
             );
         } else {
             GGML_ASSERT(false);

ggml-cuda/vecdotq.cuh

@@ -712,7 +712,6 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
 static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 
-#ifndef GGML_QKK_64
     const block_q4_K * bq4_K = (const block_q4_K *) vbq;
 
     int    v[2];
@@ -754,58 +753,11 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
     }
 
     return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
-
-#else
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    const block_q4_K * bq4_K = (const block_q4_K *) vbq;
-
-    float sumf_d = 0.0f;
-    float sumf_m = 0.0f;
-
-    uint16_t aux16[2];
-    const uint8_t * s = (const uint8_t *)aux16;
-
-    const uint16_t * a = (const uint16_t *)bq4_K->scales;
-    aux16[0] = a[0] & 0x0f0f;
-    aux16[1] = (a[0] >> 4) & 0x0f0f;
-
-    const float dall = bq4_K->dm[0];
-    const float dmin = bq4_K->dm[1];
-
-    const float d8_1 = __low2float(bq8_1[0].ds);
-    const float d8_2 = __low2float(bq8_1[1].ds);
-
-    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
-    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
-    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
-    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
-
-    const int * q4 = (const int *)bq4_K->qs + (iqs/2);
-    const int v1 = q4[0];
-    const int v2 = q4[4];
-
-    const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0));
-    const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0));
-    const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0));
-    const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0));
-
-    sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]);
-    sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]);
-
-    return dall * sumf_d - dmin * sumf_m;
-
-#else
-    NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 
-#ifndef GGML_QKK_64
     const block_q5_K * bq5_K = (const block_q5_K *) vbq;
 
     int   vl[2];
@@ -847,48 +799,6 @@ static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
     }
 
     return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
-
-#else
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    const block_q5_K * bq5_K = (const block_q5_K *) vbq;
-
-    const int8_t * s = bq5_K->scales;
-
-    const float d = bq5_K->d;
-
-    const float d8_1 = __low2half(bq8_1[0].ds);
-    const float d8_2 = __low2half(bq8_1[1].ds);
-
-    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
-    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
-    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
-    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
-
-    const int * ql = (const int *)bq5_K->qs + (iqs/2);
-    const int vl1 = ql[0];
-    const int vl2 = ql[4];
-
-    const int step = 4 * (iqs/2); // 0, 4, 8, 12
-    const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6
-    const int in = step%8; // 0, 4, 0, 4
-    const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
-
-    const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f);
-    const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f);
-    const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f);
-    const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f);
-
-    const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1])
-                       + d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]);
-
-    return d * sumf_d;
-
-#else
-    NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
@@ -919,7 +829,6 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
 
 static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-#if QK_K == 256
     const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq;
 
 #if QR2_XXS == 8
@@ -960,15 +869,11 @@ static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
     }
     return d * (sumi1 + sumi2);
 #endif
-#else
-    NO_DEVICE_CODE;
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-#if QK_K == 256
     const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq;
 
     const int ib32 = iqs;
@@ -1002,17 +907,12 @@ static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
     GGML_UNUSED(ksigns64);
     NO_DEVICE_CODE;
 #endif
-#else
-    GGML_UNUSED(ksigns64);
-    NO_DEVICE_CODE;
-#endif
 }
 
 // TODO
 static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-#if QK_K == 256
     const block_iq2_s * bq2 = (const block_iq2_s *) vbq;
 
     const int ib32 = iqs;
@@ -1048,16 +948,11 @@ static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
     GGML_UNUSED(ksigns64);
     NO_DEVICE_CODE;
 #endif
-#else
-    GGML_UNUSED(ksigns64);
-    NO_DEVICE_CODE;
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-#if QK_K == 256
     const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq;
 
     const int ib32 = iqs;
@@ -1082,16 +977,12 @@ static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
 #else
     NO_DEVICE_CODE;
 #endif
-#else
-    NO_DEVICE_CODE;
-#endif
 }
 
 // TODO: don't use lookup table for signs
 static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-#if QK_K == 256
     const block_iq3_s * bq2 = (const block_iq3_s *) vbq;
 
     const int ib32 = iqs;
@@ -1114,14 +1005,10 @@ static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
 #else
     NO_DEVICE_CODE;
 #endif
-#else
-    NO_DEVICE_CODE;
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-#if QK_K == 256
     const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
 
     const int ib32 = iqs;
@@ -1149,14 +1036,10 @@ static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
     const float d = d1q * __low2float (bq8_1[ib32].ds);
     const float m = d1q * __high2float(bq8_1[ib32].ds);
     return d * sumi + m * delta;
-#else
-    NO_DEVICE_CODE;
-#endif
 }
 
 static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-#if QK_K == 256
     const block_iq1_m * bq1 = (const block_iq1_m *) vbq;
 
     const int ib32 = iqs;
@@ -1192,9 +1075,6 @@ static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
     scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
     const float d = (float)scale.f16 * __low2float (bq8_1[ib32].ds);
     return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1));
-#else
-    NO_DEVICE_CODE;
-#endif
 }
 
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
@@ -1250,9 +1130,7 @@ static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
 static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
 
-#if QK_K == 256
 #if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-
     const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
     const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
 
@@ -1270,10 +1148,6 @@ static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
         sumi2 = __dp4a(v2, q8[j+4], sumi2);
     }
     return d * (sumi1 + sumi2);
-
-#else
-    NO_DEVICE_CODE;
-#endif
 #else
     return vec_dot_iq4_xs_q8_1(vbq, bq8_1, iqs);
 #endif

ggml-kompute.cpp

@@ -1677,6 +1677,10 @@ static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml
                     } break;
                 case GGML_OP_ROPE:
                     {
+#pragma message("TODO: implement phi3 frequency factors support")
+#pragma message("      https://github.com/ggerganov/llama.cpp/pull/7225")
+                        GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
+
                         GGML_ASSERT(ne10 == ne02);
                         GGML_ASSERT(src0t == dstt);
                         // const int n_past = ((int32_t *) dst->op_params)[0];

ggml-metal.m

@@ -381,10 +381,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
                 // dictionary of preprocessor macros
                 NSMutableDictionary * prep = [NSMutableDictionary dictionary];
 
-#ifdef GGML_QKK_64
-                prep[@"GGML_QKK_64"] = @(1);
-#endif
-
                 MTLCompileOptions* options = [MTLCompileOptions new];
                 options.preprocessorMacros = prep;
 
@@ -927,12 +923,22 @@ static enum ggml_status ggml_metal_graph_compute(
             const int64_t  ne10 = src1 ? src1->ne[0] : 0;
             const int64_t  ne11 = src1 ? src1->ne[1] : 0;
             const int64_t  ne12 = src1 ? src1->ne[2] : 0;
-            const int64_t  ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
+            const int64_t  ne13 = src1 ? src1->ne[3] : 0;
 
             const uint64_t nb10 = src1 ? src1->nb[0] : 0;
             const uint64_t nb11 = src1 ? src1->nb[1] : 0;
             const uint64_t nb12 = src1 ? src1->nb[2] : 0;
-            const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
+            const uint64_t nb13 = src1 ? src1->nb[3] : 0;
+
+            const int64_t  ne20 = src2 ? src2->ne[0] : 0;
+            const int64_t  ne21 = src2 ? src2->ne[1] : 0;
+            const int64_t  ne22 = src2 ? src2->ne[2] : 0; GGML_UNUSED(ne22);
+            const int64_t  ne23 = src2 ? src2->ne[3] : 0; GGML_UNUSED(ne23);
+
+            const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
+            const uint64_t nb21 = src2 ? src2->nb[1] : 0;
+            const uint64_t nb22 = src2 ? src2->nb[2] : 0;
+            const uint64_t nb23 = src2 ? src2->nb[3] : 0;
 
             const int64_t  ne0  =  dst ?  dst->ne[0] : 0;
             const int64_t  ne1  =  dst ?  dst->ne[1] : 0;
@@ -1763,11 +1769,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_Q3_K) {
-#ifdef GGML_QKK_64
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#else
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#endif
                             }
                             else if (src0t == GGML_TYPE_Q5_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -1785,16 +1787,6 @@ static enum ggml_status ggml_metal_graph_compute(
                         const int n_as = src0->ne[2];
 
                         // src2 = ids
-                        const int64_t  ne20 = src2->ne[0];
-                        const int64_t  ne21 = src2->ne[1];
-                        const int64_t  ne22 = src2->ne[2]; GGML_UNUSED(ne22);
-                        const int64_t  ne23 = src2->ne[3]; GGML_UNUSED(ne23);
-
-                        const uint64_t nb20 = src2->nb[0]; GGML_UNUSED(nb20);
-                        const uint64_t nb21 = src2->nb[1];
-                        const uint64_t nb22 = src2->nb[2]; GGML_UNUSED(nb22);
-                        const uint64_t nb23 = src2->nb[3]; GGML_UNUSED(nb23);
-
                         const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t);
 
                         GGML_ASSERT(src2t == GGML_TYPE_I32);
@@ -2018,12 +2010,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                     {
                                         nth0 = 4;
                                         nth1 = 16;
-                                    #if QK_K == 64
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
-                                    #else
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
-                                    #endif
-
                                     } break;
                                 default:
                                     {
@@ -2088,11 +2075,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_Q3_K) {
-#ifdef GGML_QKK_64
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#else
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#endif
                             }
                             else if (src0t == GGML_TYPE_Q5_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
@@ -2244,7 +2227,13 @@ static enum ggml_status ggml_metal_graph_compute(
                         // skip 3, n_ctx, used in GLM RoPE, unimplemented in metal
                         const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
 
-                        float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+                        float freq_base;
+                        float freq_scale;
+                        float ext_factor;
+                        float attn_factor;
+                        float beta_fast;
+                        float beta_slow;
+
                         memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
                         memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
                         memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
@@ -2252,6 +2241,15 @@ static enum ggml_status ggml_metal_graph_compute(
                         memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
                         memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
 
+                        const bool is_neox = mode & 2;
+                        const bool is_glm  = mode & 4;
+
+                        GGML_ASSERT(!is_glm && "GLM RoPE not implemented in Metal");
+
+                        if (!is_neox) {
+                            GGML_ASSERT(id_src2 == nil && "TODO: freq_factors not implemented for !is_neox");
+                        }
+
                         id<MTLComputePipelineState> pipeline = nil;
 
                         switch (src0->type) {
@@ -2263,33 +2261,38 @@ static enum ggml_status ggml_metal_graph_compute(
                         [encoder setComputePipelineState:pipeline];
                         [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
                         [encoder setBuffer:id_src1     offset:offs_src1        atIndex:1];
-                        [encoder setBuffer:id_dst      offset:offs_dst         atIndex:2];
+                        if (id_src2 != nil) {
-                        [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:3];
+                            [encoder setBuffer:id_src2 offset:offs_src2        atIndex:2];
-                        [encoder setBytes:&ne01        length:sizeof( int64_t) atIndex:4];
+                        } else {
-                        [encoder setBytes:&ne02        length:sizeof( int64_t) atIndex:5];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:2];
-                        [encoder setBytes:&ne03        length:sizeof( int64_t) atIndex:6];
+                        }
-                        [encoder setBytes:&nb00        length:sizeof(uint64_t) atIndex:7];
+                        [encoder setBuffer:id_dst      offset:offs_dst         atIndex:3];
-                        [encoder setBytes:&nb01        length:sizeof(uint64_t) atIndex:8];
+                        [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:4];
-                        [encoder setBytes:&nb02        length:sizeof(uint64_t) atIndex:9];
+                        [encoder setBytes:&ne01        length:sizeof( int64_t) atIndex:5];
-                        [encoder setBytes:&nb03        length:sizeof(uint64_t) atIndex:10];
+                        [encoder setBytes:&ne02        length:sizeof( int64_t) atIndex:6];
-                        [encoder setBytes:&ne0         length:sizeof( int64_t) atIndex:11];
+                        [encoder setBytes:&ne03        length:sizeof( int64_t) atIndex:7];
-                        [encoder setBytes:&ne1         length:sizeof( int64_t) atIndex:12];
+                        [encoder setBytes:&nb00        length:sizeof(uint64_t) atIndex:8];
-                        [encoder setBytes:&ne2         length:sizeof( int64_t) atIndex:13];
+                        [encoder setBytes:&nb01        length:sizeof(uint64_t) atIndex:9];
-                        [encoder setBytes:&ne3         length:sizeof( int64_t) atIndex:14];
+                        [encoder setBytes:&nb02        length:sizeof(uint64_t) atIndex:10];
-                        [encoder setBytes:&nb0         length:sizeof(uint64_t) atIndex:15];
+                        [encoder setBytes:&nb03        length:sizeof(uint64_t) atIndex:11];
-                        [encoder setBytes:&nb1         length:sizeof(uint64_t) atIndex:16];
+                        [encoder setBytes:&ne0         length:sizeof( int64_t) atIndex:12];
-                        [encoder setBytes:&nb2         length:sizeof(uint64_t) atIndex:17];
+                        [encoder setBytes:&ne1         length:sizeof( int64_t) atIndex:13];
-                        [encoder setBytes:&nb3         length:sizeof(uint64_t) atIndex:18];
+                        [encoder setBytes:&ne2         length:sizeof( int64_t) atIndex:14];
-                        [encoder setBytes:&n_past      length:sizeof(     int) atIndex:19];
+                        [encoder setBytes:&ne3         length:sizeof( int64_t) atIndex:15];
-                        [encoder setBytes:&n_dims      length:sizeof(     int) atIndex:20];
+                        [encoder setBytes:&nb0         length:sizeof(uint64_t) atIndex:16];
-                        [encoder setBytes:&mode        length:sizeof(     int) atIndex:21];
+                        [encoder setBytes:&nb1         length:sizeof(uint64_t) atIndex:17];
-                        [encoder setBytes:&n_orig_ctx  length:sizeof(     int) atIndex:22];
+                        [encoder setBytes:&nb2         length:sizeof(uint64_t) atIndex:18];
-                        [encoder setBytes:&freq_base   length:sizeof(   float) atIndex:23];
+                        [encoder setBytes:&nb3         length:sizeof(uint64_t) atIndex:19];
-                        [encoder setBytes:&freq_scale  length:sizeof(   float) atIndex:24];
+                        [encoder setBytes:&n_past      length:sizeof(     int) atIndex:20];
-                        [encoder setBytes:&ext_factor  length:sizeof(   float) atIndex:25];
+                        [encoder setBytes:&n_dims      length:sizeof(     int) atIndex:21];
-                        [encoder setBytes:&attn_factor length:sizeof(   float) atIndex:26];
+                        [encoder setBytes:&mode        length:sizeof(     int) atIndex:22];
-                        [encoder setBytes:&beta_fast   length:sizeof(   float) atIndex:27];
+                        [encoder setBytes:&n_orig_ctx  length:sizeof(     int) atIndex:23];
-                        [encoder setBytes:&beta_slow   length:sizeof(   float) atIndex:28];
+                        [encoder setBytes:&freq_base   length:sizeof(   float) atIndex:24];
+                        [encoder setBytes:&freq_scale  length:sizeof(   float) atIndex:25];
+                        [encoder setBytes:&ext_factor  length:sizeof(   float) atIndex:26];
+                        [encoder setBytes:&attn_factor length:sizeof(   float) atIndex:27];
+                        [encoder setBytes:&beta_fast   length:sizeof(   float) atIndex:28];
+                        [encoder setBytes:&beta_slow   length:sizeof(   float) atIndex:29];
 
                         [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                     } break;
@@ -2535,11 +2538,6 @@ static enum ggml_status ggml_metal_graph_compute(
                         GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
                                 "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
 
-                        const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
-                        const uint64_t nb21 = src2 ? src2->nb[1] : 0;
-                        const uint64_t nb22 = src2 ? src2->nb[2] : 0;
-                        const uint64_t nb23 = src2 ? src2->nb[3] : 0;
-
                         const int64_t  ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
                       //const int64_t  ne31 = src3 ? src3->ne[1] : 0;
                         const int64_t  ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32);

ggml-metal.metal

@@ -1640,6 +1640,7 @@ static void rope_yarn_corr_dims(
 typedef void (rope_t)(
         device const    void * src0,
         device const int32_t * src1,
+        device const   float * src2,
         device         float * dst,
         constant     int64_t & ne00,
         constant     int64_t & ne01,
@@ -1675,6 +1676,7 @@ template<typename T>
 kernel void kernel_rope(
         device const    void * src0,
         device const int32_t * src1,
+        device const   float * src2,
         device         float * dst,
         constant     int64_t & ne00,
         constant     int64_t & ne01,
@@ -1744,8 +1746,10 @@ kernel void kernel_rope(
 
