metal : correctly set SIMD support flags on iOS (#4923)

* Correctly set support_simdgroup_reduction and support_simdgroup_mm on iPhone/iPad

* log a little bit more info on iOS

.github/workflows/nix-ci-aarch64.yml

@@ -0,0 +1,55 @@
+name: Nix aarch64 builds
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/**', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix']
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', '**/*.sh', '**/*.py', '**/*.nix']
+
+jobs:
+  nix-build-aarch64:
+    if: ${{ vars.CACHIX_NAME != '' }}
+    runs-on: ubuntu-latest
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+    - name: Install QEMU
+      # Copy-paste from https://github.com/orgs/community/discussions/8305#discussioncomment-5888654
+      run: |
+        sudo apt-get update
+        sudo apt-get install -y qemu-user-static qemu-system-aarch64
+        sudo usermod -a -G kvm $USER
+    - name: Install Nix
+      uses: DeterminateSystems/nix-installer-action@v9
+      with:
+        github-token: ${{ secrets.GITHUB_TOKEN }}
+        extra-conf: |
+          extra-platforms = aarch64-linux
+          extra-system-features = nixos-test kvm
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+    - uses: DeterminateSystems/magic-nix-cache-action@v2
+      with:
+        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
+    - name: Set-up cachix to push the results to
+      uses: cachix/cachix-action@v13
+      with:
+        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
+        name: ${{ vars.CACHIX_NAME }}
+    - name: Show all output paths
+      run: >
+          nix run github:nix-community/nix-eval-jobs
+          -- --gc-roots-dir gcroot
+          --flake
+          ".#packages.aarch64-linux"
+    - name: Build
+      run: >
+          nix run github:Mic92/nix-fast-build
+          -- --skip-cached --no-nom
+          --systems aarch64-linux
+          --flake
+          ".#checks.aarch64-linux"

.github/workflows/nix-ci.yml

@@ -69,44 +69,3 @@ jobs:
           -- --skip-cached --no-nom
           --flake
           ".#checks.$(nix eval --raw --impure --expr builtins.currentSystem)"
-  nix-build-aarch64:
-    if: ${{ vars.CACHIX_NAME != '' }}
-    runs-on: ubuntu-latest
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v4
-    - name: Install QEMU
-      # Copy-paste from https://github.com/orgs/community/discussions/8305#discussioncomment-5888654
-      run: |
-        sudo apt-get install -y qemu-user-static qemu-system-aarch64
-        sudo usermod -a -G kvm $USER
-    - name: Install Nix
-      uses: DeterminateSystems/nix-installer-action@v9
-      with:
-        github-token: ${{ secrets.GITHUB_TOKEN }}
-        extra-conf: |
-          extra-platforms = aarch64-linux
-          extra-system-features = nixos-test kvm
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
-    - uses: DeterminateSystems/magic-nix-cache-action@v2
-      with:
-        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
-    - name: Set-up cachix to push the results to
-      uses: cachix/cachix-action@v13
-      with:
-        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
-    - name: Show all output paths
-      run: >
-          nix run github:nix-community/nix-eval-jobs
-          -- --gc-roots-dir gcroot
-          --flake
-          ".#packages.aarch64-linux"
-    - name: Build
-      run: >
-          nix run github:Mic92/nix-fast-build
-          -- --skip-cached --no-nom
-          --systems aarch64-linux
-          --flake
-          ".#checks.aarch64-linux"

CMakeLists.txt

@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.13)  # for add_link_options
+cmake_minimum_required(VERSION 3.14)  # for add_link_options and implicit target directories.
 project("llama.cpp" C CXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
@@ -76,6 +76,10 @@ if (NOT MSVC)
     option(LLAMA_F16C                        "llama: enable F16C"                               ${INS_ENB})
 endif()
 
+if (WIN32)
+    option(LLAMA_WIN_VER                     "llama: Windows Version"                           0x602)
+endif()
+
 # 3rd party libs
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
@@ -686,7 +690,7 @@ endif()
 
 if (MINGW)
     # Target Windows 8 for PrefetchVirtualMemory
-    add_compile_definitions(_WIN32_WINNT=0x602)
+    add_compile_definitions(_WIN32_WINNT=${LLAMA_WIN_VER})
 endif()
 
 #

Makefile

@@ -43,10 +43,6 @@ ifeq ($(UNAME_S),Darwin)
 	endif
 endif
 
-ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
-BUILD_TARGETS += metal
-endif
-
 default: $(BUILD_TARGETS)
 
 test: $(TEST_TARGETS)
@@ -671,11 +667,6 @@ lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-ifdef LLAMA_METAL
-metal: examples/metal/metal.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-endif
-
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
 	(cd examples/batched.swift; make build)

common/common.cpp

@@ -617,6 +617,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             params.numa = true;
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
+        } else if (arg == "--no-display-prompt") {
+            params.display_prompt = false;
         } else if (arg == "-r" || arg == "--reverse-prompt") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -936,11 +938,12 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -mg i, --main-gpu i   the GPU to use for the model (with split-mode = none),\n");
     printf("                        or for intermediate results and KV (with split-mode = row) (default: %d)\n", params.main_gpu);
 #endif
+    printf("  --verbose-prompt      print a verbose prompt before generation (default: %s)\n", params.verbose_prompt ? "true" : "false");
+    printf("  --no-display-prompt   don't print prompt at generation (default: %s)\n", !params.display_prompt ? "true" : "false");
     printf("  -gan N, --grp-attn-n N\n");
     printf("                        group-attention factor (default: %d)\n", params.grp_attn_n);
     printf("  -gaw N, --grp-attn-w N\n");
     printf("                        group-attention width (default: %.1f)\n", (double)params.grp_attn_w);
-    printf("  --verbose-prompt      print prompt before generation\n");
     printf("  -dkvc, --dump-kv-cache\n");
     printf("                        verbose print of the KV cache\n");
     printf("  -nkvo, --no-kv-offload\n");
@@ -1582,6 +1585,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
     fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
+    fprintf(stream, "display_prompt: %s # default: true\n", params.display_prompt ? "true" : "false");
 }
 
 //

common/common.h

@@ -126,6 +126,7 @@ struct gpt_params {
     bool use_mlock         = false; // use mlock to keep model in memory
     bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
+    bool display_prompt    = true;  // print prompt before generation
     bool infill            = false; // use infill mode
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading

convert-hf-to-gguf.py

@@ -23,6 +23,15 @@ if 'NO_LOCAL_GGUF' not in os.environ:
 import gguf
 
 
+# check for any of the given keys in the dictionary and return the value of the first key found
+def get_key_opts(d, keys):
+    for k in keys:
+        if k in d:
+            return d[k]
+    print(f"Could not find any of {keys}")
+    sys.exit()
+
+
 ###### MODEL DEFINITIONS ######
 
 class SentencePieceTokenTypes(IntEnum):
@@ -257,10 +266,11 @@ class Model:
                 toktypes.append(gguf.TokenType.USER_DEFINED)
             elif reverse_vocab[i] in added_vocab:
                 tokens.append(reverse_vocab[i])
-                if tokenizer.added_tokens_decoder[i].special:
-                    toktypes.append(gguf.TokenType.CONTROL)
-                else:
-                    toktypes.append(gguf.TokenType.USER_DEFINED)
+                if hasattr(tokenizer, "added_tokens_decoder"):
+                    if tokenizer.added_tokens_decoder[i].special:
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
             else:
                 tokens.append(reverse_vocab[i])
                 toktypes.append(gguf.TokenType.NORMAL)
@@ -1068,17 +1078,22 @@ class GPT2Model(Model):
 
 class Phi2Model(Model):
     def set_gguf_parameters(self):
-        block_count = self.hparams["n_layer"]
+        block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"])
+
+        rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
+        n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"])
+        n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"])
 
         self.gguf_writer.add_name("Phi2")
-        self.gguf_writer.add_context_length(self.hparams["n_positions"])
-        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
-        self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
+        self.gguf_writer.add_context_length(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"]))
+
+        self.gguf_writer.add_embedding_length(n_embd)
+        self.gguf_writer.add_feed_forward_length(4 * n_embd)
         self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_head_count(self.hparams["n_head"])
-        self.gguf_writer.add_head_count_kv(self.hparams["n_head"])
-        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
-        self.gguf_writer.add_rope_dimension_count(self.hparams["rotary_dim"])
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head)
+        self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"]))
+        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_add_bos_token(False)
 

examples/CMakeLists.txt

@@ -37,9 +37,6 @@ else()
     add_subdirectory(lookup)
     add_subdirectory(train-text-from-scratch)
     add_subdirectory(imatrix)
-    if (LLAMA_METAL)
-        add_subdirectory(metal)
-    endif()
     if (LLAMA_BUILD_SERVER)
         add_subdirectory(server)
     endif()

examples/benchmark/benchmark-matmult.cpp

@@ -194,7 +194,7 @@ int main(int argc, char ** argv)  {
     // Set up a the benchmark matrices
     // printf("Creating new tensor q11 & Running quantize\n");
     struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements, hist_cur.data());
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], hist_cur.data(), nullptr);
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
@@ -207,7 +207,7 @@ int main(int argc, char ** argv)  {
     // Set up a second graph computation to make sure we override the CPU cache lines
     // printf("Creating new tensor q12 & Running quantize\n");
     struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements, hist_cur.data());
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], hist_cur.data(), nullptr);
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);

examples/main/main.cpp

@@ -477,6 +477,7 @@ int main(int argc, char ** argv) {
 
     bool is_antiprompt        = false;
     bool input_echo           = true;
+    bool display              = true;
     bool need_to_save_session = !path_session.empty() && n_matching_session_tokens < embd_inp.size();
 
     int n_past             = 0;
@@ -491,6 +492,7 @@ int main(int argc, char ** argv) {
 
     // the first thing we will do is to output the prompt, so set color accordingly
     console::set_display(console::prompt);
+    display = params.display_prompt;
 
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
@@ -707,7 +709,7 @@ int main(int argc, char ** argv) {
         }
 
         // display text
-        if (input_echo) {
+        if (input_echo && display) {
             for (auto id : embd) {
                 const std::string token_str = llama_token_to_piece(ctx, id);
                 printf("%s", token_str.c_str());
@@ -724,6 +726,7 @@ int main(int argc, char ** argv) {
         // reset color to default if there is no pending user input
         if (input_echo && (int) embd_inp.size() == n_consumed) {
             console::set_display(console::reset);
+            display = true;
         }
 
         // if not currently processing queued inputs;
@@ -796,6 +799,7 @@ int main(int argc, char ** argv) {
 
                 // color user input only
                 console::set_display(console::user_input);
+                display = params.display_prompt;
 
                 std::string line;
                 bool another_line = true;
@@ -806,6 +810,7 @@ int main(int argc, char ** argv) {
 
                 // done taking input, reset color
                 console::set_display(console::reset);
+                display = true;
 
                 // Add tokens to embd only if the input buffer is non-empty
                 // Entering a empty line lets the user pass control back

examples/metal/CMakeLists.txt

@@ -1,4 +0,0 @@
-set(TEST_TARGET metal)
-add_executable(${TEST_TARGET} metal.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TEST_TARGET} PRIVATE ggml)