                 // simplified from `(ib * n_dims + ic) * inv_ndims`
                 const float cur_rot = inv_ndims*ic - ib;
+                const float freq_factor = src2 != src0 ? src2[ic/2] : 1.0f;
+
+                const float theta = theta_0 * pow(freq_base, cur_rot) / freq_factor;
 
-                const float theta = theta_0 * pow(freq_base, cur_rot);
                 float cos_theta, sin_theta;
                 rope_yarn(theta, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
 
@@ -2204,11 +2208,7 @@ kernel void kernel_flash_attn_ext_f16(
         // pointer to the mask
         device const half * mp = (device const half *) (mask + iq1*nb31);
 
-        // prepare diagonal scale matrix
+        float slope = 1.0f;
-        simdgroup_float8x8 mscale(scale);
-
-        // prepare diagonal slope matrix
-        simdgroup_float8x8 mslope(1.0f);
 
         // ALiBi
         if (max_bias > 0.0f) {
@@ -2217,7 +2217,7 @@ kernel void kernel_flash_attn_ext_f16(
             const float base = h < n_head_log2 ? m0 : m1;
             const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
 
-            mslope = simdgroup_float8x8(pow(base, exph));
+            slope = pow(base, exph);
         }
 
         // loop over the KV cache
@@ -2242,18 +2242,20 @@ kernel void kernel_flash_attn_ext_f16(
                         simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk);
                     }
 
+                    simdgroup_store(mqk, ss + 8*cc, TF, 0, false);
+
+                    const short tx = tiisg%4;
+                    const short ty = tiisg/4;
+
                     if (mask != q) {
                         // mqk = mqk*scale + mask*slope
-                        simdgroup_half8x8 mm;
+                        ss[8*cc + ty*TF + 2*tx + 0] = scale*ss[8*cc + ty*TF + 2*tx + 0] + slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 0];
-                        simdgroup_load(mm, mp + ic + 8*cc, nb31/sizeof(half), 0, false);
+                        ss[8*cc + ty*TF + 2*tx + 1] = scale*ss[8*cc + ty*TF + 2*tx + 1] + slope*mp[ic + 8*cc + ty*nb31/sizeof(half) + 2*tx + 1];
-                        simdgroup_multiply(mm, mslope, mm);
-                        simdgroup_multiply_accumulate(mqk, mqk, mscale, mm);
                     } else {
                         // mqk = mqk*scale
-                        simdgroup_multiply(mqk, mscale, mqk);
+                        ss[8*cc + ty*TF + 2*tx + 0] *= scale;
+                        ss[8*cc + ty*TF + 2*tx + 1] *= scale;
                     }
-
-                    simdgroup_store(mqk, ss + 8*cc, TF, 0, false);
                 }
             }
 
@@ -2816,8 +2818,7 @@ kernel void kernel_cpy_f32_f16(
     for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
         device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
 
-        // TODO: is there a better way to handle -INFINITY?
+        dst_data[i00] = src[0];
-        dst_data[i00] = src[0] == -INFINITY ? -MAXHALF : src[0];
     }
 }
 
@@ -3385,7 +3386,6 @@ void kernel_mul_mv_q2_K_f32_impl(
 
     const int step = sizeof(block_q2_K) * nb;
 
-#if QK_K == 256
     const int ix = tiisg/8;  // 0...3
     const int it = tiisg%8;  // 0...7
     const int iq = it/4;     // 0 or 1
@@ -3437,57 +3437,6 @@ void kernel_mul_mv_q2_K_f32_impl(
 
         y4 += 4 * QK_K;
     }
-#else
-    const int ix = tiisg/2;  // 0...15
-    const int it = tiisg%2;  // 0...1
-
-    device const float * y4 = y + ix * QK_K + 8 * it;
-
-    for (int ib = ix; ib < nb; ib += 16) {
-
-        float4 sumy = {0.f, 0.f, 0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
-            yl[i+ 8] = y4[i+16]; sumy[1] += yl[i+ 8];
-            yl[i+16] = y4[i+32]; sumy[2] += yl[i+16];
-            yl[i+24] = y4[i+48]; sumy[3] += yl[i+24];
-        }
-
-        device const uint8_t  * sc = (device const uint8_t  *)x[ib].scales;
-        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
-        device const half     * dh = &x[ib].d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
-            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
-                acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
-                acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
-                acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
-                acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
-                acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
-                acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
-                acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
-            }
-
-            float dall = dh[0];
-            float dmin = dh[1];
-            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
-                                 (acc1[1] + 1.f/256.f * acc2[1]) * (sc[1] & 0xF) * 1.f/ 4.f +
-                                 (acc1[2] + 1.f/256.f * acc2[2]) * (sc[2] & 0xF) * 1.f/16.f +
-                                 (acc1[3] + 1.f/256.f * acc2[3]) * (sc[3] & 0xF) * 1.f/64.f) -
-                         dmin * (sumy[0] * (sc[0] >> 4) + sumy[1] * (sc[1] >> 4) + sumy[2] * (sc[2] >> 4) + sumy[3] * (sc[3] >> 4));
-
-            qs += step/2;
-            sc += step;
-            dh += step/2;
-        }
-
-        y4 += 16 * QK_K;
-    }
-#endif
 
     for (int row = 0; row < N_DST; ++row) {
         all_sum = simd_sum(sumf[row]);
@@ -3525,7 +3474,6 @@ kernel void kernel_mul_mv_q2_K_f32(
     kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
 }
 
-#if QK_K == 256
 void kernel_mul_mv_q3_K_f32_impl(
         device const  void * src0,
         device const float * src1,
@@ -3684,84 +3632,6 @@ void kernel_mul_mv_q3_K_f32_impl(
         }
     }
 }
-#else
-void kernel_mul_mv_q3_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    const int nb = ne00/QK_K;
-
-    const int64_t r0 = tgpig.x;
-    const int64_t r1 = tgpig.y;
-    const int64_t im = tgpig.z;
-
-    const int row = 2 * r0 + sgitg;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q3_K * x = (device const block_q3_K *) src0 + row*nb + offset0;
-    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    const int ix = tiisg/4;
-    const int il = 4 * (tiisg%4);// 0, 4, 8, 12
-    const int iq = il/8;         // 0, 0, 1, 1
-    const int in = il%8;         // 0, 4, 0, 4
-
-    float2 sum = {0.f, 0.f};
-
-    for (int i = ix; i < nb; i += 8) {
-
-        const float d_all = (float)(x[i].d);
-
-        device const uint16_t * q = (device const uint16_t *)(x[i].qs + il);
-        device const uint16_t * h = (device const uint16_t *)(x[i].hmask + in);
-        device const uint16_t * s = (device const uint16_t *)(x[i].scales);
-        device const float    * y = yy + i * QK_K + il;
-
-        const float d1 = d_all * ((int32_t)(s[0] & 0x000F) - 8);
-        const float d2 = d_all * ((int32_t)(s[0] & 0x00F0) - 128) * 1.f/64.f;
-        const float d3 = d_all * ((int32_t)(s[0] & 0x0F00) - 2048) * 1.f/4096.f;
-        const float d4 = d_all * ((int32_t)(s[0] & 0xF000) - 32768) * 1.f/262144.f;
-
-        for (int l = 0; l < 4; l += 2) {
-            const uint16_t hm = h[l/2] >> iq;
-            sum[0] += y[l+ 0] * d1 * ((int32_t)(q[l/2] & 0x0003) - ((hm & 0x0001) ? 0 :  4))
-                    + y[l+16] * d2 * ((int32_t)(q[l/2] & 0x000c) - ((hm & 0x0004) ? 0 : 16))
-                    + y[l+32] * d3 * ((int32_t)(q[l/2] & 0x0030) - ((hm & 0x0010) ? 0 : 64))
-                    + y[l+48] * d4 * ((int32_t)(q[l/2] & 0x00c0) - ((hm & 0x0040) ? 0 : 256));
-            sum[1] += y[l+ 1] * d1 * ((int32_t)(q[l/2] & 0x0300) - ((hm & 0x0100) ? 0 : 1024))
-                    + y[l+17] * d2 * ((int32_t)(q[l/2] & 0x0c00) - ((hm & 0x0400) ? 0 : 4096))
-                    + y[l+33] * d3 * ((int32_t)(q[l/2] & 0x3000) - ((hm & 0x1000) ? 0 : 16384))
-                    + y[l+49] * d4 * ((int32_t)(q[l/2] & 0xc000) - ((hm & 0x4000) ? 0 : 65536));
-        }
-
-    }
-    const float sumf = sum[0] + sum[1] * 1.f/256.f;
-
-    const float tot = simd_sum(sumf);
-    if (tiisg == 0) {
-        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
-    }
-
-}
-#endif
 
 [[host_name("kernel_mul_mv_q3_K_f32")]]
 kernel void kernel_mul_mv_q3_K_f32(
@@ -3791,7 +3661,6 @@ kernel void kernel_mul_mv_q3_K_f32(
     kernel_mul_mv_q3_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
 }
 
-#if QK_K == 256
 void kernel_mul_mv_q4_K_f32_impl(
         device const  void * src0,
         device const float * src1,
@@ -3905,103 +3774,6 @@ void kernel_mul_mv_q4_K_f32_impl(
         }
     }
 }
-#else
-void kernel_mul_mv_q4_K_f32_impl(
-        device const  void * src0,
-        device const float * src1,
-        device       float * dst,
-        constant   int64_t & ne00,
-        constant   int64_t & ne01,
-        constant   int64_t & ne02,
-        constant   int64_t & ne10,
-        constant   int64_t & ne12,
-        constant   int64_t & ne0,
-        constant   int64_t & ne1,
-        constant   uint    & r2,
-        constant   uint    & r3,
-        threadgroup int8_t * shared_values [[threadgroup(0)]],
-        uint3 tgpig[[threadgroup_position_in_grid]],
-        uint  tiisg[[thread_index_in_simdgroup]],
-        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
-
-    const int ix = tiisg/4;  // 0...7
-    const int it = tiisg%4;  // 0...3
-
-    const int nb = ne00/QK_K;
-    const int r0 = tgpig.x;
-    const int r1 = tgpig.y;
-    const int im = tgpig.z;
-    const int first_row = r0 * N_DST;
-    const int ib_row = first_row * nb;
-
-    const uint i12 = im%ne12;
-    const uint i13 = im/ne12;
-
-    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
-
-    device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
-    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
-
-    float yl[8];
-    float yh[8];
-    float sumf[N_DST]={0.f}, all_sum;
-
-    const int step = sizeof(block_q4_K) * nb / 2;
-
-    device const float * y4 = y + ix * QK_K + 8 * it;
-
-    uint16_t sc16[4];
-
-    for (int ib = ix; ib < nb; ib += 8) {
-
-        float2 sumy = {0.f, 0.f};
-        for (int i = 0; i < 8; ++i) {
-            yl[i] = y4[i+ 0]; sumy[0] += yl[i];
-            yh[i] = y4[i+32]; sumy[1] += yh[i];
-        }
-
-        device const uint16_t * sc = (device const uint16_t *)x[ib].scales;
-        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 4 * it;
-        device const half     * dh = x[ib].d;
-
-        for (int row = 0; row < N_DST; row++) {
-
-            sc16[0] = sc[0] & 0x000f;
-            sc16[1] = sc[0] & 0x0f00;
-            sc16[2] = sc[0] & 0x00f0;
-            sc16[3] = sc[0] & 0xf000;
-
-            float2 acc1 = {0.f, 0.f};
-            float2 acc2 = {0.f, 0.f};
-            for (int i = 0; i < 8; i += 2) {
-                acc1[0] += yl[i+0] * (qs[i/2] & 0x000F);
-                acc1[1] += yl[i+1] * (qs[i/2] & 0x0F00);
-                acc2[0] += yh[i+0] * (qs[i/2] & 0x00F0);
-                acc2[1] += yh[i+1] * (qs[i/2] & 0xF000);
-            }
-
-            float dall = dh[0];
-            float dmin = dh[1];
-            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc16[0] +
-                                 (acc2[0] + 1.f/256.f * acc2[1]) * sc16[1] * 1.f/4096.f) -
-                         dmin * 1.f/16.f * (sumy[0] * sc16[2] + sumy[1] * sc16[3] * 1.f/256.f);
-
-            qs += step;
-            sc += step;
-            dh += step;
-        }
-
-        y4 += 8 * QK_K;
-    }
-
-    for (int row = 0; row < N_DST; ++row) {
-        all_sum = simd_sum(sumf[row]);
-        if (tiisg == 0) {
-            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
-        }
-    }
-}
-#endif
 
 [[host_name("kernel_mul_mv_q4_K_f32")]]
 kernel void kernel_mul_mv_q4_K_f32(
@@ -4069,8 +3841,6 @@ void kernel_mul_mv_q5_K_f32_impl(
 
     const int step = sizeof(block_q5_K) * nb;
 
-#if QK_K == 256
-#
     float yl[16], yh[16];
 
     const uint16_t kmask1 = 0x3f3f;
@@ -4153,54 +3923,6 @@ void kernel_mul_mv_q5_K_f32_impl(
         y1 += 4 * QK_K;
 
     }
-#else
-    float yl[8], yh[8];
-
-    const int il = 4 * (tiisg/8);  // 0, 4, 8, 12
-    const int ix = tiisg%8;
-    const int iq = il/8;         // 0, 0, 1, 1
-    const int in = il%8;         // 0, 4, 0, 4
-
-    device const float * y = yy + ix*QK_K + il;
-
-    for (int i = ix; i < nb; i += 8) {
-
-        for (int l = 0; l < 4; ++l) {
-            yl[l+0] = y[l+ 0];
-            yl[l+4] = y[l+16];
-            yh[l+0] = y[l+32];
-            yh[l+4] = y[l+48];
-        }
-
-        device const half * dh = &x[i].d;
-        device const uint8_t * q = x[i].qs + il;
-        device const uint8_t * h = x[i].qh + in;
-        device const int8_t  * s = x[i].scales;
-
-        for (int row = 0; row < 2; ++row) {
-
-            const float d = dh[0];
-
-            float2 acc = {0.f, 0.f};
-            for (int l = 0; l < 4; ++l) {
-                const uint8_t hl = h[l] >> iq;
-                acc[0] += yl[l+0] * s[0] * ((int16_t)(q[l+ 0] & 0x0F) - (hl & 0x01 ? 0 : 16))
-                        + yl[l+4] * s[1] * ((int16_t)(q[l+16] & 0x0F) - (hl & 0x04 ? 0 : 16));
-                acc[1] += yh[l+0] * s[2] * ((int16_t)(q[l+ 0] & 0xF0) - (hl & 0x10 ? 0 : 256))
-                        + yh[l+4] * s[3] * ((int16_t)(q[l+16] & 0xF0) - (hl & 0x40 ? 0 : 256));
-            }
-            sumf[row] += d * (acc[0] + 1.f/16.f * acc[1]);
-
-            q += step;
-            h += step;
-            s += step;
-            dh += step/2;
-
-        }
-
-        y += 8 * QK_K;
-    }
-#endif
 
     for (int row = 0; row < 2; ++row) {
         const float tot = simd_sum(sumf[row]);
@@ -4279,7 +4001,6 @@ void kernel_mul_mv_q6_K_f32_impl(
 
     float sumf = 0;
 
-#if QK_K == 256
     const int tid  = tiisg/2;
     const int ix   = tiisg%2;
     const int ip   = tid/8;         // 0 or 1
@@ -4315,30 +4036,6 @@ void kernel_mul_mv_q6_K_f32_impl(
 