examples/metal/metal.cpp

@@ -1,103 +0,0 @@
-// Evaluate a statically exported ggml computation graph with Metal
-//
-// - First, export a LLaMA graph:
-//
-//  $ ./bin/main -m ../models/7B/ggml-model-q4_0.gguf --export
-//
-// - Run this tool to evaluate the exported graph:
-//
-//  $ ./bin/metal llama.ggml
-//
-// The purpose of this tool is mostly for debugging and demonstration purposes.
-// The main limitation of exporting computation graphs is that their sizes are static which often
-// can be a problem for real-world applications.
-//
-
-#include "ggml.h"
-#include "ggml-metal.h"
-
-#include <cstdio>
-#include <cstring>
-#include <cstdlib>
-
-int main(int argc, char ** argv) {
-    ggml_time_init();
-
-    if (argc != 2) {
-        fprintf(stderr, "Usage: %s llama.ggml\n", argv[0]);
-        return -1;
-    }
-
-    const char * fname_cgraph = argv[1];
-
-    // load the compute graph
-    struct ggml_context * ctx_data = NULL;
-    struct ggml_context * ctx_eval = NULL;
-
-    struct ggml_cgraph * gf = ggml_graph_import(fname_cgraph, &ctx_data, &ctx_eval);
-
-    // this allocates all Metal resources and memory buffers
-    auto * ctx_metal = ggml_metal_init(1);
-
-    const size_t max_size_data = ggml_get_max_tensor_size(ctx_data);
-    const size_t max_size_eval = ggml_get_max_tensor_size(ctx_eval);
-    ggml_metal_add_buffer(ctx_metal, "data", ggml_get_mem_buffer(ctx_data), ggml_get_mem_size(ctx_data), max_size_data);
-    ggml_metal_add_buffer(ctx_metal, "eval", ggml_get_mem_buffer(ctx_eval), ggml_get_mem_size(ctx_eval), max_size_eval);
-
-    // main
-    {
-        struct ggml_tensor * input = ggml_graph_get_tensor(gf, "embd");
-        *(int32_t *) input->data = 1; // BOS
-
-        ggml_metal_set_tensor(ctx_metal, input);
-
-        // warmup
-        ggml_metal_graph_compute(ctx_metal, gf);
-
-        const int n_iter = 16;
-
-        const int64_t t0 = ggml_time_us();
-
-        // the actual inference happens here
-        for (int i = 0; i < n_iter; ++i) {
-            ggml_metal_graph_compute(ctx_metal, gf);
-        }
-
-        const int64_t t1 = ggml_time_us();
-
-        printf("time: %.2f ms, %.2f ms/tok\n", (t1 - t0) / 1000.0, (t1 - t0) / 1000.0 / n_iter);
-    }
-
-    // debug output
-    {
-        struct ggml_tensor * logits = gf->nodes[gf->n_nodes - 1];
-        ggml_metal_get_tensor(ctx_metal, logits);
-
-        float * ptr = (float *) ggml_get_data(logits);
-
-        printf("logits: ");
-        for (int i = 0; i < 10; i++) {
-            printf("%8.4f ", ptr[i]);
-        }
-        printf("\n");
-        int imax = 0;
-        double sum = 0.0;
-        double vmax = -1e9;
-        for (int i = 0; i < 32000; i++) {
-            sum += (double) ptr[i];
-            if (ptr[i] > vmax) {
-                vmax = ptr[i];
-                imax = i;
-            }
-        }
-        printf("sum: %f, imax = %d, vmax = %f\n", sum, imax, vmax);
-    }
-
-    ggml_metal_free(ctx_metal);
-
-    ggml_free(ctx_data);
-    ggml_free(ctx_eval);
-
-    return 0;
-}
-

examples/quantize/quantize.cpp

@@ -5,6 +5,10 @@
 #include <cstring>
 #include <vector>
 #include <string>
+#include <unordered_map>
+#include <fstream>
+#include <cmath>
+#include <algorithm>
 
 struct quant_option {
     std::string name;
@@ -17,6 +21,8 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q4_1",   LLAMA_FTYPE_MOSTLY_Q4_1,   " 3.90G, +0.1585 ppl @ LLaMA-v1-7B", },
     { "Q5_0",   LLAMA_FTYPE_MOSTLY_Q5_0,   " 4.33G, +0.0683 ppl @ LLaMA-v1-7B", },
     { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
+    { "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization",            },
+    { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization",            },
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
@@ -72,10 +78,14 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
     printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
+    printf("  --imatrixfile_name: use data in file_name as importance matrix for quant optimizations\n");
+    printf("  --include-weights tensor_name: use importance matrix for this/these tensor(s)\n");
+    printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
+    printf("Note: --include-weights and --exclude-weights cannot be used together\n");
     printf("\nAllowed quantization types:\n");
     for (auto & it : QUANT_OPTIONS) {
         if (it.name != "COPY") {
@@ -83,11 +93,93 @@ static void usage(const char * executable) {
         } else {
             printf("          ");
         }
-        printf("%-6s : %s\n", it.name.c_str(), it.desc.c_str());
+        printf("%-7s : %s\n", it.name.c_str(), it.desc.c_str());
     }
     exit(1);
 }
 
+static void load_imatrix(const std::string& imatrix_file, std::unordered_map<std::string, std::vector<float>>& imatrix_data) {
+    std::ifstream in(imatrix_file.c_str(), std::ios::binary);
+    if (!in) {
+        printf("%s: failed to open %s\n",__func__,imatrix_file.c_str());
+        return;
+    }
+    int n_entries;
+    in.read((char*)&n_entries, sizeof(n_entries));
+    if (in.fail() || n_entries < 1) {
+        printf("%s: no data in file %s\n", __func__, imatrix_file.c_str());
+        return;
+    }
+    for (int i = 0; i < n_entries; ++i) {
+        int len; in.read((char *)&len, sizeof(len));
+        std::vector<char> name_as_vec(len+1);
+        in.read((char *)name_as_vec.data(), len);
+        if (in.fail()) {
+            printf("%s: failed reading name for entry %d from %s\n",__func__,i+1,imatrix_file.c_str());
+            return;
+        }
+        name_as_vec[len] = 0;
+        std::string name{name_as_vec.data()};
+        auto& e = imatrix_data[std::move(name)];
+        int ncall;
+        in.read((char*)&ncall, sizeof(ncall));
+        int nval;
+        in.read((char *)&nval, sizeof(nval));
+        if (in.fail() || nval < 1) {
+            printf("%s: failed reading number of values for entry %d\n",__func__,i);
+            imatrix_data = {};
+            return;
+        }
+        e.resize(nval);
+        in.read((char*)e.data(), nval*sizeof(float));
+        if (in.fail()) {
+            printf("%s: failed reading data for entry %d\n",__func__,i);
+            imatrix_data = {};
+            return;
+        }
+        if (ncall > 0) {
+            for (auto& v : e) v /= ncall;
+        }
+    }
+    printf("%s: loaded %d importance matrix entries from %s\n",__func__,int(imatrix_data.size()),imatrix_file.c_str());
+}
+
+static void prepare_imatrix(const std::string& imatrix_file,
+        const std::vector<std::string>& included_weights,
+        const std::vector<std::string>& excluded_weights,
+        std::unordered_map<std::string, std::vector<float>>& imatrix_data) {
+    if (!imatrix_file.empty()) {
+        load_imatrix(imatrix_file, imatrix_data);
+    }
+    if (imatrix_data.empty()) {
+        return;
+    }
+    if (!excluded_weights.empty()) {
+        for (auto& name : excluded_weights) {
+            for (auto it = imatrix_data.begin(); it != imatrix_data.end(); ) {
+                auto pos = it->first.find(name);
+                if (pos != std::string::npos) it = imatrix_data.erase(it);
+                else ++it;
+            }
+        }
+    }
+    if (!included_weights.empty()) {
+        std::unordered_map<std::string, std::vector<float>> tmp;
+        for (auto& name : included_weights) {
+            for (auto& e : imatrix_data) {
+                auto pos = e.first.find(name);
+                if (pos != std::string::npos) {
+                    tmp.emplace(std::move(e));
+                }
+            }
+        }
+        imatrix_data = std::move(tmp);
+    }
+    if (!imatrix_data.empty()) {
+        printf("%s: have %d importance matrix entries\n", __func__, int(imatrix_data.size()));
+    }
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -96,6 +188,8 @@ int main(int argc, char ** argv) {
     llama_model_quantize_params params = llama_model_quantize_default_params();
 
     int arg_idx = 1;
+    std::string imatrix_file;
+    std::vector<std::string> included_weights, excluded_weights;
 
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
@@ -104,14 +198,42 @@ int main(int argc, char ** argv) {
             params.allow_requantize = true;
         } else if (strcmp(argv[arg_idx], "--pure") == 0) {
             params.pure = true;
+        } else if (strcmp(argv[arg_idx], "--imatrix") == 0) {
+            if (arg_idx < argc-1) {
+                imatrix_file = argv[++arg_idx];
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--include-weights") == 0) {
+            if (arg_idx < argc-1) {
+                included_weights.push_back(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--exclude-weights") == 0) {
+            if (arg_idx < argc-1) {
+                excluded_weights.push_back(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
         } else {
             usage(argv[0]);
         }
     }
 
     if (argc - arg_idx < 2) {
+        printf("%s: bad arguments\n", argv[0]);
         usage(argv[0]);
     }
+    if (!included_weights.empty() && !excluded_weights.empty()) {
+        usage(argv[0]);
+    }
+
+    std::unordered_map<std::string, std::vector<float>> imatrix_data;
+    prepare_imatrix(imatrix_file, included_weights, excluded_weights, imatrix_data);
+    if (!imatrix_data.empty()) {
+        params.imatrix = &imatrix_data;
+    }
 
     llama_backend_init(false);
 
@@ -163,6 +285,13 @@ int main(int argc, char ** argv) {
         }
     }
 
+    if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) && imatrix_data.empty()) {
+        fprintf(stderr, "\n===============================================================================================\n");
+        fprintf(stderr, "Please do not use IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
+        fprintf(stderr, "===============================================================================================\n\n\n");
+        return 1;
+    }
+
     print_build_info();
 
     fprintf(stderr, "%s: quantizing '%s' to '%s' as %s", __func__, fname_inp.c_str(), fname_out.c_str(), ftype_str.c_str());

examples/save-load-state/save-load-state.cpp

@@ -45,13 +45,13 @@ int main(int argc, char ** argv) {
     // save state (rng, logits, embedding and kv_cache) to file
     {
         std::vector<uint8_t> state_mem(llama_get_state_size(ctx));
+        const size_t written = llama_copy_state_data(ctx, state_mem.data());
 
-        {
-            FILE *fp_write = fopen("dump_state.bin", "wb");
-            llama_copy_state_data(ctx, state_mem.data()); // could also copy directly to memory mapped file
-            fwrite(state_mem.data(), 1, state_mem.size(), fp_write);
-            fclose(fp_write);
-        }
+        FILE *fp_write = fopen("dump_state.bin", "wb");
+        fwrite(state_mem.data(), 1, written, fp_write);
+        fclose(fp_write);
+
+        fprintf(stderr, "%s : serialized state into %zd out of a maximum of %zd bytes\n", __func__, written, state_mem.size());
     }
 
     // save state (last tokens)
@@ -100,18 +100,17 @@ int main(int argc, char ** argv) {
         std::vector<uint8_t> state_mem(llama_get_state_size(ctx2));
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
+        const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
+        fclose(fp_read);
 
-        const size_t ret = fread(state_mem.data(), 1, state_mem.size(), fp_read);
-        if (ret != state_mem.size()) {
+        if (read != llama_set_state_data(ctx2, state_mem.data())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
             return 1;
         }
 
-        llama_set_state_data(ctx2, state_mem.data());
-
-        fclose(fp_read);
+        fprintf(stderr, "%s : deserialized state from %zd out of a maximum of %zd bytes\n", __func__, read, state_mem.size());
     }
 
     // restore state (last tokens)

examples/server/server.cpp

@@ -1180,8 +1180,9 @@ struct llama_server_context
         return slot.images.size() > 0;
     }
 
-    void send_error(task_server& task, std::string error)
+    void send_error(task_server& task, const std::string &error)
     {
+        LOG_TEE("task %i - error: %s\n", task.id, error.c_str());
         std::unique_lock<std::mutex> lock(mutex_results);
         task_result res;
         res.id = task.id;
@@ -1350,14 +1351,17 @@ struct llama_server_context
             res.result_json["model"] = slot.oaicompat_model;
         }
 
+        queue_results.push_back(res);
+        condition_results.notify_all();
+
+        // done with results, unlock
+        lock.unlock();
+
         // parent multitask, if any, needs to be updated
         if (slot.multitask_id != -1)
         {
             update_multi_task(slot.multitask_id, slot.task_id, res);
         }
-
-        queue_results.push_back(res);
-        condition_results.notify_all();
     }
 
     void send_embedding(llama_client_slot &slot)
@@ -1567,12 +1571,22 @@ struct llama_server_context
                         LOG_TEE("slot unavailable\n");
                         // send error result
                         send_error(task, "slot unavailable");
-                        return;
+                        break;
                     }
 
                     if (task.data.contains("system_prompt"))
                     {
+                        if (!all_slots_are_idle) {
+                            send_error(task, "system prompt can only be updated when all slots are idle");
+                            break;
+                        }
                         process_system_prompt_data(task.data["system_prompt"]);
+
+                        // reset cache_tokens for all slots
+                        for (llama_client_slot &slot : slots)
+                        {
+                            slot.cache_tokens.clear();
+                        }
                     }
 
                     slot->reset();
@@ -1603,6 +1617,7 @@ struct llama_server_context
         }
 
         // remove finished multitasks from the queue of multitasks, and add the corresponding result to the result queue
+        std::vector<task_result> agg_results;
         auto queue_iterator = queue_multitasks.begin();
         while (queue_iterator != queue_multitasks.end())
         {
@@ -1623,8 +1638,9 @@ struct llama_server_context
                 }
                 aggregate_result.result_json = json{ "results", result_jsons };
 
-                std::lock_guard<std::mutex> lock(mutex_results);
-                queue_results.push_back(aggregate_result);
+
+                agg_results.push_back(aggregate_result);
+
                 condition_results.notify_all();
 
                 queue_iterator = queue_multitasks.erase(queue_iterator);
@@ -1634,14 +1650,20 @@ struct llama_server_context
                 ++queue_iterator;
             }
         }
+
+        // done with tasks, unlock
+        lock.unlock();
+
+        // copy aggregate results of complete multi-tasks to the results queue
+        std::lock_guard<std::mutex> lock_results(mutex_results);
+        queue_results.insert(queue_results.end(), agg_results.begin(), agg_results.end());
     }
 
     bool update_slots() {
         // attend tasks
         process_tasks();
 
-        // update the system prompt wait until all slots are idle state
-        if (system_need_update && all_slots_are_idle)
+        if (system_need_update)
         {
             LOG_TEE("updating system prompt\n");
             update_system_prompt();
@@ -1835,7 +1857,7 @@ struct llama_server_context
 
                     slot.cache_tokens = prompt_tokens;
 
-                    if (slot.n_past == slot.num_prompt_tokens)
+                    if (slot.n_past == slot.num_prompt_tokens && slot.n_past > 0)
                     {
                         // we have to evaluate at least 1 token to generate logits.
                         LOG_TEE("slot %d : we have to evaluate at least 1 token to generate logits\n", slot.id);

ggml-metal.h

@@ -36,64 +36,13 @@ struct ggml_cgraph;
 extern "C" {
 #endif
 
-//
-// internal API
-// temporary exposed to user-code
-//
-
-struct ggml_metal_context;
-
-void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
-
-// number of command buffers to use
-struct ggml_metal_context * ggml_metal_init(int n_cb);
-void ggml_metal_free(struct ggml_metal_context * ctx);
-
-void * ggml_metal_host_malloc(size_t n);
-void   ggml_metal_host_free  (void * data);
-
-// set the number of command buffers to use
-void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
-
-// creates a mapping between a host memory buffer and a device memory buffer
-// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
-// - the mapping is used during computation to determine the arguments of the compute kernels
-// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
-// - max_size specifies the maximum size of a tensor and is used to create shared views such
-//   that it is guaranteed that the tensor will fit in at least one of the views
-//
-bool ggml_metal_add_buffer(
-        struct ggml_metal_context * ctx,
-                       const char * name,
-                             void * data,
-                           size_t   size,
-                           size_t   max_size);
-
-// set data from host memory into the device
-void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
-
-// get data from the device into host memory
-void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
-
-// try to find operations that can be run concurrently in the graph
-// you should run it again if the topology of your graph changes
-void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf, bool check_mem);
-
-// if the graph has been optimized for concurrently dispatch, return length of the concur_list if optimized
-int ggml_metal_if_optimized(struct ggml_metal_context * ctx);
-
-// output the concur_list for ggml_alloc
-int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
-
-// same as ggml_graph_compute but uses Metal
-// creates gf->n_threads command buffers in parallel
-bool ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
-
 //
 // backend API
 // user-code should use only these functions
 //
 
+GGML_API void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
+
 GGML_API ggml_backend_t ggml_backend_metal_init(void);
 
 GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);

ggml-quants.c

@@ -5,6 +5,8 @@
 #include <string.h>
 #include <assert.h>
 #include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
 
 #ifdef __ARM_NEON
 
@@ -1639,6 +1641,241 @@ size_t ggml_quantize_q2_K(const float * restrict src, void * restrict dst, int n
     return (n/QK_K*sizeof(block_q2_K));
 }
 
+static float make_qkx3_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
+        uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
+        float rmin, float rdelta, int nstep, bool use_mad) {
+    float min = x[0];
+    float max = x[0];
+    float sum_w = weights ? weights[0] : x[0]*x[0];
+    float sum_x = sum_w * x[0];
+    for (int i = 1; i < n; ++i) {
+        if (x[i] < min) min = x[i];
+        if (x[i] > max) max = x[i];
+        float w = weights ? weights[i] : x[i]*x[i];
+        sum_w += w;
+        sum_x += w * x[i];
+    }
+    if (min > 0) {
+        min = 0;
+    }
+    if (max <= min) {
+        for (int i = 0; i < n; ++i) L[i] = 0;
+        *the_min = -min;
+        return 0.f;
+    }
+    float iscale = nmax/(max - min);
+    float scale = 1/iscale;
+    float best_mad = 0;
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int(iscale*(x[i] - min));
+        L[i] = MAX(0, MIN(nmax, l));
+        float diff = scale * L[i] + min - x[i];
+        diff = use_mad ? fabsf(diff) : diff*diff;
+        float w = weights ? weights[i] : x[i]*x[i];
+        best_mad += w * diff;
+    }
+    if (nstep < 1) {
+        *the_min = -min;
+        return scale;
+    }
+    for (int is = 0; is <= nstep; ++is) {
+        iscale = (rmin + rdelta*is + nmax)/(max - min);
+        float sum_l = 0, sum_l2 = 0, sum_xl = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale*(x[i] - min));
+            l = MAX(0, MIN(nmax, l));
+            Laux[i] = l;
+            float w = weights ? weights[i] : x[i]*x[i];
+            sum_l  += w*l;
+            sum_l2 += w*l*l;
+            sum_xl += w*l*x[i];
+        }
+        float D = sum_w * sum_l2 - sum_l * sum_l;
+        if (D > 0) {
+            float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
+            float this_min   = (sum_l2 * sum_x - sum_l * sum_xl)/D;
+            if (this_min > 0) {
+                this_min = 0;
+                this_scale = sum_xl / sum_l2;
+            }
+            float mad = 0;
+            for (int i = 0; i < n; ++i) {
+                float diff = this_scale * Laux[i] + this_min - x[i];
+                diff = use_mad ? fabsf(diff) : diff*diff;
+                float w = weights ? weights[i] : x[i]*x[i];
+                mad += w * diff;
+            }
+            if (mad < best_mad) {
+                for (int i = 0; i < n; ++i) {
+                    L[i] = Laux[i];
+                }
+                best_mad = mad;
+                scale = this_scale;
+                min = this_min;
+            }
+        }
+    }
+    *the_min = -min;
+    return scale;
+}
+
+static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, const float * quant_weights) {
+    float max = 0;
+    for (int i = 0; i < n; ++i) {
+        max = MAX(max, x[i]);
+    }
+    if (!max) { // all zero
+        for (int i = 0; i < n; ++i) { L[i] = 0; }
+        return 0.f;
+    }
+    float iscale = nmax / max;
+    for (int i = 0; i < n; ++i) {
+        L[i] = nearest_int(iscale * x[i]);
+    }
+    float scale = 1/iscale;
+    float best_mse = 0;
+    for (int i = 0; i < n; ++i) {
+        float diff = x[i] - scale*L[i];
+        float w = quant_weights[i];
+        best_mse += w*diff*diff;
+    }
+    for (int is = -4; is <= 4; ++is) {
+        if (is == 0) continue;
+        float iscale_is = (0.1f*is + nmax)/max;
+        float scale_is = 1/iscale_is;
+        float mse = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale_is*x[i]);
+            l = MIN(nmax, l);
+            float diff = x[i] - scale_is*l;
+            float w = quant_weights[i];
+            mse += w*diff*diff;
+        }
+        if (mse < best_mse) {
+            best_mse = mse;
+            iscale = iscale_is;
+        }
+    }
+    float sumlx = 0;
+    float suml2 = 0;
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int(iscale * x[i]);
+        l = MIN(nmax, l);
+        L[i] = l;
+        float w = quant_weights[i];
+        sumlx += w*x[i]*l;
+        suml2 += w*l*l;
+    }
+    for (int itry = 0; itry < 5; ++itry) {
+        int n_changed = 0;
+        for (int i = 0; i < n; ++i) {
+            float w = quant_weights[i];
+            float slx = sumlx - w*x[i]*L[i];
+            float sl2 = suml2 - w*L[i]*L[i];
+            if (slx > 0 && sl2 > 0) {
+                int new_l = nearest_int(x[i] * sl2 / slx);
+                new_l = MIN(nmax, new_l);
+                if (new_l != L[i]) {
+                    slx += w*x[i]*new_l;
+                    sl2 += w*new_l*new_l;
+                    if (slx*slx*suml2 > sumlx*sumlx*sl2) {
+                        L[i] = new_l; sumlx = slx; suml2 = sl2;
+                        ++n_changed;
+                    }
+                }
+            }
+        }
+        if (!n_changed) {
+            break;
+        }
+    }
+    return sumlx / suml2;
+}
+
+static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restrict y, int k, const float * restrict quant_weights) {
+    GGML_ASSERT(quant_weights);
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+    const bool requantize = true;
+
+    uint8_t L[QK_K];
+    uint8_t Laux[16];
+    float mins[QK_K/16];
+    float scales[QK_K/16];
+    float sw[QK_K/16];
+    float weight[QK_K/16];
+    uint8_t Ls[QK_K/16], Lm[QK_K/16];
+
+    for (int i = 0; i < nb; i++) {
+        memset(sw, 0, QK_K/16*sizeof(float));
+        float sumx2 = 0;
+        for (int j = 0; j < QK_K; ++j) sumx2 += x[j]*x[j];
+        float sigma2 = sumx2/QK_K;
+        for (int j = 0; j < QK_K/16; ++j) {
+            const float * restrict qw = quant_weights + QK_K * i + 16*j;
+            for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j + l]*x[16*j + l]);
+            for (int l = 0; l < 16; ++l) sw[j] += weight[l];
+            scales[j] = make_qkx3_quants(16, 3, x + 16*j, weight, L + 16*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
+        }
+
+        float dm  = make_qp_quants(QK_K/16, 15, scales, Ls, sw);
+        float mm  = make_qp_quants(QK_K/16, 15, mins,   Lm, sw);
+        y[i].d    = GGML_FP32_TO_FP16(dm);
+        y[i].dmin = GGML_FP32_TO_FP16(mm);
+        dm        = GGML_FP16_TO_FP32(y[i].d);
+        mm        = GGML_FP16_TO_FP32(y[i].dmin);
+
+        for (int j = 0; j < QK_K/16; ++j) {
+            y[i].scales[j] = Ls[j] | (Lm[j] << 4);
+        }
+
+        if (requantize) {
+            for (int j = 0; j < QK_K/16; ++j) {
+                const float d = dm * (y[i].scales[j] & 0xF);
+                if (!d) continue;
+                const float m = mm * (y[i].scales[j] >> 4);
+                for (int ii = 0; ii < 16; ++ii) {
+                    int l = nearest_int((x[16*j + ii] + m)/d);
+                    l = MAX(0, MIN(3, l));
+                    L[16*j + ii] = l;
+                }
+            }
+        }
+
+#if QK_K == 256
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
+            }
+        }
+#else
+        for (int l = 0; l < 16; ++l) {
+            y[i].qs[l] = L[l] | (L[l + 16] << 2) | (L[l + 32] << 4) | (L[l + 48] << 6);
+        }
+#endif
+
+        x += QK_K;
+
+    }
+}
+
+size_t quantize_q2_K(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
+    (void)hist;
+    int row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row);
+    if (!quant_weights) {
+        quantize_row_q2_K_reference(src, dst, nrow*n_per_row);
+    }
+    else {
+        char * qrow = (char *)dst;
+        for (int row = 0; row < nrow; ++row) {
+            quantize_row_q2_K_impl(src, (block_q2_K*)qrow, n_per_row, quant_weights);
+            src += n_per_row;
+            qrow += row_size;
+        }
+    }
+    return nrow * row_size;
+}
+
 //========================= 3-bit (de)-quantization
 
 void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k) {
@@ -2584,14 +2821,6 @@ static const uint8_t ksigns_iq2xs[128] = {
 
 static const uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128};
 
-void quantize_row_iq2_xxs_reference(const float * restrict x, block_iq2_xxs * restrict y, int k) {
-    (void)x;
-    (void)y;
-    (void)k;
-    assert(k % QK_K == 0);
-    //fprintf(stderr, "=========================== %s: not implemented\n", __func__);
-}
-
 void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
@@ -2618,33 +2847,8 @@ void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y
     }
 }
 
-void quantize_row_iq2_xxs(const float * restrict x, void * restrict vy, int k) {
-    assert(k % QK_K == 0);
-    block_iq2_xxs * restrict y = vy;
-    quantize_row_iq2_xxs_reference(x, y, k);
-}
-
-size_t ggml_quantize_iq2_xxs(const float * src, void * dst, int n, int k, int64_t * hist) {
-    assert(k % QK_K == 0);
-    (void)hist; // TODO: collect histograms
-
-    for (int j = 0; j < n; j += k) {
-        block_iq2_xxs * restrict y = (block_iq2_xxs *)dst + j/QK_K;
-        quantize_row_iq2_xxs_reference(src + j, y, k);
-    }
-    return (n/QK_K*sizeof(block_iq2_xxs));
-}
-
 // ====================== 2.3125 bpw (de)-quantization
 