     }
 
-#else
-    const int ix  = tiisg/4;
-    const int il  = 4*(tiisg%4);
-
-    for (int i = ix; i < nb; i += 8) {
-        device const float * y = yy + i * QK_K + il;
-        device const uint8_t * ql = x[i].ql + il;
-        device const uint8_t * qh = x[i].qh + il;
-        device const int8_t  * s  = x[i].scales;
-
-        const float d = x[i].d;
-
-        float4 sums = {0.f, 0.f, 0.f, 0.f};
-        for (int l = 0; l < 4; ++l) {
-            sums[0] += y[l+ 0] * ((int8_t)((ql[l+ 0] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
-            sums[1] += y[l+16] * ((int8_t)((ql[l+16] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
-            sums[2] += y[l+32] * ((int8_t)((ql[l+ 0] >>  4) | ((qh[l] & kmask3) >> 0)) - 32);
-            sums[3] += y[l+48] * ((int8_t)((ql[l+16] >>  4) | ((qh[l] & kmask4) >> 2)) - 32);
-        }
-        sumf += d * (sums[0] * s[0] + sums[1] * s[1] + sums[2] * s[2] + sums[3] * s[3]);
-    }
-
-#endif
-
     const float tot = simd_sum(sumf);
     if (tiisg == 0) {
         dst[r1*ne0 + im*ne0*ne1 + row] = tot;
@@ -5172,9 +4869,7 @@ void kernel_mul_mv_iq1_m_f32_impl(
 
     device const float * y4 = y + 32 * ix;
 
-#if QK_K != 64
     iq1m_scale_t scale;
-#endif
 
     for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
 
@@ -5195,10 +4890,7 @@ void kernel_mul_mv_iq1_m_f32_impl(
         device const uint16_t * sc = (device const uint16_t *)xr->scales;
 
         for (int row = 0; row < N_DST; row++) {
-
-#if QK_K != 64
             scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
-#endif
 
             constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
             constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
@@ -5214,14 +4906,9 @@ void kernel_mul_mv_iq1_m_f32_impl(
             }
             const float delta1 = sumy[0] * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[1] * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
             const float delta2 = sumy[2] * (qh[1] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[3] * (qh[1] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
-#if QK_K == 64
+
-            const float d = (float) *((device const half *)(sc - 1));
-            sumf[row] += d * ((sum[0] + delta1) * (2*((sc[0] >> (8*(ib%2)+0)) & 0xf) + 1) +
-                              (sum[1] + delta2) * (2*((sc[0] >> (8*(ib%2)+4)) & 0xf) + 1));
-#else
             sumf[row] += (float)scale.f16 * ((sum[0] + delta1) * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 7) + 1) +
                                              (sum[1] + delta2) * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 7) + 1));
-#endif
 
             sc += nb*sizeof(block_iq1_m)/2;
             qs += nb*sizeof(block_iq1_m);
@@ -5333,7 +5020,6 @@ void kernel_mul_mv_iq4_nl_f32_impl(
     }
 }
 
-#if QK_K != 64
 void kernel_mul_mv_iq4_xs_f32_impl(
         device const  void * src0,
         device const float * src1,
@@ -5428,7 +5114,6 @@ void kernel_mul_mv_iq4_xs_f32_impl(
         }
     }
 }
-#endif
 
 [[host_name("kernel_mul_mv_iq1_s_f32")]]
 kernel void kernel_mul_mv_iq1_s_f32(
@@ -5541,11 +5226,7 @@ kernel void kernel_mul_mv_iq4_xs_f32(
         uint tiisg[[thread_index_in_simdgroup]],
         uint sgitg[[simdgroup_index_in_threadgroup]]) {
 
-#if QK_K == 64
-    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-#else
     kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
-#endif
 }
 
 //============================= templates and their specializations =============================
@@ -5671,10 +5352,9 @@ void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg
     float dl, ml;
     uint8_t sc = xb->scales[il];
 
-#if QK_K == 256
     q = q + 32*(il/8) + 16*(il&1);
     il = (il/2)%4;
-#endif
+
     half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
     uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
     dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
@@ -5690,7 +5370,6 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
     device const uint8_t * h = (device const uint8_t *)xb->hmask;
     device const int8_t * scales = (device const int8_t *)xb->scales;
 
-#if QK_K == 256
     q = q + 32 * (il/8) + 16 * (il&1);
     h = h + 16 * (il&1);
     uint8_t m = 1 << (il/2);
@@ -5711,17 +5390,6 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
     }
-#else
-    float    kcoef = il&1 ? 1.f/16.f : 1.f;
-    uint16_t kmask = il&1 ? 0xF0     : 0x0F;
-    float    dl = d_all * ((scales[il/2] & kmask) * kcoef - 8);
-    float    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
-    uint8_t  mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
-    uint8_t  m = 1<<(il*2);
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - ((h[i%8] & (m * (1 + i/8))) ? 0 : 4.f/coef));
-    }
-#endif
 }
 
 static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
@@ -5733,7 +5401,6 @@ template <typename type4x4>
 void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
     device const uchar * q = xb->qs;
 
-#if QK_K == 256
     short is = (il/4) * 2;
     q = q + (il/4) * 32 + 16 * (il&1);
     il = il & 3;
@@ -5742,16 +5409,7 @@ void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg
     const float min = xb->dmin;
     const float dl = d * sc[0];
     const float ml = min * sc[1];
-#else
-    (void) get_scale_min_k4_just2;
 
-    q = q + 16 * (il&1);
-    device const uint8_t * s = xb->scales;
-    device const half2 * dh = (device const half2 *)xb->d;
-    const float2 d = (float2)dh[0];
-    const float dl = il<2 ? d[0] * (s[0]&0xF) : d[0] * (s[1]&0xF)/16.h;
-    const float ml = il<2 ? d[1] * (s[0]>>4)  : d[1] * (s[1]>>4);
-#endif
     const ushort mask = il<2 ? 0x0F : 0xF0;
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * (q[i] & mask) - ml;
@@ -5763,7 +5421,6 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
     device const uint8_t * q  = xb->qs;
     device const uint8_t * qh = xb->qh;
 
-#if QK_K == 256
     short is = (il/4) * 2;
     q  = q + 32 * (il/4) + 16 * (il&1);
     qh = qh + 16 * (il&1);
@@ -5780,17 +5437,6 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
     for (int i = 0; i < 16; ++i) {
         reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
     }
-#else
-    q = q + 16 * (il&1);
-    device const int8_t * s = xb->scales;
-    const float dl = xb->d * s[il];
-    uint8_t m = 1<<(il*2);
-    const float  coef = il<2 ? 1.f  : 1.f/16.f;
-    const ushort mask = il<2 ? 0x0F : 0xF0;
-    for (int i = 0; i < 16; ++i) {
-        reg[i/4][i%4] = coef * dl * ((q[i] & mask) - (qh[i%8] & (m*(1+i/8)) ? 0.f : 16.f/coef));
-    }
-#endif
 }
 
 template <typename type4x4>
@@ -5800,15 +5446,11 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
     device const uint8_t * qh = (device const uint8_t *)xb->qh;
     device const int8_t * scales = (device const int8_t *)xb->scales;
 
-#if QK_K == 256
     ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
     qh = qh + 32*(il/8) + 16*(il&1);
     float sc = scales[(il%2) + 2 * ((il/2))];
     il = (il/2) & 3;
-#else
+
-    ql = ql + 16 * (il&1);
-    float sc = scales[il];
-#endif
     const uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
     const uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
     const float       coef = il>1 ? 1.f/16.f          : 1.f;
@@ -5965,20 +5607,15 @@ void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 &
     const int ib32 = il/2;
     il = il%2;
     device const uint16_t * sc = (device const uint16_t *)xb->scales;
-#if QK_K == 64
+
-    const float d = xb->d;
-#else
     iq1m_scale_t scale;
     scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
     const float d = scale.f16;
-#endif
+
     device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
     device const uint8_t * qh = xb->qh + 2*ib32 + il;
-#if QK_K == 64
+
-    const float dl  = d * (2*((sc[ib32/2] >> (8*(ib32%2)+4*il)) & 0xf) + 1);
-#else
     const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
-#endif
     const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
     const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
     constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
@@ -6008,9 +5645,6 @@ void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4
 
 template <typename type4x4>
 void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
-#if QK_K == 64
-    dequantize_iq4_nl(xb, il, reg);
-#else
     // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
     const int ib32 = il/2;
     il = il%2;
@@ -6027,7 +5661,6 @@ void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4
         reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
         reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
     }
-#endif
 }
 
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
@@ -6532,11 +6165,7 @@ kernel void kernel_mul_mm_id(
         sgitg);
 }
 
-#if QK_K == 256
 #define QK_NL 16
-#else
-#define QK_NL 4
-#endif
 
 //
 // get rows
@@ -6576,11 +6205,7 @@ template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_t kernel_get_r
 template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_t kernel_get_rows<block_iq1_s,   QK_NL, dequantize_iq1_s>;
 template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_t kernel_get_rows<block_iq1_m,   QK_NL, dequantize_iq1_m>;
 template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_t kernel_get_rows<block_iq4_nl,  2,     dequantize_iq4_nl>;
-#if QK_K == 64
-template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_t kernel_get_rows<block_iq4_xs,  2,     dequantize_iq4_xs>;
-#else
 template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_t kernel_get_rows<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
-#endif
 
 //
 // matrix-matrix multiplication
@@ -6608,11 +6233,7 @@ template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mat_mm_t kernel_mul_m
 template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq1_s,   QK_NL, dequantize_iq1_s>;
 template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mat_mm_t kernel_mul_mm<block_iq1_m,   QK_NL, dequantize_iq1_m>;
 template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_nl,  2,     dequantize_iq4_nl>;
-#if QK_K == 64
-template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_nl,  2,     dequantize_iq4_xs>;
-#else
 template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
-#endif
 
 //
 // indirect matrix-matrix multiplication
@@ -6640,11 +6261,7 @@ template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mat_mm_id_t kernel
 template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s,   QK_NL, dequantize_iq1_s>;
 template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_m,   QK_NL, dequantize_iq1_m>;
 template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl,  2,     dequantize_iq4_nl>;
-#if QK_K == 64
-template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs,  2,     dequantize_iq4_xs>;
-#else
 template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
-#endif
 
 //
 // matrix-vector multiplication
@@ -6853,7 +6470,5 @@ template [[host_name("kernel_mul_mv_id_iq3_xxs_f32")]] kernel kernel_mul_mv_id_t
 template [[host_name("kernel_mul_mv_id_iq3_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_s_f32_impl>>;
 template [[host_name("kernel_mul_mv_id_iq2_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_s_f32_impl>>;
 template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_nl_f32_impl>>;
-#if QK_K != 64
 template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_xs_f32_impl>>;
-#endif
 

ggml-opencl.cpp

@@ -1,4 +1,4 @@
-#include "ggml.h"
+#include "ggml.h"
 #include "ggml-opencl.h"
 #include "ggml-backend-impl.h"
 

ggml-sycl.cpp

@@ -4197,7 +4197,6 @@ static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restri
     const block_q2_K * x = (const block_q2_K *) vx;
 
     const int tid = item_ct1.get_local_id(2);
-#if QK_K == 256
     const int n   = tid/32;
     const int l   = tid - 32*n;
     const int is  = 8*n + l/16;
@@ -4211,18 +4210,6 @@ static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restri
     y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
     y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
     y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
-#else
-    const int is = tid/16;  // 0 or 1
-    const int il = tid%16;  // 0...15
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    dst_t * y = yy + i*QK_K + 16*is + il;
-
-    float dall = x[i].dm[0];
-    float dmin = x[i].dm[1];
-    y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
-    y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
-#endif
-
 }
 
 template<typename dst_t>
@@ -4232,7 +4219,6 @@ static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restri
     const int i = item_ct1.get_group(2);
     const block_q3_K * x = (const block_q3_K *) vx;
 
-#if QK_K == 256
     const int r = item_ct1.get_local_id(2) / 4;
     const int tid = r/2;
     const int is0 = r%2;
@@ -4256,31 +4242,8 @@ static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restri
     const uint8_t * hm = x[i].hmask;
 
     for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
-#else
-    const int tid = item_ct1.get_local_id(2);
-    const int is  = tid/16;  // 0 or 1
-    const int il  = tid%16;  // 0...15
-    const int im  = il/8;    // 0...1
-    const int in  = il%8;    // 0...7
-
-    dst_t * y = yy + i*QK_K + 16*is + il;
-
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    const uint8_t h = x[i].hmask[in] >> (2*is + im);
-    const float   d = (float)x[i].d;
-
-    if (is == 0) {
-        y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    } else {
-        y[ 0] = d * ((x[i].scales[0] >>  4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] >>  4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    }
-#endif
-
 }
 
-#if QK_K == 256
 static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
     if (j < 4) {
         d = q[j] & 63; m = q[j + 4] & 63;
@@ -4289,7 +4252,6 @@ static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8
         m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
     }
 }
-#endif
 
 template<typename dst_t>
 static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
@@ -4298,7 +4260,6 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri
 
     const int i = item_ct1.get_group(2);
 
-#if QK_K == 256
     // assume 32 threads
     const int tid = item_ct1.get_local_id(2);
     const int il  = tid/8;
@@ -4322,15 +4283,6 @@ static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restri
         y[l + 0] = d1 * (q[l] & 0xF) - m1;
         y[l +32] = d2 * (q[l] >>  4) - m2;
     }
-#else
-    const int tid = item_ct1.get_local_id(2);
-    const uint8_t * q = x[i].qs;
-    dst_t * y = yy + i*QK_K;
-    const float d = (float)x[i].dm[0];
-    const float m = (float)x[i].dm[1];
-    y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
-    y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
-#endif
 }
 
 template<typename dst_t>
@@ -4340,7 +4292,6 @@ static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restri
 
     const int i = item_ct1.get_group(2);
 
-#if QK_K == 256
     // assume 64 threads - this is very slightly better than the one below
     const int tid = item_ct1.get_local_id(2);
     const int il  = tid/16;   // il is in 0...3
@@ -4367,18 +4318,6 @@ static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restri
     hm <<= 1;
     y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
     y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
-#else
-    const int tid = item_ct1.get_local_id(2);
-    const uint8_t q = x[i].qs[tid];
-    const int im = tid/8;  // 0...3
-    const int in = tid%8;  // 0...7
-    const int is = tid/16; // 0 or 1
-    const uint8_t h = x[i].qh[in] >> im;
-    const float d = x[i].d;
-    dst_t * y = yy + i*QK_K + tid;
-    y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
-    y[32] = d * x[i].scales[is+2] * ((q >>  4) - ((h >> 4) & 1 ? 0 : 16));
-#endif
 }
 
 template<typename dst_t>
@@ -4387,7 +4326,6 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri
     const block_q6_K * x = (const block_q6_K *) vx;
 
     const int i = item_ct1.get_group(2);
-#if QK_K == 256
 
     // assume 64 threads - this is very slightly better than the one below
     const int tid = item_ct1.get_local_id(2);
@@ -4407,24 +4345,6 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri
     y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
     y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
     y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
-#else
-
-    // assume 32 threads
-    const int tid = item_ct1.get_local_id(2);
-    const int ip  = tid/16;         // 0 or 1
-    const int il  = tid - 16*ip;    // 0...15
-
-    dst_t * y = yy + i*QK_K + 16*ip + il;
-
-    const float d = x[i].d;
-
-    const uint8_t   ql = x[i].ql[16*ip + il];
-    const uint8_t   qh = x[i].qh[il] >> (2*ip);
-    const int8_t  * sc = x[i].scales;
-
-    y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
-    y[32] = d * sc[ip+2] * ((int8_t)((ql  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
-#endif
 }
 
 template<typename dst_t>
@@ -4438,7 +4358,6 @@ static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __res
     const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
 
     const int tid = item_ct1.get_local_id(2);
-#if QK_K == 256
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -4449,10 +4368,6 @@ static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __res
     const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs_ptr[(aux32 >> 7*il) & 127];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs_ptr[j] ? -1.f : 1.f);
-#else
-    assert(false);
-#endif
-
 }
 
 template<typename dst_t>
@@ -4466,7 +4381,6 @@ static void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __rest
     const block_iq2_xs * x = (const block_iq2_xs *) vx;
 
     const int tid = item_ct1.get_local_id(2);
-#if QK_K == 256
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -4475,10 +4389,6 @@ static void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __rest
     const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
     for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    assert(false);
-#endif
-
 }
 