-void quantize_row_iq2_xs_reference(const float * restrict x, block_iq2_xs * restrict y, int k) {
-    (void)x;
-    (void)y;
-    (void)k;
-    assert(k % QK_K == 0);
-    //fprintf(stderr, "=========================== %s: not implemented\n", __func__);
-}
-
 void dequantize_row_iq2_xs(const block_iq2_xs * restrict x, float * restrict y, int k) {
     assert(k % QK_K == 0);
     const int nb = k / QK_K;
@@ -2670,23 +2874,6 @@ void dequantize_row_iq2_xs(const block_iq2_xs * restrict x, float * restrict y,
     }
 }
 
-void quantize_row_iq2_xs(const float * restrict x, void * restrict vy, int k) {
-    assert(k % QK_K == 0);
-    block_iq2_xs * restrict y = vy;
-    quantize_row_iq2_xs_reference(x, y, k);
-}
-
-size_t ggml_quantize_iq2_xs(const float * src, void * dst, int n, int k, int64_t * hist) {
-    assert(k % QK_K == 0);
-    (void)hist; // TODO: collect histograms
-
-    for (int j = 0; j < n; j += k) {
-        block_iq2_xs * restrict y = (block_iq2_xs *)dst + j/QK_K;
-        quantize_row_iq2_xs_reference(src + j, y, k);
-    }
-    return (n/QK_K*sizeof(block_iq2_xs));
-}
-
 //===================================== Q8_K ==============================================
 
 void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k) {
@@ -7730,3 +7917,666 @@ void ggml_vec_dot_iq2_xs_q8_K(const int n, float * restrict s, const void * rest
     *s = 0.125f * sumf;
 #endif
 }
+
+// ================================ IQ2 quantization =============================================
+
+typedef struct {
+    uint64_t * grid;
+    int      * map;
+    uint16_t * neighbours;
+} iq2_entry_t;
+
+static iq2_entry_t iq2_data[2] = {
+    {NULL, NULL, NULL},
+    {NULL, NULL, NULL},
+};
+
+static inline int iq2_data_index(int grid_size) {
+    GGML_ASSERT(grid_size == 256 || grid_size == 512);
+    return grid_size == 256 ? 0 : 1;
+}
+
+static int iq2_compare_func(const void * left, const void * right) {
+    const int * l = (const int *)left;
+    const int * r = (const int *)right;
+    return l[0] < r[0] ? -1 : l[0] > r[0] ? 1 : l[1] < r[1] ? -1 : l[1] > r[1] ? 1 : 0;
+}
+
+static void q2xs_init_impl(int grid_size) {
+    const int gindex = iq2_data_index(grid_size);
+    if (iq2_data[gindex].grid) {
+        return;
+    }
+    static const uint16_t kgrid_256[256] = {
+            0,     2,     5,     8,    10,    17,    20,    32,    34,    40,    42,    65,    68,    80,    88,    97,
+          100,   128,   130,   138,   162,   257,   260,   272,   277,   320,   388,   408,   512,   514,   546,   642,
+         1025,  1028,  1040,  1057,  1060,  1088,  1090,  1096,  1120,  1153,  1156,  1168,  1188,  1280,  1282,  1288,
+         1312,  1350,  1385,  1408,  1425,  1545,  1552,  1600,  1668,  1700,  2048,  2053,  2056,  2068,  2088,  2113,
+         2116,  2128,  2130,  2184,  2308,  2368,  2562,  2580,  4097,  4100,  4112,  4129,  4160,  4192,  4228,  4240,
+         4245,  4352,  4360,  4384,  4432,  4442,  4480,  4644,  4677,  5120,  5128,  5152,  5157,  5193,  5248,  5400,
+         5474,  5632,  5654,  6145,  6148,  6160,  6208,  6273,  6400,  6405,  6560,  6737,  8192,  8194,  8202,  8260,
+         8289,  8320,  8322,  8489,  8520,  8704,  8706,  9217,  9220,  9232,  9280,  9302,  9472,  9537,  9572,  9872,
+        10248, 10272, 10388, 10820, 16385, 16388, 16400, 16408, 16417, 16420, 16448, 16456, 16470, 16480, 16513, 16516,
+        16528, 16640, 16672, 16737, 16768, 16773, 16897, 16912, 16968, 16982, 17000, 17408, 17416, 17440, 17536, 17561,
+        17682, 17700, 17920, 18433, 18436, 18448, 18496, 18501, 18688, 18776, 18785, 18818, 19013, 19088, 20480, 20488,
+        20497, 20505, 20512, 20608, 20616, 20740, 20802, 20900, 21137, 21648, 21650, 21770, 22017, 22100, 22528, 22545,
+        22553, 22628, 22848, 23048, 24580, 24592, 24640, 24680, 24832, 24917, 25112, 25184, 25600, 25605, 25872, 25874,
+        25988, 26690, 32768, 32770, 32778, 32833, 32898, 33028, 33048, 33088, 33297, 33793, 33796, 33808, 33813, 33856,
+        33888, 34048, 34118, 34196, 34313, 34368, 34400, 34818, 35076, 35345, 36868, 36880, 36900, 36928, 37025, 37142,
+        37248, 37445, 37888, 37922, 37956, 38225, 39041, 39200, 40962, 41040, 41093, 41225, 41472, 42008, 43088, 43268,
+    };
+    static const uint16_t kgrid_512[512] = {
+            0,     2,     5,     8,    10,    17,    20,    22,    25,    32,    34,    37,    40,    65,    68,    70,
+           73,    80,    82,    85,    88,    97,   100,   128,   130,   133,   136,   145,   148,   153,   160,   257,
+          260,   262,   265,   272,   274,   277,   280,   282,   289,   292,   320,   322,   325,   328,   337,   340,
+          352,   360,   385,   388,   400,   512,   514,   517,   520,   529,   532,   544,   577,   580,   592,   597,
+          640,   650,  1025,  1028,  1030,  1033,  1040,  1042,  1045,  1048,  1057,  1060,  1088,  1090,  1093,  1096,
+         1105,  1108,  1110,  1120,  1153,  1156,  1168,  1280,  1282,  1285,  1288,  1297,  1300,  1312,  1345,  1348,
+         1360,  1377,  1408,  1537,  1540,  1552,  1574,  1600,  1602,  1668,  2048,  2050,  2053,  2056,  2058,  2065,
+         2068,  2080,  2085,  2113,  2116,  2128,  2136,  2176,  2208,  2218,  2305,  2308,  2320,  2368,  2433,  2441,
+         2560,  2592,  2600,  2710,  2720,  4097,  4100,  4102,  4105,  4112,  4114,  4117,  4120,  4129,  4132,  4160,
+         4162,  4165,  4168,  4177,  4180,  4192,  4202,  4225,  4228,  4240,  4352,  4354,  4357,  4360,  4369,  4372,
+         4384,  4417,  4420,  4432,  4480,  4500,  4502,  4609,  4612,  4614,  4624,  4672,  4704,  5120,  5122,  5125,
+         5128,  5137,  5140,  5152,  5185,  5188,  5193,  5200,  5220,  5248,  5377,  5380,  5392,  5440,  5632,  5652,
+         5705,  6145,  6148,  6160,  6162,  6208,  6228,  6278,  6400,  6405,  6502,  6737,  6825,  8192,  8194,  8197,
+         8200,  8202,  8209,  8212,  8224,  8257,  8260,  8272,  8320,  8352,  8449,  8452,  8464,  8512,  8520,  8549,
+         8704,  8738,  8832,  8872,  9217,  9220,  9232,  9257,  9280,  9472,  9537,  9554,  9625,  9729,  9754,  9894,
+        10240, 10248, 10250, 10272, 10325, 10376, 10402, 10600, 10640, 10760, 10784, 10882, 10888, 10890, 16385, 16388,
+        16390, 16393, 16400, 16402, 16405, 16408, 16417, 16420, 16448, 16450, 16453, 16456, 16458, 16465, 16468, 16480,
+        16485, 16513, 16516, 16528, 16640, 16642, 16645, 16648, 16657, 16660, 16672, 16705, 16708, 16720, 16768, 16773,
+        16802, 16897, 16900, 16912, 16914, 16937, 16960, 17408, 17410, 17413, 17416, 17425, 17428, 17433, 17440, 17473,
+        17476, 17488, 17536, 17556, 17665, 17668, 17680, 17700, 17728, 17818, 17920, 17930, 17988, 18000, 18433, 18436,
+        18448, 18496, 18501, 18516, 18530, 18688, 18705, 18756, 18768, 18793, 18948, 20480, 20482, 20485, 20488, 20497,
+        20500, 20512, 20520, 20545, 20548, 20560, 20608, 20737, 20740, 20752, 20757, 20800, 20802, 20992, 21060, 21162,
+        21505, 21508, 21520, 21537, 21568, 21600, 21633, 21665, 21760, 21768, 21888, 21896, 22049, 22120, 22177, 22528,
+        22548, 22593, 22608, 22681, 22810, 22848, 22850, 23173, 24577, 24580, 24592, 24640, 24660, 24674, 24710, 24745,
+        24832, 25124, 25162, 25234, 25600, 25622, 25872, 25920, 25925, 26020, 26625, 26730, 26917, 27142, 27220, 27234,
+        32768, 32770, 32773, 32776, 32785, 32788, 32800, 32810, 32833, 32836, 32848, 32896, 32898, 32936, 32938, 33025,
+        33028, 33030, 33040, 33088, 33105, 33113, 33280, 33312, 33408, 33410, 33440, 33448, 33793, 33796, 33808, 33810,
+        33813, 33856, 33888, 33929, 34048, 34116, 34213, 34328, 34410, 34816, 34824, 34853, 34906, 34944, 34946, 34984,
+        35078, 35362, 35456, 35464, 35478, 35496, 36865, 36868, 36880, 36928, 36950, 36996, 37120, 37154, 37220, 37462,
+        37513, 37888, 37893, 37956, 37968, 37976, 38185, 38288, 38290, 38465, 38993, 39078, 39241, 39445, 39520, 40960,
+        40962, 40968, 40970, 40992, 41002, 41120, 41297, 41305, 41382, 41472, 41474, 41480, 41514, 41600, 41632, 42048,
+        42133, 42597, 42648, 43018, 43040, 43042, 43048, 43168, 43176, 43268, 43396, 43398, 43560, 43562, 43665, 43690,
+    };
+    const int kmap_size = 43692;
+    const int nwant = 2;
+    const uint16_t * kgrid = grid_size == 256 ? kgrid_256 : kgrid_512;
+    uint64_t * kgrid_q2xs;
+    int      * kmap_q2xs;
+    uint16_t * kneighbors_q2xs;
+
+    printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size);
+    uint64_t * the_grid = (uint64_t *)malloc(grid_size*sizeof(uint64_t));
+    for (int k = 0; k < grid_size; ++k) {
+        int8_t * pos = (int8_t *)(the_grid + k);
+        for (int i = 0; i < 8; ++i) {
+            int l = (kgrid[k] >> 2*i) & 0x3;
+            pos[i] = 2*l + 1;
+        }
+    }
+    kgrid_q2xs = the_grid;
+    iq2_data[gindex].grid = the_grid;
+    kmap_q2xs = (int *)malloc(kmap_size*sizeof(int));
+    iq2_data[gindex].map = kmap_q2xs;
+    for (int i = 0; i < kmap_size; ++i) kmap_q2xs[i] = -1;
+    uint64_t aux64;
+    uint8_t * aux8 = (uint8_t *)&aux64;
+    for (int i = 0; i < grid_size; ++i) {
+        aux64 = kgrid_q2xs[i];
+        uint16_t index = 0;
+        for (int k=0; k<8; ++k) {
+            uint16_t q = (aux8[k] - 1)/2;
+            index |= (q << 2*k);
+        }
+        kmap_q2xs[index] = i;
+    }
+    int8_t pos[8];
+    int * dist2 = (int *)malloc(2*grid_size*sizeof(int));
+    int num_neighbors = 0, num_not_in_map = 0;
+    for (int i = 0; i < kmap_size; ++i) {
+        if (kmap_q2xs[i] >= 0) continue;
+        ++num_not_in_map;
+        for (int k = 0; k < 8; ++k) {
+            int l = (i >> 2*k) & 0x3;
+            pos[k] = 2*l + 1;
+        }
+        for (int j = 0; j < grid_size; ++j) {
+            const int8_t * pg = (const int8_t *)(kgrid_q2xs + j);
+            int d2 = 0;
+            for (int k = 0; k < 8; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+            dist2[2*j+0] = d2;
+            dist2[2*j+1] = j;
+        }
+        qsort(dist2, grid_size, 2*sizeof(int), iq2_compare_func);
+        int n = 0; int d2 = dist2[0];
+        int nhave = 1;
+        for (int j = 0; j < grid_size; ++j) {
+            if (dist2[2*j] > d2) {
+                if (nhave == nwant) break;
+                d2 = dist2[2*j];
+                ++nhave;
+            }
+            ++n;
+        }
+        num_neighbors += n;
+    }
+    printf("%s: %d neighbours in total\n", __func__, num_neighbors);
+    kneighbors_q2xs = (uint16_t *)malloc((num_neighbors + num_not_in_map)*sizeof(uint16_t));
+    iq2_data[gindex].neighbours = kneighbors_q2xs;
+    int counter = 0;
+    for (int i = 0; i < kmap_size; ++i) {
+        if (kmap_q2xs[i] >= 0) continue;
+        for (int k = 0; k < 8; ++k) {
+            int l = (i >> 2*k) & 0x3;
+            pos[k] = 2*l + 1;
+        }
+        for (int j = 0; j < grid_size; ++j) {
+            const int8_t * pg = (const int8_t *)(kgrid_q2xs + j);
+            int d2 = 0;
+            for (int k = 0; k < 8; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+            dist2[2*j+0] = d2;
+            dist2[2*j+1] = j;
+        }
+        qsort(dist2, grid_size, 2*sizeof(int), iq2_compare_func);
+        kmap_q2xs[i] = -(counter + 1);
+        int d2 = dist2[0];
+        uint16_t * start = &kneighbors_q2xs[counter++];
+        int n = 0, nhave = 1;
+        for (int j = 0; j < grid_size; ++j) {
+            if (dist2[2*j] > d2) {
+                if (nhave == nwant) break;
+                d2 = dist2[2*j];
+                ++nhave;
+            }
+            kneighbors_q2xs[counter++] = dist2[2*j+1];
+            ++n;
+        }
+        *start = n;
+    }
+    free(dist2);
+}
+
+void ggml_init_iq2_quantization(enum ggml_type type) {
+    if (type == GGML_TYPE_IQ2_XXS) {
+        q2xs_init_impl(256);
+    }
+    else if (type == GGML_TYPE_IQ2_XS) {
+        q2xs_init_impl(512);
+    }
+    else {
+        fprintf(stderr, "======================== Why are you calling %s with type %d?\n", __func__, (int)type);
+    }
+}
+
+static void q2xs_deinit_impl(int grid_size) {
+    GGML_ASSERT(grid_size == 256 || grid_size == 512 || grid_size == 1024);
+    const int gindex = iq2_data_index(grid_size);
+    if (iq2_data[gindex].grid) {
+        free(iq2_data[gindex].grid);       iq2_data[gindex].grid = NULL;
+        free(iq2_data[gindex].map);        iq2_data[gindex].map  = NULL;
+        free(iq2_data[gindex].neighbours); iq2_data[gindex].neighbours = NULL;
+    }
+}
+
+void ggml_deinit_iq2_quantization(enum ggml_type type) {
+    if (type == GGML_TYPE_IQ2_XXS) {
+        q2xs_deinit_impl(256);
+    }
+    else if (type == GGML_TYPE_IQ2_XS) {
+        q2xs_deinit_impl(512);
+    }
+    else {
+        fprintf(stderr, "======================== Why are you calling %s with type %d?\n", __func__, (int)type);
+    }
+}
+
+static int iq2_find_best_neighbour(const uint16_t * restrict neighbours, const uint64_t * restrict grid,
+        const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L) {
+    int num_neighbors = neighbours[0];
+    GGML_ASSERT(num_neighbors > 0);
+    float best_d2 = FLT_MAX;
+    int grid_index = -1;
+    for (int j = 1; j <= num_neighbors; ++j) {
+        const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+        float d2 = 0;
+        for (int i = 0; i < 8; ++i) {
+            float q = pg[i];
+            float diff = scale*q - xval[i];
+            d2 += weight[i]*diff*diff;
+        }
+        if (d2 < best_d2) {
+            best_d2 = d2; grid_index = neighbours[j];
+        }
+    }
+    GGML_ASSERT(grid_index >= 0);
+    const int8_t * pg = (const int8_t *)(grid + grid_index);
+    for (int i = 0; i < 8; ++i) L[i] = (pg[i] - 1)/2;
+    return grid_index;
+}
+
+static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict vy, int n, const float * restrict quant_weights) {
+
+    const int gindex = iq2_data_index(256);
+
+    const uint64_t * kgrid_q2xs      = iq2_data[gindex].grid;
+    const int      * kmap_q2xs       = iq2_data[gindex].map;
+    const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+    GGML_ASSERT(quant_weights);
+    GGML_ASSERT(kgrid_q2xs);
+    GGML_ASSERT(kmap_q2xs);
+    GGML_ASSERT(kneighbors_q2xs);
+    GGML_ASSERT(n%QK_K == 0);
+
+    const int kMaxQ = 3;
+
+    const int nbl = n/256;
+
+    block_iq2_xxs * y = vy;
+
+    float scales[QK_K/32];
+    float weight[32];
+    float xval[32];
+    int8_t L[32];
+    int8_t Laux[32];
+    float  waux[32];
+    bool   is_on_grid[4];
+    bool   is_on_grid_aux[4];
+    uint8_t block_signs[4];
+    uint32_t q2[2*(QK_K/32)];
+
+    for (int ibl = 0; ibl < nbl; ++ibl) {
+
+        y[ibl].d = GGML_FP32_TO_FP16(0.f);
+        memset(q2, 0, QK_K/4);
+
+        float max_scale = 0;
+
+        const float * xbl = x + QK_K*ibl;
+        float sumx2 = 0;
+        for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+        float sigma2 = sumx2/QK_K;
+
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const float * xb = xbl + 32*ib;
+            const float * qw = quant_weights + QK_K*ibl + 32*ib;
+            for (int i = 0; i < 32; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            for (int i = 0; i < 32; ++i) waux[i] = sqrtf(weight[i]);
+            for (int k = 0; k < 4; ++k) {
+                int nflip = 0;
+                uint8_t s = 0;
+                for (int i = 0; i < 8; ++i) {
+                    if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+                    else {
+                        xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i);
+                    }
+                }
+                if (nflip%2) {
+                    int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin];
+                    for (int i = 1; i < 8; ++i) {
+                        float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i];
+                        if (ax < min) {
+                            min = ax; imin = i;
+                        }
+                    }
+                    xval[8*k+imin] = -xval[8*k+imin];
+                    s ^= (1 << imin);
+                }
+                block_signs[k] = s & 127;
+            }
+            float max = xval[0];
+            for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
+            if (!max) {
+                scales[ib] = 0;
+                memset(L, 0, 32);
+                continue;
+            }
+            float best = 0;
+            float scale = max/(2*kMaxQ-1);
+            for (int is = -9; is <= 9; ++is) {
+                float id = (2*kMaxQ-1+is*0.1f)/max;
+                float this_scale = 1/id;
+                for (int k = 0; k < 4; ++k) {
+                    for (int i = 0; i < 8; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+                        Laux[8*k+i] = MAX(0, MIN(kMaxQ-1, l));
+                    }
+                    uint16_t u = 0;
+                    for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i);
+                    int grid_index = kmap_q2xs[u];
+                    is_on_grid_aux[k] = true;
+                    if (grid_index < 0) {
+                        is_on_grid_aux[k] = false;
+                        const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                        grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k);
+                    }
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 32; ++i) {
+                    float w = weight[i];
+                    float q = 2*Laux[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+                    scale = sumqx/sumq2; best = scale*sumqx;
+                    for (int i = 0; i < 32; ++i) L[i] = Laux[i];
+                    for (int k = 0; k <  4; ++k) is_on_grid[k] = is_on_grid_aux[k];
+                }
+            }
+            int n_not_ongrid = 0;
+            for (int k = 0; k < 4; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
+            if (n_not_ongrid > 0 && scale > 0) {
+                float id = 1/scale;
+                for (int k = 0; k < 4; ++k) {
+                    if (is_on_grid[k]) continue;
+                    uint16_t u = 0;
+                    for (int i = 0; i < 8; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+                        l = MAX(0, MIN(kMaxQ-1, l));
+                        u |= (l << 2*i);
+                    }
+                    int grid_index = kmap_q2xs[u];
+                    if (grid_index < 0) {
+                        const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                        grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, scale, L + 8*k);
+                    }
+                    const int8_t * pg = (const int8_t *)(kgrid_q2xs + grid_index);
+                    for (int i = 0; i < 8; ++i) L[8*k+i] = (pg[i] - 1)/2;
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 32; ++i) {
+                    float w = weight[i];
+                    float q = 2*L[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0) scale = sumqx/sumq2;
+            }
+            if (scale < 0) {
+                // This should never happen, but just in case, flip scale so that it is positive (we use uint's to encode the scale)
+                // and correspondingly flip quant signs.
+                scale = -scale;
+                for (int k = 0; k < 4; ++k) block_signs[k] = (~block_signs[k]) & 127;
+            }
+            for (int k = 0; k < 4; ++k) {
+                uint16_t u = 0;
+                for (int i = 0; i < 8; ++i) u |= (L[8*k+i] << 2*i);
+                int grid_index = kmap_q2xs[u];
+                if (grid_index < 0) {
+                    printf("Oops: found point %u not on grid:", u);
+                    for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
+                    printf("\n");
+                    GGML_ASSERT(false);
+                }
+                q2[2*ib+0] |= (grid_index << 8*k);
+                q2[2*ib+1] |= (block_signs[k] << 7*k);
+            }
+            GGML_ASSERT(scale >= 0);
+            scales[ib] = scale;
+            max_scale = MAX(max_scale, scale);
+        }
+
+        if (!max_scale) {
+            memset(y[ibl].qs, 0, QK_K/4);
+            continue;
+        }
+
+        float d = max_scale/31;
+        y[ibl].d = GGML_FP32_TO_FP16(d);
+        float id = 1/d;
+        float sumqx = 0, sumq2 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            int l = nearest_int(0.5f*(id*scales[ib]-1));
+            l = MAX(0, MIN(15, l));
+            q2[2*ib+1] |= ((uint32_t)l << 28);
+            const float * xb = xbl + 32*ib;
+            const float * qw = quant_weights + QK_K*ibl + 32*ib;
+            for (int i = 0; i < 32; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            const uint8_t * aux8 = (const uint8_t *)(q2 + 2*ib);
+            const float db = d * (1 + 2*l);
+            uint32_t u = 0;
+            for (int k = 0; k < 4; ++k) {
+                const int8_t * signs = keven_signs_q2xs + 8*((q2[2*ib+1] >> 7*k) & 127);
+                const float * xk = xb + 8*k;
+                const float * wk = weight + 8*k;
+                const uint8_t * grid = (const uint8_t *)(kgrid_q2xs + aux8[k]);
+                float best_mse = 0; int best_index = aux8[k];
+                for (int j = 0; j < 8; ++j) {
+                    float diff = db * grid[j] * signs[j] - xk[j];
+                    best_mse += wk[j] * diff * diff;
+                }
+                for (int idx = 0; idx < 256; ++idx) {
+                    grid = (const uint8_t *)(kgrid_q2xs + idx);
+                    float mse = 0;
+                    for (int j = 0; j < 8; ++j) {
+                        float diff = db * grid[j] * signs[j] - xk[j];
+                        mse += wk[j] * diff * diff;
+                    }
+                    if (mse < best_mse) {
+                        best_mse = mse; best_index = idx;
+                    }
+                }
+                u |= (best_index << 8*k);
+                grid = (const uint8_t *)(kgrid_q2xs + best_index);
+                //grid = (const uint8_t *)(kgrid_q2xs + aux8[k]);
+                for (int j = 0; j < 8; ++j) {
+                    float q = db * grid[j] * signs[j];
+                    sumqx += wk[j] * q * xk[j];
+                    sumq2 += wk[j] * q * q;
+                }
+            }
+            q2[2*ib] = u;
+            if (sumq2 > 0) y[ibl].d = GGML_FP32_TO_FP16(d*sumqx/sumq2);
+        }
+        memcpy(y[ibl].qs, q2, QK_K/4);
+    }
+}
+
+static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict vy, int n, const float * restrict quant_weights) {
+
+    const int gindex = iq2_data_index(512);
+
+    const uint64_t * kgrid_q2xs      = iq2_data[gindex].grid;
+    const int      * kmap_q2xs       = iq2_data[gindex].map;
+    const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+    GGML_ASSERT(quant_weights);
+    GGML_ASSERT(kmap_q2xs);
+    GGML_ASSERT(kgrid_q2xs);
+    GGML_ASSERT(kneighbors_q2xs);
+    GGML_ASSERT(n%QK_K == 0);
+
+    const int kMaxQ = 3;
+
+    const int nbl = n/256;
+
+    block_iq2_xs * y = vy;
+
+    float scales[QK_K/16];
+    float weight[16];
+    float xval[16];
+    int8_t L[16];
+    int8_t Laux[16];
+    float  waux[16];
+    bool   is_on_grid[2];
+    bool   is_on_grid_aux[2];
+    uint8_t block_signs[2];
+    uint16_t q2[2*(QK_K/16)];
+
+    for (int ibl = 0; ibl < nbl; ++ibl) {
+
+        y[ibl].d = GGML_FP32_TO_FP16(0.f);
+        memset(q2, 0, QK_K/4);
+        memset(y[ibl].scales, 0, QK_K/32);
+
+        float max_scale = 0;
+
+        const float * xbl = x + QK_K*ibl;
+        float sumx2 = 0;
+        for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+        float sigma2 = sumx2/QK_K;
+
+        for (int ib = 0; ib < QK_K/16; ++ib) {
+            const float * xb = xbl + 16*ib;
+            const float * qw = quant_weights + QK_K*ibl + 16*ib;
+            for (int i = 0; i < 16; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            for (int i = 0; i < 16; ++i) waux[i] = sqrtf(weight[i]);
+            for (int k = 0; k < 2; ++k) {
+                int nflip = 0;
+                uint8_t s = 0;
+                for (int i = 0; i < 8; ++i) {
+                    if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+                    else {
+                        xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i);
+                    }
+                }
+                if (nflip%2) {
+                    int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin];
+                    for (int i = 1; i < 8; ++i) {
+                        float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i];
+                        if (ax < min) {
+                            min = ax; imin = i;
+                        }
+                    }
+                    xval[8*k+imin] = -xval[8*k+imin];
+                    s ^= (1 << imin);
+                }
+                block_signs[k] = s & 127;
+            }
+            float max = xval[0];
+            for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
+            if (!max) {
+                scales[ib] = 0;
+                memset(L, 0, 16);
+                continue;
+            }
+            float best = 0;
+            float scale = max/(2*kMaxQ-1);
+            is_on_grid[0] = is_on_grid[1] = true;
+            for (int is = -9; is <= 9; ++is) {
+                float id = (2*kMaxQ-1+is*0.1f)/max;
+                float this_scale = 1/id;
+                for (int k = 0; k < 2; ++k) {
+                    for (int i = 0; i < 8; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+                        Laux[8*k+i] = MAX(0, MIN(kMaxQ-1, l));
+                    }
+                    uint16_t u = 0;
+                    for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i);
+                    int grid_index = kmap_q2xs[u];
+                    is_on_grid_aux[k] = true;
+                    if (grid_index < 0) {
+                        is_on_grid_aux[k] = false;
+                        const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                        grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k);
+                    }
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 16; ++i) {
+                    float w = weight[i];
+                    float q = 2*Laux[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+                    scale = sumqx/sumq2; best = scale*sumqx;
+                    for (int i = 0; i < 16; ++i) L[i] = Laux[i];
+                    for (int k = 0; k <  2; ++k) is_on_grid[k] = is_on_grid_aux[k];
+                }
+            }
+            int n_not_ongrid = 0;
+            for (int k = 0; k < 2; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
+            if (n_not_ongrid > 0 && scale > 0) {
+                float id = 1/scale;
+                for (int k = 0; k < 2; ++k) {
+                    if (is_on_grid[k]) continue;
+                    uint16_t u = 0;
+                    for (int i = 0; i < 8; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+                        l = MAX(0, MIN(kMaxQ-1, l));
+                        u |= (l << 2*i);
+                        L[8*k + i] = l;
+                    }
+                    int grid_index = kmap_q2xs[u];
+                    if (grid_index < 0) {
+                        const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                        grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, scale, L + 8*k);
+                    }
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 16; ++i) {
+                    float w = weight[i];
+                    float q = 2*L[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0) scale = sumqx/sumq2;
+            }
+            if (scale < 0) {
+                scale = -scale;
+                for (int k = 0; k < 2; ++k) block_signs[k] = (~block_signs[k]) & 127;
+            }
+            for (int k = 0; k < 2; ++k) {
+                uint16_t u = 0;
+                for (int i = 0; i < 8; ++i) u |= (L[8*k+i] << 2*i);
+                int grid_index = kmap_q2xs[u];
+                if (grid_index < 0) {
+                    printf("Oops: found point %u not on grid:", u);
+                    for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
+                    printf("\n");
+                    GGML_ASSERT(false);
+                }
+                q2[2*ib+k] = grid_index | (block_signs[k] << 9);
+            }
+            GGML_ASSERT(scale >= 0);
+            scales[ib] = scale;
+            max_scale = MAX(max_scale, scale);
+        }
+
+        if (!max_scale) {
+            memset(y[ibl].qs, 0, QK_K/4);
+            continue;
+        }
+
+        float d = max_scale/31;
+        y[ibl].d = GGML_FP32_TO_FP16(d);
+        float id = 1/d;
+        for (int ib = 0; ib < QK_K/16; ++ib) {
+            int l = nearest_int(0.5f*(id*scales[ib]-1));
+            l = MAX(0, MIN(15, l));
+            if (ib%2 == 0) y[ibl].scales[ib/2] = l;
+            else y[ibl].scales[ib/2] |= (l << 4);
+        }
+        memcpy(y[ibl].qs, q2, QK_K/4);
+
+    }
+}
+
+size_t quantize_iq2_xxs(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
+    (void)hist;
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int nblock = n_per_row/QK_K;
+    char * qrow = (char *)dst;
+    for (int row = 0; row < nrow; ++row) {
+        quantize_row_iq2_xxs_impl(src, qrow, n_per_row, quant_weights);
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq2_xxs);
+    }
+    return nrow * nblock * sizeof(block_iq2_xxs);
+}
+
+size_t quantize_iq2_xs(const float * src, void * dst, int nrow, int n_per_row, int64_t * hist, const float * quant_weights) {
+    (void)hist;
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int nblock = n_per_row/QK_K;
+    char * qrow = (char *)dst;
+    for (int row = 0; row < nrow; ++row) {
+        quantize_row_iq2_xs_impl(src, qrow, n_per_row, quant_weights);
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq2_xs);
+    }
+    return nrow * nblock * sizeof(block_iq2_xs);
+}
+