 template <typename dst_t>
@@ -4490,7 +4400,6 @@ dequantize_block_iq2_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
     const block_iq2_s * x = (const block_iq2_s *) vx;
 
     const int tid = item_ct1.get_local_id(2);
-#if QK_K == 256
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -4498,13 +4407,9 @@ dequantize_block_iq2_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
     const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
     const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
 #pragma unroll
-    for (int j = 0; j < 8; ++j)
+    for (int j = 0; j < 8; ++j) {
         y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
+    }
-    assert(false);
-
-#endif
-
 }
 
 template<typename dst_t>
@@ -4518,7 +4423,6 @@ static void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __res
     const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
 
     const int tid = item_ct1.get_local_id(2);
-#if QK_K == 256
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -4533,10 +4437,6 @@ static void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __res
         y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
         y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
-#else
-    assert(false);
-#endif
-
 }
 
 template <typename dst_t>
@@ -4549,7 +4449,6 @@ dequantize_block_iq3_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
     const block_iq3_s * x = (const block_iq3_s *) vx;
 
     const int tid = item_ct1.get_local_id(2);
-#if QK_K == 256
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -4563,10 +4462,6 @@ dequantize_block_iq3_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
         y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
         y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
     }
-#else
-    assert(false);
-#endif
-
 }
 
 template <typename dst_t>
@@ -4579,7 +4474,6 @@ dequantize_block_iq1_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
     const block_iq1_s * x = (const block_iq1_s  *) vx;
 
     const int tid = item_ct1.get_local_id(2);
-#if QK_K == 256
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -4593,10 +4487,6 @@ dequantize_block_iq1_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
     for (int j = 0; j < 8; ++j) {
         y[j] = d * (q[j] + delta);
     }
-#else
-    assert(false);
-#endif
-
 }
 
 template <typename dst_t>
@@ -4609,7 +4499,6 @@ dequantize_block_iq1_m(const void *__restrict__ vx, dst_t *__restrict__ yy,
     const block_iq1_m * x = (const block_iq1_m  *) vx;
 
     const int tid = item_ct1.get_local_id(2);
-#if QK_K == 256
     const int il = tid/8; // 0...3
     const int ib = tid%8; // 0...7
     dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -4627,10 +4516,6 @@ dequantize_block_iq1_m(const void *__restrict__ vx, dst_t *__restrict__ yy,
     for (int j = 0; j < 8; ++j) {
         y[j] = d * (q[j] + delta);
     }
-#else
-    assert(false);
-#endif
-
 }
 
 template <typename dst_t>
@@ -4704,7 +4589,6 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
     const int tid =
         item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...15
     const int ix =
@@ -4755,42 +4639,6 @@ static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx,
         tmp += dall * sum1 - dmin * sum2;
 
     }
-#else
-    const int tid = item_ct1.get_local_id(2) /
-                    (2 * K_QUANTS_PER_ITERATION); // 0...15 or 0...7
-    const int ix = item_ct1.get_local_id(2) %
-                   (2 * K_QUANTS_PER_ITERATION); // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;
-
-    uint32_t uaux[2];
-    const uint8_t * d = (const uint8_t *)uaux;
-
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint32_t * s = (const uint32_t *)x[i].scales;
-
-        uaux[0] = s[0] & 0x0f0f0f0f;
-        uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;
-
-        const sycl::float2 dall =
-            x[i].dm.convert<float, sycl::rounding_mode::automatic>();
-
-        float sum1 = 0, sum2 = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t ql = q[l];
-            sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
-                  + y[l+16] * d[1] * ((ql >> 2) & 3)
-                  + y[l+32] * d[2] * ((ql >> 4) & 3)
-                  + y[l+48] * d[3] * ((ql >> 6) & 3);
-            sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
-        }
-        tmp += dall.x() * sum1 - dall.y() * sum2;
-    }
-
-#endif
 
     // sum up partial sums and write back result
 #pragma unroll
@@ -4828,8 +4676,6 @@ static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
-
     const uint16_t kmask1 = 0x0303;
     const uint16_t kmask2 = 0x0f0f;
 
@@ -4882,34 +4728,6 @@ static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx,
         tmp += d * sum;
 
     }
-#else
-
-    const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
-    const int ix  = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;         // 0...15 or 0...14
-    const int in = offset/8;                                 // 0 or 1
-    const int im = offset%8;                                 // 0...7
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint8_t * s = x[i].scales;
-
-        const float dall = (float)x[i].d;
-
-        float sum = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t hl = x[i].hmask[im+l] >> in;
-            const uint8_t ql = q[l];
-            sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
-                 + y[l+16] * dall * ((s[0] >>  4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
-                 + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
-                 + y[l+48] * dall * ((s[1] >>  4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
-        }
-        tmp += sum;
-    }
-#endif
 
     // sum up partial sums and write back result
 #pragma unroll
@@ -4944,7 +4762,6 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx,
 
     const block_q4_K * x = (const block_q4_K *)vx + ib0;
 
-#if QK_K == 256
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
     const uint16_t kmask3 = 0xc0c0;
@@ -5033,36 +4850,6 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx,
 #endif
 
     }
-#else
-    const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    uint16_t aux16[2];
-    const uint8_t * s = (const uint8_t *)aux16;
-
-    float tmp = 0;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const float   * y = yy + i*QK_K + step;
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux16[0] = a[0] & 0x0f0f;
-        aux16[1] = (a[0] >> 4) & 0x0f0f;
-        const float d = (float)x[i].dm[0];
-        const float m = (float)x[i].dm[1];
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
-                 + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
-                 + y[j+32] * (d * s[1] * (q[j+ 0] >>  4) - m * s[3])
-                 + y[j+48] * (d * s[1] * (q[j+16] >>  4) - m * s[3]);
-        }
-        tmp += sum;
-    }
-
-#endif
 
     // sum up partial sums and write back result
 #pragma unroll
@@ -5097,7 +4884,6 @@ static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx,
 
     float tmp = 0; // partial sum for thread in warp
 
-#if QK_K == 256
     const uint16_t kmask1 = 0x3f3f;
     const uint16_t kmask2 = 0x0f0f;
     const uint16_t kmask3 = 0xc0c0;
@@ -5174,30 +4960,6 @@ static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx,
                dmin * smin;
     }
 
-#else
-    const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);
-    const int step = tid * K_QUANTS_PER_ITERATION;
-    const int im = step/8;
-    const int in = step%8;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const int8_t  * s = x[i].scales;
-        const float   * y = yy + i*QK_K + step;
-        const float     d = x[i].d;
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            const uint8_t h = x[i].qh[in+j] >> im;
-            sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
-                 + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
-                 + y[j+32] * d * s[2] * ((q[j+ 0] >>  4) - ((h >> 4) & 1 ? 0 : 16))
-                 + y[j+48] * d * s[3] * ((q[j+16] >>  4) - ((h >> 6) & 1 ? 0 : 16));
-        }
-        tmp += sum;
-    }
-#endif
-
     // sum up partial sums and write back result
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
@@ -5224,8 +4986,6 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa
 
     const block_q6_K * x = (const block_q6_K *)vx + ib0;
 
-#if QK_K == 256
-
     const int tid =
         item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16
     const int ix =
@@ -5282,37 +5042,6 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa
 
     }
 
-#else
-
-    const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION);  // 0...7
-    const int ix  = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);  // 0...3
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + step;
-        const uint8_t * ql = x[i].ql + step;
-        const uint8_t * qh = x[i].qh + step;
-        const int8_t  * s  = x[i].scales;
-
-        const float d = x[i+0].d;
-
-        float sum = 0;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
-                 + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
-                 + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >>  4) | ((qh[j] & 0x30) >> 0)) - 32)
-                 + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >>  4) | ((qh[j] & 0xc0) >> 2)) - 32);
-        }
-        tmp += sum;
-
-    }
-
-#endif
-
     // sum up partial sums and write back result
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
@@ -6857,7 +6586,6 @@ static __dpct_inline__ float
 vec_dot_q4_K_q8_1(const void *__restrict__ vbq,
                   const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
 
-#ifndef GGML_QKK_64
     const block_q4_K * bq4_K = (const block_q4_K *) vbq;
 
     int    v[2];
@@ -6899,52 +6627,6 @@ vec_dot_q4_K_q8_1(const void *__restrict__ vbq,
     }
 
     return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
-
-#else
-
-#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics
-    const block_q4_K * bq4_K = (const block_q4_K *) vbq;
-
-    float sumf_d = 0.0f;
-    float sumf_m = 0.0f;
-
-    uint16_t aux16[2];
-    const uint8_t * s = (const uint8_t *)aux16;
-
-    const uint16_t * a = (const uint16_t *)bq4_K->scales;
-    aux16[0] = a[0] & 0x0f0f;
-    aux16[1] = (a[0] >> 4) & 0x0f0f;
-
-    const float dall = bq4_K->dm[0];
-    const float dmin = bq4_K->dm[1];
-
-    const float d8_1 = bq8_1[0].ds[0];
-    const float d8_2 = bq8_1[1].ds[1];
-
-    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
-    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
-    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
-    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
-
-    const int * q4 = (const int *)bq4_K->qs + (iqs/2);
-    const int v1 = q4[0];
-    const int v2 = q4[4];
-
-    const int dot1 = dpct::dp4a(ui2, v2 & 0x0f0f0f0f, dpct::dp4a(ui1, v1 & 0x0f0f0f0f, 0));
-    const int dot2 = dpct::dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, dpct::dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0));
-    const int dot3 = dpct::dp4a(0x01010101, ui2, dpct::dp4a(0x01010101, ui1, 0));
-    const int dot4 = dpct::dp4a(0x01010101, ui4, dpct::dp4a(0x01010101, ui3, 0));
-
-    sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]);
-    sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]);
-
-    return dall * sumf_d - dmin * sumf_m;
-
-#else
-    bad_arch();
-#endif // __SYCL_ARCH__ >= VER_4VEC
-
-#endif
 }
 
 template <int mmq_y>
@@ -7003,11 +6685,7 @@ load_tiles_q4_K(const void *__restrict__ vx, int *__restrict__ x_ql,
 
         const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd;
 
-#if QK_K == 256
         x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm;
-#else
-        x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]};
-#endif
     }
 
 #pragma unroll
@@ -7050,7 +6728,6 @@ static __dpct_inline__ float
 vec_dot_q5_K_q8_1(const void *__restrict__ vbq,
                   const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
 
-#ifndef GGML_QKK_64
     const block_q5_K * bq5_K = (const block_q5_K *) vbq;
 
     int   vl[2];
@@ -7092,48 +6769,6 @@ vec_dot_q5_K_q8_1(const void *__restrict__ vbq,
     }
 
     return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
-
-#else
-
-#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics
-    const block_q5_K * bq5_K = (const block_q5_K *) vbq;
-
-    const int8_t * s = bq5_K->scales;
-
-    const float d = bq5_K->d;
-
-    const float d8_1 = bq8_1[0].ds[0];
-    const float d8_2 = bq8_1[1].ds[1];
-
-    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
-    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
-    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
-    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
-
-    const int * ql = (const int *)bq5_K->qs + (iqs/2);
-    const int vl1 = ql[0];
-    const int vl2 = ql[4];
-
-    const int step = 4 * (iqs/2); // 0, 4, 8, 12
-    const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6
-    const int in = step%8; // 0, 4, 0, 4
-    const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
-
-    const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f);
-    const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f);
-    const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f);
-    const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f);
-
-    const float sumf_d = d8_1 * (dpct::dp4a(ui1, v1, 0) * s[0] + dpct::dp4a(ui2, v2, 0) * s[1])
-                       + d8_2 * (dpct::dp4a(ui3, v3, 0) * s[2] + dpct::dp4a(ui4, v4, 0) * s[3]);
-
-    return d * sumf_d;
-
-#else
-    bad_arch();
-#endif // __SYCL_ARCH__ >= VER_4VEC
-
-#endif
 }
 
 template <int mmq_y>
@@ -7205,9 +6840,7 @@ load_tiles_q5_K(const void *__restrict__ vx, int *__restrict__ x_ql,
 
         const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd;
 
-#if QK_K == 256
         x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm;
-#endif
     }
 
 #pragma unroll
@@ -7387,7 +7020,6 @@ vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq,
                      const block_q8_1 *__restrict__ bq8_1, const int &iqs,
                      const uint64_t *iq2xxs_grid, const uint8_t *ksigns_iq2xs,
                      const uint8_t *kmask_iq2xs) {
-#if QK_K == 256
     const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq;
 
 #if QR2_XXS == 8
@@ -7428,10 +7060,6 @@ vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq,
     }
     return d * (sumi1 + sumi2);
 #endif
-#else
-    assert(false);
-    return 0.f;
-#endif
 }
 
 static __dpct_inline__ float
@@ -7440,7 +7068,6 @@ vec_dot_iq2_xs_q8_1(const void *__restrict__ vbq,
                     const uint64_t *iq2xs_grid, const uint64_t *ksigns64) {
 #if DPCT_COMPATIBILITY_TEMP >=                                                 \
     MIN_CC_DP4A // lowest compute capability for integer intrinsics
-#if QK_K == 256
     const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq;
 
     const int ib32 = iqs;
@@ -7478,16 +7105,11 @@ vec_dot_iq2_xs_q8_1(const void *__restrict__ vbq,
     assert(false);
     return 0.f;
 #endif
-#else
-    assert(false);
-    return 0.f;
-#endif
 }
 
 static __dpct_inline__ float
 vec_dot_iq2_s_q8_1(const void *__restrict__ vbq,
                    const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
-#if QK_K == 256
     const block_iq2_s * bq2 = (const block_iq2_s *) vbq;
 
     const int ib32 = iqs;
@@ -7531,9 +7153,6 @@ vec_dot_iq2_s_q8_1(const void *__restrict__ vbq,
     }
     const float d = (float)bq2->d * bq8_1[ib32].ds[0] * 0.25f;
     return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
-#else
-    assert(false);
-#endif
 }
 
 static __dpct_inline__ float
@@ -7542,7 +7161,6 @@ vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq,
                      const uint32_t *iq3xxs_grid, const uint64_t *ksigns64) {
 #if DPCT_COMPATIBILITY_TEMP >=                                                 \
     MIN_CC_DP4A // lowest compute capability for integer intrinsics
-#if QK_K == 256
     const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq;
 
     const int ib32 = iqs;
@@ -7570,17 +7188,12 @@ vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq,
     assert(false);
     return 0.f;
 #endif
-#else
-    assert(false);
-    return 0.f;
-#endif
 }
 
 static __dpct_inline__ float
 vec_dot_iq3_s_q8_1(const void *__restrict__ vbq,
                    const block_q8_1 *__restrict__ bq8_1, const int &iqs,
                    const uint32_t *iq3s_grid) {
-#if QK_K == 256
     const block_iq3_s * bq2 = (const block_iq3_s *) vbq;
 
     const int ib32 = iqs;
@@ -7609,16 +7222,12 @@ vec_dot_iq3_s_q8_1(const void *__restrict__ vbq,
         (1 + 2 * ((bq2->scales[ib32 / 2] >> 4 * (ib32 % 2)) & 0xf)) *
         bq8_1[ib32].ds[0];
     return d * sumi;
-#else
-    assert(false);
-#endif
 }
 
 static __dpct_inline__ float
 vec_dot_iq1_s_q8_1(const void *__restrict__ vbq,
                    const block_q8_1 *__restrict__ bq8_1, const int &iqs,
                    const uint32_t *iq1s_grid_gpu) {
-#if QK_K == 256
     const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
 
     const int ib32 = iqs;
@@ -7637,15 +7246,11 @@ vec_dot_iq1_s_q8_1(const void *__restrict__ vbq,
     const float d = d1q * bq8_1[ib32].ds[0];
     const float m = d1q * bq8_1[ib32].ds[1];
     return d * sumi + m * delta;
-#else
-    assert(false);
-#endif
 }
 
 static __dpct_inline__ float
 vec_dot_iq1_m_q8_1(const void *__restrict__ vbq,
                    const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
-#if QK_K == 256
     const block_iq1_m * bq1 = (const block_iq1_m *) vbq;
 
     const int ib32 = iqs;
@@ -7670,9 +7275,6 @@ vec_dot_iq1_m_q8_1(const void *__restrict__ vbq,
     scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
     const float d = (float)scale.f16 * bq8_1[ib32].ds[0];
     return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1));
-#else
-    assert(false);
-#endif
 }
 
 static __dpct_inline__ void get_int_from_table_16(const uint32_t &q4,
@@ -7720,7 +7322,6 @@ static __dpct_inline__ float
 vec_dot_iq4_xs_q8_1(const void *__restrict__ vbq,
                     const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
 
-#if QK_K == 256
     const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
     const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
 
@@ -7738,9 +7339,6 @@ vec_dot_iq4_xs_q8_1(const void *__restrict__ vbq,
         sumi2 = dpct::dp4a(v2, q8[j + 4], sumi2);
     }
     return d * (sumi1 + sumi2);
-#else
-    assert(false);
-#endif
 }
 
 template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x,
@@ -10203,7 +9801,6 @@ template <typename dst_t>
 static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int k,
                                      dpct::queue_ptr stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
@@ -10215,27 +9812,12 @@ static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int k,
                                  dequantize_block_q2_K(vx, y, item_ct1);
                              });
     }
-#else
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
-
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q2_K(vx, y, item_ct1);
-                             });
-    }
-
-#endif
 }
 
 template <typename dst_t>
 static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int k,
                                      dpct::queue_ptr stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
@@ -10247,19 +9829,6 @@ static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int k,
                                  dequantize_block_q3_K(vx, y, item_ct1);
                              });
     }
-#else
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
-
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q3_K(vx, y, item_ct1);
-                             });
-    }
-#endif
 }
 
 template <typename dst_t>
@@ -10320,7 +9889,6 @@ template <typename dst_t>
 static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int k,
                                      dpct::queue_ptr stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
@@ -10332,27 +9900,12 @@ static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int k,
                                  dequantize_block_q5_K(vx, y, item_ct1);
                              });
     }
-#else
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
-
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q5_K(vx, y, item_ct1);
-                             });
-    }
-
-#endif
 }
 
 template <typename dst_t>
 static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int k,
                                      dpct::queue_ptr stream) {
     const int nb = k / QK_K;
-#if QK_K == 256
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
@@ -10364,20 +9917,6 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int k,
                                  dequantize_block_q6_K(vx, y, item_ct1);
                              });
     }
-#else
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
-
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q6_K(vx, y, item_ct1);
-                             });
-    }
-
-#endif
 }
 
 template <typename dst_t>
@@ -10529,9 +10068,6 @@ template <typename dst_t>
 static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int k,
                                        dpct::queue_ptr stream) {
     const int nb = (k + QK_K - 1) / QK_K;
-#if QK_K == 64
-    dequantize_row_iq4_nl_sycl(vx, y, k, stream);
-#else
       {
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
@@ -10546,7 +10082,6 @@ static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int k,
                       });
             });
       }
-#endif
 }
 
 
@@ -12051,8 +11586,6 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
                                         const int nrows_y, const int nrows_dst,
                                         dpct::queue_ptr stream) try {
 
-#if QK_K == 256
-
     int id;
     SYCL_CHECK(
         CHECK_TRY_ERROR(id = get_current_device_id()));
@@ -12167,7 +11700,6 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
             });
         }
     }
-#endif
 }
 catch (sycl::exception const &exc) {
   std::cerr << exc.what() << "Exception caught at file:" << __FILE__
@@ -14454,6 +13986,9 @@ inline void ggml_sycl_op_rope(const ggml_tensor *src0, const ggml_tensor *src1,
                               ggml_tensor *dst, const float *src0_dd,
                               const float *src1_dd, float *dst_dd,
                               const dpct::queue_ptr &main_stream) {
+#pragma message("TODO: implement phi3 frequency factors support")
+#pragma message("      https://github.com/ggerganov/llama.cpp/pull/7225")
+    GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
 
     GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
     GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);

ggml-vulkan.cpp

@@ -290,6 +290,7 @@ struct vk_op_rope_neox_push_constants {
     float corr_dims[4];
     float theta_scale;
     float inv_ndims;
+    uint32_t has_freq_facs;
 };
 
 struct vk_op_soft_max_push_constants {
@@ -1522,8 +1523,8 @@ static void ggml_vk_load_shaders(ggml_backend_vk_context * ctx) {
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_f32, "rope_f32", rope_f32_len, rope_f32_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_f16, "rope_f16", rope_f16_len, rope_f16_data, "main", 3, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 3, sizeof(vk_op_rope_neox_push_constants), {1, 512, 1}, {}, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 4, sizeof(vk_op_rope_neox_push_constants), {1, 512, 1}, {}, 1);
-    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 3, sizeof(vk_op_rope_neox_push_constants), {1, 512, 1}, {}, 1);
+    ggml_vk_create_pipeline(ctx, ctx->device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_neox_push_constants), {1, 512, 1}, {}, 1);
 
     ggml_vk_create_pipeline(ctx, ctx->device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1);
 }
@@ -3732,7 +3733,7 @@ static void ggml_vk_op_repeat(ggml_backend_vk_context * ctx, vk_context * subctx
 }
 
 
-static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_op op) {
+static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op) {
     switch (op) {
     case GGML_OP_ADD:
         if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
@@ -3853,6 +3854,8 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
     default:
         return nullptr;
     }
+
+    GGML_UNUSED(src2);
 }
 
 static ggml_vk_func_t ggml_vk_op_get_func(ggml_op op) {
@@ -3880,12 +3883,15 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
 }
 
 template<typename PC>
-static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_op op, const PC&& pc) {
+static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, const PC&& pc) {
 #ifdef GGML_VULKAN_DEBUG
     std::cerr << "ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     if (src1 != nullptr) {
         std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
     }
+    if (src2 != nullptr) {
+        std::cerr << "), (" << src2 << ", name=" << src2->name << ", type=" << src2->type << ", ne0=" << src2->ne[0] << ", ne1=" << src2->ne[1] << ", ne2=" << src2->ne[2] << ", ne3=" << src2->ne[3] << ", nb0=" << src2->nb[0] << ", nb1=" << src2->nb[1] << ", nb2=" << src2->nb[2] << ", nb3=" << src2->nb[3];
+    }
     std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "), " << ggml_op_name(op) << ")" << std::endl;
 #endif
     GGML_ASSERT(op == GGML_OP_GET_ROWS || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
@@ -3896,6 +3902,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     const uint64_t ne02 = src0->ne[2];
     const uint64_t ne03 = src0->ne[3];
     const uint64_t ne0 = ne00 * ne01;
+
     const bool use_src1 = src1 != nullptr;
     const uint64_t ne10 = use_src1 ? src1->ne[0] : 0;
     const uint64_t ne11 = use_src1 ? src1->ne[1] : 0;
@@ -3904,7 +3911,14 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     const uint64_t ne1 = ne10 * ne11;
     // const uint64_t nb10 = use_src1 ? src1->nb[0] : 0;
 
-    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, dst, op);
+    const bool use_src2 = src2 != nullptr;
+    const uint64_t ne20 = use_src2 ? src2->ne[0] : 0;
+    const uint64_t ne21 = use_src2 ? src2->ne[1] : 0;
+    const uint64_t ne22 = use_src2 ? src2->ne[2] : 0;
+    const uint64_t ne23 = use_src2 ? src2->ne[3] : 0;
+    const uint64_t ne2 = ne20 * ne21;
+
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, dst, op);
     ggml_vk_func_t op_func;
 
     if (pipeline == nullptr) {
@@ -3927,15 +3941,18 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) dst->extra;
     ggml_tensor_extra_gpu * extra_src0 = (ggml_tensor_extra_gpu *) src0->extra;
     ggml_tensor_extra_gpu * extra_src1 = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
+    ggml_tensor_extra_gpu * extra_src2 = use_src2 ? (ggml_tensor_extra_gpu *) src2->extra : nullptr;
 
     vk_buffer d_X = nullptr;
     size_t x_buf_offset = 0;
     vk_buffer d_Y = nullptr;
     size_t y_buf_offset = 0;
     vk_buffer d_Z = nullptr;
+    size_t z_buf_offset = 0;
 
     bool src0_uma = false;
     bool src1_uma = false;
+    bool src2_uma = false;
 
     if (ctx->device->uma) {
         ggml_vk_host_get(ctx, src0->data, d_X, x_buf_offset);
@@ -3944,10 +3961,15 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
             ggml_vk_host_get(ctx, src1->data, d_Y, y_buf_offset);
             src1_uma = d_Y != nullptr;
         }
+        if (use_src2) {
+            ggml_vk_host_get(ctx, src2->data, d_Z, z_buf_offset);
+            src2_uma = d_Z != nullptr;
+        }
     }
 
     uint64_t x_sz = ggml_vk_align_size(ggml_type_size(src0->type)/ggml_blck_size(src0->type) * ne0, ctx->device->properties.limits.minStorageBufferOffsetAlignment);
     uint64_t y_sz = use_src1 ? ggml_vk_align_size(ggml_type_size(src1->type) * ne1, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : 0;
+    uint64_t z_sz = use_src2 ? ggml_vk_align_size(ggml_type_size(src2->type) * ne2, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : 0;
     uint64_t d_sz = ggml_type_size(dst->type) * ne0;
 
     vk_buffer d_D = extra->buffer_gpu.lock();
@@ -3970,10 +3992,16 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
         y_buf_offset = extra_src1->offset;
         GGML_ASSERT(d_Y != nullptr);
     }
+    if (use_src2 && !src2_uma) {
+        d_Z = extra_src2->buffer_gpu.lock();
+        z_buf_offset = extra_src2->offset;
+        GGML_ASSERT(d_Z != nullptr);
+    }
 
     if (op_supports_incontiguous) {
         x_sz = ggml_nbytes(src0);
         y_sz = use_src1 ? ggml_nbytes(src1) : 0;
+        z_sz = use_src2 ? ggml_nbytes(src2) : 0;
         d_sz = ggml_nbytes(dst);
 
         if (x_buf_offset + x_sz >= d_X->size) {
@@ -3982,6 +4010,9 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
         if (use_src1 && y_buf_offset + y_sz >= d_Y->size) {
             y_sz = VK_WHOLE_SIZE;
         }
+        if (use_src2 && z_buf_offset + z_sz >= d_Z->size) {
+            z_sz = VK_WHOLE_SIZE;
+        }
         if (d_buf_offset + d_sz >= d_D->size) {
             d_sz = VK_WHOLE_SIZE;
         }
@@ -4021,13 +4052,16 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
             if (use_src1 && y_sz != VK_WHOLE_SIZE) {
                 y_sz *= ne12 * ne13;
             }
+            if (use_src2 && z_sz != VK_WHOLE_SIZE) {
+                z_sz *= ne22 * ne23;
+            }
             if (d_sz != VK_WHOLE_SIZE) {
                 d_sz *= ne02 * ne03;
             }
         }
 
         if (op == GGML_OP_SOFT_MAX) {
-            // Empty src1 is possible on soft_max, but the shader needs a buffer
+            // Empty src1 is possible in soft_max, but the shader needs a buffer
             vk_subbuffer subbuf_y;
             if (use_src1) {
                 subbuf_y = { d_Y, y_buf_offset, y_sz };
@@ -4037,6 +4071,28 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
 
             ggml_vk_sync_buffers(subctx);
             ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, subbuf_y, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        } else if (op == GGML_OP_ROPE) {
+            const int mode          = ((int32_t *) dst->op_params)[2];
+            const bool is_neox = mode & 2;
+
+            if (is_neox) {
+                // Empty src2 is possible in rope, but the shader needs a buffer
+                vk_subbuffer subbuf_z;
+                if (use_src2) {
+                    subbuf_z = { d_Z, z_buf_offset, z_sz };
+                } else {
+                    subbuf_z = { d_X, 0, d_X->size };
+                }
+
+                ggml_vk_sync_buffers(subctx);
+                ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, subbuf_z, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+            } else {
+                ggml_vk_sync_buffers(subctx);
+                ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+            }
+        } else if (use_src2) {
+            ggml_vk_sync_buffers(subctx);
+            ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, { d_Z, z_buf_offset, z_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
         } else if (use_src1) {
             ggml_vk_sync_buffers(subctx);
             ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { { d_X, x_buf_offset, x_sz }, { d_Y, y_buf_offset, y_sz }, { d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
@@ -4047,6 +4103,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
     } else {
         GGML_ASSERT(op != GGML_OP_SOFT_MAX);
         GGML_ASSERT(op != GGML_OP_ARGSORT);
+        GGML_ASSERT(!use_src2);
 
         ggml_pipeline_allocate_descriptor_sets(ctx, pipeline, ne02 * ne03);
 
@@ -4088,7 +4145,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context * subctx, c
 }
 
 static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_REPEAT, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f });
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_REPEAT, { (uint32_t)ggml_nelements(src0), (uint32_t)ggml_nelements(src1), 0.0f, 0.0f });
 }
 
 static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -4096,7 +4153,7 @@ static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context * subctx,
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_GET_ROWS, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_GET_ROWS, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -4111,7 +4168,7 @@ static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context * subctx, cons
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_ADD, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ADD, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -4126,7 +4183,7 @@ static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context * subctx, cons
     const uint32_t src1_type_size = ggml_type_size(src1->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_MUL, {
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_MUL, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
@@ -4141,7 +4198,7 @@ static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context * subctx, co
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_SCALE, {
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SCALE, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
@@ -4154,7 +4211,7 @@ static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context * subctx, cons
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_SQR, {
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQR, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
@@ -4168,7 +4225,7 @@ static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context * subctx, co
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t dst_type_size = ggml_type_size(dst->type);
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_CLAMP, {
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CLAMP, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
@@ -4183,7 +4240,7 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context * subctx, cons
     const uint32_t dst_type_size = ggml_type_size(dst->type);
     const uint32_t d_offset = (extra->offset % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
 
-    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_CPY, {
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, {
         (uint32_t)ggml_nelements(src0),
         (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
         (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
@@ -4195,21 +4252,21 @@ static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context * subctx, cons
 static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
 
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
 }
 
 static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     float * op_params = (float *)dst->op_params;
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f });
 }
 
 static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f });
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f });
 }
 
 static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, ggml_tensor * dst) {
     int32_t * op_params = (int32_t *)dst->op_params;
-    ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] });
+    ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] });
 }
 
 static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
@@ -4228,7 +4285,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx,
     const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
     const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
-    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_SOFT_MAX, {
+    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX, {
         ncols,
         src1 != nullptr ? nrows_y : (uint32_t)0,
         scale, max_bias,
@@ -4237,7 +4294,7 @@ static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context * subctx,
     });
 }
 
-static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst) {
     const int n_dims        = ((int32_t *) dst->op_params)[1];
     const int mode          = ((int32_t *) dst->op_params)[2];
     // const int n_ctx         = ((int32_t *) dst->op_params)[3];
@@ -4260,12 +4317,13 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context * subctx, con
     if (is_neox) {
         const float theta_scale = powf(freq_base, -2.0f/n_dims);
         const float inv_ndims = -1.0f / n_dims;
-        ggml_vk_op_f32<vk_op_rope_neox_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_ROPE, {
+        ggml_vk_op_f32<vk_op_rope_neox_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_ROPE, {
             (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
-            freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1], 0.0f, 0.0f}, theta_scale, inv_ndims
+            freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1], 0.0f, 0.0f}, theta_scale, inv_ndims,
+            src2 != nullptr,
         });
     } else {
-        ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, dst, GGML_OP_ROPE, {
+        ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_ROPE, {
             (uint32_t)src0->ne[0], freq_scale, (uint32_t)src0->ne[1],
             freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1], 0.0f, 0.0f}
         });
@@ -4288,7 +4346,7 @@ static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context * subctx,
 
     std::cerr << ((ggml_sort_order) op_params[0]) << " " << GGML_SORT_ORDER_ASC << std::endl;
 