ggml-quants.h

@@ -196,8 +196,6 @@ void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict
 void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
 void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
 void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
-void quantize_row_iq2_xxs_reference(const float * restrict x, block_iq2_xxs * restrict y, int k);
-void quantize_row_iq2_xs_reference (const float * restrict x, block_iq2_xs  * restrict y, int k);
 
 void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
 void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
@@ -212,8 +210,6 @@ void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
 void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_iq2_xxs(const float * restrict x, void * restrict y, int k);
-void quantize_row_iq2_xs (const float * restrict x, void * restrict y, int k);
 
 // Dequantization
 void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
@@ -246,3 +242,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx,
 void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_iq2_xs_q8_K (int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+
+//
+// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
+//
+size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q2_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+

ggml.c

@@ -585,8 +585,8 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq2_xxs),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xxs,
-        .from_float               = quantize_row_iq2_xxs,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_iq2_xxs_reference,
+        .from_float               = NULL,
+        .from_float_reference     = NULL,
         .vec_dot                  = ggml_vec_dot_iq2_xxs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
     },
@@ -596,8 +596,8 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_iq2_xs),
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xs,
-        .from_float               = quantize_row_iq2_xs,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_iq2_xs_reference,
+        .from_float               = NULL,
+        .from_float_reference     = NULL,
         .vec_dot                  = ggml_vec_dot_iq2_xs_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
     },
@@ -11638,6 +11638,21 @@ static float ggml_rope_yarn_corr_dim(int n_dims, int n_orig_ctx, float n_rot, fl
     return n_dims * logf(n_orig_ctx / (n_rot * 2 * (float)M_PI)) / (2 * logf(base));
 }
 
+static void ggml_rope_cache_init(
+     float theta_base, float freq_scale, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
+     float * cache, float sin_sign, float theta_scale
+) {
+    float theta = theta_base;
+    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+        rope_yarn(
+            theta, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1]
+        );
+        cache[i0 + 1] *= sin_sign;
+
+        theta *= theta_scale;
+    }
+}
+
 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]
 ) {
@@ -11720,6 +11735,12 @@ static void ggml_compute_forward_rope_f32(
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
             const int64_t p = pos[i2];
+
+            float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
+            if (!is_glm && !is_neox) { // TODO: cache sin/cos for glm, neox
+                ggml_rope_cache_init(p, freq_scale, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
+
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
@@ -11753,18 +11774,13 @@ static void ggml_compute_forward_rope_f32(
                     }
                 } else if (!is_neox) {
                     for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
-                        float cos_theta, sin_theta;
-                        rope_yarn(
-                            theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta
-                        );
-                        sin_theta *= sin_sign;
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
 
                         // zeta scaling for xPos only:
                         float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), p / xpos_base) : 1.0f;
                         if (xpos_down) zeta = 1.0f / zeta;
 
-                        theta_base *= theta_scale;
-
                         const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                               float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
@@ -11888,6 +11904,12 @@ static void ggml_compute_forward_rope_f16(
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
             const int64_t p = pos[i2];
+
+            float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
+            if (!is_glm && !is_neox) { // TODO: cache sin/cos for glm, neox
+                ggml_rope_cache_init(p, freq_scale, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
+
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
@@ -11921,13 +11943,8 @@ static void ggml_compute_forward_rope_f16(
                     }
                 } else if (!is_neox) {
                     for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
-                        float cos_theta, sin_theta;
-                        rope_yarn(
-                            theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta
-                        );
-                        sin_theta *= sin_sign;
-
-                        theta_base *= theta_scale;
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
 
                         const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                               ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -16722,6 +16739,7 @@ struct ggml_cplan ggml_graph_plan(const struct ggml_cgraph * cgraph, int n_threa
                     }
                 } break;
             case GGML_OP_SOFT_MAX:
+            case GGML_OP_ROPE:
                 {
                     cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
                 } break;
@@ -18647,8 +18665,11 @@ size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t *
     return (n/QK8_0*sizeof(block_q8_0));
 }
 
-size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist) {
+size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start,
+        int nrows, int n_per_row, int64_t * hist, const float * imatrix) {
+    (void)imatrix;
     size_t result = 0;
+    int n = nrows * n_per_row;
     switch (type) {
         case GGML_TYPE_Q4_0:
             {
@@ -18683,8 +18704,11 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
         case GGML_TYPE_Q2_K:
             {
                 GGML_ASSERT(start % QK_K == 0);
-                block_q2_K * block = (block_q2_K*)dst + start / QK_K;
-                result = ggml_quantize_q2_K(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_q2_K(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_Q3_K:
             {
@@ -18713,14 +18737,22 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
         case GGML_TYPE_IQ2_XXS:
             {
                 GGML_ASSERT(start % QK_K == 0);
-                block_iq2_xxs * block = (block_iq2_xxs*)dst + start / QK_K;
-                result = ggml_quantize_iq2_xxs(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                GGML_ASSERT(imatrix);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_iq2_xxs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_IQ2_XS:
             {
                 GGML_ASSERT(start % QK_K == 0);
-                block_iq2_xs * block = (block_iq2_xs*)dst + start / QK_K;
-                result = ggml_quantize_iq2_xs(src + start, block, n, n, hist);
+                GGML_ASSERT(start % n_per_row == 0);
+                GGML_ASSERT(imatrix);
+                size_t start_row = start / n_per_row;
+                size_t row_size = ggml_row_size(type, n_per_row);
+                result = quantize_iq2_xs(src + start, (char *)dst + start_row * row_size, nrows, n_per_row, hist, imatrix);
+                GGML_ASSERT(result == row_size * nrows);
             } break;
         case GGML_TYPE_F16:
             {
@@ -19184,7 +19216,7 @@ void gguf_free(struct gguf_context * ctx) {
 
     if (ctx->kv) {
         // free string memory - not great..
-        for (uint32_t i = 0; i < ctx->header.n_kv; ++i) {
+        for (uint64_t i = 0; i < ctx->header.n_kv; ++i) {
             struct gguf_kv * kv = &ctx->kv[i];
 
             if (kv->key.data) {
@@ -19200,7 +19232,7 @@ void gguf_free(struct gguf_context * ctx) {
             if (kv->type == GGUF_TYPE_ARRAY) {
                 if (kv->value.arr.data) {
                     if (kv->value.arr.type == GGUF_TYPE_STRING) {
-                        for (uint32_t j = 0; j < kv->value.arr.n; ++j) {
+                        for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
                             struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[j];
                             if (str->data) {
                                 free(str->data);
@@ -19216,7 +19248,7 @@ void gguf_free(struct gguf_context * ctx) {
     }
 
     if (ctx->infos) {
-        for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) {
+        for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
             struct gguf_tensor_info * info = &ctx->infos[i];
 
             if (info->name.data) {

ggml.h

@@ -2067,10 +2067,13 @@ extern "C" {
     GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
     GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
-    GGML_API size_t ggml_quantize_iq2_xxs(const float * src, void * dst, int n, int k, int64_t * hist);
-    GGML_API size_t ggml_quantize_iq2_xs (const float * src, void * dst, int n, int k, int64_t * hist);
 