-    ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, dst, GGML_OP_ARGSORT, {
+    ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
         ncols,
         ncols_pad,
         op_params[0],
@@ -5404,6 +5462,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 
     const ggml_tensor * src0 = node->src[0];
     const ggml_tensor * src1 = node->src[1];
+    const ggml_tensor * src2 = node->src[2];
 
     switch (node->op) {
     case GGML_OP_UNARY:
@@ -5520,7 +5579,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
 
         break;
     case GGML_OP_ROPE:
-        ggml_vk_rope(ctx, ctx->compute_ctx, src0, src1, node);
+        ggml_vk_rope(ctx, ctx->compute_ctx, src0, src1, src2, node);
 
         break;
     case GGML_OP_ARGSORT:
@@ -6496,7 +6555,7 @@ static void ggml_vk_print_graph_origin(const ggml_tensor * tensor, std::vector<c
     for (int j = 0; j < level; j++) {
         std::cerr << " ";
     }
-    std::cerr << ggml_op_name(tensor->op) << " gpu=" << (tensor->extra != nullptr) << " backend=" << tensor->backend << std::endl;
+    std::cerr << ggml_op_name(tensor->op) << " gpu=" << (tensor->extra != nullptr) << std::endl;
 
     done.push_back(tensor);
 
@@ -6546,7 +6605,7 @@ static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * d
 static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tensor * tensor, const char * name) {
     void * tensor_data = tensor->data;
 
-    if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
+    if (ggml_backend_buffer_is_vk(tensor->buffer)) {
         const size_t tensor_size = ggml_nbytes(tensor);
         tensor_data = malloc(tensor_size);
 
@@ -6557,12 +6616,12 @@ static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tenso
     }
 
     std::cerr << "TENSOR CHECK " << name << " (" << tensor->name << "): " << ggml_op_name(tensor->op) << std::endl;
-    std::cerr << "tensor=" << tensor << " tensor->backend: " << tensor->backend << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << std::endl;
+    std::cerr << "tensor=" << tensor << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << std::endl;
     if (tensor->src[0] != nullptr) {
-        std::cerr << "tensor->src[0]=" << tensor->src[0] << " name=" << tensor->src[0]->name << " op=" << ggml_op_name(tensor->src[0]->op) << " type=" << ggml_type_name(tensor->src[0]->type) << " backend=" << tensor->src[0]->backend << " ne0=" << tensor->src[0]->ne[0] << " nb0=" << tensor->src[0]->nb[0] << " ne1=" << tensor->src[0]->ne[1] << " nb1=" << tensor->src[0]->nb[1] << " ne2=" << tensor->src[0]->ne[2] << " nb2=" << tensor->src[0]->nb[2] << " ne3=" << tensor->src[0]->ne[3] << " nb3=" << tensor->src[0]->nb[3] << std::endl;
+        std::cerr << "tensor->src[0]=" << tensor->src[0] << " name=" << tensor->src[0]->name << " op=" << ggml_op_name(tensor->src[0]->op) << " type=" << ggml_type_name(tensor->src[0]->type) << " ne0=" << tensor->src[0]->ne[0] << " nb0=" << tensor->src[0]->nb[0] << " ne1=" << tensor->src[0]->ne[1] << " nb1=" << tensor->src[0]->nb[1] << " ne2=" << tensor->src[0]->ne[2] << " nb2=" << tensor->src[0]->nb[2] << " ne3=" << tensor->src[0]->ne[3] << " nb3=" << tensor->src[0]->nb[3] << std::endl;
     }
     if (tensor->src[1] != nullptr) {
-        std::cerr << "tensor->src[1]=" << tensor->src[1] << " name=" << tensor->src[1]->name << " op=" << ggml_op_name(tensor->src[1]->op) << " type=" << ggml_type_name(tensor->src[1]->type) << " backend=" << tensor->src[1]->backend << " ne0=" << tensor->src[1]->ne[0] << " nb0=" << tensor->src[1]->nb[0] << " ne1=" << tensor->src[1]->ne[1] << " nb1=" << tensor->src[1]->nb[1] << " ne2=" << tensor->src[1]->ne[2] << " nb2=" << tensor->src[1]->nb[2] << " ne3=" << tensor->src[1]->ne[3] << " nb3=" << tensor->src[1]->nb[3] << std::endl;
+        std::cerr << "tensor->src[1]=" << tensor->src[1] << " name=" << tensor->src[1]->name << " op=" << ggml_op_name(tensor->src[1]->op) << " type=" << ggml_type_name(tensor->src[1]->type) << " ne0=" << tensor->src[1]->ne[0] << " nb0=" << tensor->src[1]->nb[0] << " ne1=" << tensor->src[1]->ne[1] << " nb1=" << tensor->src[1]->nb[1] << " ne2=" << tensor->src[1]->ne[2] << " nb2=" << tensor->src[1]->nb[2] << " ne3=" << tensor->src[1]->ne[3] << " nb3=" << tensor->src[1]->nb[3] << std::endl;
     }
     std::cerr << std::endl << "Result:" << std::endl;
     ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 0, 0);
@@ -6573,43 +6632,11 @@ static void ggml_vk_print_tensor(ggml_backend_vk_context * ctx, const ggml_tenso
     std::vector<const ggml_tensor *> done;
     ggml_vk_print_graph_origin(tensor, done);
 
-    if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
+    if (ggml_backend_buffer_is_vk(tensor->buffer)) {
         free(tensor_data);
     }
 }
 
-static void ggml_vk_check_tensor(const std::string& name, const ggml_tensor * tensor) {
-    return;
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_CPU);
-    if (tensor->type != GGML_TYPE_F32 && tensor->type != GGML_TYPE_F16) {
-        return;
-    }
-    for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
-        for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
-            for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
-                for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
-                    float val = 0.0f;
-                    if (tensor->type == GGML_TYPE_F32) {
-                        val = *(float *) ((char *) tensor->data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]);
-                    } else if (tensor->type == GGML_TYPE_F16) {
-                        val = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor->data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]));
-                    }
-                    if (std::isnan(val)) {
-                        std::cerr << "ERROR: TENSOR CHECK " << name << ": Invalid value in " << ggml_op_name(tensor->op) << " i3=" << i3 << " i2=" << i2 << " i1=" << i1 << " i0=" << i0 << " val=" << val << std::endl;
-                        std::cerr << "tensor=" << tensor << " tensor->type=" << ggml_type_name(tensor->type) << " tensor->backend: " << tensor->backend << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << std::endl;
-                        std::cerr << std::endl;
-                        ggml_vk_print_tensor_area(tensor, tensor->data, i0, i1, i2, i3);
-                        std::cerr << std::endl;
-                        std::vector<const ggml_tensor *> done;
-                        ggml_vk_print_graph_origin(tensor, done);
-                        GGML_ASSERT(false);
-                    }
-                }
-            }
-        }
-    }
-}
-
 void * comp_result;
 size_t comp_size;
 size_t comp_nb[GGML_MAX_DIMS];
@@ -6633,6 +6660,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
 
     ggml_tensor * src0 = tensor->src[0];
     ggml_tensor * src1 = tensor->src[1];
+    ggml_tensor * src2 = tensor->src[2];
 
     struct ggml_init_params iparams = {
         /*.mem_size   =*/ 1024*1024*1024,
@@ -6662,10 +6690,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
 
         src0_buffer = malloc(src0_size);
         src0_clone->data = src0_buffer;
-        if (src0->backend == GGML_BACKEND_TYPE_CPU) {
+        if (ggml_backend_buffer_is_host(src0->buffer)) {
             memcpy(src0_clone->data, src0->data, src0_size);
             memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS);
-        } else if (src0->backend == GGML_BACKEND_TYPE_GPU) {
+        } else if (ggml_backend_buffer_is_vk(src0->buffer)) {
             ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src0->extra;
             vk_buffer buffer_gpu = extra->buffer_gpu.lock();
             uint64_t offset = extra->offset;
@@ -6696,8 +6724,6 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
         if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
             ggml_vk_print_tensor(ctx, src0, "src0");
         }
-
-        ggml_vk_check_tensor(std::string(ggml_op_name(tensor->op)) + "->src0", src0_clone);
     }
     if (src1 != nullptr) {
         src1_clone = ggml_dup_tensor(ggml_ctx, src1);
@@ -6706,10 +6732,10 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
 
         src1_buffer = malloc(src1_size);
         src1_clone->data = src1_buffer;
-        if (src1->backend == GGML_BACKEND_TYPE_CPU) {
+        if (ggml_backend_buffer_is_host(src1->buffer)) {
             memcpy(src1_clone->data, src1->data, src1_size);
             memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS);
-        } else if (src1->backend == GGML_BACKEND_TYPE_GPU) {
+        } else if (ggml_backend_buffer_is_vk(src1->buffer)) {
             ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src1->extra;
             vk_buffer buffer_gpu = extra->buffer_gpu.lock();
             uint64_t offset = extra->offset;
@@ -6740,12 +6766,12 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
         if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
             ggml_vk_print_tensor(ctx, src1, "src1");
             std::cerr << "TENSOR CHECK: " << ggml_op_name(src1_clone->op) << " (check " << check_counter << ")" << std::endl;
-            std::cerr << "src1_clone=" << tensor << " src1_clone->backend: " << src1_clone->backend << " src1_clone->type: " << ggml_type_name(src1_clone->type) << " ne0=" << src1_clone->ne[0] << " nb0=" << src1_clone->nb[0] << " ne1=" << src1_clone->ne[1] << " nb1=" << src1_clone->nb[1] << " ne2=" << src1_clone->ne[2] << " nb2=" << src1_clone->nb[2] << " ne3=" << src1_clone->ne[3] << " nb3=" << src1_clone->nb[3] << std::endl;
+            std::cerr << "src1_clone=" << tensor << " src1_clone->type: " << ggml_type_name(src1_clone->type) << " ne0=" << src1_clone->ne[0] << " nb0=" << src1_clone->nb[0] << " ne1=" << src1_clone->ne[1] << " nb1=" << src1_clone->nb[1] << " ne2=" << src1_clone->ne[2] << " nb2=" << src1_clone->nb[2] << " ne3=" << src1_clone->ne[3] << " nb3=" << src1_clone->nb[3] << std::endl;
             if (src1->src[0] != nullptr) {
-                std::cerr << "src1->src[0]=" << src1->src[0] << " op=" << ggml_op_name(src1->src[0]->op) << " type=" << ggml_type_name(src1->src[0]->type) << " backend=" << src1->src[0]->backend << " ne0=" << src1->src[0]->ne[0] << " nb0=" << src1->src[0]->nb[0] << " ne1=" << src1->src[0]->ne[1] << " nb1=" << src1->src[0]->nb[1] << " ne2=" << src1->src[0]->ne[2] << " nb2=" << src1->src[0]->nb[2] << " ne3=" << src1->src[0]->ne[3] << " nb3=" << src1->src[0]->nb[3] << std::endl;
+                std::cerr << "src1->src[0]=" << src1->src[0] << " op=" << ggml_op_name(src1->src[0]->op) << " type=" << ggml_type_name(src1->src[0]->type) << " ne0=" << src1->src[0]->ne[0] << " nb0=" << src1->src[0]->nb[0] << " ne1=" << src1->src[0]->ne[1] << " nb1=" << src1->src[0]->nb[1] << " ne2=" << src1->src[0]->ne[2] << " nb2=" << src1->src[0]->nb[2] << " ne3=" << src1->src[0]->ne[3] << " nb3=" << src1->src[0]->nb[3] << std::endl;
             }
             if (src1->src[1] != nullptr) {
-                std::cerr << "src1->src[1]=" << src1->src[1] << " op=" << ggml_op_name(src1->src[1]->op) << " type=" << ggml_type_name(src1->src[1]->type) << " backend=" << src1->src[1]->backend << " ne0=" << src1->src[1]->ne[0] << " nb0=" << src1->src[1]->nb[0] << " ne1=" << src1->src[1]->ne[1] << " nb1=" << src1->src[1]->nb[1] << " ne2=" << src1->src[1]->ne[2] << " nb2=" << src1->src[1]->nb[2] << " ne3=" << src1->src[1]->ne[3] << " nb3=" << src1->src[1]->nb[3] << std::endl;
+                std::cerr << "src1->src[1]=" << src1->src[1] << " op=" << ggml_op_name(src1->src[1]->op) << " type=" << ggml_type_name(src1->src[1]->type) << " ne0=" << src1->src[1]->ne[0] << " nb0=" << src1->src[1]->nb[0] << " ne1=" << src1->src[1]->ne[1] << " nb1=" << src1->src[1]->nb[1] << " ne2=" << src1->src[1]->ne[2] << " nb2=" << src1->src[1]->nb[2] << " ne3=" << src1->src[1]->ne[3] << " nb3=" << src1->src[1]->nb[3] << std::endl;
             }
             std::cerr << std::endl << "Result:" << std::endl;
             ggml_vk_print_tensor_area(src1_clone, src1_clone->data, 5, 5, 0, 0);
@@ -6756,8 +6782,64 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
             std::vector<const ggml_tensor *> done;
             ggml_vk_print_graph_origin(src1_clone, done);
         }
+    }
+    if (src2 != nullptr) {
+        src2_clone = ggml_dup_tensor(ggml_ctx, src2);
 
-        ggml_vk_check_tensor(std::string(ggml_op_name(tensor->op)) + "->src1", src1_clone);
+        src2_size = ggml_nbytes(src2);
+
+        src2_buffer = malloc(src2_size);
+        src2_clone->data = src2_buffer;
+        if (ggml_backend_buffer_is_host(src2->buffer)) {
+            memcpy(src2_clone->data, src2->data, src2_size);
+            memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS);
+        } else if (ggml_backend_buffer_is_vk(src2->buffer)) {
+            ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src2->extra;
+            vk_buffer buffer_gpu = extra->buffer_gpu.lock();
+            uint64_t offset = extra->offset;
+            if (!ggml_is_contiguous(src2) && ggml_vk_dim01_contiguous(src2)) {
+                for (int i3 = 0; i3 < src2->ne[3]; i3++) {
+                    for (int i2 = 0; i2 < src2->ne[2]; i2++) {
+                        const int idx = i3*src2->ne[2] + i2;
+                        ggml_vk_buffer_read(ctx, buffer_gpu, offset + idx * src2->nb[2], ((char *)src2_clone->data + idx * src2_clone->nb[2]), src2->ne[1] * src2->nb[1]);
+                    }
+                }
+
+                src2_clone->nb[0] = src2->nb[0];
+                src2_clone->nb[1] = src2->nb[1];
+                for (int i = 2; i < GGML_MAX_DIMS; i++) {
+                    src2_clone->nb[i] = src2_clone->nb[i - 1]*src2_clone->ne[i - 1];
+                }
+            } else {
+                if (offset + src2_size >= buffer_gpu->size) {
+                    src2_size = buffer_gpu->size - offset;
+                }
+                ggml_vk_buffer_read(ctx, buffer_gpu, offset, src2_clone->data, src2_size);
+                memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS);
+            }
+        } else {
+            GGML_ASSERT(false);
+        }
+
+        if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
+            ggml_vk_print_tensor(ctx, src2, "src2");
+            std::cerr << "TENSOR CHECK: " << ggml_op_name(src2_clone->op) << " (check " << check_counter << ")" << std::endl;
+            std::cerr << "src2_clone=" << tensor << " src2_clone->type: " << ggml_type_name(src2_clone->type) << " ne0=" << src2_clone->ne[0] << " nb0=" << src2_clone->nb[0] << " ne1=" << src2_clone->ne[1] << " nb1=" << src2_clone->nb[1] << " ne2=" << src2_clone->ne[2] << " nb2=" << src2_clone->nb[2] << " ne3=" << src2_clone->ne[3] << " nb3=" << src2_clone->nb[3] << std::endl;
+            if (src2->src[0] != nullptr) {
+                std::cerr << "src2->src[0]=" << src2->src[0] << " op=" << ggml_op_name(src2->src[0]->op) << " type=" << ggml_type_name(src2->src[0]->type) << " ne0=" << src2->src[0]->ne[0] << " nb0=" << src2->src[0]->nb[0] << " ne1=" << src2->src[0]->ne[1] << " nb1=" << src2->src[0]->nb[1] << " ne2=" << src2->src[0]->ne[2] << " nb2=" << src2->src[0]->nb[2] << " ne3=" << src2->src[0]->ne[3] << " nb3=" << src2->src[0]->nb[3] << std::endl;
+            }
+            if (src2->src[1] != nullptr) {
+                std::cerr << "src2->src[1]=" << src2->src[1] << " op=" << ggml_op_name(src2->src[1]->op) << " type=" << ggml_type_name(src2->src[1]->type) << " ne0=" << src2->src[1]->ne[0] << " nb0=" << src2->src[1]->nb[0] << " ne1=" << src2->src[1]->ne[1] << " nb1=" << src2->src[1]->nb[1] << " ne2=" << src2->src[1]->ne[2] << " nb2=" << src2->src[1]->nb[2] << " ne3=" << src2->src[1]->ne[3] << " nb3=" << src2->src[1]->nb[3] << std::endl;
+            }
+            std::cerr << std::endl << "Result:" << std::endl;
+            ggml_vk_print_tensor_area(src2_clone, src2_clone->data, 5, 5, 0, 0);
+            std::cerr << std::endl;
+            std::cerr << std::endl << "Result:" << std::endl;
+            ggml_vk_print_tensor_area(src2_clone, src2_clone->data, 5, 5, 1, 0);
+            std::cerr << std::endl;
+            std::vector<const ggml_tensor *> done;
+            ggml_vk_print_graph_origin(src2_clone, done);
+        }
     }
 
     if (tensor->op == GGML_OP_MUL_MAT) {
@@ -6795,7 +6877,7 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
         float attn_factor     = ((float *)   tensor->op_params)[8];
         float beta_fast       = ((float *)   tensor->op_params)[9];
         float beta_slow       = ((float *)   tensor->op_params)[10];
-        tensor_clone = ggml_rope_custom(ggml_ctx, src0_clone, src1_clone, n_dims, mode, n_ggml_ctx, n_orig_ggml_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+        tensor_clone = ggml_rope_ext(ggml_ctx, src0_clone, src1_clone, src2_clone, n_dims, mode, n_ggml_ctx, n_orig_ggml_ctx, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
     } else if (tensor->op == GGML_OP_UNARY) {
         switch (ggml_get_unary_op(tensor)) {
         case GGML_UNARY_OP_SILU:
@@ -6843,7 +6925,6 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
 
     ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 8);
 
-    ggml_vk_check_tensor(ggml_op_name(tensor->op), tensor_clone);
     if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
         ggml_vk_print_tensor(ctx, tensor_clone, "tensor_clone");
     }
@@ -6884,7 +6965,7 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
 
     void * tensor_data = tensor->data;
 
-    if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
+    if (ggml_backend_buffer_is_vk(tensor->buffer)) {
         size_t tensor_size = ggml_nbytes(tensor);
         tensor_data = malloc(tensor_size);
 
@@ -6932,12 +7013,12 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
 
                     if ((std::isnan(correct) != std::isnan(result)) || (std::isinf(correct) != std::isinf(result)) || !buffer_size_fit) {
                         std::cerr << "ERROR: Invalid value in " << ggml_op_name(tensor->op) << " i3=" << i3 << " i2=" << i2 << " i1=" << i1 << " i0=" << i0 << " result=" << result << " correct=" << correct << " avg_err=" << (avg_err / counter) << std::endl;
-                        std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->backend: " << tensor->backend << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
+                        std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
                         if (src0 != nullptr) {
-                            std::cerr << "src0=" << src0 << " src0->name=" << src0->name << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " backend=" << src0->backend << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
+                            std::cerr << "src0=" << src0 << " src0->name=" << src0->name << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
                         }
                         if (src1 != nullptr) {
-                            std::cerr << "src1=" << src1 << " src1->name=" << src1->name << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " backend=" << src1->backend << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
+                            std::cerr << "src1=" << src1 << " src1->name=" << src1->name << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
                         }
                         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
                         std::cerr << std::endl << "Result:" << std::endl;
@@ -6973,12 +7054,12 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
 
     if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
         std::cerr << "TENSOR CHECK: avg_err=" << avg_err << " in " << ggml_op_name(tensor->op) << " (check " << check_counter << ")" << std::endl;
-        std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->backend: " << tensor->backend << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
+        std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
         if (src0 != nullptr) {
-            std::cerr << "src0=" << src0 << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " backend=" << src0->backend << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
+            std::cerr << "src0=" << src0 << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
         }
         if (src1 != nullptr) {
-            std::cerr << "src1=" << src1 << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " backend=" << src1->backend << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
+            std::cerr << "src1=" << src1 << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
         }
         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
         std::cerr << std::endl << "Result:" << std::endl;
@@ -6997,12 +7078,12 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
 
     if (avg_err > 0.05 || std::isnan(avg_err)) {
         std::cerr << "ERROR: avg_err=" << avg_err << " in " << ggml_op_name(tensor->op) << " (check " << check_counter << ")" << std::endl;
-        std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->backend: " << tensor->backend << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
+        std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
         if (src0 != nullptr) {
-            std::cerr << "src0=" << src0 << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " backend=" << src0->backend << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
+            std::cerr << "src0=" << src0 << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
         }
         if (src1 != nullptr) {
-            std::cerr << "src1=" << src1 << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " backend=" << src1->backend << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
+            std::cerr << "src1=" << src1 << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
         }
         std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
         std::cerr << std::endl << "Result:" << std::endl;
@@ -7014,14 +7095,14 @@ static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_
         ggml_vk_print_graph_origin(tensor, done);
         GGML_ASSERT(false);
     } else {
-        std::cerr << check_counter << " " << tensor->name << " op=" << ggml_op_name(tensor->op) << " backend=" << tensor->backend << " avg_err=" << avg_err << std::endl;
+        std::cerr << check_counter << " " << tensor->name << " op=" << ggml_op_name(tensor->op) << " avg_err=" << avg_err << std::endl;
     }
 
     free(comp_result);
     comp_result = nullptr;
     comp_size = 0;
 
-    if (tensor->backend == GGML_BACKEND_TYPE_GPU) {
+    if (ggml_backend_buffer_is_vk(tensor->buffer)) {
         free(tensor_data);
     }
 }

ggml.h

@@ -481,9 +481,7 @@ extern "C" {
         GGML_OP_ARGSORT,
         GGML_OP_LEAKY_RELU,
 
-        GGML_OP_FLASH_ATTN,
         GGML_OP_FLASH_ATTN_EXT,
-        GGML_OP_FLASH_FF,
         GGML_OP_FLASH_ATTN_BACK,
         GGML_OP_SSM_CONV,
         GGML_OP_SSM_SCAN,
@@ -1460,11 +1458,12 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements (DEPRECATED)
+    // if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
     // if mode & 2 == 1, GPT-NeoX style
     // if mode & 4 == 1, ChatGLM style
     //
     // b is an int32 vector with size a->ne[2], it contains the positions
+    // c is freq factors (e.g. phi3-128k), (optional)
     GGML_API struct ggml_tensor * ggml_rope(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1483,10 +1482,11 @@ extern "C" {
             int                   n_ctx);
 
     // custom RoPE
-    GGML_API struct ggml_tensor * ggml_rope_custom(
+    GGML_API struct ggml_tensor * ggml_rope_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
@@ -1499,7 +1499,23 @@ extern "C" {
             float                 beta_slow);
 
     // in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
+    GGML_API struct ggml_tensor * ggml_rope_ext_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx,
+            int                   n_orig_ctx,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
@@ -1512,20 +1528,28 @@ extern "C" {
             float                 ext_factor,
             float                 attn_factor,
             float                 beta_fast,
-            float                 beta_slow);
+            float                 beta_slow),
+        "use ggml_rope_ext instead");
 
-    // compute correction dims for YaRN RoPE scaling
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
-    GGML_CALL void ggml_rope_yarn_corr_dims(
-        int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
-
-    // xPos RoPE, in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             int                   n_dims,
-            float                 base,
+            int                   mode,
-            bool                  down);
+            int                   n_ctx,
+            int                   n_orig_ctx,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow),
+        "use ggml_rope_ext_inplace instead");
+
+    // compute correction dims for YaRN RoPE scaling
+    GGML_CALL void ggml_rope_yarn_corr_dims(
+        int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
 
     // rotary position embedding backward, i.e compute dx from dy
     // a - dy
@@ -1533,6 +1557,7 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
             int                   n_dims,
             int                   mode,
             int                   n_ctx,
@@ -1734,13 +1759,6 @@ extern "C" {
             struct ggml_tensor  * a,
             int                   k);
 
-    GGML_API struct ggml_tensor * ggml_flash_attn(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * q,
-            struct ggml_tensor  * k,
-            struct ggml_tensor  * v,
-            bool                  masked);
-
 #define GGML_KQ_MASK_PAD 32
 
     // q:    [n_embd, n_batch,     n_head,    1]
@@ -1761,6 +1779,7 @@ extern "C" {
             struct ggml_tensor * a,
             enum ggml_prec       prec);
 
+    // TODO: needs to be adapted to ggml_flash_attn_ext
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,
            struct ggml_tensor  * q,
@@ -1769,14 +1788,6 @@ extern "C" {
            struct ggml_tensor  * d,
            bool                  masked);
 
-    GGML_API struct ggml_tensor * ggml_flash_ff(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b0,
-            struct ggml_tensor  * b1,
-            struct ggml_tensor  * c0,
-            struct ggml_tensor  * c1);
-
     GGML_API struct ggml_tensor * ggml_ssm_conv(
             struct ggml_context * ctx,
             struct ggml_tensor  * s,

ggml_vk_generate_shaders.py

@@ -2609,7 +2609,8 @@ layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
 layout (binding = 1) readonly buffer Y {int data_b[];};
-layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+layout (binding = 2) readonly buffer Z {float data_freq_factors[];};
+layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
 
 layout (push_constant) uniform parameter {
     uint ncols;
@@ -2622,6 +2623,7 @@ layout (push_constant) uniform parameter {
     float corr_dims[4];
     float theta_scale;
     float inv_ndims;
+    uint has_freq_facs;
 } p;
 
 float rope_yarn_ramp(const float low, const float high, const uint i0) {
@@ -2671,7 +2673,8 @@ void main() {
     const float cur_rot = p.inv_ndims * ic - ib;
 
     const int pos = data_b[i2];
-    const float theta_base = pos*p.freq_scale*pow(p.theta_scale, col/2.0f);
+    const float freq_factor = p.has_freq_facs != 0 ? data_freq_factors[ic/2] : 1.0f;
+    const float theta_base = pos*p.freq_scale*pow(p.theta_scale, col/2.0f) / freq_factor;
 
     float cos_theta, sin_theta;
     rope_yarn(theta_base, uint(cur_rot), cos_theta, sin_theta);

gguf-py/gguf/constants.py

@@ -61,6 +61,7 @@ class Keys:
         FREQ_BASE               = "{arch}.rope.freq_base"
         SCALING_TYPE            = "{arch}.rope.scaling.type"
         SCALING_FACTOR          = "{arch}.rope.scaling.factor"
+        SCALING_ATTN_FACTOR     = "{arch}.rope.scaling.attn_factor"
         SCALING_ORIG_CTX_LEN    = "{arch}.rope.scaling.original_context_length"
         SCALING_FINETUNED       = "{arch}.rope.scaling.finetuned"
 
@@ -148,6 +149,8 @@ class MODEL_TENSOR(IntEnum):
     OUTPUT             = auto()
     OUTPUT_NORM        = auto()
     ROPE_FREQS         = auto()
+    ROPE_FACTORS_LONG  = auto()
+    ROPE_FACTORS_SHORT = auto()
     ATTN_Q             = auto()
     ATTN_K             = auto()
     ATTN_V             = auto()
@@ -225,6 +228,8 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.OUTPUT_NORM:        "output_norm",
     MODEL_TENSOR.OUTPUT:             "output",
     MODEL_TENSOR.ROPE_FREQS:         "rope_freqs",
+    MODEL_TENSOR.ROPE_FACTORS_LONG:  "rope_factors_long",
+    MODEL_TENSOR.ROPE_FACTORS_SHORT: "rope_factors_short",
     MODEL_TENSOR.ATTN_NORM:          "blk.{bid}.attn_norm",
     MODEL_TENSOR.ATTN_NORM_2:        "blk.{bid}.attn_norm_2",
     MODEL_TENSOR.ATTN_QKV:           "blk.{bid}.attn_qkv",
@@ -900,9 +905,8 @@ class GGUFValueType(IntEnum):
             raise ValueError(f"Unknown type: {type(val)}")
 
 
-# Note: Does not support GGML_QKK_64
-QK_K = 256
 # Items here are (block size, type size)
+QK_K = 256
 GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
     GGMLQuantizationType.F32:     (1, 4),
     GGMLQuantizationType.F16:     (1, 2),

gguf-py/gguf/gguf_writer.py

@@ -433,6 +433,9 @@ class GGUFWriter:
     def add_rope_scaling_factor(self, value: float) -> None:
         self.add_float32(Keys.Rope.SCALING_FACTOR.format(arch=self.arch), value)
 
+    def add_rope_scaling_attn_factors(self, value: Sequence[float]) -> None:
+        self.add_float32(Keys.Rope.SCALING_ATTN_FACTOR.format(arch=self.arch), value)
+
     def add_rope_scaling_orig_ctx_len(self, value: int) -> None:
         self.add_uint32(Keys.Rope.SCALING_ORIG_CTX_LEN.format(arch=self.arch), value)
 

llama.h

@@ -759,6 +759,12 @@ extern "C" {
     // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
     LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
 
+    // Get the number of threads used for generation of a single token.
+    LLAMA_API uint32_t llama_n_threads(struct llama_context * ctx);
+
+    // Get the number of threads used for prompt and batch processing (multiple token).
+    LLAMA_API uint32_t llama_n_threads_batch(struct llama_context * ctx);
+
     // Set whether to use causal attention or not
     // If set to true, the model will only attend to the past tokens
     LLAMA_API void llama_set_causal_attn(struct llama_context * ctx, bool causal_attn);

tests/test-backend-ops.cpp

@@ -1142,20 +1142,22 @@ struct test_rope : public test_case {
     int n_dims;
     int mode;
     int n_ctx;
+    bool ff;
 
     std::string vars() override {
-        return VARS_TO_STR5(type, ne, n_dims, mode, n_ctx);
+        return VARS_TO_STR6(type, ne, n_dims, mode, n_ctx, ff);
     }
 
     test_rope(ggml_type type = GGML_TYPE_F32,
             std::array<int64_t, 4> ne = {10, 10, 10, 1},
-            int n_dims = 10, int mode = 0, int n_ctx = 512)
+            int n_dims = 10, int mode = 0, int n_ctx = 512, bool ff = false)
-        : type(type), ne(ne), n_dims(n_dims), mode(mode), n_ctx(n_ctx) {}
+        : type(type), ne(ne), n_dims(n_dims), mode(mode), n_ctx(n_ctx), ff(ff) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
         ggml_tensor * pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, ne[2]);
-        ggml_tensor * out = ggml_rope(ctx, a, pos, n_dims, mode, n_ctx);
+        ggml_tensor * freq = ff ? ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_dims/2) : nullptr;
+        ggml_tensor * out = ggml_rope_ext(ctx, a, pos, freq, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f);
         return out;
     }
 
@@ -1168,11 +1170,16 @@ struct test_rope : public test_case {
                     data[i] = rand() % n_ctx;
                 }
                 ggml_backend_tensor_set(t, data.data(), 0, ne[2] * sizeof(int));
+            } else {
+                if (t->ne[0] == n_dims/2) {
+                    // frequency factors in the range [0.9f, 1.1f]
+                    init_tensor_uniform(t, 0.9f, 1.1f);
                 } else {
                     init_tensor_uniform(t);
                 }
             }
         }
+    }
 };
 
 // GGML_OP_POOL2D
@@ -1763,14 +1770,14 @@ struct test_llama : public test_llm {
                 struct ggml_tensor * Kcur = ggml_mul_mat(ctx, wk, cur);
                 struct ggml_tensor * Vcur = ggml_mul_mat(ctx, wv, cur);
 
-                Qcur = ggml_rope_custom(
+                Qcur = ggml_rope_ext(
-                    ctx, ggml_reshape_3d(ctx, Qcur, hp.n_embd_head, hp.n_head,    hp.n_tokens), inp_pos,
+                    ctx, ggml_reshape_3d(ctx, Qcur, hp.n_embd_head, hp.n_head,    hp.n_tokens), inp_pos, nullptr,
                     hp.n_rot, 0, 0, hp.n_orig_ctx, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
 