-    GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);
+    GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst,
+            int start, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+
+    // These are needed for IQ2_XS and IQ2_XXS quantizations
+    GGML_API void ggml_init_iq2_quantization(enum ggml_type type);
+    GGML_API void ggml_deinit_iq2_quantization(enum ggml_type type);
 
     //
     // Importance matrix

gguf-py/gguf/constants.py

@@ -389,6 +389,9 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.OUTPUT,
         MODEL_TENSOR.ATTN_NORM,
         MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
         MODEL_TENSOR.ATTN_OUT,
         MODEL_TENSOR.FFN_NORM,
         MODEL_TENSOR.FFN_DOWN,

gguf-py/gguf/tensor_mapping.py

@@ -191,6 +191,7 @@ class TensorNameMap:
             "transformer.h.{bid}.mlp.w1",                             # qwen
             "h.{bid}.mlp.c_fc",                                       # gpt2
             "transformer.h.{bid}.mlp.fc1",                            # phi2
+            "model.layers.{bid}.mlp.fc1",                             # phi2
             "model.layers.layers.{bid}.mlp.up_proj",                  # plamo
         ),
 
@@ -232,6 +233,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.dense_4h_to_h",                   # persimmon
             "h.{bid}.mlp.c_proj",                                     # gpt2
             "transformer.h.{bid}.mlp.fc2",                            # phi2
+            "model.layers.{bid}.mlp.fc2",                             # phi2
             "model.layers.layers.{bid}.mlp.down_proj",                # plamo
         ),
 

llama.cpp

@@ -574,6 +574,9 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
             { LLM_TENSOR_OUTPUT,          "output" },
             { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
             { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_Q,          "blk.%d.attn_q" },
+            { LLM_TENSOR_ATTN_K,          "blk.%d.attn_k" },
+            { LLM_TENSOR_ATTN_V,          "blk.%d.attn_v" },
             { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
             { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
             { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
@@ -984,6 +987,7 @@ struct llama_mmap {
         }
 
         if (prefetch > 0) {
+#if _WIN32_WINNT >= 0x602
             // PrefetchVirtualMemory is only present on Windows 8 and above, so we dynamically load it
             BOOL (WINAPI *pPrefetchVirtualMemory) (HANDLE, ULONG_PTR, PWIN32_MEMORY_RANGE_ENTRY, ULONG);
             HMODULE hKernel32 = GetModuleHandleW(L"kernel32.dll");
@@ -1001,6 +1005,9 @@ struct llama_mmap {
                             llama_format_win_err(GetLastError()).c_str());
                 }
             }
+#else
+            throw std::runtime_error("PrefetchVirtualMemory unavailable");
+#endif
         }
     }
 
@@ -1263,7 +1270,7 @@ static ggml_backend_buffer_type_t llama_default_buffer_type_split(int fallback_g
 struct llama_state {
     llama_state() {
 #ifdef GGML_USE_METAL
-        ggml_metal_log_set_callback(log_callback, log_callback_user_data);
+        ggml_backend_metal_log_set_callback(log_callback, log_callback_user_data);
 #endif
     }
 
@@ -3676,8 +3683,19 @@ static bool llm_load_tensors(
                         layer.attn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
                         layer.attn_norm_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd});
 
-                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa});
-                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa});
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, false);
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa}, false);
+
+                        if (layer.wqkv == nullptr) {
+                            layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
+                            layer.bq = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_Q, "bias", i),   {n_embd});
+
+                            layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
+                            layer.bk = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_K, "bias", i),   {n_embd_gqa});
+
+                            layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
+                            layer.bv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_V, "bias", i),   {n_embd_gqa});
+                        }
 
                         layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
                         layer.bo   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd});
@@ -5637,15 +5655,25 @@ struct llm_build_context {
 
             // self-attention
             {
-                cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, attn_norm_output);
-                cb(cur, "wqkv", il);
+                struct ggml_tensor * Qcur = nullptr;
+                struct ggml_tensor * Kcur = nullptr;
+                struct ggml_tensor * Vcur = nullptr;
 
-                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
-                cb(cur, "bqkv", il);
+                if (model.layers[il].wqkv) {
+                    cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, attn_norm_output);
+                    cb(cur, "wqkv", il);
 
-                struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
-                struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                    cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                    cb(cur, "bqkv", il);
+
+                    Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                    Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                    Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+                } else {
+                    Qcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq);
+                    Kcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk);
+                    Vcur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv);
+                }
 
                 cb(Qcur, "Qcur", il);
                 cb(Kcur, "Kcur", il);
@@ -8405,9 +8433,23 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         if (arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
             new_type = GGML_TYPE_Q8_0;
         }
+        else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
+            new_type = GGML_TYPE_Q5_K;
+        }
         else if (new_type != GGML_TYPE_Q8_0) {
             new_type = GGML_TYPE_Q6_K;
         }
+    } else if (ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS || ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS) {
+        if (name.find("attn_v.weight") != std::string::npos) {
+            if (qs.model.hparams.n_gqa() >= 4 || qs.model.hparams.n_expert >= 4) new_type = GGML_TYPE_Q4_K;
+            else new_type = GGML_TYPE_Q2_K;
+            ++qs.i_attention_wv;
+        }
+        else if (name.find("ffn_down") != std::string::npos) {
+            if (qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) new_type = GGML_TYPE_Q2_K;
+            ++qs.i_feed_forward_w2;
+        }
+        else if (name == "token_embd.weight") new_type = GGML_TYPE_Q2_K;
     } else if (name.find("attn_v.weight") != std::string::npos) {
         if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
@@ -8465,9 +8507,16 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         ++qs.i_feed_forward_w2;
     } else if (name.find("attn_output.weight") != std::string::npos) {
         if (arch != LLM_ARCH_FALCON) {
-            if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K  ) new_type = GGML_TYPE_Q3_K;
-            else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
-            else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+            if (qs.model.hparams.n_expert == 8) {
+                if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K   || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M ||
+                    ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
+                    new_type = GGML_TYPE_Q5_K;
+                }
+            } else {
+                if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K  ) new_type = GGML_TYPE_Q3_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
+                else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
+            }
         } else {
             if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
         }
@@ -8570,6 +8619,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     if (params->only_copy) {
         ftype = model.ftype;
     }
+    const std::unordered_map<std::string, std::vector<float>> * imatrix_data = nullptr;
+    if (params->imatrix) {
+        imatrix_data = static_cast<const std::unordered_map<std::string, std::vector<float>>*>(params->imatrix);
+        if (imatrix_data) {
+            printf("================================ Have weights data with %d entries\n",int(imatrix_data->size()));
+        }
+    }
 
     const size_t align = GGUF_DEFAULT_ALIGNMENT;
     struct gguf_context * ctx_out = gguf_init_empty();
@@ -8627,6 +8683,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     // placeholder for the meta data
     ::zeros(fout, meta_size);
 
+    std::set<ggml_type> used_iq2;
+
     for (int i = 0; i < ml.n_tensors; ++i) {
         struct ggml_tensor * tensor = ml.get_tensor_meta(i);
 
@@ -8679,6 +8737,35 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         } else {
             const size_t nelements = ggml_nelements(tensor);
 
+            if ((new_type == GGML_TYPE_IQ2_XXS || new_type == GGML_TYPE_IQ2_XS) && used_iq2.find(new_type) == used_iq2.end()) {
+                ggml_init_iq2_quantization(new_type);
+                used_iq2.insert(new_type);
+            }
+
+            const float * imatrix = nullptr;
+            if (imatrix_data) {
+                auto it = imatrix_data->find(tensor->name);
+                if (it == imatrix_data->end()) {
+                    printf("\n====== %s: did not find weights for %s\n", __func__, tensor->name);
+                } else {
+                    if (it->second.size() == (size_t)tensor->ne[0]) {
+                        imatrix = it->second.data();
+                    } else {
+                        printf("\n====== %s: imatrix size %d is different from tensor size %d for %s\n", __func__,
+                                int(it->second.size()), int(tensor->ne[0]), tensor->name);
+                    }
+                }
+            }
+            if ((new_type == GGML_TYPE_IQ2_XXS ||
+                 new_type == GGML_TYPE_IQ2_XS  ||
+                (new_type == GGML_TYPE_Q2_K && params->ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S && strcmp(tensor->name, "token_embd.weight") != 0)) && !imatrix) {
+                fprintf(stderr, "\n\n============================================================\n");
+                fprintf(stderr, "Missing importance matrix for tensor %s in a very low-bit quantization\n", tensor->name);
+                fprintf(stderr, "The result will be garbage, so bailing out\n");
+                fprintf(stderr, "============================================================\n\n");
+                throw std::runtime_error(format("Missing importance matrix for tensor %s in a very low-bit quantization", tensor->name));
+            }
+
             float * f32_data;
 
             if (tensor->type == GGML_TYPE_F32) {
@@ -8699,21 +8786,28 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             new_data = work.data();
             std::array<int64_t, 1 << 4> hist_cur = {};
 
-            static const int chunk_size = 32 * 512;
+            const int n_per_row = tensor->ne[0];
+            const int nrows = nelements / n_per_row;
+
+            static const int min_chunk_size = 32 * 512;
+            const int chunk_size = n_per_row >= min_chunk_size ? n_per_row : n_per_row * ((min_chunk_size + n_per_row - 1)/n_per_row);
+
             const int nchunk = (nelements + chunk_size - 1)/chunk_size;
             const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1;
             if (nthread_use < 2) {
-                new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nelements, hist_cur.data());
+                new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, hist_cur.data(), imatrix);
             } else {
-                size_t counter = 0;
+                int counter = 0;
                 new_size = 0;
-                auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, nelements]() {
+                auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, chunk_size,
+                     nrows, n_per_row, imatrix]() {
                     std::array<int64_t, 1 << 4> local_hist = {};
+                    const int nrows_per_chunk = chunk_size / n_per_row;
                     size_t local_size = 0;
                     while (true) {
                         std::unique_lock<std::mutex> lock(mutex);
-                        size_t first = counter; counter += chunk_size;
-                        if (first >= nelements) {
+                        int first_row = counter; counter += nrows_per_chunk;
+                        if (first_row >= nrows) {
                             if (local_size > 0) {
                                 for (int j=0; j<int(local_hist.size()); ++j) {
                                     hist_cur[j] += local_hist[j];
@@ -8723,8 +8817,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                             break;
                         }
                         lock.unlock();
-                        size_t last = std::min(nelements, first + chunk_size);
-                        local_size += ggml_quantize_chunk(new_type, f32_data, new_data, first, last - first, local_hist.data());
+                        const int this_nrow = std::min(nrows - first_row, nrows_per_chunk);
+                        local_size += ggml_quantize_chunk(new_type, f32_data, new_data,
+                                first_row * n_per_row, this_nrow, n_per_row, local_hist.data(), imatrix);
                     }
                 };
                 for (int it = 0; it < nthread_use - 1; ++it) {
@@ -8735,7 +8830,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
                 workers.clear();
             }
 
-            LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB | hist: ", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
+            LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
             int64_t tot_count = 0;
             for (size_t i = 0; i < hist_cur.size(); i++) {
                 hist_all[i] += hist_cur[i];
@@ -8743,6 +8838,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             }
 
             if (tot_count > 0) {
+                LLAMA_LOG_INFO(" | hist: ");
                 for (size_t i = 0; i < hist_cur.size(); i++) {
                     LLAMA_LOG_INFO("%5.3f ", hist_cur[i] / float(nelements));
                 }
@@ -8771,6 +8867,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
     fout.close();
 
+    for (auto type : used_iq2) {
+        ggml_deinit_iq2_quantization(type);
+    }
+
     gguf_free(ctx_out);
 
     LLAMA_LOG_INFO("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
@@ -9135,6 +9235,7 @@ struct llama_model_quantize_params llama_model_quantize_default_params() {
         /*.quantize_output_tensor      =*/ true,
         /*.only_copy                   =*/ false,
         /*.pure                        =*/ false,
+        /*.imatrix                     =*/ nullptr,
     };
 
     return result;
@@ -9355,12 +9456,8 @@ struct llama_context * llama_new_context_with_model(
                 ggml_type_name(type_v), (float)memory_size_v / (1024.0f * 1024.0f));
         }
 