-                Kcur = ggml_rope_custom(
+                Kcur = ggml_rope_ext(
-                    ctx, ggml_reshape_3d(ctx, Kcur, hp.n_embd_head, hp.n_head_kv, hp.n_tokens), inp_pos,
+                    ctx, ggml_reshape_3d(ctx, Kcur, hp.n_embd_head, hp.n_head_kv, hp.n_tokens), inp_pos, nullptr,
                     hp.n_rot, 0, 0, hp.n_orig_ctx, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
@@ -1889,13 +1896,13 @@ struct test_falcon : public test_llm {
                 Kcur = ggml_reshape_3d(ctx, Kcur, hp.n_embd_head, hp.n_head_kv, hp.n_tokens);
 
                 // using mode = 2 for neox mode
-                Qcur = ggml_rope_custom(
+                Qcur = ggml_rope_ext(
-                    ctx, Qcur, inp_pos, hp.n_rot, 2, 0, hp.n_orig_ctx,
+                    ctx, Qcur, inp_pos, nullptr, hp.n_rot, 2, 0, hp.n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
 
-                Kcur = ggml_rope_custom(
+                Kcur = ggml_rope_ext(
-                    ctx, Kcur, inp_pos, hp.n_rot, 2, 0, hp.n_orig_ctx,
+                    ctx, Kcur, inp_pos, nullptr, hp.n_rot, 2, 0, hp.n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
 
@@ -2188,16 +2195,20 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {32, 2, 32, 1}, true,  0.1f, 8.0f));
 
     for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
-        test_cases.emplace_back(new test_rope(type, {128,  32, 10, 1}, 128, 0, 512)); // llama 7B
+        // TODO: ff not supported yet for !neox
-        test_cases.emplace_back(new test_rope(type, {128,  40, 10, 1}, 128, 0, 512)); // llama 13B
+        test_cases.emplace_back(new test_rope(type, {128,  32, 10, 1}, 128, 0, 512, false)); // llama 7B
-        test_cases.emplace_back(new test_rope(type, {128,  52, 10, 1}, 128, 0, 512)); // llama 30B
+        test_cases.emplace_back(new test_rope(type, {128,  40, 10, 1}, 128, 0, 512, false)); // llama 13B
-        test_cases.emplace_back(new test_rope(type, {128,  64, 10, 1}, 128, 0, 512)); // llama 65B
+        test_cases.emplace_back(new test_rope(type, {128,  52, 10, 1}, 128, 0, 512, false)); // llama 30B
-        test_cases.emplace_back(new test_rope(type, { 64,   1, 10, 1},  64, 2, 512)); // neox (falcon 7B)
+        test_cases.emplace_back(new test_rope(type, {128,  64, 10, 1}, 128, 0, 512, false)); // llama 65B
-        test_cases.emplace_back(new test_rope(type, { 64,  71, 10, 1},  64, 2, 512)); // neox (falcon 7B)
+
-        test_cases.emplace_back(new test_rope(type, { 64,   8, 10, 1},  64, 2, 512)); // neox (falcon 40B)
+        for (bool ff : {false, true}) { // freq_factors
-        test_cases.emplace_back(new test_rope(type, { 64, 128, 10, 1},  64, 2, 512)); // neox (falcon 40B)
+            test_cases.emplace_back(new test_rope(type, { 64,   1, 10, 1},  64, 2, 512, ff)); // neox (falcon 7B)
-        test_cases.emplace_back(new test_rope(type, { 80,  32, 10, 1},  20, 2, 512)); // neox (stablelm)
+            test_cases.emplace_back(new test_rope(type, { 64,  71, 10, 1},  64, 2, 512, ff)); // neox (falcon 7B)
-        test_cases.emplace_back(new test_rope(type, { 80,  32, 10, 1},  32, 2, 512)); // neox (phi-2)
+            test_cases.emplace_back(new test_rope(type, { 64,   8, 10, 1},  64, 2, 512, ff)); // neox (falcon 40B)
+            test_cases.emplace_back(new test_rope(type, { 64, 128, 10, 1},  64, 2, 512, ff)); // neox (falcon 40B)
+            test_cases.emplace_back(new test_rope(type, { 80,  32, 10, 1},  20, 2, 512, ff)); // neox (stablelm)
+            test_cases.emplace_back(new test_rope(type, { 80,  32, 10, 1},  32, 2, 512, ff)); // neox (phi-2)
+        }
     }
 
     test_cases.emplace_back(new test_concat(GGML_TYPE_F32));

tests/test-chat-template.cpp

@@ -49,8 +49,14 @@ int main(void) {
         "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif false == true %}{% set loop_messages = messages %}{% set system_message = 'You are Command-R, a brilliant, sophisticated, AI-assistant trained to assist human users by providing thorough responses. You are trained by Cohere.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% if system_message != false %}{{ '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' + system_message + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|START_OF_TURN_TOKEN|><|USER_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% elif message['role'] == 'assistant' %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>'  + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}{% endif %}",
         // Llama-3
         "{% set loop_messages = messages %}{% for message in loop_messages %}{% set content = '<|start_header_id|>' + message['role'] + '<|end_header_id|>\n\n'+ message['content'] | trim + '<|eot_id|>' %}{% if loop.index0 == 0 %}{% set content = bos_token + content %}{% endif %}{{ content }}{% endfor %}{{ '<|start_header_id|>assistant<|end_header_id|>\n\n' }}",
-        // Phi-3
+        //Phi-3-mini
-        "{{ bos_token }}{% for message in messages %}{{'<|' + message['role'] + '|>' + ' ' + message['content'] + '<|end|> ' }}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|> ' }}{% else %}{{ eos_token }}{% endif %}"
+        "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif (message['role'] == 'assistant') %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}",
+        //Phi-3-small
+        "{{ bos_token }}{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>\n' }}{% else %}{{ eos_token }}{% endif %}",
+        //Phi-3-medium
+        "{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif (message['role'] == 'assistant') %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}",
+        //Phi-3-vision
+        "{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}{{- '<|assistant|>\n' -}}{% endif %}"
     };
     std::vector<std::string> expected_output = {
         // teknium/OpenHermes-2.5-Mistral-7B
@@ -79,7 +85,13 @@ int main(void) {
         "<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>You are a helpful assistant<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Hello<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>Hi there<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Who are you<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>I am an assistant<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|USER_TOKEN|>Another question<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>",
         // Llama 3
         "<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nHello<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nHi there<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nWho are you<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\nI am an assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\n\nAnother question<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n",
-        // Phi 3
+        //Phi-3-mini
+        "<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n   I am an assistant   <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
+        //Phi-3-small
+        "<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n   I am an assistant   <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
+        //Phi-3-medium
+        "<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n   I am an assistant   <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
+        //Phi-3-vision
         "<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n   I am an assistant   <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
     };
     std::vector<char> formatted_chat(1024);

tests/test-grad0.cpp

@@ -1515,90 +1515,50 @@ int main(int argc, const char ** argv) {
         }
 
         // flash_attn f32
-        {
+        // TODO: adapt to ggml_flash_attn_ext() changes
-            srand(seed);
+        //{
-            const int nargs = 3;
+        //    srand(seed);
+        //    const int nargs = 3;
 
-            int64_t ne2[4];
+        //    int64_t ne2[4];
 
-            get_random_dims(ne2, 4);
+        //    get_random_dims(ne2, 4);
-            int64_t D = ne2[0];
+        //    int64_t D = ne2[0];
-            int64_t N = ne2[1];
+        //    int64_t N = ne2[1];
-            int64_t M = ne2[2] + N;
+        //    int64_t M = ne2[2] + N;
-            int64_t B = ne2[3];
+        //    int64_t B = ne2[3];
 
-            for (int masked = 0; masked <= 1; ++masked) {
+        //    for (int masked = 0; masked <= 1; ++masked) {
-                for (int ndims = 2; ndims <= 4; ++ndims) {
+        //        for (int ndims = 2; ndims <= 4; ++ndims) {
-                    int max_nrep = (ndims >= 3) ? 2 : 1;
+        //            int max_nrep = (ndims >= 3) ? 2 : 1;
-                    for (int nrep = 1; nrep < max_nrep; ++nrep) {
+        //            for (int nrep = 1; nrep < max_nrep; ++nrep) {
-                        int64_t neq[4] = { D, N, B*nrep, ne[3] };
+        //                int64_t neq[4] = { D, N, B*nrep, ne[3] };
-                        int64_t nek[4] = { D, M, B, ne[3] };
+        //                int64_t nek[4] = { D, M, B, ne[3] };
-                        int64_t nev[4] = { M, D, B, ne[3] };
+        //                int64_t nev[4] = { M, D, B, ne[3] };
-                        if (ndims == 2) {
+        //                if (ndims == 2) {
-                            neq[2] = 1; neq[3] = 1;
+        //                    neq[2] = 1; neq[3] = 1;
-                            nek[2] = 1; nek[3] = 1;
+        //                    nek[2] = 1; nek[3] = 1;
-                            nev[2] = 1; nev[3] = 1;
+        //                    nev[2] = 1; nev[3] = 1;
-                        } else if (ndims == 3) {
+        //                } else if (ndims == 3) {
-                            neq[3] = 1;
+        //                    neq[3] = 1;
-                            nek[3] = 1;
+        //                    nek[3] = 1;
-                            nev[3] = 1;
+        //                    nev[3] = 1;
-                        }
+        //                }
-                        x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
+        //                x[0] = get_random_tensor_f32(ctx0, ndims, neq, -0.1250f, 0.1250f);
-                        x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
+        //                x[1] = get_random_tensor_f32(ctx0, ndims, nek, -0.1250f, 0.1250f);
-                        x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
+        //                x[2] = get_random_tensor_f32(ctx0, ndims, nev, -0.1250f, 0.1250f);
-                        ggml_set_param(ctx0, x[0]);
+        //                ggml_set_param(ctx0, x[0]);
-                        ggml_set_param(ctx0, x[1]);
+        //                ggml_set_param(ctx0, x[1]);
-                        ggml_set_param(ctx0, x[2]);
+        //                ggml_set_param(ctx0, x[2]);
 
-                        struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
+        //                struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
 
-                        check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
+        //                check_gradient("flash_attn f32", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
-                    }
+        //            }
-                }
+        //        }
-            }
+        //    }
-        }
+        //}
 
-        // flash_attn f16, not yet fully implemented
-        if(0)
-        {
-            srand(seed);
-            const int nargs = 3;
-
-            int64_t ne2[4];
-
-            get_random_dims(ne2, 4);
-            int64_t D = ne2[0];
-            int64_t N = ne2[1];
-            int64_t M = ne2[2] + N;
-            int64_t B = ne2[3];
-
-            for (int masked = 0; masked <= 1; ++masked) {
-                for (int ndims = 2; ndims <= 4; ++ndims) {
-                    int64_t neq[4] = { D, N, B, ne[3] };
-                    int64_t nek[4] = { D, M, B, ne[3] };
-                    int64_t nev[4] = { M, D, B, ne[3] };
-                    if (ndims == 2) {
-                        neq[2] = 1; neq[3] = 1;
-                        nek[2] = 1; nek[3] = 1;
-                        nev[2] = 1; nev[3] = 1;
-                    } else if (ndims == 3) {
-                        neq[3] = 1;
-                        nek[3] = 1;
-                        nev[3] = 1;
-                    }
-                    x[0] = get_random_tensor_f16(ctx0, ndims, neq, -0.1250f, 0.1250f);
-                    x[1] = get_random_tensor_f16(ctx0, ndims, nek, -0.1250f, 0.1250f);
-                    x[2] = get_random_tensor_f16(ctx0, ndims, nev, -0.1250f, 0.1250f);
-                    ggml_set_param(ctx0, x[0]);
-                    ggml_set_param(ctx0, x[1]);
-                    ggml_set_param(ctx0, x[2]);
-
-                    struct ggml_tensor * f = ggml_sum(ctx0, ggml_flash_attn(ctx0, x[0], x[1], x[2], (masked == 0)));
-
-                    check_gradient("flash_attn f16", ctx0, x, f, ndims, nargs, 1.5e-4f, 1e-3f, INFINITY);
-                }
-            }
-        }
         ggml_free(ctx0);
     }
 

tests/test-tokenizer-0.sh

@@ -17,10 +17,15 @@ make -j tests/test-tokenizer-0
 
 printf "Testing %s on %s ...\n" $name $input
 
-python3 ./tests/test-tokenizer-0.py ./models/tokenizers/$name --fname-tok $input > /tmp/test-tokenizer-0-$name-py.log 2>&1
+set -e
-cat /tmp/test-tokenizer-0-$name-py.log | grep "tokenized in"
 
+printf "Tokenizing using (py)  Python AutoTokenizer ...\n"
+python3 ./tests/test-tokenizer-0.py ./models/tokenizers/$name --fname-tok $input > /tmp/test-tokenizer-0-$name-py.log 2>&1
+
+printf "Tokenizing using (cpp) llama.cpp ...\n"
 ./tests/test-tokenizer-0 ./models/ggml-vocab-$name.gguf $input > /tmp/test-tokenizer-0-$name-cpp.log 2>&1
+
+cat /tmp/test-tokenizer-0-$name-py.log | grep "tokenized in"
 cat /tmp/test-tokenizer-0-$name-cpp.log | grep "tokenized in"
 
 diff $input.tok $input.tokcpp > /dev/null 2>&1

tests/test-tokenizer-random.py

@@ -154,19 +154,22 @@ def generator_custom_text_edge_cases() -> Iterator[str]:
         '\uFEFF//',   # unicode_ranges_control, 0xFEFF (BOM)
         'Cửa Việt',   # llama-3, ignore_merges = true
         '<s>a',       # Phi-3 fail
-        '<unk><|endoftext|><s>'  # Phi-3 fail
+        '<unk><|endoftext|><s>',  # Phi-3 fail
         'a\na',       # TODO: Bert fail
     ]
 
 
-def generator_random_special_tokens(special_tokens:list[str], iterations=100) -> Iterator[str]:
+def generator_random_special_tokens(tokenizer, iterations=100) -> Iterator[str]:
-    special_tokens = set(special_tokens)
+    special_tokens = set(tokenizer.all_special_tokens)
     special_tokens.update([" ", "\n", "\t", "-", "!", "one", "1", "<s>", "</s>"])
     special_tokens = list(sorted(special_tokens))
     rand = random.Random()
     for m in range(iterations):
         rand.seed(m)
         words = rand.choices(special_tokens, k=500)
+        if tokenizer.add_bos_token:  # skip spam warning of double BOS
+            while words and words[0] == tokenizer.bos_token:
+                words.pop(0)
         yield "".join(words)
 
 
@@ -290,18 +293,19 @@ def main(argv: list[str] = None):
     model = LibLlamaModel(LibLlama(), args.vocab_file, mparams=dict(vocab_only=True), cparams=dict(n_ctx=4096))
     tokenizer = AutoTokenizer.from_pretrained(args.dir_tokenizer)
 
-    def func_tokenize2(text: str):
+    tokenizer.add_bos_token = getattr(tokenizer, "add_bos_token", True)
-        return tokenizer.encode(text, add_special_tokens=False)
+    tokenizer.add_eos_token = getattr(tokenizer, "add_eos_token", False)
-
-    parse_special = all(len(func_tokenize2(t)) == 1 for t in tokenizer.all_special_tokens)
 
     def func_tokenize1(text: str):
-        return model.tokenize(text, add_special=False, parse_special=parse_special)
+        return model.tokenize(text, add_special=True, parse_special=True)
+
+    def func_tokenize2(text: str):
+        return tokenizer.encode(text, add_special_tokens=True)
 
     vocab = list(sorted(tokenizer.batch_decode(list(tokenizer.get_vocab().values()), skip_special_tokens=True)))
     test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text())
     test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_custom_text_edge_cases())
-    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_special_tokens(tokenizer.all_special_tokens, 10_000))
+    test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_special_tokens(tokenizer, 10_000))
     test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_vocab_words(vocab))
     test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_chars(10_000))
     test_compare_tokenizer(func_tokenize1, func_tokenize2, generator_random_vocab_chars(vocab, 10_000))