-        // resized during inference
-        if (params.logits_all) {
-            ctx->logits.reserve(cparams.n_ctx*hparams.n_vocab);
-        } else {
-            ctx->logits.reserve(hparams.n_vocab);
-        }
+        // resized during inference, reserve maximum
+        ctx->logits.reserve(hparams.n_vocab*cparams.n_batch);
 
         if (params.embedding){
             ctx->embedding.resize(hparams.n_embd);
@@ -9707,8 +9804,8 @@ size_t llama_get_state_size(const struct llama_context * ctx) {
     // for reference, std::mt19937(1337) serializes to 6701 bytes.
     const size_t s_rng_size        = sizeof(size_t);
     const size_t s_rng             = LLAMA_MAX_RNG_STATE;
-    const size_t s_logits_capacity = sizeof(size_t);
     const size_t s_logits_size     = sizeof(size_t);
+    // assume worst case for logits although only currently set ones are serialized
     const size_t s_logits          = ctx->logits.capacity() * sizeof(float);
     const size_t s_embedding_size  = sizeof(size_t);
     const size_t s_embedding       = ctx->embedding.size() * sizeof(float);
@@ -9719,7 +9816,6 @@ size_t llama_get_state_size(const struct llama_context * ctx) {
     const size_t s_total = (
         + s_rng_size
         + s_rng
-        + s_logits_capacity
         + s_logits_size
         + s_logits
         + s_embedding_size
@@ -9788,37 +9884,27 @@ struct llama_data_file_context : llama_data_context {
 static void llama_copy_state_data_internal(struct llama_context * ctx, llama_data_context * data_ctx) {
     // copy rng
     {
-        std::stringstream rng_ss;
+        std::ostringstream rng_ss;
         rng_ss << ctx->rng;
 
-        const size_t rng_size = rng_ss.str().size();
-        char rng_buf[LLAMA_MAX_RNG_STATE];
+        const std::string & rng_str = rng_ss.str();
+        const size_t        rng_size = rng_str.size();
 
-        memset(&rng_buf[0], 0, LLAMA_MAX_RNG_STATE);
-        memcpy(&rng_buf[0], rng_ss.str().data(), rng_ss.str().size());
+        GGML_ASSERT(rng_size <= LLAMA_MAX_RNG_STATE);
 
-        data_ctx->write(&rng_size,   sizeof(rng_size));
-        data_ctx->write(&rng_buf[0], LLAMA_MAX_RNG_STATE);
+        data_ctx->write(&rng_size,      sizeof(rng_size));
+        data_ctx->write(rng_str.data(), rng_size);
     }
 
     // copy logits
     {
-        const size_t logits_cap  = ctx->logits.capacity();
         const size_t logits_size = ctx->logits.size();
 
-        data_ctx->write(&logits_cap,  sizeof(logits_cap));
         data_ctx->write(&logits_size, sizeof(logits_size));
 
         if (logits_size) {
             data_ctx->write(ctx->logits.data(), logits_size * sizeof(float));
         }
-
-        // If there is a gap between the size and the capacity, write padding
-        size_t padding_size = (logits_cap - logits_size) * sizeof(float);
-        if (padding_size > 0) {
-            std::vector<uint8_t> padding(padding_size, 0); // Create a buffer filled with zeros
-            data_ctx->write(padding.data(), padding_size);
-        }
     }
 
     // copy embeddings
@@ -9901,13 +9987,13 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
     // set rng
     {
         size_t rng_size;
-        char   rng_buf[LLAMA_MAX_RNG_STATE];
+        memcpy(&rng_size, inp, sizeof(rng_size)); inp += sizeof(rng_size);
 
-        memcpy(&rng_size,   inp, sizeof(rng_size));    inp += sizeof(rng_size);
-        memcpy(&rng_buf[0], inp, LLAMA_MAX_RNG_STATE); inp += LLAMA_MAX_RNG_STATE;
+        GGML_ASSERT(rng_size <= LLAMA_MAX_RNG_STATE);
 
-        std::stringstream rng_ss;
-        rng_ss.str(std::string(&rng_buf[0], rng_size));
+        std::string rng_str((char *)inp, rng_size); inp += rng_size;
+
+        std::istringstream rng_ss(rng_str);
         rng_ss >> ctx->rng;
 
         GGML_ASSERT(!rng_ss.fail());
@@ -9915,20 +10001,18 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
 
     // set logits
     {
-        size_t logits_cap;
         size_t logits_size;
 
-        memcpy(&logits_cap,  inp, sizeof(logits_cap));  inp += sizeof(logits_cap);
         memcpy(&logits_size, inp, sizeof(logits_size)); inp += sizeof(logits_size);
 
-        GGML_ASSERT(ctx->logits.capacity() == logits_cap);
+        GGML_ASSERT(ctx->logits.capacity() >= logits_size);
 
         if (logits_size) {
             ctx->logits.resize(logits_size);
-            memcpy(ctx->logits.data(), inp, logits_size * sizeof(float));
-        }
 
-        inp += logits_cap * sizeof(float);
+            memcpy(ctx->logits.data(), inp, logits_size * sizeof(float));
+            inp += logits_size * sizeof(float);
+        }
     }
 
     // set embeddings
@@ -10298,6 +10382,8 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token
     if (0 <= token && token < llama_n_vocab(model)) {
         switch (llama_vocab_get_type(model->vocab)) {
         case LLAMA_VOCAB_TYPE_SPM: {
+            // NOTE: we accept all unsupported token types,
+            // suppressing them like CONTROL tokens.
             if (llama_is_normal_token(model->vocab, token)) {
                 std::string result = model->vocab.id_to_token[token].text;
                 llama_unescape_whitespace(result);
@@ -10306,6 +10392,13 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token
                 }
                 memcpy(buf, result.c_str(), result.length());
                 return result.length();
+            } else if (llama_is_user_defined_token(model->vocab, token)) {
+                std::string result = model->vocab.id_to_token[token].text;
+                if (length < (int) result.length()) {
+                    return -result.length();
+                }
+                memcpy(buf, result.c_str(), result.length());
+                return result.length();
             } else if (llama_is_unknown_token(model->vocab, token)) { // NOLINT
                 if (length < 3) {
                     return -3;
@@ -10320,14 +10413,12 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token
                 }
                 buf[0] = llama_token_to_byte(model->vocab, token);
                 return 1;
-            } else {
-                // TODO: for now we accept all unsupported token types,
-                // suppressing them like CONTROL tokens.
-                // GGML_ASSERT(false);
             }
             break;
         }
         case LLAMA_VOCAB_TYPE_BPE: {
+            // NOTE: we accept all unsupported token types,
+            // suppressing them like CONTROL tokens.
             if (llama_is_normal_token(model->vocab, token)) {
                 std::string result = model->vocab.id_to_token[token].text;
                 result = llama_decode_text(result);
@@ -10336,12 +10427,15 @@ int32_t llama_token_to_piece(const struct llama_model * model, llama_token token
                 }
                 memcpy(buf, result.c_str(), result.length());
                 return result.length();
+            } else if (llama_is_user_defined_token(model->vocab, token)) {
+                std::string result = model->vocab.id_to_token[token].text;
+                if (length < (int) result.length()) {
+                    return -result.length();
+                }
+                memcpy(buf, result.c_str(), result.length());
+                return result.length();
             } else if (llama_is_control_token(model->vocab, token)) {
                 ;
-            } else {
-                // TODO: for now we accept all unsupported token types,
-                // suppressing them like CONTROL tokens.
-                // GGML_ASSERT(false);
             }
             break;
         }
@@ -10453,7 +10547,7 @@ void llama_log_set(ggml_log_callback log_callback, void * user_data) {
     g_state.log_callback = log_callback ? log_callback : llama_log_callback_default;
     g_state.log_callback_user_data = user_data;
 #ifdef GGML_USE_METAL
-    ggml_metal_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
+    ggml_backend_metal_log_set_callback(g_state.log_callback, g_state.log_callback_user_data);
 #endif
 }
 

llama.h

@@ -43,7 +43,7 @@
 #define LLAMA_FILE_MAGIC_GGSN 0x6767736eu // 'ggsn'
 
 #define LLAMA_SESSION_MAGIC   LLAMA_FILE_MAGIC_GGSN
-#define LLAMA_SESSION_VERSION 3
+#define LLAMA_SESSION_VERSION 4
 
 #if defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST) || defined(GGML_USE_METAL)
 // Defined when llama.cpp is compiled with support for offloading model layers to GPU.
@@ -249,6 +249,7 @@ extern "C" {
         bool quantize_output_tensor; // quantize output.weight
         bool only_copy;              // only copy tensors - ftype, allow_requantize and quantize_output_tensor are ignored
         bool pure;                   // disable k-quant mixtures and quantize all tensors to the same type
+        void * imatrix;              // pointer to importance matrix data
     } llama_model_quantize_params;
 
     // grammar types

scripts/compare-llama-bench.py

@@ -10,15 +10,15 @@ import sqlite3
 try:
     import git
     from tabulate import tabulate
-except ImportError:
+except ImportError as e:
     print("ERROR: the following Python libraries are required: GitPython, tabulate.")
-    sys.exit(1)
+    raise e
 
 # Properties by which to differentiate results per commit:
 KEY_PROPERTIES = [
-    "cuda", "opencl", "metal", "gpu_blas", "blas", "cpu_info", "gpu_info", "model_filename",
-    "model_type", "model_size", "model_n_params", "n_batch", "n_threads", "type_k", "type_v",
-    "n_gpu_layers", "main_gpu", "no_kv_offload", "mul_mat_q", "tensor_split", "n_prompt", "n_gen"
+    "cpu_info", "gpu_info", "n_gpu_layers", "main_gpu", "cuda", "opencl", "metal", "gpu_blas",
+    "blas", "model_filename", "model_type", "model_size", "model_n_params", "n_batch", "n_threads",
+    "type_k", "type_v", "no_kv_offload", "mul_mat_q", "tensor_split", "n_prompt", "n_gen"
 ]
 
 # Properties that are boolean and are converted to Yes/No for the table:
@@ -37,6 +37,7 @@ PRETTY_NAMES = {
 DEFAULT_SHOW = ["model_type"]  # Always show these properties by default.
 DEFAULT_HIDE = ["model_filename"]  # Always hide these properties by default.
 GPU_NAME_STRIP = ["NVIDIA GeForce ", "Tesla ", "AMD Radeon "]  # Strip prefixes for smaller tables.
+MODEL_SUFFIX_REPLACE = {" - Small": "_S", " - Medium": "_M", " - Large": "_L"}
 
 DESCRIPTION = """Creates tables from llama-bench data written to an SQLite database. Example usage (Linux):
 
@@ -308,8 +309,13 @@ else:
         if gpu_blas and "gpu_info" not in properties_different:
             show.append("gpu_info")
 
-    show += DEFAULT_SHOW
     show += properties_different
+
+    index_default = 0
+    for prop in ["cpu_info", "gpu_info", "n_gpu_layers", "main_gpu"]:
+        if prop in show:
+            index_default += 1
+    show = show[:index_default] + DEFAULT_SHOW + show[index_default:]
     for prop in DEFAULT_HIDE:
         try:
             show.remove(prop)
@@ -334,6 +340,12 @@ for bool_property in BOOL_PROPERTIES:
         for row_table in table:
             row_table[ip] = "Yes" if int(row_table[ip]) == 1 else "No"
 
+if "model_type" in show:
+    ip = show.index("model_type")
+    for (old, new) in MODEL_SUFFIX_REPLACE.items():
+        for row_table in table:
+            row_table[ip] = row_table[ip].replace(old, new)
+
 if "model_size" in show:
     ip = show.index("model_size")
     for row_table in table:
@@ -341,10 +353,16 @@ if "model_size" in show:
 
 if "gpu_info" in show:
     ip = show.index("gpu_info")
-    for gns in GPU_NAME_STRIP:
-        for row_table in table:
+    for row_table in table:
+        for gns in GPU_NAME_STRIP:
             row_table[ip] = row_table[ip].replace(gns, "")
 
+        gpu_names = row_table[ip].split("/")
+        num_gpus = len(gpu_names)
+        all_names_the_same = len(set(gpu_names)) == 1
+        if len(gpu_names) >= 2 and all_names_the_same:
+            row_table[ip] = f"{num_gpus}x {gpu_names[0]}"
+
 headers  = [PRETTY_NAMES[p] for p in show]
 headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"]
 

scripts/sync-ggml.last

@@ -1 +1 @@
-400c07f00508e6f60fb25405444b5669c365b0a9
+1890780da4ea10db88736fcde85f285abf6c64b0

tests/test-backend-ops.cpp

@@ -56,7 +56,7 @@ static void init_tensor_uniform(ggml_tensor * tensor, float min = -1.0f, float m
         GGML_ASSERT(size % ggml_blck_size(tensor->type) == 0);
         std::vector<uint8_t> dataq(ggml_row_size(tensor->type, size));
         int64_t hist[16];
-        ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size, hist);
+        ggml_quantize_chunk(tensor->type, data.data(), dataq.data(), 0, size/tensor->ne[0], tensor->ne[0], hist, nullptr);
         ggml_backend_tensor_set(tensor, dataq.data(), 0, dataq.size());
     } else if (tensor->type == GGML_TYPE_I8 || tensor->type == GGML_TYPE_I16 || tensor->type == GGML_TYPE_I32) {
         // This is going to create some weird integers though.